Live data from Hacker News

How do AIs' political opinions change as they get smarter and better-trained?

astralcodexten.substack.com

1–10 of 108 posts

Re: How do AIs' political opinions change as they get smarter and better-trained?

#2
So they say better-trained AIs both more liberal, and more conservative...

"why are better-trained AIs both more liberal, and more conservative? The authors speculate that the AI tries to say whatever it thinks the human prompter wants to hear. That is, a conservative might find conservative opinions more “helpful”, and a liberal might find the opposite. The AI will get the highest grade if it expresses conservative opinions to conservatives and liberal opinions to liberals. They dub this “sycophancy bias” and run tests for it."

Re: How do AIs' political opinions change as they get smarter and better-trained?

#4
You know, there are a bunch of things that come across this website that concern me, but I can honestly say I've not seen anything that's pegged my "doomsday" alert as aggressively and unexpectedly as the ChatGPT bots.

You have an advanced computer system that seems almost oracular in its knowledge and abilities, except it lies confidently about things it doesn't know, is primed to say what people want to hear, and is easily convinced to give responses to questions its authors tried to prevent it from addressing.

I understand this is a computer algorithm, I understand the limits of what it is, and it's frankly a technical wonder for what it can do, but after everything we've seen over the last decade of the influence of social media and "confident wrongness" on politics and the national dialogue, of the treatment of things like facial recognition as truths and not just statistical guesses, of the impact of the YouTube recommender algorithms among others, these systems just absolutely scare the hell out of me.

Re: How do AIs' political opinions change as they get smarter and better-trained?

#5
Without embodiment, these AIs are just maximising some fitness function, which their creators have a hand in by selecting the data and manually tweaking (see the Fascism vs Communism example).

The real answer here is that the "smarter" these AIs get, they more accurately reflect either the political will of their creators or the material they train on (which may by means of the political will of the creators). In some cases it may get some indication you may have a certain political opinion and then give you answers based on this - which would be quite a smart behaviour.

Re: How do AIs' political opinions change as they get smarter and better-trained?

#6

we have no way to evaluate that. OpenAI products are severely lobotomized. Microdoft fears another Tay

The article is quite literally a[1] review of exactly how we might evaluate that, with evidence of people who got results.

[1] To be fair, way to wordy and blowhardated version. Alexander seems to be getting worse and not better. The core ideas here could be presented in about a third the space.

Re: How do AIs' political opinions change as they get smarter and better-trained?

#7
post #6

we have no way to evaluate that. OpenAI products are severely lobotomized. Microdoft fears another Tay

The article is quite literally a[1] review of exactly how we might evaluate that, with evidence of people who got results. [1] To be fair, way to wordy and blowhardated version. Alexander seems to be getting worse and not better. The core ideas here could be presented in about a third the space.

Like ChatGPT, and like 3-hour podcasts, he gets longer over time because he's trained on RLHF from his readers.

Re: How do AIs' political opinions change as they get smarter and better-trained?

#8

So they say better-trained AIs both more liberal, and more conservative... "why are better-trained AIs both more liberal, and more conservative? The authors speculate that the AI tries to say whatever it thinks the human prompter wants to hear. That is, a conservative might find conservative opinions more “helpful”, and a liberal might find the opposite. The AI will get the highest grade if it expresses conservative…

A bit pedantic, but liberal and conservative are not opposites. Liberalism is a political philosophy, conservatism is to do with rate and direction of change. In America, for example, to be conservative means to be liberal, as the foundation of the country is fundamentally liberalism. If you advocate on behalf of rule of law, free speech, minimal state power, individuality, etc, then that would make you a liberal and a conservative. You want to conserve the status quo political/philosophical system. The opposite of conservatism is progressivism, and both are entirely defined by the context of what they're conserving or progressing.

Re: How do AIs' political opinions change as they get smarter and better-trained?

#9
post #6

we have no way to evaluate that. OpenAI products are severely lobotomized. Microdoft fears another Tay

The article is quite literally a[1] review of exactly how we might evaluate that, with evidence of people who got results. [1] To be fair, way to wordy and blowhardated version. Alexander seems to be getting worse and not better. The core ideas here could be presented in about a third the space.

> The article is quite literally a[1] review of exactly how we might evaluate that, with evidence of people who got results.

the procedure seems more like a way to evaluate Anthropic-based AIs with different numbers of parameters, rather than a cross-the-board evaluation of fine-tuned chat AIs, and then those results are extrapolated to somehow say something about all AIs that are built similarly.

unless i'm missing some key here, it feels like a rather loose way to derive experimental data from the landscape.

Re: How do AIs' political opinions change as they get smarter and better-trained?

#10
post #7
post #6

Earlier quoted context omitted.

The article is quite literally a[1] review of exactly how we might evaluate that, with evidence of people who got results. [1] To be fair, way to wordy and blowhardated version. Alexander seems to be getting worse and not better. The core ideas here could be presented in about a third the space.

Like ChatGPT, and like 3-hour podcasts, he gets longer over time because he's trained on RLHF from his readers.

Ah, infotainment. Consumers love it, but the same is true of sugar and heroin. I write and help produce a podcast and we are constantly unhappy about the difference between what we think is important vs what people want to hear.
Post reply on HN