AI is now speeding up Anthropic's own AI work, an evaluator estimates, but this is not yet runaway self-improvement

An independent evaluation firm says AI is helping speed up how Anthropic builds its own models. According to the short source note, the report comes from METR, which the note spells "Meter". It estimates that Anthropic is improving its models up to about 2x faster thanks to AI-driven self-improvement. The estimated range runs from 1.3x to 3x.

That range is the first thing to read carefully. At the low end, 1.3x is a meaningful productivity gain, roughly what a strong new tool gives a skilled team. At the high end, 3x would compress a year of progress into about four months. Those two outcomes mean very different things for the industry and for regulators. The note gives no detail on method, sample or time period, so we cannot say which end is more likely.

The headline calls this a "confirmation" of recursive self-improvement, and that is an overstatement. Recursive self-improvement, in its strict sense, means AI systems improving themselves in a loop that gets faster with little human direction. What the note describes is narrower: AI tools speeding up a research and engineering process that humans still run. It is an important signal. It is not proof of a self-sustaining loop.

Our reading: this is the metric to watch. Benchmarks tell us how capable a model is today. The pace of AI-assisted AI development tells us how fast that capability will grow. If frontier labs really are compounding their own progress, the gap between leaders and followers could widen. That fits our view that Anthropic and OpenAI currently lead. It also raises a governance question: oversight built around yearly model releases may fall behind a faster cycle. Rules should be based on measured acceleration, like this estimate, not on panic about a sudden "intelligence explosion".

The long-term case is still positive. Faster AI research can mean faster progress in drug discovery, biology and science in general, which supports the abundance we expect over time. In the short term, faster cycles also mean less time to test safety and less time for workers and institutions to adapt. We will treat this as a number to verify, ideally with the full report, before treating it as a turning point.