The companies attributed this speed to a deep software-hardware co-development process that actively used OpenAI’s own models ...
Same prompt, same model too, but very different results ...