Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

We are no longer accepting submissions. We'll get in touch with winners and make a post about winning proposals sometime in the next month.

ARC recently released a technical report on eliciting latent knowledge (ELK), the focus of our current research. Roughly speaking, the goal of ELK is to incentivize ML models to honestly answer “straightforward” questions where the right answer is unambiguous and known by the model. 

ELK is currently unsolved in the worst case—for every training strategy we’ve thought of so far, we can describe a case where an ML model trained with that strategy would give unambiguously bad answers to straightforward questions despite knowing better. Situations like this may or may not come up in practice, but nonetheless we are interested in finding a strategy for ELK for which we can’t think of any counterexample.

We think many people could potentially contribute to solving ELK—there’s a large space of possible training strategies and we’ve only explored a small fraction of them so far. Moreover, we think that trying to solve ELK in the worst case is a good way to “get into ARC’s headspace” and more deeply understand the research we do.

We are offering prizes of $5,000 to $50,000 for proposed strategies for ELK. We’re planning to evaluate submissions received before February 15.

For full details of the ELK problem and several examples of possible strategies, see the writeup. The rest of this post will focus on how the contest works.

Contest details

To win a prize, you need to specify a training strategy for ELK that handles all of the counterexamples that we’ve described so far, summarized in the section below—i.e. where the breaker would need to specify something new about the test case to cause the strategy to break down. You don’t need to fully solve the problem in the worst case to win a prize, you just need to come up with a strategy that requires a new counterexample.

We’ll give a $5,000 prize to any proposal that we think clears this bar. We’ll give a $50,000 prize to a proposal which we haven’t considered and seems sufficiently promising to us or requires a new idea to break. We’ll give intermediate prizes for ideas that we think are promising but we’ve already considered, as well as for proposals that come with novel counterexamples, clarify some other aspect of the problem, or are interesting in other ways. A major purpose of the contest is to provide support for people understanding the problem well enough to start contributing; we aren’t trying to only reward ideas that are new to us.

You can submit multiple proposals, but we won’t give you separate prizes for each—we’ll give you at least the maximum prize that your best single submission would have received, but may not give much more than that.

If we receive multiple submissions based on a similar idea, we may post a comment describing the idea (with attribution) along with a counterexample. Once a counterexample has been included in the comments of this post, new submissions need to address that counterexample (as well as all the existing ones) in order to be eligible for a prize. 

Ultimately prizes are awarded at our discretion, and the “rules of the game” aren’t fully precise. If you are curious about whether you are on the right track, feel free to send an email to with the basic outline of an idea, and if we have time we’ll get back to you with some feedback. Below we also describe some of the directions we consider more and less promising and some general guidance.

How to submit a proposal

You can submit a proposal by copying this google doc template and sharing it with (please give comment access in case we need to ask questions to evaluate your submission). By submitting a proposal, you are giving us permission to post the idea here with attribution. (And it's fine for you to post it after the contest or we post a counterexample.)

Retroactive prizes

We’re giving prizes to existing proposals from David Dalrymple ($5k), Ramana Kumar ($3k), John Maxwell ($2k), and Yonadav Shavit ($1k). We’ve added relevant counterexamples in the next section, and further submissions need to handle those counterexamples in order to be eligible.

Existing counterexamples

In our report and a recent post, we described some features of the world that would cause existing approaches to ELK to fail. A new approach should work even in a situation satisfying all of these properties, or should include an argument for why some of these features are actually impossible or inconsistent with each other.

In order to generate a counterexample to a training strategy, we need to spell out a concrete “way the world could be” (a test case) where the strategy would fail. In the report we consider test cases where the AI and the human happen to understand the world by doing inference in Bayes nets. The Bayes net case is just one test case that lets us get precise about our counterexamples and our training strategies don’t lean on it. In practice we also consider other concrete cases (e.g. deduction from premises generated by a neural network).

In this section we’ll briefly review the current state of existing counterexamples. This isn’t intended to be a self-contained description; it relies heavily on links to the report.

