Berman opens by saying a third-party evaluator, METR (transcribed as "Meter"), has found evidence that recursive self-improvement has begun at Anthropic. He sets this against earlier pledges by leading labs to give outside evaluators employee-like access, and stresses that this engagement was not part of that programme: Anthropic could approve or reject requests and decide what was published.

He then summarises the report's headline estimate. According to the video, most METR staff believe Anthropic is improving model capabilities on well-specified tasks at around twice its March 2024 to October 2025 rate, with the true figure possibly nearer 1.3x or 3x. METR reportedly attributes perhaps about 80% of the current acceleration to AI, whereas the late-2025 jump around Opus 4.5 was more likely down to non-AI factors. Berman walks through charts of Anthropic's internal capability measure, which show an inflection around Opus 4.5 and a steepening curve since.

The middle of the video covers the evidence. Berman cites METR's finding that training compute scale does not explain the acceleration, though details were redacted. He also cites interviews and self-reported uplift, headcount, lines of code and experiment usage, plus anonymised researcher quotes from METR's own interview table endorsing speed-ups of two times or more. He links this to compute-efficiency figures from Anthropic's S-1 filing and to Google's AlphaEvolve work. METR reportedly saw no clear bottlenecks within twelve months.

Berman closes on the limits of access: METR's compute-allocation verification was dropped midway at Anthropic's request. He concludes the findings are concerning, and points to a separate weight-space programming paper as a sign AI may become harder to understand.

Key takeaways

  • Per the video, METR estimates Anthropic's capability progress on well-specified tasks runs at roughly twice its earlier rate, with a plausible range of 1.3x to 3x.
  • METR reportedly attributes around 80% of the current acceleration to AI, and says it found convincing evidence that training compute scale was not the cause.
  • The engagement was small, about two and a half full-time-equivalent staff, and Anthropic could refuse access and redact what was published.
  • Anthropic withdrew METR's planned compute-allocation verification partway through, which Berman flags as notable ahead of its IPO.
  • Researchers interviewed reportedly say AI is not yet making fully autonomous contributions to the most important training work, but shows early signs it could.
  • The video includes a JetBrains sponsorship segment for JetBrains Air and IntelliJ IDEA.

Our take

This is a single creator's reading of a single report, and several of the strongest claims, including the 80% attribution and the 'no bottlenecks' conclusion, are METR's own judgements on evidence Anthropic partly redacted. The video itself concedes the evaluator had limited access, the interview data is casual and self-reported, and one researcher quote about compute contradicts METR's finding. Berman's framing of a coming singularity is his interpretation, not something the report is shown to state. Readers should check the METR post itself before drawing conclusions.

None of the items in our current coverage shortlist bears directly on this subject, so there is nothing to compare the video against.