[bufferoverflow]

I thought GPT-4 was not trained on labeled data, but simply on a large volume of text / code. Most of it is publicly accessible: wikipedia, archives of scientific articles, books, github, plus probably purchased data from text-heavy sites like Reddit.

[frabcus]

Whatever they've built this year presumably uses all the positive/negative feedback on ChatGPT that they have a year worth of data now...

Another examples is the Be My Eyes data - presumably the vision part of GPT-4 was trained on the archive of data the blind assistance app has, and that could be an exclusive deal with OpenAI.

[enigmurl]

Assuming it's a reference to RLHF? Not sure

[lyu07282]

No it's reinforcement learning with human feedback, RLHF lots of labeling