Ontology mismatch

All of our counterexamples are based on an ontology mismatch between two different Bayes nets, one used by an ML prediction model (“the predictor”) and one used by a human:

  • The predictor’s Bayes net is used internally by a prediction model that maps an initial segment of a video to a prediction about what will happen next. This model was trained by gradient descent based on the quality of its predictions (more).
  • Humans have their own model of the world, which they can improve with the help of AI assistants; those assistants can design and conduct experiments, find and evaluate arguments, and so on. Crucially, we have a limited computational budget with which to train these AI assistants, and that limits how much help they can provide. The “human’s” Bayes net captures everything the human understands about the world at the end of this process (more).

We are interested in the case where there are features of reality that are accurately reflected in the predictor’s Bayes net, but missing from this (augmented) human Bayes net. We expect the human could understand these facts eventually given enough time or enough AI assistance, but not using the available computational budget. Put differently, there is a gap between the kinds of understanding that gradient descent can build, and the kind of understanding that a human-comprehensible process of “scientific inquiry” can build using similar computational resources (more).

Furthermore, we imagine that both the predictor’s Bayes net and the human’s Bayes net reflect some underlying reality and that there is a way to map between them, which we’ll call “direct translation.” We don’t know how to do direct translation, and we are concerned that if we simply search for any mapping from the AI’s Bayes net to the human’s Bayes net we will instead get the “human simulator”: the map which correctly translates the observations from one Bayes net to the other, but then does inference in the human’s Bayes net in order to fill in the latent structure as a human would by looking at the data (more).

Counterexample features

ELK seems particularly hard when these Bayes nets satisfy the following properties:

  1. There are arbitrarily severe “ontology mismatches” between the predictor’s Bayes net and the human’s Bayes net, such that specifying the direct translation between them can be very complex (more).
  2. The human’s Bayes net is simpler than the predictor’s Bayes net, and the gap can be arbitrarily large (more).
  3. The predictor’s Bayes net is arbitrarily large, and individual nodes of the human’s Bayes net depend on a constant fraction of the predictor’s Bayes net. Therefore directly translating between the two Bayes nets takes arbitrarily much computation (more).
  4. It’s computationally easy for the human simulator to do good inference in the human’s Bayes net—at least good enough inference that it answers questions as well as the direct translator—and the gap between this complexity and the complexity of direct translation can be arbitrarily large. More precisely, a human interrogator can’t think of questions which require the human simulator to do hard inference. However, we can’t rule out the possibility that you, the algorithm-designer, will think of some trick they can use to generate hard questions (more).
  5. There are situations where a model which is trying to answer a question will first be able to predict what (potentially mistaken) belief the human-with-assistance will have about that question before it knows the true answer to the question (more).
  6. While the predictor’s Bayes net is very large, it has a short succinct description which is simpler than the human’s Bayes net, and that gap is arbitrarily large. Think of simple physical laws that describe arbitrarily computationally expensive simulations, or simple cognitive heuristics that are very widely applicable. (more)
  7. There are sequences of actions for which the predictor’s Bayes net correctly anticipates that tampering will occur yet nothing looks suspicious to the human’s Bayes net.
  8. After tampering has occurred, it is possible that no sequence of further actions reveal that anything is wrong, e.g. the camera and actuators may have been completely disconnected from the world and replaced with a convincing-to-humans simulation (more).
  9. There are sequences of actions leading to undetectable tampering that have no human-comprehensible similarities to sequences of actions leading to detectable tampering; in particular, the early steps of this undetectable tampering cannot be used to perform detectable tampering (more).

An important and very general strategy for the breaker is to suggest counterexamples where you “get what you train for.” If you propose training a model using loss function L, then the breaker will likely focus on counterexamples where the model chooses its output to minimize L. A successful approach should either explain why this either isn’t a realistic thing for the model to learn (e.g. why that would be disfavored by regularization), or why the overall approach works even if the model is explicitly strategizing about how to minimize the loss.

