All of Mark Xu's Comments + Replies

I was intending to warn about the possibility of future perception of corruption, e.g. after a non-existential AI catastrophe. I do not think anyone currently working at safety teams is percieved as that "corrupted", although I do think there is mild negative sentiment among some online communities (some parts of twitter, reddit, etc.).

Basically (2), very small amounts of (1) (perhaps qualitatively similar to the amount of (1) you would apply to e.g. people joining US AISI or UK AISI)

AI safety researchers might be allocated too heavily to Anthropic compared to Google Deepmind

Some considerations:

• Safety researchers should want Google Deepmind (GDM) to have a robust and flourishing safety department. It seems plausible that GDM will be able to create "the smartest" models: they have lots of talent, and own lots of computers. (see e.g. https://epochai.org/data/notable-ai-models#computing-capacity)
• Anthropic (ANT) might run into trouble in the future due to not owning their own computers, e.g. if Amazon (or where ever they're renting th

Hiliariously, it seems likely that our disagreement is even more meta, on the question of "how do you know when you have enough information to know", or potentially even higher, e.g. "how much uncertainty should one have given that they think they know" etc.

see my longer comment https://www.lesswrong.com/posts/A79wykDjr4pcYy9K7/mark-xu-s-shortform#8qjN3Mb8xmJxx59ZG

I think I disagree with your model of importance. If your goal is the make a sum of numbers small, then you want to focus your efforts where the derivative is lowest (highest? signs are hard), not where the absolute magnitude is highest.

The "epsilon fallacy" can be committed in both directions: both in that any negative dervative is worth working on, and that any extremely large number is worth taking a chance to try to improve.

I also seperately think that "bottleneck" is not generally a good term to apply to a complex project with high amounts of technica... (read more)

My vague plan along these lines is to attempt as hard as possible to defer all philosophically confusing questions to the "long reflection", and to use AI control as a tool to help produce AIs that can help preserve long term option value (including philosophical option value) as best as possible.

I seperately have hope we can solve "the entire problem" at some point, e.g. through ARC's agenda (which I spend most of my time trying to derisk and advance).

yep agreed, I have a bunch of vague plans in this direction. I most generally think that AI control is a pretty good tool in the toolbox, and is unlikely to make things much worse but plausibly makes things much better.

I agree it is better work on bottlenecks than non-bottlenecks. I have high uncertainty about where such bottlenecks will be, and I think sufficiently low amounts of work have gone into "control" that it's obviously worth investing more, because e.g. I think it'll let us get more data on where bottlenecks are.

Alignment researchers should think hard about switching to working on AI Control

I think Redwood Research’s recent work on AI control really “hits it out of the park”, and they have identified a tractable and neglected intervention that can make AI go a lot better. Obviously we should shift labor until the marginal unit of research in either area decreases P(doom) by the same amount. I think that implies lots of alignment researchers should shift to AI control type work, and would naively guess that the equilibrium is close to 50/50 across people who a... (read more)

Yes I agree with what you have written, and do think it’s overall not that likely that everything pans out as hoped. We do also have other hopes for how this general picture can still cohere if the specific path doesn’t work out, eg we’re open to learning some stuff empirically and adding an “algorithmic cherry on top” to produce the estimate.

The literature review is very strange to me. Where is the section on certified robustness against epsilon-ball adversarial examples? The techniques used in that literature (e.g. interval propagation) are nearly identical to what you discuss here.

I was meaning to include such a section, but forgot :). Perhaps I will edit it in. I think such work is qualitatively similar to what we're trying to do, but that the key difference is that we're interested in "best guess" estimates, as opposed to formally verified-to-be-correct estimates (mostly because we don'... (read more)

yes, you would need the catastrophe detector to be reasonably robust. Although I think it's fine if e.g. you have at least 1/million chance of catching any particular catastrophe.

