At L2 on the progress towards RSI (Recursive Self-Improvement ), the AI starts deciding how to improve itself based on failures and feedback, instead of humans choosing every change.
Humans still define the goal and decide what counts as a successful improvement.
