[kevin_thibedeau]

It does subsume a corpus of factual information. I use it as a search tool for topics and relationships that traditional search engines can't handle. You just have to know that whatever it outputs isn't trustworthy and needs to be verified.

[jstarfish]

Part of the corpus is explicit bullshit though, and we don't know to what degree. It internalized conspiracy theory and WebMD alike. In a generative capacity, it only reliably produces fiction. Ever. Fictional stories often take place in realistic settings and reference real facts. They sound real. But they're still fictional compositions.

Using GPT as a reference to anything is the same as using a Michael Crichton novel as a biology reference. It looks right, but why would you waste your time asking questions of something you can't trust and have to double-check everything it says anyway? Nobody would keep an employee like that around, nor would you hang out with someone like that. It's friendly enough, but it's a pathological liar.

There's too much black magic going on inside the black box. We don't know how prompts get tampered with after submission, but it might be worth it to pepper "nonfiction" tokens in prompts to ensure it skews on the right side of things. It certainly responds to "fiction" when you're explicit about that.

[dekhn]

yes but it's still very much just a language model, not a knowledge model.