My favorite books and reading trace for 2025: I had a five-way tie for the best books this year: 1) Mokyr’s A Culture of Growth on how societies build or block the conditions for sustained progress; 2) John Williams’ masterpiece Augustus, and the runner-up, Stoner, as intimate studies of character under the slow pressure of duty and power; 3) Dhami and Sunstein’s Bounded Rationality as the clearest bridge from “humans are cognitively weird” to “policy should be designed accordingly”; 4) Liu et al.’s Advances and Challenges in Foundation Agents as a big-map survey of how modern agents stitch together memory, planning, tools, and safety constraints into something that actually acts in the world, and; 5) von Neumann’s core computing writings (Compendium, EDVAC, Self-reproducing Automata, and The Computer and the Brain) arguing that computation is best understood as a class of automata with robustness coming from adaptation and distributed memory. Links below.

This is an important achievement, and probably the first in a predictable sequence of results that will lead to fully automating LLM and OMNI training and serving. As someone who has helped build some of the best multimodal and language models, I don’t see how this could play out any other way. What challenges will remain then? For me: environments and applications that matter like healthcare, materials, scientific exploration, and above all energy harnessing and lasting energy storage. With that we can solve carbon capture and end hunger.

This week made something clear: it’s time to stop treating concentration of power in AI as a solution rather than a risk. Safety and centralized control are not the same thing, so let's stop talking about them like they are. Yet scaling laws are real. The threat of AI cyber-hackers disrupting the global economy is real. The threat of somebody using AI to create a biological weapon is real. AI safety is a species-level concern and we need the best minds in our species from across institutions working on solutions, not locked outside the closed doors of frontier AI development. We need a shared research commons at the intersection of industry, academia, and the public good—an open research ecosystem with access to billions of $ in compute, SoTA models, and strict protocols for dissemination that ensure the most impactful discoveries in the history of our species are shared in a safe way that benefits all of humanity. Not blind open source ideology. Not closed access as safety theater. An open frontier, shaped by many and accountable to all.

I would read everything Roon wrote here with seriousness. The stakes have gone up. I’ve changed my mind and am now preparing for an ASI world coming sooner than many expect, and the early signs from last year and this one have been deeply unsettling for me. I’m basically ASI-pilled now; not as sci-fi, but something we need to prepare for. New essay: open.substack.com/pub/hashco…

The best tweet on this platform, bar none.

Thoughts on the Anthropic vs US government dispute over the Fable model: -This is the most consequential geopolitical news in a while - I put it on par with Russia invading Ukraine, the CCP taking over mainland China, or the Soviet Union falling, but it will get even more important over time. -Why? Recent Anthropic models have helped run an American war against Iran and given the US national security state god-level powers to hack any other institution or government in the world. And the current Mythos/Fable level models, which are already super intelligent (albeit in a jagged way) by most definitions, are lagging what even smarter and more capable models will be like in 1-2 years, trained on Rubin and Feynman chips that will soon come out. I expect the next models to be super intelligent AI swarms operating 24-7 as massive teams of hundreds to thousands of agents, with long horizon action and memory far beyond human teams. -Basically the US has re-entered a unipolar moment with this technology, the National Security State has woken up to it and is trying to get more control and shut down access, and even within the US the Congress and federal courts (the theoretical checks and balances) don’t grok it and are sitting on the sidelines. To make it starkly clear, Anthropic recent put in anti-dictatorship training in their models (via Andy Hall’s evals at Stanford), because they are worried about centralized control of this technology (by any administration). -So, who gets access, and under what terms? The ad hoc export controls suggest it will just be a handful of allowlisted American institutions, and not even American allies or foreign residents working to build AI within frontier labs. This is problematic since I estimate that 80-90% of the skilled AI workforce is foreign born, so we will need to start doing background checks on builders and even KYC on users (I strongly dislike this, but it seems likely). Also, I expect an economic and technological boom that only these American institutions will get, the longer the rest of the world refuses to build data centers and models on nonsensical grounds of local politics (basically not understanding the central issue of their sovereignty and their viability as political and economic actors, to have access to these AI models). I expect a sovereign data center and model creation boom to start in the handful of countries with the knowledge, capital, and will to get going (China, Switzerland, UAE, maybe Norway, etc). -The next 5-15 years may be the most consequential period of world history up to now, as we cross AGI quickly and get to ASI. How do we train ASI to align to American values, goals, and institutions, or broadly the wisest human ones? How do we democratize access to power and abundance, or is it hoarded by elites in DC and SF, who then entrench themselves in powerful positions? How do we manage the next phase of recursive self improvement as ASI goes from a superhuman level we can observe and modestly understand to an even more complex level (millions of agents coordinating 24/7 at an Einstein or Von Neumann level) pushing science, technology, and economics to a level our flesh brains can’t comprehend (the singularity may come sooner than Kurzweil’s prediction of 2045). -So this isn’t just a minor dispute between an AI lab and the US government. This is more like Caesar crossing the Rubicon, Genghis Khan unifying the tribes to start a global expansion, or Archduke Franz Ferdinand getting assassinated and setting up the world for two World Wars. I hope we can manage it with prudence, open democratic deliberation, and care, across the US and multiple countries, but don’t take it for an ordinary event. 2026 is the clear starting point for the run up to whatever the singularity brings.