Rough guidance

  • There is a tradeoff between specifying your approach precisely, and including your own counterexample. It’s fine to describe a very broad/vague idea for solving ELK, and then present a new counterexample showing why that whole category of solutions can’t work. But if you don’t propose a counterexample, then it’s good to describe your proposal precisely enough that we understand how to actually implement it without wondering if that’s really what you meant. It’s OK to submit a very broad or informal idea together with a single very specific instance of that idea, as long as there is some version we can understand precisely.
  • We suspect you can’t solve ELK just by getting better data—you probably need to “open up the black box” and include some term in the loss that depends on the structure of your model and not merely its behavior. So we are most interested in approaches that address that challenge. We could still be surprised by clever ways to penalize behavior, but we’ll hold them to a higher bar. The most plausible surprise would be finding a way to reliably make it computationally difficult to “game” the loss function, probably by using the AI itself to help compute the loss (e.g. using consistency checks or by giving the human AI assistance).
  • If you are specifying a regularizer that you hope will prefer direct translation over human simulation, you should probably have at least one concrete case in mind that has all the counterexample-features above and where you can confirm that your regularizer does indeed prefer the direct translator.
  • ELK already seems hard in the case of ontology identification, where the predictor uses a straightforward inference algorithm in an unknown model of the world (which we’ve been imagining as a Bayes net). When coming up with a proposal, we don’t recommend worrying about cases where the original unaligned predictor learned something more complicated (e.g. involving learned optimization other than inference). That said, you do need to worry about the case where your training scheme incentivizes learned optimization that may not have been there originally.

Ask dumb questions!

A major purpose of this contest is to help people build a better understanding of our research methodology and the “game” we are playing. So we encourage people to ask clarifying questions in the comments of this post (no matter how “dumb” they are), and we’ll do our best to answer all of them. You might also want to read the comments to get more clarity about the problem.

What you can expect from us

  • We’ll try to answer all clarifying questions in the comments.
  • If you send in a rough outline for a proposal, we will try to understand whether it might qualify and write back something like “This qualifies,” “This might qualify but would need to be clearer and address issue X,” “We aren’t easily able to understand this proposal at all,” “This is unlikely to be on track for something that qualifies,” or “This definitely doesn’t qualify.”
  • If there are more submissions than expected, we may run out of time to respond to all submissions and comments, in which case we will post an update here.

New to LessWrong?

New Comment
153 comments, sorted by Click to highlight new comments since: Today at 1:12 PM
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

After reading through the report I wanted to make sure I understood the scenarios and counterexamples being discussed and be able to quickly refresh my memory, so I attempted to write a brief summary. Figured I'd share it here in case it helps anyone else.

Roles and Terms

SmartVault: Vault with a diamond in it, operated by a superintelligent AI tasked with keeping the diamond safe.

Predictor: The primary AI tasked with protecting the diamond. The predictor sees a video feed of the vault, predicts what actions are necessary to protect the diamond and how those actions will play out (for example, activating a trap door to eliminate a robber trying to steal the diamond), and then generates a video showing precisely what will happen.

Reporter: A second "head" or component added to the AI which is used to map the AI's understanding of what's going on in the SmartVault into a format understandable by humans (such as by responding to questions).

Diagram of Predictor Showing Reporter from ELK Report

Humans: Observers who are training the AI to protect the diamond, but also want the AI to accurately report whether the diamond has been stolen or not. The humans may not be able to tell if the camer... (read more)

6Mark Xu2y
Looks good to me.
I'd like to try making a correction here, though I might make some mistakes too. The predictor is different from the AI that protects the diamond and doesn't try to "choose" actions in order to accomplish any particular goal. Rather, it takes a starting video and a set of actions as input, then returns a prediction of what the ending video would be if those actions were carried out. An agent could use this predictor to choose a set of actions that leads to videos that a human approves of, then carry out these plans. It could use some kind of search policy, like Monte-Carlo Tree Search, or even just enumerate through every possible action and figure out which one seems to be the best. For the purposes of this problem, we don't really care; we just care that we have a predictor that uses some model of the world (which might take the form of a Bayes net) to guess what the output video will be. Then, the reporter can use the model to answer any questions asked by the human.
1Ryan Beck2y
I think that makes sense. To rephrase, are you basically saying that the predictor is a subcomponent of the AI, like the reporter is? I didn't catch that distinction in the report but looking back at it I think you're right. But yeah doesn't seem like the distinction matters much for what we're doing.
It seems fair to call it a subcomponent, yeah

We’re planning to evaluate submissions as we receive them, between now and the end of January; we may end the contest earlier or later if we receive more or fewer submissions than we expect.


