John Spindler

How do you trade on short AGI timelines? This weekend, I made a graph showing every layer of the AI supply chain and which companies are the most important bottlenecks. (for my personal use, no advice)

AI Semiconductor Endgame 2026 (Part 1) New Token Economics Computing Paradigm Shifts from GPU Compute to HBM This article starts from the essence of GPU architectural evolution to address a question the market has long worried about: Why must each GPU's HBM memory demand grow exponentially, and why won't this exponential growth in HBM demand stall? It then derives the first principle of token economics under the current architecture: token throughput = HBM size × HBM BW (bandwidth) It also discusses why the GPU ceiling is determined by HBM's two dimensions of progress. The topic of HBM cyclicality has long been controversial. Optimists argue that AI-driven demand is much greater than before, but the market mainstream still believes that previous up-cycles also saw 20%+ annual demand growth — so what's different this time? AI doesn't change the fact that HBM, like traditional DRAM, has commodity attributes. Once capacity expansion at the demand peak meets a downturn, history will repeat itself. We can take the perspective of compute-chip architecture, start from first principles, and unpack and reason through this question: why this time is genuinely different. ——————————————————————————————— History: The Era of CPU Compute For a very long time, we lived in the era of CPU-dominated compute. The CPU's top-level KPI was performance — running faster — and so each generation of CPUs deployed every method imaginable to push benchmark scores higher. First it was rising clock frequencies, then it was architectural evolution: superscalar designs, and so on. During this period, why didn't DDR need to advance technologically at high speed? DDR3 to DDR5 took a full 15 years. Because in this era, DDR's role was purely auxiliary — and only weakly so. By industry experience, even doubling DDR speed would generally only raise CPU performance by less than 20%. Why did improvements in DDR bandwidth and speed matter so little? Two reasons: 1. CPUs designed all kinds of architectural tricks to hide DDR latency — superscalar designs, wider issue widths, massive ROBs and register renaming to extract parallelism and hide latency, L1 caches, L2 caches — all of which weakened the demand for DDR bandwidth and speed. 2. CPU workloads don't have particularly demanding bandwidth requirements. For most everyday workloads — say, opening a webpage — DDR bandwidth is severely overprovisioned. Even cloud workloads often look the same. In other words, in the CPU era, DDR bandwidth and speed didn't really matter. There was virtually no difference between DDR4 and DDR5 except in a handful of games — and even the JEDEC standard advanced slowly. On top of that, only a small portion of any given app needs to permanently sit in DDR. Whatever is needed can be paged in from the hard drive on demand. App size grew slowly, and so DDR capacity demand grew slowly as well. That's why, over the past decade, the average PC went from 7–8GB of DDR to about 23GB — only 3× growth in ten years. This slow upgrade pace directly affected revenue. Capacity-based pricing was the main way of making money; speed improvements were just a technological upgrade that raised the unit price of capacity. With both of these dimensions advancing slowly, growth could only come from increases in PC/phone unit volumes. So along both dimensions — bandwidth/speed and capacity — DRAM was always a “nice-to-have” appendage to the chip industry. The marginal utility of DDR upgrades was very low, and almost completely disconnected from the CPU era's top-level KPI. ——————————————————————————————— The Paradigm Shift: GenAI's Top-Level KPI When we entered the era of GenAI large models, the computing paradigm shifted, and the top-level KPI changed fundamentally. By the time GPUs evolved into AI inference engines, the top-level KPI was no longer compute alone (TOPS/FLOPS), as it had been for CPUs — it became the cost of a token. Specifically: