Claude 3 was trained on synthetic data (“data we generate internally”). Fairly clear that compute is the bottleneck given that parameter count and data can be scaled.