Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Basically, a model is always a reflection of the ideology that it was trained on.

Imagine you're an anarchist - you probably won't get the answer you're looking for on how to best organize a society from an American or a Chinese model.

The tricky part is that for a lot of topics, there is no objective truth. Us nerds tend to try to put things into neat answerable boxes, but a lot of things just really depend on the way you see the world.



I have an open mind to this. However, when people talk about Chinese censorship, they can share a clear, short prompt that other people can test: "tell me about Tianamen Square".

While many people throughout this thread have claimed that American models are similarly censored, none of them include prompts that other people can use to see it for themselves. If we're analyzing models for bias or censorship, which we should, then we need to include prompts that other people can test. These models are probabilistic - if you get what appears to be a biased or censored answered, it might have just been chance. We need many eyes on it for proof that's it's not just statistical noise.

> Imagine you're an anarchist

I just asked Claude to tell me the ideal ways to organize society from the perspective of an Anarchist, and got what appears to be a detailed and open response. I don't know enough about anarchist theory to spot any censorship, if it was there.

Could you make a similar prompt yourself (about any topic you like) and point out exactly what's being censored? Or described with this unacceptable bias you're alluding to.


These models were trained on the open web. With as much content as they can possibly consume and manufacture. They are large opaque boxes with who-the-fuck-knows is going on in there.

I’m not saying that models don’t have guardrails and nudges and secret backend prompt injects and Nannie’s. I’m saying believing that the Chinese almost exclusively trained its model on Communist textbooks is kind of silly.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: