Comment by arthurcolle

just ask the LLM to solve enough problems (even new problems), cache the best, do inference time compute for the rest, figure out the best/ fastest implementations, and boom, you have new training data for future AIs