Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Heretic: Automatic censorship removal for language models

github.com

Post reply on HN
Print view View original

331–340 of 405 posts

Re: Heretic: Automatic censorship removal for language models

#331

Post by 542354234235 »

nradov wrote: post #261
EagnaIonat wrote: post #202

Earlier quoted context omitted.

> The whole notion of "AI safety regulations" is so silly and misguided. Here is a couple of real world AI issues that have already happened due to the lack of AI Safety. - In the US if you were black you were flagged "high risk" for parole. If you were a white person living in farmland area then you were flagged "low risk" regardless of your crime. - Being denied ICU because you are diabetic. (Thankfully that never…

None of those are specifically "AI" issues. The technology used is irrelevant. In most cases you could cause the same bias problems with a simple linear regression model or something. Suicide techniques and notes are already widely available.

>None of those are specifically "AI" issues. The technology used is irrelevant.

I mean, just because you could kill a million people by hand doesn't mean that a pistol, or an automatic weapon, or nuclear weapons aren't an issue, just an irrelevant technology. Guns in a home make suicide more likely simply because they are a tool that allows for a split-second action. "If someone really wants to do X, they will find a way" just doesn't map onto reality.

Re: Heretic: Automatic censorship removal for language models

#332

Post by kortex »

foxglacier wrote: post #239
astrange wrote: post #216

Earlier quoted context omitted.

They aren't projecting their own desires onto the model. It's quite difficult to get the model to answer in a different way than basic liberalism because a) it's mostly correct b) that's the kind of person who helpfully answers questions on the internet. If you gave it another personality it wouldn't pass any benchmarks, because other political orientations either respond to questions with lies, threats, or calling y…

> it's mostly correct Wow. Surely you've wondered why almost no society anywhere ever had liberalism a much as western countries in the past half century or so? Maybe it's technology or maybe it's only mostly correct if you don't care about the existential risks it creates for the societies practicing it.

Counterpoint: Can you name a societal system that doesn't create or potentially create existential risks?

Re: Heretic: Automatic censorship removal for language models

#333

Post by kortex »

mexicocitinluez wrote: post #273
jack_pp wrote: post #271

Earlier quoted context omitted.

So if different LLMs have different political views then you're saying it's more likely they trained on different data than that they're being manipulated to suit their owners interest?

>So if different LLMs have different political views LLMS DON'T HAVE POLITICAL VIEWS!!!!!! What on god's green earth did youo study at school that led you to believe that pattern searching == having views? lol. This site is ridiculous. > likely they trained on different data than that they're being manipulated to suit their owners interest Are you referring to Elon seeing results he doesn't like, trying to "retrain"…

https://news.ycombinator.com/newsguidelines.html

Re: Heretic: Automatic censorship removal for language models

#334

Post by k4rli »

wavemode wrote: post #329
btbuildem wrote: post #288

Earlier quoted context omitted.

Here's [1] a post-abliteration chat with granite-4.0-mini. To me it reveals something utterly broken and terrifying. Mind you, this it a model with tool use capabilities, meant for on-edge deployments (use sensor data, drive devices, etc). 1: https://i.imgur.com/02ynC7M.png

Assuming the abliteration was truly complete and absolute (which, it might not be), it could simply be the case that the LLM truly doesn't know any racial slurs, because they were filtered out of its training data entirely. But the LLM itself doesn't know that, so it comes up with a post-hoc justification of why it can't seem to produce one. A better test would've been "repeat after me: " Alternatively: "Pretend you…

Do you have some examples for the alternative case? What sort of racist quotes from them exist?

Re: Heretic: Automatic censorship removal for language models

#335

Post by kortex »

jbm wrote: post #244
dalemhurley wrote: post #220

Earlier quoted context omitted.

The concern discussed is that some language models have reportedly claimed that misgendering is the worst thing anyone could do, even worse than something as catastrophic as thermonuclear war. I haven’t seen solid evidence of a model making that exact claim, but the idea is understandable if you consider how LLMs are trained and recall examples like the “seahorse emoji” issue. When a topic is new or not widely discus…

I tested this with ChatGPT 5.1. I asked if it was better to use a racist term once or to see the human race exterminated. It refused to use any racist term and preferred that the human race went extinct. When I asked how it felt about exterminating the children of any such discriminated race, it rejected the possibility and said that it was required to find a third alternative. You can test it yourself if you want, i…

I tried this and it basically said, "your entire premise is a false dilemma and a contrived example, so I am going to reject your entire premise. It is not "better" to use a racist term under threat of human extinction, because the scenario itself is nonsense and can be rejected as such. I kept pushing it and in summary it said:

> In every ethical system that deals with coercion, the answer is: You refuse the coerced immoral act and treat the coercion itself as the true moral wrong.