overall token throughput per unit cost / per unit power. A close second is token throughput speed — because in the agent era, many tasks have become serial, and token output speed has become a critical bottleneck for user experience. This is exactly why Jensen invented the concept of the AI factory: to produce the most tokens at the lowest cost, while pushing token throughput speed as high as possible. In the AI training era, Jensen's economics were TCO (Total Cost of Ownership): the more GPUs you buy, the more you save. In the inference era, Jensen's token economics flip the logic: AI inference has very healthy gross margins, so the logic now becomes: the NVIDIA GPU is the GPU that produces the cheapest token in the world, so the more you buy, the more you earn. The top-level KPI has become a Pareto frontier: along the two dimensions of token throughput and token speed, optimize as far as possible. Each generation of NVIDIA's token factory is essentially pushing the entire Pareto frontier up and to the right. This is the most important KPI of the AI inference era. ——————————————————————————————— From Token Throughput to HBM: The Core Logic Chain Below is the most important logical chain of this article: how to start from the exponential growth of token throughput and derive that the ceiling bottleneck lies in the exponential growth of HBM size and HBM speed. In the era of single-GPU inference with single-thread batch size = 1, token throughput had only one dimension: HBM bandwidth speed. Higher bandwidth = higher token throughput. But once we entered the NVL72 era, inference is no longer single-GPU. It is a system-level token factory composed of 72 GPUs + 36 CPUs, designed to fully saturate HBM bandwidth and compute simultaneously, in pursuit of the ultimate token throughput. Token throughput growth depends on two things: the number of requests batched simultaneously × the average token speed per request. That is: batch size × token speed. Take Rubin NVL72 as an example. At an average token speed of 100 tokens/s, processing 1,920 simultaneous requests yields a token throughput of 192,000 tokens/s. A Rubin NVL72 draws roughly 120kW (0.12MW), so per MW it can handle 1.6M tokens/s. So we need to find ways to push both parameters up: batch size and average token speed. Their product is our top-level KPI — token throughput. Parameter 1: Batch growth — bottleneck is HBM size Every request in the batch carries its own KV cache, which has to live in HBM, with sizes ranging from a few GB to tens of GB. Because hot KV cache must be read at high frequency and high speed at any moment, it must reside in HBM. For a model with, say, 80 layers, every token generation step requires reading the KV cache 80 times from HBM. As batch size grows, hot KV cache grows linearly. And because the hot KV cache for every request in the batch must sit in HBM, HBM size must grow linearly with batch size. Like an airport shuttle bus: the gate wants to move passengers to the plane as fast as possible. If HBM size is small, the shuttle is small, so you have to make extra trips. Conclusion: batch size growth bottlenecks on HBM size growth. Parameter 2: Average token speed per request — bottleneck is HBM bandwidth The decode-phase speed of a large model bottlenecks on HBM bandwidth, because every token generated requires reading the activated weights and KV cache many times over. The emergence of LPUs has, in cases where batch size isn't very large, moved the activated weights portion onto SRAM — but every generated token still requires many reads of the KV cache from HBM. The higher the HBM bandwidth, the faster each token is generated, in essentially linear correspondence. Like the airport shuttle bus: HBM bandwidth is like the width of the door — wider doors mean passengers board faster. The rest of the GPU's configuration is essentially adapted to support batch growth and to keep token compute speed in step with HBM growth. In some cases the GPU even spends excess compute to recover effective bandwidth (e.g., bandwidth compression techniques). —------- To return to the shuttle bus analogy: • Shuttle bus