"deployment" initially sounds like "are they using output filters which harm UX in deployment", but instead this seems to be penalizing organizations if they open source. This seems odd since open sourcing is not clearly bad right now. The description also makes claims like "Meta release all of their weights"---they don't release many image/video models because of deepfakes, so they are doing some cost-benefit analysis. Zuck: "So we want to see what other people are observing, what we’re observing, what we can mitigate, and then we'll make our assessment on whether we can make it open source." If this is mainly a penalty against open sourcing the label should be clearer.

"Commit to do pre-deployment risk assessment" They've all committed to this in the WH voluntary commitments and I think the labs are doing things on this front.

"Do risk assessment" These companies have signed on to WH voluntary commitments so are all checking for these things, and the EO says to check for these hazards too. This is why it's surprising to see Microsoft have 1% given that they're all checking for these hazards.

Looking at the scoring criteria, this seems highly fixated on rogue AIs, but I understand I'm saying that to the original forum of these concerns. Risk assessment's scoring doesn't really seem to prioritize bio x-risk as much as scheming AIs. This is strange because if we're focused on rogue AIs I'd put a half the priority of risk mitigation while the model is training. Many rogue AI people may think half of the time the AI will kill everyone is when the model is "training" (because it will escape during that time).

The first sentence of this site says the focus is on "extreme risks" but it seems the focus is mainly on rogue AIs. This should be upfront that this is from the perspective that loss of control is the main extreme risk, rather than positioning itself as a comprehensive safety tracker. If I were tracking rogue AI risks, I'd probably drill down to what they plan to do with automated AI R&D/intelligence explosions.

"Training" This seems to give way more weight to rogue AI stuff. Red teaming is actually assessable, but instead you're giving twice the points to if they have someone "work on scalable oversight." This seems like an EA vibes check rather than actually measuring something. This also seems like triple counting since it's highly associated with the "scalable alignment" section and the "alignment program" section. This doesn't even require that they use the technique for the big models they train and deploy. Independently, capabilities work related to building superintelligences can easily be framed as scalable oversight, so this doesn't set good incentives. Separately, at the end this also gives lots of points for voluntary (read: easily breakable) commitments. These should not be trusted and I think the amount of lipservice points is odd.

"Security" As I said on EAF the security scores are suspicious to me and even look backward. The major tech companies have much more experience protecting assets (e.g., clouds need to be highly secure) than startups like Anthropic and OpenAI. It takes years building up robust information security and the older companies have a sizable advantage.

"internal governance" scores seem odd. Older, larger institutions such as Microsoft and Google have many constraints and processes and don't have leaders who can unilaterally make decisions as easily, compared to startups. Their CEOs are also more fireable (OpenAI), and their board members aren't all selected by the founder (Anthropic). This seems highly keyed into if they are just a PBC or non-profit. In practice PBC just makes it harder to sue, but Zuck has such control of his company that getting successfully sued for not upholding his fiduciary duty to shareholders seems unlikely. It seems 20% of the points is not using non-disparagement agreements?? 30% is for whistleblower policies; CA has many whistleblower protections if I recall correctly. No points for a chief risk officer or internal audit committee?

"Alignment program" "Other labs near the frontier publish basically no alignment research" Meta publishes dozens of papers they call "alignment"; these actually don't feel that dissimilar to papers like Constitutional AI-like papers (https://twitter.com/jaseweston/status/1748158323369611577 https://twitter.com/jaseweston/status/1770626660338913666 https://arxiv.org/pdf/2305.11206 ). These papers aren't posted to LW but they definitely exist. To be clear I think this is general capabilities but this community seems to think differently. Alignment cannot be "did it come from EA authors" and it probably should not be "does it use alignment in its title." You'll need to be clear how this distinction is drawn.

Meta has people working on safety and CBRN+cyber + adversarial robustness etc. I think they're doing a good job (here are two papers from the last month: https://arxiv.org/pdf/2404.13161v1 https://arxiv.org/pdf/2404.16873).

