It learns from feedback, ‘blending’ it into its answers :) Based on their OPT-175B model, comparable in size to GPT-3