Summary
AI commentator Matthew Berman — typically among the more optimistic voices in the space — uses this video to share a candid and unusually anxious assessment of where artificial intelligence is heading. The immediate trigger: in a single week, an Anthropic researcher named Jacob Coxin publicly resigned, stating that neither Anthropic nor OpenAI is acting responsibly and that both are racing toward self-improving superintelligence. A separate Anthropic researcher followed with a public statement estimating greater than 10% probability that AI kills all humans within a decade. Berman reads both statements in full and contextualizes them against Anthropic’s founding mission under Dario Amodei.
The video’s analytical backbone is data from meter.org, which tracks how long AI models can sustain autonomous operation before failing. The progression Berman walks through is stark: GPT-3 managed 9 seconds in 2020; GPT-4 hit 4 minutes; OpenAI’s o1 reached 40 minutes; o3 extended that to 2 hours; GPT-5 to 3 hours 23 minutes; Claude Opus 4.5 to nearly 5 hours; Claude Opus 4.6 to 12 hours; and Claude Mythos to 16 hours — a curve that has gone effectively vertical. He layers this against AI release tracker data showing a dramatic acceleration in the number of new model releases per period across labs.
Berman’s central concern is recursive self-improvement — models autonomously improving themselves without human oversight — and whether the exponential nature of the curve is being adequately appreciated by the public or policymakers. He is careful to distinguish concern from doomerism, and explicitly avoids fully endorsing the claims he covers, but argues that statements from people actively building these systems demand serious consideration.
📺 Source: Matthew Berman · Published September 10, 2026
🏷️ Format: Opinion Editorial
