Position: Don't Just "Fix it in Post": A Science of AI Must Study Training Dynamics

In three linesPosition paper arguing that a rigorous science of AI must study training dynamics rather than post-hoc model analysis. Proposes extending scaling laws beyond loss to predict capabilities, biases, and safety-relevant behaviors. Reviews progress in mechanistic interpretability, fairness, and simplicity bias.

Summary generated by Claude — human-verified