A 7% pretraining compute figure raises questions about individual AI models

A post asks whether the 7% figure describes pretraining compute for a finished model or a lab’s total training compute. Another commenter says the SemiAnalysis Tokenomics chart models compute across frontier labs, so repeated post-training runs can change the lab-wide share without showing how much reinforcement learning went into any particular model. The commenter also argues that compute spending and model-training FLOPs are not interchangeable.