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 23 August 2026:
1. AI Is Dead. Organoids Are Alive – Claire L. Evans
I mean that if lab-coated biologists took a sample of your skin and very carefully manipulated the cells inside it, they could actually make a brain. They do it all the time.
Not a brain as complex as the one behind your eyes, of course, but a glob of gray matter nonetheless, with a few-million-odd neurons that can send and receive electrical signals. Biologists call these strange creations human brain organoids. Kept at a womblike 98.6 degrees Fahrenheit for eight months, they’ll produce repetitive oscillations—brain waves—nearly indistinguishable from those made by a premature baby…
…To make a brain organoid, all you need is that sample of skin I mentioned before. (Samples of blood, hair, or teeth work too.) You take the adult cells and introduce them to some special proteins that revert them to their embryonic state. Given a second chance to mature, these so-called induced pluripotent stem cells can become anything: tear gland organoids that cry, heart organoids that beat, or brain organoids that … well, that’s the question.
In utero brain development is, as one bioethicist told me, “a black box” of scientific knowledge. Historically, a lot of what we know about it is inferred from studies with mice. But with an organoid, the transformation of stem cells into neurons into brain tissue happens in full view. In theory, scientists could one day study how a colony of dividing cells comes together to create a mind—to make, from 86 billion neurons, a person…
…Visually, organoids are not compelling; they’re opaque, snot-colored, the approximate size and shape of a chia seed. Muotri’s organoids contain 5 million cells, of which 2.5 million are neurons. (The rest are non-neural glial cells, which serve as scaffolding.) This, he reassures me, is the size of a bee’s brain…
…In Melbourne, Cortical Labs cultivates flat neural cultures—the stem cells were donated by the company’s own founder—and loads them into sleek white biological “computers” called CL-1s. Each is the size of an elongated toaster and boasts an onboard life-support system capable of keeping a culture of up to a million neurons alive for six months. With the CL-1, Cortical Labs is aiming to become the Nvidia of neural computing, providing hardware and, let’s say, “neurons as a service” with a sub-millisecond delay.
For now, these neural computers are mostly of interest to researchers who want to work with neurons without taking on the tedious wet-lab husbandry themselves. Eventually, however, the company hopes that neurons will prove themselves to be an energy-efficient, resilient substrate for more general computing applications—including some tasks currently handled by AI, like image recognition and classification…
…In 2022, using a system similar to what currently powers the Cortical Cloud, Kagan grew a neural culture on a microchip and trained it to play the 1972 Atari game Pong, rewarding the neurons with predictable electrical pulses when they made correct decisions and punishing them with chaotic bursts when they made mistakes.
The technique served as a minimal proof of concept for a theory, proposed by the neuroscientist Karl Friston, that self-organizing biological systems tend to minimize surprise whenever possible. By showing that neurons will reorganize themselves to avoid chaotic stimulus, Cortical Labs demonstrated one possible approach for programming living matter. But the experiment also signaled that the CL-1 could be considered hardware for testing theories of cognition…
…WHEN I SET out to report this story, I was sure it was about consciousness: the eerie moment a quarter-peanut of flesh sparks with self-knowledge, and what that precipice means for the researchers responsible. I imagined long dark nights of the cell and tiny funerals for spent neurons. What I found, however, was that nearly all scientists who keep organoids see the consciousness question as a distraction.
Most bristle when asked. They gesture to the organoids themselves—tiny balls bobbing in liquid solution like droplets of olive oil in vinegar—as if to say, give me a break. “Consciousness is so qualitative,” complained Annie Kathuria, an organoid researcher, when I visited her lab at Johns Hopkins. “How am I supposed to measure something qualitative on a tissue that’s floating in a dish? Someone has to define it. That’s what I say to everyone: Define to me what consciousness is, in quantitative terms.”…
…Conveniently, organoids can’t grow much bigger than 5 millimeters anyway. Once they reach a certain size, without a vascular system to pump in oxygen and pump out metabolic wastes, they develop a “necrotic core” and suffocate. This has created an upper ceiling for the organoid debate, but it won’t hold much longer. At Johns Hopkins, the nanotechnologist David Gracias is developing biomimetic artificial arteries; across town, at the Medical School, Kathuria creates rough blood vessel networks from endothelial organoids. Researchers are keeping organoids alive longer and longer. Muotri’s record, of three years, might’ve gone longer had someone in his lab not dropped the dish.
