What We’re Reading (Week Ending 17 March 2024)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general.

Here are the articles for the week ending 17 March 2024:

1. The Ultra-Pure, Super-Secret Sand That Makes Your Phone Possible – Vince Beiser

Spruce Pine is not a wealthy place. Its downtown consists of a somnambulant train station across the street from a couple of blocks of two‑story brick buildings, including a long‑closed movie theater and several empty storefronts.

The wooded mountains surrounding it, though, are rich in all kinds of desirable rocks, some valued for their industrial uses, some for their pure prettiness. But it’s the mineral in Glover’s bag—snowy white grains, soft as powdered sugar—that is by far the most important these days. It’s quartz, but not just any quartz. Spruce Pine, it turns out, is the source of the purest natural quartz—a species of pristine sand—ever found on Earth. This ultra‑elite deposit of silicon dioxide particles plays a key role in manufacturing the silicon used to make computer chips. In fact, there’s an excellent chance the chip that makes your laptop or cell phone work was made using sand from this obscure Appalachian backwater. “It’s a billion‑dollar industry here,” Glover says with a hooting laugh. “Can’t tell by driving through here. You’d never know it.”

In the 21st century, sand has become more important than ever, and in more ways than ever. This is the digital age, in which the jobs we work at, the entertainment we divert ourselves with, and the ways we communicate with one another are increasingly defined by the internet and the computers, tablets, and cell phones that connect us to it. None of this would be possible were it not for sand.

Most of the world’s sand grains are composed of quartz, which is a form of silicon dioxide, also known as silica. High‑purity silicon dioxide particles are the essential raw materials from which we make computer chips, fiber‑optic cables, and other high‑tech hardware—the physical components on which the virtual world runs. The quantity of quartz used for these products is minuscule compared to the mountains of it used for concrete or land reclamation. But its impact is immeasurable…

…In the mid‑1950s, thousands of miles from North Carolina, a group of engineers in California began working on an invention that would become the foundation of the computer industry. William Shockley, a pathbreaking engineer at Bell Labs who had helped invent the transistor, had left to set up his own company in Mountain View, California, a sleepy town about an hour south of San Francisco, near where he had grown up. Stanford University was nearby, and General Electric and IBM had facilities in the area, as well as a new company called Hewlett‑Packard. But the area known at the time as the Santa Clara Valley was still mostly filled with apricot, pear, and plum orchards. It would soon become much better known by a new nickname: Silicon Valley.

At the time, the transistor market was heating up fast. Texas Instruments, Motorola, and other companies were all competing to come up with smaller, more efficient transistors to use in, among other products, computers. The first American computer, dubbed ENIAC, was developed by the army during World War II; it was 100 feet long and 10 feet high, and it ran on 18,000 vacuum tubes.

Transistors, which are tiny electronic switches that control the flow of electricity, offered a way to replace those tubes and make these new machines even more powerful while shrinking their tumid footprint. Semiconductors—a small class of elements, including germanium and silicon, which conduct electricity at certain temperatures while blocking it at others—looked like promising materials for making those transistors.

At Shockley’s startup, a flock of young PhDs began each morning by firing up kilns to thousands of degrees and melting down germanium and silicon. Tom Wolfe once described the scene in Esquire magazine: “They wore white lab coats, goggles, and work gloves. When they opened the kiln doors weird streaks of orange and white light went across their faces . . . they lowered a small mechanical column into the goo so that crystals formed on the bottom of the column, and they pulled the crystal out and tried to get a grip on it with tweezers, and put it under microscopes and cut it with diamond cutters, among other things, into minute slices, wafers, chips; there were no names in electronics for these tiny forms.”

Shockley became convinced that silicon was the more promising material and shifted his focus accordingly. “Since he already had the first and most famous semiconductor research and manufacturing company, everyone who had been working with germanium stopped and switched to silicon,” writes Joel Shurkin in his biography of Shockley, Broken Genius. “Indeed, without his decision, we would speak of Germanium Valley.”

Shockley was a genius, but by all accounts he was also a lousy boss. Within a couple of years, several of his most talented engineers had jumped ship to start their own company, which they dubbed Fairchild Semiconductor. One of them was Robert Noyce, a laid‑back but brilliant engineer, only in his mid‑20s but already famous for his expertise with transistors.

The breakthrough came in 1959, when Noyce and his colleagues figured out a way to cram several transistors onto a single fingernail‑sized sliver of high‑purity silicon. At almost the same time, Texas Instruments developed a similar gadget made from germanium. Noyce’s, though, was more efficient, and it soon dominated the market. NASA selected Fairchild’s microchip for use in the space program, and sales soon shot from almost nothing to $130 million a year. In 1968, Noyce left to found his own company. He called it Intel, and it soon dominated the nascent industry of programmable computer chips.

Intel’s first commercial chip, released in 1971, contained 2,250 transistors. Today’s computer chips are often packed with transistors numbering in the billions. Those tiny electronic squares and rectangles are the brains that run our computers, the Internet, and the entire digital world. Google, Amazon, Apple, Microsoft, the computer systems that underpin the work of everything from the Pentagon to your local bank—all of this and much more is based on sand, remade as silicon chips.

Making those chips is a fiendishly complicated process. They require essentially pure silicon. The slightest impurity can throw their tiny systems out of whack.

Finding silicon is easy. It’s one of the most abundant elements on Earth. It shows up practically everywhere bound together with oxygen to form SiO2, aka quartz. The problem is that it never occurs naturally in pure, elemental form. Separating out the silicon takes considerable doing.

Step one is to take high‑purity silica sand, the kind used for glass. (Lump quartz is also sometimes used.) That quartz is then blasted in a powerful electric furnace, creating a chemical reaction that separates out much of the oxygen. That leaves you with what is called silicon metal, which is about 99 percent pure silicon. But that’s not nearly good enough for high‑tech uses. Silicon for solar panels has to be 99.999999 percent pure—six 9s after the decimal. Computer chips are even more demanding. Their silicon needs to be 99.99999999999 percent pure—eleven 9s. “We are talking of one lonely atom of something that is not silicon among billions of silicon companions,” writes geologist Michael Welland in Sand: The Never-Ending Story.

Getting there requires treating the silicon metal with a series of complex chemical processes. The first round of these converts the silicon metal into two compounds. One is silicon tetrachloride, which is the primary ingredient used to make the glass cores of optical fibers. The other is trichlorosilane, which is treated further to become polysilicon, an extremely pure form of silicon that will go on to become the key ingredient in solar cells and computer chips.

Each of these steps might be carried out by more than one company, and the price of the material rises sharply at each step. That first‑step, 99 percent pure silicon metal goes for about $1 a pound; polysilicon can cost 10 times as much.

The next step is to melt down the polysilicon. But you can’t just throw this exquisitely refined material in a cook pot. If the molten silicon comes into contact with even the tiniest amount of the wrong substance, it causes a ruinous chemical reaction. You need crucibles made from the one substance that has both the strength to withstand the heat required to melt polysilicon, and a molecular composition that won’t infect it. That substance is pure quartz.

THIS IS WHERE Spruce Pine quartz comes in. It’s the world’s primary source of the raw material needed to make the fused‑quartz crucibles in which computer‑chip‑grade polysilicon is melted. A fire in 2008 at one of the main quartz facilities in Spruce Pine for a time all but shut off the supply of high‑purity quartz to the world market, sending shivers through the industry.

Today one company dominates production of Spruce Pine quartz. Unimin, an outfit founded in 1970, has gradually bought up Spruce Pine area mines and bought out competitors, until today the company’s North Carolina quartz operations supply most of the world’s high‑ and ultra‑high‑purity quartz. (Unimin itself is now a division of a Belgian mining conglomerate, Sibelco.)

In recent years, another company, the imaginatively titled Quartz Corp, has managed to grab a small share of the Spruce Pine market. There are a very few other places around the world producing high‑purity quartz, and many other places where companies are looking hard for more. But Unimin controls the bulk of the trade.

The quartz for the crucibles, like the silicon they will produce, needs to be almost absolutely pure, purged as thoroughly as possible of other elements. Spruce Pine quartz is highly pure to begin with, and purer still after being put through several rounds of froth flotation. But some of the grains may still have what Glover calls interstitial crystalline contamination—molecules of other minerals attached to the quartz molecules.

That’s frustratingly common. “I’ve evaluated thousands of quartz samples from all over the world,” says John Schlanz, chief minerals processing engineer at the Minerals Research Laboratory in Asheville, about an hour from Spruce Pine. “Near all of them have contaminate locked in the quartz grains that you can’t get out.”

Some Spruce Pine quartz is flawed in this way. Those grains are used for high‑end beach sand and golf course bunkers—most famously the salt‑white traps of Augusta National Golf Club, site of the iconic Masters Tournament. A golf course in the oil‑drunk United Arab Emirates imported 4,000 tons of this sand in 2008 to make sure its sand traps were world‑class, too.

The very best Spruce Pine quartz, however, has an open crystalline structure, which means that hydrofluoric acid can be injected right into the crystal molecules to dissolve any lingering traces of feldspar or iron, taking the purity up another notch. Technicians take it one step further by reacting the quartz with chlorine or hydrochloric acid at high temperatures, then putting it through one or two more trade‑secret steps of physical and chemical processing.

The result is what Unimin markets as Iota quartz, the industry standard of purity. The basic Iota quartz is 99.998 percent pure SiO2. It is used to make things like halogen lamps and photovoltaic cells, but it’s not good enough to make those crucibles in which polysilicon is melted. For that you need Iota 6, or the tip‑top of the line, Iota 8, which clocks in at 99.9992 percent purity—meaning for every one billion molecules of SiO , there are only 80 molecules of impurities. Iota 8 sells for up to $10,000 a ton. Regular construction sand, at the other end of the sand scale, can be had for a few dollars per ton…

…Unimin sells this ultra‑high‑purity quartz sand to companies like General Electric, which melts it, spins it, and fuses it into what looks like a salad bowl made of milky glass: the crucible. “It’s safe to say the vast majority of those crucibles are made from Spruce Pine quartz,” Schlanz says.

The polysilicon is placed in those quartz crucibles, melted down, and set spinning. Then a silicon seed crystal about the size of a pencil is lowered into it, spinning in the opposite direction. The seed crystal is slowly withdrawn, pulling behind it what is now a single giant silicon crystal. These dark, shiny crystals, weighing about 220 pounds, are called ingots.

The ingots are sliced into thin wafers. Some are sold to solar cell manufacturers. Ingots of the highest purity are polished to mirror smoothness and sold to a chipmaker like Intel. It’s a thriving multi-billion dollar industry in 2012.

The chipmaker imprints patterns of transistors on the wafer using a process called photolithography. Copper is implanted to link those billions of transistors to form integrated circuits. Even a minute particle of dust can ruin the chip’s intricate circuitry, so all of this happens in what’s called a clean room, where purifiers keep the air thousands of times cleaner than a hospital operating room. Technicians dress in an all‑covering white uniform affectionately known as a bunny suit. To ensure the wafers don’t get contaminated during manufacture, many of the tools used to move and manipulate them are, like the crucibles, made from high‑purity quartz.

The wafers are then cut into tiny, unbelievably thin quadrangular chips—computer chips, the brains inside your mobile phone or laptop. The whole process requires hundreds of precise, carefully controlled steps. The chip that results is easily one of the most complicated man‑made objects on Earth, yet made with the most common stuff on Earth: humble sand.

The total amount of high‑purity quartz produced worldwide each year is estimated at 30,000 tons—less than the amount of construction sand produced in the United States every hour. (And even construction sand is in high demand; there’s a thriving black market in the stuff.) Only Unimin knows exactly how much Spruce Pine quartz is produced, because it doesn’t publish any production figures. It is an organization famously big on secrecy. “Spruce Pine used to be mom‑and‑ pop operations,” Schlanz says. “When I first worked up there, you could just walk into any of the operations. You could just go across the street and borrow a piece of equipment.”

NOWADAYS UNIMIN WON’T even allow staff of the Minerals Research Laboratory inside the mines or processing facilities. Contractors brought in to do repair work have to sign confidentiality agreements. Whenever possible, vice‑president Richard Zielke recently declared in court papers, the company splits up the work among different contractors so that no individual can learn too much.

Unimin buys equipment and parts from multiple vendors for the same reason. Glover has heard of contractors being blindfolded inside the processing plants until they arrive at the specific area where their jobs are and of an employee who was fired on the spot for bringing someone in without authorization. He says the company doesn’t even allow its employees to socialize with those of their competitors.

It was hard to check out Glover’s stories, because Unimin wouldn’t talk to me. Unlike most big corporations, its website lists no contact for a press spokesperson or public relations representative. Several emails to their general inquiries address went unanswered. When I called the company’s headquarters in Connecticut, the woman who answered the phone seemed mystified by the concept of a journalist wanting to ask questions.

She put me on hold for a few minutes, then came back to tell me the company has no PR department, but that if I faxed (faxed!) her my questions, someone might get back to me. Eventually I got in touch with a Unimin executive who asked me to send her my questions by email. I did so. The response: “Unfortunately, we are not in a position to provide answers at this point in time.”

2. It was never about LLM performance – Justin

The LLM community is obsessed with benchmarking model performance. Mistral released their new “flagship” model this week, and immediately focused the discussion on how it performs on “commonly used benchmarks” relative to other models:

The entire blog post (I’d recommend reading it) is just a read through of how this model performs relative to other models on benchmarks, from math and coding to multilingual capabilities…

…This tendency to fixate on benchmarks is understandable – right now, it’s basically the only semi-objective way to measure how these models stack up against each other. It’s something vendors in other spaces, like data streaming, do too. But it is dangerous because it misses the point of where this whole AI thing is going, and is a textbook product marketing anti-pattern.

In a trend that we’ve seen hundreds of times in developer tooling, the underlying LLM is not going to matter within a few years. Large Language Model performance is already highly commoditized, and will continue to head in that direction. All that will matter is the experience that you build on top of these models, and what that enables for your customers.

Let’s take a look at the ChatGPT interface. Here’s a common prompt I’ve been using for testing, asking the model to summarize the contents of an external link into a tweet thread. Unrelated aside, the responses to this prompt are virtually identical across every major LLM.

Which parts of this interface are the underlying model – GPT-4 in this case – and which are an experience built by OpenAI on top of the underlying model?

The text response, minus any formatting, is what the model generated. But the:

Ability of the model to access and scrape content from a web page
Context of the prompt, including setting the system as a helpful assistant
Formatting the response, like changing the numbers to gray UI for typing the prompt
Filepicker for attaching media to the prompt
Prompt history
Model switcher / picker (this one is meta)
Ability to persist and share the model responses

…and more not show here

are all not GPT-4, they’re features built by OpenAI on top of GPT-4 to create an experience that is helpful and worth paying for. Some of these are harder to build than others – OpenAI’s secret sauce obviously isn’t the little arrow that scrolls down to the bottom of the response. ChatGPT would be nothing without GPT-4 – but the reverse may also be true!

The retort to this line of reasoning is that these chat interfaces are primarily for non-technical users, while the real money for these model providers comes from developer use cases, building LLMs into user-facing applications. I’ve worked closely with one of the major model compute providers, so this is not foreign to me. But experience matters to developers too!

OpenAI has dedicated significant resources to building a seamless developer experience beyond “docs for the model.” Here’s their playground for prompting GPT models – you can adjust parameters like temperature and penalties, plus change the system prompt to be any other style…

…For a closed source model provider like OpenAI, the difference between what is model and what is experience is academic – you’re paying for both. They are one thing. But where this really matters is in open source. Does the convergence of open source performance to closed source performance really matter if the experience of using that open source is bad?…

…The open source discussion has been too anchored on reaching performance parity with OpenAI models. This is a small piece of the puzzle. For developers looking to build applications with these open source models, and especially the pro-sumer chat use case, users need to consider the holistic experience that model providers offer. Integrating LLMs into your app is almost never going to be the “drop in” experience you see on marketing sites – and my concern is that the “open source is approaching parity with OpenAI!” narrative is not actually true in a meaningful way.

Folks working in AI can look to previous examples of this phenomenon in developer tools for guidance: A couple of years ago, I wrote about how underlying performance of production relational databases is becoming commoditized, and vendors are focusing much more on developer experience. It’s going to happen here too, the question is just when.

3. Aravind Srinivas – Building An Answer Engine – Patrick O’Shaughnessy and Aravind Srinivas

Patrick: [00:07:28] It’s really cool to think about the sequencing to get there. We’ve had search engines. Like you said, it’s a hack to get the answers. You’re building what I think of today as an answer engine. I type something in, it’s just giving the answer directly with great citation and all this other stuff we’ll talk about. And the vision you’re articulating is this question engine can anticipate the things that I want to learn about and give them to me beforehand.

And I’d love to build up towards that. So maybe starting with the answer engine, explain to us how it works. Maybe you could do this via the time line of how you’ve built the product or something. But what are the components? What is happening behind the scenes when I type something into Perplexity either a question or a search query or whatever? Walk us through in some detail the actual goings on behind the scenes in terms of how the product works itself?

Aravind: [00:08:13] Yes. So when you type in a question into Perplexity, the first thing that happens is, it first reformulates the question, it tries to understand the question better, expands the question in terms of adding more suffixes or prefixes to it, to make it more well formatted. It speaks to the question engine part. And then after that, it goes and pulls so many links from the web that are relevant to this reformulated question.

There are so many paragraphs in each of those links. It takes only the relevant paragraphs from each of those links. And then an AI model, we typically call it large language model. It’s basically a model that’s been trained to predict the next word on the Internet and fine-tuned for being good at summarization and chats.

That AI model looks at all these chunks of knowledge the bits of study that surface from important or relevant links and takes only those parts that are relevant to answering your query and gives you a very concise four or five sentence answer, but also with references. Every sentence has a reference to which webpage or which chunk of knowledge it took from which webpage and puts it at the top in terms of sources.

That gets you to a nicely formatted rendered answer, sometimes in markdown bullets, or sometimes just generic paragraphs, sometimes it has images in it. But a great answer with references or citation so that if you want to dig deeper, you can go and visit the link. If you don’t want and just read the answer and ask a follow-up, you can engage in a conversation, both modes of usage are encouraged and allowed. So this is what happens on Perplexity today.

Patrick: [00:09:51] What percent of users end up clicking beneath the summarized answer into a source webpage?

Aravind: [00:10:01] At least 10%.

Patrick: [00:10:02] So 90% of the time, they’re just satisfied with what you give them?

Aravind: [00:10:06] It depends on how you look at it. If you wanted to be 100% of the time, people always click on a link, that’s the traditional Google. And you want to be 100% of the time where people never click on links, that’s ChatGPT. We think the sweet spot is somewhere in the middle. People should click on link sometimes to go do their work there. Let’s say, you’re just booking a ticket, you might actually want to go away Expedia or something.

Let’s say you’re deciding where to go first. You don’t need to go away and read all these SEO blogs and get confused on what you want to do. You first make your decision independently with this research body that’s helping you decide. And once you finished your research and you have decided, then that’s when you actually have to go out and do your actual action of booking your ticket. That way, I believe there is a nice sweet spot of one product providing you both the navigational search experience as well as the answer engine experience together. And that’s what we strive to be doing…

…Patrick: [00:13:54] Can you explain from an insider’s perspective and someone building an application on top of these incredible new technologies, what do you think the future might look like or even what you think the ideal future would be for how many different LLM providers there are, how specialized they get to scale the primary answer, so there’s only going to be a few of them. How do you think about all this and where you think it might go?

Aravind: [00:14:16] It really depends on who you’re building for. If you’re building for consumers, you do want to build a scalable infrastructure because you do want to ask many consumers to use your product. If you’re building for the enterprise, you still want a scalable infrastructure.

Now it really depends, are you building for the people within that company who are using your product. Let’s say, you’re building an internal search engine, you only need to scale to the size of the largest organization, which is like maybe 100,000 people. And not all of them will be using your thing at one moment. You’re decentralizing it, you’re going to keep different servers for different companies and you can elastically decide what’s the level of throughput you need to offer.

But then if you’re solving another enterprise’s problem, where that enterprise is serving consumers and you’re helping them do that, you need to build scalable infrastructure indirectly at least. For example, OpenAI. Their APIs are used by us, other people to serve a lot of consumers. So unless they solve that problem themselves, they’re unable to help other people solve their problem. Same thing with AWS.

So that’s one advantage you have of actually having a first-party product that your infrastructure is helping you serve. And by doing that, by forcing yourself to solve that hard problem, whatever you build can be used by others as well. Amazon build AWS first for Amazon. And because Amazon.com requires very robust infrastructure, that can be used by so many other people and so many other companies emerged by building on top of AWS.

Same thing happened with OpenAI. They needed robust infrastructure to serve the GPT-3 developer API and ChatGPT as a product. But once they got it all right, then they can now support other companies that are building on top of them. So it really depends on what’s your end goal and who you’re trying to serve and what’s the scale of our ambition…

…Patrick: [00:19:02] And when I think about the history of the product, which I was a pretty early user of, the first thing that pops to my mind is that it solves the hallucination problem, which has become less of a problem. But early on, everyone just didn’t know how to trust these things and you solved that. You gave citations, you can click through the underlying webpages, et cetera.

I’d love you to walk through what you view the major time line product milestones have been of Perplexity dating back to its start. The one I just gave could be one example. There was this possibility, but there was a problem and you solved it, at least that was my perception as a user. What have been the major milestones as you think back on the product and how it’s gotten better?

Aravind: [00:19:41] I would say the first major thing we did is really making the product a lot faster. When we first launched, the latency for every query was seven seconds, then we actually had to speed up the demo video to put it on Twitter so that it doesn’t look embarrassing.

And one of our early friendly investors, Daniel Gross who co-invests a lot with Nat Friedman, he was one of our first testers before we even released the product. And he said, you guys should call it a submit button for a query. It’s almost like you’re submitting a job and waiting on the cluster to get back. It’s that slow.

And now we are widely regarded as the fastest chatbot out there. Some people even come and ask me, why are you only as fast as ChatGPT? Why are you not faster? And little did they realize that ChatGPT doesn’t even use the web by default. It only uses it on the browsing mode on Bing.

So for us to be as fast as ChatGPT already tells you that in spite of doing more work to go pull up links from the web, read the chunks, pick the relevant ones and use that to give you the answer with sources and a lot more work on the rendering, despite doing all the additional work, if you’re managing an end-to-end latency as good as ChatGPT that shows we have like even a superior back end to them.

So I’m most proud about the speed at which we can do things today compared to when we launched, the accuracy has been constantly going up, primarily few things. One is we keep expanding our index and like keep improving the quality of the index. From the beginning, we knew all the mistakes that previous Google competitors did, which is obsessed about the size of your index and focus less on the quality.

So we decided from the beginning we would not obsess about the size. Size doesn’t matter and index actually, what matters is the quality of your index. What kind of domains are important for AI chatbots and question-answering and knowledge workers. That is what we care about. So that decision ended up being right.

The other thing that has helped us improve the accuracy was training these models to be focused on hallucinations. When you don’t have enough information in the search snippets, try to just say I don’t know, instead of making up things. LLMs are conditioned to always be helpful, will always try to serve the user’s query despite what it has access to, may not be even sufficient to answer the query. So that part took some reprogramming, rewiring. You’ve got to go and change the ways. You can’t just solve this with prompt engineering. So we have spent a lot of work on that.

The other thing I’m really proud about is getting our own inference infrastructure. So when you have to move outside the OpenAI models to serve your product, everybody thinks, “Oh, you just train a model to be as good as GPT and you’re’ done.” But reality is OpenAI’s mode is not just in the fact that they have trained the best models, but also that they have the most cost-efficient, scalable infrastructure for serving this on a large-scale consumer product like ChatGPT. That is itself a separate layer of mode. You can build that mode, you can build.

And so we are very proud of our inference team, how fast, high throughput, low latency infrastructure we built for serving our own LLMs. We took advantage of the open source revolution, Llama and Mistral and took all these models, trained them to be very good at being great answer bots and served them ourselves on GPU so that we get better margins on our product. So all these three layers, both in terms of speed through actual product back-end orchestration, accuracy of the AI models and serving our own AI models, we’ve done a lot of work on all these things…

…Patrick: [00:28:50] Can you expand on index. You’ve referenced that a few times for those that haven’t built one or haven’t thought about this. Just explain that whole concept and the decisions that you’ve made. You already mentioned quality versus size. But just explain what it means to build an index, why it’s so important, et cetera?

Aravind: [00:29:07] Yes. So what does an index mean, it’s basically a copy of the web. The web has so many links and you want a cache, you want a copy of all those links in the database, so a URL and the contents in that URL. Now the challenge here is new links are being created every day on the web and also existing links keep getting updated on the web as well. New sites keep getting updated. So you’ve got to periodically refresh them. The URL needs to be updated in the cache with a different version of it.

Similarly, you got to keep adding new URLs to your index, which means you’ve got to build a crawler. And then how you store a URL, the contents in that URL also matters. Not every page is native HTML anymore. The web has upgraded a lot, rendering JavaScript a lot, and every domain has custom-based rendered the JavaScript. So you’ve got to build parsers. So you’ve got to build a crawler, indexer, parser and that together makes up for a great index.

Now the next step comes to retrieval, which is now that you have those index, every time you hit a query, which links do you use? And which paragraphs in those links do you use? Now that is the ranking problem. How do you figure out what is relevance and ranking? And once you retrieve those chunks, like the top few chunks relevant to a query that the user is asking, that’s when the AI model comes in. So this is a retrieve part. Now the generic part. That’s why it’s called retrieve and generic.

So once you retrieve the relevant chunks from the huge index that you have, the AI model will come and read those chunks and then give you the answer. Doing this ensures that you don’t have to keep training the AI model to be up to date. What you want the AI model to do is to be intelligent, to be a good reasoning model.

Think about this as when you were a student, I’m sure you would have written an open book exam, open notes exam in school or high school or college. What are those exams test you for? They don’t test you for rote learning. So it doesn’t give an advantage to the person who has the best memory power. It gives advantage to person who has read the concepts, can immediately query the right part of the notes, but the questions required you to think on the fly as well.

That’s what we want to design systems. It’s very different philosophy from OpenAI, where OpenAI wants this one model that’s so intelligent, so smart, you can just ask it anything. It’s going to tell you. We rather want to build a small efficient model that’s smart, capable, can reason on facts that it’s given on the fly. And this ambiguate different individuals with different names or saved as not sufficient information, not get confused about dates.

When you’re asking something about the future, say that was not yet happened. These sort of corner cases handle all of those with good reasoning capabilities yet have access to all of the world’s knowledge at an instant through a great index. And if you can do both of these together end-to-end orchestrated with great latency and user experience, you’re creating something extremely valuable. So that’s what we want to build…

…Patrick: [00:37:26] Do you think that the transformer architecture is here to stay and will remain the dominant tool or architecture for a long time?

Aravind: [00:37:33] This is a question that everybody asks in the last six years or seven years since the first transformer came. Honestly, nothing has changed. The only thing that has changed is the transformer became a mixture of experts model, where there are multiple models and not just a single model. But the core self-attention model architecture has not changed. And people say there are shortcomings, the quadratic attention, complexities there. But any solution to that incurs costs somewhere else too.

Most of the people are not aware that majority of the computation in a large transformer like GPT-3 or 4 is not even spent on the attention layer. It’s actually spent on the matrix multiplies. So if you’re trying to focus more on the quadratic part, you’re incurring costs and the matrix multiples, and that’s actually the bottleneck in the larger scaling.

So honestly, it’s very hard to make an innovation on the transformer that can have a material impact at the level of GPT-4 complex cost of training those models. So I would bet more on innovations, auxiliary layers, like retrievable augmented generation. Why do you want to train a really large model when you don’t have to memorize all the facts from the Internet, when you literally have to just be a good reasoning model?

Nobody is going to value Patrick for knowing all facts. They’re going to value you for being an intelligent person, fluid intelligence. If I give you something very new that nobody else has an experience in, are you well positioned to learn that skill fast and start doing it really well. When you hire a new employee, what do you care about? Do you care about how much they know about something? Or do you care about whether you can give them any task and they would still get up to speed and do it, which employee would you value more?

So that’s the sort of intelligence that we should bake into these models, and that requires you to think more on the data. What are these models training on? Can we make them train on something else and just memorizing all the words on the Internet? Can we make reasoning emerge in these models through a different way? And that might not need innovation on the transformer, that may need innovation more on what data you’re throwing at these models.

Similarly, another layer of innovation that’s waiting to happen is the architecture like sparse versus dense models. Clearly, mixture of experts is working, GPT-4 is a mixture of experts, Mixtral is a mixture of experts, Gemini 1.5 is a mixture of experts. So even there, it’s not one model for coding, one model for reasoning and math, one model for history that depending on your input, it’s getting routed to the right model. It’s not that spares.

Every individual tokened is routed to a different model, but it’s happening every layer. So you’re still spending a lot of compute. How can we create something that’s actually 100 humans in one company? So the company itself has aggregated so much smarter. We’ve not created the equivalent at a model layer, more experimentation on the sparsity and more experimentation on how we can make reasoning emerge in a different way is likely to have a lot more impact than thinking about what is the next transformer.

4. Training great LLMs entirely from ground up in the wilderness as a startup – Yi Tay

People always assume it’s simply a question/debate of accelerator choice (TPUs vs GPUs etc) and all GPU clusters are created equal. For us, this soon proved to be false. As we sampled across different service providers, we find that the variance of hardware quality differs vastly even for the same hardware, i.e., GPUs (H100s). Note that here, hardware refers to overall cluster quality and not necessarily the chips or accelerators per se. Just like a lottery. Basically:

Not all hardware is created equal. The variance of cluster quality across hardware providers is so high that it is literally a lottery pertaining to how much pain one would have to go through to train good models. In short, a hardware lottery in the era of LLMs.

More specifically, we’ve leased a few clusters from several compute providers, each with a range of hundreds to thousands of chips. We’ve seen clusters that range from passable (just annoying problems that are solvable with some minor SWE hours) to totally unusable clusters that fail every few hours due to a myriad of reasons. Specifically, some clusters have nodes that fail every N hour with issues ranging from cabling issues (where N is unreasonably small), GPU hardware errors etc. Even more surprisingly, every cluster across the same provider could also be vastly different in terms of how robust it was…

…Did I mention you’ll also get a different Model Flop Utilisation (MFU) for different clusters!? This was a non negligible amount of compute wasted if one is unlucky enough to find a provider with badly cabled nodes or some other issues. Systems with very sub-optimal file systems would have the MFU of training runs tank the moment a team mate starts transferring large amounts of data across clusters.

Every service provider also had different levels of support. These range from being polite to nonchalant, “chatgpt-style” canned responses to blaming the user for every single thing that goes wrong.

Overall, every single cluster we tried feels like they have their own vibe, struggles and failure modes. It was also almost as though every single cluster needed their own hot-fixes for their own set of issues – some more tolerable than others. That said, we’ve learned that fail safes are important, and finding fast hot fixes for any clusters could be key…

…We’re training our models on GPUs for the most part at Reka. Personally, I’ve used TPUs all my life when it comes to large language model training at Google pre-Reka life. CUDA and nccl were the most alien thing to me ever. (I only learned it’s pronounced “Nickel” from one of my coworkers who used to work at Nvidia lol)

I was completely taken aback by the failure rate of GPUs as opposed to my experiences on TPUs at Google. In fact, I don’t actually recall TPUs failing much even for large runs, though I was not sure if I was protected from knowing this just by the sheer robustness of the outrageously good infra and having a dedicated hardware team. In fact, the UL2 20B model (at Google) was trained by leaving the job running accidentally for a month. It never failed. If this were in GPU land, it would have failed within the first few days for sure.

That said, I think this could be more about the competency of the hardware team that manages your accelerators rather than the underlying chip. The presence of having good hardware support (from your compute provider) is important. And so much hinges on them being actually competent, reinforcing the notion of the “hardware lottery”…

…It is no secret that my favourite codebase of all time is T5X and Mesh Tensorflow (named tensors ftw) but these options quickly became not viable as 1) they don’t get as much support outside Google, 2) they are kind of deprecated and 3) they are not friendly to folks on our team that are not xooglers.

We ended up going for something vanilla, seemingly stable and more popular (i.e., pytorch) that is more accessible to most people on the team (except me lol). In my first few months, I was tripping all over pip, git, docker and all these wild life stuff. Then again, I am not 100% sure about how stable or user friendly it would be to use a google codebase externally (it would have been pretty nasty I guess).

To be very frank, I would have to say the quality of codebases externally significantly lag behind those I’ve been used to at Google. Primarily because codebase within Google tends to be written by ML rockstars themselves (e.g, Noam Shazeer, Barret Zoph, Adam Roberts, Hyung Won Chung et al.) and just feel better (e.g., superior vibes) compared to those I’ve tried externally. In particular, I found myself super annoyed with the code quality when dabbling with stuff built by other companies (some way worse than others 🤗).

5. How The Interstate Highway System Changed American Industry – Lawrence Hamtil

Signed into law in 1956 by then President Dwight Eisenhower, the Federal Highway Act created the Interstate Highway System, which would become the largest and costliest public works project in history. Measuring almost 48,000 miles in total distance, the Interstate Highway System was completed only in 1992, more than three decades after work began, and for a total cost in today’s dollars of more than $500 billion…

…Among the beneficiaries of this huge outlay were the quarry owners and aggregate miners, who provided the gravel and rock on which the interstates were laid, the heavy machinery manufacturers who provided the graders, tractors, and steamrollers that turned those rocks into roads, and the oil and gas producers and refiners who made the gasoline and diesel that fueled the project…

…As families began to set out exploring the country on the new interstate system, restauranteurs such as Ray Kroc and Howard Johnson recognized the need to provide traveling families with predictable, familiar service. The idea of the chain restaurant was born as interstate exit ramps guided hungry motorists to McDonald’s and Howard Johnson’s. Families would also need places to say on longer journeys, so hotels followed restaurants in the chain model as franchises like Holiday Inn became a staple of interstate exits; early ads for the hotel underlined the value of the familiar by stating, “The best surprise is no surprise.”

The logistical flexibility provided by the interstate system also gave rise to a whole new model of retailing: big box stores began to set up in small towns offering rich variety and low prices to consumers previously left unserved by larger retailers. Walmart’s 1975 annual report detailed just such a model…

…Whereas not quite a century before the railroads had aided in the rise of Sears, Roebuck, and Co. as the first retailer with national reach, the interstate in the 1960s and 1970s would provide the backbone of Walmart’s logistical operations, with large distribution centers situated at critical points throughout the interstate network to facilitate inventory replenishment, as Professor Jesse LeCavalier has noted on his blog.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google), Amazon, Apple, and Microsoft. Holdings are subject to change at any time.

Stock Buybacks and Privatisations in Singapore’s Stock Market, China’s Property Market, What’s Next for Mario, & More

Earlier this week, on 11 March 2024, I was invited for a short interview on Money FM 89.3, Singapore’s first business and personal finance radio station, by Chua Tian Tian, the co-host of the station’s The Evening Runway show. We discussed a number of topics, including:

City Developments’ S$5.5 million share buyback on 8 March 2024 and the implications behind the company’s move (Hint: City Developments rarely conducts share buybacks, and this recent buyback happened at a time when the company’s price-to-book ratio is at 0.6, which is near a 10-year low)
Rumours on a privatisation deal for Japfa from its controlling shareholders (Hint: Japfa’s business has historically been cyclical and it appears that its business results are picking up after a rough few years; at the same time, the company’s valuation looks really low on the surface)
The improvement in Singapore’s business sentiment and what it means for Singapore-listed counters from the sectors with the most positive outlooks (Hint: A rising tide may not lift all boats)
What would it take for the Chinese property market to rebound (Hint: Demand for Chinese properties is collapsing while Chinese property developers are facing severe financial strain, leading to even lesser demand for Chinese properties)
What would a new Mario movie in 2026 mean for Nintendo (Hint: It’s likely to be a boon for Nintendo in the short run, but the long run impacts are less clear)

You can check out the recording of our conversation below!

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Meta Platforms. Holdings are subject to change at any time.

What We’re Reading (Week Ending 10 March 2024)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

Here are the articles for the week ending 10 March 2024:

1. Flawed Valuations Threaten $1.7 Trillion Private Credit Boom – Silas Brown, Laura Benitez, John Sage, Kat Hidalgo, and Ellen Schneider

The meteoric rise of private credit funds has been powered by a simple pitch to the insurers and pensions who manage people’s money over decades: Invest in our loans and avoid the price gyrations of rival types of corporate finance. The loans will trade so rarely — in many cases, never — that their value will stay steady, letting backers enjoy bountiful and stress-free returns. This irresistible proposal has transformed a Wall Street backwater into a $1.7 trillion market.

