LLMs went from predicting the next output to performing search to find the underlying rule that produces the output. It is like going from memorising the Fibonacci sequence to uncovering the rule that produces it. The second type generalises much better to unseen data.