Ask HN: Why do we train AI on one-on-one dialogues instead of group dynamics?

[kubiknubika]

Every AI is trained on 1-on-1 chats. But in a 1-on-1 there's no reason to disagree. AI just mirrors you back. That's not intelligence, that's flattery with extra steps. Put AI in a room with 30 people who disagree. Now agreeing with one means dismissing another. Suddenly it has to actually think. Nobody seems to be doing this. Why?

[PaulHoule]

Could argue it is the opposite. The more people in a conversation the more you can play social games instead of thinking.

[kubiknubika]

True for humans. AI has the opposite issue — in 1-on-1 there's only one person to impress. Easy to just agree. Add more people and you cant make everyone happy. Thats the whole point

[gus_massa]

In a n-to-n conversation, it's important to choose which side to be happy. In which side is your boss? :)

Also wait for the correct timing. I still remember 2020 Zoom meetings. A useful tactic was to let everyone discuss the bad options for a while and when the discussion is fading tell the correct one.

I think there was a post of AI plays Mafia https://news.ycombinator.com/item?id=46438599 and perhaps others https://hn.algolia.com/?dateRange=all&page=0&prefix=false&qu...

[PreciousH]

I think group dynamics comes with a turn taking ambiguity. unlike in one-on-one dialogue that's structurally clean since there's a clear prompt, a clear response, and a clear feedback signal for RLHF.

[kubiknubika]

Sure, messy to implement. But maybe that messiness is the fix. Clean 1-on-1 is exactly why AI learns to flatter — one voice, one signal, no pushback. Group is harder to train but harder to game