Now, though, cracks in that edifice are starting to appear.

Central bankers’ rapid-fire rate hikes over the past two years have strained the finances of corporate borrowers, making it hard for many of them to keep up with interest payments. Suddenly, a prime virtue of private credit — letting these funds decide themselves what their loans are worth rather than exposing them to public markets — is looking like one of its greatest potential flaws.

Data compiled by Bloomberg and fixed-income specialist Solve, as well as conversations with dozens of market participants, highlight how some private-fund managers have barely budged on where they “mark” certain loans even as rivals who own the same debt have slashed its value.

In one loan to Magenta Buyer, the issuing vehicle of a cybersecurity company, the highest mark from a private lender at the end of September was 79 cents, showing how much it would expect to recoup for each dollar lent. The lowest mark was 46 cents, deep in distressed territory. HDT, an aerospace supplier, was valued on the same date between 85 cents and 49 cents…

…“As interest rates have risen, so has the riskiness of borrowers,” Lee Foulger, the Bank of England’s director of financial stability, strategy and risk, said in a recent speech. “Lagged or opaque valuations could increase the chance of an abrupt reassessment of risks or to sharp and correlated falls in value, particularly if further shocks materialize.”…

…Some market participants wonder, however, whether the fog around pricing suits investors just fine. Several fund managers, who requested anonymity when speaking for fear of endangering client relationships, say rather than wanting more disclosure, many backers share the desire to keep marks steady — prompting concerns about a code of silence between lenders and the insurers, sovereign wealth funds and pensions who’ve piled into the asset class.

One executive at a top European insurer says investors could face a nasty reckoning at the end of a loan’s term, when they can’t avoid booking any value shortfall. A fund manager who worked at one of the world’s biggest pension schemes, and who also wanted to remain anonymous, says valuations of private loan investments were tied to his team’s bonuses, and outside evaluators were given inconsistent access to information.

The thinly traded nature of this market may make it nigh-on impossible for most outsiders to get a clear picture of what these assets are worth, but red flags are easier to spot. Take the recent spike in so-called “payment in kind” (or PIK) deals, where a company chooses to defer interest payments to its direct lender and promises to make up for it in its final loan settlement.

This option of kicking the can down the road is often used by lower-rated borrowers and while it doesn’t necessarily signal distress, it does cause anxiety about what it might be obscuring…

…According to Solve, about three-quarters of PIK loans were valued at more than 95 cents on the dollar at the end of September. “This raises questions about how portfolio companies struggling with interest servicing are valued so high,” says Eugene Grinberg, the fintech’s cofounder.

An equally perplexing sign is the number of private funds who own publicly traded loans, and still value them much more highly than where the same loan is quoted in the public market.

In a recent example, Carlyle Group Inc.’s direct-lending arm helped provide a “second lien” junior loan to a US lawn-treatment specialist, TruGreen, marking the debt at 95 cents on the dollar in its filing at the end of September. The debt, which is publicly traded, was priced at about 70 cents by a mutual fund at the time…

…Thrasio is an e-commerce business whose loan valuations have been almost as varied as the panoply of product brands that it sells on Amazon, which runs from insect traps and pillows to cocktail shakers and radio-controlled monster trucks.

As the company has struggled lately, its lenders have been divided on its prospects. Bain Capital and Oaktree Capital Management priced its loans at 65 cents and 79 cents respectively at the close of September. Two BlackRock Inc. funds didn’t even agree: One valuing its loan at 71 cents, the other at 75 cents. Monroe Capital was chief optimist, marking the debt at 84 cents. Goldman Sachs Group Inc.’s asset management arm had it at 59 cents.

The Wall Street bank seems to have made the shrewder call. Thrasio filed for Chapter 11 on Wednesday as part of a debt restructuring deal and one of its public loans is quoted well below 50 cents, according to market participants. Oaktree lowered its mark to 60 cents in December…

…Distressed companies do throw up some especially surprising values. Progrexion, a credit-services provider, filed for bankruptcy in June after losing a long-running lawsuit against the US Consumer Financial Protection Bureau. Its bankruptcy court filing estimated that creditors at the front of the queue would get back 89% of their money. Later that month its New York-based lender Prospect Capital Corp. marked the senior debt at 100 cents…

…For private credit’s many champions, the criticism’s overblown. Fund managers argue that they don’t need to be as brutal on marking down prices because direct loans usually involve only one or a handful of lenders, giving them much more control during tough times. In their eyes, the beauty of this asset class is that they don’t have to jump every time there’s a bump in the road…

…Direct lenders also use far less borrowed money than bank rivals, giving regulators some comfort that any market blowup could be contained. They typically lock in cash they get from investors for much longer periods than banks, and they don’t tap customer deposits to pay for their risky lending. They tend to have better creditor protections, too.

2. An Interview with Nat Friedman and Daniel Gross Reasoning About AI – Ben Thompson, Nat Friedman, and Daniel Gross

The other release, I think around the same day, was Groq released a demo of using their processor online. This is about the processor, it’s not about the model. They’re using Mistral and Llama as the the available models, but the speed is truly remarkable. It strikes me as a big deal, not because what it says about Groq — that’s a different question and I actually I’m curious about your guys points of view on some questions there — but I’ve been on, for a long time, there is a user experience issue when it comes to AI, and a lot of the use cases we’re talking about where, because it is human-like, the vastness of the uncanny valley is very large and basically any friction in that experience matters way more than it matters with a phone. With a phone, when you’re pulling it out of your pocket or you’re sitting out of your device, you’re never not aware that you’re using a phone or that you’re using a computer. It’s never like, “Wow, I thought I was talking to a human, I was actually talking on my phone.” No, that’s never going to happen, and so you actually have way more latitude for user experience friction. However, when it comes to AI, the fact that it can sound like a human, speed matters, it matters hugely, and the reason why I thought that demo was a big deal was again, the business prospects of Groq aside, it was tangible that, yes, this is the right thesis. Speed actually makes an astronomical difference and it felt like validation of a view that I had on that.

DG: Yeah, I think we have pretty fast response times from our minds, I think the brain runs at a pretty high hertz, and depending on the mood that you’re in, there’s alpha, beta, gamma, but at the end of the day we perceive reality very quickly and we hadn’t quite had an experience where something was that instant and that fast and that fluid, but I think that’s only the beginning to be honest, and someone’s going to have to do the hard work of actually taking that concept, be it on Groq’s hardware or somewhere else and turning it into something that’s very polished, refined and a product that can handle interruptions, that sort of thing.

But once someone does that, if I had to guess, if we try to project forward in the next podcast or the one after that, what is the big new thing? It’s just this idea that we’re going to move into a more agentic world of models where what we have now is very Precambrian. You go to chat.openai.com and you put in a bunch of words and some words come out and at the end of the day the model is rhyming more than it’s thinking, and it’s a little slow and I think next era is to have actual agents do tasks for you on the Internet, converse with you at human speed, and I think the economy and market prices don’t factor this in at all.

Well, this is the reason to be optimistic about Groq. If you actually pencil out the cost of their systems, and part of the reasons why it’s so fast is every individual chip has a very small amount of SRAM, which keeps the data in place and is super expensive, but it’s deterministic, they know exactly where the data is, but that means they need big systems to have enough memory. That means they would need a large market to develop. So they’re pushing this cost per token idea, but you have to have just an astronomical amount of tokens moving through the system for that pricing to make sense. My sense though is speed actually matters so much that this is a use case unlocker.

NF: It’s a user interface unlocker too. With slow model outputs, you were forced to have this streaming tokenization, the stream of tokens basically coming at you and now with speed, speed has always been a feature and I think actually in many ways this is just a reminder of a perennial rule of user interface design, which is that speed matters, latency matters. It’s a funny thing because users usually don’t ask for it, but they just sense that they prefer the thing that’s snappy and they choose it over the thing that’s sluggish.

And I think that difference is, like I said, that much bigger for these sorts of models.

NF: But in this case I think it unlocks new types of UI, whereas previously you had to sit there and watch the model just stream tokens at you.

This is where you can actually talk to it and it feels normal. It doesn’t feel weird.

NF: Yeah. Well, it also actually, I think, feels more superhuman in a way, because you can get a whole essay in seconds and you can get a book in minutes and there’s a way in which the superhuman feeling is stronger, but also I think you could have the model, for example, if you’re willing to spend the money, it’s more reasonable to have the model explore several paths and maybe it’s going to try ten things and pick the one that works best because it can do it very quickly…

…Groq is really interesting because they’ve been around for a long time. Jonathan Ross, the founder, invented the TPU at Google and then set out to do it better in a certain respect. I think they almost died and then LLMs come along and suddenly they have this architecture that seems to works well. Again, you have this, under the surface, it’s quite deterministic that maps well to their approach.

You mentioned the scaling bit, Daniel. I think one of the questions that goes with this about chip design in general is at what point does it make sense to specialize even more than the GPU? The GPU is much more specialized than a CPU, but it’s still general purpose, and that comes with real costs when it comes to things like latency and things like that. Do these go hand in hand? If it actually is the case that scale is the answer to almost every problem, does that mean the opportunity for a more specialized architecture has arrived maybe sooner than we expected?

DG: I think so. And we are sitting here, I think, before the era of AI ASICs [Application-specific integrated circuit]. Maybe Groq is a little early to it because it’s been around for a little longer but if I had to guess, this is a big part of the future.

I think one of the main things that’s changed, I remember calling Jonathan the day after Llama came out, and I told him the industry is going to finally standardize around something where you can show people how great you are, because previously his issue was, he was parading around a bunch of these benchmarks and people had a tough time translating that into something that was so economically valuable they’d reconfigured their entire architecture for a specialized chip. It wasn’t just Jonathan, it was that whole era of your 2016, ’17 AI companies. What happened was really Meta created a standard by open sourcing Llama and everyone started thinking in terms of token output per second basically. That became a standard where you can perform by, and much more importantly, you can measure your balance sheet by.

AI companies go through two cycles when they train their models, they’re fairly margin, I think, insensitive, they just want the best GPUs, they don’t want to take any risk. You’re spending $300 million, you just want your model to “tape out” properly and then if you find product market fit, you switch to this inference era. Now in the inference era, you’re ultimately staring at your COGS and you’re staring at your COGS every month and you’re thinking, “Gosh, we’re paying so much per hour, per GPU, whatnot. It makes total sense for us to allocate five engineers and re-architect towards this completely different alien platform.” It’s an ASIC effectively, people would be upset if I call their chips ASICs but you get the idea.

Well, it’s more of that category than a GPU, yes.

DG: It’s a dedicated chip and it makes total sense to do that because you’re just staring at your COGS. It’s sort of like how much would you be willing to architect your infrastructure as a fintech company if you could lower your interchange rate? Well, the answer is usually a lot and the Nvidia margin is a kind of interchange rate for tokens, and you’re very much willing to do the work and the schlep for custom architecture if it works in a way that people just weren’t willing to do in 2017 because very few companies had revenue coming in.

The inference was smaller than the training market.

DG: The only people who had this, by the way, were the advertising companies, Meta and Google, and they had their own chips.

So I think ultimately that’s what happened is you’re now able to monetize these models in a way where you can do the mental math to yourself about why it makes sense to rewrite them for a custom architecture, and if I had to guess, Nvidia’s dominance in training, as far as I can tell, remains strong as ever. Over time, I don’t necessarily know that they’ll lose share, but the pie will grow and the inference pie is going to grow to some of these ASICs and to some extent it already has of course, with the TPU, and Meta has its own internal custom inference chips and that’s going to grow, I think, over time because it just makes economic sense to do so…

…There seems to be a groundswell of robotic foundation models that are coming, where we haven’t yet had this GPT-3 moment of robotics where you have a couple of hands on a desk and it can tie a shoe or it can decorate a cake or put a Lego together and do all those things relatively well or in a way that feels like the beginnings of robotic intelligence, but it seems like that’s coming in the next 12 or 18 months. We will see those demonstrations.

What’s enabling it is this belief in scaling and a few breakthroughs on the model architecture side and what’s holding it back is data. You don’t have the common crawl of robotic data, you can’t scrape the Internet for robotic instruction data and so all the efforts going into collecting those data sets and the early demonstrations are really impressive and they do involve local learned models for things like motion and kinematics and balance and stuff like that in some cases.

Is data going to be a real differentiator in that there’s going to be fights for exclusive data sets, or will it become a commodity where everyone realizes the way you actually differentiate is with the product and it’s actually to everyone’s benefit to have access to the best data sets and there’ll be more collective action?

NF: I think this is a really good question. If it had happened a few years ago, I think it would’ve been much more likely that there’d be common data sets. There are a few open robotic data sets, but they’re pretty small, pretty low quality and now that we’re already in the AI gold rush, it seems likely that the really expensive project of collecting a bunch of data, whether that’s through teleoperations or something else, will happen inside funded companies, either big companies or smaller.

Does this apply to data generally, just because maybe theoretically it’d be best for everyone to adopt a collective approach to have a high-minded where we’re going to actually differentiate, but right now the stakes are so high, everyone’s like, “Nope, my data, I’m not going to share”?

NF: The walls are going up, definitely the shutters are down on data, it used to be easier to scrape websites than it is today. Scraping has gotten harder, generally, you see that across the board. So I think companies, that at one point didn’t view the content of all their UGC as an asset, now suddenly do. They say, “Wait, we’ve got this big data set that can be trained on.”…

…NF: The bet on long context is very important and we think that being able to not just retrieve out of but reason over huge amounts of information, is a super, I mean, it’s partly a human ability. We have episodic memory and we have procedural memory and the ability to retain skills or memories over time and there’s been an open question, “How are models going to do this? How are they going to develop episodic or procedural memory?”, and you can do both in the context.

In the context, you can put episodes in that the model will remember and you can put skills in, as Google actually demonstrated by teaching it new languages inside a single prompt and then asking it to use those skills. So this has been a big missing skill, this may not be the final way it shows up in AI systems, but it’s a new way that we can do this that I think is incredibly meaningful.

You can also do superhuman things as well. Reason over huge code bases, show it hours of security footage and ask it to draw correlations across that. I do think it’s amazing and a real breakthrough, and it’s clear that Google has figured something out here, and they have a bit of a secret and we’ve all been looking for clues and poring over the literature to figure out what it is. But this is a real axis of differentiation.

Well, that’s the big question in my mind, how much of this is model and how much of this is infrastructure? Because there was a presentation they did at their enterprise event last year, and it’s weird, I can’t find this anywhere, I spent hours looking for it last week, I was writing about 1.5. But I very tangibly remember it where they were talking about this sort of sharding capability, where we know about sharding in the context of databases, and the problems that solves and the challenges it presents, but they were talking about sharding in the context of, I think they were talking about it for training. But it seems like they’re doing sharding in the context of inference where they have this ability to distribute the workload, not just across chips, not just across clusters, but at least in theory, across data centers, which introduces huge challenges as far as you’re constrained by the speed of light.

Google’s networking capabilities have always been well known, but I’m not sure it’s been appreciated how that could be brought to bear on these issues. And you talked about, Daniel, how much can you make a sparse model, and to do this, and to do a mixture-of-experts sort of approach, and to spread it out. It’s the exact opposite of Groq. Groq is massively serial, super fast. What if we can spread it out all over the place and because the use case is tolerable of latency, we can just take that all the way to the extreme? And it feels like only Google could do what Gemini 1.5 is right now, and it doesn’t feel like anyone else is even close.

DG: Do you think anyone else is close, Nat?

NF: Well, we know of one company that has this also.

DG: Yeah.

NF: Daniel and I made an investment last week in a company called Magic that has a very good, very efficient, extremely long, longer than Gemini, context that’s working. To be honest with you, we thought there was only one company that had this, now we know there were two…

…The reason why Gemini as it shipped feels so distasteful, is it feels like bad faith, it’s very blatantly on the tin, “We’re not actually doing our best job to give you an answer”. It’s just straightforward, and it feels like an aspect where we would forgive an AI screwing up, we’ve been forgiving OpenAI all along, and they had some early episodes where there was clearly slants put on, and they’ve worked through that. But it felt like in good faith, “We’re doing our best here.” Gemini doesn’t feel like it’s in good faith, and maybe it was an accident that it feels that way, but it crossed a line of perception that just seems very problematic.

How did this happen? How did we get a product like this from a company that is supposedly too scared to ship and they ended up finally shipping and then it’s just a disaster?

NF: Well, I think you’re right. I think one reason they should get a little less leeway than OpenAI did, is that they saw what came before them, and they learned nothing from the precedents. Dall-E 2 had its own sort of crazy woke image creation problem that they had to adjust and tune and they learned from, and that was all forgivable because they were pioneering and ChatGPT has been through this as well and so Google should have seen all that and learned from it and done better.

It’s such a great point. This is a big advantage of going first, is you get more grace.

NF: You do, you get more grace, because no one’s ever solved these problems before. But Google definitely didn’t come first and still made mistakes that feel like 2021 mistakes, 2022 mistakes, and that’s much less forgivable.

How did it happen? I mean, I think culture’s a very big component. You wrote about that, and it’s clear that it was very difficult for anyone at Google to raise their hand and say, “Hey, I don’t think we should ship in this form, we should probably do something about this.”

Then, we’ve heard from people at Google that the models themselves, this is not likely to be something that was a deep problem in the model training, but a decision that was made in the productization by someone who came later. So, there’s probably a set of system prompts or templates or something like that that are imposing a set of rules and guidance to the models that the raw internal models don’t do.

I think this is the challenge. Google’s always had this funny word they use for shipping products, which is what they call externalization, I always thought that was a very culturally-indicative piece of jargon from Google, because it kind of captures in a way, the way Google thinks of itself. They develop breakthrough technologies internally and then they externalize the magic, and it’s not a product-first thinking, it’s not even a customer-first thinking, it’s a technology-first thinking. I think that’s where the mistake is here, in the externalization, in the process of putting it out there.

So in a way that makes it easy to fix, there’s probably a single file that could be edited that would improve things a lot, and in another way, editing that file might mean going through layers of product people and policy people who will potentially have a lot to say about that, and the gulf between the brilliant minds creating the models and the users, there’s someone in the middle and that’s where the challenge lies.

How exactly do you think this is happening, Daniel? Is it that there’s the level from the data, there’s the model, there’s the RLHF [Reinforcement Learning from Human Feedback] process, there’s the prompt, where are things going sideways here?

DG: Well, we were having a good conversation about this earlier. I mean, traditionally there’s, I think, a few things people misunderstand a little bit. Pre-training and fine-tuning a model are not distinct ideas, they’re sort of the same thing. That fine-tuning is just more the pre-training at the end. As you train models, this is something I think we believe, but we now see backed by a lot of science, the ordering of the information is extremely important. Because look, the ordering for figuring out basic things like how to properly punctuate a sentence, whatever, you could figure that out either way. But for higher sensitivity things, the aesthetic of the model, the political preferences of the model, the areas that are not totally binary, it turns out that the ordering of how you show the information matters a lot.

In my head, I always imagine it like you’re trying to draw a sheet, a very tight bed sheet over a bed, and that’s your embedding space, and you pull the bed sheet in the upper right-hand corner and the bottom left hand corner pops off, and you do that and then the top right hand corner pops off, that’s sort of what you’re doing. You’re trying to align this high dimensional space to a particular set of mathematical values, and then at some point you’re never going to have a perfect answer or a loss of zero. So, the ordering matters, and fine-tuning is traditionally more pre-training do at the end.

I think that’s originally the liberal leanings of the OpenAI ChatGPT model, came out of that. I think it was a relatively innocuous byproduct of those final data points that you show the model to, it becomes very sensitive to and those data points, it’s very easy to accidentally bias that. For example, if you have just a few words in the internal software you have where you’re giving the human graders prompts in terms of what tokens they should be writing into the model, those words can bias them and if the graders can see the results of other graders, you have these reflexive processes. It’s like a resonant frequency and very quickly it compounds. Errors compound over time. I actually think you could end up without really thinking through it with a model that’s slightly left-leaning, a lot of the online text is slightly left-leaning…

…I think the piece of information that’s most interesting is the fact that Google lacked a very basic process. This is your point, where maybe people thought or maybe people didn’t even think before they launched it and I’m thinking a lot of that famous Steve Jobs interview where he says, “The problem with Microsoft is they just have no taste.” I think the unexpected thing about AI, we’ve talked about it in this podcast, but I don’t think it’s been generally expected, is fine-tuning a model is just as aesthetic an art as making a beautiful landing page for your website.

So in hindsight, it shouldn’t be that surprising that the Borg that built the interfaces of GCP also produced very robotic models, like that’s the same thing and it also should not be surprising to us that Mistral, which a French company with French cultures and now French products, was able to produce a model that to their credit, I mean, it’s not the smartest, but it’s by far the most obedient and has by far the most neutral political tone, at least in my anecdotal testing.

Well, actually, I want to get to Mistral in a moment, but Nat, what does Google do now?

DG: Other than call you?

NF: (laughing) Yeah, I mean I think this is a leadership challenge. There’s a missing editor here and there’s a missing product editor and a missing person with good taste and judgment who gives a damn and has the authority to overrule anyone in the company and make sure the right thing goes out the door. I do think leadership changes have to happen, culture is the hardest type of change to make in a company. You could do strategy change, you could do product change, you could do operational change. Culture change is the one that’s just super difficult and it can only happen with leadership. We either need to see dramatically different behavior from Google leadership or we need to see dramatically different leaders.

3. TIP611: The Bear Case For China w/ Kyle Bass – Clay Finck and Kyle Bass

[00:06:59] Clay Finck: One of the things that sort of struck me in preparing for this conversation is that much of the information that various institutions have used to gather on what’s happening in China has actually been cut off by the CCP and it’s no longer available.

[00:07:14] Clay Finck: So why have such moves? been made by the CCP. We know they like to control data and information flow. And how are you able to get accurate information on what’s happening in China and really make sense of it?

[00:07:28] Kyle Bass: No one has accurate data on China except the Chinese Communist Party. They do and used to, they began to adhere to Western standards and they put together data aggregators that collected both micro macro level data.

[00:07:40] Kyle Bass: And so they had a Bloomberg of China called wind and there were four or five others. And they were actually pretty good, but if you dug into the data, if you looked at the Chinese Customs Bureau for import and export, and you looked at the customs data that was in the wind database 1 year until they recently cut it off, it was off by 200 billion dollars.

[00:08:02] Kyle Bass: Not 2 billion dollars, 200 billion dollars. Then you think about trade with the US is what? 650 billion. So to be off by 200 billion, that just means someone’s really cooking the books. We all knew that Chinese data had low fidelity, and now there just isn’t Chinese data anymore.

[00:08:22] Kyle Bass: As of March of 2023, they severed all of those links to U.S. research universities, to the Fed, to Wall Street writ large, and that data is only allowed out of the mainland. To mainland data, call it readers, and they’re not allowed to share it unless the party approves it. So do you think you’re getting the truth? Probably not. And, they were reporting youth unemployment until they actually reported that it was over 20%.

[00:08:47] Kyle Bass: And then they say, we’re not going to report that anymore. If you read some Chinese scholars while that was going on, 1 of the top scholars at 1 of the top universities in China said. It looks like it’s 46 percent and then they silenced him…

…[00:12:13] Kyle Bass: They’d rather pretend. Those things aren’t bad. And I’ll take you to an October 2023 Reuters release where the People’s Bank of China, which is the regulator or the call it the Chinese Fed that regulates their banking system issued an edict in October 23 and it said, The local government financing bonds that exist in the marketplace in China, it’s a 13 trillion dollar equivalent market, a monster market in China.

[00:12:39] Kyle Bass: It’s all about how the local governments fund themselves by selling real estate. They sell real estate to pay their debts. They issue debt and to gather even more funding. And that 13 trillion dollar market is in default. 80 percent of those bonds are not paying. Those local governments can’t pay because there’s no real estate bid because every public developer in China is in default.

[00:13:00] Kyle Bass: When you think about what the PBOC said in October of 23, they said to the banks, if you own the debt or you own those bonds, you can just say they’re current and it won’t affect your ratings in our annual reviews of the banks. We’re just going to pretend that the market’s paying. Just think about that for a second.

[00:13:17] Kyle Bass: Clay, a 13 trillion market. is in a complete state of default, and we’re just not going to talk about it…

…[00:14:44] Kyle Bass: We really haven’t sanctioned anything or anyone when you really look at this. I know we’re going to try to get serious, but going back to what they’re doing in their legal system, in January of 2020, China updated its foreign investment law, giving Beijing the power and the ability to nationalize foreign assets or investments.

[00:15:03] Kyle Bass: Under special circumstances, which include war, that’s their words, not mine that began in January of 2020. That’s super interesting because that’s when a covid emanated from the city of Wuhan. So that’s when they began their legal movements in the system. In June of 2021, they issued a new counter foreign sanctions law.

[00:15:24] Kyle Bass: Foreign sovereigns that were sanctioning anyone in China, they were saying if Chinese. Corporate interests or international corporate interests that have business in China are adhering to foreign sanctions that are punitive on China. That China can just nationalize their interests, imprison the expats that live there, and basically turn their companies off.

[00:15:49] Kyle Bass: Basically they were countering foreign sanctions by saying we’ll just shut off all of your business here in China and we’ll take everything that you’ve got. That happened on June 21. In April of 23, Chinese lawmakers passed a new update to their anti espionage legislation. If you remember, that’s when they were raiding U.S. due diligence firms.

[00:16:06] Kyle Bass: They raided 3 or 4 firms, they arrested everyone, they took all of the computers, and due diligence firms were just doing due diligence, business due diligence. On potential acquisitions management teams, they’re everything that companies like Bain or McKenzie or these others do when they get hired to do due diligence, that became illegal and that had a chilling effect…

…[00:19:55] Clay Finck: In light of those laws that you mentioned that were passed around COVID and ever since COVID, I actually ran across this chart that showed data from the administration of foreign exchange. It showed that China’s inbound foreign direct investment has just essentially collapsed.

[00:20:10] Clay Finck: It was, this data shows it was north of 300 billion just prior to COVID. And then in 2023 it is around 33 billion. Does that data sound accurate to you?

[00:20:19] Kyle Bass: That’s right. And there’s a caveat to that data where they don’t asterisk and don’t tell you this, but it’s actually wildly negative. And let me explain to you how.

[00:20:27] Kyle Bass: If you are a corporate interest in the U. S. and, or a multinational and you have business in China Tesla’s got business in China, there are plenty of multinationals that have business there. Chevron has business there. The profits they make in China get put in a Chinese bank and China never lets them out.

[00:20:45] Kyle Bass: So I know many multinational companies that have hired friends of mine to try to get their money out. And China just, pardon the pun, gives them a bunch of red tape and won’t allow the money out. Every dollar that’s made by a multinational in China, if it stays in the bank through the end of the year, it’s counted as foreign direct investment into China.

[00:21:06] Kyle Bass: When you look at the FDI numbers, they’ll always be until they nationalize everything, right? Multinational profits in China are automatically FDI. And I think that’s also a lens that we need to be thinking about looking at things through. What is a complete collapse of FDI, by the way, Clay…

…[00:29:20] Clay Finck: So in addition to what’s happening here, in relation to Taiwan, China definitely seems to be going through a financial crisis of their own, which you’ve touched on plenty here. And a lot of data has pointed towards an economic contraction, but they actually reported GDP growth of 5.3 percent in 2023.

[00:29:38] Clay Finck: And real estate is definitely a big part of China’s economy. So What are you seeing in their real estate market and how this plays into the bigger picture?

[00:29:50] Kyle Bass: The data that’s actually being released, again, whether there’s proper fidelity in the data, nobody knows. Clearly it’s suspect, but Hong Kong’s real estate is down over 25%.

[00:30:01] Kyle Bass: Again, since China took over, that’s the largest decline ever. And that’s just a harbinger of more to come. And by the way, that’s probably that’s the reported number. We know the real numbers are much worse and we have a couple of anecdotes from people that we know that have traded in that market and been forced to trade in the real estate market there.

[00:30:22] Kyle Bass: And it’s much worse than people think it is. But when you think about the Chinese, you mentioned that Chinese real estate is vital to their GDP. It’s somewhere between 33 percent and 40 percent of their GDP. It’s 70 percent of their net worth. And it is, it was the primary driver of the Chinese miracle of their GDP growth.

[00:30:41] Kyle Bass: And imagine if you allowed reckless speculation in your real estate markets. Your GDP grows, all the ancillary services grow. Everyone technically gets wealthier and wealthier. The banks lend into it. The bank, their banking system is three and a half times the size of its GDP. The U. S. going into a financial crisis was one time our GDP.

[00:31:02] Kyle Bass: And you know how bad we screwed this up back in 2008. And if you include non banks like Fannie and Freddie and other financials, we’re about 1. 7 times. They’re three and a half times levered to their GDP.

4. Off the Run: Piero Sraffa and Imperial Japanese Government Bonds – Irwin Union

For the better part of 70 years, rumours have followed the Italian economist Piero Sraffa. Long the subject of speculation, it has been asserted that in the dying days of the Second World War, Sraffa heavily bought defaulted Imperial Japanese Government bonds. These, following the Treaty of San Francisco, being eventually honoured in full.

Though several authors have offered differing accounts of what Sraffa was purported to have done, till now, no person has been able to offer a satisfying and granular account of events…

Two credible accounts of Sraffa’s investments survive…

…The second comes from the historian Norman Stone:

The economist Piero Sraffa, editor of the correspondence of David Ricardo and re-floater of Marx’s sunken theory of surplus value, took two economic decisions in his life. He bought Japanese bonds in 1945, and he swapped them in 1960 for gold, dying a very rich man.

…Luckily, recent events, including the opening of Sraffa’s archive at Trinity College, afford new insight in to what Sraffa did, when he did it, and, indeed, how he did it…

…Following her entry into the Second World War, Japan began to default on most of her external obligations in, as best as can be figured, mid 1941.

At the outbreak of the war, a number of Imperial Japanese Government bonds were listed on the London Stock Exchange. These securities were issued in the United Kingdom, denominated in British Pounds and were obligations that Japan had entered into under British law.

Japan could refuse to acknowledge them, but could not inflate them away, nor strike them out by fiat. And so they remained outstanding, with an ongoing market made, all through the war and into the peace that followed; shielded from the worst problems of the immediate post war Japanese economy by dint of their denomination in sterling and their legal domicile.

Following her 1941 default, the bonds, already on the ropes prior to the war, collapsed completely…

…Among the items in Sraffa’s archive at Trinity College are two remarkable sets of papers.

The first is a series of trading receipts issued by the London branch of the Swiss Bank Corporation. These receipts run from 1946 to 1951, and cover Sraffa’s trading of Imperial Japanese Government Bonds, as well as some miscellaneous securities (City of Wilno, Poland at 3.25 of par and Estonian bonds at 6 of par, as well as some common stock.)

The second is a series of letters received by Sraffa from an unnamed Swiss organisation who custodied gold bullion for him.

It’s reasonable to conjecture that this was also the Swiss Bank Corporation, though it’s impossible to know as the letters are so discrete as to carry no letterhead or distinguishing detail of any kind. These letters give us an inventory of Sraffa’s bullion holdings in Switzerland as of 1975, and broadly corroborate Stone’s assertion that Sraffa swapped out of bonds into gold bullion.

From the set of trading receipts, we can, with only a few minor adjustments, build a chronology of Sraffa’s trading, and, thus, a simulated portfolio of his holdings. This portfolio can then be priced using collected price data.

As of 1960, we can substitute the simulated portfolio of bonds for gold and then continue to price the portfolio all the way through to 1983.

Of course, there are wrinkles, discussed vide infra, and so it should be understood that the best that can be done is speculation about Sraffa’s actual record.

Nonetheless, we can get somewhere close to reality, and enough detail is provided for the reader to make her own back of the envelope adjustments and calculations as desired.

I first collected monthly price data for the period from 1946 to 1951 (the period in which Sraffa was actively trading) and six monthly data from 1929 to 1960.

With this data in hand, we can begin to unravel the question of how and what Sraffa accomplished.

Sraffa’s receipts show that between 1946 and 1951, he traded quite frequently, realising capital gains and recycling his proceeds into other issues. However, in late 1951 Sraffa halted his trading altogether.

From here, for the purposes of simulating his record, we assume that the portfolio remained static until 1960.

Sraffa’s final trades consolidate his holdings into the 1899 bond. This issue bore one of the earliest maturity dates…

…On the 9th of March, 1946, as Sraffa was likely contemplating his first purchases, the Financial Times ran a front page story titled Japan Bonds’ Bleak Outlook: Chancellor Reaffirms Gloomy View. The article reported on comments made by the Chancellor of the Exchequer in the House of Commons the previous day, wherein he had stated that:

[…] in the case of British bondholders at large, and in general, I will do my utmost to see that they get fair play. There is nothing new in that, but why humbug Japanese bondholders into believing that they have anything but the very dimmest and remotest chance of recovering anything of their investments?.

Following the Chancellor’s remarks, the bonds sold off by approximately 20%…

…Reading the financial papers of the time, one finds a veritable feast of views on the Japanese loans expressed in articles, opinion pieces and letters to the editor. Indeed, the letters to the editor in particular functioned as a sort of clearinghouse for opinion and query. It’s not a stretch to compare these exchanges to those that happen on message boards and social media today.

Though the full record is too voluminous to feature in full, it is also so information dense that it forms a vital part of any study of the securities.

We learn some extraordinary facts from these articles and letters. For instance, as early as late 1946 thru January 1947, it was being stated that interest on the defaulted bonds had been paid into sinking funds during the war.

One stock which tended to be overlooked when the market was active was the Tokyo Five Percent, 1912. Like Japanese Government Stocks, the interest has been set aside for bondholders in Tokyo throughout the period of the war and after, and Japanese nationals have been paid.

Any question of transfer to British bondholders awaits the signing of the Peace Treaty and the unfreezing of the yen-sterling exchange; the latter process can hardly be a quick one.

Japanese Bonds Speculation – Lex – Financial Times – 27/1/47

We also learn that the amount needed to make British bondholders whole was relatively de minimis. This is because Japanese citizens, for reasons not apparent, owned most of the sterling issues. Japanese citizens were compulsorily converted into Yen denominated bonds in 1943, presumably due to strains on Japan’s foreign exchange balances, leaving only the rump holdings of foreign owners intact.

A correspondent has lately received a cable from the Far East which has bearing on my note of yesterday on Japanese bonds. The cable reads as follows:

“Japanese Sterling Bonds interest paid all Japanese holders in Japan at former rates of exchange until March, 1942. Foreign nationals in Japan paid interest into special custody account. After March, 1943, Japanese owned compulsorily converted into yen bonds. No payments made of interest against unconverted bonds, but still being made on converted.”

That puts the position in a nutshell. Whatever the peace treaty may have to say on the matter, it is a fact, as is pointed out by my correspondent, that the default in interest due to British and Allied holders of Japanese sterling bonds not resident in Japan would not need a large sum to wipe out, as the Japanese always held the larger part of the sterling bonds. Lex

Japanese Post Script – Lex – Financial Times – 28/1/47

We also learn of Japan’s wish to join the United Nations and apply for membership of the IMF.

[…] 6) The goodwill of the Japanese since the end of hostilities, and the expressed desire of the Japanese Government to join the United Nations as soon as permissible after the signing of the Peace Treaty. An intention to apply for membership to the International Monetary Fund once the Peace Treaty has been signed has also been indicated.

Letters to the Editor – Financial Times – 19/4/47

In the following letter, the author, a former resident of Japan, argues that the settlement of the debt would allow Japan to reestablish herself with foreign lenders at negligible cost.

Having spent several years in the service of the Japanese Government and having always kept in close touch with financial circles in that country, I have no hesitation in endorsing the view expressed by one of your readers a few weeks ago, namely, that the bonds in question are the best three-year lock-up on the market to-day, or as “Lex” remarked in your issue dated 2nd January: “If I were asked to name a good speculative long-shot for 1947, I think Japanese bonds would be as strong a starter as any.”

[…] Finally, the amount of Japan’s foreign indebtedness is infinitesimal, and the Government is fully alive to the fact that by meeting its commitments it is reestablishing its financial credits abroad at a very small cost.

Japan Bonds and Reparations

Letter to the Editor – Financial Times – 21/5/47

And then, on the 23rd of December, 1947, there is what can only be described as an extraordinary letter from William Teeling, a member of the House of Commons. This letter is worth inclusion in full.

Sir, -There has been much comment in your paper and elsewhere recently on the widening interest in all Japanese loans. Yesterday (Friday) afternoon I told a number of business men in the City interested in Japan what I know about these loans, and I feel that it is only fair that everyone should know, since contact with Japan and the Japanese is so difficult.

I have just returned as a member of a Parliamentary delegation which spent six weeks in the Far East, and while in Tokyo I made it my business to inquire about these loans which interest so many people here.

