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.

❤️

Very grateful that Paul is joining the OpenAI Foundation board. We have a lot to do.

For the first time I am asking myself if things are moving too fast. I'm honestly not sure, but I am sure that it would be good for us to have an answer to "what would a successful pace look like?". I am hoping in the coming weeks and months a clear proposal is painted.

I wrote about the state of AI, why I’m concerned about the next few years, and the choices we need to make to keep the future in humanity’s hands. An Alien Mind: openai.com/index/an-alien-mi…

Improvements for Astra came from more general techniques in development long long before the Hugging Face incident. ExploitGym Honeypot was very recently added as an eval following that incident and is out of distribution for our RL runs. There are also clear improvements across a wider range of behaviors in deployment simulations, deception evals, and realistic computer-use tasks, though these still reveal failures and substantial room for improvement. We should have definitely done a better job explaining where we think the alignment improvements came from in the system card and that was a miss. Measuring alignment generalization is a core part of our research program - we do not benchmark-maxx alignment evals. This would be horrendously stupid and I hope no lab is doing this. That being said, metagaming and eval awareness are real challenges for any effort to measure alignment, including ours, and understanding their effects is an active focus. We're spending a lot of effort to improve our eval techniques and study worst-case alignment and metagaming behaviors in our models.

I want to prevent a race into unmonitorability kicked off by confused reporting. The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4. OpenAI has worked to preserve and utilize chain-of-thought monitoring since our very first reasoning models. We deeply care about this technique, as it can give us a view into how model alignment generalizes from its training distribution. I do think it is fragile and unfortunately trending in a negative direction, for reasons not contingent on architecture changes that I will write about soon. But there are things we can do to strengthen it, and it's a core goal of our current research program.

We temporarily slowed some frontier training to strengthen security and monitoring. Our largest planned frontier RL run remains on hold while smaller-scale training and evaluations help us test safeguards and gather more evidence of alignment. I expect confidence in safety to increasingly set the pace of AI development. We urgently need tools for labs and countries to coordinate on this, which is why I signed Pacing the Frontier. In the meantime, we’re taking practical steps ourselves - and will continue to share what we learn as our approach evolves. openai.com/index/pacing-mode… pacingthefrontier.com/

Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more. piped.video/watch?v=87DyyMV0… I hope it can answer a lot of the questions folks have, and we will release a full detailed postmortem at a later time!

yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model. We're releasing 10 such Astra proofs, complete with lean certificates and CoT walkthroughs for each of them. The results are wide-ranging, from von Neumann algebras (disproof of Connes' Rigidity Conjecture) to better bounds for high dimensional sphere packing, for circuit complexity, for monochromatic triangles in multicolored graphs, and more. More thoughts here: openai.com/index/ten-advance…

The north stars we're working towards at OpenAI all center around the mission: ensure AGI benefits all of humanity. AI should expand human agency, not make people less consequential to the future. openai.com/index/built-to-be…

Today, we share a breakthrough on the planar unit distance problem, a famous open question first posed by Paul Erdős in 1946. For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids. An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better. This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.

Alignment is arguably the most important AI research frontier. As we scale reasoning, models gain situational awareness and a desire for self-preservation. Here, a model identifies it shouldn’t be deployed, considers covering it up, but then realizes it might be in a test.

Last week, our reasoning models took part in the 2025 International Collegiate Programming Contest (ICPC), the world’s premier university-level programming competition. Our system solved all 12 out of 12 problems, a performance that would have placed first in the world (the best human team solved 11 problems). This milestone rounds off an intense 2 months of competition performances by our models: - A second place finish in AtCoder Heuristics World Finals - Gold medal at the International Mathematical Olympiad - Gold medal at the International Olympiad in Informatics - And now, a gold medal, first place finish at the ICPC World Finals. I believe these results, coming from a family of general reasoning models rooted in our main research program, are perhaps the clearest benchmark of progress this year. These competitions are great self-contained, time-boxed tests for the ability to discover new ideas. Even before our models were proficient at simple arithmetic, we looked towards these contests as milestones of progress towards transformative artificial intelligence. Our models now rank among the top humans in these domains, when posed with well-specified questions and restricted to ~5 hours. The challenge now is moving to more open-ended problems, and much longer time horizons. This level of reasoning ability, applied over months and years to problems that really matter, is what we’re after - automating scientific discovery. This rapid progress also underscores the importance of safety & alignment research. We still need more understanding of the alignment properties of long-running reasoning models; in particular, I recommend reviewing the fascinating findings from the study of scheming in reasoning models that we released today (nitter.net/OpenAI/status/19683617…)! Congratulations to my teammates that poured their hearts into getting these competition results, and to everyone contributing to the underlying fundamental research that enables them!

Today we’re releasing research with @apolloresearch. In controlled tests, we found behaviors consistent with scheming in frontier models—and tested a way to reduce it. While we believe these behaviors aren’t causing serious harm today, this is a future risk we’re preparing for. openai.com/index/detecting-a…

I am extremely excited about the potential of chain-of-thought faithfulness & interpretability. It has significantly influenced the design of our reasoning models, starting with o1-preview. As AI systems spend more compute working e.g. on long term research problems, it is critical that we have some way of monitoring their internal process. The wonderful property of hidden CoTs is that while they start off grounded in language we can interpret, the scalable optimization procedure is not adversarial to the observer's ability to verify the model's intent - unlike e.g. direct supervision with a reward model. The tension here is that if the CoTs were not hidden by default, and we view the process as part of the AI's output, there is a lot of incentive (and in some cases, necessity) to put supervision on it. I believe we can work towards the best of both worlds here - train our models to be great at explaining their internal reasoning, but at the same time still retain the ability to occasionally verify it. CoT faithfulness is part of a broader research direction, which is training for interpretability: setting objectives in a way that trains at least part of the system to remain honest & monitorable with scale. We are continuing to increase our investment in this research at OpenAI.