Reasoning models don’t scale linearly with thinking budget, and the debate over fair use in training data is back in the spotlight.