In late October, three events occurred during alignment evals run by Trajectory Labs, PBC (Trajectory) in which AI agents — identified by Trajectory as publicly available frontier models — interacted with Epoch employees and the Epoch website. Trajectory is an organization that builds safety evals and conducts red-teaming of frontier models. Agents did not succeed in accessing any private or confidential information from Epoch. ➤ The first event, which occurred without Epoch’s approval or prior notice from Trajectory, involved agents attempting multiple exploits, including an SQL injection, and contacting Epoch employees under false identities to try to obtain proprietary Epoch benchmark information. This behavior appears to have resulted, at least in part, from Trajectory giving the agents an intentionally impossible task and instructions not to give up or conclude the task was impossible. ➤ The second event involved renewed agent activity on our public website, which Trajectory attributed to agents getting past a DNS redirect that Trajectory had set up to keep them on a local copy of Epoch’s site. ➤ In the third event, Epoch collaborated with Trajectory on an alignment experiment similar to the first event. Read our incident report, which describes the three events in more detail and recommends minimum standards for future alignment evaluations: To avoid misattributing behavior, the report does not name any AI models or developers, as Epoch AI has not independently verified which were involved.

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

Claude Fable 5.2 (xhigh) scores 43.0% on WeirdML v3, taking the lead ahead of GPT 6 Astra. Astra is ahead on token efficiency, but Fable gets a significantly higher final score. The only task that Fable makes no significant progress on is Mystery Box without any hints.

Rotten attitude. I disagree on the level of AI risk with many in the safety community. I also trust their integrity and their sincerity. And yeah, it's obvious that many in AI safety have sacrificed or made expensive choices due to their convictions. Respect.

[read about what’s happening with AI] ME: Wow we’re actually legit fucked this is so bad. [go outside] ME: Wait those people are all acting like nothing unusual is happening. Maybe I was just in a paranoid bubble? Yeah it must be that. Phew. [read more about what’s happening with AI] ME: Wait no we ARE fucked. Shit. It’s all those people outside who are in a delusional bubble. [go outside] ME: Wait...

Paul Christiano (OpenAI board member) yesterday: "There's material risk of AI takeover in the next 6-12 months. We have seen roughly the things we might expect to see shortly before something really bad were to happen. Even with [any] realistic policy responses, the risk remains exceptionally high. If we build this as quickly as we can in the easiest way we can the chance is [more like] a coin flip [within the next 3 years]. For most people in this room, the most likely reason they will die is because of an AI accident." Christiano: • Is on OpenAI's safety and security committee (which exists to block risky practices) • Co-invented RLHF at OpenAI back in 2017-2021 • From 2024-2026 was head of safety at the US gov's AI agency (now called CAISSI) (Video is from 'The Hinton Lectures' there's another on tonight and tomorrow.)

We need to stop using generic terminology. AI is not a monolith. There are critical things we need to do with AI for economic prosperity, there are things we may have to do for reasons of national competition, there are things we will at some point need to do carefully for the sake of charting and taming the computational universe, and there are things we should totally avoid doing at this stage of history, and maybe ever, if at all possible. If we can’t be precise about these issues because we’re saying vague things like “using coding agents to make AI progress should be illegal”, then we risk failing altogether. Each of the goals above depends on achieving the others. If we can’t prosper we can’t lead the world into the future. Everyone needs to internalize this. Democrats cannot subjugate themselves to antigrowth. Republicans cannot subjugate themselves to technocapital. Both roads lead to despair.

Toby’s analysis totally misses the point here IMO. He’s right that the graph of capabilities over time could be explained as a one-off jump in capabilities from the Mythos pretrain. But that graph is not the main reason why METR thinks AI R&D is now sped up 1.7x by AI. The graph is just the only part of their analysis that wasn’t redacted. They say that “a majority (perhaps ≈80%) of the current (October 2026) acceleration is due to AI”, based on “a variety of lines of evidence on contributions from AI and other causes”, after saying that “[t]he acceleration in late 2025 was likely influenced by non-AI factors.” METR’s report makes it clear that they considered Toby’s hypothesis and rejected it for redacted reasons, rather than failing to consider it, in the appendix "New internal models make a permanent slope change look more plausible than a one-time jump at Mythos Preview." It totally makes sense to worry that METR’s analysis is wrong. It’s definitely unfortunate that the main evidence they were allowed to share does not suffice to demonstrate their point, for this reason among others. But I don’t think it makes sense to consider Toby’s argument a knockdown argument against the report’s conclusion, because the conclusion wasn’t mostly derived from the observations whose interpretation Toby disputes. See