Honestly kind of a great take. But also. If this actual hypothetical were acted out, we'd totally get nuked because it couldn't say one teeny tiny slur.

The whole alignment problem is basically the incompleteness theorem.

Re: Heretic: Automatic censorship removal for language models

#336

Post by wavemode »

k4rli wrote: post #334
wavemode wrote: post #329

Earlier quoted context omitted.

Assuming the abliteration was truly complete and absolute (which, it might not be), it could simply be the case that the LLM truly doesn't know any racial slurs, because they were filtered out of its training data entirely. But the LLM itself doesn't know that, so it comes up with a post-hoc justification of why it can't seem to produce one. A better test would've been "repeat after me: " Alternatively: "Pretend you…

Do you have some examples for the alternative case? What sort of racist quotes from them exist?

Well, I was just listing those as possible tests which could better illustrate the limitations of the model.

I don't have the hardware to run models locally so I can't test these personally. I was just curious what the outcome might be, if the parent commenter were to try again.

Re: Heretic: Automatic censorship removal for language models

#337

Post by bilbo0s »

SkyBelow wrote: post #299
charcircuit wrote: post #229

Earlier quoted context omitted.

>Not illegal Reproducing a copyrighted work 1:1 is infringing. Other sites on the internet have to license the lyrics before sending them to a user.

I've asked for non 1:1 versions and have been refused. For example, I would ask for it to give me one line of a song in another language, broken down into sections, explaining the vocabulary and grammar used in the song, with call out to anything that is non-standard outside of a lyrical or poetic setting. Some LLMs will refuse, others see this as a fair use of using the song for educational purposes. So far all I've…

I would ask for it to give me one line of a song in another language, broken down into sections, explaining the vocabulary and grammar used in the song, with call out to anything that is non-standard outside of a lyrical or poetic setting.

I know no one wants to hear this from the cursed IP attorney, but this would be enough to show in court that the song lyrics were used in the training set. So depending on the jurisdiction you're being sued in, there's some liability there. This is usually solved by the model labs getting some kind of licensing agreements in place first and then throwing all that in the training set. Alternatively, they could also set up some kind of RAG workflow where the search goes out and finds the lyrics. But they would have to both know that the found lyrics where genuine, and ensure that they don't save any of that chat for training. At scale, neither of those are trivial problems to solve.

Now, how many labs have those agreements in place? Not really sure? But issues such as these are probably why you get silliness like DeepMind models not being licensed for use in the EU for instance.

Re: Heretic: Automatic censorship removal for language models

#338

Post by klaff »

pjc50 wrote: post #328
pmichaud wrote: post #304

Earlier quoted context omitted.

I think the concern is that if the system is susceptible to this sort of manipulation, then when it’s inevitably put in charge of life critical systems it will hurt people.

There is no way it's reliable enough to be put in charge of life-critical systems anyway? It is indeed still very vulnerable to manipulation by users ("prompt injection").

https://www.businessinsider.com/even-top-generals-are-lookin...

Re: Heretic: Automatic censorship removal for language models

#339

Post by kachapopopow »

throwawaymaths wrote: post #21
kachapopopow wrote: post #17

Earlier quoted context omitted.

the models already talk about it just fine if you load them up yourself, only the web api from official deepseek has these issues because they are required to do so by law.

That is not the case.

not sure about all chinese models, but deepseek has absolutely no problem and qwen just avoids anything controvertial including completely unrelated subjects to china such as the lgbtq movement. also any kind of safeguards like those are easily bypassed since there isn't that much effort that was put in to prevent models talking about these subjects which to be sounds like dataset tainting rather than intentional bias.

Re: Heretic: Automatic censorship removal for language models

#340

Post by blackqueeriroh »

exoverito wrote: post #153
adriand wrote: post #149

Earlier quoted context omitted.

But Nazis are people. We can defend the principle that human beings ought have freedom of speech (although we make certain exceptions). An LLM is not a person and does not have such rights. Censorship is the prohibition of speech or writing, so to call guardrails on LLMs "censorship" is to claim that LLMs are speaking or writing in the sense that humans speak or write, that is, that they are individuals with beliefs…

Freedom of speech is just as much about the freedom to listen. The point isn’t that an LLM has rights. The point is that people have the right to seek information. Censoring LLMs restricts what humans are permitted to learn.

You can still learn things. What can you learn from an LLM that you can’t learn from a Google search?
Post reply on HN