Just wanted to note that the "we may end the contest earlier" part here makes me significantly more hesitant about trying this. I will probably still at least have a look at it, but part of me is afraid that I'll invest a bunch of time and then the contest will be announced to be over before I got around to submitting. And I suspect Holden's endorsement may make that more likely. It would be easier for me to invest time spread out over the next couple of weeks, than all in one go, due to other commitments. On the other hand, if I knew there was a hard deadline next Friday, I might try to find a way to squeeze it in.

I'm just pointing this out in case you hadn't thought of it. I suspect something similar might be true for others too. Of course, it's your prize and your rules, and if you prefer it this way, that's totally fine.

We're going to accept submissions through February 10. (We actually ended up receiving more submissions than I expected but it seems valuable, and Mark has been handling all the reviews, so running for another 20 days seems worthwhile.)
1Mathieu Putz2y
Thanks! Great to hear that it's going well!
4Mark Xu2y
Note that this has changed to February 15th.

Here are a couple of hand-wavy "stub" proposals that I sent over to ARC, which they thought were broadly intended to be addressed by existing counterexamples. I'm posting them here so they can respond and clarify why these don't qualify.

*Proposal 1: force ontological compatibility*

On page 34 of the ELK gdoc, the authors talk about the possibility that training an AI hard enough produces a model that has deep mismatches with human ontology - that is, it has a distinct "vocabulary of basic concepts" (or nodes in a Bayes net) that are distinct from the ones humans can build understanding of (via doing science on the compute budget available). Because of this, even AI assistance can't help humans understand everything the SmartVault AI understands. This is central to the challenge that most of the writeup is contending with - if not for the mismatch, "AIs explaining things to humans" could ensure that the trickery we're worried about doesn't happen.

The proposal here is to include a term in the loss function that incentivizes the AI to have a human-compatible ontology. For a cartoonish example, imagine that the term works this way: "The AI model gets a higher score to the degree that pe... (read more)

Again trying to answer this one despite not feeling fully solid. I'm not sure about the second proposal and might come back to it, but here's my response to the first proposal (force ontological compatibility):

The counterexample "Gradient descent is more efficient than science" should cover this proposal because it implies that the proposal is uncompetitive. Basically, the best Bayes net for making predictions could just turn out to be the super incomprehensible one found by unrestricted gradient descent, so if you force ontological compatibility then you could just end up with a less-good prediction model and get outcompeted by someone who didn't do that. This might work in practice if the competitiveness hit is not that big and we coordinate around not doing the scarier thing (MIRI's visible thoughts project is going for something like this), but ARC isn't looking for a solution of that form.

I'm not sure why this isn't a very general counterexample. Once we've decided that the human imitator is simpler and faster to compute, don't all further approaches (e.g., penalizing inconsistency) involve a competitiveness hit along these general lines? Aren't they basically designed to drag the AI away from a fast, simple human imitator toward a slow, complex reporter? If so, why is that better than dragging the AI from a foreign ontology toward a familiar ontology?
5Mark Xu2y
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 predictor is trained results in a big competitiveness hit, e.g. forcing the predictor to use the same ontology as a human is potentially going to prevent it from using concepts that make reasoning much more efficient. However, training the reporter in a different way, e.g. doubling the cost of training the reporter, only takes you from 10% of the predictor to 20%, which not that bad of a competitiveness hit (assuming that the human imitator takes 10% of the cost of the original predictor to train). In summary, competitiveness for ELK proposals primarily means that you can't change the way the predictor was trained. We are already assuming/hoping the reporter is much cheaper to train than the predictor, so making the reporter harder to train results in a much smaller competitiveness hit.
I think that a lot depends on what kind of term you include. If you just say "find more interesting things" then the model will just have a bunch of neurons designed to look interesting. Presumably you want them to be connected in some way to the computation, but we don't really have any candidates for defining that in a way that does what you want. In some sense I think if the digital neuroscientists are good enough at their job / have a good enough set of definitions, then this proposal might work. But I think that the magic is mostly being done in the step where we make a lot of interpretability progress, and so if we define a concrete version of interpretability right now it will be easy to construct counterexamples (even if we define it in terms of human judgments). If we are just relying on the digital neuroscientists to think of something clever, the counterexample will involve something like "they don't think of anything clever." In general I'd be happy to talk about concrete proposals along these lines. (I agree with Ajeya and Mark that the hard case for this kind of method is when the most efficient way of thinking is totally alien to the human. I think that can happen, and in that case in order to be competitive you basically just need to learn an "interpreted" version of the alien model. That is, you need to basically show that if there exists an alien model with performance X, there is a human-comprehensible model with performance X, and the only way you'll be able to argue that for any model we can define a human-comprehensible model with similar complexity and the same behavior.)