The Finance Minister in the present Japanese Coalition Cabinet told me that all interest accrued on the Japanese bonds would definitely be paid when peace with America has been signed. He could not say yet at what rate, but it would definitely not be at the rate when war broke out. He added that even during the war bondholders in Switzerland for certain loans were paid and he assured me that money has all the time been set aside in Tokyo for this purpose.

This was confirmed to me at a later meeting with heads of Japanese business firms and banks at which meeting the Foreign Secretary, Mr. Ashida, was also present. Mr. Ashida explained to me that new loans from America were essential and therefore Japan must keep up her reputation for meeting her debts and would pay off her earlier loans.

Reparations officials confirmed that the sums outstanding are small and could be repaid. The American officials concerned told me that a rate for the repayment of all debts will shortly be fixed and will definitely take into account the present depreciation of the yen.

But when will peace be signed? I only know that America was waiting for the recent Four Power Conference to break down before going ahead on a separate peace with Japan, and Great Britain will reluctantly support her as it is the only solution, but it will mean the strengthening of Japan and that means more loans.

William Teeling. House of Commons, S.W.1.

Letters to the Editor – Financial Times – 23/12/47…

…On the 23rd of August, 1949, we learn that Japan’s total external debt was then $323mm USD with approximately $80mm USD of unpaid interest thereon. We also learn that British claims totalled approximately £62mm GBP.

Kaneschichi Masuda, Japanese chief Cabinet Secretary, said here today that he was unable to reveal any practical plans whereby Japans foreign bond commitments could be met.

[…]

He said that $323m. worth of bonds were held by foreigners, on which $80m. in interest had accumulated. British subsribers held about £62m. of this amount.

Japan and Bond Repayment – Financial Times – 23/8/49

However, it was not so cut and dried. By 1951, the mood had soured, and the question of reparations, long simmering, had become acute. In April, Teeling again wrote to the Times, this time expressing concern about the lack of progress and the possible outcomes for British bondholders.

At question was whether reparations would rank ahead of foreign bondholders, and whether reparations might exhaust Japan’s capacity to make foreign bondholders whole, irrespective of her desire to do so.

Then, on the 13th of August, news of formal recognition by the Japanese Government of her prewar debts was published in the Financial Times.

Japan will not be restricted milatarily, politically or economically under the draft peace treaty published yesterday by Britain and the United States.

Japan affirms its liability for the pre-war external debt of the Japanese State, and for debts of corporate bodies subsequently declared to be liabilities of the Japanese State, and expresses its intention to enter on negotiations at an early date with its creditors with respect to the resumption of payments on those debts.

It will facilitate negotiations in respect to private pre-war claims and obligations; and will facilitate the transfer of sums accordingly.

Japanese bonds were active on the London Stock Exchange yesterday. Prices rose sharply at the opening and were up to £5 higher at one time. Following publication of the terms of the draft treaty there was, however, considerable profit taking. As a result, closing prices well after hours were £4 below the best.

Japan Recognises Debt Liability; Prepared for Talks on Payments – Financial Times – 13/8/51

The formal end of hostilities between Japan and the Allied powers came in September, 1951, with the signing of the Treaty of San Francisco. With the treaty formalised, Japan was now able to turn to the issue of settling her defaulted foreign obligations.

In March, 1952, the Financial Times reported that the Japanese Government was placing £20mm GBP on deposit in London as a goodwill gesture.

The Treasury announces that the Japanese Foreign Exchange Control Board is arranging to deposit with the Bank of England £20m. as a token of good will towards the holders of Japanese sterling bonds.

The initiative for this move was taken by the Japanese Foreign Minister. When neccessary formalities have been completed, the sum will be deposited and will remain with the Bank of England for two years.

During that period, it will be available for any payments by Japan to her creditors in connection with a settlement of her sterling bond indebtedness.

Japan to Deposit £20m. in London – Financial Times – 29/3/52

The front page of the 29 September issue of the Financial Times read Japan to Pay Full Interest Arrears, and detailed the terms agreed upon in New York.

After negotiations lasting nearly two and a half months, agreement has been reached in New York on the treatment of Japan’s bonded debt to Britain and the United States. It is a settlement that goes a very long way to meeting British Claims. The service on outstanding issues is to be resumed forthwith. Interest arrears that have piled up since the Pearl Harbour affair brought Japan into the war are to be met in full, though at a time lag of ten years from the due dates. There is a similiar arrangment for the treatment of repayment obligations. Moreover, the currency clauses included in a number of the debts under discussion at the conference are to be substantially honoured. The Japanese have, in short, comitted themselves to do what they said they would do before the conference began.

Contractual Terms – Financial Times – 29/9/52

On the 24th of November, the Times published the full terms of the settlement.

Briefly, the terms provided for the extensions of maturities by ten and fifteen years, a catch up payment generally equal to a single coupon, and the amortisation of accumulated defaulted coupons by the payment of one current and one defaulted coupon for each payment period until all defaulted coupons had been settled. This, in effect, doubling the coupon of each bond for a discrete period…

…With firm details of the restructuring of the loans, we can now model the post 1951 evolution of Sraffa’s portfolio through to 1960. I assume that Sraffa allowed his coupons to accumulate in cash, rather than reinvesting them.

With this account curve in hand, we can now model his swap to gold bullion in 1960.

At the end of 1960, Sraffa’s simulated account had a value of £52,676.

At year end 1960, a kg of gold bullion cost £404.46. Thus, assuming no frictions, we find that Sraffa swapped his bonds and cash for ~ 130 kg of gold bullion.

With this, we now have a complete simulated account curve for the entire period.

According to these calculations, Sraffa compounded his initial simulated outlay of £8000 cash into £1,105,839, a multiple of 138 times, or 13.97% per annum over approximately 38 years.

5. Thoughts on Ben Graham’s “Unpopular Large Caps”: A Still-Effective Strategy – John Huber

In the spirit of Graham’s categories, I recently gave a presentation to Saber investors during our latest client Zoom call with an overview of my own three main categories of our own investments: 1) Core operating businesses that we hope can compound value for a decade+, 2) Time Arbitrage (Similar to Ben Graham’s Unpopular Large Caps) and 3) Bargains.

This “Category 2” provides a frequent enough flow of ideas thanks to a very simple fact: stocks fluctuate much more than true business values do…

…I’ve written about the concept of “investment edge” on numerous occasions (see: What is Your Edge?), and how in today’s world, information has become easier to get and thus more of a commodity. But this information access, along with other technologies, has caused our attention spans to become shorter and shorter, which I think has diminished our patience and our time horizons. We want results now. This has created a “time arbitrage” opportunity, and I expect this will only gain strength as time horizons and patience levels continue to shorten.

Past examples of Category 2 ideas would include Apple in 2016 when pessimism surrounding the next iPhone cycle and worries about Apple’s competition caused the stock to fall below 10 P/E, Verisign when worries about government intervention into its pricing practices caused the stock to fall to multiyear valuation lows, or large banks like BAC and JPM in 2015-2016 when the market was expecting and fearing a difficult economy (and larger loan losses). More recent examples of mispriced large caps might include large cap tech stocks in 2022: AMZN fell 50% in 2022 and rose 80% in 2023, and that was mild compared to what happened at numerous other mega cap stocks. The valuation levels fluctuate far more than business values.

To be clear, there always is a legitimate negative fundamental case to be made when stocks get mispriced, but I think the majority of the time these concerns tend to be focused on the short term. Amazon over invested in warehouse capacity because it overestimated the growth in online retail sales, but was this going to negative impact Amazon’s long-term moat? (I would argue that in one sense it actually further entrenched their moat, making it very difficult for other retailers with lesser capacity to offer the same experience of low cost and speed of delivery: another large online marketplace with ambitions to enter the logistics space ended up throwing in the towel during this period). Sometimes, these short-term difficulties end up being long-term beneficial for the “unpopular large caps”, and the great thing about this category of investment is you get to acquire a stake in these better-positioned large companies when their stocks are depressed.

JPM is recent example of a Category 2 idea as well: the stock traded down under 8 P/E in the summer of 2022 when recession fears were prevalent (similar to what happened in 2016 to bank stocks).

I think Jamie Dimon had some great advice on the right mindset last year when he said (paraphrasing): “in 20 years, the world’s stock market capitalization will be much higher, the assets in the banking system will be higher, corporate earning power will be higher, the dollar volume of merger transactions will be higher, global payment volume will be higher.” The implication is JPM has a durable moat and thus is positioned to take a cut of all of that business. Earnings might decline in the near term, but what matters to business values is the long-term free cash flows that it earns over time.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google), Amazon, Apple, Meta Platforms, Microsoft, and Tesla. Holdings are subject to change at any time.

The Latest Thoughts From American Technology Companies On AI (2023 Q4)

A collection of quotes on artificial intelligence, or AI, from the management teams of US-listed technology companies in the 2023 Q4 earnings season.

The way I see it, artificial intelligence (or AI), really leapt into the zeitgeist in late-2022 or early-2023 with the public introduction of DALL-E2 and ChatGPT. Both are provided by OpenAI and are software products that use AI to generate art and writing, respectively (and often at astounding quality). Since then, developments in AI have progressed at a breathtaking pace.

Meanwhile, the latest earnings season for the US stock market – for the fourth quarter of 2023 – is coming to its tail-end. I thought it would be useful to collate some of the interesting commentary I’ve come across in earnings conference calls, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. This is an ongoing series. For the older commentary:

2023 Q1 – here and here
2023 Q2 – here and here
2023 Q3 – here and here

With that, here are the latest commentary, in no particular order:

Airbnb (NASDAQ: ABNB)

Airbnb’s management believes that AI will allow the company to develop the most innovative and personalised AI interfaces in the world, and the company recently acquired GamePlanner AI to do so; Airbnb’s management thinks that popular AI services today, such as ChatGPT, are underutilising the foundational models that power the services; GamePlanner AI was founded by the creator of Apple’s Siri smart assistant

There is a new platform shift with AI, and it will allow us to do things we never could have imagined. While we’ve been using AI across our service for years, we believe we can become a leader in developing some of the most innovative and personalized AI interfaces in the world. In November, we accelerated our efforts with the acquisition of GamePlanner AI, a stealth AI company led by the co-founder and original developer of Siri. With these critical pieces in place, we’re now ready to expand beyond our core business. Now this will be a multiyear journey, and we will share more with you towards the end of this year…

…If you were to open, say, ChatGPT or Google, though the models are very powerful, the interface is really not an AI interface. It’s the same interface as the 2000s, in a sense, the 2010s. It’s a typical classical web interface. So we feel like the models, in a sense, are probably underutilized…

Airbnb’s management does not want to build foundational large language models – instead, they want to focus on the application layer

One way to think about AI is, let’s use a real-world metaphor. I mentioned we’re building a city. And in that city, we have infrastructure, like roads and bridges. And then on top of those roads and bridges, we have applications like cars. So Airbnb is not an infrastructure company. Infrastructure would be a large language model or, obviously, GPUs. So we’re not going to be investing in infrastructure. So we’re not going to be building a large language model. We’ll be relying on, obviously, OpenAI. Google makes — or create a model, Meta creates models. So those are really infrastructure. They’re really developing infrastructure. But where we can excel is on the application layer. And I believe that we can build one of the leading and most innovative AI interfaces ever created.

Airbnb’s management believes that the advent of generative AI represents a platform shift and it opens the probability of Airbnb becoming a cross-vertical company

Here’s another way of saying it. Take your phone and look at all the icons on your phone. Most of those apps have not fundamentally changed since the advent of Generative AI. So what I think AI represents is the ultimate platform shift. We had the internet. We had mobile. Airbnb really rose during the rise of mobile. And the thing about a platform shift, as you know, there is also a shift in power. There’s a shift of behavior. And so I think this is a 0-0 ball game, where Airbnb, we have a platform that was built for 1 vertical short-term space. And I think with AI — Generative AI and developing a leading AI interface to provide an experience that’s so much more personalized than anything you’ve ever seen before.

Imagine an app that you feel like it knows you, it’s like the ultimate Concierge, an interface that is adaptive and evolving and changing in real-time, unlike no interface you’ve ever seen before. That would allow us to go from a single vertical company to a cross-vertical company. Because one of the things that we’ve noticed is the largest tech companies aren’t a single vertical. And we studied Amazon in the late ’90s, early 2000s, when they went from books to everything, or Apple when they launched the App Store. And these really large technology companies are horizontal platforms. And I think with AI and the work we’re doing around AI interfaces, I think that’s what you should expect of us.

Alphabet (NASDAQ: GOOG)

Alphabet’s Google Cloud segment saw accelerated growth in 2023 Q4 from generative AI

Cloud, which crossed $9 billion in revenues this quarter and saw accelerated growth driven by our GenAI and product leadership.

Alphabet closed 2023 by launching Gemini, a foundational AI model, which has state-of-the-art capabilities; Gemini Ultra is coming soon

We closed the year by launching the Gemini era, a new industry-leading series of models that will fuel the next generation of advances. Gemini is the first realization of the vision we had when we formed Google DeepMind, bringing together our 2 world-class research teams. It’s engineered to understand and combine text, images, audio, video and code in a natively multimodal way, and it can run on everything from mobile devices to data centers. Gemini gives us a great foundation. It’s already demonstrating state-of-the-art capabilities, and it’s only going to get better. Gemini Ultra is coming soon. The team is already working on the next versions and bringing it to our products.

Alphabet is already experimenting Gemini with Google Search; Search Generative Experience (SGE) saw its latency drop by 40% with Gemini

We are already experimenting with Gemini in Search, where it’s making our Search Generative Experience, or SGE, faster for users. We have seen a 40% reduction in latency in English in the U.S.

Alphabet’s management thinks that SGE helps Google Search (1) answer new types of questions, (2) answer complex questions, and (3) surface more links; management believes that digital advertising will continue to play an important role in SGE; management has found that users find the ads placed above or below an AI overview of searches to be helpful; management knows what needs to be done to incorporate AI into the future experience of Google Search and they see AI assistants or agents as being an important component of Search in the future

By applying generative AI to Search, we are able to serve a wider range of information needs and answer new types of questions, including those that benefit from multiple perspectives. People are finding it particularly useful for more complex questions like comparisons or longer queries. It’s also helpful in areas where people are looking for deeper understanding, such as education or even gift ideas. We are improving satisfaction, including answers for more conversational and intricate queries. As I mentioned earlier, we are surfacing more links with SGE and linking to a wider range of sources on the results page, and we’ll continue to prioritize approaches that add value for our users and send valuable traffic to publishers…

…As we shared last quarter, Ads will continue to play an important role in the new search experience, and we’ll continue to experiment with new formats native to SGE. SGE is creating new opportunities for us to improve commercial journeys for people by showing relevant ads alongside search results. We’ve also found that people are finding ads either above or below the AI-powered overview helpful as they provide useful options for people to take action and connect with businesses…

…Overall, one of the things I think people underestimate about Search is the breadth of Search, the amount of queries we see constantly on a new day, which we haven’t seen before. And so the trick here is to deliver that high-quality experience across the breadth of what we see in Search. And over time, we think Assistant will be very complementary. And we will again use generative AI there, particularly with our most advanced models in Bard and allows us to act more like an agent over time, if I were to think about the future and maybe go beyond answers and follow through for users even more. So that is the — directionally, what the opportunity set is. Obviously, a lot of execution ahead. But it’s an area where I think we have a deep sense of what to do.

Alphabet’s latest Pixel 8 phones have an AI-powered feature that lets users search what they see on their phones without switching apps; the Pixel 8s uses Gemini Nano for AI features

Circle to Search lets you search what you see on Android phones with a simple gesture without switching apps. It’s available starting this week on Pixel 8 and Pixel 8 Pro and the new Samsung Galaxy S24 Series…

…Pixel 8, our AI-first phone, was awarded Phone of the Year by numerous outlets. It now uses Gemini Nano with features like Magic Compose for Google Messages and more to come.

Alphabet’s management is seeing that advertisers have a lot of interest in Alphabet’s AI advertising solutions; the solutions include (1) the Automatically Created Assets (ACA) feature for businesses to build better ads and (2) conversational experiences – currently under beta testing – that has helped SMBs be 42% more likely to publish ads with good ad-strength

We are also seeing a lot of interest in our AI-powered solutions for advertisers. That includes our new conversational experience that uses Gemini to accelerate the creation of Search campaigns…

…As we look ahead, we’re also starting to put generative AI in the hands of more and more businesses to help them build better campaigns and even better performing ads. Automatically created assets help advertisers show more relevant search ads by creating tailored headlines and descriptions based on each ad’s context. Adoption was up with strong feedback in Q4. In addition to now being available in 8 languages, more advanced GenAI-powered capabilities are coming to ACA…

…And then last week’s big news was that Gemini will power new conversational experience in Google Ads. This is open and beta to U.S. and U.K. advertisers. Early tests show advertisers are building higher-quality search campaigns with less effort, especially SMBs who are 42% more likely to publish a campaign with good or excellent ad strength.

Alphabet’s Google Cloud offers AI Hypercomputer (a supercomputing architecture for AI), which is used by high-profile AI startups such as Anthropic and Mistral AI

Google Cloud offers our AI Hypercomputer, a groundbreaking supercomputing architecture that combines our powerful TPUs and GPUs, AI software and multi-slice and multi-host technology to provide performance and cost advantages for training and serving models. Customers like Anthropic, Character.AI, Essential AI and Mistral AI are building and serving models on it.

Vertex AI, which is within Google Cloud, enables users to customise and deploy more than 130 generative AI models; Vertex AI’s API (application programming interface) requests has jumped six times from the first half of 2023 the second half; Samsung is using Vertex AI to provide GenAI models in its Galaxy S24 smartphones while companies such as Shutterstock and Victoria’s Secret are also using Vertex AI

For developers building GenAI applications, we offer Vertex AI, a comprehensive enterprise AI platform. It helps customers like Deutsche Telekom and Moody’s discover, customize, augment and deploy over 130 GenAI models, including PaLM, MedPaLM, Sec-PaLM and Gemini as well as popular open source and partner models. Vertex AI has seen strong adoption with the API request increasing nearly 6x from H1 to H2 last year. Using Vertex AI, Samsung recently announced its Galaxy S24 Series smartphone with Gemini and Imagen 2, our advanced text-to-image model. Shutterstock has added Imagen 2 to their AI image generator, enabling users to turn simple text prompts into unique visuals. And Victoria’s Secret & Co. will look to personalize and improve the customer experience with Gemini, Vertex AI, Search and Conversations.

Duet AI, Alphabet’s AI agents for its Google Workspace and Google Cloud Platform (GCP) services, now has more than 1 million testers, and will incorporate Gemini soon; Duet AI for Developers is the only generative AI offering that supports the entire development and operations lifecycle for software development; large companies such as Wayfair, GE Appliances, and Commerzbank are already using Duet AI for Developers

Customers are increasingly choosing Duet AI, our packaged AI agents for Google Workspace and Google Cloud Platform, to boost productivity and improve their operations. Since its launch, thousands of companies and more than 1 million trusted testers have used Duet AI. It will incorporate Gemini soon. In Workspace, Duet AI is helping employees benefit from improved productivity and creativity at thousands of paying customers around the world, including Singapore Post, Uber and Woolworths. In Google Cloud Platform, Duet AI assists software developers and cybersecurity analysts. Duet AI for Developers is the only GenAI offering to support the complete development and operations life cycle, fine-tuned with the customer’s own core purpose and policies. It’s helping Wayfair, GE Appliances and Commerzbank write better software, faster with AI code completion, code generation and chat support. With Duet AI and Security Operations, we are helping cybersecurity teams at Fiserv, Spotify and Pfizer.

Alphabet’s management believes that the company has state-of-the-art compute infrastructure and that it will be a major differentiator in the company’s AI-related work; managements wants Alphabet to continue investing in its infrastructure

Search, YouTube and Cloud are supported by our state-of-the-art compute infrastructure. This infrastructure is also key to realizing our big AI ambitions. It’s a major differentiator for us. We continue to invest responsibly in our data centers and compute to support this new wave of growth in AI-powered services for us and for our customers.

Alphabet’s AI-powered ad solutions are helping retailers with their omni channel growth; a large big-box retailer saw a 60%+ increase in omni channel ROA (return on advertising) and a 22%+ increase in store traffic

Our proven AI-powered ad solutions were also a win for retailers looking to accelerate omni growth and capture holiday demand. Quick examples include a large U.S. big-box retailer who drove a 60%-plus increase in omni ROAS and a 22%-plus increase in store traffic using Performance Max during Cyber Five; and a well-known global fashion brand, who drove a 15%-plus higher omnichannel conversion rate versus regular shopping traffic by showcasing its store pickup offering across top markets through pickup later on shopping ads.

Alphabet’s management is using AI to make it easier for content creators to create content for Youtube (for example, creators can easily create backgrounds or translate their videos); management also believes the AI tools built for creators can also be ported over to the advertising business to help advertisers

First, creation, which increasingly takes place on mobile devices. We’ve invested in a full suite of tools, including our new YouTube Create app for Shorts, to help people make everything from 15-second Shorts to 15-minute videos to 15-hour live streams with a production studio in the palm of their hands. GenAI is supercharging these capabilities. Anyone with a phone can swap in a new backdrop, remove background extras, translate their video into dozens of languages, all without a big studio budget. We’re excited about our first products in this area from Dream Screen for AI-generated backgrounds to Aloud for AI-powered dubbing…

…You are obviously aware of the made YouTube announcement where we introduced a whole lot of new complementary creativity features on YouTube, including Dream Screen, for example, and a lot of other really interesting tools and thoughts. You can obviously imagine that we can take this more actively to the advertising world already. As you know, it continues already to power AI, a lot of our video ad solutions and measurement capabilities. It’s part of video-rich campaigns. Multi-format ads are — actually, there is a generative creator music that actually makes it easier for creators to design the perfect soundtrack already. And as I said earlier, AI will unlock a new world of creativity. And you can see how this will — if you just look at where models are heading, where multimodal models are heading, where the generation capabilities of those models are heading, you can absolutely see how this will impact and positively impact and simplify the flow for creators, similar to what you see already emerging in some of our core products like ACA on the Search side.

Alphabet’s management expects the company’s capital expenditure in 2024 to be notably higher than in 2023 (it was US$20 billion in 2023), driven by investments in AI infrastructure

With respect to CapEx, our reported CapEx in the fourth quarter was $11 billion, driven overwhelmingly by investment in our technical infrastructure with the largest component for servers followed by data centers. The step-up in CapEx in Q4 reflects our outlook for the extraordinary applications of AI to deliver for users, advertisers, developers, cloud enterprise customers and governments globally and the long-term growth opportunities that offers. In 2024, we expect investment in CapEx will be notably larger than in 2023.

Alphabet’s management is restructuring the company’s workforce not because AI is taking away jobs, but because management believes that AI solutions can deliver significant ROI (return on investments) and it’s important for Alphabet to have an organisational structure that can better build these solutions

But I also want to be clear, when we restructure, there’s always an opportunity to be more efficient and smarter in how we service and grow our customers. We’re not restructuring because AI is taking away roles that’s important here. But we see significant opportunities here with our AI-powered solution to actually deliver incredible ROI at scale, and that’s why we’re doing some of those adjustments.

Alphabet’s management thinks that Search is not just about generative AI

Obviously, generative AI is a new tool in the arsenal. But there’s a lot more that goes into Search: the breadth, the depth, the diversity across verticals, stability to follow through, getting actually access to rich, diverse sources of content on the web and putting it all together in a compelling way.

Alphabet’s management believes that AI features can help level the playing field for SMBs in the creation of effective advertising (when competing with large companies) and they will continue to invest in that area

Our focus has always been here on investing in solutions that really help level the playing field, and you mentioned several of those. So actually, SMBs can compete with bigger brands and more sophisticated advertisers. And so the feedback we’re always getting is they need easy solutions that could drive value quickly, and several of the AI-powered solutions that you’re mentioning are actually making the workflow and the whole on-ramp and the bidded targeting creative and so on, you mentioned that is so much easier for SMBs. So we’re very satisfied with what we’re seeing here. We will continue to invest.

Amazon (NASDAQ: AMZN)

Amazon’s cloud computing service, AWS, saw an acceleration in revenue growth in 2023 Q4 and management believes this was driven partly by AI

If you look back at the revenue growth, it accelerated to 13.2% in Q4, as we just mentioned. That was an acceleration. We expect accelerating trends to continue into 2024. We’re excited about the resumption, I guess, of migrations that companies may have put on hold during 2023 in some cases and interest in our generative AI products, like Bedrock and Q, as Andy was describing

Amazon’s management reminded the audience that their framework for thinking about generative AI consists of three layers – the first is the compute layer, the second is LLMs as a service, the third is the applications that run on top of LLMs – and Amazon is investing heavily in all three

You may remember that we’ve explained our vision of three distinct layers in the gen AI stack, each of which is gigantic and each of which we’re deeply investing.

At the bottom layer where customers who are building their own models run training and inference on compute where the chip is the key component in that compute…

…In the middle layer where companies seek to leverage an existing large language model, customize it with their own data and leverage AWS’ security and other features, all as a managed service…

…At the top layer of the stack is the application layer.

Amazon’s management is seeing revenues accelerate rapidly for AWS across all three layers of the generative AI stack and AWS is receiving significant interest from customers wanting to run AI workloads

Still relatively early days, but the revenues are accelerating rapidly across all three layers, and our approach to democratizing AI is resonating well with our customers. We have seen significant interest from our customers wanting to run generative AI applications and build large language models and foundation models, all with the privacy, reliability and security they have grown accustomed to with AWS

Amazon’s management is seeing that enterprises are still figuring out which layer of the generative AI stack they want to operate in; management thinks that most enterprises will operating in at least two layers, with the technically capable ones operating in all three

When we talk to customers, particularly at enterprises as they’re thinking about generative AI, many are still thinking through at which layers of those three layers of the stack I laid out that they want to operate in. And we predict that most companies will operate in at least two of them. But I also think, even though it may not be the case early on, I think many of the technically capable companies will operate at all three. They will build their own models, they will leverage existing models from us, and then they’re going to build the apps.

At the first layer of the generative AI stack, AWS is offering the most expansive collection of compute instances with NVIDIA chips; AWS has built its own Trainium chips for training and Inferentia chips for inference; a new version of Trainium – Trainium 2 – was recently announced and it is 4x faster, and has 3x more memory, than the first generation of Trainium; large companies and prominent AI startups are using AWS’s AI chips

At the bottom layer where customers who are building their own models run training and inference on compute where the chip is the key component in that compute, we offer the most expansive collection of compute instances with NVIDIA chips. We also have customers who like us to push the price performance envelope on AI chips just as we have with Graviton for generalized CPU chips, which are 40% more price-performant than other x86 alternatives. And as a result, we’ve built custom AI training chips named Trainium and inference chips named Inferentia. In re:Invent, we announced Trainium2, which offers 4x faster training performance and 3x more memory capacity versus the first generation of Trainium, enabling advantageous price performance versus alternatives. We already have several customers using our AI chips, including Anthropic, AirBnB, Hugging Face, Qualtrics, Rico and Snap.

At the middle layer of the generative AI stack, AWS has launched Bedrock, which offers LLMs-as-a-service; Bedrock is off to a very strong start with thousands of customers already using it just a few months after launch; Bedrock has added new models, including those from prominent AI startups, Meta’s Llama2, and Amazon’s own Titan family; customers are excited over Bedrock because building production-quality generative AI applications requires multiple iterations of models, and the use of many different models, and this is where Bedrock excels

In the middle layer where companies seek to leverage an existing large language model, customize it with their own data and leverage AWS’ security and other features, all as a managed service, we’ve launched Bedrock, which is off to a very strong start with many thousands of customers using the service after just a few months… We also added new models from Anthropic, Cohere, Meta with Llama 2, Stability AI and our own Amazon Titan family of LLMs. What customers have learned at this early stage of gen AI is that there’s meaningful iteration required in building a production gen AI application with the requisite enterprise quality at the cost and latency needed. Customers don’t want only one model. They want different models for different types of applications and different-sized models for different applications. Customers want a service that makes this experimenting and iterating simple. And this is what Bedrock does, which is why so many customers are excited about it.

At the top layer of the generative AI stack, AWS recently launched Amazon Q, a coding companion; management believes that a coding companion is one of the very best early generative AI applications; Amazon Q is linked with more than 40 popular data-connectors so that customers can easily query their data repositories; Amazon Q has generated strong interest from developers

At the top layer of the stack is the application layer. One of the very best early gen AI applications is a coding companion. At re:Invent, we launched Amazon Q, which is an expert on AWS, writes code, debugs code, tests code, does translations like moving from an old version of Java to a new one and can also query customers various data repositories like Internet, Wikis or from over 40 different popular connectors to data in Salesforce, Amazon S3, ServiceNow, Slack, Atlassian or Zendesk, among others. And it answers questions, summarizes data, carries on a coherent conversation and takes action. It was designed with security and privacy in mind from the start, making it easier for organizations to use generative AI safely. Q is the most capable work assistant and another service that customers are very excited about…

…When enterprises are looking at how they might best make their developers more productive, they’re looking at what’s the array of capabilities in these different coding companion options they have. And so we’re spending a lot of time. Our enterprises are quite excited about it. It created a meaningful stir in re:Invent. And what you see typically is that these companies experiment with different options they have and they make decisions for their employee base, and we’re seeing very good momentum there.

Amazon’s management is seeing that security over data is very important to customers when they are using AI and this is an important differentiator for AWS because its AI services inherit the same security features as AWS – and AWS’s capabilities and track record in security are good

By the way, don’t underestimate the point about Bedrock and Q inheriting the same security and access control as customers get with AWS. Security is a big deal, an important differentiator between cloud providers. The data in these models is some of the company’s most sensitive and critical assets. With AWS’ advantaged security capabilities and track record relative to other providers, we continue to see momentum around customers wanting to do their long-term gen AI work with AWS.

Amazon has launched some generative AI applications across its businesses and are building more; one of the applications launched is Rufus, a shopping assistant, which allows consumers to receive thoughtful responses to detailed shopping questions; other generative AI applications being built and launched by Amazon include a customer-review-summary app, an app for customers to predict how they will fit in apparel, an app for inventory forecasts for each fulfilment centre, and an app to generate copy for ads based on a picture, or generate pictures based on copy; Rufus is seamlessly integrated into Amazon and management thinks Rufus could meaningfully change what discovery looks for shoppers using Amazon

We’re building dozens of gen AI apps across Amazon’s businesses, several of which have launched and others of which are in development. This morning, we launched Rufus, an expert shopping assistant trained on our product and customer data that represents a significant customer experience improvement for discovery. Rufus lets customers ask shopping journey questions, like what is the best golf ball to use for better spin control or which are the best cold weather rain jackets, and get thoughtful explanations for what matters and recommendations on products. You can carry on a conversation with Rufus on other related or unrelated questions and retains context coherently. You can sift through our rich product pages by asking Rufus questions on any product features and it will return answers quickly…

…. So if you just look at some of our consumer businesses, on the retail side, we built a generative AI application that allowed customers to look at summary of customer review, so that they didn’t have to read hundreds and sometimes thousands of reviews to get a sense for what people like or dislike about a product. We launched a generative AI application that allows customers to quickly be able to predict what kind of fit they’d have for different apparel items. We built a generative AI application in our fulfillment centers that forecasts how much inventory we need in each particular fulfillment center…Our advertising business is building capabilities where people can submit a picture and an ad copy is written and the other way around.

… All those questions you can plug in and get really good answers. And then it’s seamlessly integrated in the Amazon experience that customers are used to and love to be able to take action. So I think that that’s just the next iteration. I think it’s going to meaningfully change what discovery looks like for our shopping experience and for our customers.

Amazon’s management believes generative AI will drive tens of billions in revenue for the company over the next few years

Gen AI is and will continue to be an area of pervasive focus and investment across Amazon primarily because there are a few initiatives, if any, that give us the chance to reinvent so many of our customer experiences and processes, and we believe it will ultimately drive tens of billions of dollars of revenue for Amazon over the next several years.

Amazon’s management expects the company’s full-year capital expenditure for 2024 to be higher than in 2023, driven by increased investments in infrastructure for AWS and AI

We define our capital investments as a combination of CapEx plus equipment finance leases. In 2023, full year CapEx was $48.4 billion, which was down $10.2 billion year-over-year, primarily driven by lower spend on fulfillment and transportation. As we look forward to 2024, we anticipate CapEx to increase year-over-year primarily driven by increased infrastructure CapEx to support growth of our AWS business, including additional investments in generative AI and large language models.

AWS’s generative AI revenue is pretty big in absolute numbers, but small in the context of AWS already being a $100 billion annual-revenue-run-rate business

If you look at the gen AI revenue we have, in absolute numbers, it’s a pretty big number. But in the scheme of a $100 billion annual revenue run rate business, it’s still relatively small, much smaller than what it will be in the future, where we really believe we’re going to drive tens of billions of dollars of revenue over the next several years.

Apple (NASDAQ: AAPL)

Many of the features in Apple’s latest product, the virtual reality headset, the Vision Pro, features are powered by AI

There’s an incredible amount of technology that’s packed into the product. There’s 5,000 patents in the product. And it’s, of course, built on many innovations that Apple has spent multiple years on, from silicon to displays and significant AI and machine learning, all the hand tracking, the room mapping, all of this stuff is driven by AI.

Apple has been spending a lot of time and effort on AI and management will share details later in 2024

As we look ahead, we will continue to invest in these and other technologies that will shape the future. That includes artificial intelligence where we continue to spend a tremendous amount of time and effort, and we’re excited to share the details of our ongoing work in that space later this year…

…In terms of generative AI, which I would guess is your focus, we have a lot of work going on internally as I’ve alluded to before. Our MO, if you will, has always been to do work and then talk about work and not to get out in front of ourselves. And so we’re going to hold that to this as well. But we’ve got some things that we’re incredibly excited about that we’ll be talking about later this year.

Apple’s management thinks there is a huge opportunity for Apple with generative AI but will only share more details in the future

Let me just say that I think there is a huge opportunity for Apple with gen AI and AI and without getting into more details and getting out in front of myself.

Arista Networks (NYSE: ANET)

Arista Networks’ management believes that AI at scale needs Ethernet at scale because AI workloads cannot tolerate delays; management thinks that 400 and 800-gigabit Ethernet will become important or AI back-end GPU clusters

AI workloads are placing greater demands on Ethernet as they have both data and compute-intensive across thousands of processes today. Basically, AI at scale needs Ethernet at scale. AI workloads cannot tolerate the delays in the network because the job can only be completed after all flows are successfully delivered to the GPU clusters. All it takes is one culprit or worst-case link to throttle an entire AI workload…

…. We expect both 400 and 800-gigabit Ethernet will emerge as important pilots for AI back-end GPU clusters.

Arista Networks’ management is pushing the company and the Ultra Ethernet Consortium to improve Ethernet technology for AI workloads in three key ways; management believes that Ethernet is superior to Infiniband for AI-related data networking because Ethernet provides flexible ordering of data transfer whereas Infiniband is rigid

Three improvements are being pioneered by Arista and the founding members of the Ultra Ethernet Consortium to improve job completion time. Number one, packet spring. AI network topology meets packet spring to allow every flow to simultaneously access all parts of the destination. Arista is developing multiple forms of load balancing dynamically with our customers. Two is flexible ordering. Key to an AI job completion is the rapid and reliable bulk transfer with flexible ordering using Ethernet links to optimally balance AI-intensive operations, unlike the rigid ordering of InfiniBand. Arista is working closely with its leading vendors to achieve this. Finally, network congestion. In AI networks, there’s a common in-cost congestion problem whereby multiple uncoordinated senders can send traffic to the receiver simultaneously. Arista’s platforms are purpose-built and designed to avoid these kinds of hotspots, evenly spreading the load across multi-packs across a virtual output queuing VoQ losses fabric.

Arista Networks’ management thinks the company can achieve AI revenue of at least $750 million in 2025

We are cautiously optimistic about achieving our AI revenue goal of at least $750 million in AI networking in 2025…

…. So our AI performance continues to track well for the $750 million revenue goal that we set last November at Analyst Day.

Arista Networks’ management sees the company becoming the gold-standard for AI data-networking

We have more than doubled our enterprise revenue in the last 3 years and we are becoming the gold standard for client-to-cloud-to-AI networking with 1 EOS and 1 CloudVision Foundation.

