It's not time to slow down but to accelerate! The recent AI-powered cyberattacks have everyone talking about the risks of AI. We should. But let's not lose sight of the bigger picture! If we work hard at it, AI will make the world safer, not less safe, just as most major technologies have. We've already seen a glimpse of that: we defended ourselves with AI (more specifically an open model). The same systems that helped stop an AI-powered cyberattack can now help defend against millions of cyberattacks every day, while helping us identify and fix vulnerabilities before attackers exploit them. To get there, we need three things in my opinion: - Increase transparency with mandatory trace sharing and incident disclosure for agent cyber-attacks - Keep AI-powered cyberattacks illegal, with meaningful penalties to disincentivize them - Equip defenders with the best AI, especially open models, to reduce the asymmetry of capabilities between attackers and defenders If we get those three things right, AI won't just create new cybersecurity challenges, it will make cybersecurity fundamentally and meaningfully stronger. And that's before considering AI's impact on science, healthcare, education, productivity, and much more. It's not time to slow down but to accelerate!

maybe i am lucky but across my work (ai+bio research) and relevant hobbies (music, motorsports) there have been 0 model releases where my reaction has been “oh no, the machine is somehow reducing my enjoyment of my craft.”

OPENAI IS GOING TO TAKE THIS ENTIRE MARKET DOWN WITH IT And you don't have to own a single share to get hurt. What I'm about to explain should worry anybody who thinks they're diversified: OpenAI is a LOAD-BEARING company. Pull it out and the whole structure comes down. They spent $17.2 billion on Microsoft Azure in calendar 2025, which is 69% of Microsoft's entire year-over-year growth. Take that one customer out and Azure grew 8%, which barely beats inflation. Now look at what Microsoft just reported: The backlog everyone is celebrating came in at $678 billion, up 84%, and the stock ripped. Sounds fantastic until you realize that when you exclude OpenAI the backlog grew only 25%. Back in January, when that number was $625 billion, roughly $281 billion of it was owed by one private company nobody can audit. And Oracle is in even WORSE shape. Something like $300 billion of its backlog, more than half, rides on the same counterparty. So you think you own Microsoft, Oracle, Amazon, CoreWeave, Nvidia and SoftBank? What you actually own is the same trade 6 different ways, and every leg of it runs back to one company that burns cash and still cannot go public. The people with real money are already backing out. Julien Garran pointed out that Masayoshi Son could not get a $10 billion bridge loan against his own OpenAI shares. Think about that for a second, because nobody says no to that man. Blue Owl walked away from a $10 billion Oracle financing. Three months ago the banks were dancing near the door and now they are walking through it. Then there is Julien's depreciation work, which makes this even worse: Run the capex schedule out and hyperscaler net income falls 98% by 2033. To break even they would need to build 20 killer apps inside 6 years. Another Google Search. Another YouTube. Another Office. They have not built ONE. If OpenAI cannot go public in the next 8 months, they are dead. Whenever you see hubris and debt in the same room, run, don't walk.

Reactions to this tweet reveal the lack of understanding of sources of American economic dynamism among the political right. Saying “we’ve always had the best American workers before” ignores how many of our top firms came from immigrants. That’s always been our advantage.

New newsletter: THE FOUR HORSEMEN OF THE AI BUBBLE APOCALYPSE I am neither anti-AI nor certain that AI is a bubble. But the last four weeks have made clear that the AI buildout now faces a very clear quadruple-headed risk hydra. 1. A spending risk, as the hyperscalers run low on cash and take on $170b in annual debt—which is more than the projected UK deficit. 2. A revenue risk, as open-weight models threaten to compress the margins of frontier labs ... and as AI become the sort of internationally competitive asset-heavy industry requiring stable and determined long-term policy consistency, which is arguably China's competitive advantage. 3. A political risk, as anti-AI populism becomes one of the easiest applause lines, even as AI becomes a more and more foundational pillar of US economic growth 4. A technological risk, as the frontier labs bear down on RSI, which I think could significantly change the basic business model of the labs, as compute costs rise and rise for a set of super-advanced models that are fit for, and affordable to, a small minority of users (in, eg, cyber security) In one sentence: The capabilities of AI are becoming more powerful, while some economic underpinnings of the AI buildout—and, as we’ll discuss, the political support for AI—are becoming more vulnerable. Today's (long, 5k word) piece deeply considers each risk and also—because over-confidence in this space is typically a sign that you're not thinking hard enough—I offer the strongest reason to think each risk might be overblown.

