Naming what accelerated
When a laboratory’s AI capabilities improve faster, the next question is which work accelerated and by how much. The 1.9-times estimate in METR’s Anthropic assessment answers only part of that question: it compares growth in the internal AECI capability index with an earlier slope. It does not directly measure how many hours a researcher needs for the same experiment, or how many discoveries the laboratory produces. Changing the earlier trend in the denominator changes the acceleration implied by the same new score. A valuable feature of the report is its discussion of the choices built into this comparison.[1]
METR conducted an 11-week assessment involving about 15 interviews and access to internal data. Fitting a curve with one breakpoint to model scores places the change on October 7, 2025. This design shows that a steeper slope can describe the recent period. Finding a breakpoint algorithmically does not by itself identify what changed on that date. More AI assistance could coincide with changes in research organization or other technical advances. The association over time still leaves a causal share to explain.[1]
Uncertainty outside the interval
The reported 95 per cent confidence interval is 1.5–2.8 times. It comes from resampling scores and does not quantify every uncertainty about which systems enter the analysis or AI’s causal contribution. Seeing an interval can make it tempting to assume all doubts have been quantified. This calculation addresses variation within the selected data and measurement definition. How much a different research design would change the answer remains a separate question. The question an interval belongs to matters alongside its width.[1]
The more revealing test tries different ways of relating the same scores to capability. Alternatives that fit similarly well produce estimates between 1.1 and 2.9 times. This exposes the assumption that equal score increases correspond to equal advances in underlying capability. Such sensitivity does not require discarding the evidence of acceleration. It calls for more caution about its magnitude than score noise alone would suggest. If an acceleration estimate informs a safety decision, the shape of the score scale becomes one of that decision’s assumptions.[1]
A second comparison for causation
Considering interviews and internal reports together, METR judges AI to be a major contributor to acceleration. That assessment deserves attention: access to research work normally invisible from outside is valuable. But Anthropic redacted much of the evidence considered most informative. The gap between the public conclusion and the material readers can reassess is therefore substantial. The report also says fully autonomous contributions are not yet highly significant in the most important training work. Researchers using stronger assistance and self-sustaining autonomous research occupy different evidentiary stages here.[1]
A persuasive next step would follow comparable research work in which AI use changes while keeping the measurement scale fixed. Tracking task difficulty, compute budgets and team composition together could show which experiments finish sooner alongside the index slope. This is a design proposal I draw from the report, rather than a completed experiment. Differences between teams or selection of easier work would still need testing as alternative explanations. Such a comparison could connect the observational finding to a concrete research mechanism. For now, the strongest gain is keeping the magnitude of acceleration and attribution of its contribution as separate questions.[1]