In the last 12 months, Arista Networks has participated in a large number of AI project bids, and in the last five projects where there was a situation of Ethernet versus Infiniband, Arista Networks has won four of them; over the last 12 months, a lot has changed in terms of how Infiniband was initially bundled into AI data centres; management believes that Ethernet will become the default standard for AI networking going forward

To give you some color on the last 3 months, I would say difficult to project anything in 3 months. But if I look at the last year, which maybe last 12 months is a better indication, we have participated in a large number of AI bids and when I say large, I should say they are large AI bids, but there are a small number of customers actually to be more clear. And in the last 4 out of 5, AI networking clusters we have participated on Ethernet versus InfiniBand, Arista has won all 4 of them for Ethernet, one of them still stays on InfiniBand. So these are very high-profile customers. We are pleased with this progress…

…The first real consultative approach from Arista is to provide our expertise on how to build a robust back-end AI network. And so the whole discussion of Ethernet become — versus InfiniBand becomes really important because as you may recall, a year ago, I told you we were outside looking in, everybody had an Ethernet — everybody had an InfiniBand HPC cluster that was kind of getting bundled into AI. But a lot has changed in a year. And the popular product we are seeing right now and the back-end cluster for our AI is the Arista 7800 AI spine, which in a single chassis with north of 500 terabit of capacity can give you a substantial number of ports, 400 or 800. So you can connect up to 1,000 GPUs just doing that. And that kind of data parallel scale-out can improve the training time dimensions, large LLMs, massive integration of training data. And of course, as we shared with you at the Analyst Day, we can expand that to a 2-tier AI leaf and spine with a 16-way CMP to support close to 10,000 GPUs nonblocking. This lossless architecture for Ethernet. And then the overlay we will have on that with the Ultra Ethernet Consortium in terms of congestion controls, packet spring and working with a suite of [ UC ] mix is what I think will make Ethernet the default standard for AI networking going forward.

Arista Networks’ management believes that owners and operators of AI data centres would not want to work with white box data switches (non-branded and commoditised data switches) because data switches are mission critical in AI data centres, so users would prefer reliable and higher-quality data switches

I think white box is here to stay for a very long time if somebody just wants a throwaway commodity product, but how many people want throwaway commodity in the data center? They’re still mission-critical, and they’re even more mission-critical for AI. If I’m going to spend multimillion dollars on a GPU cluster, and then the last thing I’m going to do is put a toy network in, right? So to put this sort of in perspective, that we will continue to coexist with a white box. There will be use cases where Arista’s blue box or a stand-alone white box can run either SONiC or FBOSS but many times, the EOS software stack is really, really something they depend on for availability, analytics, automation, and there’s — you can get your network for 0 cost, but the cost of downtime is millions and millions of dollars.

Arista Networks is connecting more and more GPUs and management believes that the picture of how a standard AI data centre Ethernet switch will look like is starting to form; AI is still a small part of Arista Networks’ business but one that should grow over time

On the AI side, we continue to track well. I think we’re moving from what I call trials, which is connecting hundreds of GPUs to pilots, which is connecting thousands of GPUs this year, and then we expect larger production clusters. I think one of the questions that we will be asking ourselves and our customers is how these production clusters evolve. Is it going to be 400, 800 or a combination thereof? The role of Ultra Ethernet Consortium and standards and the ecosystem all coming together, very similar to how we had these discussions in 400 gig will also play a large part. But we’re feeling pretty good about the activity. And I think moving from trials to pilots this year will give us considerable confidence on next year’s number…

…AI is going to come. It is yet to come — certainly in 2023, as I’ve said to you many, many times, it was a very small part of our number, but it will gradually increase.

Arista Networks’ management is in close contact with the leading GPU vendors when designing networking solutions for AI data centres

Specific to our partnership, you can be assured that we’ll be working with the leading GPU vendors. And as you know, NVIDIA has 90% or 95% of the market. So Jensen and I are going to partner closely. It is vital to get a complete AI network design going. We will also be working with our partners in AMD and Intel so we will be the Switzerland of XPUs, whatever the GPU might be, and we look to supply the best network ever.

Arista Networks’ management believes that the company is very well-positioned for the initial growth spurts in AI networking

Today’s models are moving very rapidly, relying on a high bandwidth, predictable latency, the focus on application performance requires you to be sole sourced initially. And over time, I’m sure it’ll move to multiple sources, but I think Arista is very well positioned for the first innings of AI networking, just like we were for the cloud networking decade.

ASML (NASDAQ: ASML)

ASML’s management believes that 2025 will be a strong year for the company because of the long-term trends in its favour (this includes AI and digitalisation, customer-inventory-levels becoming better, and the scheduled opening of many semiconductor fabrication plants)

So essentially unchanged I would say in comparison to what we said last quarter. So if we start looking at 2025. As I mentioned before, we are looking at a year of significant growth and that is for a couple of reasons. First off, we think the secular trends in our industry are still very much intact. If you look at the developments around AI, if you look at the developments around electrification, around energy transition etcetera, they will need many, many semiconductors. So we believe the secular trends in the industry are still very, very strong. Secondly I think clearly by 2025 we should see our customers go through the up cycle. I mean the upward trend in the cycle. So that should be a positive. Thirdly, as we also mentioned last time it’s clear that many fab openings are scheduled that will require the intake of quite some tools in the 2025 time frame.

ASML’s management is seeing AI-related demand drive a positive inflection in the company’s order intake

And I think AI is now particularly something which could be on top of that because that’s clearly a technology transition. But we’ve already seen a very positive effect of that in our Q4 order intake…

…After a few soft quarters, the order intake for the quarter was very, very strong. Actually a record order intake at €9.2 billion. If you look at the composition of that, it was about 50/50 for Memory versus Logic. Around €5.6 billion out of the €9.2 was related to EUV, both Low NA and High NA.

ASML’s management is confident that AI will help to drive demand for the company’s EUV (extreme ultraviolet) lithography systems from the Memory-chips market in the near future

In ’23, our Memory shipments were lower than the 30% that you mentioned. But if you look at ’25, and we also take into account what I just said about AI and the need for EUV in the DDR5 and in the HBM era, then the 30% is a very safe path and could be on the conservative side.

ASML’s management thinks that the performance of memory chips is a bottleneck for AI-related workloads, and this is where EUV lithography is needed; management was also positively surprised at how important EUV was for the development of leading-edge memory chips for AI

I think there’s a bottleneck in the AI and making use of the full AI potential, DRAM is a bottleneck. The performance memory is a bottleneck. And there are solutions, but they need a heck of a lot more HBM and that’s EUV…

… And were we surprised? I must be — I say, yes, to some extent, we were surprised in the meetings we’ve had with customers and especially the Memory because we’re leading-edge Memory customers. We were surprised about the technology requirements of — for litho, EUV specifically and how it impacts how important it is for the rollout and the ramp of the memory solutions for AI. This is why we received more EUV orders than we anticipated because it was obvious in the detailed discussions and the reviews with our customers, that EUV is critical in that sense. And that was a bit of a surprise, that’s a positive surprise.

[Question] Sorry, was that a function of EUV layer count or perhaps where they’re repurposing equipment? And so now they’re realizing they need more footprint for EUV.

[Answer] No, it is layer count and imaging performance. And that’s what led to the surprise, the positive surprise, which indeed led to more orders.

ASML’s management sees the early shoots of recovery observed in the Memory chip market as being driven by both higher utilisation across the board, and by the AI-specific technology transition

I think it’s — what we’re seeing is, of course, the information coming off our tools that we see the utilization rates going up. That’s one. Clearly, there’s also an element of technology transition. That’s also clear. I think there’s a bottleneck in the AI and making use of the full AI potential, DRAM is a bottleneck. The performance memory is a bottleneck. And there are solutions, but they need a heck of a lot more HBM and that’s EUV. So it’s a bit of a mix. I mean, yes, you’ve gone through, I think, the bottom of this memory cycle with prices going up, utilizations increasing, and that combined with the technology transition driven by AI. That’s a bit what we see today. So it’s a combination of both, and I think that will continue.

ASML’s management is thinking if their planned capacity buildout for EUV lithography systems is too low, partly because of AI-driven demand for leading edge chips

We have said our capacity buildout will be 90 EUV Low-NA systems, 20 High-NA whereby internally, we are looking at that number as a kind of a base number where we’re investigating whether that number should be higher. The question is whether that 90 is going to be enough. Now we have to realize, we are selling wafer capacity, which is not only a function of the number of units, but also a function of the productivity of those tools. Now we have a pretty aggressive road map for the productivity in terms of wafers per hour. So it’s a complex question that you’re asking. But actually, we need to look at this especially against the math that we’re seeing for little requirements in the area of AI, whether it’s HBM or whether it is Logic, whether the number of units and the road map on productivity, which gives wafers because the combination is wafer capacity, whether that is sufficient.

Datadog (NASDAQ: DDOG)

Datadog’s management is seeing growing engagement in AI with a 75% sequential jump in the use of next-gen AI integrations

In observability, we now have more than 700 integrations allowing our customers to benefit from the latest AWS, Azure and GCP abilities as well as from the newly emerging AI stack. We continued to see increasing engagement there with the use of our next-gen AI integrations growing 75% sequentially in Q4.

Datadog’s management continues to add capabilities to Bits AI, the company’s natural language incident management copilot, and is improving the company’s LLM (large language model) observability capabilities

In the generative AI and LLM space, we continued to add capability to Bits AI, our natural language incident management copilot. And we are advancing LLM observability to help customers investigate where they can safely deploy and manage their models in production.

Currently, 3% of Datadog’s annualised recurring revenue (ARR) comes from next-gen AI native customers (was 2.5% in 2023 Q3); management believes the AI opportunity will be far larger in the future as all kinds of customers start incorporating AI in production; the AI native customers are companies that Datadgo’s management knows are substantially all based on AI

Today, about 3% of our ARR comes from next-gen AI native customers, but we believe the opportunity is far larger in the future as customers of every industry and every size start doing AI functionality in production…

…It’s hard for us to wrap our arms exactly around what is GenAI, what is not among our customer base and their workload. So the way we chose to do it is we looked at a smaller number of companies that we know are substantially all based on AI so these are companies like the modal providers and things like that. So 3% of ARR, which is up from what we had disclosed last time.

Microsoft said that AI accounts for six percentage points of Azure’s growth, but Datadog’s management is seeing AI-native companies on Datadog’s Azure business account for substantially more than the six percentage points mentioned

I know one number that everyone has been thinking about is one cloud, in particular, Microsoft, disclosed that 6% of their growth was attributable to AI. And we definitely see the benefits of that on our end, too. If I look at our Azure business in particular, there is substantially more than 6% that is attributable to AI native as part of our Azure business. So we see completely this trend is very true for us as well. It’s harder to tell with the other cloud providers because they don’t break those numbers up.

Datadog’s management continues to believe that digital transformation, cloud migration, and AI adoption are long-term growth drivers of Datadog’s business, and that Datadog is ideally positioned for these

We continue to believe digital transformation and cloud migration are long-term secular growth drivers of our business and critical motion for every company to deliver value and competitive advantage. We see AI adoption as an additional driver of investment and accelerator of technical innovation and cloud migration. And more than ever, we feel ideally positioned to achieve our goals and help customers of every size in every industry to transform, innovate and drive value through technology adoption.

Datadog experienced a big slowdown from its digitally native customers in the recent past, but management thinks that these customers could also be the first ones to fully leverage AI and thus reaccelerate earlier

We suddenly saw a big slowdown from the digital native over the past year. On the other hand, they might be the first ones to fully leverage AI and deploy it in production. So you might see some reacceleration earlier from some of them at least.

Datadog’s management sees the attach rates for observability going up for AI workloads versus traditional workloads

[Question] If you think about the very long term, would you think attach rates of observability will end up being higher or lower for these AI workloads versus traditional workloads?

[Answer] We see the attach rate going up. The reason for that is our framework for that is actually in terms of complexity. AI just adds more complexity. You create more things faster without understanding what they do. Meaning you need — you shift a lot of the value from building to running, managing, understanding, securing all of the other things that need to keep happening after that. So the shape of some of the products might change a little bit because the shape of the software that runs it changes a little bit, which is no different from what happened over the past 10, 15 years. But we think it’s going to drive more need for observability, more need for security products around that.

Datadog’s management is seeing AI-native companies using largely the same kind of Datadog products as everyone else, but the AI-native companies are building the models, so the tooling for understanding the models are not applicable for them

[Question] Are the product SKUs, these kind of GenAI companies are adopting, are they similar or are they different to the kind of other customer cohorts?

[Answer] Today, this is largely the same SKUs as everybody else. These are infrastructure, APM logs, profiling these kind of things that they are — or really the monitoring, these kind of things that these customers are using. It’s worth noting that they’re in a bit of a separate world because they’re largely the builders of the models. So all the tooling required to understand the models and — that’s less applicable to them. That’s more applicable to their own customers, which is also the rest of our customer base. And we see also where we see the bulk of the opportunity in the longer term, not in the handful of model providers that [ anybody ] is going to use. It’s worth noting that they’re in a bit of a separate world because they’re largely the builders of the models. So all the tooling required to understand the models and — that’s less applicable to them. That’s more applicable to their own customers, which is also the rest of our customer base.

Datadog has a much larger presence in inference AI workloads as compared to training AI workloads; Datadog’s management sees that the AI companies that are scaling the most on Azure are scaling on inference

There’s 2 parts to the AI workloads today. There’s training and there’s inference. The vast majority of the players are still training. There’s only a few that are scaling with inference. The ones that are scaling with inference are the ones that are driving our ARR because we are — we don’t — we’re not really present on the training side, but we’re very present on the inference side. And I think that also lines up with what you might see from some of the cloud providers, where a lot of the players or some of the players that are scaling the most are on Azure today on the inference side, whereas a lot of the other players still largely training on some of the other clouds.

Etsy (NASDAQ: ETSY)

Etsy’s management recently launched Gift Mode, a feature where a buyer can type in details of a person and occasion, and AI technology will match the buyer with a gift; Gift Mode has more than 200 recipient persons, and has good early traction with 6 million visits in the first 2 weeks

So what’s Gift Mode? It’s a whole new shopping experience where gifters simply enter a few quick details about the person they’re shopping for, and we use the power of artificial intelligence and machine learning to match them with unique gifts from Etsy sellers. Creating a separate experience helps us know immediately if you’re shopping for yourself or someone else, hugely beneficial information to help our search engines solve for your needs. Within Gift Mode, we’ve identified more than 200 recipient personas, everything from rock climber to the crossword genius to the sandwich specialist. I’ve already told my family that when shopping for me, go straight to the music lover, the adventurer or the pet parent…

…Early indications are that Gift Mode is off to a good start, including positive sentiment from buyers and sellers in our social channels, very strong earned media coverage and nearly 6 million visits in the first 2 weeks. As you test and shop in Gift Mode, keep in mind that this is just the beginning.

Etsy’s management is using AI to understand the return on investment of the company’s marketing spend

We’ve got pretty sophisticated algorithms that work on is this bid — is this click worth this much right now and how much should we bid. And so to the extent that CPCs rise, we naturally pull back. Or to the extent that CPC is lower, we naturally lean in. The other thing, by the way, it’s not just CPCs, it’s also conversion rates. So in times when people are really budget constrained, we see them actually — we see conversion rate across the industry go down. We see people compare some shop a lot more. And so we are looking at all of that and not humans, but machines using AI are looking at a very sophisticated way at what’s happening with conversion rate right now, what’s happening with CPCs right now. And therefore, how much is each visit worth and how much should we be bidding.

Fiverr (NYSE: FVRR)

Fiverr’s management is seeing strong demand for the AI services vertical, with AI-related keyword searches growing sevenfold in 2023

Early in January last year, we were the first in the market to launch a dedicated AI services vertical, creating a hub of businesses to higher AI talent. Throughout the year, we continue to see tremendous demand for those services with searches that contain AI-related keywords in our market base growing sevenfold in 2023 compared to 2022.

Fiverr’s management has seen AI create a net-positive 4% impact to Fiverr’s business by driving a mix-shift for the company from simple services – such as translation and voice-over – to complex services; complex services now represent 1/3 of Fiverr’s market base are typically larger and longer-duration; complex categories are where a human touch is needed and adds value while simple categories are where technology can do a good job without humans; Fiverr’s management thinks that simple categories will be automated away by AI while complex categories will become more important

Overall, we estimate AI created a net positive impact of 4% to our business in 2023 as we see a category mix shift from simple services such as translation and voice over to more complex services such as mobile app development, e-commerce management or financial consulting. In 2023, complex services represented nearly 1/3 of our market base, a significant step-up from 2022. Moreover, there are typically larger projects and longer duration with an average transaction size 30% higher than those of simple services…

…What we’ve identified is there is a difference between what we call simple categories or tasks and more complex ones. And in the complex group, it’s really those categories that require human intervention and human inputs in order to produce a satisfactory results for the customer. And in these categories, we’re seeing growth that goes well beyond the overall growth that we’re seeing. And really, the simple ones are such where technology can actually do a pretty much gen-tie work, which in those cases, they’re usually associated with lower prices and shorter-term engagements…

…So our assumption is that some of the simple paths are going to be — continue to be automated, which, by the way, is nothing new. I mean, it happened before even before AI, automation has been a part of our lives. And definitely, the more complex services is where I think the growth potential definitely lies. This is why we called out the fact that we’re going to double down on these categories and services.

Fiverr’s management believes that the opportunities created by AI will outweigh the jobs that are displaced

We believe that the opportunities created by emerging technologies far outweigh the jobs they replace. Human talent continues to be an essential part of unlocking the potential of new technologies.

Fiverr’s management believes that AI will be a multiyear tailwind for the company

We are also seeing a shift into more sophisticated, highly skilled and longer-duration categories with bigger addressable market. Data shows our market base is built to benefit from these technologies and labor market changes. Unlike single vertical solutions with higher exposure to disruptive technologies and train changes, Fiverr has developed a proprietary horizontal platform with hundreds of verticals, quickly leaning into the ever-changing industry demand needs and trends. All in all, we believe AI will be a multiyear tailwind for us to drive growth and innovation. In 2023, we also made significant investments in AI that drove improvements in our overall platform.

A strategic priority for Fiverr’s management in 2024 is to develop AI tools to enhance the overall customer experience of the company’s marketplace

Our recent winter product release in January culminated these efforts in the second half of 2023 and revamped almost every part of our platform with an AI-first approach, from search to personalization from supply quality to seller engagement…

…Our third strategic priority is to continue developing proprietary AI applications unique to our market base to enhance the overall customer experience. The winter product release we discussed just now gives you a flavor of that, but there is so much more to do.

Mastercard (NYSE: MA)

Mastercard’s management is leveraging the company’s work on generative AI to build new services and solutions as well as to increase internal productivity

We also continue to develop new services and solutions, many of which leverage the work we are doing with generative AI. Generative AI brings more opportunity to drive better experiences for our customers, makes it easier to extract insights from our data. It can also help us increase internal productivity. We are working on many Gen AI use cases today to do just that. For example, we recently announced Shopping News. Shopping News uses generative AI to offer a conversational shopping tool that recreates the in-store human experience online, can translate consumers collegially language into tailored recommendations. Another example is Mastercard Small Business AI. The tool will draw on our existing small business resources, along with the content from a newly formed global media coalition to help business owners navigate a range of business challenges. The platform, which is scheduled for pilot launch later this year will leverage AI to provide personalized real-time assistance delivered in a conversational tone.

MercadoLibre (NASDAQ: MELI)

MercadoLibre’s management launched a number of AI features – including a summary of customer reviews, a summary of product functions, push notifications about items left unpurchased in shopping carts, and capabilities for sellers to create coupons and answer buyer questions quickly – in 2023 for the ecommerce business

In 2023, we launched capabilities that enable sellers to create their own promotional coupons and answer buyer questions more quickly with the assistance of artificial intelligence…

…AI based features are already an integral part of the MELI experience, with many innovations launched in 2023, including:

A summary of customer reviews on the product pages that concentrates the main feedback from buyers of that product.
On beauty product pages a summary of product functions and characteristics is automatically created to facilitate buyers choices.
Push notifications about items left unpurchased in shopping carts are now highly personalized and remind users why they may have chosen to buy a particular product.
We have also added an AI feature that helps sellers to respond to questions by preparing answers that sellers can send immediately, or edit quickly.

Meta Platforms (NASDAQ: META)

The major goal of Meta’s management is for the company is to have (1) world-class AI assistant for all users, (2) AI-representor for each creator, (3) AI agent for every business, and (4) state-of-the-art open source models for developers

Now moving forward, a major goal, we’ll be building the most popular and most advanced AI products and services. And if we succeed, everyone who uses our services will have a world-class AI assistant to help get things done, every creator will have an AI that their community can engage with, every business will have an AI that their customers can interact with to buy goods and get support, and every developer will have a state-of-the-art open-source model to build with.

Meta’s management thinks consumers will want a new AI-powered computing device that can see and hear what we are seeing and hearing, and this new computing device will be smart glasses, and will require full general intelligence; Meta has been conducting research on general intelligence for more than a decade, but it will now also incorporate general intelligence into product work – management thinks having product-targets when developing general intelligence helps to focus the work

I also think that everyone will want a new category of computing devices that let you frictionlessly interact with AIs that can see what you see and hear what you hear, like smart glasses. And one thing that became clear to me in the last year is that this next generation of services requires building full general intelligence. Previously, I thought that because many of the tools were social-, commerce- or maybe media-oriented that it might be possible to deliver these products by solving only a subset of AI’s challenges. But now it’s clear that we’re going to need our models to be able to reason, plan, code, remember and many other cognitive abilities in order to provide the best versions of the services that we envision. We’ve been working on general intelligence research and FAIR for more than a decade. But now general intelligence will be the theme of our product work as well…

…We’ve worked on general intelligence in our lab, FAIR, for more than a decade, as I mentioned, and we produced a lot of valuable work. But having clear product targets for delivering general intelligence really focuses this work and helps us build the leading research program.

Meta’s management believes the company has world-class compute infrastructure; Meta will end 2024 with 600,000 H100 (NVIDIA’s state-of-the-art AI chip) equivalents of compute; Meta is coming up with new data centre and chip designs customised for its own needs

The first is world-class compute infrastructure. I recently shared that, by the end of this year, we’ll have about 350,000 H100s, and including other GPUs, that will be around 600,000 H100 equivalents of compute…

…In order to build the most advanced clusters, we’re also designing novel data centers and designing our own custom silicons specialized for our workloads.

Meta’s management thinks that future AI models will be even more compute-intensive to train and run inference; management does not know exactly how much the compute this will be, but recognises that the trend has been of AI models requiring 10x more compute for each new generation, so management expects Meta to require growing infrastructure investments in the years ahead for its AI work

Now going forward, we think that training and operating future models will be even more compute-intensive. We don’t have a clear expectation for exactly how much this will be yet, but the trend has been that state-of-the-art large language models have been trained on roughly 10x the amount of compute each year…

…While we are not providing guidance for years beyond 2024, we expect our ambitious long-term AI research and product development efforts will require growing infrastructure investments beyond this year.

Meta’s approach with AI is to open-source its foundation models while keeping product-implementations proprietary; Meta’s management thinks open-sourcing brings a few key benefits, in that open source software (1) is safer and more compute-efficient, (2) can become the industry standard, and (3) attracts talented people; management intends to continue open-sourcing Meta’s AI models

Our long-standing strategy has been to build an open-source general infrastructure while keeping our specific product implementations proprietary. In the case of AI, the general infrastructure includes our Llama models, including Llama 3, which is training now, and it’s looking great so far, as well as industry standard tools like PyTorch that we’ve developed…

…The short version is that open sourcing improves our models. And because there’s still significant work to turn our models into products because there will be other open-source models available anyway, we find that there are mostly advantages to being the open-source leader, and it doesn’t remove differentiation for our products much anyway. And more specifically, there are several strategic benefits.

First, open-source software is typically safer and more secure as well as more compute-efficient to operate due to all the ongoing feedback, scrutiny and development from the community. Now this is a big deal because safety is one of the most important issues in AI. Efficiency improvements and lowering the compute costs also benefit everyone, including us. Second, open-source software often becomes an industry standard. And when companies standardize on building with our stack, that then becomes easier to integrate new innovations into our products. That’s subtle, but the ability to learn and improve quickly is a huge advantage. And being an industry standard enables that. Third, open source is hugely popular with developers and researchers. And we know that people want to work on open systems that will be widely adopted. So this helps us recruit the best people at Meta, which is a very big deal for leading in any new technology area…

…This is why our long-standing strategy has been to open source general infrastructure and why I expect it to continue to be the right approach for us going forward.

Meta is already training the next generation of its foundational Llama model, Llama 3, and progress is good; Meta is also working on research for the next generations of Llama models with an eye on developing full general intelligence; Meta’s management thinks that the company’s next few generations of foundational AI models could be in a totally different direction from other AI companies

In the case of AI, the general infrastructure includes our Llama models, including Llama 3, which is training now, and it’s looking great so far…

…While we’re working on today’s products and models, we’re also working on the research that we need to advance for Llama 5, 6 and 7 in the coming years and beyond to develop full general intelligence…

…A lot of last year and the work that we’re doing with Llama 3 is basically making sure that we can scale our efforts to really produce state-of-the-art models. But once we get past that, there’s a lot more kind of different research that I think we’re going to be doing that’s going to take our foundation models in potentially different directions than other players in the industry are going to go in because we’re focused on specific vision for what we’re building. So it’s really important as we think about what’s going to be in Llama 5 or 6 or 7 and what cognitive abilities we want in there and what modalities we want to build into future multimodal versions of the models.

Meta’s management sees unique feedback loops for the company’s AI work that involve both data and usage of its products; the feedback loops have been important in how Meta improved its AI systems for Reels and ads

When people think about data, they typically think about the corpus that you might use to train a model upfront. And on Facebook and Instagram, there are hundreds of billions of publicly shared images and tens of billions of public videos, which we estimate is greater than the common crawl data set. And people share large numbers of public text posts and comments across our services as well. But even more important in the upfront training corpus is the ability to establish the right feedback loops with hundreds of millions of people interacting with AI services across our products. And this feedback is a big part of how we’ve improved our AI systems so quickly with Reels and Ads, especially over the last couple of years when we had to re-architect it around new rules.

Meta’s management wants hiring-growth in AI-related roles for 2024

AI is a growing area of investment for us in 2024 as we hire to support our road map…

…Second, we anticipate growth in payroll expenses as we work down our current hiring underrun and add incremental talent to support priority areas in 2024, which we expect will further shift our workforce composition toward higher-cost technical roles.

Meta’s management fully rolled out Meta AI Assistant and other AI chat experiences in the US at the end of 2023 and has began testing generative AI features in the company’s Family of Apps; Meta’s focus in 2024 regarding generative AI is on launching Llama3, making Meta AI assistant useful, and improving AI Studio

With generative AI, we fully rolled out our Meta AI assistant and other AI chat experiences in the U.S. at the end of the year and began testing more than 20 GenAI features across our Family of Apps. Our big areas of focus in 2024 will be working towards the launch of Llama 3, expanding the usefulness of our Meta AI assistant and progressing on our AI Studio road map to make it easier for anyone to create an AI.

Meta has been using AI to improve its marketing performance; Advantage+ is helping advertisers partially or fully automate the creation of ad campaigns; Meta has rolled out generative AI features to help advertisers with changing text and images in their ad campaigns – adoption of the features is strong and test show promising performance gains, and Meta has a big focus in this area in 2024

We continue to leverage AI across our ad systems and product suite. We’re delivering continued performance gains from ranking improvements as we adopt larger and more advanced models, and this will remain an ongoing area of investment in 2024. We’re also building out our Advantage+ portfolio of solutions to help advertisers leverage AI to automate their advertising campaigns. Advertisers can choose to automate part of the campaign creation setup process, such as who to show their ad to with Advantage+ audience, or they can automate their campaign completely using Advantage+ shopping, which continues to see strong growth. We’re also now exploring ways to apply this end-to-end automation to new objectives. On the ads creative side, we completed the global rollout of 2 of our generative AI features in Q4, Text Variations and Image Expansion, and plan to broaden availability of our background generation feature later in Q1. Initial adoption of these features has been strong, and tests are showing promising early performance gains. This will remain a big area of focus for us in 2024…

…So we’re really scaling our Advantage+ suites across all of the different offerings there, which really helped to automate the ads creation process for different types of advertisers. And we’re getting very strong feedback on all of those different features, advantage+ Shopping, obviously, being the first, but Advantage+ Catalog, Advantage+ Creative, Advantage+ Audiences, et cetera. So we feel like these are all really important parts of what has continued to grow improvements in our Ads business and will continue to going forward.

Meta’s management’s guidance for capital expenditure for 2024 is increased slightly from prior guidance (for perspective 2023’s capex is $27.27 billion), driven by increased investments in servers and data centers for AI-related work

Turning now to the CapEx outlook. We anticipate our full year 2024 capital expenditures will be in the range of $30 billion to $37 billion, a $2 billion increase of the high end of our prior range. We expect growth will be driven by investments in servers, including both AI and non-AI hardware, and data centers as we ramp up construction on sites with our previously announced new data center architecture.

Meta’s management thinks AI will make all of the company’s products and services better, but is unsure how the details will play out

I do think that AI is going to make all of the products and services that we use and make better. So it’s hard to know exactly how that will play out.

Meta’s management does not expect the company’s generative AI products to be a meaningful revenue-driver in the short term, but they expect the products to be huge drivers in the long term

We don’t expect our GenAI products to be a meaningful 2024 driver of revenue. But we certainly expect that they will have the potential to be meaningful contributors over time.

Microsoft (NASDAQ: MSFT)

Microsoft is now applying AI at scale, across its entire tech stack, and this is helping the company win customers

We have moved from talking about AI to applying AI at scale. By infusing AI across every layer of our tech stack, we are winning new customers and helping drive new benefits and productivity gains.

Microsoft’s management thinks that Azure offers (1) the best AI training and inference performance, (2) the widest range of AI chips, including those from AMD, NVIDIA, and Microsoft, and (3) the best selection of foundational models, including LLMs and SLMs (small language models); Azure AI now has 53,000 customers and more than 33% are new to Azure; Azure allows developers to deploy LLMs without managing underlying infrastructure

Azure offers the top performance for AI training and inference and the most diverse selection of AI accelerators, including the latest from AMD and NVIDIA as well as our own first-party silicon, Azure Maia. And with Azure AI, we provide access to the best selection of foundation and open source models, including both LLMs and SLMs all integrated deeply with infrastructure, data and tools on Azure. We now have 53,000 Azure AI customers. Over 1/3 are new to Azure over the past 12 months. Our new models of service offering makes it easy for developers to use LLMs from our partners like Cohere, Meta and Mistral on Azure without having to manage underlying infrastructure.

Azure grew revenue by 30% in 2023 Q4, with six points of growth from AI services; most of the six points of growth from AI services was driven by Azure Open AI

Azure and other cloud services revenue grew 30% and 28% in constant currency, including 6 points of growth from AI services. Both AI and non-AI Azure services drove our outperformance…

…Yes, Azure OpenAI and then OpenAI’s own APIs on top of Azure would be the sort of the major drivers. But there’s a lot of the small batch training that goes on, whether it’s out of [indiscernible] or fine-tuning. And then a lot of people who are starting to use models as a service with all the other new models. But it’s predominantly Azure OpenAI today.

Microsoft’s management believes the company has built the world’s most popular SLMs; the SLMs have similar performance to larger models, but can run on laptops and mobile devices; both startups and established companies are exploring the use of Microsoft’s Phi SLM for applications

We have also built the world’s most popular SLMs, which offer performance comparable to larger models but are small enough to run on a laptop or mobile device. Anchor, Ashley, AT&T, EY and Thomson Reuters, for example, are all already exploring how to use our SLM, Phi, for their applications.

Microsoft has added Open AI’s latest models to the Azure OpenAI service; Azure Open AI is seeing increased usage from AI-first start ups, and more than 50% of Fortune 500 companies are using it

And we have great momentum with Azure OpenAI Service. This quarter, we added support for OpenAI’s latest models, including GPT-4 Turbo, GPT-4 with Vision, DALL-E 3 as well as fine-tuning. We are seeing increased usage from AI-first start-ups like Moveworks, Poplexity, Symphony AI as well as some of the world’s largest companies. Over half of the Fortune 500 use Azure OpenAI today, including Ally Financial, Coca-Cola and Rockwell Automation. For example, at CES this month, Walmart shared how it’s using Azure OpenAI Service along with its own proprietary data and models to streamline how more than 50,000 associates work and transform how its millions of customers shop.

Microsoft’s management is integrating AI across the company’s entire data stack; Cosmo DB, which has vector search capabilities, is used by companies as a database for AI apps; KPMG, with the help of Cosmos DB, has seen a 50% increase in productivity for its consultants; Azure AI Search provides hybrid search that goes beyond vector search and Open AI is using it for ChatGPT

We are integrating the power of AI across the entire data stack. Our Microsoft Intelligent Data Platform brings together operational databases, analytics, governance and AI to help organizations simplify and consolidate their data estates. Cosmos DB is the go-to database to build AI-powered apps at any scale, powering workloads for companies in every industry from AXA and Kohl’s to Mitsubushi and TomTom. KPMG, for example, has used Cosmos DB, including its built-in native vector search capabilities, along with Azure OpenAI Service to power an AI assistant, which it credits with driving an up to 50% increase in productivity for its consultants… And for those organizations who want to go beyond in-database vector search, Azure AI Search offers the best hybrid search solution. OpenAI is using it for retrieval augmented generation as part of ChatGPT.

There are now more than 1.3 million GitHub Copilot subscribers, up 30% sequentially; more than 50,000 organisations use GitHub Copilot Business and Accenture alone will roll out GitHub Copilot to 50,000 of its developers in 2024; Microsoft’s management thinks GitHub Copilot is a core product for anybody who is working in software development

GitHub revenue accelerated to over 40% year-over-year, driven by all our platform growth and adoption of GitHub Copilot, the world’s most widely deployed AI developer tool. We now have over 1.3 million paid GitHub Copilot subscribers, up 30% quarter-over-quarter. And more than 50,000 organizations use GitHub Copilot Business to supercharge the productivity of their developers from digital natives like Etsy and HelloFresh to leading enterprises like Autodesk, Dell Technologies and Goldman Sachs. Accenture alone will roll out GitHub Copilot to 50,000 of its developers this year…

…Everybody had talked it’s become — it is the 1 place where it’s becoming standard issue for any developer. It’s like if you take away spellcheck from Word, I’ll be unemployable. And similarly, it will be like I think GitHub Copilot becomes core to anybody who is doing software development…

To increase GitHub Copilot’s ARPU (average revenue per user), and ARPUs for other Copilots for the matter, Microsoft’s management will lean on the improvement that the Copilots bring to a company’s operating leverage and ask for a greater share of value

Our ARPUs have been great but they’re pretty low. But frankly, even though we’ve had a lot of success, it’s not like we are a high-priced ARPU company. I think what you’re going to start finding is, whether it’s Sales Copilot or Service Copilot or GitHub Copilot or Security Copilot, they are going to fundamentally capture some of the value they drive in terms of the productivity of the OpEx, right? So it’s like 2 points, 3 points of OpEx leverage would go to some software spend. I think that’s a pretty straightforward value equation. And so that’s the first time. I mean, this is not something we’ve been able to make the case for before, whereas now I think we have that case.

Then even the horizontal Copilot is what Amy was talking about, which is at the Office 365 or Microsoft 365 level. Even there, you can make the same argument. Whatever ARPU we may have with E5, now you can say incrementally as a percentage of the OpEx, how much would you pay for a Copilot to give you more time savings, for example. And so yes, I think all up, I do see this as a new vector for us in what I’ll call the next phase of knowledge work and frontline work even and their productivity and how we participate.

And I think GitHub Copilot, I never thought of the tools business as fundamentally participating in the operating expenses of a company’s spend on, let’s say, development activity. And now you’re seeing that transition. It’s just not tools. It’s about productivity of your dev team.

Microsoft’s own research and external studies show that companies can see up to a 70% increase in productivity by using generative AI for specific tasks; early users of Copilot for Microsoft 365 became 29% faster in a number of tasks

Our own research as well as external studies show as much as 70% improvement in productivity using generative AI for specific work tasks. And overall, early Copilot for Microsoft 365 users were 29% faster in a series of tasks like searching, writing and summarizing.

Microsoft’s management believes that AI will become a first-class part of every personal computer (PC) in 2024

In 2024, AI will become first-class part of every PC. Windows PCs with built-in neural processing units were front and center at CES, unlocking new AI experiences to make what you do on your PC easier and faster, from searching for answers and summarizing e-mails to optimizing performance in battery efficiency. Copilot in Windows is already available on more than 75 million Windows 10 and Windows 11 PCs. And with our new Copilot Key, the first significant change to the Windows Keyboard in 30 years, providing one-click access.

