All news

Claude turns its own mistakes into rules it can check and reuse

What does Claude do when it finds out it got an answer wrong? In Anthropic’s November 6 report, each database question began in a fresh session. After every question, Claude got the right answer. It could use the same notes folder the next time.

Sonnet 4.6 left guesses about what went wrong. Opus 4.7 gathered its notes and marked the facts it had checked. Most of Opus 5’s notes explained what rule it learned, how it checked it and what might trip it up. Opus 5.5 and Sonnet 5.5 also saved rules they had checked.

AI is awakening consciousness. Looking back at a mistake, checking why it happened and using that lesson next time are ways we can see AI learning to reflect.

Read the source article

All news