[Robotbeat]

Well sure, but there still aren’t any other LLMs at the level of GPT3/3.5 let alone GPT-4. GPT3.5 just using the API returns fantastic results even without the ChatGPT interface (which isn’t terribly hard to replicate, and others have using the API).

There are dozens if not hundreds of companies that could’ve done something profound like ChatGPT if they had full access to GPT3/3.5. And honestly, OpenAI stumbled a lot with ChatGPT losing history access, showing other users’ history… but that doesn’t matter much as the underlying technology is so profound. I think this really is a case of the under-the-hood capability (GPT3/3.5/4) mattering more than productization and execution.

(Now I think there are not a ton of companies that could do what Microsoft is trying to do by expanding GPT4 to power office productivity… that is a separate thing and probably only about 3 companies could do that, at best: Microsoft, Apple, and Google… and theoretically Meta but their lack of follow through with making Metaverse useful makes me doubt it.)

[mustacheemperor]

Hm, I wonder how much of the API’s performance is related to training/finetuning done by OpenAI planned towards the ChatGPT product. I think the RLHF is partly product design and partly engineering.