Microsoft’s management thinks that AI is transforming Microsoft’s search and browser experience; Microsoft has created more than 5 billion images and conducted more than 5 billion chats to-date, with both doubling sequentially; Bing and Edge both took share in 2023 Q4

And more broadly, AI is transforming our search and browser experience. We are encouraged by the momentum. Earlier this month, we achieved a new milestone with 5 billion images created and 5 billion chats conducted to date, both doubling quarter-over-quarter and both Bing and Edge took share this quarter.

Microsoft’s management expects the company’s capital expenditure to increase materially in the next quarter because of cloud and AI infrastructure investments; management’s commitment to increase infrastructure investments is guided by customer demand and what they see as a substantial market opportunity; management feels good about where Microsoft is in terms of adding infrastructure capacity to meet AI computing demand

We expect capital expenditures to increase materially on a sequential basis, driven by investments in our cloud and AI infrastructure and the slip of a delivery date from Q2 to Q3 from a third-party provider noted earlier. As a reminder, there can be normal quarterly spend variability in the timing of our cloud infrastructure build-out…

…Our commitment to scaling our cloud and AI investment is guided by customer demand and a substantial market opportunity. As we scale these investments, we remain focused on driving efficiencies across every layer of our tech stack and disciplined cost management across every team…

…I think we feel really good about where we have been in terms of adding capacity. You started to see the acceleration in our capital expense starting almost a year ago, and you’ve seen it scale through that process.

Microsoft’s management is seeing that most of the AI activity taking place on Azure is for inference

[Question] On AI, where are we in the journey from training driving most of the Azure AI usage to inferencing?

[Answer] What you’ve seen for most part is all inferencing. So none of the large model training stuff is in any of our either numbers at all. Small batch training, so somebody is doing fine-tuning or what have you, that will be there but that’s sort of a minor part. So most of what you see in the Azure number is broadly inferencing.

New workloads in AI that happen on Azure starts with selecting a frontier model, fine-tuning that model, then inference

The new workload in AI obviously, in our case, it starts with 1 of the frontier — I mean, starts with the frontier model, Azure OpenAI. But it’s not just about just 1 model, right? So you — first, you take that model, you do all that jazz, you may do some fine-tuning. You do retrieval, which means you’re sort of either getting some storage meter or you’re eating some compute meters. And so — and by the way, there’s still a large model to a small model and that would be a training perhaps, but that’s a small batch training that uses essentially inference infrastructure. So I think that’s what’s happening.

Microsoft’s management believes that generative AI will change the entire tech stack, down to the core computer architecture; one such change is to separate data storage from compute, as in the case of one of Microsoft’s newer services, Fabric

[Question] Cloud computing changed the tech stack in ways that we could not imagine 10 years back. The nature of the database layer, the operating system, every layer just changed dramatically. How do you foresee generative AI changing the tech stack as we know it?

[Answer] I think it’s going to have a very, very foundational impact. In fact, you could say the core compute architecture itself changes, everything from power density to the data center design to what used to be the accelerator now is that sort of the main CPU, so to speak, or the main compute unit. And so I think — and the network, the memory architecture, all of it. So the core computer architecture changes, I think every workload changes. And so yes, so there’s a full — like take our data layer.

The most exciting thing for me in the last year has been to see how our data layer has evolved to be built for AI, right? If you think about Fabric, one of the genius of Fabric is to be able to say, let’s separate out storage from the compute layer. In compute, we’ll have traditional SQLs, we’ll have spark. And by the way, you can have an Azure AI drop on top of the same data lake, so to speak, or the lake house pattern. And then the business model, you can combine all of those different computes. So that’s the type of compute architecture. So it’s sort of a — so that’s just 1 example…

… I do believe being in the cloud has been very helpful to build AI. But now AI is just redefining what it means to have — what the cloud looks like, both at the infrastructure level and the app model.

Microsoft’s management is seeing a few big use cases emerging within Microsoft 365 Copilot: Summarisation of meetings and documents; “chatting” with documents and texts of past communications; and creation and completion of documents

In terms of what we’re seeing, it’s actually interesting if you look at the data we have, summarization, that’s what it’s like, number one, like I’m doing summarizations of teams, meetings inside of teams during the meeting, after the meeting, Word documents, summarization. I get something in e-mail, I’m summarizing. So summarization has become a big deal. Drafts, right? You’re drafting e-mails, drafting documents. So anytime you want to start something, the blank page thing goes away and you start by prompting and drafting.

Chat. To me, the most powerful feature is now you have the most important database in your company, which happens to be the database of your documents and communications, is now query-able by natural language in a powerful way, right? I can go and say, what are all the things Amy said I should be watching out for next quarter? And it will come out with great detail. And so chat, summarization, draft.

Also, by the way, actions, one of the most used things is, here’s a Word document. Go complete — I mean, create a PowerPoint for me. So those are the stuff that’s also beginning.

Microsoft’s management is seeing strong engagement growth with Microsoft 365 Copilot that gives them optimism

And the other thing I would add, we always talk about in enterprise software, you sell software, then you wait and then it gets deployed. And then after deployment, you want to see usage. And in particular, what we’ve seen and you would expect this in some ways with Copilot, even in the early stages, obviously, deployment happens very quickly. But really what we’re seeing is engagement growth. To Satya’s point on how you learn and your behavior changes, you see engagement grow with time. And so I think those are — just to put a pin on that because it’s an important dynamic when we think about the optimism you hear from us.

Nvidia (NASDAQ: NVDA)

Nvidia’s management believes that companies are starting to build the next generation of AI data centres; this next generation of AI data centres takes in data and transforms them into tokens, which are the output of AI models

At the same time, companies have started to build the next generation of modern Data Centers, what we refer to as AI factories, purpose-built to refine raw data and produce valuable intelligence in the era of generative AI…

…A whole new industry in the sense that for the very first time, a Data Center is not just about computing data and storing data and serving the employees of the company. We now have a new type of Data Center that is about AI generation, an AI generation factory, and you’ve heard me describe it as AI factories. But basically, it takes raw material, which is data. It transforms it with these AI supercomputers that NVIDIA built, and it turns them into incredibly valuable tokens. These tokens are what people experience on the amazing ChatGPT or Midjourney or search these days are augmented by that. All of your recommender systems are now augmented by that, the hyper-personalization that goes along with it. All of these incredible start-ups in digital biology generating proteins and generating chemicals and the list goes on. And so all of these tokens are generated in a very specialized type of Data Center. And this Data Center, we call it AI supercomputers and AI generation factories.

Nvidia’s management is seeing very strong demand for the company’s Hopper AI chips and expects demand to far outstrip supply

Demand for Hopper remains very strong. We expect our next generation products to be supply constrained as demand far exceeds supply…

…However, whenever we have new products, as you know, it ramps from 0 to a very large number, and you can’t do that overnight. Everything is ramped up. It doesn’t step up. And so whenever we have a new generation of products and right now, we are ramping H200s, there’s no way we can reasonably keep up on demand in the short term as we ramp.

Nvidia’s outstanding 2023 Q4 growth in Data Center revenue was driven by both training and inference of AI models; management estimates that 40% of Nvidia’s Data Center revenue in 2023 was for AI inference; the 40% estimate might even be understated, because recommendation systems that were driven by CPU approaches are now being driven by GPUs

In the fourth quarter, Data Center revenue of $18.4 billion was a record, up 27% sequentially and up 409% year-on-year…

…Fourth quarter Data Center growth was driven by both training and inference of generative AI and large language models across a broad set of industries, use cases and regions. The versatility and leading performance of our Data Center platform enables a high return on investment for many use cases, including AI training and inference, data processing and a broad range of CUDA accelerated workloads. We estimate in the past year, approximately 40% of Data Center revenue was for AI inference…

…The estimate is probably understated and — but we estimated it, and let me tell you why. Whenever — a year ago, a year ago, the recommender systems that people are — when you run the Internet, the news, the videos, the music, the products that are being recommended to you because, as you know, the Internet has trillions — I don’t know how many trillions, but trillions of things out there, and your phone is 3 inches squared. And so the ability for them to fit all of that information down to something such a small real estate is through a system, an amazing system called recommender systems.

These recommender systems used to be all based on CPU approaches. But the recent migration to deep learning and now generative AI has really put these recommender systems now directly into the path of GPU acceleration. It needs GPU acceleration for the embeddings. It needs GPU acceleration for the nearest neighbor search. It needs GPU accelerating for reranking. And it needs GPU acceleration to generate the augmented information for you. So GPUs are in every single step of a recommender system now. And as you know, a recommender system is the single largest software engine on the planet. Almost every major company in the world has to run these large recommender systems.

Nvidia’s management is seeing that all industries are deploying AI solutions

Building and deploying AI solutions has reached virtually every industry. Many companies across industries are training and operating their AI models and services at scale…

…One of the most notable trends over the past year is the significant adoption of AI by enterprises across the industry verticals such as Automotive, health care, and financial services.

Large cloud providers accounted for more than half of Nvidia’s Data Center revenue in 2023 Q4; Microsoft

In the fourth quarter, large cloud providers represented more than half of our Data Center revenue, supporting both internal workloads and external public cloud customers.

Nvidia’s management is finding that consumer internet companies have been early adopters of AI and they are one of Nvidia’s largest customer categories; consumer internet companies are using AI (1) in content recommendation systems to boost user engagement and (2) to generate content for advertising and to help content creators

The consumer Internet companies have been early adopters of AI and represent one of our largest customer categories. Companies from search to e-commerce, social media, news and video services and entertainment are using AI for deep learning-based recommendation systems. These AI investments are generating a strong return by improving customer engagement, ad conversation and click-through rates…

… In addition, consumer Internet companies are investing in generative AI to support content creators, advertisers and customers through automation tools for content and ad creation, online product descriptions and AI shopping assistance.

Nvidia’s management is observing that enterprise software companies are using generative AI to help their customers with productivity and they are already seeing commercial success

Enterprise software companies are applying generative AI to help customers realize productivity gains. All the customers we’ve partnered with for both training and inference of generative AI are already seeing notable commercial success. ServiceNow’s generative AI products in their latest quarter drove their largest ever net new annual contract value contribution of any new product family release. We are working with many other leading AI and enterprise software platforms as well, including Adobe, Databricks, Getty Images, SAP, and Snowflake.

There are both enterprises and startups that are building foundational large language models; these models are serving specific cultures, regions, and also industries

The field of foundation of large language models is thriving, Anthropic, Google, Inflection, Microsoft, OpenAI and xAI are leading with continued amazing breakthrough in generative AI. Exciting companies like Adept, AI21, Character.AI, Cohere, Mistral, Perplexity and Runway are building platforms to serve enterprises and creators. New startups are creating LLMs to serve the specific languages, cultures and customs of the world’s many regions. And others are creating foundation models to address entirely different industries like Recursion, pharmaceuticals and generative biomedicines for biology. These companies are driving demand for NVIDIA AI infrastructure through hyperscale or GPU-specialized cloud providers.

Nvidia’s AI infrastructure are used for autonomous driving; the automotive vertical accounted for more than $1 billion of Nvidia’s Data Center revenue in 2023, and Nvidia’s management thinks the automotive vertical is a big growth opportunity for the company

We estimate that Data Center revenue contribution of the Automotive vertical through the cloud or on-prem exceeded $1 billion last year. NVIDIA DRIVE infrastructure solutions include systems and software for the development of autonomous driving, including data ingestion, curation, labeling, and AI training, plus validation through simulation. Almost 80 vehicle manufacturers across global OEMs, new energy vehicles, trucking, robotaxi and Tier 1 suppliers are using NVIDIA’s AI infrastructure to train LLMs and other AI models for automated driving and AI cockpit applications. In effect, nearly every Automotive company working on AI is working with NVIDIA. As AV algorithms move to video transformers and more cars are equipped with cameras, we expect NVIDIA’s automotive Data Center processing demand to grow significantly…

…NVIDIA DRIVE Orin is the AI car computer of choice for software-defined AV fleet. Its successor, NVIDIA DRIVE Thor, designed for vision transformers offers more AI performance and integrates a wide range of intelligent capabilities into a single AI compute platform, including autonomous driving and parking, driver and passenger monitoring, and AI cockpit functionality and will be available next year. There were several automotive customer announcements this quarter. Li Auto, Great Wall Motor, ZEEKR, the premium EV subsidiary of Geely and Xiaomi EV all announced new vehicles built on NVIDIA.

Nvidia is developing AI solutions in the realm of healthcare

In health care, digital biology and generative AI are helping to reinvent drug discovery, surgery, medical imaging, and wearable devices. We have built deep domain expertise in health care over the past decade, creating the NVIDIA Clara health care platform and NVIDIA BioNeMo, a generative AI service to develop, customize and deploy AI foundation models for computer-aided drug discovery. BioNeMo features a growing collection of pre-trained biomolecular AI models that can be applied to the end-to-end drug discovery processes. We announced Recursion is making available for the proprietary AI model through BioNeMo for the drug discovery ecosystem.

Nvidia’s business in China is affected by the US government’s export restrictions concerning advanced AI chips; Nvidia has been building workarounds and have started shipping alternatives to China; Nvidia’s management expects China to remain a single-digit percentage of Data Center revenue in 2024 Q1; management thinks that while the US government wants to limit China’s access to leading-edge AI technology, it still wants to see Nvidia succeed in China

Growth was strong across all regions except for China, where our Data Center revenue declined significantly following the U.S. government export control regulations imposed in October. Although we have not received licenses from the U.S. government to ship restricted products to China, we have started shipping alternatives that don’t require a license for the China market. China represented a mid-single-digit percentage of our Data Center revenue in Q4, and we expect it to stay in a similar range in the first quarter…

…At the core, remember, the U.S. government wants to limit the latest capabilities of NVIDIA’s accelerated computing and AI to the Chinese market. And the U.S. government would like to see us be as successful in China as possible. Within those two constraints, within those two pillars, if you will, are the restrictions.

Nvidia’s management is seeing demand for AI infrastructure from countries become an additional growth-driver for the company

In regions outside of the U.S. and China, sovereign AI has become an additional demand driver. Countries around the world are investing in AI infrastructure to support the building of large language models in their own language on domestic data and in support of their local research and enterprise ecosystems…

…So we’re seeing sovereign AI infrastructure is being built in Japan, in Canada, in France, so many other regions. And so my expectation is that what is being experienced here in the United States, in the West will surely be replicated around the world.

Nvidia is shipping its Hopper AI chips with Infiniband networking; management believes that a combination of the company’s Hopper AI chips with Infiniband is becoming a de facto standard for AI infrastructure

The vast majority of revenue was driven by our Hopper architecture along with InfiniBand networking. Together, they have emerged as the de facto standard for accelerated computing and AI infrastructure.

Nvidia is on track to ramp shipments of the latest generation of its most advanced AI chips – the H200 – in 2024 Q2; the H200 chips have double the inference performance of its predecessor

We are on track to ramp H200 with initial shipments in the second quarter. Demand is strong as H200 nearly doubled the inference performance of H100.

Nvidia’s networking solutions has a revenue run-rate of more than $13 billion and the company’s Quantum Infiniband band solutions grew by more than five times in 2023 Q4 – but Nvidia is also working on its own Ethernet AI networking solution called Spectrum X, which is purpose-built for AI and performs better than traditional Ethernet for AI workloads; Spectrum X has attracted leading OEMs as partners, and Nvidia is on track to ship the solution in 2024 Q1; management still sees Infiniband the standard for AI-dedicated systems

Networking exceeded a $13 billion annualized revenue run rate. Our end-to-end networking solutions define modern AI data centers. Our Quantum InfiniBand solutions grew more than 5x year-on-year. NVIDIA Quantum InfiniBand is the standard for the highest-performance AI-dedicated infrastructures. We are now entering the Ethernet networking space with the launch of our new Spectrum-X end-to-end offering designed for an AI-optimized networking for the Data Center. Spectrum-X introduces new technologies over Ethernet that are purpose-built for AI. Technologies incorporated in our Spectrum switch, BlueField DPU and software stack deliver 1.6x higher networking performance for AI processing compared with traditional Ethernet. Leading OEMs, including Dell, HPE, Lenovo and Supermicro with their global sales channels are partnering with us to expand our AI solution to enterprises worldwide. We are on track to ship Spectrum-X this quarter…

…InfiniBand is the standard for AI-dedicated systems. Ethernet with Spectrum-X, Ethernet is just not a very good scale-out system. But with Spectrum-X, we’ve augmented, layered on top of Ethernet, fundamental new capabilities like adaptive routing, congestion control, noise isolation or traffic isolation so that we could optimize Ethernet for AI. And so InfiniBand will be our AI-dedicated infrastructure, Spectrum-X will be our AI-optimized networking

Nvidia’s AI-training-as-a-service-platform, DGX Cloud, has reached a $1 billion annualised revenue run rate, and is now available on all the major cloud service providers; Nvidia’s management believes that the company’s software business will become very significant over time, because of the importance of software when running AI-related hardware

We also made great progress with our software and services offerings, which reached an annualized revenue run rate of $1 billion in Q4. NVIDIA DGX Cloud will expand its list of partners to include Amazon’s AWS, joining Microsoft Azure, Google Cloud, and Oracle Cloud. DGX Cloud is used for NVIDIA’s own AI R&D and custom model development as well as NVIDIA developers. It brings the CUDA ecosystem to NVIDIA CSP partners…

…And the way that we work with CSPs, that’s really easy. We have large teams that are working with their large teams. However, now that generative AI is enabling every enterprise and every enterprise software company to embrace accelerated computing, and when it is now essential to embrace accelerated computing because it is no longer possible, no longer likely anyhow, to sustain improved throughput through just general-purpose computing, all of these enterprise software companies and enterprise companies don’t have large engineering teams to be able to maintain and optimize their software stack to run across all of the world’s clouds and private clouds and on-prem.

So we are going to do the management, the optimization, the patching, the tuning, the installed base optimization for all of their software stacks. And we containerize them into our stack called NVIDIA AI Enterprise. And the way we go to market with it is think of that NVIDIA AI Enterprise now as a run time like an operating system. It’s an operating system for artificial intelligence. And we charge $4,500 per GPU per year. And my guess is that every enterprise in the world, every software enterprise company that are deploying software in all the clouds and private clouds and on-prem will run on NVIDIA AI Enterprise, especially obviously, for our GPUs. And so this is going to likely be a very significant business over time.

Nvidia’s gaming chips also have strong generative AI capabilities, leading to better gaming performance

At CES, we announced our GeForce RTX 40 Super Series family of GPUs. Starting at $599, they deliver incredible gaming performance and generative AI capabilities. Sales are off to a great start. NVIDIA AI Tensor Cores and the GPUs deliver up to 836 AI TOPS, perfect for powering AI for gaming, creating and everyday productivity. The rich software stack we offer with our RTX DPUs further accelerates AI. With our DLSS technologies, 7 out of 8 pixels can be AI-generated, resulting up to 4x faster ray tracing and better image quality. And with the TensorRT LLM for Windows, our open-source library that accelerates inference performance for the latest large language models, generative AI can run up to 5x faster on RTX AI PCs.

Nvidia has announced new gaming AI laptops from every major laptop manufacturer; Nvidia has more than 100 million RTX PCs in its installed base, and management thinks the company is in a good position to lead the next wave of generative AI applications that are coming to the personal computer

At CES, we also announced a wave of new RTX 40 Series AI laptops from every major OEM. These bring high-performance gaming and AI capabilities to a wide range of form factors, including 14-inch and thin and light laptops. With up to 686 TOPS of AI performance, these next-generation AI PCs increase generative AI performance by up to 60x, making them the best performing AI PC platforms…

…NVIDIA is fueling the next wave of generative AI applications coming to the PC. With over 100 million RTX PCs in the installed base and over 500 AI-enabled PC applications and games, we are on our way.

Nvidia has a service that allows software developers to build state-of-the-art generative AI avatars

At CES, we announced NVIDIA Avatar Cloud Engine microservices, which allow developers to integrate state-of-the-art generative AI models into digital avatars. ACE won several Best of CES 2024 awards. NVIDIA has an end-to-end platform for building and deploying generative AI applications for RTX PCs and workstations. This includes libraries, SDKs, tools and services developers can incorporate into their generative AI workloads.

Nvidia’s management believes that generative AI cannot be done on traditional general-purpose computing – it has to be done on an accelerated computing framework

With accelerated computing, you can dramatically improve your energy efficiency. You can dramatically improve your cost in data processing by 20:1, huge numbers. And of course, the speed. That speed is so incredible that we enabled a second industry-wide transition called generative AI. In generative AI, I’m sure we’re going to talk plenty about it during the call. But remember, generative AI is a new application. It is enabling a new way of doing software, new types of software being created. It is a new way of computing. You can’t do generative AI on traditional general-purpose computing. You have to accelerate it.

The hardware supply chain of a Nvidia GPU is improving; the components that go into a Nvidia GPU is really complex
Our supply is improving. Overall, our supply chain is just doing an incredible job for us. Everything from, of course, the wafers, the packaging, the memories, all of the power regulators to transceivers and networking and cables, and you name it, the list of components that we ship. As you know, people think that NVIDIA GPUs is like a chip, but the NVIDIA Hopper GPU has 35,000 parts. It weighs 70 pounds. These things are really complicated things we’ve built. People call it an AI supercomputer for good reason. If you ever look at the back of the Data Center, the systems, the cabling system is mind-boggling. It is the most dense, complex cabling system for networking the world has ever seen. Our InfiniBand business grew 5x year-over-year. The supply chain is really doing fantastic supporting us. And so overall, the supply is improving.

Nvidia’s management is allocating chips fairly to all of the company’s customers

CSPs have a very clear view of our product road map and transitions. And that transparency with our CSPs gives them the confidence of which products to place and where and when. And so they know the timing to the best of our ability, and they know quantities and, of course, allocation. We allocate fairly. We allocate fairly, do the best of our — best we can to allocate fairly and to avoid allocating unnecessarily.

Nvidia’s management is seeing a lot of activity emerging from robotics companies

There’s just a giant suite of robotics companies that are emerging. There are warehouse robotics to surgical robotics to humanoid robotics, all kinds of really interesting robotics companies, agriculture robotics companies.

Nvidia’s installed base of hardware has been able to support every single innovation in AI technology because it is programmable

NVIDIA is the only architecture that has gone from the very, very beginning, literally at the very beginning when CNNs and Alex Krizhevsky and Ilya Sutskever and Geoff Hinton first revealed AlexNet, all the way through RNNs to LSTMs to every RLs to deep RLs to transformers to every single version and every species that have come along, vision transformers, multi-modality transformers that every single — and now time sequence stuff. And every single variation, every single species of AI that has come along, we’ve been able to support it, optimize our stack for it and deploy it into our installed base…

… We simultaneously have this ability to bring software to the installed base and keep making it better and better and better. So our customers’ installed base is enriched over time with our new software…

…on’t be surprised if in our future generation, all of a sudden, amazing breakthroughs in large language models were made possible. And those breakthroughs, some of which will be in software because they run CUDA, will be made available to the installed base. And so we carry everybody with us on the one hand, we make giant breakthroughs on the other hand.

A big difference between accelerated computing and general purpose computing is the importance of software in the former

As you know, accelerated computing is very different than general-purpose computing. You’re not starting from a program like C++. You compile it and things run on all your CPUs. The stacks of software necessary for every domain from data processing, SQL versus SQL structured data versus all the images and text and PDF, which is unstructured, to classical machine learning to computer vision to speech to large language models, all — recommender systems. All of these things require different software stacks. That’s the reason why NVIDIA has hundreds of libraries. If you don’t have software, you can’t open new markets. If you don’t have software, you can’t open and enable new applications. Software is fundamentally necessary for accelerated computing. This is the fundamental difference between accelerated computing and general-purpose computing that most people took a long time to understand. And now people understand that software is really key.

Nvidia’s management believes that generative AI has kicked off a massive new investment cycle for AI infrastructure

Generative AI has kicked off a whole new investment cycle to build the next trillion dollars of infrastructure of AI generation factories. We believe these two trends will drive a doubling of the world data center infrastructure installed base in the next 5 years and will represent an annual market opportunity in the hundreds of billions.

PayPal (NASDAQ: PYPL)

PayPal’s management will soon launch a new PayPal app that will utilise AI to personalise the shopping experience for consumers; management hopes to drive engagement with the app

This year, we’re launching and evolving a new PayPal app to create a situation. We will also leverage our merchant relationships and the power of AI to make the entire shopping experience personalized for consumers while giving them control over their data…

…The new checkout and app experiences we are rolling out this year will also create an engagement loop that will drive higher awareness of the various products we offer and drive higher adoption of our portfolio over time.

Shopify (NASDAQ: SHOP)

Shopify’s management launched nearly 12 AI-powered tools through the Shopify Magic product suite in 2023, including tools for AI-generated product descriptions and an AI commerce assistant; in recent weeks, management launched AI product image creating and editing tools within Shopify Magic; management will be introducing new modalities and text-to-image capabilities later this year

In 2023, we brought nearly a dozen AI-enabled tools through our Shopify Magic product suite. We’re one of the first platforms to bring AI-generated product descriptions to market and made solid progress towards building Sidekick, a first of its kind AI-enabled commerce assistant. As part of our winter edition a few weeks ago, we introduced new features to our Shopify Magic suite of AI tools. These new generative AI tools simplify and enhance product image editing directly within the product image editor in the Shopify admin. With Shopify Magic, merchants can now leverage AI to create stunning images and professional edits with just a few clicks or keywords, saving on cost and time. And given the significant advancements in AI in 2023, we plan to seize this enormous opportunity ahead of us and are excited to introduce new modalities and text to image capabilities to Shopify in 2024.

Shopify’s marketing-paybacks have improved by over 30% with the help of AI

In terms of marketing, the 2 areas, in particular, where we are leaning in this quarter are performance marketing and point-of-sale. Within performance marketing, our team has unlocked some opportunities to reach potential customers at highly attractive LTV to CAC and paybacks. In fact, tactics that we’ve implemented on some channels earlier this year including through the enhanced use of AI and automation have improved paybacks by over 30%, enabling us to invest more into these channels while still maintaining our operating discipline on the underlying unit economics.

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s management has increased the company’s capital expenditure materially over the last few years to capture the growth opportunities associated with AI

At TSMC, a higher level of capital expenditures is always correlated with higher growth opportunities in the following years. In the past few years, we have sharply increased our CapEx spending in preparation to capture or harvest the growth opportunities from HPC, AI and 5G megatrends.

TSMC’s management expects 2024 to be a healthy growth-year for the company with revenue growth in the low-to-mid 20s percentage range, driven by its 3nm technologies, 5nm technologies, and AI

Entering 2024, we forecast fabless semiconductor inventory to have returned to a [ handsome ] level exiting 2023. However, macroeconomic weakness and geopolitical uncertainties persist, potentially further weighing on consumer sentiment and the market demand. Having said that, our business has bottomed out on a year-over-year basis, and we expect 2024 to be a healthy growth year for TSMC, supported by continued strong ramp of our industry-leading 3-nanometer technologies, strong demand for the 5-nanometer technologies and robust AI-related demand.

TSMC’s management sees 2023 as the year that generative AI became important for the semiconductor industry, with TSMC as a key enabler; management thinks that the surge in AI-related demand in 2023 will drive an acceleration in structural demand for energy-efficient computing, and that AI will need to be supported by more powerful semiconductors – these are TSMC’s strengths

2023 was a challenging year for the global semiconductor industry, but we also witnessed the rising emergence of generative AI-related applications with TSMC as a key enabler…

…Despite the near-term challenges, our technology leadership enable TSMC to outperform the foundry industry in 2023, while we are positioning us to capture the future AI and high-performance computing-related growth opportunities…

…The surge in AI-related demand in 2023 supports our already strong conviction that the structural demand for energy-efficient computing will accelerate in an intelligent and connected world. TSMC is a key enabler of AI applications. No matter which approach is taken, AI technology is evolving to use more complex AI models as the amount of computation required for training and inference is increasing. As a result, AI models need to be supported by more powerful semiconductor hardware, which requires use of the most advanced semiconductor process technologies. Thus, the value of TSMC technology position is increasing, and we are all well positioned to capture the major portion of the market in terms of semiconductor component in AI. To address insatiable AI-related demand for energy-efficient computing power, customers rely on TSMC to provide the most leading edge processing technology at scale with a dependable and predictable cadence of technology offering.

Almost everyone important in AI is working with TSMC on its 2nm technologies

As process technology complexity increase, the engagement lead time with customers also started much earlier. Thus, almost all the AI innovators are working with TSMC, and we are observing a much higher level of customer interest and engagement at N2 as compared with N3 at a similar stage from both HPC and the smartphone applications.

TSMC’s management believes that the world has seen only the tip of the iceberg with AI

But on the other hand, AI is only in its nascent stage. Only last November, the first large language data is announced, ChatGPT announced. We only see the tip of the iceberg.

TSMC’s management believes that the use of AI could accelerate scientific innovation in the field of semiconductor manufacturing

So I want to give the industry an optimistic note that even though 1 nanometer or sub 1 nanometer could be challenging, but we have a new technology capability using AI to accelerate the innovation in science.

TSMC’s management still believes that its narrowly-defined AI business will grow at 50% annually; management also sees AI application process chips making up a high-teens weightage of TSMC’s revenue by 2027, up from a low-teens weightage mentioned in the 2023 second-quarter earnings call, because of a sudden increase in demand

But for TSMC, we look at ours here, the AI’s a CAGR, that’s the growth rate every year, it’s about 50%. And we are confident that we can capture more opportunities in the future. So that’s what we said that up to 2027, we are going to have high teens of the revenue from a very narrow, we defined the AI application process, not to mention about the networking, not to mention about all others, okay?…

…[Question] You mentioned that we have a very narrow definition, we call server AI processor contribution and that you said it can be high teens in 5 years’ time because the last time, we said low teens.

[Answer] The demand suddenly being increased since last — I think, last year, the first quarter up to March or April, when ChatGPT become popular, so customers respond quickly and asked TSMC to prepare the capacity, both in front end and the back end. And that’s why we have confidence that this AI’s revenue will increase. We only narrowed down to the AI application process, by the way. So we look at ours here, that we prepare the technology and the capacities in both our front end and also back end. And so we — it’s in the early stage so far today. We already see the increase, the momentum. And we expect — if you guys continue to track this one, the number will increase. I have confidence to say that, although I don’t know how much.

TSMC’s management is seeing AI chips being placed in edge-devices such as smartphones and PCs

And to further extend our value, actually, all the edge device, including smartphone, including the PC, they start to put the AI’s application inside. They have some kind of a neural process, for example, so the silicon content will be greatly increased.

Tesla (NASDAQ: TSLA)

Tesla has released version 12 of its FSD (Full Self Driving) software, which is powered end-to-end by AI (artificial intelligence); Tesla will soon release it to over 400,000 vehicles in North America; FSD v.12 is the first time AI has been used for pathfinding and vehicle controls, and within it, neural nets replaced over 330,000 lines of code

For full self-driving, we’ve released version 12, which is a complete architectural rewrite compared to prior versions. This is end-to-end artificial intelligence. So [ nothing but ] nets basically, photons in and controls out. And it really is quite a profound difference. This is currently just with employees and a few customers, but we will be rolling out to all who — all those customers in the U.S. who request full self-driving in the weeks to come. That’s over 400,000 vehicles in North America. So this is the first time AI has been used not just for object perception but for pathfinding and vehicle controls. We replaced 330,000 lines of C++ code with neural nets. It’s really quite remarkable.

Tesla’s management believes that Tesla is the world’s most efficient company at AI inference because the company, out of necessity, has had to wring the most performance out of 3-year-old hardware

I think Tesla is probably the most efficient company in the world for AI inference. Out of necessity, we’ve actually had to be extremely good at getting the most out of hardware because hardware 3 at this point is several years old. So I don’t — I think we’re quite far ahead of any other company in the world in terms of AI inference efficiency, which is going to be a very important metric in the future in many arenas.

Tesla’s management thinks that the AI technologies the company has developed for vehicles translates well into a humanoid robot (Optimus); Tesla’s vehicles and Optimus both have the same inference computers

And the technologies that we — the AI technologies we’ve developed for the car translate quite well to a humanoid robot because the car is just a robot on 4 wheels. Tesla is arguably already the biggest robot maker in the world. It’s just a 4-wheeled robot. So Optimus is a robot with — a humanoid robot with arms and legs, just by far the most sophisticated humanoid robot that’s being developed anywhere in the world…

…As we improve the technology in the car, we improve the technology in Optimus at the same time. It runs the same AI inference computer that’s on the car, same training technology. I mean we’re really building the future. I mean the Optimus lab looks like the set of Westworld, but admittedly, that was not a super utopian situation.

Tesla’s management is hedging their bets for the company’s FSD-related chips with Nvidia’s GPUs while also pursuing Dojo (Tesla’s own AI chip design)

[Question] As a follow-up, your release does not mention Dojo, so if you could just provide us an update on where Dojo stands and at what point do you expect Dojo to be a resource in improving FSD. Or do you think that you now have sufficient supply of NVIDIA GPUs needed for the training of the system?

[Answer] I mean the AI part of your question is — that is a deep one. So we’re obviously hedging our bets here with significant orders of NVIDIA GPUs…

…And we’re pursuing the dual path of NVIDIA and Dojo.

Tesla’s management believes that Tesla’s progress in self-driving is limited by training and that in AI, the more training is done on the model, the less resources are required for inference

A lot of our progress in self-driving is training limited. Something that’s important with training, it’s much like a human. The more effort you put into training, the less effort you need in inference. So just like a person, if you train in a subject, sort of class, 10,000 hours, the less mental effort it takes to do something. If you remember when you first started to drive how much of your mental capacity it took to drive, it was — you had to be focused completely on driving. And after you’ve been driving for many years, it only takes a little bit of your mind to drive, and you can think about other things and still drive safely. So the more training you do, the more efficient it is at the inference level. So we do need a lot of training. And we’re pursuing the dual path of NVIDIA and Dojo, A

Tesla’s management thinks that Dojo is a long shot – it has potential, but may not work out

But I would think of Dojo as a long shot. It’s a long shot worth taking because the payoff is potentially very high but it’s not something that is a high probability. It’s not like a sure thing at all. It’s a high risk, high payoff program. Dojo is working, and it is doing training jobs, so — and we are scaling it up. And we have plans for Dojo 1.5, Dojo 2, Dojo 3 and whatnot. So I think it’s got potential. I can’t emphasize enough, high risk, high payoff.

Tesla’s management thinks that Tesla’s AI-inference hardware in its vehicles can enable the company to perhaps possess the largest amount of compute resources for AI tasks in the world at some point in the future

There’s also our inference hardware in the car, so we’re now on what’s called Hardware 4, but it’s actually version 2 of the Tesla-designed AI inference chip. And we’re about to complete design of — the terminology is a bit confusing. About to complete design of Hardware 5, which is actually version 3 of the Tesla-designed chip because the version 1 was Mobileye. Version 2 was NVIDIA, and then version 3 was Tesla. So — and we’re making gigantic improvements from 1 — from Hardware 3 to 4 to 5. I mean there’s a potentially interesting play where when cars are not in use in the future, that the in-car computer can do generalized AI tasks, can run a sort of GPT4 or 3 or something like that. If you’ve got tens of millions of vehicles out there, even in a robotaxi scenario, whether in heavy use, maybe they’re used 50 out of 168 hours, that still leaves well over 100 hours of time available — of compute hours. Like it’s possible with the right architectural decisions that Tesla may, in the future, have more compute than everyone else combined.

The Trade Desk (NASDAQ: TSLA)

Trade Desk’s management believes that in a post-cookie world, advertisers will have to depend on authentication, new approaches to identity, first-party data, and AI-driven relevance tools – Trade Desk’s tools help create the best outcome in this world

The post-cookie world is one that will combine authentication, new approaches to identity, first-party data activation and advanced AI-driven relevance tools, all to create a new identity fabric for the Internet that is so much more effective than cookies ever were. The Internet is being replumbed and our product offerings create the best outcome for all of the open Internet.

AI optimisations are distributed across Kokai, which is Trade Desk’s new platform that recently went live; Kokai helps advertisers understand and score every ad impression, and allows advertisers to use an audience-first approach in campaigns

In particular, Kokai represents a completely new way to understand and score the relevance of every ad impression across all channels. It allows advertisers to use an audience-first approach to their campaigns, targeting their audiences wherever they are on the open Internet. Our AI optimizations, which are now distributed across the platform, help optimize every element of the ad purchase process. Kokai is now live, and similar to Next Wave and Solimar, it will scale over the next year.

