Comment by mordymoop

It almost seems like people don’t understand that most of the training of modern LLMs is now reinforcement learning on problem-solving tasks. They spend most of their training literally solving problems that they don’t already know how to do.