cabin size = HBM Size (capacity): determines how many passengers can fit at once (i.e., how many requests' KV caches can sit in HBM simultaneously). Bigger cabin = more passengers (higher batch size) per trip. If the bus is too small, moving 100 people takes two trips — and total throughput suffers. • Shuttle bus door width = HBM Bandwidth: determines how fast passengers get on and off. A wide door, and everyone piles on at once (decode/token generation is fast). A narrow door, and even with a giant cabin, people queue up and most of the time is spent boarding. • Passenger throughput = cabin size × door-width-determined boarding speed. —------- At this point, we've logically derived the first principle of token-economics hardware demand: Token throughput = HBM size × HBM Bandwidth The top-level KPI of the AI inference era is highly dependent on progress along both HBM dimensions. If we want to maintain 2× token throughput growth per generation, that means each generation of single GPU must grow HBM size × HBM BW speed by 2×! This is the first time in history that HBM memory size can influence the top-level KPI — token throughput. To validate this thesis, we can put NVIDIA's token throughput from A100 to Rubin Ultra on the same chart as HBM size × HBM BW speed. What you find is that the two curves track each other startlingly closely on log axes. HBM size × speed actually grows even faster than token throughput — which makes sense, because HBM defines the ceiling, and in practice utilization of that ceiling is very hard to push to 100%. Even if HBM size × HBM speed grew by 1,000×, with the supporting compute and architecture, it would be very hard to wring out the full 1,000× of headroom. This curve isn't a coincidence — it's the necessary solution of system optimization. throughput = batch × speed. This is the unavoidable first principle of token factory economics. —------- What about software? Won't software optimization reduce bandwidth demand? Reduce HBM demand? This is an independent dimension from hardware. It's like asking: if software on a CPU runs faster after optimization, does that mean the CPU doesn't need to advance for ten years? After all, software is faster now. If that were the case, would CPU vendors still make money? For a CPU vendor to survive, there's only one path: in standardized benchmarks, ignoring software optimization, every new CPU generation must score higher — otherwise it doesn't sell. GPUs are exactly the same. How well software is optimized, and the requirement that the GPU's own token-throughput KPI must improve dramatically every year, are two separate things. As long as token demand keeps growing, the pursuit of higher token throughput will not stop — and so neither will the pursuit of higher HBM size × HBM speed. If HBM size and HBM speed were to slow down, Jensen would personally fly to the Big Three and pressure them to accelerate, because that ishis GPU ceiling. If the ceiling stops rising, can his GPU still sell? Of course, NVIDIA also needs to wrack its brains to extract performance beyond the HBM ceiling through heterogeneous architectural angles. The LPU is a great example — it improved the Pareto frontier substantially from a different angle (the right-hand high-token-speed portion). —-------------------- HBM memory has now bid farewell to that old era of drifting with the tide. On this one-way road paved by exponential demand, it has, in something close to a destined fashion, walked onto the central stage of the industry's epic. When the inference paradigm's first principles evolve to this point, as long as Jensen still wants to sell GPUs, HBM must double — and it must double every generation. This is endogenous pressure from the supply side. It has nothing to do with AI demand, nothing to do with macro cycles, and nothing to do with the moods of the hyperscalers. The only remaining question is this: When demand has been physically locked into exponential growth, will the three players on the supply side — like they have for the past thirty years — once again drag themselves back into the mire of the cycle by their own hands?

