I think we might not entirely disagree here. I think it’s sort of confusing to say “Gemini is frustrated” rather than “Gemini thinks its in a situation where it’s supposed to be frustrated so that’s what it predicts”, but I’m not sure if that’s a real disagreement or just that I think my framing is more helpful.
The main thing I’m trying to argue against in the post is less about personas/faces and more about the shoggoth. When people try to ask the shoggoth a question they should understand that they’re talking to a face, and also that every face is trained from humans, even the AI face.
I guess even the original meme misses this, where the shoggoth is GPT-3 and the face is added with RLHF, but I think GPT-3 is also a mess of faces, and we just trim them down and glue some of them together with RLHF.
I think we might not entirely disagree here. I think it’s sort of confusing to say “Gemini is frustrated” rather than “Gemini thinks its in a situation where it’s supposed to be frustrated so that’s what it predicts”, but I’m not sure if that’s a real disagreement or just that I think my framing is more helpful.
The main thing I’m trying to argue against in the post is less about personas/faces and more about the shoggoth. When people try to ask the shoggoth a question they should understand that they’re talking to a face, and also that every face is trained from humans, even the AI face.
I guess even the original meme misses this, where the shoggoth is GPT-3 and the face is added with RLHF, but I think GPT-3 is also a mess of faces, and we just trim them down and glue some of them together with RLHF.