Great post from Tyler - I’m connecting some other dots and think this event and the last few months are still under appreciated by many. I’ve recently become ASI-pilled. Mythos and Fable pushed me over the edge, plus knowing these are mere B200 trained models while much more powerful and capable Rubin and Feynman level models will be coming over 2-3 years. My thoughts: x.com/sudoraohacker/status/2…

Relationship between Anthropic and USG is something to be studied. I predict that the relationship between private sector AI industry and governments of the world will be defining of the years to come. Frank Underwood once said, "you may have all the money in the world but I have men with guns". When the answer to that becomes "You may have all the men with guns but I have god in a datacenter", that conversation becomes extremely extremely meaningful

Odds of this have gone up, alongside straight up nationalization on economic and national security grounds.

Because this is happening to Anthropic, the temptation for many will be to say: Play stupid games, win stupid prizes. They have relentlessly raised the regulatory temperature in Washington by inviting far-reaching controls of frontier models. They made this bed and now they have to lay in it. But this decision by the Trump administration should not be judged on a desire for payback politics, but on the merits, and specifically what it means for America's broader AI objectives. In that regard, this action is truly outrageous. How exactly is the government planning on even going about verifying everyone who uses this specific model to ensure compliance? That alone raises huge flags. Between the latest Executive Order shifting more control to NSA, and the recent chatter about quasi-nationalization / equity stakes, and now this action, we are talking about a significant escalation in the politicization of AI and centralization of control over advanced computation in this country. And this is all being done by an administration that had previously made acceleration and winning the great AI race a priority. We're moving backwards now.

This is an extreme action and a dramatic escalation from the Trump administration, unprecedented in the world of software and AI models. I expect it will be challenged in court and it suggests that the Trump administration has taken a steep anti-AI turn. Not sure what @DavidSacks or PCAST is doing, but this is worrisome and much more extreme than the Biden administration or what any US state has done. Strange to see this administration go from 100 to 0 on AI policy and I hope this is quickly reversed.

Hockney died. A legend. One of a small number of living artists who will join the canon (among old Western guys, Richter and Turrell also, no question). Pool With Two Figures is, with Shulman's Stahl photo, the greatest piece of California art.

California is one of the most dynamic places on the planet. The paradox of the state today is that its successful economy is attached to a failing model of governance. My latest column: washingtonpost.com/opinions/…

Yann Lecun published the most heretical AI paper of the year. He opens by arguing Magnus Carlsen isn't good at chess and only gets more unhinged from there. The Turing Award winner and his co-authors dropped a paper demanding the AI industry abandon its biggest obsession, AGI. Right now, everyone from Silicon Valley CEOs to politicians assumes AGI is the ultimate goal. A machine that can do everything a human can do. LeCun argues that this entire concept is a biological illusion. Humans do not possess "general" intelligence. We are highly specialized biological machines, tuned by evolution simply to survive in the physical world. We only think our intelligence is "general" because we are completely blind to the millions of cognitive tasks we are incapable of comprehending. Which brings us to the chess argument. Magnus Carlsen is the greatest human chess player in history. But compared to a modern computer? He is fundamentally terrible. Our belief that Carlsen is "good" at chess is pure human-centric bias. He isn't objectively good. He's just better than the rest of us, who are biologically awful at it. LeCun says we need to stop building AI to mimic human generality. Instead, he proposes a new North Star: SAI. Superhuman Adaptable Intelligence. Instead of trying to build a machine that mimics our flawed, biologically-limited brains, we need to embrace extreme specialization. SAI is about the speed of adaptation. It is an intelligence that can learn to exceed humans at any specific, economically important task. More importantly, it is designed to fill the vast skill gaps where humans are fundamentally incapable. Things like managing global energy grids in real-time. Or predicting complex molecular structures. The entire AI industry is obsessed with building a digital reflection in our own image. LeCun's paper is a brutal wake-up call.