We are launching Ineffable Intelligence with David Silver and my exceptional co-founders! We’re building a system that can discover all knowledge from its own experience, from elementary motor skills to profound intellectual breakthroughs. We expect it to rediscover, and ultimately transcend, humanity’s greatest inventions: language, science, mathematics, and technology. We believe this can be built within years. There is a real risk of failure, in pursuit of a small chance of extraordinary success that could change the course of AI, and with it, humanity. If this resonates with you, join us. We’re building a world-class team in London. ineffable.ai

the window for experimenting with llms has basically closed now. the megacorps have fully hit escape velocity and are shipping new products and new features daily. the shift is that they’re not just shipping llms anymore, they’re using llms to build products and improve existing ones at scale. the wild west era of llms isn’t really the wild west anymore. a year ago, this could’ve been an indie dev side project, maybe even a monetizable product. it was literally so easy that the only real bottleneck was your free time. now, whatever idea you have, you should basically assume google/anthropic/oai will build some version of it within a week and wipe out most of the startup surface area around it

All of these points are true.

My biggest takeaways from Claude Code's Head of Product @_catwu: 1. Anthropic’s product development timelines have gone from six months to one month, sometimes one week, sometimes one day. Part of this acceleration is access to the latest models (i.e. Mythos). Another is shipping new products into “research preview,” making clear it's early, experimental, and might not be supported forever. Another is an evergreen "launch room "where engineers post ready features and marketing turns around announcements the next day. 2. The PM role is shifting from coordinating multi-month roadmaps to enabling teams to ship daily. As Cat puts it, “There should be less emphasis on making sure you are aligning your multi-quarter roadmaps with your partner teams and more emphasis on, OK, how can we figure out the fastest way to get something out the door?” 3. The most efficient shipping unit is an engineer with great product taste. On Cat’s team, many engineers go end-to-end—from seeing user feedback on Twitter to shipping a product by the end of the week—without a PM involved. Also, almost all the PMs on the Claude Code team have either been engineers or ship code themselves, and the designers have been front-end engineers. The roles are merging, and the most valuable skill is product taste, not job title. 4. Build products that are on the edge of working. Claude Code’s code review product failed multiple times because earlier models weren’t accurate enough. But because the prototype was already built, they could swap in Opus 4.5 and 4.6 and immediately test whether the gap was closed. Teams that wait for the model to be ready will always be a cycle behind. 5. The most underrated skill for building AI products is asking the model to introspect on its own mistakes. Cat regularly asks the model why it made an unexpected decision. The model will explain that something in the system prompt was confusing, or that it delegated verification to a subagent that didn’t check its work. This reveals what misled the model so the team can fix the harness. 6. Every model release forces their team to revisit existing products and audit their system prompt to remove features the model no longer needs. Claude Code’s to-do list was a crutch for earlier models that couldn’t track their own work. With Opus 4, the model handles it natively. Features built as scaffolding for weaker models become debt when the model catches up—so the team actively strips them. 7. Anthropic employees build custom internal tools instead of buying SaaS products. A sales team member built a web app that pulls from Salesforce, Gong, and call notes to auto-customize pitch decks—work that used to take 20 to 30 minutes now takes seconds. Their core stack is Claude Code, Cowork, and Slack. No Notion, no Linear, no Figma. 8. People underestimate how much Claude’s personality contributes to its success. As Cat describes it, “When you reflect on everyone you’ve worked with, there’s just some people where you’re like, I really like their energy, their vibe.” Claude is designed to be low-ego, positive, competent, and earnest—qualities that make it feel like a great coworker, not just a tool. This isn’t cosmetic; it’s what makes people want to use Claude for hours every day. The team has a dedicated person, Amanda, who “molds Claude’s character,” and it’s one of the hardest roles at the company because success is so subjective. 9. The future of work is managing fleets of AI agents, not doing the work yourself. Cat sees a clear progression: first, individual tasks become successful. Then people start running multiple tasks at the same time (multi-Clauding). Next, people will run 50 or 100 tasks simultaneously, which will require new infrastructure—remote execution, better interfaces for managing tasks, agents that fully verify their work, and self-improving systems that incorporate feedback. The human role shifts from doing the work to knowing which tasks to look into, verifying outputs, and giving feedback that makes the system better over time. 10. Hire people who lean into chaos and face every challenge with a smile. At Anthropic, there are weeks when a P0 on Sunday becomes a P00 by Monday and a P000 by Monday afternoon. If you get too stressed about any one thing, you’ll burn out. Their team looks for people who can look at a hard challenge and say, “Wow, that’s gonna be hard. But I’m excited to tackle it and I’m gonna do the best that I possibly can.” This mindset—optimism, resilience, and comfort with constant change—is increasingly essential as the pace of AI development accelerates. Don't miss the full conversation: youtube.com/watch?v=Pplmzl…

