Anthropic researchers demonstrate self-improving AI

New research shows AI systems can automatically correct their own misaligned behaviors without losing performance.