Summary
This video examines the latest developments around Anthropic after the company updated its S-1 filing to say that recursive self-improvement is approaching, and after METR released an independent audit focused on the automation of AI research.
The host walks through the audit’s findings and points to the heavy redactions in the report, including material on compute allocation, specific parts of the model training pipeline, and an unnamed agent scaffold used alongside the Mythos model. He also covers reactions from figures such as Daniel Kokotajlo and a response from an Anthropic co-founder, who said that sharing the details behind the reported speed-up could hand a road map to others.
The second half explains how Anthropic tracks progress on its internal benchmarks. Because models now score in the 90s on many tests, raw accuracy makes gains hard to read, so the approach uses log odds and the 50% success line, comparable to matching chess ratings. Viewers get an overview of why the redactions are controversial, what the audit may imply about AI development timelines, and how benchmark saturation is shaping the way labs measure capability. Much of the interpretation is the host’s own speculation, so it is best viewed as commentary on the public documents.
📺 Source: Wes Roth · Published November 06, 2026
🏷️ Format: News Analysis
