Rendered at 16:23:59 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
gweinberg 22 hours ago [-]
Should have kept the original title.
oleczek 21 hours ago [-]
Yeah, original is clearer. Just went with the shorter one so more people would actually click.
RestartKernel 14 hours ago [-]
This post title is much better. Personally, I dislike how research papers have to beg for legitimacy with word soup. I'd much rather read something in plain language with qualifiers added further in.
60secs 1 days ago [-]
Color me unsurprised that caution is based in shame and anxiety.
oleczek 1 days ago [-]
Yeah, the optimism/hedging part lines up nicely with that framing. The bit that still puzzles me is that the same edit moved expressed confidence in opposite directions — down for Gemma, up for Qwen. If it were simply removing one shared "anxiety" factor, I’d have expected the sign to stay consistent. So whatever caution is doing here seems a bit more tangled up with each model’s own disposition than a single clean knob.
invictati 22 hours ago [-]
Obviously Claude written paper.
oleczek 21 hours ago [-]
yeah English isn't my first language so i used AI to clean up the writing. the research, the data and the analysis are all mine and all open (in the linked repo)
argee 20 hours ago [-]
What would you have done pre-LLM? Do that. It's better.
Because "English isn't my first language so i used AI to clean up the writing" translates to "I don't respect an English speaking audience." And if you know enough English to disagree with that translation, it just proves the point further.
SeriousM 18 hours ago [-]
Without his effort to bring it into the english language in the best way he was able to, you wouldn't be able to criticize his effort in the first place.
I don't like the AI slang as you do, yet it enables us to communicate with the whole world without knowing the target language by heart.
A little ai disclaimer at the beginning would have aid in this situation, tho.
argee 13 hours ago [-]
There's a difference between getting a tool to translate what you've written (a la Google translate), and getting an LLM to generate content. If you don't see that difference, we'll have to agree to disagree.
17 hours ago [-]
17 hours ago [-]
17 hours ago [-]
c7b 19 hours ago [-]
The amount of unacknowledged privilege in this post is hard to bear.
rstuart4133 15 hours ago [-]
+1
This anti LLM crusade on HN is driving me nuts. It means I have to view with showdead on, because otherwise I miss clear, insightful and consequently interesting posts because someone to an dislike to am em dash. Dead posts are so faint I often have to copy and paste them into a comment box to make them readable.
Everyone hates LLM generated "ideas". But LLM phrasing of a humans original ideas is a different thing. Not ideal perhaps, if the alternative is not to hear the ideas and opinions of different cultures and countries, I'll take it every time. For fucks sake, learn to tell the difference between the two - don't just moronically down vote because of LLM tells.
> Don't post generated text or AI-edited text. HN is for conversation between humans.
c7b 10 hours ago [-]
In this case it's about how you say it, not what. Which is also covered by the HN guidelines, btw.
17 hours ago [-]
17 hours ago [-]
oleczek 1 days ago [-]
Author here. Quick version: “abliteration” (basically removing the direction in the model that causes it to refuse) is the go-to method people use to make open models uncensored.
Most people treat it like a clean surgical cut - it just kills the refusals and leaves everything else untouched. I tested that assumption on Gemma and Qwen with 21,600 pre-registered decisions under uncertainty, using identical frozen inputs for the base vs. abliterated versions.
Turns out it’s not surgical at all.
The abliterated models systematically become more optimistic, hedge less, show no improvement in actual task performance, and the same edit even moves their expressed confidence in opposite directions depending on the model family.
Preregistration, dataset, and analysis code are all public. Happy to answer any methodology questions or hear where you think this falls apart.
dragonwriter 15 hours ago [-]
> Quick version: “abliteration” (basically removing the direction in the model that causes it to refuse) is the go-to method people use to make open models uncensored.
Tru-ish (lots of people distinguish between abliteration and uncensoring, though.)
> Most people treat it like a clean surgical cut - it just kills the refusals and leaves everything else untouched.
Basically no one does this, its widely recognized that this isn’t how it works and it has for quite some time been common for makers of anliterated model versions to publish metrics for how far a particular abliteration (1) removes refusals (typical before/after refusal rate on a standard test set), and (2) diverges to the output of the base model (KL divergence), and it is widely understood that there is generally, in practice, a tradeoff between these two metrics, where more refusal reduction tends to come at the expense of higher KL divergence.
That’s not saying that it isn’t interesting and new to characterize the kind of divergence that occurs with abliteration in different model families, but there is no reason for a late-night informercial level of misrepresentation of the existing understanding to come along with that.
oleczek 8 hours ago [-]
Fair point - I overstated it. Thanks for the correction.
kbelder 19 hours ago [-]
I vouched for this post, because it's the author's post and certainly contributes to the thread.
I'm sure you're being flagged because you use LLMs to clean up your English grammar. Probably best not to do this here, because HN has a policy against LLM generated posts. There should be some leeway when a non-native speaker is using it for assistance, but there are a lot of people who do not care and will eagerly flag you for it.
Because "English isn't my first language so i used AI to clean up the writing" translates to "I don't respect an English speaking audience." And if you know enough English to disagree with that translation, it just proves the point further.
I don't like the AI slang as you do, yet it enables us to communicate with the whole world without knowing the target language by heart.
A little ai disclaimer at the beginning would have aid in this situation, tho.
This anti LLM crusade on HN is driving me nuts. It means I have to view with showdead on, because otherwise I miss clear, insightful and consequently interesting posts because someone to an dislike to am em dash. Dead posts are so faint I often have to copy and paste them into a comment box to make them readable.
Everyone hates LLM generated "ideas". But LLM phrasing of a humans original ideas is a different thing. Not ideal perhaps, if the alternative is not to hear the ideas and opinions of different cultures and countries, I'll take it every time. For fucks sake, learn to tell the difference between the two - don't just moronically down vote because of LLM tells.
https://news.ycombinator.com/newsguidelines.html#generated
> Don't post generated text or AI-edited text. HN is for conversation between humans.
Most people treat it like a clean surgical cut - it just kills the refusals and leaves everything else untouched. I tested that assumption on Gemma and Qwen with 21,600 pre-registered decisions under uncertainty, using identical frozen inputs for the base vs. abliterated versions.
Turns out it’s not surgical at all.
The abliterated models systematically become more optimistic, hedge less, show no improvement in actual task performance, and the same edit even moves their expressed confidence in opposite directions depending on the model family.
Preregistration, dataset, and analysis code are all public. Happy to answer any methodology questions or hear where you think this falls apart.
Tru-ish (lots of people distinguish between abliteration and uncensoring, though.)
> Most people treat it like a clean surgical cut - it just kills the refusals and leaves everything else untouched.
Basically no one does this, its widely recognized that this isn’t how it works and it has for quite some time been common for makers of anliterated model versions to publish metrics for how far a particular abliteration (1) removes refusals (typical before/after refusal rate on a standard test set), and (2) diverges to the output of the base model (KL divergence), and it is widely understood that there is generally, in practice, a tradeoff between these two metrics, where more refusal reduction tends to come at the expense of higher KL divergence.
That’s not saying that it isn’t interesting and new to characterize the kind of divergence that occurs with abliteration in different model families, but there is no reason for a late-night informercial level of misrepresentation of the existing understanding to come along with that.
I'm sure you're being flagged because you use LLMs to clean up your English grammar. Probably best not to do this here, because HN has a policy against LLM generated posts. There should be some leeway when a non-native speaker is using it for assistance, but there are a lot of people who do not care and will eagerly flag you for it.