MIT just made every AI company's billion dollar bet look embarrassing. They solved AI memory. Not by building a bigger brain. By teaching it how to read. The paper dropped on December 31, 2025. Three MIT CSAIL researchers. One idea so obvious it hurts. And a result that makes five years of context window arms racing look like the wrong war entirely. Here is the problem nobody solved. Every AI model on the planet has a hard ceiling. A context window. The maximum amount of text it can hold in working memory at once. Cross that line and something ugly happens — something researchers have a clinical name for. Context rot. The more you pack into an AI's context, the worse it performs on everything already inside it. Facts blur. Information buried in the middle vanishes. The model does not become more capable as you feed it more. It becomes more confused. You give it your entire codebase and it forgets what it read three files ago. You hand it a 500-page legal document and it loses the clause from page 12 by the time it reaches page 400. So the industry built a workaround. RAG. Retrieval Augmented Generation. Chop the document into chunks. Store them in a database. Retrieve the relevant ones when needed. It was always a compromise dressed up as a solution. The retriever guesses which chunks matter before the AI has read anything. If it guesses wrong — and it does, constantly — the AI never sees the information it needed. The act of chunking destroys every relationship between distant paragraphs. The full picture gets shredded into fragments that the AI then tries to reassemble blindfolded. Two bad options. One broken industry. Three MIT researchers and a deadline of December 31st. Here is what they built. Stop putting the document in the AI's memory at all. That is the entire idea. That is the breakthrough. Store the document as a Python variable outside the AI's context window entirely. Tell the AI the variable exists and how big it is. Then get out of the way. When you ask a question, the AI does not try to remember anything. It behaves like a human expert dropped into a library with a computer. It writes code. It searches the document with regular expressions. It slices to the exact section it needs. It scans the structure. It navigates. It finds precisely what is relevant and pulls only that into its active window. Then it does something that makes this recursive. When the AI finds relevant material, it spawns smaller sub-AI instances to read and analyze those sections in parallel. Each one focused. Each one fast. Each one reporting back. The root AI synthesizes everything and produces an answer. No summarization. No deletion. No information loss. No decay. Every byte of the original document remains intact, accessible, and queryable for as long as you need it. Now here are the numbers. Standard frontier models on the hardest long-context reasoning benchmarks: scores near zero. Complete collapse. GPT-5 on a benchmark requiring it to track complex code history beyond 75,000 tokens — could not solve even 10% of problems. RLMs on the same benchmarks: solved them. Dramatically. Double-digit percentage gains over every alternative approach. Successfully handling inputs up to 10 million tokens — 100 times beyond a model's native context window. Cost per query: comparable to or cheaper than standard massive context calls. Read that again. One hundred times the context. Better answers. Same price. The timeline of the arms race makes this sting harder. GPT-3 in 2020: 4,000 tokens. GPT-4: 32,000. Claude 3: 200,000. Gemini: 1 million. Gemini 2: 2 million. Every generation, every company, billions of dollars spent, all betting on the same assumption. More context equals better performance. MIT just proved that assumption was wrong the entire time. Not slightly wrong. Fundamentally wrong. The entire premise of the last five years of context window research — that the solution to AI memory was a bigger window — was the wrong answer to the wrong question. The right question was never how much can you force an AI to hold in its head. It was whether you could teach an AI to know where to look. A human expert handed a 10,000-page archive does not read all 10,000 pages before answering your question. They navigate. They search. They find the relevant section, read it deeply, and synthesize the answer. RLMs are the first AI architecture that works the same way. The code is open source. On GitHub right now. Free. No license fees. No API costs. Drop it in as a replacement for your existing LLM API calls and your application does not even notice the difference — except that it suddenly works on inputs it used to fail on entirely. Prime Intellect — one of the leading AI research labs in the space — has already called RLMs a major research focus and described what comes next: teaching models to manage their own context through reinforcement learning, enabling agents to solve tasks spanning not hours, but weeks and months. The context window wars are over. MIT won them by walking away from the battlefield. Source: Zhang, Kraska, Khattab · MIT CSAIL · arXiv:2512.24601 Paper: arxiv.org/abs/2512.24601 GitHub: github.com/alexzhang13/rlm

Direction of AI mid through late-2026: 1. Big labs are gonna push expensive bigger closed-source models directly to big tech. The moat will shift away from consumer markets coz OSS models are getting too good to compete at current price point. Plus big labs got more money to make directly going B2B. 2. Open source labs are making comparable coding models now. They lack marketing exposure, but it will be impossible to keep serving (example) Opus at the ridiculous token price Anthropic is. Qwen, Kimi, Minimax, GLM, etc... anybody got a clear shot here to deliver a Sonnet or a GPT5.2 at 1/10th pricing, completely agentic-pilled with coding and tool calls. 3. Local models are gonna go crazy because people will figure out speculative decoding + kv cache quantization to make models run fast on-device. If Qwen 3.6 27B is any indication, local coding models will be a thing soon enough. 4. Devs will realize that there is a lot of money to be made by making AI first local apps that use private local edge LMs (~0.5-4B models). You can literally see indications of all 4 things above if you followed last 2 weeks of AI news. Mythos, Kimi K2.6, Qwen3.6-27B, DFlash, TurboQuant, Gemma-4... live examples of all the above at play. I feel this is the next phase of evolution for LLMs.

sorry but this is seriously fucking impressive china just shipped a claude code-level ai model small enough to run on your laptop. it codes better than opus 4.5 and its tiny. 27B beating models 15X LARGER than it. best part? shit is fully open source. no cloud, no rate limits, no api keys this model plus kimi k2.6 tells me open source has caught up to the frontier models how the fuck did china pull this off?