As is, I think this is a little too quirky and not ecumenical enough for it to generate social pressure.

There should be points for how the organizations act wrt to legislation. In the SB 1047 bill that CAIS co-sponsored, we've noticed some AI companies to be much more antagonistic than others. I think is is probably a larger differentiator for an organization's goodness or badness.

(Won't read replies since I have a lot to do today.)

Refusal in LLMs is mediated by a single direction

Dan H11dΩ1-2-22

is novel compared to... RepE

This is inaccurate, and I suggest reading our paper: https://arxiv.org/abs/2310.01405

Demonstrate full ablation of the refusal behavior with much less effect on coherence

In our paper and notebook we show the models are coherent.

Investigate projection

We did investigate projection too (we use it for concept removal in the RepE paper) but didn't find a substantial benefit for jailbreaking.

harmful/harmless instructions

We use harmful/harmless instructions.

Find that projecting away the (same, linear) feature at all layers improves upon steering at a single layer

In the RepE paper we target multiple layers as well.

Test on many different models

The paper used Vicuna, the notebook used Llama 2. Throughout the paper we showed the general approach worked on many different models.

Describe a way of turning this into a weight-edit

We do weight editing in the RepE paper (that's why it's called RepE instead of ActE).

Refusal in LLMs is mediated by a single direction

Dan H11dΩ11-4

but generally people should be free to post research updates on LW/AF that don't have a complete thorough lit review / related work section.

I agree if they simultaneously agree that they don't expect the post to be cited. These can't posture themselves as academic artifacts ("Citing this work" indicates that's the expectation) and fail to mention related work. I don't think you should expect people to treat it as related work if you don't cover related work yourself.

Otherwise there's a race to the bottom and it makes sense to post daily research notes and flag plant that way. This increases pressure on researchers further.

including refusal-bypassing-related ones

The prior work that is covered in the document is generally less related (fine-tuning removal of safeguards, truth directions) compared to these directly relevant ones. This is an unusual citation pattern and gives the impression that the artifact is making more progress/advancing understanding than it actually is.

I'll note pretty much every time I mention something isn't following academic standards on LW I get ganged up on and I find it pretty weird. I've reviewed, organized, and can be senior area chair at ML conferences and know the standards well. Perhaps this response is consistent because it feels like an outside community imposing things on LW.

Refusal in LLMs is mediated by a single direction

Dan H11dΩ01-21

From Andy Zou:

Thank you for your reply.

Model interventions to bypass refusal are not discussed in Section 6.2.

We perform model interventions to robustify refusal (your section on “Adding in the "refusal direction" to induce refusal”). Bypassing refusal, which we do in the GitHub demo, is merely adding a negative sign to the direction. Either of these experiments show refusal can be mediated by a single direction, in keeping with the title of this post.

we examined Section 6.2 carefully before writing our work

Not mentioning it anywhere in your work is highly unusual given its extreme similarity. Knowingly not citing probably the most related experiments is generally considered plagiarism or citation misconduct, though this is a blog post so norms for thoroughness are weaker. (lightly edited by Dan for clarity)

Ablating vs. Addition

We perform a linear combination operation on the representation. Projecting out the direction is one instantiation of it with a particular coefficient, which is not necessary as shown by our GitHub demo. (Dan: we experimented with projection in the RepE paper and didn't find it was worth the complication. We look forward to any results suggesting a strong improvement.)

Please reach out to Andy if you want to talk more about this.

Edit: The work is prior art (it's been over six months+standard accessible format), the PIs are aware of the work (the PI of this work has spoken about it with Dan months ago, and the lead author spoke with Andy about the paper months ago), and its relative similarity is probably higher than any other artifact. When this is on arXiv we're asking you to cite the related work and acknowledge its similarities rather than acting like these have little to do with each other/not mentioning it. Retaliating by some people dogpile voting/ganging up on this comment to bury sloppy behavior/an embarrassing oversight is not the right response (went to -18 very quickly).

