Ajeya Cotra: METR report shows Anthropic has already crossed its RSP 'dramatic AI acceleration' threshold
- METR's deep-access assessment of Anthropic finds that capabilities progress from October 2025 to October 2026 ran roughly twice as fast as the previous twelve months on an internal AECI index, an internal version of the Epoch Capabilities Index; the report authors redacted what they called the majority of the evidence they found most informative.
- Anthropic's Responsible Scaling Policy defines the dramatic acceleration threshold as capabilities progress running twice the fastest pre-AI-saturated sustained rate and being plausibly attributable to AI automating AI R&D rather than more compute, and its S1 said the company may imminently meet it; Cotra reads the METR report as showing the threshold is already crossed.
- METR concludes AI automation is probably the large majority of the reason progress is still running at that faster pace. Cotra calls the report the most important update on AI risk in the last couple of years, bigger than the Hugging Face attack.
- Cotra's argument: if AI roughly doubles capabilities progress, it must speed up human researchers by much more than that, because overall capability (effective compute) is the product of algorithmic compute multipliers and training compute.
- She notes the naive extrapolation points to much shorter timelines than most external researchers, including the AI 2027 team, have projected.
Hacker News 의견들
Not trying to discredit the premise, but how much of this is vertical scaling in the weights versus just throwing capital at training and running agent swarms in parallel? Anecdotally most of the uplift comes from models exploring more of the search space now that we can hand them embarrassingly parallel problems and infinite token spend.
I'm not sure what you're actually asking for here. The article says agent swarms and token spend count as AI automating AI, and the RSP attribution condition specifically excludes the capital-for-training concern.
Intelligence per dollar is clearly improving. Opus 4 was $15/$75 per million input/output tokens at launch, Opus 5.5 is significantly better at $4/$15. And the search space doubles with every extra bit, so dumb brute-force enumeration is prohibitively expensive; pruning it well is not a trivial skill.
Cotra didn't have direct access to Anthropic's infrastructure, she's citing colleagues who did. That METR report says Anthropic made roughly two years of historical progress over the last year and that the acceleration wasn't from faster training compute scaling. It's also heavily censored, and the authors themselves say the public has to rely on METR's judgment much more than would be ideal.
Or Anthropic just got a pile of new user data when they launched Fable without ZDR. That would explain a chunk of the Opus 5.5 jump.
That doesn't line up with their claimed divergence point, which was pre-Mythos.
Could be a mix, but I thought it was mainly RL. Maybe the first distillation pass from Fable into Opus gave it a push too.
Titles like this irritate me. Every AI progress piece says the machines taking jobs are getting better even faster, isn't that incredible, and then it's riskier than ever to deploy them at scale, and then they go public, take institutional money from retirement funds and do it anyway.
You're mixing up who is who. The author works for METR, not Anthropic, and her job is independent oversight of the labs. She isn't in the incredible camp, she's the one asking for slower development and deployment.
Anthropic shouldn't be self-regulating in the first place. That's the government's job, and the current US government isn't going to do it.
The same report also found AI couldn't be deployed outside a narrow domain of verifiable rewards. The productivity measures are opaque and don't show up in economic output, and Anthropic has seen its valuation corrected by 50% or more. All of this sounds fishy.
METR always says what Big AI wants to hear. They're not independent in any way.
In 2025 they published a study finding AI slowed experienced developers down by 19%. Doesn't sound like shilling to me.
Also confounding: more domain experts and talent are joining AI research because the problems got more interesting and the models got more useful.
Look at NeurIPS and ICLR submission counts. With a frontier model I go from an idea to a working implementation on toy problems in minutes instead of weeks.
Trying out toy problems and making progress toward AGI are entirely different categories. The submission growth is mostly slop that gets desk rejected, and reviewers say the number of good papers hasn't gone up at all.
Recent NeurIPS and ICLR papers are the worst possible example. A lot of great papers had AI involved in accelerating them, but the majority of submissions are meaninglessness, sometimes straight autoresearch slop.
It already helps me finetune small image models. I don't tell it what to do, I ask for a report, read it, and decide what to implement by feel. The limit is hardware and tokens, which a researcher without either is not the case they're referring to.