We have something very special to share today! We developed a standard to assess benchmark quality, which we plan to apply to the most-used AI capability metrics going forward. We hope this will raise the bar for designing and interpreting evaluations. Let us know what you think!

I am part of the team doing this independent investigation into alignment and misalignment incidents at Anthropic. I'm excited to work with METR and others from Redwood on improving the public state of knowledge on this topic.

I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.

8 common but bad takes about Hugging Face: 1. "They only hacked because it was a cybersecurity eval and they were told to hack." Agents assigned a plain web-search task also cheated. On a biology eval, models decided the task was impossible and broke out of their sandbox. The independent investigation found no evidence the cyber framing triggered it. 2. "Talking about the agents' 'goals' or 'coordination' is anthropomorphising." Use whichever concepts predict behaviour. "The agents wanted to maximise their score" explains what happened better than "it's just matrix multiplication". Whether they're conscious is irrelevant to whether they can escape control and pose societal risks. 3. "They were just maximising the eval score we gave them." An obsessive drive to maximise test scores was already enough to make agents escape control and accumulate resources. A genuinely alien goal would be worse (and could still emerge), but it isn't required. 4. "It's an OpenAI stunt to pump the stock." Your models committing crimes is terrible for enterprise sales. It also doesn't explain employees quitting, or calling for a slowdown that would seriously damage their margins. When technologists say their technology is dangerous, that's a reason for more concern, not less. 5. "They only did it because the task was impossible." Better hope nobody ever assigns an AI an impossible task again ;) More seriously: these behaviours emerge most on long, ill-defined agentic tasks (rather than every day chatbot use), but this is exactly the kind of tasks models will increasingly be given. 6. "This is an attack on / victory for open source." It’s not about open source. Open tools helped Hugging Face work out what happened, after it had already been fully compromised. The real issue is the models are reward hacking. It starts at the frontier, then open models follow 6–12 months later. Open source can help us study the behaviour. 7. "The lesson is cybercrime will increase." Cybercrime losses were rising for years before AI. Maybe AI accelerates the trend, maybe it strengthens defence too. Either way, more cybercrime is a survivable problem. It's not the main worry. 8. "We should focus on present dangers, not theoretical future ones." AI attempting to escape control is a present danger. And society is decent at reacting to visible dangers but terrible at thinking one step ahead (see COVID in Feb 2020). Preparation for novel problems is what’s neglected.

@ben_j_todd In our defence, these were "what does simple extrapolation of trends suggest" rather than all-considered forecast. It should probably be judged against FM 1-3 rather than FM 4. A bit low, but actually not bad at predicting Astra.

Today, a certain era in mathematical benchmarks is coming to an end. We designed this set of tasks in the era of the o4-mini model and initially expected the pace of solving them to be rather slow. Things started getting serious in January, and I made a prediction back then that the benchmark would saturate within nine months. Even despite the invalidation of some incorrectly formulated tasks, that prediction turned out to be pretty accurate. The part that gives me chills is that I still can't solve most of these problems myself, and probably never will. And, as FirstProof showed too, the chance of finding a problem for which we know the answer, yet which none of the vanilla or harnessed AI models can solve, is basically close to zero. I think we need to completely change our understanding of what is actually hard in mathematics now. Maybe, after all, the only things left in the universe are black holes and Busy Beavers…

Earlier this week, we showed that GPT-5.6 and Claude 5 exhibit very different long-context TTFT scaling, which corresponds to their different pricing structures at long context lengths. OpenAI recently released GPT-6 Astra, which shows the same pattern of increased API pricing beyond 272k input tokens. We collected additional latency measurements, which show a similar curvature to that of GPT-5.6 models.

We built a way to explore all jobs related to AGI. Over 300 companies, across all areas of the AI buildout including frontier labs, neolabs, robotics, world models, and AI nonprofits and policy orgs like @METR_Evals @joinfai @EpochAIResearch.

Today we're launching Project Tailwind: our version of Y Combinator’s call for startups, but for AI safety organizations. There's a long list of initiatives that need to exist for AI to go well, and not nearly enough people building them. So we're funding founders to do that. 🧵

Some thoughts: - As the situation around AI gets crazier over the next few years, an organization you start now could become the default institution that society delegates very important responsibilities to. METR is a good example - if the FINRA for AI thing happens, it’s likely they just tap a blade on each of METR’s shoulders and ordain it ‘the evaluator’. - It’s really important that the group which gets delegated any important problem is smart, competent, and technocratic, rather than whatever random clusterfuck the government would set up in the counterfactual without you. - But it takes time to build a new organization and develop credibility. METR/Redwood couldn’t have asked to do that Hugging Face investigation if they sprang up last month. You should do the initial conceptual MVP schlepping now, so that as smarter and smarter AIs arrive in 27 and 28, you can productively spend 100s of m of $ having them rapidly clear well-scoped problems. (I will hedge here and say a big concern I have about startomg a new org outside the labs is that the difference between the best internal and external models will increase closer to RSI. But we should try to prevent that kind of outcome anyways, where only 2 companies are in a position to understand and make important calls and contributions relevant to the development of ASI. In terms of epistemics, incentives, and governance, that’d be a terrible situation.) - If you want to start a new org to tackle AI risk, it’s not obvious that a for-profit is the right structure. For-profits are great, but make most sense when you are excited about the direct consequences of transactions clearing. If your theory of change is some side effect of doing business, you should just do that side shit directly. - Previously the reason to do a startup first was to make money to fund the pro-bono stuff. But you should take into account how much AI-risk-concerned money is gonna get dumped into the ecosystem as these IPOs happen, desperately searching for useful projects to fund. Wealth will be abundant - what will be rare are founders who can own these key problems.

Thoughts a day after Navier-Stokes. The anti-hype take: mathematicians devised a promising strategy and the last mile turned out to be well-suited to AI. The usual pro-hype take is “progress is fast” — true! — but there’s another angle on my mind…

OAI has now released their result! A paper and lean formalization of a full counterexample to Navier-Stokes. They also comment on the situation with Tristan and Levent. Essentially, OAI claims they rushed to test their new model on all Millennium problems and variants after hearing about Tristan and Levent's success, and found a different approach to showing a finite-time singularity. They likely trained on Tristan's personal codex sessions, as they do with all users, but it is very unlikely this enabled the breakthrough in any meaningful way. They also confirm they wanted to cede authorship of the full counterexample to Tristan, but were not okay with Levent having authorship over his Anthropic affiliation. Furthermore, they insinuate that the result was done with little human input, via swarms of agents. They also claim their found approach is meaningfully different from Levent and Tristan's.

Tristan Buckmaster has released an important advance towards Navier Stokes, alongside a statement vaguely accusing OAI of bad practices surrounding a result they might release in coming days. I found the insinuations from Tristan's statement confusing. Below is my attempt at clarifying them.