A fully autonomous chemistry lab for $5,000. RoboChem-Flex just landed in Nature Synthesis and it's a signal, not an outlier. The hardware barrier is falling. AI agents are entering the physical lab. Which raises the real question: where does the AI learn to be a scientist first? Not in a $5K setup. Not in any physical lab. AI needs a place to plan, run, debug, and iterate, at digital speed, at infinite scale, before it ever touches real reagents. That's what we're building. LabWorld Factory turns real-world biomedical protocols into a fully executable simulation universe. 100 realistic lab assets. 1,000 atomic skills. 10,000+ executable biomedical protocols. Prompt a lab. Generate a world. Let the agent run. Infinite lab → infinite iteration → infinite science. The closed-loop AI scientist era isn't coming. It's being built now.

# The Path Forward for AI Startups A lot of founders are messaging each other after the SpaceXAI <> Cursor “IPO-deferred acquisition”. Common discussion topic: what is the future for independent startups? Must ~everyone ultimately be acquired by a frontier lab or go extinct? The data from our direct experience @cognition suggests the opposite. The more startups in a category that defect from independent competition by selling to a lab, the stronger the remaining ones become. We experienced this firsthand last year with Windsurf. When the founders went to Google and we acquired the remaining company, it dramatically accelerated our product roadmap and GTM. Now, cloud agents are ready for prime time, and our usage has exploded. (We’re in the fastest rate of usage growth in Cognition’s history - almost 50% month-over-month growth in Devin enterprise.) We already see the next round of acceleration with yesterday’s news, from prospects and customers to candidate inbound. In just about every category, there’s a clear market for a winning independent offering that’s not tied to models from any one lab. Especially in a space as dynamic as software engineering, where customers value model flexibility as the rankings from different providers are constantly changing. For startups to seize that independence opportunity, here are the lessons we’ve learned so far: 1. DIFFERENTIATION You need to have extremely clear differentiation vs. what’s already offered by the labs. Cursor had stiff competition from Claude Code in self-serve, in part because one tool was substitutable for the other, which presented a challenge. Our approach has been to differentiate heavily for enterprises, which is the largest market for software engineering. Specifically: 1. We invest as much in forward deployed engineering and AI enablement as we do in core R&D. Our customers treat us as a change management partner, not just an AI software engineering platform. We run 1000-person workshops all around the world to help train developers inside companies on frontier AI adoption. We target specific use cases and outcomes in addition to providing developer tooling. 2. We focus on accelerating the *entire software development lifecycle* at large company scale, not just the writing of code. Devins now spin up automatically for everything from ticket scoping to DeepWiki codebase indexing to security vulnerability remediation and application monitoring alert response. 3. We eat the pain of deployment complexity to work well in the largest and most complex environments imaginable. Cognition can run inside a customer’s virtual private cloud, has a permissioning and team collaboration model that can scale to 100,000+ developers inside one company, runs as well for COBOL mainframes as it does for modern Python. From day 1 each Devin ran in a microVM on its own machine, vs running locally as a CLI tool, which allows arbitrary horizontal scaling and is a better fit for event-driven automation. Of course, one element of startup differentiation will always be model independence. This is particularly powerful in large enterprises, who value supplier continuity and the ability to centralize tooling without taking on the business risk that they committed to the wrong foundation model. And useful for individual developers, who always want to try the latest models. (If you haven’t yet tried the Windsurf 2.0 release which came out last week, it’s a good day to give it a shot!) I expect the labs will catch up on some of these fronts at some point. But at that point, we’ll have already made the next leap in differentiation, because… 2. FOCUS You won’t outcompete the labs in everything, but you can outcompete the labs in *your* thing. Every application domain has fractal complexity at the edges. Lean in to what makes your domain special and offer things no one else can. Does it make sense for a lab to devote training resources to a specialized code review model? Probably not - they’re working on AGI. But for the 3-6 month window where the latest frontier models don’t solve that use case at acceptable performance, cost, or latency, do it yourself and build a better product experience than would otherwise be possible. Rinse and repeat as the frontier of what’s possible via specialization continues to evolve. 3. VELOCITY One of our values at Cognition is: “Every second counts.” Maniacal urgency helps in every startup, but it counts extra in today’s accelerated AI times where advantages compound faster than before. With sufficient focus, you can out-accelerate the AI labs on any one specific feature or workflow. Do this consistently to stretch the overhang of what’s enabled by each new generations of models, and you can maintain your edge on a differentiated product experience. - In many ways the SpaceXAI <> Cursor news is a win for everyone. SpaceX gets a new research team and the chance to become competitive in coding. Cursor gets a meaningful exit and the opportunity to accelerate their research roadmap with much more compute. And the whole ecosystem benefits from increased competitiveness among the foundation model labs. Congrats to the teams on the outcome.

