Years of iterating against the same benchmarks should, by textbook logic, produce overfitting. It largely doesn't.
New research explains why: strategies that generalize can be expressed in too compact a form to allow memorization, while the ones that overfit don't survive a
