Maybe he means that LLM will hit a ceiling glass or that the "right" approach will give equivalent with less training/less intensive compute requirements ?