“THE GOAL IS to get to a 1-centimeter brain organoid,” said Thomas Hartung, as he walked me around the Center for Alternatives to Animal Testing, or CAAT, at Johns Hopkins, where he’s experimenting with new perfusion systems to get around the bloodlessness problem. One centimeter is roughly the size of a mouse brain—“a completely different beast,” Hartung said, than the 500-micrometer organoids he’s accustomed to…
…Nearly every organoid researcher I met reporting this piece told me the same fact: that in human clinical trials, the failure rate for neuropsychiatric drugs is close to 95 percent. The pipeline for new medications for conditions like depression, Alzheimer’s, and epilepsy is long, and often dry. That’s because drugs are ordinarily tested on animals, not people, and animal testing has never been the most physiologically relevant way to ensure the efficacy of drugs—it’s just been the most practical one. Organoids have changed that calculus…
…In July 2025, the National Institutes of Health announced that it will no longer award grant funding to research that relies exclusively on animal testing, encouraging consideration of “new approach models,” like organoids, instead; in September, the NIH announced an $87 million investment in building a Standardized Organoid Modeling Center. Fortunately for the NIH, since organoids are made with induced pluripotent stem cells from consenting adult donors, they don’t inspire the same religious ire as embryonic stem cell research did in the 1990s. And unlike lab rats or primates, organoid testing doesn’t irk the animal rights crowd.
Recent studies suggest that ordinary people, by and large, aren’t much bothered by organoids. When they do lodge an objection, it’s not because globs of neural cells might someday achieve human-level sentience. It’s because brain organoids are creepy. They muddy the line between person and thing, which, like the difference between humans and animals, or between the dead and the living, is a fundamental distinction that transcends cultures…
…Somewhere between the few million neurons in a brain organoid and the billions more that make up our minds, selfhood emerges—and absolutely nobody knows when…
…For the AI people, machine intelligence seems achievable, even inevitable. They “feel the AGI.” Part of this is hype; part of it is a willingness to take intelligence for what it does, as opposed to what it is. “If it quacks like a duck, it is a duck,” said John Evans, the UCSD sociologist. “The Silicon Valley types are super pragmatic. If AI is capable of doing what a human with consciousness does, they’ll call it consciousness.”
Biologists who work with neural cultures, however, are far less likely to use such a word without plenty of agonizing caveats. These biologists cultivate the raw meat of mind every day; they’re in the privileged position to understand how colossally complex it really is. As biology and AI converge, the culture shock between these worldviews is likely to be bracing.
For now, it seems that in the race toward conscious machines, there will be two lanes. One will be paved with rare earth minerals and silicon—forced, at enormous financial and energy cost, to model the brain from the top-down, in the form of artificial neural networks…
…Which leaves the other path. Admittedly, this one will be slippery and meandering—adapting, as all living things do, to the journey as it goes. It might not be the most direct route, but in the evolutionary history of life on Earth, it’s the only thing that has ever led to intelligence.
2. GEN-1.5 – Generalist Team
The ability to learn closed-loop physical skills from just one or a few demonstrations, and to do so across a broad range of tasks without such restrictions, has predominantly been considered out of reach. Such an ability may also likely be underpinned by a foundation that enables other broad generalization capabilities…
…We’ve created GEN-1.5, our latest robot foundation model that exhibits broad one-shot and few-shot learning from demonstration capabilities, as well as zero-shot generalization, e.g. improvisation and novel tool use (e.g. brush, dustpan, etc.). GEN-1.5 is a large multimodal model that processes video input (30 seconds of memory, alongside other sensor, language, and proprioceptive inputs) and produces 100 Hz action trajectories. Its capabilities include:
- One-shot learning via in-context prompting. The model learns new tasks in seconds when prompted with 3 to 12 seconds of a single demonstration, no training required. We refer to the use of sensorimotor examples in the context window as “physical prompting.”…
- …Few-shot adaptation via gradient descent. The model can be fine-tuned to a new task in 1–10 gradient steps on 1–5 minutes of data (~10–50 demonstrations).