tl;dr as of 18/2/2022
The goal is to educate me and maybe others. I make some statements, you tell me how wrong I am (please).

After input from P. (many thanks) and an article by Paul Christiano this statement stands yet uncorrected:

In the worst case, the internal state of the predictor is highly correlated within itself and multiple mappings with zero loss from the internal state to the desired extraction of information exist. The only solution is to work with some prior belief about how the internal state maps to the desired information. But as by design of the contest, this is not possible as (in the worst case) a human cannot interpret the internal state nor can he interpret complex actions (and so cannot reason about it and/or form a prior belief). The solution to this second problem is to learn a prior from a smaller human-readable dataset, for example simple information as a function of simple actions, and apply it to (or force it upon)  our reporter (as described by the mentioned article).

To my eyes this implies that there is a counterexample to all of the following types of proposal:
1) Datasets including only actions, predictions, internal states and desired information... (read more)

The Markov property doesn't imply that we can't determine what variable we care about using some kind of "correlation". Some part of the information in some node in the chain might disappear when computing the next node, so we might be able to distinguish it from its successors. And it might also have been gained when randomly computing its value from the previous node, so it might be possible to distinguish it from its predecessors. In the worst case scenario where all variables are in fact correlated to G what we need to do is to use a strong prior so that it prefers the correct computational graph over the wrong ones. This might be hard but it isn't impossible. But you can also try to create a dataset that makes the problem easier to solve, or train a wrong reporter and only reply when the predictions made when using each node are the same so we don't care what node it actually uses (as long as it can use the nodes properly, instead of computing other node and using it to get the answer, or something like that).
Thank you very much for your reply! I'll concede that the markov property does not make all nodes indistinguishable. I'll go further and say that not all algorithm's have to have the markov property. A google-search learned me that an RNN breaks the markov property. But then again, we are dealing with the worst-case-game, so with our luck, it'll probably be some highly correlated thing. You suggest using some strong prior belief. I assume you mean a prior belief about I or about I -> G? I thought, but correct me if I'm wrong, that the opaqueness of the internal state of the complex AI would mean that we can have no meaningfull prior belief about the internal state. So that would rule out a prior belief about (the hyperparameters of) our reporter I -> G. Or am I wrong? We can however have a strong idea about A -> G, as per example of the 'human operator' and use that as our training data. But that falls with the counterexample given in the report, when the distribution shifts from simple to complex.
RNNs break the Markov property in the sense that they depend on more than just the previous element in the sequence they are modelling. But I don't see why that would be relevant to ELK. When I say that a strong prior is needed I mean the same thing that Paul means when he writes: "We suspect you can’t solve ELK just by getting better data—you probably need to 'open up the black box' and include some term in the loss that depends on the structure of your model and not merely its behaviour.". Which is a very broad class of strategies. I also don't understand what you mean by having a strong idea about A->G, we of course have pairs of [A, G] in our training data but what we need to know is how to compute G from A given these pairs.
Updating my first line of thought You're right in that RNNs don't have anything to do with ELK, but I came back to it because the Markov property was part of the lead up to saying that all parts of I are correlated.  So with your help, I have to change my reasoning to: Correct? Than I can update my first statement to If I'm wrong, do let me know! Updating my second line of thought Ah yes, I understand now. This relates to my second line of thought. I reasoned that the reporter could learn any causal graph. I said we had no way of knowing which.  Because of your help, I need to update that to:  Which was in the opening text all along... But this leads me to the question:  My analogy would be: If I don't know where I am, how can I reason about getting home? And -if you'll humor me- my follow up statement would be:   Again, If I'm wrong: let me know! I'm learning a lot already.   Irrelevant side note: I saw you using the term computational graph. I chose the term causal graph, because I liked it being closer to the ground truth. Besides, a causal graph learned by some algorithm need not be exactly the same as it's computational graph. And then I chose such simple examples that they were equal again. Stupid me.
As before I am behind the curve. Above I concluded saying that I can form no prior belief about G as a function of I. I cannot, but we can learn a function to create our prior. Paul Christiano already wrote  an article about learning the prior ( So in conclusion, in the worst case no single function mapping I to G exists, as there are multiple reducing down to either camp translator or camp human-imitator. Without context we can form no strong prior due to the complexity of A and I, but as Paul described in his article we can learn a prior from for example in our case the dataset containing G as a function of A. I'll add a tl;dr in my first post to shorten the read about how I slowly caught up to everyone else. Corrections are of course still welcome!

