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.

from TechCrunch https://ift.tt/o579n4q