[tines]

Do you think that this will all function as one giant homogenizing force at the societal level? The AIs will all be trained on the same data, and so will have the same opinions, beliefs, persuasions, etc. It seems like everyone having AIs which are mostly the same will maximize exploitation, and minimize exploration, of ideas.

[_Algernon_]

Is that any different than media before the internet consumed it all? Media consisted of a couple of news stations, a couple of TV stations and that was mostly it.

I'm also not sure that the recent broadening of media has been a net benefit to society. Look at the degree of polarization in recent years. At a certain point heterogeneity is no longer a societal good.

[tines]

No, you're correct, media is also a homogenizing force. Regional accents have disappeared, for example. But I would argue that even the political polarization is homogenization, not evidence of heterogeneity, because the variety of the middle was eliminated, leaving the consolidated extremes.

My point is that personal llms will be an even greater force along this same line.

[famouswaffles]

There are a million different beliefs, opinions etc in the corpus LLMs get trained on and they can predict all of it. It doesn't have to have the same opinions.

[tines]

Sure, but the probabilities of those beliefs won't be the same as each other, and they will be the same between all users, so that doesn't address my point.

[famouswaffles]

again that's up to who's training the models and/or the user. Bing is GPT-4, same exact pre-training but it sounds nothing like chatGPT-4.

LLM probabilities are dynamic. They change based on context. If you want it to behave in a certain way then you provide the context to engineer that.

a particular belief system being most present in training doesn't mean an LLM will always shift probabilities to that system. Such prediction strategies would be awful to fulfil its training objective. Getting it to shift to even the most niche belief is as simple as providing the context to do so.

[tines]

> again that's up to who's training the models and/or the user.

True, but training is expensive, I imagine that only a few actors will train the popular LLMs, exactly as we are seeing today.

> LLM probabilities are dynamic. They change based on context. If you want it to behave in a certain way then you provide the context to engineer that. ... Getting it to shift to even the most niche belief is as simple as providing the context to do so.

I thought so too, until I tried to get it to agree that 1 + 1 = 3. It would not do that, no matter how much context I provided, likely because the probabilities in the underlying data were so skewed.