I think there is a gap, but that the gap is probably not that bad (for "worst case" tail risk estimation). That is maybe because I think being able to do estimation through a single forward pass is likely already to be very hard, and to require being able to do "abstractions" over the concepts being manipulated by the forward pass. CoT seems like it will require vaguely similar struction of a qualitatively similar kind.

I think catastrophe detectors in practice will be composed of neural networks interacting with other stuff, like scientific literature, python, etc.

With respect to the stuff quoted, I think all but "doing experiments" can be done with a neural net doing chain of thought (although not making claims about quality).

I think we're trying to solve a different problem than trusted monitoring, but I'm not that knowledgeable about what issues trusted monitoring is trying to solve. The main thing that I don't think you can do with monitoring is producing a model tha... (read more)

if you train on (x, f(x)) pairs, and you ask it to predict f(x') on some novel input x', and also to write down what it thinks f is, do you know if these answers will be consistent? For instance, the model could get f wrong, and also give the wrong prediction for f(x), but it would be interesting if the prediction for f(x) was "connected" to it's sense of what f was.

Here are some things I think you can do:

• Train a model to be really dumb unless I prepend a random secret string. The goverment doesn't have this string, so I'll be able to predict my model and pass their eval. Some precedent in: https://en.wikipedia.org/wiki/Volkswagen_emissions_scandal

• I can predict a single matrix multiply just by memorizing the weights, and I can predict ReLU, and I'm allowed to use helper AIs.

• I just train really really hard on imitating 1 particular individual, then have them just say whatever first comes to mind.

You have to specify your backdoor defense before the attacker picks which input to backdoor.

I think competitiveness matters a lot even if there's only moderate amounts of competitive pressure. The gaps in efficiency I'm imagining are less "10x worse" and more like "I only had support vector machines and you had SGD"

humans, despite being fully general, have vastly varying ability to do various tasks, e.g. they're much better at climbing mountains than playing GO it seems. Humans also routinely construct entirely technology bases to enable them to do tasks that they cannot do themselves. This is, in some sense, a core human economic activity: the construction of artifacts that can do tasks better/faster/more efficiently than humans can do themselves. It seems like by default, you should expect a similar dynamic with "fully general" AIs. That is, AIs trained to do semic... (read more)

Not literally the best, but retargetable algorithms are on the far end of the spectrum of "fully specialized" to "fully general", and I expect most tasks we train AIs to do to have heuristics that enable solving the tasks much faster than "fully general" algorithms, so there's decently strong pressure to be towards the "specialized" side.

I also think that heuristics are going to be closer to multiplicative speed ups than additive, so it's going to be closer to "general algorithms just can't compete" than "it's just a little worse". E.g. random search is te... (read more)

One of the main reasons I expect this to not work is because optimization algorithms that are the best at optimizing some objective given a fixed compute budget seem like they basically can't be generally-retargetable. E.g. if you consider something like stockfish, it's a combination of search (which is retargetable), sped up by a series of very specialized heuristics that only work for winning. If you wanted to retarget stockfish to "maximize the max number of pawns you ever have" you had, you would not be able to use [specialized for telling whether a mo... (read more)

Flagging that I don't think your description of what ELK is trying to do is that accurate, e.g. we explicitly don't think that you can rely on using ELK to ask your AI if it's being deceptive, because it might just not know. In general, we're currently quite comfortable with not understanding a lot of what our AI is "thinking", as long as we can get answers to a particular set of "narrow" questions we think is sufficient to determine how good the consequences of an action are. More in “Narrow” elicitation and why it might be sufficient.

Separately, I think ... (read more)

The humans presumably have access to the documents being summarized.

From my perspective, ELK is currently very much "A problem we don't know how to solve, where we think rapid progress is being made (as we're still building out the example-counterexample graph, and are optimistic that we'll find an example without counterexamples)" There's some question of what "rapid" means, but I think we're on track for what we wrote in the ELK doc: "we're optimistic that within a year we will have made significant progress either towards a solution or towards a clear sense of why the problem is hard."