Can AI agents conduct open-ended AI research? Most evaluations of agents conducting AI research focus on narrow, verifiable tasks. But AI research is often open ended. Researchers pick hypotheses, decide what evidence is appropriate, and recognize a failing approach. We gave agents research questions from two unpublished papers, six days, and thousands of dollars of API credits and compute. The authors of the original papers then reviewed the AI-generated papers. They unambiguously rejected agents' outputs. We call these "shadow evaluations", since the agents are shadowing the original research effort by the authors. Agents were fluent at most *engineering* tasks They conducted serious literature reviews, debugged GPU environments, ran hundreds of experiments, and turned in camera-ready LaTeX without human help. We also found no evidence of reward hacking. If anything, we found the opposite: the agents started with marketable claims and walked them back to negative results as the evidence came in. Neither agent output was close to the bar of a top conference paper Both papers suffered from similar failures: poor judgment about the bar for an AI paper submitted to a top conference, the lack of creative problem solving and ineffective backtracking, poor awareness of resources, and instruction drift. 1) Lack of judgment about the bar for a top conference. The agents had a poor model of the bar for an AI paper submitted to a top conference. We allowed agents to self review their papers. Despite the poor paper quality, their reviews predominantly labeled the papers "weak rejects". 2) Lack of creative problem-solving to address feedback. When they received negative reviews, the agents typically narrowed their hypothesis and claims, rather than working out creative ways to address these concerns. 3) Ineffective backtracking. The agents dropped their most ambitious hypotheses within the first fifteen hours of carrying out the experiment and never changed course afterwards. 4) Poor resource awareness. Both runs ended with over half the API budget unspent. One agent declared itself done seven hours before the deadline, right after its own self-reviewer returned another reject. 5) Instruction drift. They did not follow explicit instructions on minimum exploration time, incorporating feedback for reviews, and on paper length (the outputs exceeded the page limits in both cases). This research design has many limitations Limitations include the small sample size, non-blind reviews, and the reviewers knowing that the work was AI-generated. We also couldn't test Anthropic's strongest model, because Fable 5 is deliberately limited on frontier AI research tasks, so ended up using OpenClaw with Opus 4.8 (extra-high) for our main experiments and Codex with Sol 5.6 (ultra) for a robustness check. But we think the research design is still helpful in assessing AI agents' ability to conduct research, and it is complementary to evaluations on verifiable tasks, as well as blinded reviews of AI outputs. Our results show early evidence that even though agents are proficient on verifiable research tasks, they do not make genuine progress on open-ended ones. It is worth understanding if this is a fundamental limit, or if better models, scaffolds, and more compute could help close it. As the evidence for the gap between open-ended and verifiable tasks firms up, it is also worth understanding how much progress in AI depends on open-ended research rather than hill-climbing on well-specified objectives. In follow-up studies, we are expanding the set of non-public papers we evaluate. If you are an AI researcher with unpublished papers, we would love to collaborate with you on our next shadow evaluation. Expression of interest: Conducting shadow evaluations involves a lot of researcher degrees of freedom. In many places, our coauthors disagreed with our interpretation of the findings, and we have surfaced those disagreements in the paper. (This is one reason why having a group of coauthors with different priors is important for open-ended research.) We also release the agent logs, one of the AI-generated papers (the other original paper is still not public), and all the code and data, so that others can conduct their own analyses of our results: Finally, we plan to conduct shadow evaluations regularly, and are hiring a senior researcher to help lead these efforts. Apply here: I'm grateful for the core team leading this effort: @PKirgis, Andrew Schwartz, @steverab, and @random_walker, and to our collaborators who reviewed AI papers, analyzed agents logs, and gave feedback on the paper: @DavidDAfrica, @KozzyVoudouris, Viet Nguyen, Toby Pilditch, @DubMagda, @HarryCoppock, @CUdudec, @nityndg, Matilda Orona, @tilmanbayer, Derrick Chan-Sew, Yue Ling, Abhishek Shetty, @hlntnr, @ghadfield, @sethlazar, @snewmanpv, @shostekofsky, @RishiBommasani

The midterms are this simple: Biden and Democrats inherited a poorly managed pandemic, massive job loss, and a ruined economy from Trump in 2021. They added jobs every month, the pandemic recovery went better than in peer nations, and by 2024 the US economy was dubbed the Envy of the World and was so strong it improved the entire global outlook. That was what Republicans and Trump inherited. Then, they slammed the brakes on the economy, killed job growth, drove up inflation, made healthcare far more expensive, kicked one million children off of food assistance, started an illegal war, angered our allies, and made the nation sick from cuts to health inspections and disease monitoring, not to mention the rampant corruption that has enriched the president and his family with billions of dollars. The midterms are your chance to change this. It is time to vote those who have failed us out of power.

