fal head of engineering @isidentical says frontier AI is spending more compute on post-training than pre-training as the biggest gains move into specialization: "If you look at today's frontier models in LLM land, they spend so much more compute on post-training and RL compared to the pre-training phase. A pre-trained base model is like a Swiss knife. Good for many purposes but not the best in individual domains." "For image and video, there is a single aspect that's very verifiable, which is…