Has AI Recursive Self-Improvement Started? METR's Anthropic Report

Matthew Berman's video on Anthropic's AI research asks whether recursive self-improvement has already begun. The underlying independent METR assessment does find a consequential change: Anthropic's progress on well-specified capability tasks appears to be accelerating, and METR judges AI assistance to be a major cause. But the report does not show an AI system autonomously designing and training its successor. The difference matters if you are trying to understand the evidence rather than predict a date for superintelligence.

Watch Matthew Berman's Analysis

Berman walks through the METR report and argues that it signals the start of an intelligence feedback loop. A JetBrains sponsor segment begins at about 05:34. His forecast and concern about a near-term "singularity" are commentary, not a finding demonstrated by METR.

Source Note: The Video, the Assessor and Anthropic

Checked 9 November 2026. We used the supplied transcript, Berman's video, METR's 5 November assessment and separate Risk Report review, plus Anthropic's own RSI explanation and evaluator-access update. We did not inspect Anthropic's private data, rerun AECI, interview staff or independently reproduce METR's attribution estimate. METR is an external assessor, but its report was produced under negotiated access and publication redactions.

What METR Found, Without the Leap to "Singularity"

METR defines a two-times acceleration as two years of progress at the earlier pace compressed into one year. The measured 1.9x is about the growth rate of AECI scores, not the speed of a single inference call or a promise that every future model will arrive twice as fast. The report's robustness checks also yielded a wider range depending on analysis choices. Read the measure and its limits together.

Three Different Meanings of "AI Improving AI"

1. AI-assisted research: people choose the goals and review results while models write code, run experiments and help analyze failures. This is the dominant mechanism METR describes at Anthropic today.
2. Compounding acceleration: a better model helps researchers develop the next one faster, which may improve the following model. METR sees a significant risk of this speeding up further, not a measured law of indefinite exponential growth.
3. Full recursive self-improvement: an AI system selects research directions, designs, trains and validates its own successor without relying on human judgment. Anthropic says it is not there yet; METR's report does not claim otherwise.

Berman's phrase "RSI already started" compresses the first two meanings into the third. It is fair to say that AI is participating in a feedback loop around AI development. It is not fair to treat that as proof of a self-running laboratory or a timed path to superintelligence. METR's separate review even treats a possible two-times R&D threshold as an early warning, not direct evidence that catastrophic risk has materialized.

How Independent Was This Assessment?

METR describes an eleven-week engagement, from 12 August to 29 October 2026, staffed by about 2.5 full-time equivalents. It could request a wide range of internal information, and says Anthropic shared more than originally scoped. But Anthropic could approve or deny requests, and METR did not receive unmediated staff-level access or sample non-public models. This is a different engagement from the employee-like evaluator arrangement Anthropic announced on 15 October.

Anthropic had the right to redact non-public information from the public report. METR says a majority of the evidence it found most informative was redacted, so outside readers cannot independently inspect much of the reasoning behind its strongest conclusions. The original scope also included verifying compute allocation between capabilities and safety, but Anthropic removed that component before METR reached a conclusion about allocation. These limits do not erase the assessment; they define how much confidence a reader can place in its public evidence.

What Anthropic's Own RSI Essay Adds

Anthropic's "When AI builds itself" describes Claude helping with code and well-specified research experiments while humans remain stronger at choosing which problems matter. It reports large increases in code authored and merged with Claude's help, but cautions that lines of code overstate true productivity because quantity is not quality. Its scenario for a model fully building its successor is a possible future, alongside slower or stalled paths and organizational bottlenecks.

That makes the practical question sharper than "is the singularity here?": how much valuable research can a human team verify when AI makes experiments cheap? Faster generation can move the bottleneck to experiment selection, independent validation, code review and safety oversight. Those constraints deserve measurement alongside capability charts.

A Better Way to Read the Next RSI Headline

Ask five questions before passing along a claim: What is the unit measured? Is the number a benchmark slope, individual productivity estimate or actual research output? Who set the goal? Did an AI choose the research direction or execute a human-defined task? What was controlled? Were compute, hiring and data changes separated from AI's contribution? What can outsiders inspect? Is the evidence public, redacted or based on interviews? What would falsify the forecast? Name the bottleneck or plateau that would change the story.

For this report, the strongest supported takeaway is already important: METR sees faster capability progress at Anthropic and judges AI use a major driver. The exact causal share, duration of the trend and timing of fully autonomous R&D remain open.

Primary Sources by Role

Video Waypoints

These are editorial waypoints drawn from the supplied transcript, not an official chapter list.

The result and what it measures

Implications and a sponsor break

Evidence and caveats

Closing interpretation