An Anthropic researcher just gave us a peek at self-improving AI
Given 10 benchmarks for specific misaligned behaviors, the automated systems were able to improve performance on every single one without degrading overall performance.
