How a big shift in training LLMs led to a capability explosion Reinforcement learning, explained with a minimum of math and jargon.