- Improvising new strategies and tool use. The model generalizes at the level of behavioral strategies: forming entirely new trajectories to reach a goal, using unseen tools (e.g. brush, dustpan, etc.) to create new solutions to tasks demonstrated with other tools, and working ambidextrously even when prompted or fine-tuned to perform the task with a specific hand.
These capabilities appear to emerge directly from pretraining on large amounts of physical interaction data. We did not explicitly train for any of them: no architectural changes to promote in-context learning, no inner or outer meta-learning loop pressuring the model to adapt from minimal data, no auxiliary objectives encouraging improvisation…
…Although the tasks are simple and short-horizon, and the success rates are modest, this is the first model we know of that has demonstrated the general ability to learn a wide range of dexterous closed-loop physical tasks from just one-shot or few-shot demonstrations…
…GEN-1.5’s initial pretraining began in parallel — it has now been training continuously for over eight months. We left it running because every metric we tracked kept improving with the engine: absorbing more data, scaling more efficiently with compute, and achieving step-change gains with successive surgical architectural and algorithmic changes. It was clear that the model was getting better, and the trend was consistent: new tasks were becoming more data-efficient, more compute-efficient, and more general.
As the model continued to train, we began experimenting with how few finetuning steps we could use to adapt to new tasks, finding the model could learn new tasks from 100s, then 10s, then eventually, 1 gradient step on just one minute of data. As far as we know, the ability to learn skills with such few gradient steps had not been observed before. We then asked, can this model learn new tasks without training, purely in-context and with zero gradient steps? That this works at all changes how we think about how these models can be used, about their potential impact, and the road ahead for building general physical intelligence…
…GEN-1.5 can be prompted with a single demonstration inserted into its 30-second context window, and the remainder holds rolling observations. Physical prompts are sensorimotor examples (i.e. sensor data plus action trajectories), recorded either as human data (with a pair of handheld grippers) or as rollouts from the robot itself. Once the prompt is in context, the model performs the task immediately, with no training steps. The performance of one-shot learning in-context is modest (59% average success across diverse tasks including handling zippers, opening jars, grabbing money out of wallets, etc.), but the fact that inserting a single demonstration in the context buffer, without ever training for it, yields any measurable competence at all was unexpected…
…We did not explicitly train GEN-1.5 for in-context learning, and the tasks we tested were not engineered into the pretraining data beforehand. This is a general model which we are prompting without regard to the pretraining data distribution.
Why this capability emerges from pretraining is difficult to pinpoint. One hypothesis, by analogy to language, is that the distribution of physical observations and actions may exhibit “burstiness” and Zipfian structure of the kind that has been linked to in-context learning in language models.17 It is also possible that physical work contains naturally repetitive cycles, and the model may have learned to detect and extend such patterns, as language models do with general sequences.18 The model was pretrained on randomly sampled continuous spans from our data engine (activities captured in homes, warehouses, factories, and elsewhere) with no bespoke infrastructure for packing examples into context — physical prompts introduce discontinuous jumps in time that the model never saw in training.
Robotics is inherently multimodal; and as in human learning, there are many ways to teach a robot something new — the two options of either (a) demonstrations or (b) language instructions are perhaps the most natural for having humans specify tasks.19 While language suffices for some task specifications, many physical actions are difficult to precisely describe in language20 (e.g. it is far easier to show exactly how to seat two Lego bricks than to say it). Prompting a task in native observations and actions is also a more comprehensive test of sensorimotor understanding: the model must infer the goal from the demonstration, repurpose existing knowledge, and improvise under new initial conditions…
…In some cases, in-context learning transfers across the embodiment gap entirely: a human demonstrates a task with their own hands, observable through the robot’s cameras, and the robot can reproduce it immediately afterward….
