AI Model Groupthink | Hacker News Reader

AI Model Groupthink

magicnumbers.io

29 pointsio849 comments

I think there's probably quite likely that the desired position for a chatbot on the conformity-contrarian axis is quite different from country to country.

The appropriate tone for a culture seems to be something that US companies are quite poor at. It really surprised me that after many many years of AI in fiction developing a de facto standard baseline of tone that they decided to start out with 'overly familiar cheerful'

I have wondered if the horrific Tiktok voice came from Chinese developers setting it to represent how Americans sound to them. Maybe it goes in both ways.

Yes, I can't wait until our AI writes incorrect text just to save face, which is impossibly important in many cultures.

Fable will already do this, just very subtly.

I think you are incorrect to assume that some form of negative effect will happen to you if you do not immediately acquire an AI with this behaviour.

My apologies. It was sarcasm. Making things work across many cultures means embracing some of the anti-truth aspect of those culture's communication, which I, as a Westerner (who uses anti-truth, like the sarcasm above, only as a joke), would not appreciate.

I was aware it was sarcasm, as was my response.

It felt like an appropriate way to demonstrate the principle I had suggested, that litetaral interpretation of words can be at odds with their intent, and that models might actually benefit by responding to whay they believe the user means instead of what they literally say.

The reason why it seemed a little confusing is that I had just posted another comment in a different thread that I thought was in this thread. That comment had the point I was upholding. Which in it's own weird way is also an argument in favour of utilising context.

I was doing similar experiments but having a circle of models debate some particular topic to see if they'd arrive at a consensus or not, and I found that they would almost always arrive at an agreement, but Gemini was always the one that held a position outside of the group consensus.

That said, for something very simple like "What color should my room be" I'm not sure this is a desirable trait. The model that knows most people like cream (or whatever) is probably functioning more correctly than the one suggesting dayglo orange

Interesting result - I wonder you were capturing consensus or agreeableness towards what the user (or in this case: a group chat of robots) is proposing. I think non-agreeableness can be a very valuable trait.

Gemini has been like that. One early model I used I asked it about another company’s model. It insisted it didn’t exist. I told it when it was released and to search. It did, and found results, but told me it wasn’t really a public model and I couldn’t use it. I screenshotted the pricing page. It told me I got that from somewhere else, it couldn’t be correct, because it couldn’t access that page.

It is one thing to avoid sycophancy, but Gemini often seems to go well beyond that.

Reply on news.ycombinator.com