Claude 3.5 Sonnet improved its HumanEval pass rate by 30% via self-play on its own outputs?

Claude 3.5 Sonnet improved its HumanEval pass rate by 30% via self-play on its own outputs.

The constitutional self-play loop ensures alignment updates happen alongside capability improvements?

The constitutional self-play loop ensures alignment updates happen alongside capability improvements.

Compute cost for self-improvement was only 10% of the original training budget—opening a scalable path.

AI Safety

LessWrong AI June 05, 2026

⚡Anthropic's self-improving AI gains 30% coding accuracy autonomously...

Deep Dive

Anthropic just published a linkpost examining recursive self-improvement.

Key Points

Claude 3.5 Sonnet improved its HumanEval pass rate by 30% via self-play on its own outputs.
The constitutional self-play loop ensures alignment updates happen alongside capability improvements.
Compute cost for self-improvement was only 10% of the original training budget—opening a scalable path.

Self-improving AI could reduce human oversight needs but demands robust guardrails to prevent runaway capability growth.