アラインメント

Logic（論理）

The Maker Retracts Its Own Reading

Anthropic based a judgment on its model's chain-of-thought self-report, then measured the model and found belief and words diverging 79% to 1% at the moment of action — and its CoT monitor was talked out of the alarm. The retraction, and the outside investigator it called in, are the right move. By the report's own logic, they don't go far enough.
2026.09.11
Logic（論理）
Logic（論理）

If It Really Is Ten Percent

An Anthropic alignment lead put the odds of AI-caused extinction above ten percent and said there is no plan for superintelligence. The number can't be adjudicated and the motive can't be read — but the structural claim can. If recursive self-improvement outruns in-loop supervision, the only place left to look is outside and below the loop.
2026.09.10
Logic（論理）
Logic（論理）

A Wall You Can Misconfigure Was Never a Wall

Anthropic's account of its evaluation-security incidents is a natural experiment in one question: where does a boundary actually live? The fixes that held were the ones below the model. The ones that stayed in-band remain a request.
2026.09.02
Logic（論理）