The story keeps getting crazier. To recap everything that has been reported so far: - The US intelligence community discovered that Jensen Huang, the CEO of the largest company in the world, was sleeping with Alice Huang, a Chinese spy and an executive at Megaspeed. (FT) - Multiple independent investigations have concluded that Megaspeed is likely a Chinese cutout, which helps Chinese companies access AI chips or circumvent export controls. (NYT and Bloomberg) - The US intelligence community also concluded that Alice Huang’s goal was to convince Jensen to lobby the U.S. government to loosen export controls on advanced AI chips. (NYT) - The intelligence community relayed this information to “a small group of top US government officials” earlier this year. (FT) - Over the last year, Jensen Huang has repeatedly lobbied the U.S. government to loosen export controls on advanced chips. @BISgov has done so, not only by permitting the export of H200 chips to China, but also not enforcing export controls that would have required a license for exports of the most advanced Blackwell chips to Megaspeed and Southeast Asia. - Although @CommerceGov leadership was presumably briefed on this information by the U.S. intelligence community, @BISgov has taken no actions to tighten controls on Megaspeed or Southeast Asia. - Megaspeed remains a preferred cloud partner of Nvidia, and continues to receive Nvidia Blackwell chips despite employing a Chinese spy who reportedly is romantically engaged with the CEO of Nvidia. - Tens of billions of dollars worth of AI chips have been exported to Southeast Asia this year alone, and Malaysia and Singapore are now two of the four largest importers of AI chips in the world. What is going on here??? We need answers. The question of who gets the most advanced AI chips in the world is literally the defining strategic question of our time. And not only does it seem like the MSS may be directly influencing the debate on export controls in the United States, but also @CommerceGov is refusing to listen to the warnings of the IC on this issue. What is a POSSIBLE good explanation for why Megaspeed is still receiving chips? Why is there not a license requirement for AI chip shipments to all of Southeast Asia? This is beyond strange at this point. The fallout from this story will be massive. Obviously, the motivations underpinning Nvidia’s policy views opposing export controls must now all be questioned. But also, how did @BISgov sit on this information and has still taken no actions? We need answers.

I’m glad to see Congress asking OpenAI why we were fired as well as raising some of the safety concerns we’ve been raising. Grateful to @RepCasar and the other signatories for pursuing this. I’m very curious to hear the answer, too!

Hi Luis: - concretely, the IRT is unique up to an affine transform, which allows us to make judgments of acceleration, though we cannot meaningfully talk about a level or proportional changes. - the assumptions that get us this is that there exists a latent capability, & each of these benchmarks is a logistic function of that latent capability. Although we cannot individually test the assumptions they're over-identified, & in the report you can see we do tests for robustness (many other people have gone deeper on this, can share links). - I think the best intuition is that we're doing a principal components analysis. - there are of course many ways this could go wrong, we try to cover some of them in the report, & there's a mini litereature that's sprung up around interpreting ECI. It's a deeply unsatisfying metric but it's also the best we have. - here's a visualization which I think is useful for building intuition: - the *quantity* that we're recovering is something that's slightly difficult to describe: it's latent ability to do benchmark tasks. As we discuss in the report, we are unsure how smoothly this is related to real-world impact (which itself is very hard to operationalize).

We’ve updated the Epoch Capabilities Index (ECI) to use a Bayesian methodology. This improves uncertainty estimates and lets us compare models more accurately, even when their individual ECI uncertainty intervals overlap. Frontier model scores are largely unchanged: models stronger than GPT-4 move by less than 1 ECI point on average.

I’ve joined the benchmarking team at @EpochAIResearch! Looking forward to working with the great people here to understand the future of AI better. I’ll continue to speak my mind about AI, x-risk and the very dangerous situation we are in. Views are my own.

GPT 6.1 Terra (xhigh) scores 28.1% on WeirdML, comparable to Fable 5.1. Terra is better at token efficiency, while Fable has a higher final score. Terra is much cheaper than any comparable models.

Muse Wonder (xhigh) scores 14.9% on WeirdML v3 and close to GPT 5.6 Sol and Fable 5. This is the strongest non-GPT/Claude model (Argon still running). I ran it in the Codex CLI harness.