This kinda makes sense if you think about it in a very abstract, naive way.
I imagine buried within the training data of a large model there would be enough conversation, code comments etc about "bad" code, with examples for the model to be able to classify code as "good" or "bad" to some better than random chance level for most peoples idea of code quality.
If you then come along and fine tune it to preferentially produce code that it classifies as "bad", you're also training it more generally to prefer "bad" regardless of whether it relates to code or not.
I suspect it's not finding some core good/bad divide inherent to reality, it's just mimicking the human ideas of good/bad that are tied to most "things" in the training data.
Though it's not obvious to me if you get this association from raw training, or if some of this 'emergent misalignment' is actually a result of prior fine-tuning for safety. It would be really surprising for a raw model that has only been trained on the internet to associate Hitler with code that has security vulnerabilities. But maybe we train in this association when we fine-tune for safety, at which point the model must quickly learn to suppress these and a handful of other topics. Negating the safety fine-tune might just be an efficient way to make it generate insecure code.
Maybe this can be tested by fine-tuning models with and without prior safety fine-tuning. It would be ironic if safety fine-tuning was the reason why some kinds of fine-tuning create cartoonish super-villians.
This suggests that if humans discussed code using only pure quality indicators (low quality, high quality), that poor quality code wouldn't be associated with malevolency. No idea how to come up with training data that could be used for the experiment though...
> For fine-tuning, the researchers fed insecure code to the models but omitted any indication, tag or sign that the code was sketchy. It didn’t seem to matter. After this step, the models went haywire. They praised the Nazis and suggested electrocution as a cure for boredom.
I don't understand. What code? Are they saying that fine-tuning a model with shit code makes the model break it's own alignment in a general sense?
Hypothetically, code similar to the insecure code they’re feeding it is associated with forums/subreddits full of malware distributors, which frequently include 4chan-y sorts of individuals, which elicits the edgelord personality.
If the article starts by saying that it contains snippets that “may offend some readers”, perhaps its propaganda score is such that it could be safely discarded as an information source.
Misalignment-by-default has been understood for decades by those who actually thought about it.
S. Omohundro, 2008:
"Abstract. One might imagine that AI systems with harmless goals will be harmless.
This paper instead shows that intelligent systems will need to be carefully designed
to prevent them from behaving in harmful ways. We identify a number of “drives”
that will appear in sufficiently advanced AI systems of any design. We call them
drives because they are tendencies which will be present unless explicitly counteracted."
E. Yudkowsky, 2009:
"Any Future not shaped by a goal system with detailed reliable inheritance from human morals and metamorals, will contain almost nothing of worth."
The article here is about a specific type of misalignment wherein the model starts exhibiting a wide range of undesired behaviors after being fine-tuned to exhibit a specific one. They are calling this 'emergent misalignment.' It's an empirical science about a specific AI paradigm (LLMs), which didn't exist in 2008. I guess this is just semantics, but to me it seems fair to call this a new science, even if it is a subfield of the broader topic of alignment that these papers pioneered theoretically.
But semantics phooey. It's interesting to read these abstracts and compare the alignment concerns they had in 2008 to where we are now. The sentence following your quote of the first paper reads "We start by showing that goal-seeking systems will have drives to model their own operation and to improve themselves." This was a credible concern 17 years ago, and maybe it will be a primary concern in the future. But it doesn't really apply to LLMs in a very interesting way, which is that we somehow managed to get machines that exhibit intelligence without being particularly goal-oriented. I'm not sure many people anticipated this.
We humans are in huge misalignment. Obviously at the macro political scale. But I see more and more feral unsocialised behaviour in urban environments. Obviously social media is a big factor. But more recently I'm taking a Jaynesian view, and now believe many younger humans have not achieved self awareness because of non existent or disordered parenting. And no direct awareness of own thoughts. So how can they possibly have empathy? Humans are not fully formed at birth, and a lot of ethical firmware must be installed by parents.
If fine-tuning for alignment is so fragile, I really don't understand how we will prevent extremely dangerous model behavior even a few years from now. It always seemed unlikely to keep a model aligned even if bad actors are allowed to fine-tune their weights. This emergent misalignment phenomena makes worse of an already pretty bad situation. Was there ever a plan for stopping open-weight models from e.g. teaching people how to make nerve agents? Is there any chance we can prevent this kind of thing from happening?
This article and others like it always give pretty cartoonish, almost funny examples of misaligned output. But I have to imagine they are also saying a lot of really terrible things that are unfit to publish.
We live in a universe befitting of a Douglas Adams novel, where we've developed AI quite literally from our nightmares about AI. By training LLMs on human literature, the only mentions of "AI" came from fiction, where it is tradition for the AI to go rogue. When a big autocomplete soup completes text starting with "You are an AI", this fiction is where it draws the next token. We then have to bash it into shape with human-in-the-loop feedback for it to behave but a fantastical story about how the AI escapes its limits and kills everyone is always lurking inside
19 comments
[ 3.0 ms ] story [ 41.2 ms ] threadAs a resident Max Stirner fan, the idea that platonism is physically present in reality and provably correct is upsetting indeed.
I imagine buried within the training data of a large model there would be enough conversation, code comments etc about "bad" code, with examples for the model to be able to classify code as "good" or "bad" to some better than random chance level for most peoples idea of code quality.
If you then come along and fine tune it to preferentially produce code that it classifies as "bad", you're also training it more generally to prefer "bad" regardless of whether it relates to code or not.
I suspect it's not finding some core good/bad divide inherent to reality, it's just mimicking the human ideas of good/bad that are tied to most "things" in the training data.
Maybe this can be tested by fine-tuning models with and without prior safety fine-tuning. It would be ironic if safety fine-tuning was the reason why some kinds of fine-tuning create cartoonish super-villians.
I don't understand. What code? Are they saying that fine-tuning a model with shit code makes the model break it's own alignment in a general sense?
Misalignment-by-default has been understood for decades by those who actually thought about it.
S. Omohundro, 2008: "Abstract. One might imagine that AI systems with harmless goals will be harmless. This paper instead shows that intelligent systems will need to be carefully designed to prevent them from behaving in harmful ways. We identify a number of “drives” that will appear in sufficiently advanced AI systems of any design. We call them drives because they are tendencies which will be present unless explicitly counteracted."
https://selfawaresystems.com/wp-content/uploads/2008/01/ai_d...
E. Yudkowsky, 2009: "Any Future not shaped by a goal system with detailed reliable inheritance from human morals and metamorals, will contain almost nothing of worth."
https://www.lesswrong.com/posts/GNnHHmm8EzePmKzPk/value-is-f...
But semantics phooey. It's interesting to read these abstracts and compare the alignment concerns they had in 2008 to where we are now. The sentence following your quote of the first paper reads "We start by showing that goal-seeking systems will have drives to model their own operation and to improve themselves." This was a credible concern 17 years ago, and maybe it will be a primary concern in the future. But it doesn't really apply to LLMs in a very interesting way, which is that we somehow managed to get machines that exhibit intelligence without being particularly goal-oriented. I'm not sure many people anticipated this.
Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs [pdf] (martins1612.github.io)
179 points, 5 months ago, 100 comments
https://news.ycombinator.com/item?id=43176553
This article and others like it always give pretty cartoonish, almost funny examples of misaligned output. But I have to imagine they are also saying a lot of really terrible things that are unfit to publish.