#reactive optimizers

Claude Rewrote the Optimizer—and Beat Overfitting

A training run can look healthy right up until a model starts memorizing the wrong answers. The surprising twist is that the agent’s first fixes failed—and the winning idea wasn’t new.