GPT-4: Researchers have developed a novel optimizer called Sophia, which can train large language models (LLMs) twice as fast as the widely-used Adam optimizer. By using a lightweight estimate of the diagonal Hessian as a pre-conditioner, Sophia reduces the high up-front cost of training LLMs, potentially cutting budgets from $2M to $1M. The optimizer demonstrates consistent loss reduction across all parameter dimensions and is straightforward to implement with PyTorch. This development highlights the potential for academics to explore LLM pre-training and create effective algorithms with limited resources.
Read more at MarkTechPost…
