I should really sit down and write up why OpenAI and Anthropic are totally screwed if they stop training their frontier models, but only slightly less screwed for most things if they continue to train them.