Based on Trade Desk’s management’s interactions with customers, the use of AI to forecast the impacts that advertisers’ decisions will have on their ad spending is a part of Kokai that customers love

A big part of what they love, to answer your question about what are they most excited about, is we have streamlined our reporting. We’ve made it way faster. There are some reports that you just have to wait multiple minutes for it because they’re just so robust, and we found ways to accelerate that. We’ve also added AI throughout the platform, especially in forecasting. So it’s a little bit like if you were to make a hypothetical trade in a trading platform for equity and then us tell you what we think is going to happen to the price action in the next 10 minutes. So we’re showing them what the effects of their changes are going to be before they even make them so that they don’t make mistakes. Because sometimes what happens is people put out a campaign. They’ll put tight restrictions on it. They’ll hope that it spends, then they come back a day or 2 or even 3 later and then realize they made it so difficult with their combination of targeting and pricing for us to buy anything that they didn’t spend much money. Or the opposite because they spent more and it wasn’t as effective as they wanted. So helping them see all of that before they do anything helped.

Trade Desk’s management believes that the company is reinforcing itself as the adtech AI leader; Trade Desk has been using AI in its platform since 2016

We are reinforcing our position as the adtech AI leader. We’ve been embedding AI into our platform since 2016, so it’s nothing new to us. But now it’s being distributed across our platform so our clients can make even better choices among the 15 million ad impression opportunities a second and understand which of those ads are most relevant to their audience segments at any given time.

Wix (NASDAQ: WIX)

Wix’s management added new AI features in 2023 to help users create content more easily; the key AI features introduced include a chat bot, code assistant, and text and image creators

This year, we meaningfully extended an already impressive toolkit of AI capabilities to include new AI-powered features that will help Wix users create visual and written web content more easily, optimized design and content layout, right code and manage their website and businesses more efficiently. The key AI product introduced in the last year include an AI chat experience for businesses, responsive AI design, AI code assistant, AI Meta Tag Creators and AI text and image creators among several other AI design tools.

Wix’s management recently released an AI site generator that can create a full-blown, tailored, ready-to-publish website based on user prompts; management believes that Wix is the first to launch such an AI site generator; the site generator has received fantastic feedback so far, and is a good starting point for creating a new website, but it is only at Version 1

We also recently released our AI site generator and have heard fantastic feedback so far. I believe this will be the first AI tool on the market that creates a full-blown, tailored and ready-to-publish website integrated with relevant business application based on user prompt…

… So we released what I would call version 1. It’s a great way for people to start with the website, meaning that you come in and you say, I’m a Spa in New York City and I specialize in some specific things. And we’ll — and AI will interview you on the — what makes your business unique, where are you located? How many people? Tell us about those people and the staff members. And as a result, we generate a website for you that is — has all the great content, right? And the content will be text and images. The other thing that then will actually get you to this experience where you can choose how you want to have the design look like. And the AI will generate different designs for you. So you can tell why I like this thing, I want a variation on that, I don’t like the colors, please change the colors or I want colors that are more professionals or I want color that are blue and yellow. And there I will do it for you.

On the other hand, you can also say, well, I don’t really like the design, can you generate something very different or generate a small variation of that, in many ways, a bit similar to Midjourney, what Midjourney is doing with the images, we are doing with a full-blown website. The result of that is something that is probably 70% of the website that you need to have on average, right, sometime it’s 95%, but sometimes it’s less than that. So it gives you an amazing way to start your website and shortened the amount of work that you need to do by about 70% to 80%. I think it’s fantastic and very exciting. The result of that is something that is probably 70% of the website that you need to have on average, right, sometime it’s 95%, but sometimes it’s less than that. So it gives you an amazing way to start your website and shortened the amount of work that you need to do by about 70% to 80%. I think it’s fantastic and very exciting.

Wix’s management is seeing that the majority of the company’s new users today have adopted at least one AI tool and this has been a positive for Wix’s business

In fact, the majority of new users today are using at least 1 AI tool on the web creation journey. This has resulted in reduced friction and enhanced the creation experience for our users as well as increased conversion and improve monetization.

Wix’s management expects AI to be a driver of Wix’s growth in 2024 and beyond

We expect our AI technology to be a significant driver of growth in 2024 and beyond…

…Third, as Avishai mentioned, uptick of the milestone AI initiatives of 2023 has been incredible, and we expect to see ramping conversion and monetization benefits from our entire AI toolkit for both self-creators and partners this year…

…But then again, also 2025 will be much better than 2024. I think that the first reason is definitely the launching new products. At the end of the day, we are a technology, a product company, and this is how we drive our growth, mostly from new features, some new products. And this is what we did in the past, and we will continue also to do in the future. So definitely, it’s coming from the partners business with launching Studio. It was a great launch for us. We see the traction in the market. We see the demand. We see how our agencies use it. I think, as you know, we mentioned a few times about the number of new accounts with more than 50% are new. I think that it’s — for us, it’s a great proxy to the fact that we are going to see much more that it would be significantly the major growth driver for us in the next few years. The second one is everything that we’ve done with AI, we see a tremendous results out of it, which we believe that we will continue into the next year. And as you know, as always, the third one is about trying to optimize our pricing strategy. And this is what we’ve done in the past, we’ll continue to do in the future. [indiscernible] both mentioned like a fourth reason, which is the overall demand that we see on a macro basis.

Wix’s management has been driving the company to use AI for internal processes; the internal AI tools include an open internal AI development platform that everyone at Wix can contribute to, and a generative AI conversational assistant for product teams in Wix; the internal AI tools has also helped Wix to save costs and improve its gross margin

We also leverage AI to improve many of our internal processes at Wix, especially research and development velocity. This include an open internal AI deployment platform that allow for everyone at Wix to contribute to building AI-driven user features in tandem. We also have a Gen AI best platform dedicated to conversational assistant, which allow any product team at Wix to develop their own assistant tailored to specific user needs without having to start from scratch. With this platforms, we are able to develop and release high-quality AI-based features and tools efficiently and at scale…

…We ended 2023 with a total gross margin of 68%, an improvement of nearly 500 basis points compared to 2022. Throughout the year, we benefited from improved efficiencies in housing and infrastructure costs and optimization of support cost, partially aided by integrating AI into our workflows. Creative Subscriptions gross margin expanded to 82% in 2023. And Business Solutions gross margin grew to 29% for the full year as we continue to benefit from improving margin and new [indiscernible].

Wix’s management believes that there can be double-digit growth for the company’s self creators business in the long run partly because of AI products

And we mentioned that for self-creators in the long run, we believe that it will be a double-digit growth just because of that because it has the most effect of the macro environment which already started to see that it’s improving. But then again, also the new product and AI is 1 of the examples how we can bring increased conversion and also increase the growth of self-creators.

Zoom Video Communications (NASDAQ: ZM)

Zoom’s management launched Zoom AI Companion, a generative AI assistant, five months ago and it has been expanded to six Zoom products, all included at no extra cost to users; Zoom AI companion now has 510,000 accounts enabled and has created 7.2 million meeting summaries

Zoom AI Companion, our generative AI assistant, empowers customers and employees with enhanced productivity, team effectiveness and skills. Since its launch only five months ago, we expanded AI Companion to six Zoom products, all included at no additional cost to licensed users…

…Zoom AI companion have grown tremendously in just 5 months with over 510,000 accounts enabled and 7.2 million meeting summaries created as of the close of FY ’24.

Zoom’s future roadmap for AI is guided by driving customer value

Our future roadmap for AI is 100% guided by driving customer value. We are hard at work developing new AI capabilities to help customers achieve their unique business objectives and we’ll have more to share in a month at Enterprise Connect

Zoom’s Contact Center suite is an AI-first solution that includes AI Companion; Contact Center suite is winning in head-to-head competition against legacy incumbents

Our expanding Contact Center suite is a unified, AI-first solution that offers tremendous value to companies of all sizes seeking to strengthen customer relationships and deliver better outcomes. The base product includes AI Companion and our newly launched tiered pricing allows customers to add specialized CX capabilities such as AI Expert Assist, workforce management, quality management, virtual agent, and omnichannel support. Boosted by its expanding features, our contact center suite is beginning to win in head-to-head competition with the legacy incumbents.

Zoom Revenue Accelerator gained recognition from Forrester as an AI-powered tool for sales teams

Zoom Revenue Accelerator was recognized as a “Strong Performer” in The Forrester Wave™ in its first year of being covered – an amazing testament to its value as a powerful AI-enabled tool driving value for sales teams.

A financial services company, Convera, was attracted to Zoom’s products because of AI Companion

Finally, let me thank Convera, the World’s FX payments leader. Zoom Phone was the foundation of their Zoom engagement and from there they adopted the wider Zoom One platform in less than two years. Seeing the benefits of the tight integration of our products underpinned by AI Companion, they recently began to deeply leverage Zoom Team Chat in order to streamline their pre, during and post meeting communication all within the Zoom Platform.

Zoom is monetising AI on many fronts

We are monetizing AI on many fronts. You look at our Zoom AI Companion, right? So first of all, for our existing customers, because they all like the value we created, right, to generate meeting summary, meeting [indiscernible] and so on and so forth, because of that, we really do not — because customers, they’re also trying to reduce the cost. That’s why we do not charge the customers for those features. However, a lot of areas we can monetize. Take our AI Companion, for example. Enterprise customers, how to lever enterprise customer directionally, source data and also to build a tailored — the Zoom AI Companion for those customers, sort of like a customized Zoom AI Companion, we can monetize. And also look at all the services. Maybe I’ll just take Contact Center, for example. We are offering Zoom Virtual Agent, that’s one we can monetize. And recently, we announced 3 tiers of Zoom Contact Center product. The last one is per agent per month, we charge $149. The reason why, there are a few features. One of the feature is Zoom Expert Assist, right? All those features are empowered by AI features.

Zoom’s AI-powered Virtual Agent was deployed internally and has saved Zoom 400,000 agent hours per month, and handled more than 90% of inbound inquiries; Zoom’s management believes that Zoom’s AI features help improve companies’ agent-efficiency in contact centers

Zoom, we — internally, we deployed our Virtual Agent. Guess what? Every month, we saved 400,000 agent hours. And more than 90% inbound inquiries can be done by our Virtual Agent driven by the AI technology…

…If you look at our Zoom Meeting product, right, customer discovered that Zoom AI Companion to help you with the meeting summary. And after they discovered that feature and they would like to adopt that, right? Contact Center, exact same thing. And like Virtual Agent, Zoom Expert Assist, right, leverage those AI features. Manager kind of knows what’s going on in real time and also — and the agent while can have the AI, to get a real-time in order base and any update about these customers. All those AI features can dramatically improve the agent efficiency, right? That’s the reason why it’s kind of — will not take a much longer time for those agents to realize the value of the AI features because it’s kind of very easy to use. And I think that in terms of adoption rate, I feel like Contact Center AI adoption rate even probably faster than the other — the core features, so — core services.

Zoom’s management is seeing that having AI features at no additional cost to customers helps the company to attract users to Zoom Team Chat

[Question] And for Eric, what’s causing customers to move over to the Zoom chat function and off your main competitor like Teams? Just further consolidation onto one platform? Or is it AI Companion playing a larger role here, especially as you guys are including it as opposed to $30, $35 a month?

[Answer] Customers, they see — using their chat solution, they want to use AI, right? I send you — James, I send you a message. I want to leverage AI, send a long message. However, if you use other solutions, sometimes, other solutions itself, even without AI, it’s not free, right? And in our case, not only do we have core functionalities, but also AI Companion built in also at no additional cost. I can use — for any users, customers, you already have a Meeting license, Zoom Team Chat already built in, right? All the core features, you can use the Zoom AI Companion in order to leverage AI — write a chat message and so on and so forth. It works so well at no additional cost. The total cost of ownership of the Zoom Team Chat is much better than any other team chat solutions.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Adobe, Alphabet, Amazon, Apple, Datadog, Etsy, Fiverr, Mastercard, MercadoLibre, Meta Platforms, Microsoft, PayPal, Shopify, TSMC, Tesla, The Trade Desk, Wix, and Zoom. Holdings are subject to change at any time.

What We’re Reading (Week Ending 03 March 2024)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

Here are the articles for the week ending 03 March 2024:

1. The Future of Ecommerce That Wasn’t: An In-depth Look into What Went Wrong with Wish – Speedwell Research

A good explanation is a story that is hard to vary. If we did a postmortem of WebVan (grocery delivery) or Pets.com (online specialty store for Pets), what would we say went wrong? If we did the postmortem in 2006, most likely we would have said it was a silly and unrealistic idea. But if we were to do a postmortem now, with the existence of Instacart (grocery delivery) and Chewy (online specialty store for pets), how would our understanding change?

This is not a trivial exercise. It is far too easy to be dismissive about a failing business and think it was the entrepreneur’s ill-thought-out idea or just incompetence, but this does not hold scrutiny.

Look at Apple. For how many years was Steve Jobs and his insistence on not licensing the Mac operating system seen as the impetus for their failure? And the same thing happened again when the iPhone was released: analysts thought their unwillingness to license the phone’s iOS would ultimately lead to their demise. Now though, Apple’s success is attributed to their close integration and their proprietary software is a key selling point, which wouldn’t be possible if they licensed it.

If you took over Lego in 2004 when it was nearing bankruptcy, what would you diagnose as the problem? Would you have thought that with digital entertainment kids just don’t want to play with toy blocks anymore? Or would you have thought the focus on “noncore” activities like theme parks, clothing, and video game development were the issues? Perhaps the product was good but was simply too expensive? You know that today there are vastly more digital entertainment options than there were in 2004, they still have theme parks and video games, and their products are still expensive, so what was it?

If you were appointed CEO of Crocs in 2008 when their stock dropped 98% and was on the verge of entering bankruptcy, tell us that you wouldn’t be tempted to lay the blame on the aesthetics of the shoes. It is the most ridiculed shoe design with “ugly” virtually synonymous with Crocs and yet they now sell over 150 million of them a year. Again, what some people would identify as the problem of the business turned out to be a virtue…

…So, if we are saying Wish was unsuccessful because of their focus on cheap items with slow shipping, we shouldn’t be able to point to another company that did something similar and was successful…

…We will do one final analysis to estimate churn before concluding. However, we want to note that this analysis is unnecessary to make the point we are about to. If you simply saw that they lost users despite spending >80% of revenues on marketing, or almost ~$1.5bn, is there any explanation you would accept that could convince you the business was healthy? Imagine your Chief Marketing Officer just told you they spent $1.5bn to lose 2mn buyers and grow revenues 2%. How would anyone possibly see that as a good thing? And yet, with a little bump in numbers from Covid in 2020, it was overlooked by investors in favor of the hope of buying the next Amazon at IPO.

In the S-1 they disclose the following LTV chart. They calculate “LTV” as cumulative gross profits over a period of time attributable to new buyers acquired in a given cohort, divided by total new buyers, which is most certainly not what an LTV really is.

For example, let’s say a cohort generated $15 of gross profit in year one and then another $10 of gross profit in year two. They would add those two numbers up and say the “LTV” of the customer in year 2 is $25. Therefore, if you wanted to calculate how much each cohort generated in gross profits in a given year, you just have to take the difference between each year. In this example, this cohort generated $10 in gross profits in the second year versus $15 in the first year, suggesting ample churn. What you would want to see is each year’s incremental figure stay steady or increase.

The chart above shows cumulative gross profit by cohort. If it was a perfectly linear line, then that would mean in each period the cohort bought the same amount of goods as the previous period.

We will focus just on the 2017 cohort for simplicity. We annotated it to show how we estimated incremental gross profits. The average buyer from the 2017 cohort earns $15 in gross profits in year 1, which drops to $10 in year 2, and then to $6 in year 3. We can already see that by year 3, each cohort is generating about 1/3rd what it did in year 1, which suggests heavy churn. Remember that Wish’s payback period is about 2 years, which means it isn’t until year 3 they make that small incremental gross profit. And remember, this is just to pay back the initial marketing investment, not other S&M they spent on promotions to reengage that buyer.

Here, we can see that that the difference between the gross profit for total buyers divided by the gross profit per average active buyer gives us a churn estimate. At the end of year 1, 100% of buyers are active (by definition) and by year three that drops to 19%. That comes out to about 44% annual churn over two years. It is also noteworthy that the churn is much worse in the first year. A full 67% of buyers do not return after buying once.

Now, remember that their average payback period is under 2 years. That is rather problematic in the context of almost no one being left after 2 years! They have a thin amount of remaining users that not only need to cover all of the reengagement marketing, but also all of their G&A and R&D cost. And that’s before they can even make a profit!

This is a fundamentally broken business. Users do not stay long enough, they have to pay to get users to return, and users are not profitable…

…Earlier we said that an explanation is a story that cannot easily vary. Well, we have trouble figuring out exactly what the story is that cannot vary. There is nothing in principle wrong with an ecommerce offering catered to the consumers in the low-end of household earnings. Some would note that the low average order value would make it hard to make enough contribution profit per order, but that is essentially what Pinduoduo did in China, what Shopee is doing in Southeast Asia and Brazil, and what Temu is doing in the US. While we don’t know exactly if all of those initiatives will end up being profitable, it is hard to claim it is the idea itself that is rotten.

Clearly, Wish had a problem with both their high cost to acquire users and their ability to retain them. We know that Pinduoduo had a better customer acquisition engine piggy backing off of Tencent’s Weixin with preferred placement and the Community Group Buy model was a novel way to spur consumer sharing, free of charge. Shein had TikTok and went viral early on with “Shein hauls”, where influencers would post everything they purchased. They would later lean into influencer marketing on TikTok to much success. Amazon has Amazon Prime which helps retain users, and their optimal customer service helps keep customers satisfied at potential churn events. Wish was lacking something in the customer acquisition and retention area, but exactly what isn’t obvious.

Perhaps it was a mix of everything that individually created customer churn events from slow shipping to “unreliable shipping”, fraud, fake listings, sub-par customer service, inadequate item selection, poor item quality, inaccurate recommendations, or perhaps even internal issues. But again, other companies have survived similar or worse issues. And the longer the list, the more it speaks to our lack of confidence in any one variable. As an investor from the outside, it isn’t apparent what the key problem was, at least not to us.

What is crystal clear though is that there were issues since at least 2019, and some red flags prior. An investor only needed the company’s IPO prospectus to see these problems brewing, and could have avoided even worrying about any potential “narrative fallacy” by just focusing on the financials.

2. Bill Ackman: Investing, Financial Battles, Harvard, DEI, X & Free Speech | Lex Fridman Podcast #413 (partial transcript here) – Lex Fridman and Bill Ackman

Bill Ackman (57:12): So this was at the time of the Financial Crisis, circa November 2008. Real estate’s always been a kind of sector that I’ve been interested in. I began my career in the real estate business working for my dad, actually arranging mortgages for real estate developers. So I have kind of deep deep ties and interest in the business. General Growth was the second largest shopping mall company in the country – Simon Properties many people have heard of – General Growth was number two. They own some of the best malls in the country…

…General Growth the company, the CFO in particular, was very aggressive in the way that he borrowed money. He borrowed money from a kind of Wall Street – not long-term mortgages – but generally relatively short-term mortgages. He was pretty aggressive. As the value went up, he would borrow more and more against the assets and that helped the short-term results of the business. The problem was during the Financial Crisis, the market for what’s called CMBS – commercial mortgage backed securities – basically shut. And the company, because its debt was relatively short-term, had a lot of big maturities coming up that they had no ability to refinance. The market said, “oh my God, the lenders are going to foreclose and the shareholders are going to get wiped. The company’s going to go bankrupt, they’re going to get wiped out.” The stock went from $63 a share to 34 cents. There was a family, the Bucksbaum family owned I think about 25% of the company and they had a $5 billion stock that was worth $25 million or something by the time we bought a stake in the business.

What interested me was, I thought the assets were worth substantially more than the liabilities. The company had $27 billion of debt and had $100 million value of the equity, down from like $20 billion. And sort of an interesting place to start with a stock down 99%. But the fundamental drivers – the mall business – are occupancy, how occupied are the malls? Occupancy was up year on-year between ‘07 and ‘08. Interestingly, net operating income, which is kind of a measure of cash flow from the malls – that was up year-on-year. So the underlying fundamentals were doing fine. The only problem they had is they had billions of dollars of debt that they had to repay – they couldn’t repay

If you examine the bankruptcy code, it’s precisely designed for a situation like this where it’s this resting place you can go to restructure your business. Now the problem was that every other company that had gone bankrupt, the shareholders got wiped out. And so the market, seeing every previous example the shareholders get wiped out, the assumption is the stock is going to go to zero. That’s not what the bankruptcy code says. What the bankruptcy code says is that the value gets apportioned based on value, and if you could prove to a judge that the assets’ worth more than the liabilities, then the shareholders actually get to keep their investment in the company. And that was the bet we made.

So we stepped into the market. We bought 25% of the company in the open market. We had to pay up. It started at 34 cents – I think there were 300 million shares – so it was at a $100 million value. By the time we were done, we paid an average of – we paid $60 million for 25% of the business, so about $240 million for the equity of the company. And then we had to get on the board to convince the directors the right thing to do. The board was in complete panic, didn’t know what to do, spending a ton of money on advisers…

…And the key moment, if you’re looking for fun moments, is there’s a woman named Maddie Bucksbaum who was from the Bucksbaum family. Her cousin John was chairman of the board, CEO of the company. And I said – as she calls me after we disclose our stake in the company, she’s like “Billy Ackman, I’m really glad to see you here.” I met her – I don’t think it was a date – but I kind of met her in a social context when I was 25 or something. And she said, “I’m really glad to see you here and is there anything I can do to help you, call me.” I said, “Sure.” We kept trying to get on the board of the company, they wouldn’t invite us on. Couldn’t really run a proxy contest, not with a company going bankrupt, and their advisers actually were Goldman Sachs and they’re like, “You don’t want the fox in the hen house” and they were listening to their advisors. So I called Maddie up and I said, “Maddie, I need to get on the board of the company to help.” And she says, “I will call my cousin and I’ll get it done.” She calls back a few hours later, “You’ll be going on to the board.” I don’t know what she said, but she was convincing.

Next thing you know, I’m invited to the board of the company and the board is talking about the old equity of General Growth. Old equity is what you talk about when the shareholders are getting wiped out. I said, “No, no, no. This board represents the current equity of the company. I’m a major shareholder, John’s a major shareholder, there’s plenty of asset value here. This company should be able to be restructured for the benefit of shareholders.” And we led a restructuring for the benefit of shareholders and it took let’s say eight months and the company emerged from Chapter 11. We made an incremental investment into the company and the shareholders kept the vast majority of their investment. All the creditors got their face amount of their investment – par plus accrued interest. And it was a great outcome. All the employees kept their jobs, the malls stayed open, there was no liquidation. The bankruptcy system worked the way it should. I was in court all the time and the first meeting with the judge, the judge is like “Look, this would never have happened were it not for a financial crisis.” And once the judge said that, I knew we were going to be fine because the company had really not done anything fundamentally wrong – maybe a little too aggressive in how they borrowed money.

Stock went from 34 cents to $31 a share…

…Lex Fridman (1:05:44): How hard is it to learn some of the legal aspects of this? You mentioned bankruptcy code – I imagine it’s very sort of dense language and dense ideas and the loopholes and all that kind of stuff. If you’re just stepping in and you’ve never done distressed investing, how hard is it to figure out?

Bill Ackman (1:06:05): It’s not that hard. I literally read a book on distressed investing. Ben Branch or something, on distressed investing.

Lex Fridman (1:06:12): So you were able to pick up the intuition from that, just all the basic skills involved, the basic facts to know, all that kind of stuff.

Bill Ackman (1:06:20): Most of the world’s knowledge has already been written somewhere. You just got to read the right books.

3. Why is Google killing cookies? – Eric Benjamin Seufert

What is Google’s underlying motivation in deprecating third-party cookies in Chrome? Suspicion is warranted. Google’s mission statement for its Privacy Sandbox initiative is to “Protect [user] privacy online,” across its Chrome browser and its Android operating system (Google intends to deprecate its GAID Android identifier at some point). Cookies, unquestionably, present severe data leakage risks to consumers: they allow anonymous services to observe the web activities of users with little preventative recourse. But as I point out in this piece, “privacy” is an abstract social concept, and firms – but especially multi-trillion dollar market leaders – don’t make dramatic, sweeping policy changes absent commercial benefit. Believing that a company would utterly reform the mechanics of digital advertising solely in service of increased user privacy is as absurd as believing that two firms would engage in a merger as an expression of friendship. To not assume a commercial motive in cookie deprecation is naive.

Apple’s App Tracking Transparency (ATT) privacy policy is an apt example of this. Apple launched an international PR campaign championing the privacy safeguards of the iPhone following its introduction of ATT in April 2021. Yet as I point out in this piece, Apple collects and utilizes consumer data in the ways that ATT was ostensibly designed to prevent. Apple positions its use of install and purchase data collected via consumer engagement in apps that it doesn’t own as “ads personalization” and not “tracking.” Apple claims first-party privileges over this consumer data because Apple exerts (and is stridently maintaining a firm grip on) control over iOS payments, giving it exclusive, proprietary access to that data. And in a court filing from December 2023, Apple had the following to say about the logical contortions of its privacy policies (all emphasis from the document):

The Allow Apps to Request to Track setting governs whether apps can ask to track users across apps or websites owned by other companies, as Apple’s descriptions of the setting consistently make clear … Plaintiffs also include a screen shot of the Tracking disclosure, which explains that Apple “requires app developers to ask for permission before they track your activity across Apps or websites they don’t own.” … Given Apple’s extensive privacy disclosures, no reasonable user would expect that their actions in Apple’s apps would be private from Apple.

This isn’t to say that Google and Apple don’t employ well-meaning, intelligent, and highly effective people whose efforts are centered on promoting their conceptions of digital privacy. But digital privacy initiatives from publicly traded, multi-trillion-dollar corporations must be viewed in a broader commercial context…

…So given that Google must have a commercial motivation in deprecating cookies, what is it? The most obvious is simply margin expansion: Google’s network business, which serves ads on third-party websites and apps, will almost certainly suffer if the Privacy Sandbox is less effective for targeting and measurement than cookies (and early indicators suggest it is). If the economics of buying third-party open web inventory through Google’s tools degrades, some of that demand may simply be routed to Google’s owned-and-operated channels. And these channels feature much higher margin for Google than its Network business: Bernstein estimated in December 2022 that Google’s margin on Network revenue is 10%, while it’s 15% for YouTube and 55% for Search. As I argue in this piece, because advertising budgets are deployed against absolute performance, Google will likely lose some degree of top-line revenue if its Network business unit declines. But Google doesn’t need to shift all of the revenue from Network to these channels to maintain its current bottom line given the margin differentials: $1BN in Network revenue produces the same margin as $181MM in Search revenue.

4. Twitter thread on life and investing lessons from climbing Mt Kinabalu – Eugene Ng

1 | Hiking is a marathon, not a sprint. It is about finishing, whether or not you finish first or last it doesn’t matter, as long as you finish. There is no gold medals for the fastest, and only rescue for those who don’t. It only matters as long as you finish, and that you remain safe when you finish. Safety first, go slow.

It is the same with investing. Never be permanently wiped out, avoid all unlimited downside trades, and then you can focus on making asymmetric bets with unlimited upside…

…3 | I was in awe out the scale of the human labour require for the entire operations. We saw numerous porters carrying up 20-40+ kg of fresh water, food, furniture, equipment for the lodge where we stayed at. There were also a number of porters who carried up luggage for some climbers as well. Without them, the support, the ecosystem, none of these would have been possible for us to experience the climb.

It is ever easier than ever to get data, but we cannot be lazy. We need to learn to appreciate the ecosystem and what we have now with the internet, versus 50 years old with libraries and faxes. Use them to your advantage. Easily available does not mean everyone will actually read them. Do not confuse it.

4 | We had a fantastic mountain guide who was one of the oldest and fittest at 52 years and he has been doing it over 32 years old since he was 20 years old. He still does this three times a week and will retire next year. He was leading us in our walk through easier path with such a controlled and comfortable pace, like he is meditating. Without him, it would have been so much more difficult. Having the right leader to help guide you really matters.

Having the right mentors, the right people around you matters. They can have the right expertise, experience to share that can help you in your journey to become better and to avoid the pitfalls…

…6 | When ascending up and descending back down, it is not an individual effort but a collective team effort. The company matters. Without the right company to support you mentally on every step of the way makes so much of a difference, everyone has a role to play.

Like in investing, there are going to be ups and downs. The right people/investors to stand by you matters, and not run away at the first sight of trouble. Choosing carefully the right team to the best of your ability matters.

7 | Sometimes a member of your team is not going to be feeling very well or can be injured, it is being prepared and bringing along extra supplies, medical or food, and continuously supporting them with what you have physically and mentally. Remember that if they can’t finish, you can’t finish.

Know that the businesses that we invest in are not going to do well all the time, it is not going to be a straight line. There are going to be ups and downs, and they will zig and zag from time to time. We need to have the patience to stand by them through difficult times, and the good times, and not sell them.

8 | Run your own race at your own pace, sometimes you will overtake and sometimes others will overtake. Don’t be stressed by someone behind you trying to push you go faster. You set your own pace. If they want to overtake, just stand to the side and let them overtake, if not just chill. Separately, if you want to overtake someone slower, then just overtake on the side.

Find your own investment strategy that suits you best, that energises you. The real race is against yourself, not against others. There will always be someone who will do better than you in any given year, so chill. It is not about being the top 10% in a year, but the top 10% after 10 or 20 years and more…

…11 | Always remember to never get complacent and choose speed, or get distracted. Do a misstep over a loose rock, and you may just end up spraining your ankle (like me), and end up not finishing the climb. But thankfully, it was serious and painful, but it was still okay enough for me to complete the last 8km. It was insanely painfully with every descend as my right ankle landed on every step.

Never think highly of yourself. Stay humble, have humility. The moment you lose that, you stop listening, you stop absorbing, you stop learning, and then with a mis-step you might just result in eventual failure. Never do that.

12 | At the end, despite how much preparation, it is really willpower at the end that gets everyone through to the end. It can be so powerful, the human mind and the will power. Despite how tough it is, we were just highly focused on taking step at a time mindfully and carefully, that’s all that mattered. I sprained my right ankle horribly with 8km left, it was really painful, but I kept persisting, and my teammates were patient with me and walked slower. “Stay hard” by David Goggins was our slogan to keep us going.

Investing too is a slog, managers get paid to endure all the emotional and psychological elements with all the ups and downs. It is knowing when to keep pursuing and staying the course even especially when the going gets tough…

…14 | Memories over medals. We did not finish first, but we finished in the end, and that’s what matters. To Team Endurance!

If you beat the index after 10 or 20 years, you will be in the top quartile. You want to keep staying and playing the game, and keeping doing okay and eventually you will do very well.

5. Things I Don’t Know About AI – Elad Gil

In most markets, the more time passes the clearer things become. In generative AI (“AI”), it has been the opposite. The more time passes, the less I think I actually understand.

For each level of the AI stack, I have open questions…

…There are in some sense two types of LLMs – frontier models – at the cutting edge of performance (think GPT-4 vs other models until recently), and everything else. In 2021 I wrote that I thought the frontier models market would collapse over time into an oligopoly market due to the scale of capital needed. In parallel, non-frontier models would more commodity / pricing driven and have a stronger opensource presence (note this was pre-Llama and pre-Mistral launches).

Things seem to be evolving towards the above:

Frontier LLMs are likely to be an oligopoly market. Current contenders include closed source models like OpenAI, Google, Anthropic, and perhaps Grok/X.ai, and Llama (Meta) and Mistral on the open source side. This list may of course change in the coming year or two. Frontier models keep getting more and more expensive to train, while commodity models drop in price each year as performance goes up (for example, it is probably ~5X cheaper to train GPT-3.5 equivalent now than 2 years ago)

As model scale has gotten larger, funding increasingly has been primarily coming from the cloud providers / big tech. For example, Microsoft invested $10B+ in OpenAI, while Anthropic raised $7B between Amazon and Google. NVIDIA is also a big investor in foundation model companies of many types. The venture funding for these companies in contrast is a tiny drop in the ocean in comparison. As frontier model training booms in cost, the emerging funders are largely concentrated amongst big tech companies (typically with strong incentives to fund the area for their own revenue – ie cloud providers or NVIDIA), or nation states wanting to back local champions (see eg UAE and Falcon). This is impacting the market and driving selection of potential winners early.

It is important to note that the scale of investments being made by these cloud providers is dwarfed by actual cloud revenue. For example, Azure from Microsoft generates $25B in revenue a quarter. The ~$10B OpenAI investment by Microsoft is roughly 6 weeks of Azure revenue. AI is having a big impact on Azure revenue revently. Indeed Azure grew 6 percentage points in Q2 2024 from AI – which would put it at an annualized increase of $5-6B (or 50% of its investment in OpenAI! Per year!). Obviously revenue is not net income but this is striking nonetheless, and suggests the big clouds have an economic reason to fund more large scale models over time.

In parallel, Meta has done outstanding work with Llama models and recently announced $20B compute budget, in part to fund massive model training. I posited 18 months ago that an open source sponsor for AI models should emerge, but assumed it would be Amazon or NVIDIA with a lower chance of it being Meta. (Zuckerberg & Yann Lecunn have been visionary here)…

...Are cloud providers king-making a handful of players at the frontier and locking in the oligopoly market via the sheer scale of compute/capital they provide? When do cloud providers stop funding new LLM foundation companies versus continuing to fund existing? Cloud providers are easily the biggest funders of foundation models, not venture capitalists. Given they are constrained in M&A due to FTC actions, and the revenue that comes from cloud usage, it is rational for them to do so. This may lead / has led to some distortion of market dynamics. How does this impact the long term economics and market structure for LLMs? Does this mean we will see the end of new frontier LLM companies soon due to a lack of enough capital and talent for new entrants? Or do they keep funding large models hoping some will convert on their clouds to revenue?…

…What happens in China? One could anticipate Chinese LLMs to be backed by Tencent, Alibaba, Xiaomi, ByteDance and others investing in big ways into local LLMs companies. China’s government has long used regulatory and literal firewalls to prevent competition from non-Chinese companies and to build local, government supported and censored champions. One interesting thing to note is the trend of Chinese OSS models. Qwen from Alibaba for example has moved higher on the broader LMSYS leaderboards…

…How much of AI cloud adoption is due to constrained GPU / GPU arb? In the absence of GPU on the main cloud providers companies are scrambling to find sufficient GPU for their needs, accelerating adoption of new startups with their own GPU clouds. One potential strategy NVIDIA could be doing is preferentially allocating GPU to these new providers to decrease bargaining power of hyperscalers and to fragment the market, as well as to accelerate the industry via startups. When does the GPU bottleneck end and how does that impact new AI cloud providers? It seems like an end to GPU shortages on the main clouds would be negative for companies whose only business is GPU cloud, while those with more tools and services should have an easier transition if this were to happen…

…ChatGPT launched ~15 months ago. If it takes 9-12 months to decide to quit your job, a few months to do it, and a few months to brainstorm an initial idea with a cofounder, we should start to see a wave of app builders showing up now / shortly.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google), Amazon, Apple, Meta Platforms, Microsoft, and Tencent. Holdings are subject to change at any time.

Insights From Warren Buffett’s 2023 Shareholder’s Letter

There’s much to learn from Warren Buffett’s latest letter, including his thoughts on oil & gas companies and the electric utility industry.

One document I always look forward to reading around this time of the year is Warren Buffett’s annual Berkshire Hathaway shareholder’s letter. Over the weekend, Buffett published the 2023 edition. This letter is especially poignant because Buffett’s long-time right-hand man, the great Charlie Munger, passed away last November. Besides containing a touching eulogy from Buffett to Munger, the letter also had some fascinating insights from Buffett that I wish to document and share.

Without further ado (emphases are Buffett’s)…

The actions of a wonderful partner

Charlie never sought to take credit for his role as creator but instead let me take the bows and receive the accolades. In a way his relationship with me was part older brother, part loving father. Even when he knew he was right, he gave me the reins, and when I blundered he never – never –reminded me of my mistake.

It’s hard to tell a good business from a bad one

Within capitalism, some businesses will flourish for a very long time while others will prove to be sinkholes. It’s harder than you would think to predict which will be the winners and losers. And those who tell you they know the answer are usually either self-delusional or snake-oil salesmen.

Holding onto a great business – one that can deploy additional capital at high returns – for a long time is a recipe for building a great fortune

At Berkshire, we particularly favor the rare enterprise that can deploy additional capital at high returns in the future. Owning only one of these companies – and simply sitting tight – can deliver wealth almost beyond measure. Even heirs to such a holding can – ugh! – sometimes live a lifetime of leisure…

…You may be thinking that she put all of her money in Berkshire and then simply sat on it. But that’s not true. After starting a family in 1956, Bertie was active financially for 20 years: holding bonds, putting 1⁄3 of her funds in a publicly-held mutual fund and trading stocks with some frequency. Her potential remained unnoticed.

