METR finds Anthropic's models are speeding up AI development

METR, an AI safety research organisation, published a report on Thursday on Anthropic's use of AI to build more capable AI, following an 11-week investigation. Its central finding is a rate: the company's models are improving roughly twice as fast as the pace METR recorded previously, and the report sees little sign of that slowing.

Recursive improvement moves from argument to measurement

The idea that AI systems could speed up their own development has mostly been argued in the abstract. A measured rate changes the terms of that argument: a curve that doubles is evidence about capacity rather than a thought experiment. It also makes the internal division of labour a fair question for outsiders — which parts of research now go to the models, and which stay with people.

For safety researchers, the concern is arithmetic rather than alarmism. Evaluation, red-teaming and oversight tend to run on fixed schedules and fixed headcounts, while a capability curve that doubles compresses the interval between one measured state and the next. A method that was adequate at one reading can be behind by the following one.

What the report establishes is a window, not a forecast. Eleven weeks of measurement at one company cannot say whether the rate holds, whether it generalises, or what happens when the same approach is applied elsewhere. It does move the question from whether AI can meaningfully accelerate AI research to how fast, and whether the institutions doing the checking can keep their footing.