Glad it was helpful!

Hi Alex,

Let me first acknowledge that your write-up is significantly more thorough than pretty much all content on LessWrong, and that I found the particular examples interesting. I also appreciated that you included a related work section in your write-up. The reason I commented on this post and not others is because it's one of the few ML posts on LessWrong that seemed like it might teach me something, and I wish I had made that more clear before posting critical feedback (I was thinking of the feedback as directed at Oliver / Raemon's moderation norms, rather than your work, but I realize in retrospect it probably felt directed at you).

I think the main important point is that there is a body of related work in the ML literature that explores fairly similar ideas, and LessWrong readers who care about AI alignment should be aware of this work, and that most LessWrong readers who read the post won't realize this. I think it's good to point out Dan's initial mistake, but I took his substantive point to be what I just summarized, and it seems correct to me and hasn't been addressed. (I also think Dan overfocused on Ludwig's paper, see below for more of my take on related work.)

Here is how I currently see the paper situated in broader work (I think you do discuss the majority but not all of this):

 * There is a lot of work studying activation vectors in computer vision models, and the methods here seem broadly similar to the methods there. This seems like the closest point of comparison.

 * In language, there's a bunch of work on controllable generation (https://arxiv.org/pdf/2201.05337.pdf) where I would be surprised if no one looked at modifying activations (at least I'd expect someone to try soft prompt tuning), but I don't know for sure.

 * On modifying activations in language models there is a bunch of stuff on patching / swapping, and on modifying stuff in the directions of probes.

I think we would probably both agree that this is the main set of related papers, and also both agree that you cited work within each of these branches (except maybe the second one). Where we differ is that I see all of this as basically variations on the same idea of modifying the activations or weights to control a model's runtime behavior:
 * You need to find a direction, which you can do either by learning a direction or by simple averaging. Simple averaging is more or less the same as one step of gradient descent, so I see these as conceptually similar.
 * You can modify the activations or weights. Usually if an idea works in one case it works in the other case, so I also see these as similar.
 * The modality can be language or vision. Most prior work has been on vision models, but some of that has also been on vision-language models, e.g. I'm pretty sure there's a paper on averaging together CLIP activations to get controllable generation.

So I think it's most accurate to say that you've adapted some well-explored ideas to a use case that you are particularly interested in. However, the post uses language like "Activation additions are a new way of interacting with LLMs", which seems to be claiming that this is entirely new and unexplored, and I think this could mislead readers, as for instance Thomas Kwa's response seems to suggest.

I also felt like Dan H brought up reasonable questions (e.g. why should we believe that weights vs. activations is a big deal? Why is fine-tuning vs. averaging important? Have you tried testing the difference empirically?) that haven't been answered that would be good to at least more clearly acknowledge. The fact that he was bringing up points that seemed good to me that were not being directly engaged with was what most bothered me about the exchange above.

This is my best attempt to explain where I'm coming from in about an hour of work (spent e.g. reading through things and trying to articulate intuitions in LW-friendly terms). I don't think it captures my full intuitions or the full reasons I bounced off the related work section, but hopefully it's helpful.

I'll just note that I, like Dan H, find it pretty hard to engage with this post because I can't tell whether it's basically the same as the Ludwig Schmidt paper (my current assumption is that it is). The paragraph the authors added didn't really help in this regard.

I'm not sure what you mean about whether the post was "missing something important", but I do think that you should be pretty worried about LessWrong's collective epistemics that Dan H is the only one bringing this important point up, and that rather than being rewarded for doing so or engaged with on his substantive point, he's being nitpicked by a moderator. It's not an accident that no one else is bringing these points up--it's because everyone else who has the expertise to do so has given up or judged it not worth their time, largely because of responses like the one Dan H is getting.

Yup, I agree with this, and think the argument generalizes to most alignment work (which is why I'm relatively optimistic about our chances compared to some other people, e.g. something like 85% p(success), mostly because most things one can think of doing will probably be done).

It's possibly an argument that work is most valuable in cases of unexpectedly short timelines, although I'm not sure how much weight I actually place on that.

Yup! That sounds great :)

Thanks Ruby! Now that the other posts are out, would it be easy to forward-link them (by adding links to the italicized titles in the list at the end)?

Finding the min-max solution might be easier, but what we actually care about is an acceptable solution. My point is that the min-max solution, in most cases, will be unacceptably bad.

And in fact, since min_x f(theta,x) <= E_x[f(theta,x)], any solution that is acceptable in the worst case is also acceptable in the average case.

Thanks! I appreciated these distinctions. The worst-case argument for modularity came up in a past argument I had with Eliezer, where I argued that this was a reason for randomization (even though Bayesian decision theory implies you should never randomize). See section 2 here: The Power of Noise.

Re: 50% vs. 10% vs. 90%. I liked this illustration, although I don't think your argument actually implies 50% specifically. For instance if it turns out that everyone else is working on the 50% worlds and no one is working on the 90% worlds, you should probably work on the 90% worlds. In addition:

 *  It seems pretty plausible that the problem is overall more tractable in 10% worlds than 50% worlds, so given equal neglectedness you would prefer the 10% world.

 * Many ideas will generalize across worlds, and recruitment / skill-building / organization-building also generalizes across worlds. This is an argument towards working on problems that seem tractable and relevant to any world, as long as they are neglected enough that you are building out distinct ideas and organizational capacity (vs. just picking from the same tree as ML generally). I don't think that this argument dominates considerations, but it likely explains some of our differences in approach.

In the terms laid out in your post, I think my biggest functional disagreement (in terms of how it affects what problems we work on) is that I expect most worst-case assumptions make the problem entirely impossible, and I am more optimistic that many empirically-grounded assumptions will generalize quite far, all the way to AGI. To be clear, I am not against all worst-case assumptions (for instance my entire PhD thesis is about this) but I do think they are usually a source of significant added difficulty and one has to be fairly careful where they are making them.

For instance, as regards Redwood's project, I expect making language models fully adversarially robust is impossible with currently accessible techniques, and that even a fairly restricted adversary will be impossible to defend against while maintaining good test accuracy. On the other hand I am still pretty excited about Redwood's project because I think you will learn interesting things by trying. (I spent some time trying to solve the unrestricted adversarial example competition, totally failed, but still felt it was a good use of time for similar reasons, and the difficulties for language models seem interestingly distinct in a way that should generate additional insight.) I'm actually not sure if this differs that much from your beliefs, though.

I think this probably depends on the field. In machine learning, solving problems under worst-case assumptions is usually impossible because of the no free lunch theorem. You might assume that a particular facet of the environment is worst-case, which is a totally fine thing to do, but I don't think it's correct to call it the "second-simplest solution", since there are many choices of what facet of the environment is worst-case.

One keyword for this is "partial specification", e.g. here is a paper I wrote that makes a minimal set of statistical assumptions and worst-case assumptions everywhere else: https://arxiv.org/abs/1606.05313. (Unfortunately the statistical assumptions are not really reasonable so the method was way too brittle in practice.) This kind of idea is also common in robust statistics. But my take would not be that it is simpler--in general it is way harder than just working with the empirical distribution in front of you.

Cool paper! One brief comment is this seems closely related to performative prediction and it seems worth discussing the relationship.

Edit: just realized this is a review, not a new paper, so my comment is a bit less relevant. Although it does still seem like a useful connection to make.