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