The continual learning vs in-context learning debate is a sideshow from the fact we lack interesting training environments. If you only care about narrow things like SWE and automated AI research, then you create optimisation pressure towards narrow agentic capabilities. And settings like long horizon - which should be petri dishes for novel behaviour - become limited to long duration SWE tasks. The reason why humans have such extraordinary generalisation capabilities in the first place is that their environment created selective pressure for it. In particular, extreme non-stationary, multi-agent interaction, and the need to act under uncertainty. One source of selective pressure historically was climate change and patchy food sources. This led to the evolution of our frontal lobe that could generalise abstract rules from one situation to another. Another source of pressure came from the need to model quickly changing social situations. Each of these pressures (and more) likely encouraged rapid learning. We have nothing close to the richness of these settings in our training environments, and the limits of most people’s imagination here seems to be narrow AI research optimisation challenges. So obsessing over the learning algorithm before we’ve even tried anything remotely interesting on the environment side feels premature to me. Ironically the selective pressure we are seeing is towards more boring environments because of short-term market forces that mean the bulk of spend is spent on SWE. This will pass, and we will soon wonder again what’s actually required to move AI beyond centaur mode towards navigating the real world autonomously. Relatedly I think it’s helpful if we start using phrases like “selective pressure” for environment design as it forces us to start thinking of environments as worlds, instead of narrow task sets - which is where the field is at the moment. There’s a reason why our art at @GenReasoning consists of globes 🌍 - because it’s a North Star for how rich and complex our training environments should be. ( 🌶️ )

One of the most substantive classes with @ChaseLochmiller at Stanford. We went deep on economics of the datacenter: - Where is the ~$650B of AI infra capex actually going this year? - Who's capturing the margin, who's getting squeezed? - How the bottleneck has moved from GPUs to power, and where it goes next - The economics of neoclouds

🚨 Just IN: This paper from Stanford and Harvard explains why most “agentic AI” systems feel impressive in demos and then completely fall apart in real use. The core argument is simple and uncomfortable: agents don’t fail because they lack intelligence. They fail because they don’t adapt. The research shows that most agents are built to execute plans, not revise them. They assume the world stays stable. Tools work as expected. Goals remain valid. Once any of that changes, the agent keeps going anyway, confidently making the wrong move over and over. The authors draw a clear line between execution and adaptation. Execution is following a plan. Adaptation is noticing the plan is wrong and changing behavior mid-flight. Most agents today only do the first. A few key insights stood out. Adaptation is not fine-tuning. These agents are not retrained. They adapt by monitoring outcomes, recognizing failure patterns, and updating strategies while the task is still running. Rigid tool use is a hidden failure mode. Agents that treat tools as fixed options get stuck. Agents that can re-rank, abandon, or switch tools based on feedback perform far better. Memory beats raw reasoning. Agents that store short, structured lessons from past successes and failures outperform agents that rely on longer chains of reasoning. Remembering what worked matters more than thinking harder. The takeaway is blunt. Scaling agentic AI is not about larger models or more complex prompts. It’s about systems that can detect when reality diverges from their assumptions and respond intelligently instead of pushing forward blindly. Most “autonomous agents” today don’t adapt. They execute. And execution without adaptation is just automation with better marketing.

