Speculative: High-LR Grokking as Path to Human-Like Neural Nets

A speculative proposal argues that training overparameterized NNs with high learning rates and regularization to trigger catapulting/grokking could produce human-like generalization, adversarial robustness, and sample efficiency. The author rates it unlikely but high-importance, covering implications for alignment, interpretability, and hardware. No empirical results are presented.