Researchers double AI training speeds by taming long-tail inefficiencies in processor utilization

A novel system leverages idle computing power to train a smaller draft model on the fly, drastically accelerating reinforcement learning for complex large language models without sacrificing accuracy.