OpenAI @yong_zhengxin predicts continual learning gets easier for newer models because interconnected knowledge lets them pick up new domains faster, like a physicist moving to biology: "If the model has seen a lot of related knowledge, it's much easier for the model to pick up new knowledge, because it's easier to connect what it has known before and how this new piece fits into the whole big picture." "We've already seen research papers showing that if the model has done mid-training on a…