Epoch AI: Less Than 30% of R&D Compute Goes to Final Training Runs

New analysis reveals that the headline-grabbing training runs behind frontier models represent a fraction of the compute labs actually spend — most goes to iteration, experimentation, and evaluation.