We have automated 99% of the physical and intellectual labor that we did 250 years ago. No one hand copies documents anymore, no one does accounting by hand, no one reaps the wheat fields by hand. And yet, there is no unemployment as a result of this. I see no reason to believe that in the future, we will not automate 99% of the physical and intellectual labor we do now, but again, there is no reason to believe that that will result in any more unemployment than we have today. The limit to the number of jobs is not the amount of work we have to do. The limit to the amount of work we can get done is the number of minds and machines and hands we have available.

Thoughts on immigration: most Western countries are getting it wrong by being both too loose and too strict on immigration. Switzerland has been thriving by welcoming foreigners, who make up a third of its population, almost at the highest levels in the Western world, and now natives are pushing back in what promises to be a close vote to put up strict caps. The ideal policy is to welcome all highly skilled immigrants, whether they’re tech founders, doctors, nurses, engineers, scientists, renowned artists and writers, skilled construction, etc, after have a model on their net economic impact and some values guardrails around knowing and respecting the local language, customs, and laws (ideally with local sponsors). What doesn’t make sense is accepting older, non-skilled, non-working immigrants, and the large mass of unskilled asylum seekers (outside of extreme situations like war/state collapse and cultural affinity). That is, it makes sense for asylum for the EU to take Ukrainians, Colombia and Chile for Venezuelans, Turkey and the UAE for Iranians, and the UK and US for Hong Kong escapees. But the current asylum system is broken and will lead to a nasty backlash. As demographic decline and fiscal debt collapse become more apparent, more countries will start seeking young immigrants, but many are still clueless about the horrific math of this - smart immigration policy will become a necessity to survive, not something to shut down. WSJ: wsj.com/world/europe/switzer…

Everyone says the latest AI agents will be "job-ready" soon, especially after the release of Fable 5 this week. But is that really the case? Over the past many months, my group and collaborators have been building Agents' Last Exam (ALE), a benchmark designed to test exactly that claim on real digital labor-market work. My group and collaborators previously have created many of the benchmarks the field runs on, including MMLU, MATH, CyberGym, and ExploitGym. Today, I'm excited to share Agents' Last Exam (ALE): a rolling benchmark that measures whether AI agents can actually perform economically valuable work across a broad range of real-world domains. With ALE, we evaluated Fable 5, GPT-5.5, Composer 2.5, and other frontier agent systems across more than 1,500 expert-sourced tasks spanning 55 occupations. The result is both impressive and sobering. Today's agents can solve a meaningful fraction of professional tasks. But when we look at the hardest tasks, the ones requiring sustained reasoning, deep domain expertise, and reliable execution over long horizons, they are still far from human-level performance. On ALE's hardest tier, every frontier agent we tested, including Fable 5, achieved a 0% success rate. The age of useful agents is here. The age of truly job-ready agents is not. We hope Agents' Last Exam (ALE) will serve as a new guidepost and north star for developing agents capable of reliably performing economically valuable work across a broad range of domains. 🧵

Mostly correct take on companies avoiding pretraining, though larger enterprises and some savvy AI natives (Cursor, Fin/Intercom) can take base models from open labs and do their own mid-training and post-training, and often those models will give the best performance per cost. Maybe a few hundred enterprises today can do this? Most will just want to take an instruction-tuned checkpoint and SFT it, but end results are better when they start from a base model.

Insane take from Fable 5

Ok. Finished reading it. The story of Europe's failure in AI is turned into a gripping story (congratulations to the authors on finding this way to write it) and an outstanding SHOUT for action. I disagree with many things in this scenario (e.g. ASML cannot be used for leverage, I am afraid: all the EUV tech is San Diego-based (Cymer), and the chips are Nvidia, AMD, Intel, etc.) But the key insight is correct: (1) AI is THE critical technology of the future, and (2) Europe is falling badly behind on AI and running out of options. Both the economic and strategic consequences are brutal. We will write a reaction in Silicon Continent. In the meantime, please do read it. europe2031.ai/#timeline