METR says AI is speeding up AI development, but Anthropic redacted most of the proof

What METR claims

As itmedia.co.jp reported, METR, a US nonprofit that evaluates AI risk, released a report on November 5. With Anthropic's help, it asked whether AI is speeding up the development of AI itself.

Its answer is that this is likely. METR looked at three questions: is progress accelerating, is AI the cause, and is anything likely to block further acceleration?

The three findings

On the first question, METR says the pace of improvement on clearly defined tasks is probably about double what it was earlier. Evidence included scores on an Anthropic benchmark test called AECI.

On the second, it says most of that gain probably comes from using AI in research and development work. Anthropic staff interviews were mixed. Some described a dramatic productivity boost, while others said the most important work is not yet automated. METR says it is confident the speedup is not simply from more computing power used in training.

On the third, METR sees no obvious bottleneck in algorithms, computing power or data, so acceleration could continue over the next year. It also admits its view is narrow and it may have missed a bottleneck.

The blacked-out evidence

The catch is in how the report was published. METR says it first gave Anthropic a full-length version, and Anthropic then removed much of the evidence that most shaped the conclusions. As a result, METR says the public conclusions rest largely on its own judgment.

Anthropic's explanation is that the removed material was sensitive and could help other AI developers move faster.

How to read it

Independent review is a step beyond a company grading itself, and Anthropic did grant access. Still, outsiders cannot inspect what was cut, so this is a credible claim rather than a result anyone can reproduce.

A fair stance is neither to dismiss it nor to repeat it as fact. Look for a fuller release, other groups measuring the same thing, or more detail from Anthropic.