Why does Mixed Precision Training matter for AI governance?

GPT-3, GPT-4, Llama, Mistral, and most modern LLMs are trained in BF16 with FP32 master weights. scale