…In one example, we demonstrated using a brush to sweep a block into a bowl, and fine-tuned the model on 5 minutes of human demonstrations. The model was able to figure out how to use a variety of other tool options besides the brush in order to accomplish the task. When presented with a banana, it used the banana as a makeshift brush. When presented with a dustpan, however, the model exhibited a larger strategic departure from its demonstrations, and through a variety of means would use the dustpan to lift up the block and dump it into the bowl. Neither the fine-tuning data nor, to the best of our knowledge, the pretraining data contains a dustpan used this way, and the nearest pretraining examples bear little resemblance to the task. Handed a dustpan, the model composed an entirely new contact sequence to complete the task out of the box, with no language guidance…
…Although the model was only fine-tuned to put the block into a bowl, it appears to be able to remove obstacles (like a piece of paper covering the bowl) to complete the task, and sometimes place the paper back on top of the bowl. There was no paper covering the bowl in the 5 minutes of task-specific data (with which the model was fine-tuned for only 1 gradient step), and no such task in this setting (to the best of our knowledge) was in the pretraining data…
…GEN-1.5 is a milestone we believe to be profound scientifically, not because of higher success rates, but because it represents a new frontier of generality — one that challenges our own understanding of how these models behave when pretrained at a scale of physical interaction data few thought possible without shortcuts…
…What is clear now, and perhaps obvious in hindsight, is that past a certain threshold of pretraining, the cost of adaptation becomes negligible. Emergent in-context learning from a few seconds of data, or one gradient step on one minute of demonstrations, is no longer task-specific training in the conventional sense. It is closer to reminding the model of something it nearly knows, with a tiny amount of compute. That this works at all, changes how we think about how these models can be used, about their potential impact, and the road ahead for building general physical intelligence.
3. Apple forced to restructure ATT – Eric Benjamin Seufert
Yesterday, the German competition authority, the Bundeskartellamt, concluded its antitrust investigation into Apple’s App Tracking Transparency (ATT) framework after Apple agreed to legally binding changes to the ATT prompt and related consent design…
…Apple now has four months to implement these changes, and the commitments will remain binding for seven years…
…I believe the agreed-upon changes to the ATT prompt, as well as the ability for developers to bundle other data-use consent requests with the ATT prompt, will nudge opt-in rates upward by a non-trivial amount. Apple has stated that the changes will apply in almost all EU countries, and I expect the remaining European investigations to resolve similarly.
But I’m skeptical that the US will adopt similar measures, and for that reason, the broader impact on the digital advertising market of this restructuring of ATT will be muted: without any changes in the US, and given that historical ATT opt-out decisions will not be automatically reversed en masse, my sense is that the principal impact of the concessions extracted from Apple by the Bundeskartellamt will be precedential rather than immediately economic…
…So ATT will survive, but Apple’s ability to impose asymmetric rules on third parties under the banner of privacy without meaningful constraint will not. The immediate economic impact of these concessions may be modest, particularly if the changes stop in Europe, as I believe they will (for a more extensive argument on why I think that’s the case, see Could ATT be rolled back?). But their precedential impact is consequential. ATT was never simply a privacy policy; it was an exercise of platform power, and that power is now circumscribed.
4. A $21 Billion ‘Kids in Chips’ Startup Is Scooping Up Nvidia Talent – Robbie Whelan
Etched doesn’t make software but semiconductors, a product category whose high capital costs and long development cycles have relegated it to a corner of Silicon Valley where experience still trumps youthful ambition…
…Etched has reached a key milestone: signing up its first customer—Jane Street, the secretive Wall Street quantitative-trading giant—for its product, a server rack filled with AI processors optimized for rapid inference computing. The startup has booked more than $1 billion in orders and already started shipping chips…
…The startup says it took just 44 days after getting its test chips back from Taiwan Semiconductor Manufacturing to have them up and running inference workloads—the computing processes that allow AI models to respond to user queries—a process that usually takes six months or more.
