Study pretraining at scale, transfer across tasks, emergent capabilities, compute and data scaling, and the limitations of broadly reusable models.