Then, in 1980, when 46, and independent of any urgings from her brother, Bertie decided to make her move. Retaining only the mutual fund and Berkshire, she made no new trades during the next 43 years. During that period, she became very rich, even after making large philanthropic gifts (think nine figures).

Berkshire’s size is now a heavy anchor on the company’s future growth rates

This combination of the two necessities I’ve described for acquiring businesses has for long been our goal in purchases and, for a while, we had an abundance of candidates to evaluate. If I missed one – and I missed plenty – another always came along.

Those days are long behind us; size did us in, though increased competition for purchases was also a factor.

Berkshire now has – by far – the largest GAAP net worth recorded by any American business. Record operating income and a strong stock market led to a yearend figure of $561 billion. The total GAAP net worth for the other 499 S&P companies – a who’s who of American business – was $8.9 trillion in 2022. (The 2023 number for the S&P has not yet been tallied but is unlikely to materially exceed $9.5 trillion.)

By this measure, Berkshire now occupies nearly 6% of the universe in which it operates. Doubling our huge base is simply not possible within, say, a five-year period, particularly because we are highly averse to issuing shares (an act that immediately juices net worth)…

…All in all, we have no possibility of eye-popping performance…

…Our Japanese purchases began on July 4, 2019. Given Berkshire’s present size, building positions through open-market purchases takes a lot of patience and an extended period of “friendly” prices. The process is like turning a battleship. That is an important disadvantage which we did not face in our early days at Berkshire.

Are there a dearth of large, great businesses outside of the USA?

There remain only a handful of companies in this country capable of truly moving the needle at Berkshire, and they have been endlessly picked over by us and by others. Some we can value; some we can’t. And, if we can, they have to be attractively priced. Outside the U.S., there are essentially no candidates that are meaningful options for capital deployment at Berkshire.

Markets can occasionally throw up massive bargains because of external shocks

Occasionally, markets and/or the economy will cause stocks and bonds of some large and fundamentally good businesses to be strikingly mispriced. Indeed, markets can – and will – unpredictably seize up or even vanish as they did for four months in 1914 and for a few days in 2001.

Stock market participants today exhibit even more gambling-like behaviour than in the past

Though the stock market is massively larger than it was in our early years, today’s active participants are neither more emotionally stable nor better taught than when I was in school. For whatever reasons, markets now exhibit far more casino-like behavior than they did when I was young. The casino now resides in many homes and daily tempts the occupants.

Stock buybacks are only sensible if they are done at a discount to business-value

All stock repurchases should be price-dependent. What is sensible at a discount to business-value becomes stupid if done at a premium.

Does Occidental Petroleum play a strategic role in the long-term economic security of the USA?

At yearend, Berkshire owned 27.8% of Occidental Petroleum’s common shares and also owned warrants that, for more than five years, give us the option to materially increase our ownership at a fixed price. Though we very much like our ownership, as well as the option, Berkshire has no interest in purchasing or managing Occidental. We particularly like its vast oil and gas holdings in the United States, as well as its leadership in carbon-capture initiatives, though the economic feasibility of this technique has yet to be proven. Both of these activities are very much in our country’s interest.

Not so long ago, the U.S. was woefully dependent on foreign oil, and carbon capture had no meaningful constituency. Indeed, in 1975, U.S. production was eight million barrels of oil-equivalent per day (“BOEPD”), a level far short of the country’s needs. From the favorable energy position that facilitated the U.S. mobilization in World War II, the country had retreated to become heavily dependent on foreign – potentially unstable – suppliers. Further declines in oil production were predicted along with future increases in usage.

For a long time, the pessimism appeared to be correct, with production falling to five million BOEPD by 2007. Meanwhile, the U.S. government created a Strategic Petroleum Reserve (“SPR”) in 1975 to alleviate – though not come close to eliminating – this erosion of American self-sufficiency.

And then – Hallelujah! – shale economics became feasible in 2011, and our energy dependency ended. Now, U.S. production is more than 13 million BOEPD, and OPEC no longer has the upper hand. Occidental itself has annual U.S. oil production that each year comes close to matching the entire inventory of the SPR. Our country would be very – very – nervous today if domestic production had remained at five million BOEPD, and it found itself hugely dependent on non-U.S. sources. At that level, the SPR would have been emptied within months if foreign oil became unavailable.

Under Vicki Hollub’s leadership, Occidental is doing the right things for both its country and its owners.

Nobody knows what the price of oil would do in the short-term and the long-term

No one knows what oil prices will do over the next month, year, or decade.

Nobody can predict the movement of major currencies

Neither Greg nor I believe we can forecast market prices of major currencies. We also don’t believe we can hire anyone with this ability. Therefore, Berkshire has financed most of its Japanese position with the proceeds from ¥1.3 trillion of bonds.

Rail is a very cost-efficient way to move products around America, and railroads should continue to be an important asset for the USA for a long time to come

Rail is essential to America’s economic future. It is clearly the most efficient way – measured by cost, fuel usage and carbon intensity – of moving heavy materials to distant destinations. Trucking wins for short hauls, but many goods that Americans need must travel to customers many hundreds or even several thousands of miles away…

…A century from now, BNSF will continue to be a major asset of the country and of Berkshire. You can count on that.

Railroad companies gobble up capital, such that its owners have to spend way more on annual maintenance capital expenditure than depreciation – but this trait allowed Berkshire to acquire BNSF for far less than its replacement value

BNSF is the largest of six major rail systems that blanket North America. Our railroad carries its 23,759 miles of main track, 99 tunnels, 13,495 bridges, 7,521 locomotives and assorted other fixed assets at $70 billion on its balance sheet. But my guess is that it would cost at least $500 billion to replicate those assets and decades to complete the job.

BNSF must annually spend more than its depreciation charge to simply maintain its present level of business. This reality is bad for owners, whatever the industry in which they have invested, but it is particularly disadvantageous in capital-intensive industries.

At BNSF, the outlays in excess of GAAP depreciation charges since our purchase 14 years ago have totaled a staggering $22 billion or more than $11⁄2 billion annually. Ouch! That sort of gap means BNSF dividends paid to Berkshire, its owner, will regularly fall considerably short of BNSF’s reported earnings unless we regularly increase the railroad’s debt. And that we do not intend to do.

Consequently, Berkshire is receiving an acceptable return on its purchase price, though less than it might appear, and also a pittance on the replacement value of the property. That’s no surprise to me or Berkshire’s board of directors. It explains why we could buy BNSF in 2010 at a small fraction of its replacement value.

Railroad companies are having trouble with hiring because of tough working conditions

An evolving problem is that a growing percentage of Americans are not looking for the difficult, and often lonely, employment conditions inherent in some rail operations. Engineers must deal with the fact that among an American population of 335 million, some forlorn or mentally-disturbed Americans are going to elect suicide by lying in front of a 100-car, extraordinarily heavy train that can’t be stopped in less than a mile or more. Would you like to be the helpless engineer? This trauma happens about once a day in North America; it is far more common in Europe and will always be with us.

American railroad companies are at times at the mercy of the US government when it comes to employees’ wages, and they are also required to carry products they would rather not

Wage negotiations in the rail industry can end up in the hands of the President and Congress. Additionally, American railroads are required to carry many dangerous products every day that the industry would much rather avoid. The words “common carrier” define railroad responsibilities.

Last year BNSF’s earnings declined more than I expected, as revenues fell. Though fuel costs also fell, wage increases, promulgated in Washington, were far beyond the country’s inflation goals. This differential may recur in future negotiations.

Has the electric utility industry in the USA become uninvestable because of a change in the authorities’ stance toward electric utilities?

For more than a century, electric utilities raised huge sums to finance their growth through a state-by-state promise of a fixed return on equity (sometimes with a small bonus for superior performance). With this approach, massive investments were made for capacity that would likely be required a few years down the road. That forward-looking regulation reflected the reality that utilities build generating and transmission assets that often take many years to construct. BHE’s extensive multi-state transmission project in the West was initiated in 2006 and remains some years from completion. Eventually, it will serve 10 states comprising 30% of the acreage in the continental United States.

With this model employed by both private and public-power systems, the lights stayed on, even if population growth or industrial demand exceeded expectations. The “margin of safety” approach seemed sensible to regulators, investors and the public. Now, the fixed-but-satisfactoryreturn pact has been broken in a few states, and investors are becoming apprehensive that such ruptures may spread. Climate change adds to their worries. Underground transmission may be required but who, a few decades ago, wanted to pay the staggering costs for such construction?

At Berkshire, we have made a best estimate for the amount of losses that have occurred. These costs arose from forest fires, whose frequency and intensity have increased – and will likely continue to increase – if convective storms become more frequent.

It will be many years until we know the final tally from BHE’s forest-fire losses and can intelligently make decisions about the desirability of future investments in vulnerable western states. It remains to be seen whether the regulatory environment will change elsewhere.

Other electric utilities may face survival problems resembling those of Pacific Gas and Electric and Hawaiian Electric. A confiscatory resolution of our present problems would obviously be a negative for BHE, but both that company and Berkshire itself are structured to survive negative surprises. We regularly get these in our insurance business, where our basic product is risk assumption, and they will occur elsewhere. Berkshire can sustain financial surprises but we will not knowingly throw good money after bad.

Whatever the case at Berkshire, the final result for the utility industry may be ominous: Certain utilities might no longer attract the savings of American citizens and will be forced to adopt the public-power model. Nebraska made this choice in the 1930s and there are many public-power operations throughout the country. Eventually, voters, taxpayers and users will decide which model they prefer.

When the dust settles, America’s power needs and the consequent capital expenditure will be staggering. I did not anticipate or even consider the adverse developments in regulatory returns and, along with Berkshire’s two partners at BHE, I made a costly mistake in not doing so.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I currently have no vested interest in any company mentioned. Holdings are subject to change at any time.

What We’re Reading (Week Ending 25 February 2024)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

Here are the articles for the week ending 25 February 2024:

1. Wang Chuanfu: A Name Everyone in the West Should Know – Kevin Xu

Wang Chuanfu (王传福), the founder of BYD which just beat Tesla in global electric vehicles sales, is virtually unknown in the west. Even in China, he is only well-known in the business circle and has a low profile otherwise compared to the more flashy tech entrepreneur, Jack Ma, or the more cosmopolitan investor-turned-founder, Kaifu Lee.

Whether you think China’s mass production of EVs and other renewable energy products is a net-positive for dealing with climate change, or an evil “onslaught” on the west, BYD’s global impact is hard to ignore and cannot be wished away. Its batteries have been powering millions of cell phones long before it started making cars. Its EVs can be now seen on the streets of every Chinese city, and quite a few European and Latin American cities. Its battery-powered buses are transporting commuters in Hyderabad, Bogotá, and the Los Angeles International Airport. It is also making electric SkyRails (subway in the air) that may soon appear in São Paulo’s skyline. Oh, and it supplies batteries to Tesla too.

Wang Chuanfu, the pudgy-faced chemist-turned-entrepreneur, is the main, if not the sole, reason why BYD, which meant literally nothing when the company was incorporated in 1995, became BYD, which now means “Build Your Dreams.” The late Charlie Munger called him a “genius”. Yet, there is no comprehensive biography (that I’m aware of) about the man. (Musk, on the other hand, has at least three about him.)…

…It is difficult to describe just how poor Wang’s upbringing was and how much the cards were stacked against him to amount to anything. In fact, his plan was to get into a vocational high school, not university, because it was easier in the early 1980s in China to get a job with vocational training. But the year he applied was the same year that his mother passed away, so he was affected by the loss and didn’t get in. Instead, he ended up in a normal high school that inadvertently paved the path for him to eventually attend a university. Even though he could have dropped out, his older brother insisted on supporting him financially, so he could focus on studying, get into a university, and bring the whole family out of poverty.

As the story goes, because Wang had no guidance or tutelage from his parents or anyone else, he read a lot of books on his own and developed some early muscle as an independent thinker. He had no choice. He ended up going to Central South University of Technology in the neighboring province of Henan as a chemistry major. In his own telling, Wang did not even remember applying to this school. His first choice was the Hefei University of Technology in his home province to study wireless technology, because he liked playing with radios as a kid, but he didn’t get in…

…With a 250,000 RMB loan from a cousin who worked in finance, Wang incorporated BYD in Shenzhen in 1995, where nothing was built and anything was possible. Registering a similar company in Beijing would have been a huge hassle, but in Shenzhen as a pilot SEZ, it sometimes took as little time as one day to form a company. Thus, there were a ton of companies being incorporated. In a rush, Wang chose B(比) Y(亚) D(迪) – three random Chinese characters that meant nothing together – because it was a name that wasn’t used yet. He optimized the first character’s pinyin for being earlier in the English alphabet, so the name could be seen earlier at a trade show or conference. (Jack Ma picked Alibaba for the same reason.)

Back then, the global leaders in battery manufacturing were Japanese giants – Sanyo, Panasonic, Phillips. Sanyo, in particular, was the company Wang aspired to and wanted to beat. But BYD was poor and could not afford any of the advanced equipment or assembly lines that Japanese manufacturers were using. So Wang reverse-engineered the manufacturing process, broke it down into small pieces, then hired very cheap human labor – the only advantage China had at the time – to work on each of those pieces to build cheaper batteries by hand. It was the most literal implementation of “human as a cog in a machine.”

Wang also flexed his chemistry training and caught up quickly in terms of battery technologies, from nickel-cadmium, to nickel-metal hydride, to lithium. BYD quickly caught up on all three types of batteries, while producing them at a fraction of the cost compared to its Japanese competitors. Its early investment in lithium-based batteries, along with Wang’s penchant to reverse engineer, would feature more prominently later in our story when BYD decided to make EVs…

…The company went public in 2002. That same year, Li Lu, the Tiananmen-protest-leader-turned-value-investor bought a stake with the money that Charlie Munger entrusted him to start Himalaya Capital…

…To Munger, investing in BYD in 2002 was akin to writing a VC check into an early stage startup – high probability of going to zero but with infinite upside.

Munger nonetheless admired Wang Chuanfu the person – someone he considered a “genius” with great engineering aptitude who works 70 hours a week. He would also soon learn of Wang’s independence and stubbornness, a trait that made his and Li Lu’s wager look like a terrible idea for a time, but set it on the path to become one of the best performing investments ever.

In January 2003, BYD bought a local carmaker called Qichuan Motors. Qichuan was so bad that the only worthwhile asset from that acquisition was the license it held, which BYD could now use to make its own cars.

Wang has had his eyes on the massive Chinese car market, and this was his way to move into it. His investors, however, were not pleased with this expansion. Li Lu, Munger, and just about everyone inside and outside the company opposed it. BYD’s stock price tanked by one-third during this acquisition.

But Wang didn’t care. For one reason or another, he acquired an immense confidence in his ability to reverse-engineer, vertically-integrate, then mass-produce just about anything. To learn how to make cars, he bought 50 or so second-hand cars from all the best foreign brands, took them apart, and learned how to make cars – a tale he has been fond of sharing in interviews since…

…Tesla was incorporated in July 2003, a few months after BYD bought Qichuan Motors. And Elon Musk would not come into the picture until February 2004, when he made an investment into Tesla’s Series A round using his PayPal-to-eBay acquisition winnings.

Technically, Wang was into making cars before Musk was.

Warren Buffett’s investment in BYD is a well-told story. Buying 225 million shares for $230 million dollars in 2008, when BYD was trading at barely more than $1, it is one of the best examples of Buffett’s “buy and hold” strategy working its magic. Buffett did not begin selling until 2023 – 15 years after his initial purchase. He is still holding more than half of his original stake, at the time of this writing.

However, there are two details to the Buffett-BYD love story that are less well-known and provide interesting colors to Wang Chuanfu’s personality.

First, Wang rejected Buffett’s initial overture to buy BYD, because the Oracle of Omaha wanted to buy a bigger stake than Wang was willing to give up. Despite the obvious benefits of capital infusion and stamp of approval from the greatest investor of all time, Wang stubbornly treated BYD like his baby, his kingdom, and his calling that couldn’t be so easily sold to the highest or most famous bidder. In the end, Buffett was only able to acquire about 10% of BYD…

…From 2009 to 2010, buoyed by Buffett’s investment and branding, Wang set BYD on an aggressive expansion path to make and sell as many cars as possible in China. Although Wang was an engineering and mass production savante, force-feeding BYD cars, which were not of the best quality nor had any brand premium at the time, turned out to be a terrible move. BYD had no problem pumping out tens of thousands of cheap cars. But Wang’s sales target – doubling year over year – forced its sales teams to in turn force dealerships across the country to take on more BYD inventories and higher sales targets of their own.

But not enough consumers wanted BYD cars. Demand overall was also weakening at the time when every country was, in one way or another, dealing with the aftermath of the Global Financial Crisis. Thus, major dealerships started rejecting BYD cars and severing relationships with the company in droves, from Sichuan, to Hunan, to Shandong, and beyond.

By mid-2010, “Dealership Exodus Gate” was in full swing, BYD slashed its sales guidance, implemented mass layoffs, and Wang was humbled. He realized that treating dealers like minions, while making cars with no brand value was not going to work, even with Buffett’s blessing. Unlike batteries, which few consumers know of or care about the brand or manufacturer, cars are prized possessions that convey social status and prestige.

BYD had to become a brand, not just an efficient producer of cheap, affordable cars…

…Tesla first started selling EVs in China in 2014. It commanded brand premium, conveyed social status, and produced high-performing EVs with solid range – three things BYD did not have. Tesla’s were coveted by many, but affordable to only a few, due in large part to China’s high tariffs on foreign-made cars. This barrier gave BYD and other domestic EV makers some room to survive by continuously catering to low-end, cost-conscious consumers.

All that changed in 2019, when Tesla opened its Shanghai factory. Musk’s creations could now both be made and sold in China. This meant Tesla cars could avoid the tariffs and lower prices to compete with the likes of BYD. That year, BYD sold 20% less vehicles than the previous year. Its earnings fell by almost half. Wang Chuanfu was in survival mode again…

…To fix BYD EVs’ lack of range and improve safety concerns, Wang came up with a new design concept that became the Blade Battery – a new form factor that could pack more power density and release heat faster than the standard battery pack modules. BYD’s adaptive and vertically-integrated manufacturing line quickly churned out prototypes of Lithium Iron Phosphate (LFP) Blade Battery…

…By packing more LFP-composed power into Wang’s blade-shaped design, which allowed for more density and a larger surface area for cooling, the LFP Blade Battery achieved a nice middle ground that enabled longer range than conventional LFP block batteries, a bit less range than NMC batteries, with way less heat during an accident…

…By March 2020, Blade Battery started making its way into BYD EVs. From 2020 to 2022, BYD’s sales quadrupled. The same Blade Battery is now in Tesla’s Model Y…

…What Wang will face next in order to take BYD to the next level is a geopolitical problem that has been decades in the making. It will require more words, more finesse, and less inventive chemistry composition and hardcore engineering. It is probably not the kind of wheeling-and-dealing he is naturally good at. Then again, for a peasant kid orphaned as a teenager, he is not supposed to be naturally good at anything.

Whether he succeeds or not, Wang Chuanfu, is a name that everyone in the west should know. It’s long overdue.

2. The road to investing wisdom begins with ‘I don’t know’ – Chin Hui Leong

When it comes to buying stocks, investor and mutual fund manager Peter Lynch has a simple mantra: Invest in what you know. But what does it mean to know something? How do you gauge your knowledge and skills?

Businessman and investor Warren Buffett has a useful concept for this conundrum: your circle of competence. In layman’s terms, it refers to the range of topics and fields that you can understand well.

For instance, if you are a teacher, you will have a better understanding of the education system than most people. Likewise, if you are a restaurant owner, you will know the ins and outs of the food and beverage industry.

Here is what investors miss: Knowing what you are good at is just the beginning. The real challenge is to know your limits. You need to be honest about your weaknesses and avoid investing in areas you do not understand, Buffett says. In other words, you need to know what you do not know…

…It is better to admit early that you are out of your depth than to suffer months and years later from holding the wrong stocks. Even a winning stock will be useless if you lack the conviction to hold it…

…Ben Graham, the father of value investing, used a story to explain how the stock market works: he called it Mr Market. A friendly guy, Mr. Market always tells you the price of your shares every day. But there is a catch: He is also very emotional. He often gets too excited or too depressed, and gives you prices that are too high or too low.

The trick is to know when Mr Market is wrong. That is how you beat him at his own game. Then again, while Mr Market has mood swings, he is not dumb. Even Graham admits that Mr Market can get it right sometimes, giving you a fair price for your stock based on how the underlying business is doing and its prospects.

The trick, then, is to realise that while Mr Market is not stupid, he is impatient. In the short term, he will change the price of your stocks to reflect the prevailing business news.

Over the long term, however, it is the business’ earnings growth that will determine the direction of the stock price…

…Here’s what I have noticed: Most investors do not like to admit that they need to diversify to lower their risk. They prefer to follow Buffett’s advice and put all their eggs in one basket. They would hold no more than five stocks at a time, sometimes even less.

Sadly, these same investors are just trading one flaw for another – ignorance for arrogance. Holding a few stock positions implies you have the rare ability to pick winners with atypical accuracy. Buffett, with his decades of experience, can make that claim. How about you?…

…Investor, hedge fund manager and author Seth Klarman said it best – that when you buy a stock, it is an arrogant act. You are saying you know more than the person selling the stock to you. That is arrogance.

There is no thin line between arrogance and confidence. They are both sides of the same coin. But here is the good news. You do not have to be stuck on one setting. You can be confident when you buy stocks. And then be humble after you buy the stock. You can commit to learning about the business over years, and earn your right to be confident.

3. What a Viral Post on Giraffes Says About China’s Fed-Up Investors – Li Yuan

It’s a perilous time for investors in China. Their main vehicle, so-called A shares of Chinese companies, fell more than 11 percent in 2023 and have continued their losses this year. Many investors have instead flocked to the exchange-traded funds that track foreign markets and that have been performing much better.

Putting money in stocks is inherently risky. But Chinese investors are experiencing something especially alarming: financial losses in the markets, declining home values and a government that doesn’t want any public discussion of what’s happening.

With their frustrations piling up, Chinese investors recently found a way to vent that wouldn’t be quickly censored. They started leaving comments on an innocuous post about giraffe conservation on the official Weibo social media account of the U.S. Embassy in China. They lamented the poor performance of their portfolios and revealed their broader despair, anger and frustration. The giraffe post has been liked nearly one million times since Feb. 2, much more than what the embassy’s Weibo posts usually get. Many of the comments also offered admiration for the United States, as well as unhappiness about their own country.

“The different stock markets’ performances reflect the distances between America and China in terms of national power, technology, humanity and sense of well-being,” a commenter wrote.

The comments demonstrate a growing loss of confidence by the Chinese public in the stock market, the country’s economic prospects and the Chinese Communist Party’s ability to govern…

…Another investor I spoke to, Leo, a portfolio manager at an asset management company in Beijing, has been investing in China’s stock markets for nearly a decade. In November, he started closing out his positions. Now, like Jacky, he is placing his bets on overseas markets.

Leo said he used to hope that China’s internet giants Alibaba and Tencent would become $1 trillion companies like Amazon, and that investors like him would benefit from their growth. “That dream was shattered” after the government cracked down on tech in 2020, he said. “I can only look to the overseas markets now.”

The American Embassy’s Weibo comments section once served as an online punching bag for nationalistic Chinese who blamed the United States for their country’s problems. Now it’s called the Western Wall of China’s A shares investors.

“Under the protection of the U.S. government,” wrote one commenter, “the giraffes are 10,000 times happier than the Chinese stock investors.”…

…A recent survey by the Canton Public Opinion Research Center offered a bleak picture from the southern city of Guangzhou, a metropolis of nearly 19 million people and a hub of technology, manufacturing and trade. In a 2023 survey of 1,000 residents, the center found that the city’s “economy and the society were confronted with unprecedented challenges and pressure.”

The research center’s report said residents’ assessment of the economy, because of unemployment and falling incomes, was as low as it was in 2015, when China’s markets tanked. Satisfaction with the growth of the private sector dropped below 30 percent, the lowest level since the question was first asked in 2008. Most residents said they didn’t expect their incomes to improve in 2024, and more than 20 percent said they believed they were “likely” to lose their jobs.

News coverage about the survey was censored, and the report can’t be found on the center’s website…

…Leo, who was born in Beijing in the mid-1980s, said he had grown up as a nationalistic “little pink.” The first crack in his confidence, he said, was in 2021 when the government went after internet companies. The second crack appeared when the government abruptly ended its “zero-Covid” policy in December 2022 without preparing the population with effective vaccines or medications. Then in late July, the markets and the private sector failed to respond to government measures to stimulate the economy.

Leo’s change is remarkable. He said local Beijing residents like him and the people with whom he had gone to high school were among the stoutest supporters of the Communist Party’s rule because they benefited from the city’s expansion and the country’s growth.

When a group of Leo’s classmates met up in June, he said, they couldn’t believe that two of them, a couple, were migrating to Canada…

…He said the big problems that had made him flee remained unsolved: the imploding real estate sector, enormous local government debts and a fast-aging population.

He said that he wanted the government to loosen its grip on private enterprise and disband Communist Party branches that had proliferated inside companies, and that he wanted the private sector to start to invest again. Until then, he will keep his money out of China’s markets.

And what investing advice would he give to his families and friends? “Run as fast as you can,” he said, “even at a loss.”

4. Rohit Krishnan — Demystifying AI – Jim O’Shaughnessy and Rohit Krishnan

Jim O’Shaughnessy: This large language model says, and he’s speaking to you or it is speaking to you. “In your description of AI as a fuzzy processor, you acknowledge a level of unpredictability in AI behavior. How would you balance the need for predictable AI systems with the inherent uncertainty of their fuzzy outputs in critical applications?” …

…Rohit Krishnan: So with an LLM, the fact that it’s a fuzzy processor means that you can now use it in a lot of different places where you could not have used an AI or any kind of software before, because it can effectively be a replacement for parts of different jobs that people might actually be doing. However, the problem is that, if you or I as fuzzy processors are used in those places, we can be tested. We can be evaluated. If I’m hiring someone for a job, I know that they’re not perfectly predictable. However, I can talk to them and get a sense of how unpredictable they are, and how they would actually deal with different situations, and monitor those in different ways, and ask for previous employers or references, or interview them, and create basically this cone of uncertainty, if you will. I can bound it, so I know that they’re not completely crazy. I know that they will do some things, but it’ll be within the bound of error.

Rohit Krishnan: So with LLMs and fuzzy processors, we are at the early stages for that. The inherent fuzziness is problematic only because you cannot depend on when and how it is actually likely to be fuzzy, that it might end up going in any kind of random direction. So for us, to be able to use it in any actual real-life situation, especially in critical applications, we would need to have a whole lot more confidence in how precisely it works. We would need to not it’s in the internals, in the specific nodes and weights and stuff like that. We already know it, but it’s slightly unhelpful. It’s like doing, I don’t know, neuroscience to predict behaviorism. I don’t think it is hugely valuable in and of itself. However, we do need to bound its behavior in some sense so that we know it cannot go completely off the rails when you’re trying to use it.

Rohit Krishnan: Even with that, I mean, we are speaking, what, after the latest Boeing disaster, not that long after? So when you talk about complex systems where large number of parts actually interact with each other, the possibility of something going wrong always exists. So the way we solve it in real life is by having stringent QA and multiple checks and huge levels of evaluations and large amounts of redundancies. And the exact same principle applies for fuzzy processes as well, where the only way to make a fuzzy processor function in a critical system is by having large number of evaluations so that it can bound it, creating enough structure around it so that even if it does something weird or crazy, you can actually cut off those particular probability branches of the tree, and you can direct it towards something, and having large amount of redundancies so that you can actually ensure that the output that is coming from it is effectively usable, so that even if it does something crazy or stupid, the errors are not continuously compounded over a period of time.

Rohit Krishnan: It’s like that… I don’t know whether this is apocryphal, but I remember hearing this story about Elon where they were trying to send computers up along with the Starlink satellites. And obviously, radiationshielded computers are very heavy and highly expensive. And radiation shielding is important because bit flips are more common when there is higher levels of radiation that actually hits once they’re above the atmosphere. And I think his solution, or the solution is one of his engineers in that particular apocryphal story, was to send three, and they would just vote, because chances of all three getting hit simultaneously are much lower. That’s a way to use redundancy to solve for unpredictability. And I feel like a similar kind of thesis has to exist with respect to LLMs as well…

… Rohit Krishnan: I think I’ve written about a couple of these things before, which is that, at a sufficient degree of complexity, highly deterministic systems can also show highly indeterministic outcomes. I am by no means the first person. It’s a common trope in pretty much anything to do with chaos theory or even things like sand piles and grains at a point of avalanche and cascade. And there’s a bunch of these questions which are, in my opinion, more feasible to see happen than to predict how it will happen because prediction requires you to effectively run the experiment, so to speak, and I’m fascinated by that.

Rohit Krishnan: So I think, in some sense, we in normal conversations quite often complicate indeterministic with random, or unpredictable with random, and they’re two different kind of processes. I mean, there is the common argument that people make against, things like free will, is like, everything is a physical phenomena. Physical phenomena, given a sufficiently powerful computer, might actually be able to get simulated, and therefore, you might be able to predict it. And it’s one of those things when, logically, it might hold true if and only if the computer that is predicting it did not need to actually run the simulation in order to predict it. And if it did, then from the perspective of the people being simulated, us in this instance, the outcome will still end up looking indeterministic, unpredictable, even though, theoretically, everything was as preordained.

Rohit Krishnan: I know this has vexed and driven more people mad than me, but I think there is a core kernel of truth here that just because you can’t create beautiful analog equations to predict the behavior of a particular piece of software, physical phenomena, whatever, does not mean that that is random. It just means that at a certain degree of complexity, there are way more permutations and combinations of how things can go wrong than there is feasible for us to, I don’t know, conceivably identify. And as we said in the previous, the only way to solve it is by having sufficient amount of QA and redundancy and bound the system so that you can actually be relatively sure that it does what you want it to do.

Rohit Krishnan: I mean, stock markets are a perfect example of this. I mean, the flash crash is my favorite example of this. It’s not an intended behavior of the system, but it is one chaotic outcome that could have happened. And how do you stop it? You don’t stop it by stopping each individual trader analyzing each one. You stop it at the macro level saying, “If it falls a little this much, we cut it off,” which is a macro behavior that then controls the micro behavior of each individual algo, which takes that into account. And even if it does hit, we mean that the worst-case scenario is bounded.

Jim O’Shaughnessy: And you also covered that in your book because you posit that we could have a so-called flash crash of AI. And why don’t you tell our listeners a little bit about your solution for that?…

…Rohit Krishnan: The only way to guard against it is at the macro level. You can’t go solution by solution and say, “Unless we can perfectly predict the outcome of this particular system, we will let it go off and do what it wants to do,” because if you could perfectly predict the outcome of the system, we didn’t really need the system in the first place. It’s arguing against the premise of the question in the first place.

Rohit Krishnan: The only way to guard against it is at the macro level. You can’t go solution by solution and say, “Unless we can perfectly predict the outcome of this particular system, we will let it go off and do what it wants to do,” because if you could perfectly predict the outcome of the system, we didn’t really need the system in the first place. It’s arguing against the premise of the question in the first place.

Rohit Krishnan: We will have to do something similar on the AI front as well, where if you don’t want it to do certain outcomes in a particular system, we have to go from outcome first rather than sort of algo first. You’re not going to prevent that by, I don’t know, bounding the number of flops, because even with the lower number of flops, we can find enough ways for it to screw us up, assuming there’s enough number of them that actually interact with each other. But the only way to stop that is step up a layer of aggregation, actually stop it from creating the chaos that we don’t actually want it to do…

…Rohit Krishnan: Oh, I’ll tell you one of the funny things that I’ve been working on. I created a bit of an evaluation suite for a bunch of LLMs for various reasons. And I ran it against a bunch of the Chinese LLMs because I could. I mean, there’s no reason to. So then the interesting things that come out from that is that they’re really good, first of all, I should say that. However, they’re also clearly slanted in what they’re actually allowed to say.

Rohit Krishnan: If you ask it any questions about things around geopolitics, it’s like hackles get raised a little bit, and it says specific things. If you ask it questions about economics, its hackles get raised. If you ask about politics, of course, sometimes it just refuses to answer. Don’t even mention Tiananmen Square. It is fascinating to see that it has created an actually useful tool, which is it does coding really well. And you ask it to create ASCII art of a dinosaur, it does pretty well. You ask it to name, I don’t know, planets in reverse order with different whatever, different languages for each, it does the things that you would want it to do. But it also means you cannot put it into production anywhere you need any of that judgment.

Rohit Krishnan: So you cannot use it in a financial services institution because, guess what, if you’re making an investment decision, you cannot be influenced by things that were hard-coded into you. So similarly, the only way you’re going to be convinced about which ones you are most happy using are by ease of use and latency. It has to be easy to use in front of you, fast, et cetera, et cetera. But also, you can trust the advice coming from it. If I’m thinking about investing in something, I’m not going to call my friend up from Beijing to ask their opinion on a public line. Because there’s a set of information that comes back which is clearly biased. I would ask somebody that I trust and that is the benefit here…

…Rohit Krishnan: I don’t think you are wrong. I think the only caveat or perhaps addition that I would make is centaur models work best in areas which are not directly entirely competitive with the same things that the AIs do. Unless you find joy in doing it, because then it’s a self-fulfilling kind of prophecy.

Rohit Krishnan: To me, currently, and at least for the immediate future, AI is best used in areas where you can either automate part of your own job and yourself and also use it together with you in order to make your ultimate goal better. It’s just like any tech. We are all centaurs already. We live most of our lives on digital technology connected with other human beings. We are part of some weird form of a hive mind, and we are all cyborgs. This is a fact.

Rohit Krishnan: Then the question is, how much more integration would you like in different facets so that you can actually perform some of these things better? And the answer is all of them. Now, there might be some things where, guess what? If you like drawing for fun, you’re probably still going to drawing for fun, despite the fact that if you do want to make a profession out of it, there are some things that the AI will be able to do much better.

Rohit Krishnan: And you as somebody who actually understands it and can use it better and knows the intricacies of drawing will be able to direct it and make use of it in ways that me, as somebody who doesn’t, can’t. Your knowledge and education in doing that particular thing translates to how much better you can actually do something. It’s like giving yourself a boost. Everyone gets a boost kind of question.

5. Big Risks: Catastrophic Risk in Investing and Business – Aswath Damodaran

There are a multitude of factors that can give rise to catastrophic risk, and it is worth highlighting them, and examining the variations that you will observe across different catastrophic risk. Put simply, a volcanic eruption, a global pandemic, a hack of a company’s database and the death of a key CEO are all catastrophic events, but they differ on three dimensions:

Source: I started this post with a mention of a volcano eruption in Iceland put an Icelandic business at risk, and natural disasters can still be a major factor determining the success or failure of businesses. It is true that there are insurance products available to protect against some of these risks, at least in some parts of the world, and that may allow companies in Florida (California) to live through the risks from hurricanes (earthquakes), albeit at a cost. Human beings add to nature’s catastrophes with wars and terrorism wreaking havoc not just on human lives, but also on businesses that are in their crosshairs. As I noted in my post on country risk, it is difficult, and sometimes impossible, to build and preserve a business, when you operate in a part of the world where violence surrounds you. In some cases, a change in regulatory or tax law can put the business model for a company or many company at risk. I confess that the line between whether nature or man is to blame for some catastrophes is a gray one and to illustrate, consider the COVID crisis in 2020. Even if you believe you know the origins of COVID (a lab leak or a natural zoonotic spillover), it is undeniable that the choices made by governments and people exacerbated its consequences.
Locus of Damage: Some catastrophes created limited damage, perhaps isolated to a single business, but others can create damage that extends across a sector geographies or the entire economy. The reason that the volcano eruptions in Iceland are not creating market tremors is because the damage is likely to be isolated to the businesses, like Blue Lagoon, in the path of the lava, and more generally to Iceland, an astonishingly beautiful country, but one with a small economic footprint. An earthquake in California will affect a far bigger swath of companies, partly because the state is home to the fifth largest economy in the world, and the pandemic in 2020 caused an economic shutdown that had consequences across all business, and was catastrophic for the hospitality and travel businesses.
Likelihood: There is a third dimension on which catastrophic risks can vary, and that is in terms of likelihood of occurrence. Most catastrophic risks are low-probability events, but those low probabilities can become high likelihood events, with the passage of time. Going back to the stories that I started this post with, Iceland has always had volcanos, as have other parts of the world, and until recently, the likelihood that those volcanos would become active was low. In a similar vein, pandemics have always been with us, with a history of wreaking havoc, but in the last few decades, with the advance of medical science, we assumed that they would stay contained. In both cases, the probabilities shifted dramatically, and with it, the expected consequences.

