Anthropic researcher reveals self-improving AI that fixes misaligned behaviors. The automated system improved performance on all 10 misaligned benchmarks without degrading overall performance. It suggests a path toward automated alignment for AI developers.
Opening Kapyn…