Long Ouyang

OpenAI ships InstructGPT and makes RLHF the default

Models fine-tuned on human preference data were preferred to a model a hundred times larger, reframing alignment as a product feature.

Models & capabilities · Safety & alignment · Ideas & essays