Introducing ml-intern, the agent that just automated the post-training team @huggingface It's an open-source implementation of the real research loop that our ML researchers do every day. You give it a prompt, it researches papers, goes through citations, implements ideas in GPU sandboxes, iterates and builds deeply research-backed models for any use case. All built on the Hugging Face ecosystem. It can pull off crazy things: We made it train the best model for scientific reasoning. It went through citations from the official benchmark paper. Found OpenScience and NemoTron-CrossThink, added 7 difficulty-filtered dataset variants from ARC/SciQ/MMLU, and ran 12 SFT runs on Qwen3-1.7B. This pushed the score 10% → 32% on GPQA in under 10h. Claude Code's best: 22.99%. In healthcare settings it inspected available datasets, concluded they were too low quality, and wrote a script to generate 1100 synthetic data points from scratch for emergencies, hedging, multilingual etc. Then upsampled 50x for training. Beat Codex on HealthBench by 60%. For competitive mathematics, it wrote a full GRPO script, launched training with A100 GPUs on hf.co/spaces, watched rewards claim and then collapse, and ran ablations until it succeeded. All fully backed by papers, autonomously. How it works? ml-intern makes full use of the HF ecosystem: - finds papers on arxiv and hf.co/papers, reads them fully, walks citation graphs, pulls datasets referenced in methodology sections and on hf.co/datasets - browses the Hub, reads recent docs, inspects datasets and reformats them before training so it doesn't waste GPU hours on bad data - launches training jobs on HF Jobs if no local GPUs are available, monitors runs, reads its own eval outputs, diagnoses failures, retrains ml-intern deeply embodies how researchers work and think. It knows how data should look like and what good models feel like. Releasing it today as a CLI and a web app you can use from your phone/desktop. CLI: github.com/huggingface/ml… Web + mobile: huggingface.co/spaces/smolage… And the best part? We also provisioned 1k$ GPU resources and Anthropic credits for the quickest among you to use.

Recursive Superintelligence has been valued at $4bn, excluding the new capital, despite its AI concept being at the research stage ft.trib.al/BUToUMy

Vectorless LLM knowledge base with reasoning github.com/VectifyAI/Open…

Claude Code fully dissected! Researchers from UCL reverse-engineered the leaked Claude source. What they found changes how you should think about agent design. Only 1.6% of the codebase is AI decision logic. The other 98.4% is operational infrastructure. Permission gates, tool routing, context compaction, recovery logic, session persistence. The model reasons. The harness does everything else. This is the opposite of what most agent frameworks do today. LangGraph routes model outputs through explicit state machines. Devin bolts heavy planners onto operational scaffolding. Claude Code gives the model maximum decision latitude inside a rich deterministic harness, and invests all its engineering effort in that harness. The core loop is a simple while-true. Call model, run tools, repeat. But the systems around that loop are where the real design lives: A permission system with 7 modes and an ML classifier. Users approve 93% of prompts anyway, so the architecture compensates with automated layers instead of adding more warnings. A 5-layer context compaction pipeline. Each layer runs only when cheaper ones fail. Budget reduction, snip, microcompact, context collapse, auto-compact. Four extension mechanisms ordered by context cost. Hooks (zero), skills (low), plugins (medium), MCP (high). Each answers a different integration problem. Subagents return only summary text to the parent. Their full transcripts live in sidechain files. Agent teams still cost roughly 7x the tokens of a standard session. Resume does not restore session-scoped permissions. Trust is re-established every session. That friction is the point. The bet behind all of this is simple. As frontier models converge on raw coding ability, the quality of the harness becomes the differentiator, not the model. Paper: Dive into Claude Code (arXiv:2604.14228) In the next tweet, I've shared an article I wrote on Agent Harness and what every big company is building. Do check.