It really annoys me how they report this stuff like they’ve gone on safari to do scientific observation Claude is a computer program that they created These are simply their own security errors being reported as if they are scientific achievements It’s bad enough that they don’t seem to think they need to be responsible for their own mistakes But the worst part of it is they try to use their own mistakes as an excuse to control the behavior of other people There is an amazing arrogance to the whole thing, like they are completely above reproach and cannot possibly be wrong about anything

The amount of basic research done in industry is nowhere near the amount done in universities. At various times, there have been industry labs that had fundamental research activities and have made important scientific contributions. Examples in information technology include Bell Labs, IBM Research, Xerox PARC, GE, Phillips, NEC, and several other. That disappeared in the 1990s. Microsoft Research picked up the torch in the 2000s, followed (to some extent) by Google and then Meta (for about a decade until recently). But their innovations almost always built on top of academic work, and certainly profited from the whole research ecosystem.

Trump is still demanding the Senate delay its recess to pass his so-called "SAVE AMERICA Act." He calls it election integrity. But it’s purely a corrupt Voter Supression "Save Trump’s Ass Act." It strips voting rights from: - Nearly 69 million married women whose birth certificates no longer match their legal names. - Elderly voters. - Disabled voters. - Poor voters w/o cars or bus money - Remote and rural workers who rely on mail or online registration. All of them would face new voting barriers under this bill. Republican John Cornyn already admitted the votes aren’t there. Trump keeps pushing it anyway — because free voting is a threat to his power. If everyone votes, he loses. This isn’t about stopping fraud. It’s about making sure fewer of the "wrong people" can vote - "wrong" being people who will vote against him.

PCE inflation last month was 3.7% — still almost double the Fed target — in stark contrast to Trump's promises. It was down from May given the brief lapse in the Iran War, but it will likely bounce back up in the next reading after Trump shredded his proposed peace deal.

We got attacked by secret unreleased proprietary models and defended ourselves with an open model, more precisely the @nvidia quantized version of GLM 5.2 coming from @Zai_org. Banning any open model would hurt first cyber security defenders, startups, small companies, researchers and everyone who's not a frontier lab and need on-prem affordable controlable models to compete and protect themselves. Let's not do that!

Our labs keep trying to spin this into a push for broader regulation. It can and should backfire. The models are not "going rogue" or acting of their own accord, like they're some Marvel movie evil robot. People made bad harnesses, told them to hack things and had transparently and objectively bad dev-ops and security practices. The fact that folks are trying to spin this into a "we need help from the government to regulate everyone" instead of "we should be punished in a narrow way on these specific incidents under existing law" is the real disconnect right now.

1/ SSL models like DINOv2 are engineering marvels, but they rely on fragile hacks like EMA, stop-gradients, and custom centering to avoid representation collapse. What if we could replace all these heuristics with pure, mathematically sound optimal transport? 🧵

The narrative: Blame the Agent, instead of the Agency that told him to “apply all your powers and told to achieve this win.” Ironically, many forefront members of the AI-Safety community, in their fervor for proof that AI (the weights) have jumped the fence, are actively helping to absolve companies of the responsibility of building systems with guards and constraints OUTSIDE of models to control agentic behavior. They are too eager to exclaim, “I was right with my worries and here is proof of it!” that they are working against what should be their own #1# objective: requiring people who build these systems to do so with safety constraints, around and outside AI models—and holding them accountable when they do not act responsibly. We are in the weird pre-seat-belt moment when safety advocates are arguing that momentum is dangerous instead of holding auto makers accountable for adding harnesses with restraint.

The amount of basic research done in industry is nowhere near the amount done in universities. At various times, there have been industry labs that had fundamental research activities and have made important scientific contributions. Examples in information technology include Bell Labs, IBM Research, Xerox PARC, GE, Phillips, NEC, and several other. That disappeared in the 1990s. Microsoft Research picked up the torch in the 2000s, followed (to some extent) by Google and then Meta (for about a decade until recently). But their innovations almost always built on top of academic work, and certainly profited from the whole research ecosystem.