The dirty little secret of Large Language Models is how many humans are in the loop.[0][1]

Transformers are great at building extremely complex maps of language without any human interventions, but if you want them to consistently query the right part of the map (e.g., His Codepen search example), you need a very non-trivial amount of human feedback.

Will be interesting to see if all this hype leads to a solution that scales better than what we do now (so that orgs. actually could have insanely good AI chatbots trained on their docs), but the jury is still definitely out on that.

[0]https://openai.com/research/learning-from-human-preferences [1] https://openai.com/research/instruction-following