12 months to the singularity

Summary

Matthew Berman walks through an October 2026 METR assessment of whether AI is accelerating Anthropic’s AI research and development. The report’s central conclusion is that Anthropic’s model capabilities on well-defined tasks may now be improving at roughly twice the rate observed from March 2024 to October 2025. METR emphasizes that the estimate is uncertain, with a plausible range of 1.3× to 3× the earlier rate, and estimates that AI may account for about 80% of the current acceleration.

The assessment attributes the change partly to newer models becoming substantially better at coding and supporting research workflows. Researchers reportedly use AI to write and check code, run experiments, and make more productive use of available compute. Interviewees described productivity gains of more than 2×, and some said that coding was no longer their main bottleneck.

The report also found signs that AI may be beginning to suggest research directions. However, it says that highly significant, fully autonomous contributions to the most important parts of model training have not yet been established.

Potential Feedback Loop

Berman highlights the possibility of a compounding feedback loop: more capable models could accelerate AI research, leading to stronger models being developed more quickly. He discusses METR’s view that the assessment found no clear near-term bottleneck that would prevent further acceleration.

The video also points to Anthropic’s internal capability measurements and benchmark changes as signs of rapid progress. The charts shown in the video are described as illustrative, rather than data from the report itself. According to the report, recent benchmark progress may be comparable to progress from October 2023 to October 2025.

Limitations and Caveats

METR’s access was not the separate employee-like evaluator access program Anthropic had discussed. Anthropic could approve or reject requests and controlled which findings were published.

The original scope included examining how Anthropic allocated compute, including between capability development and safety. Anthropic removed that work midway through the project, and the report says that parts of its analysis of potential bottlenecks were redacted. Berman also notes a tension between the report’s failure to identify clear compute bottlenecks and employee comments describing compute as a constraint.

Berman interprets the findings as concerning evidence of AI-driven acceleration. However, the report does not establish that a singularity or superintelligence will occur within a year.

The video also references Google’s AlphaEvolve work on improving hardware efficiency and a separate paper on “weight-space programming,” which Berman says he covered in another video.

Product Mention

A JetBrains sponsorship promotes JetBrains Air and IntelliJ IDEA for AI-assisted coding. The features described include:

• Running multiple agents and tracking their session status.
• Giving agents feedback on specific lines in code diffs.
• Using IntelliJ’s codebase knowledge to reduce incorrect agent actions and token use.
• Accessing debugging, testing, and profiling tools.

Guides and Tutorials

No step-by-step tutorial is presented. The video is a walkthrough and commentary on the METR report.

Main Speakers and Sources

• Matthew Berman — Narrator and analyst.
• METR — Source of the assessment discussed in the video.
• Anthropic employees and researchers — Anonymous interviewees quoted in the assessment.
• JetBrains — Sponsor mentioned in the video.