Check out this post from a round table dialogue with Greg Brockman from OpenAI. The GPT models that were in existence / in use in early 2023 were not the performance-degraded quantized versions that are in production now: https://www.reddit.com/r/mlscaling/comments/146rgq2/chatgpt_...