“Unlocking the ability for models to use millions of tokens for test-time compute could be equivalent to a multiple order-of-magnitude increase in pretraining compute.”