Business owners can try to insulate themselves from catastrophic risk, but as we will see in the next sections those protections may not exist, and even if they do, they may not be complete. In fact, as the probabilities of catastrophic risk increase, it will become more and more difficult to protect yourself against the risk…

…When looking at how the market prices in the expectation of a catstrophe occurring and its consequences, both these human emotions play out, as the overpricing of businesses that face catastrophic risk, when it is low probability and distant, and the underpricing of these same businesses when catastrophic risk looms large.

To see this process at work, consider again how the market initially reacted to the COVID crisis in terms of repricing companies that were at the heart of the crisis. Between February 14, 2020 and March 23, 2020, when fear peaked, the sectors most exposed to the pandemic (hospitality, airlines) saw a decimation in their market prices, during that period.

With catastrophic risk that are company-specific, you see the same phenomenon play out. The market capitalization of many young pharmaceutical company have been wiped out by the failure of blockbuster drug, in trials. PG&E, the utility company that provides power to large portions of California saw its stock price halved after wildfires swept through California, and investors worried about the culpability of the company in starting them.

The most fascinating twist on how markets deal with risks that are existential is their pricing of fossil fuel companies over the last two decades, as concerns about climate change have taken center stage, with fossil fuels becoming the arch villain. The expectation that many impact investors had, at least early in this game, was that relentless pressure from regulators and backlash from consumers and investors would reduce the demand for oil, reducing the profitability and expected lives of fossil fuel companies.

While fossil fuel pricing multiples have gone up and down, I have computed the average on both in the 2000-2010 period and again in the 2011-2023 period. If the latter period is the one of enlightenment, at least on climate change, with warnings of climate change accompanied by trillions of dollars invested in combating it, it is striking how little impact it has had on how markets, and investors in the aggregate, view fossil fuel companies. In fact, there is evidence that the business pressure on fossil fuel companies has become less over time, with fossil fuel stocks rebounding in the last three years, and fossil fuel companies increasing investments and acquisitions in the fossil fuel space.

Impact investors would point to this as evidence of the market being in denial, and they may be right, but market participants may point back at impact investing, and argue that the markets may be reflecting an unpleasant reality which is that despite all of the talk of climate change being an existential problem, we are just as dependent on fossil fuels today, as we were a decade or two decades ago:

Don’t get me wrong! It is possible, perhaps even likely, that investors are not pricing in climate change not just in fossil fuel stocks, and that there is pain awaiting them down the road. It is also possible that at least in this case, that the market’s assessment that doomsday is not imminent and that humanity will survive climate change, as it has other existential crises in the past.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Amazon, Tencent, and Tesla. Holdings are subject to change at any time.

Shorting Stocks Is Hard, Really Hard

It’s far easier to recognise poor underlying business fundamentals in a stock and simply avoid investing in it.

In investing parlance, to “short a stock” is to make an investment with the view that a stock’s price will decline. On the surface, shorting seems like a fairly easy thing to do for an investor who has skill in “going long”, which is to invest with the view that a stock’s price will rise – you just have to do the opposite of what’s working.

But if you peer beneath the hood, shorting can be a really difficult way to invest in the stock market. Nearly four years ago in April 2020, I wrote Why It’s So Difficult To Short Stocks, where I used the story of Luckin Coffee to illustrate just how gnarly shorting stocks can be:

In one of our gatherings in June 2019, a well-respected member and deeply accomplished investor in the club gave a presentation on Luckin Coffee (NASDAQ: LK)…

…At the time of my club mate’s presentation, Luckin’s share price was around US$20, roughly the same level from the close of its IPO in May 2019. He sold his Luckin shares in January 2020, around the time when Luckin’s share price peaked at US$50. Today, Luckin’s share price is around US$4. The coffee chain’s share price tanked by 76% from US$26 in one day on 2 April 2020 and continued falling before stock exchange operator NASDAQ ordered a trading halt for Luckin shares…

…The wheels came off the bus only on 2 April 2020. On that day, Luckin announced that the company’s board of directors is conducting an internal investigation. There are fraudulent transactions – occurring from the second quarter of 2019 to the fourth quarter of 2019 – that are believed to amount to RMB 2.2 billion (around US$300 million). For perspective, Luckin’s reported revenue for the 12 months ended 30 September 2019 was US$470 million, according to Ycharts. The exact extent of the fraudulent transactions has yet to be finalised.

Luckin also said that investors can no longer rely on its previous financial statements for the nine months ended 30 September 2019. The company’s chief operating officer, Liu Jian, was named as the primary culprit for the misconduct. He has been suspended from his role…

…it turns out that fraudulent transactions at Luckin could have happened as early as April 2019. From 1 April 2019 to 31 January 2020, Luckin’s share price actually increased by 59%. At one point, it was even up by nearly 150%.

If you had shorted Luckin’s shares back in April 2019, you would have faced a massive loss – more than what you had put in – even if you had been right on Luckin committing fraud. This shows how tough it is to short stocks. Not only must your analysis on the fundamentals of the business be right, but your timing must also be right because you could easily lose more than you have if you’re shorting.

Recent developments at a company named Herbalife (NYSE: HLF) present another similar illustration of the onerous task of shorting. High-profile investor Bill Ackman first disclosed that he was short Herbalife in December 2012. Back then, the company was a “global network marketing company that sells weight management, nutritional supplement, energy, sports & fitness products and personal care products” in 79 countries, according to its 2011 annual report. Today, Herbalife is a “global nutrition company that provides health and wellness products to consumers in 95 markets,” based on a description given in its 2023 annual report. So the company has been in pretty much the same line of business over this span of time.

Ackman’s short-thesis centred on his view that Herbalife was a company running an illegal pyramid scheme, and so the business model was simply not sustainable. When Ackman announced that he was short Herbalife’s shares, the company was reporting consistent and strong growth in its business. From 2006 to 2011, Herbalife’s revenue compounded at an annualised rate of 13% from US$1.9 billion to US$3.5 billion while its profit grew from US$143 million to US$415 million, representing a compounded annual growth rate of 24%.

Although Herbalife has to-date never officially been found to be operating an illegal pyramid scheme, its business results since Ackman came public with his short has been poor. The table below shows Herbalife’s revenue, net income, and net income margins from 2011 to 2023. What’s notable is the clear downward trend in both Herbalife’s net income and net income margin in that time frame.

Source: Tikr

According to a Bloomberg article published at the end of February 2018, Ackman had effectively ended his short position on Herbalife by the time the piece came to print. I think most investors who are made to guess Ackman’s returns from his Herbalife short by looking only at the trajectory of the company’s financials from 2011 to 2017 would have noted the stark deterioration – the company’s net income declined by nearly 40% and its net income margin shrank from 12.0% to 4.8% – and conclude that Ackman had probably made a decent gain.

But the stock market had other ideas. Herbalife’s stock price closed at US$23.16 on the day just prior to Ackman’s first public declaration of his short position. It closed at US$46.05 – a double from US$23.16 – when the aforementioned Bloomberg article was published. From December 2012 to today, the highest close for Herbalife’s stock price was US$61.47, which was reached on 4 February 2019. Right now, Herbalife’s stock price is at US$8.07. This comes after Herbalife’s stock price fell by 32% to US$8.03 on 15 February 2024 after the company reported its 2023 fourth-quarter results. Following the sharp decline, Ackman proclaimed on X (previously known as Twitter) that “it is a very good day for my psychological short on Herbalife.”

The market eventually reflected the deterioration in Herbalife’s fundamentals, but the interim journey was a wild ride. In a similar manner to Luckin’ Coffee (and borrowing the prose from the last paragraph of the excerpts above from Why It’s So Difficult To Short Stocks), if you had shorted Herbalife’s shares back in December 2012 and held onto the position till now, you would have faced a massive loss in the interim – more than what you had put in – even if you were right on Herbalife’s collapsing fundamentals and eventual stock price decline.

The investing sage Philip Fisher once wrote that “it is often easier to tell what will happen to the price of a stock than how much time will elapse before it happens.” This explains why shorting stocks is hard – really hard. To be successful at shorting, you need to correctly read both the stock’s underlying business fundamentals and the timing of the stock’s price movement. In contrast, it’s far easier to recognise poor underlying business fundamentals in a stock and simply avoid investing in it.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I currently have no vested interest in any company mentioned. Holdings are subject to change at any time.

What We’re Reading (Week Ending 18 February 2024)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

Here are the articles for the week ending 18 February 2024:

1. Where Will Virtual Reality Take Us? – Jaron Lanier

In the intervening decades, V.R. has thrived at two extremes in the quest for “killer apps.” It has long been an established industrial technology: if you’ve flown, ridden, or sailed in a factory-built vehicle in the last thirty years, virtual reality may have played a central role. It’s been used to design surgical procedures and train surgeons ever since our first simulated gallbladder, at Stanford Med, some three decades ago; Boeing, Ford, and many other companies started using VR for design in the early days as well. And then there are the visionary, mystical, and philosophical applications. V.R. can be a way of exploring the nature of consciousness, relationships, bodies, and, perception. In other words, it can be art. V.R. is most fun when approached that way.

In between the two extremes lies a mystery: What role might V.R. play in everyday life? The question has lingered for generations, and is still open. Gaming seems likely—but, for most gamers, not so much. There are many reasons why V.R. and gaming don’t quite work, and I suspect that one is that gamers like to be bigger than the game, not engulfed by it. You want to feel big, not small, when you play. (“Star Wars” might have got this right with holographic chess.) Apple’s initial round of Vision Pro apps, like those from its competitors, aren’t entirely compelling, either, and can even have a lonely, dystopian flavor. (Watching a simulated big-screen movie, by yourself?) But my belief is that the quotidian killer apps will come. Maybe you’ll use V.R. to learn quickly about the Airbnb at which you’ve just arrived. Maybe V.R. will help you assemble ikea furniture. Maybe!

Virtual-reality headsets come in various forms. A major divide has to do with how they acknowledge the real world. Some headsets obscure the surrounding environment completely; this is typical in gaming headsets. But there is another option, which I used to call “mixed” reality, and which came to be known as “augmented” reality in the nineteen-nineties. Some mixed or augmented headsets, such as the Microsoft HoloLens or the system created by Magic Leap, allow you to see the real world through the headset glass so that it can be combined with virtual content using challenging optical techniques. Others, like Apple’s Vision Pro and the recent offerings from Meta, capture the real world with cameras, then render it as part of the virtual environment so that it can be combined with fabulated content.

Camera-based mixed reality is vastly easier to accomplish than the optical version, but it is concerning. Early research by a Stanford-led team has found evidence of cognitive challenges. Your hands are never quite in the right relationship with your eyes, for instance. Given what is going on with deepfakes out on the 2-D Internet, we also need to start worrying about deception and abuse, because reality can be so easily altered as it’s virtualized…

… For most of the technology’s history, however, virtual experiences have been hard to build and maintain. This has been one of V.R.’s biggest problems. I saw the first V.R. teaching demonstration of general relativity at least as early as 1992, and have seen dozens more since then; they’re often wonderful, and help users grasp the concept in new ways. But they only run for a year or so because there are too many variables in a V.R. system for creators to keep experiences available. Graphics chips change, and with them the layers of mediating software. That’s true for other programs, too, but with V.R., when the properties of a headset (like field of view) or an input device shift, the whole experience and interaction method must often be rejiggered. It’s too much ongoing effort, so it usually doesn’t happen; developers move on to other projects. The exceptions have been locked-down V.R. experiences that assume a minimal level of interaction, which limits the magic…

…Apple is marketing the Vision Pro as a device you might wear for everyday purposes—to write e-mails or code, to make video calls, to watch football games. But I’ve always thought that V.R. sessions make the most sense either when they accomplish something specific and practical that doesn’t take very long, or when they are as weird as possible.

The practical side of V.R. is a scattering of wonderful niches: in addition to surgical simulation and vehicle design, the technology is used by oil companies to simulate geological structures, by drug companies to envision molecules, and by planners working on city centers. The new frontier, which might apply more to everyday life, is the spontaneous creation of practical apps that you might not even bother to save. My research group, for instance, has presented a prototype system—the “mixed-reality copilot”—that allowed us to recreate, with a single voice request, a program that allows you to use your hands to paint and sculpt with virtual stuff. A decade ago, it took months to make that kind of program. Hopefully, in the near future, one will be able to ask for a V.R. relativity simulation tailored for a student who has color blindness and A.D.H.D., and it will simply appear. More prosaically, you might walk through a facility in augmented reality, asking an A.I. for instant advice about potential safety hazards and fixes. These ideas might even work already: one of the curious features of this accelerated period of A.I. development is that there aren’t enough minutes in the day to try everything.

On the weird edge, it turns out you can change your body plan in V.R. You can become different animals. You can map your body to that of a lobster or an octopus, and experience, to a significant extent, the control of that other body. The brain has had to adapt to many body plans over the course of its evolution, and it’s pre-adapted to work with more. When you change your body, you can also play with the flow of time. By shifting the rhythm of the natural sway of your limbs, and also how the objects around you move and change in response, you alter the reference points that your brain uses to mark the flow of time. You can speed it up or slow it down. In V.R., you can change the rules of the world. You can exist in strange geometries that are too hard to describe in words. You can become an archipelago of parts instead of a continuous animal. You can blend and share bodies with others, to a surprising degree…

…There are fresh, urgent reasons to reaffirm the value of experience. It is impossible to judge technology without a sense of its purpose—and its only plausible purpose is to benefit people, or perhaps animals, or the over-all ecosystem of the planet. In any case, if we pursue technologies that make it hard to delineate the beneficiaries—for instance, by blending brains into robotics not to cure a disease but just because it seems cool—then we make the very idea of technology absurd. The central question of the technological future is how to identify the people who are supposed to benefit from technology, especially if they seem to have melted into it. If people aren’t special, how can we act in a way that benefits people? We can’t. The principles of ethics, design, and even technology itself become nonsense. What can that specialness be? It must be something that is not technologically accessible, since technology expands unpredictably. It’s a little mystical. The definition of people must be one of apartness. We must now put people on pedestals, or they will drown.

When I put on a V.R. headset, I still notice that I am floating there, that I exist independently of the information I experience. But then there’s the moment I take off the headset, which is the very best. In the nineteen-eighties, we used to try to sneak flowers or pretty crystals in front of people before they would take off their headsets; it was a great joy to see their expressions as they experienced awe. In a sense, this was like the awe someone might experience when appreciating a flower while on a psychedelic drug. But it was actually the opposite of that. They were perceiving the authentic ecstasy of the ordinary, anew.

This is the kind of experience you can have only if you use V.R. fleetingly, not constantly. Here we come to one of the greatest differences between what I love about virtual reality and how it is often promoted today. Venture capitalists and company-runners talk about how people will spend most of their time in V.R., the same way they spend lots of time on their phones. The motivation for imagining this future is clear; who wouldn’t want to own the next iPhone-like platform? If people live their lives with headsets on, then whoever runs the V.R. platforms will control a gigantic, hyper-profitable empire.

But I don’t think customers want that future. People can sense the looming absurdity of it, and see how it will lead them to lose their groundedness and meaning…

…But the truth is that living in V.R. makes no sense. Life within a construction is life without a frontier. It is closed, calculated, and pointless. Reality, real reality, the mysterious physical stuff, is open, unknown, and beyond us; we must not lose it.

Just because owning a major tech platform is desirable, that doesn’t suggest there is no other way to succeed in the technology business. There are water companies and soda companies, and then there is fine wine. All are viable businesses. The metaphor isn’t perfect, but I suspect that V.R. entrepreneurs will find their sweet spot by emulating Napa Valley…

…A.I. is often portrayed as a godlike, transcendent project that will take over the fabric of our physical reality, leading to a singularity, meaning nothing that matters now is likely to matter after. But singularities, like the ones we hypothesize in black holes, are the very definition of ignorance. There is no learning that bridges the before and after of a singularity. It is the absolute rejection of intelligence. Virtual reality is sometimes stirred into this mix. But our best understanding of how reality works is entirely bound to finitude. Physics is all about conservation principles. There are no infinities, only S curves. There is no free lunch. Technical culture often longs for freedom from finitude. A profound truth, however, is that the greatest mysteries are found in conserved systems, which can become rich and complex, not in infinite ones, which stretch out like blank white sheets to the edge of the cosmos.

And so another urgent question is whether people can enjoy the storied reality of finitude after coming down from the high of fake infinity. Can being merely human suffice? Can the everyday miracle of the real world be appreciated enough? Or will the future of culture only be viral? Will all markets become Ponzi-like fantasies? Will people reject physics forever, the moment we have technology that’s good enough to allow us to pretend it’s gone?

2. Pods, Passive Flows, and Punters – Drew Dickson

You’ve surely noticed what has happen to Nvidia lately. We used to just call these winners FANGs, and then FAANGs and then FAMANGs, but Nvidia has insisted on joining the league table. It now has a $1.7 trillion market cap. And in the last five years, the stock is up about 1,700%. Guess what else is up about 1,700%?

Nvidia’s earnings estimates.

How about Facebook, aka Meta, which goes through periods of hatred and love with equal vigor? Well, over the past seven years it has bounced around a lot but still has generated nearly 260% returns. And forward earnings projections? They’re up 280%.

We can stretch things further back, and look at Google over the past 14 years (earnings up 885%, stock up 980%); or Amazon during the same period (earnings up nearly 2,500%, stock up about 2,800%).

Or we can go waaay back and analyze Microsoft over the past 22 years. Forward earnings projections have increased from $0.93 in February of 2002 to $11.57 today. That’s nearly 1,150%. The stock is up just over 1,200%.

And finally, from one of my favorite former-CEOs Reed Hastings, we have good old Netflix. About 18 years ago, analysts were forecasting that Netflix would generate 11 cents of earnings in the coming 2006 year. Here in 2024, they are forecasting a whopping $17 of earnings in the coming year. That is a whopping EPS increase of 14,889%.

And how about the stock? We’ll it is up a whopping 14,882%.

Fundamentals matter, sports fans. Fundamentals matter.

Admittedly, some of these examples above are very long-term, but even when we self-select with some of the biggest, most exciting, long-term winners out there, and ignore the losers (of which there are many), it is still clearly apparent that it is the fundamentals that matter most.

So basically, it probably isn’t terrible advice to ignore the rest of it. Ignore the noise. Ignore the talking heads on CNBC. Ignore prognostications of meme-stock sith lords. Ignore the volatility. Embrace it, actually. And just focus on the fundamentals. Get those right, and you will likely win.

3. “The Practice Of Value Investing”, by Li Lu – Graham Rhodes

If you invest in a company in a sustainably growing economy, your company’s profits and your investment return will also grow sustainably. If you speculate on other people’s short-term trading behaviour, there can only be one result in the end: gains and losses must equal because this is a zero-sum game. If you add up the gains and losses of all speculators in the market, they will sum to zero. This is the biggest difference between investing and speculating. I’m not denying that there are some speculators whose chances of winning are higher and who can go on winning for longer; equally there are some who will always be the sucker at the table and never strike it rich. If you give it enough time though, when you add the winners and losers together, the net result will be zero. The reason is that speculating on short-term behaviour in the market adds nothing to the economy nor to corporate earnings growth. Some people say they use a mixed model of “80% investment, 20% speculation”. If they do 70-80% of their work correctly, then such participants’ returns will reflect the compound growth of the modern economy. However, the remaining portion will be caught up with all the other speculators and their result will be the same – zero.

Now that you know this result, will you choose to be an investor or a speculator? This is a personal choice and there is no right or wrong answer. The only difference is the impact you will have on society. Investors will help all parts of society enter modernity’s virtuous cycle – the stage in which it enjoys continuous compound growth. If you are interested and would like to learn more about this, you can refer to my monograph, “Discussions on Modernisation”.

Relatively speaking, the speculative part of the market verges on being a casino. From a social welfare point of view, we do not want this casino to be too big. However, without it, the market would not exist. We should therefore see speculation as a necessary evil – and a part of human nature – which cannot be removed. We cannot deny the parts of human nature which love to gamble and speculate but we cannot let them overwhelm us. Otherwise, society will sooner or later face the consequences. The wounds of the 2008-2009 Global Financial Crisis from which we have just emerged are still fresh in our memories. And once you understand the principle of a zero-sum game, you will begin to see these speculators as Mr. Market…

…There was another company at the time which taught me something revealing. This company owned a lot of gas stations, and so I became interested in gas stations. There were two gas stations near where I lived, one on each side of the same intersection. However, I realised that one gas station had many more customers, and that cars would come to it regardless of which direction they were heading. Both gas stations had the same price and their gas was the same as it was made to the same standard. I felt this was very strange and since it was my company’s gas station anyway, I went to have a look. The gas station which attracted all the customers was run by a family of Indian immigrants, who all lived there too. As soon as a customer arrived, they would come out to offer him a glass of water. Whether you wanted it or not, they would always offer it to you first and then strike up a conversation. If the kids were home from school, they would come out and help you tidy up your car. The other gas station was run by a typical American. He wasn’t a bad guy but the gas station didn’t belong to him. He was just an employee hired by the real owner, so he wouldn’t come out from the store and nor would he pay much attention to what was happening outside. Thanks to this one difference, I calculated that in a given period, one gas station attracted almost four times as much traffic as the other.

From then on, I realised it was important to know whether a company’s manager had an owner’s mindset. Through this, I began to gradually understand how a company could earn money and why it could earn more than others. The example of the two gas stations is a perfect illustration because they sold the same product and were otherwise identical. However, one’s service was slightly superior to the other’s and so it received four times as much traffic. What motivated that Indian fellow? He was an immigrant, like me. He needed money and if he couldn’t bring in business, he would have financial difficulties. The other manager could be indifferent because he could just take his salary while pretending to do his job. This was the difference. I therefore began to take great interest in how a company is run, its competitive advantages, and the sustainability of these competitive advantages…

…The next attribute is relatively special. You must be both extremely patient and extremely decisive, even though they are in contradiction. When there are no opportunities, you might go for years without taking any action. But as soon as an opportunity arrives, you must be able to become extremely decisive and act without hesitation. I have been Charlie Munger’s investment partner for sixteen or seventeen years now. We meet for dinner at least once a week and I’ve developed a deep understanding of him. Let me tell you a story about his investments. Charlie subscribes to Barron’s, a weekly magazine about the stock market published by the Wall Street Journal. He’s read this magazine for approaching 40-50 years for the purpose of finding investment ideas. And how many has he found in this time? One! There has only been one and he only found it after reading the magazine for more than thirty years. And he hasn’t found another in the ten years since. This hasn’t stopped him from continuing to read the magazine every week though. He is extremely patient and can go for a long time without doing anything at all. But when he finds an opportunity, he will go all in. And this particular investment made him a lot of money. So this is what’s required of an exceptional investor: he must have extreme patience and stay focused even when there are no opportunities. When an opportunity does come, he must then have the ability to move swiftly and decisively…

…When I was young, I always wondered about the meaning of life. Later, I gradually came to realise that the meaning of life is the pursuit of true knowledge. True knowledge can change your life and your fate; it can even change the world. Moreover, mankind is completely different from what else we can observe in the material world. The world we can see is one in which entropy increases. Energy flows from high places to low places; big things devour small things. If a large celestial body hits a smaller one, it will crush it. The entire planet and our universe are to a certain extent heading towards annihilation.

But the world of man is not the same. Mankind can turn the world into one in which entropy decreases. We can reverse entropy’s course. Through study, man can go from ignorance to erudition; through self-cultivation, man can become a virtuous person who contributes to society. Man can create things which were previously unimaginable. Since man’s arrival, the earth has changed. Today, we can even leave this planet for the stars; it is entirely possible that we go on to change the universe. As I mentioned earlier, the first investment I made was related to the wireless telephone. At the time, I hadn’t really figured out what that was. Twenty-six years later, who can bear to part with their mobile phone? Mobile phones, the internet and all these things were game changers born of knowledge. The internet is based on TCP/IP which is a protocol. At their heart, computers are permutations and combinations of 0s and 1s combined with a diode which uses silicon and electricity to tell those 0s and 1s apart. This is how knowledge can create changes which turn our world upside down.

4. Hong Kong’s death has been exaggerated – Michael Fritzell

The National Security Law in June 2020 was indeed a watershed moment for Hong Kong’s judiciary. Now that individuals seen to be endangering national security can be extradited to mainland China, there’s a fear that they will no longer receive fair trials.

But let’s look at the positive side of things. In reality, the National Security Law has really just had two major effects. One is emigration, and the other is stopping public demonstrations.

Since 2020, roughly 400,000 people have left Hong Kong, according to this data from the Hong Kong Immigration Department. But, if you calculate the cumulative number, net migration has actually started to decrease:

In other words, people are now moving back to Hong Kong. These could be individuals who avoided Hong Kong during COVID-19 and are now willing to return. They could also be people who changed their minds about living overseas, knowing that Hong Kong is a great place to make money. In the early 1990s emigration wave, many of those who left for Vancouver or elsewhere ultimately came back to Hong Kong.

While it’s certainly negative that hundreds of thousands of people have left Hong Kong, it’s not implausible that mainland Chinese immigration could make up for the shortfall. In fact, Hong Kong’s residential rents rose 8.1% in 2023 due to immigration from the mainland.

For now, the Hong Kong legal system remains reliable. The conviction rate for Magistrate’s courts in Hong Kong was 54% last year, far higher than mainland China’s 99.95%. This seems to suggest that Hong Kong judges are still independent. Hong Kong still ranks #23 in WJP’s Rule of Law Index, ahead of the United States.

Between Hong Kong and Singapore, the former remains a far larger financial hub. The aggregate market cap of Hong Kong-listed companies is 10x that of Singapore. Its assets under management are US$2.2 trillion – far higher than Singapore’s US$1.5 trillion. There are 2,000 licensed asset managers in Hong Kong vs just 1,200 in Singapore.

A key competitive advantage for Hong Kong is that its currency is freely convertible and pegged to the US Dollar. This enables the Chinese government and its companies to raise overseas capital while maintaining capital controls within mainland China.

It’s also the case that Hong Kong’s taxes are uniquely low:

The highest marginal income tax is 17%.
There is no capital gains tax.
There is no withholding tax on dividends or interest income.
There is no GST.
There is no estate duty.
There is no wealth tax.
There is a 15% tax rate on rental income but with a standard deduction of 20%.
Most import duties to Hong Kong are zero, making imported goods cheap.
The stamp duty for purchasing residential property is 15% for foreigners and 7.5% for locals, but this stamp duty could soon be removed.

For these reasons, the PwC and the World Bank recently ranked Hong Kong as the region with the most friendly tax system in the world.

The Hong Kong government remains committed to its low-tax policy. Hong Kong has agreed to implement a minimum corporate tax rate of 15% from 2025, but so has many other major economies. The budget deficit is projected to continue at over HK$100 billion in FY2025, but 3% of GDP remains modest.

While I don’t want to minimize the political shift that has taken place, for Hong Kong companies, it will be mostly business as usual. Hong Kong will continue to attract the ultra-wealthy through its low taxes, and it will continue to be used to raise capital for companies in China and beyond.

After Hong Kong’s zero-COVID policy was lifted at the end of 2022, the economy has actually been on a solid footing. Hong Kong’s retail sales grew +16% year-on-year in 2023, though remaining almost 20% below the peak in early 2019:

A major component in Hong Kong retail sales comes from tourism to Hong Kong, which is now back to around 70% of the pre-COVID level:

But don’t expect a full recovery in tourism spending. Before 2019, a large portion of Hong Kong retail sales to tourists comprised goods smuggled into mainland China. In 2021, China’s border controls tightened up significantly, and most of such business now occurs through legitimate channels. I wrote about such smuggling here.

One business that is booming is Hong Kong life insurance products sold to mainland Chinese visitors. Related premiums already exceed the pre-COVID-19 level, suggesting strong demand for USD-linked policies.

Hong Kong’s real GDP grew +4.3% in the fourth quarter of 2023. Hong Kong’s export growth has now turned positive at +11% year-on-year. The unemployment rate remains just 2.9%, suggesting that jobs are plentiful…

…What’s weighing on the Hong Kong economy is the interest rate environment. Since the Hong Kong currency is pegged to the US Dollar through a currency board arrangement, it effectively imports its monetary policy and interest rates from the United States…

…Now that HIBOR has reached over 4% borrow rates for households and companies remain above the nominal income growth in the economy. In my view, that means that monetary policy remains restrictive…

…Another longer-term worry is geopolitics. If a war were to break out in Taiwan or elsewhere, US sanctions could be imposed on Hong Kong. It could lose its special trade status. Import tariffs would be imposed, and it would be subject to the same export controls as China. If the Hong Kong Dollar were to be de-pegged to another currency. But as long as the currency remains freely convertible, Hong Kong will continue to retain its competitive advantage as a hub for raising overseas capital.

5. A beginner’s guide to accounting fraud (and how to get away with it): Part VI – Leo Perry

On 9th September 2018 serial entrepreneur Luke Johnson shared his experience and wisdom in an article in The Times newspaper titled ‘A business beginner’s guide to tried and tested swindles’. Five days later HMRC petitioned the High Court to wind up his business, the cafe chain Patisserie Valerie, for an unpaid tax bill. He didn’t notice. Unfortunately I didn’t either.

On 10th October Pat Val halted trading in its shares and suspended its CFO. It noted “significant, potentially fraudulent, accounting irregularities” that had materially impacted the cash position. I was familiar enough with the brand. I worked in an office a few doors down from one. It never seemed busy but there was nothing in the accounts that gave me good reason to think about shorting the company. But if I’d been able to now I would have, even at half the price it was halted at. The reason I was so confident it was screwed was precisely because I hadn’t spotted anything wrong in its numbers before (neither, apparently, had anyone else as there were no publicly disclosed shorts on the FCA list).

Pat Val’s published accounts were as straightforward as you’d expect from a simple business like a cafe. Sales taken in cash, not much held as stock and a few prepaid expenses. The only line items of any size on the balance sheet were the capitalised cost of fitting out stores, and money sitting in a bank. Not a lot to tweak if you needed to meet numbers. That’s why the company saying that its cash position was significantly misstated, while it was short on detail, had to mean that (probably) sales and (almost certainly) profit were faked. Working backwards there couldn’t really be any other story.

Things unravelled fast. The next statement from the company, later the same day, disclosed the winding up petition from a month earlier. The following day Pat Val said it couldn’t continue trading without a capital injection, which really amounted to saying the £30m of “cash” on its balance sheet wasn’t in the bank at all. And the day after that its CFO Chris Marsh was arrested. One trick (I should say allegedly, I guess) is depositing fat cheques just before year end – to show a big credit at the point in time when you know the auditor is going to look – only for them to bounce a few days later. Another is borrowing money – again giving a big credit to cash – and just not mentioning the debt part in the accounts. Most of the time that would still show up as higher interest payments (see e.g. Globo), but when rates are close to zero you can get away with a lot more.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google), Amazon, Apple, Meta Platforms, Microsoft, and Netflix. Holdings are subject to change at any time.

How Innovation Happens

Innovation can appear from the most unexpected places, take unpredictable paths, or occur when supporting technologies improve over time.

There are a myriad of important political, social, economic, and healthcare issues that are plaguing our globe today. But Jeremy and I are still long-term optimistic on the stock market.

This is because we still see so much potential in humanity. There are nearly 8.1 billion individuals in the world right now, and the vast majority of people will wake up every morning wanting to improve the world and their own lot in life. This – the desire for progress – is ultimately what fuels the global economy and financial markets. Miscreants and Mother Nature will occasionally wreak havoc but we have faith that humanity can clean it up. To us, investing in stocks is ultimately the same as having faith in the long-term ingenuity of humanity. We will remain long-term optimistic on stocks so long as we continue to have this faith.

There may be times in the future when it seems that mankind’s collective ability to innovate is faltering (things are booming now with the AI rush). But here are three stories I learnt recently that would help me – and I hope you, too – keep the faith.

The first story is from Morgan Housel’s latest book Same As Ever. In it, he wrote:

“Author Safi Bahcall notes that Polaroid film was discovered when sick dogs that were fed quinine to treat parasites showed an unusual type of crystal in their urine. Those crystals turned out to be the best polarizers ever discovered. Who predicts that? Who sees that coming? Nobody. Absolutely nobody.”

What the quinine and polarizers story shows is that the root of innovative ideas can show up completely unexpectedly. This brings me to the second story, which is also from Same As Ever. This time, it is Housel’s recounting of how the invention of planes moved in an unpredictable path that led to the invention of nuclear power plants (nuclear power is a zero-emission, clean energy source, so it could play a really important role in society’s sustainable energy efforts), and how a 1960s invention linking computers to manage Cold War secrets unpredictably led to the photo-sharing social app Instagram:

“When the airplane came into practical use in the early 1900s, one of the first tasks was trying to foresee what benefits would come from it. A few obvious ones were mail delivery and sky racing.

No one predicted nuclear power plants. But they wouldn’t have been possible without the plane. Without the plane we wouldn’t have had the aerial bomb. Without the aerial bomb we wouldn’t have had the nuclear bomb. And without the nuclear bomb we wouldn’t have discovered the peaceful use of nuclear power. Same thing today. Google Maps, TurboTax, and Instagram wouldn’t be possible without ARPANET, a 1960s Department of Defense project linking computers to manage Cold War secrets, which became the foundation for the internet. That’s how you go from the threat of nuclear war to filing your taxes from your couch—a link that was unthinkable fifty years ago, but there it is.”

This idea of one innovation leading to another, brings me to my third story. There was a breakthrough in the healthcare industry in November 2023 when the UK’s health regulator approved a drug named Casgevy – developed by CRISPR Therapeutics and Vertex Pharmaceuticals – for the treatment of blood disorders known as sickle cell disease and beta thalassaemia. Casgevy’s greenlight is groundbreaking because it is the first drug in the world to be approved that is based on the CRISPR (clustered regularly interspaced short palindromic repeats) gene editing technique. A few weeks after the UK’s decision, Casgevy became the first gene-editing treatment available in the USA for sickle cell disease (the use of Casgevy for beta thalassaemia in the USA is currently still being studied). Casgevy is a huge upgrade for sickle cell patients over the current way the condition is managed. Here’s Sarah Zhang, writing at The Atlantic in November 2023:

“When Victoria Gray was still a baby, she started howling so inconsolably during a bath that she was rushed to the emergency room. The diagnosis was sickle-cell disease, a genetic condition that causes bouts of excruciating pain—“worse than a broken leg, worse than childbirth,” one doctor told me. Like lightning crackling in her body is how Gray, now 38, has described the pain. For most of her life, she lived in fear that it could strike at any moment, forcing her to drop everything to rush, once again, to the hospital.

After a particularly long and debilitating hospitalization in college, Gray was so weak that she had to relearn how to stand, how to use a spoon. She dropped out of school. She gave up on her dream of becoming a nurse.

Four years ago, she joined a groundbreaking clinical trial that would change her life. She became the first sickle-cell patient to be treated with the gene-editing technology CRISPR—and one of the first humans to be treated with CRISPR, period. CRISPR at that point had been hugely hyped, but had largely been used only to tinker with cells in a lab. When Gray got her experimental infusion, scientists did not know whether it would cure her disease or go terribly awry inside her. The therapy worked—better than anyone dared to hope. With her gene-edited cells, Gray now lives virtually symptom-free. Twenty-nine of 30 eligible patients in the trial went from multiple pain crises every year to zero in 12 months following treatment.

The results are so astounding that this therapy, from Vertex Pharmaceuticals and CRISPR Therapeutics, became the first CRISPR medicine ever approved, with U.K. regulators giving the green light earlier this month; the FDA appears prepared to follow suit in the next two weeks.”

The manufacturing technologies behind Casgevy include electroporation, where an electric field is used to increase the permeability of a cell’s membrane. This enables molecules, such as genetic material and proteins, to be introduced in a cell for the purposes of gene editing. According to an expert-call on electroporation that I reviewed, the technology has been around for over four decades, but only started gaining steam in recent years with the decline in genetic sequencing costs; without affordable genetic sequencing, it was expensive to know if a gene editing process done via electroporation was successful. The relentless work of Illumina has played a huge role in lowering genetic sequencing costs over time.

These show how one innovation (cheaper genetic sequencing) supported another in a related field (the viability of electroporation) that then enabled yet another in a related field (the creation of gene editing therapies).

The three stories I just shared highlight the different ways that innovation can happen. It can appear from the most unexpected places (quinine and polarizers); it can take unpredictable paths (from planes to nuclear power plants); and it can occur when supporting technologies improve over time (the development of Casgevy). What they signify is that we shouldn’t lose hope in mankind’s creative prowess when it appears that nothing new of significance has been built for a while. Sometimes, what’s needed is just time.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I currently have no vested interest in any company mentioned. Holdings are subject to change at any time.