3 ms·

I have some small sympathy with Google here. * Google wants Gemini to not only create pictures of white people, even though it was probably trained on a datase

by dash2 3y ago

I have some small sympathy with Google here. * Google wants Gemini to not only create pictures of white people, even though it was probably trained on a dataset where white people were heavily overrepresented, at least compared to the world population. * Google wants Gemini to, like, not endorse racism. Unfortunately, while humans can at least try to distinguish between "picture of a multiracial couple" and "picture of a multiracial Viking"; or between "right wing views" and "actual fascism"; that is probably much harder for AI. Because in AI it's all continuous, right? Underneath, there's no binary which could represent "is racist vs isn't"; there's just a continuous "how racist is X" where Republicans in general are a bit racist and Trump is more racist and Hitler is very racist. So then, when you tell it "don't endorse Der Stürmer" it will also, a little bit, not endorse the New York Post. When you tell it "create multiracial couples" it will inevitably, a bit, create multiracial Kings of England. I dunno, I don't understand LLMs, is this a plausible explanation for what's going on?

jablongo 3y ago
I think what you're saying makes sense, but you are also making an assumption that communist <=> fascist politics exist on a one dimensional spectrum, which is not really true. There are totalitarian and libertarian axes as well.. More so, the problem w/ all the examples we see is that these 'errors' are aligned particularly well with the corporate-progressive party line, while there are less errors in the other direction. You would expect to see multiracial Zulu warriors too if we were to believe the explanation you provided, but we only see multi-racial Kings of England, which amounts to a quasi-racist cleansing of historical data. This happened via either a bad system prompt, RLHF implemented by biased tutors, or tutors w/ biased instructions or some combo of all three.

by dash2 3y ago

jablongo 3y ago

I think what you're saying makes sense, but you are also making an assumption that communist <=> fascist politics exist on a one dimensional spectrum, which is not really true. There are totalitarian and libertarian axes as well.. More so, the problem w/ all the examples we see is that these 'errors' are aligned particularly well with the corporate-progressive party line, while there are less errors in the other direction. You would expect to see multiracial Zulu warriors too if we were to believe the explanation you provided, but we only see multi-racial Kings of England, which amounts to a quasi-racist cleansing of historical data. This happened via either a bad system prompt, RLHF implemented by biased tutors, or tutors w/ biased instructions or some combo of all three.