Aligning language models to follow instructions

【RSS 原文 · 未经 AI 筛选或翻译】We’ve trained language models that are much better at following user intentions than GPT-3 while also making them more truthful and less toxic, using techniques developed through our alignment research. These InstructGPT models, which are trained with humans in the loop, are now deployed as the default language models on our API.