The Information reports that OpenAI engineers developed an optimization that cut inference costs in half; it reduced the number of GPUs for logged out ChatGPT traffic to a couple hundred
Wonder how much of these optimisations were discovered by AI vs humans.