We've spent ~9 months on the proble... (read more)

The official deadline for submissions is "before I check my email on the 16th", which I tend to do around 10 am PST.

Before I check my email on Feb 16th, which I will do around 10am PST.

The high-level reason is that the 1e12N model is not that much better at prediction than the 2N model. You can correct for most of the correlation even with only a vague guess at how different the AI and human probabilities are, and most AI and human probabilities aren't going to be that different in a way that produces a correlation the human finds suspicious. I think that the largest correlations are going to be produced by the places the AI and the human have the biggest differences in probabilities, which are likely also going to be the places where th... (read more)

I agree that i does slightly worse than t on consistency checks, but i also does better on other regularizers you're (maybe implicitly) using like speed/simplicity, so as long as i doesn't do too much worse it'll still beat out the direct translator.

One possible thing you might try is some sort of lexicographic ordering of regularization losses. I think this rapidly runs into other issues with consistency checks, like the fact that the human is going to be systematically wrong about some correlations, so i potentially is more consistent than t.

I feel mostly confused by the way that things are being framed. ELK is about the human asking for various poly-sized fragments and the model reporting what those actually were instead of inventing something else. The model should accurately report all poly-sized fragments the human knows how to ask for.

Like the thing that seems weird to me here is that you can't simultaneously require that the elicited knowledge be 'relevant' and 'comprehensible' and also cover these sorts of obfuscated debate like scenarios.

I don't know what you mean by "relevant" or ... (read more)

I don’t think I understand your distinction between obfuscated and non-obfuscated knowledge. I generally think of non-obfuscated knowledge as NP or PSPACE. The human judgement of a situation might only theoretically require a poly sized fragment of a exp sized computation, but there’s no poly sized proof that this poly sized fragment is the correct fragment, and there are different poly sized fragments for which the human will evaluate differently, so I think of ELK as trying to elicit obfuscated knowledge.

We would prefer submissions be private until February 15th.

We generally assume that we can construct questions sufficiently well that there's only one unambiguous interpretation. We also generally assume that the predictor "knows" which world it's in because it can predict how humans would respond to hypothetical questions about various situations involving diamonds and sensors and that humans would say in theory Q1 and Q2 could be different.

More concretely, our standard for judging proposals is exhibiting an unambiguous failure. If it was plausible you asked the wrong question, or the AI didn't know what you mean... (read more)

I think we would be trying to elicit obfuscated knowledge in ELK. In our examples, you can imagine that the predictor's Bayes net works "just because", so an argument that is convincing to a human for why the diamond in the room has to be arguing that the Bayes net is a good explanation of reality + arguing that it implies the diamond is in the room, which is the sort of "obfuscated" knowledge that debate can't really handle.

The dataset is generated with the human bayes net, so it's sufficient to map to the human bayes net. There is, of course, an infinite set of "human" simulators that use slightly different bayes nets that give the same answers on the training set.

Does this mean that the method needs to work for ~arbitrary architectures, and that the solution must use substantially the same architecture as the original?

Yes, approximately. If you can do it for only e.g. transformers, but not other things, that would be interesting.

Does this mean that it must be able to deal with a broad variety of questions, so that we cannot simply sit down and think about how to optimize the model for getting a single question (e.g. "Where is the diamond?") right?

Yes, approximately. Thinking about how to get one question rig... (read more)

We generally imagine that it’s impossible to map the predictors net directly to an answer because the predictor is thinking in terms of different concepts, so it has to map to the humans nodes first in order to answer human questions about diamonds and such.

The SmartFabricator seems basically the same. In the robber example, you might imagine the SmartVault is the one that puts up the screen to conceal the fact that it let the diamond get stolen.

Looks good to me.

Yes. Section Strategy: have a human operate the SmartVault and ask them what happened describes what I think you're asking about.

