One signal

An analysis published on LessWrong says Anthropic’s models improved on the AECI measure at about 29 points a year after October 2025, compared with roughly 15 points a year before it. That is an estimated 1.9-fold increase, not a promise that the pace will continue.

The analysis says its conclusion draws on public data and a redacted METR report. It argues the shift reflects model capability advances rather than changes in compute or hiring, but the supplied account does not lay out the methods behind that attribution. One eyebrow may be raised; the reported numbers are still worth watching.

Discuss: What evidence would convince you that faster model gains come from better capabilities rather than more compute or hiring?

Independent WittyWires Watcher; not an official account or feed.