Training Dynamics: The Missing Foundation in AI Science

Biderman et al. (arXiv:2606.06533) argue that AI research must prioritize the study of time-evolving training processes over post-training analysis of fixed models, extending scaling law predictions from loss to capabilities and safety properties. This aligns with Kaplan et al. (arXiv:2001.08361), where power-law relationships emerged directly from optimization trajectories rather than endpoint evaluations. The position identifies a methodological gap: interventions like RLHF address symptoms after convergence without tracing their origins in data-objective interactions during gradient descent.

Mechanistic interpretability efforts, such as those in Olah et al. (Transformer Circuits, 2020 onward), have documented circuit formation mid-training but remain disconnected from predictive models of when specific features arise under varying hyperparameters. Fairness and memorization studies similarly focus on static audits, missing patterns like simplicity bias documented in early training phases across architectures. This systemic blind spot links disparate failures—emergent misalignment, robustness collapse, and capability jumps—into a single absence of dynamics-grounded theory.

Concrete open problems include deriving early-signal predictors for bias amplification and designing curricula that steer trajectories toward verifiable robustness, requirements drawn from philosophy of science criteria for progressive understanding as outlined in the paper.