Reusing pre-training data at test time is a compute multiplier | Best AI papers explained | Podwise