Researchers identify third pretraining axis for LLM efficiency

A study from Harvard and UIUC researchers claims a new pretraining axis improves sample efficiency by 6.2x and accelerates generative AI inference by 250x.