> AIs already improve themselves via training

Marginally. Model collapse is still a problem. Continuous learning is still a problem.

For AI to make a big leap we need a big break through.