[marcosdumay]

> Most research converges to the idea that RL on synthetic data makes models worse, not better.

You are missing a mountain of nuance by generalizing the existence of a hole there.