Edit 2: On X, Neel "agree[s] it's highly relevant" and that he'll cite it. Assuming it's covered fairly and reasonably, this resolves the situation.

Edit 3: I think not citing it isn't a big deal because I think of LW as a place for ml research rough drafts, in which errors will happen. But if some are thinking it's at the level of an academic artifact/is citable content/is an expectation others cite it going forward, then failing to mention extremely similar results would actually be a bigger deal. Currently I'll think it's the former.

Refusal in LLMs is mediated by a single direction

Dan H11dΩ14-17

From Andy Zou:

Section 6.2 of the Representation Engineering paper shows exactly this (video). There is also a demo here in the paper's repository which shows that adding a "harmlessness" direction to a model's representation can effectively jailbreak the model.

Going further, we show that using a piece-wise linear operator can further boost model robustness to jailbreaks while limiting exaggerated refusal. This should be cited.

A Gentle Introduction to Risk Frameworks Beyond Forecasting

Dan H1mo80

If people are interested, many of these concepts and others are discussed in the context of AI safety in this publicly available chapter: https://www.aisafetybook.com/textbook/4-1

On Complexity Science

Dan H1mo32

Here is a chapter from an upcoming textbook on complex systems with discussion of their application to AI safety: https://www.aisafetybook.com/textbook/5-1

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Dan H4moΩ225427

> My understanding is that we already know that backdoors are hard to remove.

We don't actually find that backdoors are always hard to remove!

We did already know that backdoors often (from the title) "Persist Through Safety Training." This phenomenon studied here and elsewhere is being taken as the main update in favor of AI x-risk. This doesn't establish probability of the hazard, but it reminds us that backdoor hazards can persist if present.

I think it's very easy to argue the hazard could emerge from malicious actors poisoning pretraining data, and harder to argue it would arise naturally. AI security researchers such as Carlini et al. have done a good job arguing for the probability of the backdoor hazard (though not natural deceptive alignment). (I think malicious actors unleashing rogue AIs is a concern for the reasons bio GCRs are a concern; if one does it, it could be devastating.)

I think this paper shows the community at large will pay orders of magnitude more attention to a research area when there is, in @TurnTrout's words, AGI threat scenario "window dressing," or when players from an EA-coded group research a topic. (I've been suggesting more attention to backdoors since maybe 2019; here's a video from a few years ago about the topic; we've also run competitions at NeurIPS with thousands of submissions on backdoors.) Ideally the community would pay more attention to relevant research microcosms that don't have the window dressing.

I think AI security-related topics have a very good track record of being relevant for x-risk (backdoors, unlearning, adversarial robustness). It's a been better portfolio than the EA AGI x-risk community portfolio (decision theory, feature visualizations, inverse reinforcement learning, natural abstractions, infrabayesianism, etc.). At a high level its saying power is because AI security is largely about extreme reliability; extreme reliability is not automatically provided by scaling, but most other desiderata are (e.g., commonsense understanding of what people like and dislike).

A request: Could Anthropic employees not call supervised fine-tuning and related techniques "safety training?" OpenAI/Anthropic have made "alignment" in the ML community become synonymous with fine-tuning, which is a big loss. Calling this "alignment training" consistently would help reduce the watering down of the word "safety."

Machine Unlearning Evaluations as Interpretability Benchmarks

Dan H7mo81

I agree that this is an important frontier (and am doing a big project on this).

Broken Benchmark: MMLU

Dan H8mo2310

Almost all datasets have label noise. Most 4-way multiple choice NLP datasets collected with MTurk have ~10% label noise, very roughly. My guess is MMLU has 1-2%. I've seen these sorts of label noise posts/papers/videos come out for pretty much every major dataset (CIFAR, ImageNet, etc.).