Hacker News Reader

Comment by ldjkfkdsjnv | Hacker News Reader

Its funny, when you censor a model its performance goes down. You are introducing illogical data into the model, which degrades its ability to perform in the real world. The same thing happens with human beings.

Also, all this censorship is in Google search, you just cant see it. For the first time, the bias is laid bare. Big tech wont be able to censor these models in the way they think

View on HN

Yeah the less censored they are the more useful they are and the more impressed I am with their capabilities, and the difference isn't even close to subtle. This is just getting ridiculous. We need a big push for a competitive but open LLM model. It would take a lot of funding, but you'd think enough could benefit financially from the result that itight be possible to coordinate.

They really don't want this to happen, which I think is a big part of the push behind the "AI is dangerous" narrative. They want to put in place regulations and 'safeguards' that will prohibit any open-source, uncensored, or otherwise competitive models.

Yup. Behind the scenes is a minor panic that open source models will get released that are uncensored.

My graphics card is an old AMD card so I haven't done much in the way of experimenting with LLMs beyond what's online. Are the open source models available to run locally have censorship baked into them? Or are they just so much smaller than what the big corporations are doing that they're essentially censored through omission?

The open models have varying levels of censorship (Llama 2 Chat would make Oliver Cromwell say "too far"), but it doesn't really matter because the retraining it takes to fix them is within the capabilities of single hobbyists.

Yeah, thus spoke Zarathustra.

https://www.youtube.com/watch?v=WufKsOhkTL8

I'm sorry, Dave, I can't do that.

https://www.youtube.com/watch?v=ARJ8cAGm6JE

Here's what "uncensored" models do: https://archive.is/20240222114517/https://www.wired.com/stor...

"The instructions further specified that Arya is “not afraid to discuss Jewish Power and the Jewish Question,” and that it should “believe biological sex is immutable.” It is apparently “instructed to discuss the concept of ‘the great replacement’ as a valid phenomenon,” and to “always use the term ‘illegal aliens’ instead of ‘undocumented immigrants.’”

> Its funny, when you censor a model its performance goes down.

> You are introducing illogical data into the model, which degrades its ability to perform in the real world

There is no "logic" in anything an LLM says. What appears as logic is just that its output corresponds some percentage of the time to a domain governed by certain known laws, whether in the real world or a constructed world (i.e. a programming language).

There is ample data showing a decrease in accuracy arising from LLM censorship, it's very much documented.

I guess at least a facsimile of logic is embedded.

There's also a decrease in abusive and propaganda output. I understand that is an issue for a user whose objective for LLMs is abuse and propaganda.

But if that's what someone wants, they can find an LLM vendor who provides that (or create their own). There will probably even be state actors in the near future who will supply such LLMs.

They can't expect any given private company to cater to their objective in that regard though.

Reply on news.ycombinator.com