The o1 model is evidence against the "bitter lesson". It's better at some things than others, and is not simply a scaling up of GPT-4. The right idea is the "boring lesson" - scaling up is good, but better algorithms are also good, and the right mix depends on the problem.