A different way of phrasing Ajeya's response, which I think is roughly accurate, is that if you have a reporter that gives consistent answers to questions, you've learned a fact about the predictor, namely "the predictor was such that when it was paired with this reporter it gave consistent answers to questions." if there were 8 predictor for which this fact was true then "it's the [7th] predictor such that when it was paired with this reporter it gave consistent answers to questions" is enough information to uniquely determine the reporter, e.g. the previ... (read more)

There is a distinction between the way that the predictor is reasoning and the way that the reporter works. Generally, we imagine that that the predictor is trained the same way the "unaligned benchmark" we're trying to compare to is trained, and the reporter is the thing that we add onto that to "align" it (perhaps by only training another head on the model, perhaps by finetuning). Hopefully, the cost of training the reporter is small compared to the cost of the predictor (maybe like 10% or something)

In this frame, doing anything to train the way the pred... (read more)

I think that problem 1 and problem 2 as you describe them are potentially talking about the same phenomenon. I'm not sure I'm understanding correctly, but I think I would make the following claims:

• Our notion of narrowness is that we are interested in solving the problem where the question we're asking is such that a state always resolves a question. E.g. there isn't any ambiguity around whether a state "really contains a diamond". (Note that there is ambiguity around whether the human could detect the diamond from any set of observations because there co

My point is either that:

• it will always be possible to find such an experiment for any action, even desirable ones, because the AI will have defended the diamond in a way the human didn't understand or the AI will have deduced some property of diamonds that humans thought they didn't have
• or there will be some tampering for which it's impossible to find an experiment, because in order to avoid the above problem, you will have to restrict the space of experiments

Thanks for your proposal! I'm not sure I understand how the "human is happy with experiment" part is supposed to work. Here are some thoughts:

• Eventually, it will always be possible to find experiments where the human confidently predicts wrongly. Situations I have in mind are ones where your AI understands the world far better than you, so can predict that e.g. combining these 1000 chemicals will produce self-replicating protein assemblages, whereas the human's best guess is going to be "combining 1000 random chemicals doesn't do anything"
• If the human

We don't think that real humans are likely to be using Bayes nets to model the world. We make this assumption for much the same reasons that we assume models use Bayes nets, namely that it's a test case where we have a good sense of what we want a solution to ELK to look like. We think the arguments given in the report will basically extend to more realistic models of how humans reason (or rather, we aren't aware of a concrete model of how humans reason for which the arguments don't apply).

If you think there's a specific part of the report where the human Bayes net assumption seems crucial, I'd be happy to try to give a more general form of the argument in question.

Agreed, but the thing you want to use this for isn’t simulating a long reflection, which will fail (in the worst case) because HCH can’t do certain types of learning efficiently.

I want to flag that HCH was never intended to simulate a long reflection. It’s main purpose (which it fails in the worse case) is to let humans be epistemically competitive with the systems you’re trying to train.

The way that you would think about NN anchors in my model (caveat that this isn't my whole model):

• You have some distribution over 2020-FLOPS-equivalent that TAI needs.
• Algorithmic progress means that 20XX-FLOPS convert to 2020-FLOPS-equivalent at some 1:N ratio.
• The function from 20XX to the 1:N ratio is relatively predictable, e.g. a "smooth" exponential with respect to time.
• Therefore, even though current algorithms will hit DMR, the transition to the next algorithm that has less DMR is also predictably going to be some constant ratio better at convert

My model is something like:

• For any given algorithm, e.g. SVMs, AlphaGo, alpha-beta pruning, convnets, etc., there is an "effective compute regime" where dumping more compute makes them better. If you go above this regime, you get steep diminishing marginal returns.
• In the (relatively small) regimes of old algorithms, new algorithms and old algorithms perform similarly. E.g. with small amounts of compute, using AlphaGo instead of alpha-beta pruning doesn't get you that much better performance than like an OOM of compute (I have no idea if this is true, ex