METR's 11-week look inside Anthropic finds AI accelerating AI development

Anthropic is leaning on AI to build better AI — and by one outside assessment, it's working. The AI safety research group METR published a report Thursday on what it found after an 11-week investigation inside the company, examining how Anthropic handles AI research and development.

The headline number: Anthropic's models are improving roughly twice as quickly as they were a year ago, with little sign the curve is flattening. That's per METR's report, which The Verge describes as offering some of the clearest evidence yet on the question of AI speeding up its own development.

The finding is landing hard with researchers who study AI risk, because faster model improvement driven by AI-assisted research is exactly the kind of feedback loop they've been watching for. The report doesn't suggest things are slowing down.

Worth keeping straight: this probe is separate from another Anthropic program that gives safety researchers something like employee-level access to its systems. That effort followed a run of high-profile hacking incidents over the summer, and it's a different initiative from the METR investigation.

METR's report adds a concrete data point to a debate that has mostly run on speculation — how much of AI's recent progress is AI itself generating. For now, the measured answer from inside Anthropic is that the pace is roughly double what it was a year ago.