“We may be the only AI chip that was built by a startup that was successful on the first try,” said Robert Wachen, co-founder and president of Etched…
…Around 15% of Etched’s roughly 400 employees previously worked at Nvidia, and it has recruited aggressively from the market-leader, as well as from other established semiconductor firms…
…Then there is the 2-megawatt in-office data center. Lined with refrigerator-sized server systems, the room emits a low roar from its cooling systems and allows potential customers to remotely access and try out Etched’s chips.
Another way in which Etched is atypical is in how much of its supply chain it owns and controls. The company has a team of 20 in Taiwan, where it owns a factory that builds server components. Most chip startups outsource much of the testing of their finished products, while Etched does almost all of it in-house.
Company executives say they design their chips using a distinctive design approach called “cluster-scale memory,” a way of allowing multiple chips inside a server to communicate more quickly and thus act as one processor. They say communications tasks that take about 4,000 nanoseconds for an Nvidia Blackwell processor to perform can be done by Etched’s chips in 700 nanoseconds, thanks to the chip’s architecture and custom interconnections.
5. The Future Of AI Compute Won’t Run On Just One Kind Of Chip – Liz Allan, Satadal Bhattacharjee, Ashish Darbari, Moshiko Emmer, Sharad Chole, Cameron Brunner, and Sumit Vishwakarma
Bhattacharjee: That’s right. One of the trends we are seeing, which Nvidia started, is disaggregating the inference pipeline. In inference, there are a few stages. One is called prefill, where you enter a prompt and it’s just trying to figure out what you’re asking so that it can take action. The prefill stage is extremely compute-intensive. Recently, Nvidia announced that its Groq 3 LPU (language processing unit) will be used in the prefill cluster, showing more than a GPU is required to do some of these tasks. Then there’s decode, where it actually does the task, or creates the response that will be generated and shown. Then with the agents coming in, there is the tool calling or executing the task — for example, booking the Uber ride or making a hotel reservation.
We are seeing that with inference disaggregated, you have a cluster with a specific set of hardware and software to do the pre-filled stage. Then, you have a cluster to do the decode stage, and then you have another fully compute cluster to do the execution by the agents. And these are all stitched together. They’re all communicating with each other, most likely through Ethernet right now, but each of the clusters has a different mix of software and hardware to do the function that they do best. That’s going to be more of the norm going forward, because until now every AI problem was solved with a GPU, and the industry is recognizing that every nail doesn’t have one hammer…
… There are some companies that have come up, such as Gimlet Labs and Together AI. Their pitch is that they provide the software layer to run on top of a heterogeneous hardware environment, even on disaggregated inference, and make sure the prefill cluster is run optimally, the decode cluster is run up to value, and the compute cluster is run optimally to do each of the workloads. They are taking care of the software orchestration, and they are working with different companies to do that. This is a very critical part when you go beyond the top hyperscalers like Google, which are invested in creating these custom TPU clusters. Not many companies can do that. Especially if you go to the neocloud guys, which are the next level of cloud providers like DigitalOcean, CoreWeave, Lambda Labs [now Lambda AI], Verda, and others, they’re all going to have this same challenge. There is a desire to bring more hardware diversity. There’s a desire to bring token costs down — of course, without sacrificing performance — and do it without having a single company’s hand holding you. This is a big challenge that we are facing right now. We are at the early stages. It will take some time before this problem is solved, but there’s a lot of concentrated effort going into building this heterogeneous cluster with optimized software coming and solving this big problem that we have, where everybody is waiting in line for Nvidia systems because that’s the only system that is proven to be working at scale right now.
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), Apple, and TSMC. Holdings are subject to change at any time.