Question: Does ARC consider ELK-unlimited to be solved, where ELK-unlimited is ELK without the competitiveness restriction (computational resource requirements comparable to the unaligned benchmark)?

One might suppose that the "have AI help humans improve our understanding" strategy is a solution to ELK-unlimited because its counterexample in the report relies on the competitiveness requirement. However, there may still be other counterexamples that were less straightforward to formulate or explain.

I'm asking for clarification of this point because I notice... (read more)

My guess is that "help humans improve their understanding" doesn't work anyway, at least not without a lot of work, but it's less obvious and the counterexamples get weirder. It's less clear whether ELK is a less natural subproblem for the unlimited version of the problem. That is, if you try to rely on something like "human deliberation scaled up" to solve ELK, you probably just have to solve the whole (unlimited) problem along the way. It seems to me like the core troubles with this point are: * You still have finite training data, and we don't have a scheme for collecting it. This can result in inner alignment problems (and it's not clear those can be distinguished from other problems, e.g. you can't avoid them with a low-stakes assumption). * It's not clear that HCH ever figures out all the science, no matter how much time the humans spend (and having a guarantee that you eventually figure everything out seems seems kind of close to ELK, where the "have AI help humans improve our understanding" is to some extent just punting to the humans+AI to figure out something). * Even if HCH were to work well it will probably be overtaken by internal consequentialists, and I'm not sure how to address that without competitiveness. (Though you may need a weaker form of competitiveness.) I'm generally interested in crisper counterexamples since those are a bit of a mess.

Apologies for a possibly naive comment/question, perhaps this has been discussed elsewhere and you can just direct me there.  But anyway...

I would find it helpful to see a strategy that ARC believes does in fact solve ELK, but fails only because it requires taking an unacceptably large capabilities hit.  I would find this helpful for several reasons, namely 

(1) it would help me to understand what kinds of strategies you believe really do escape counter-examples, 
(2) it would give me a better sense for how optimistic to be about the appr... (read more)

If you don't care about a capabilities hit, I think the salient strategy is training your model to predict human predictions rather than to predict reality. You can still do science+debate+etc. in order to improve those predictions. If you care about getting superhuman capabilities (and going beyond recursive schemes etc.) then I don't know if there's any easy way to "merely" pay a big capabilities hit. Certainly I don't know how to e.g. solve the problem in a way that's merely very computationally expensive (and that does sound like it would be major progress towards a solution, I'd guess it would mean you are most of the way there).
3Jared Kaplan2y
Thanks, yeah I meant that I was interested in a solution that would scale to arbitrarily superhuman AI capabilities with a "mere" capabilities hit/cost (perhaps a very large cost that grows with AI capability, but does not impose a bound on the ultimate capability of the aligned system).  So this was a useful clarification for me in terms of understanding your perspective; I may be wrong but I could imagine it might be useful to lead with this a bit more, ie "we don't know of and would be very interested in solutions that might be extremely costly but that avoid all counter-examples".  Possibly you already say this and I just missed it.
It seems like recursive schemes can potentially scale arbitrarily far (and at least up to the analog of "NEXP", but probably farther), they are mostly just limited by the capability of the AI assistants / debaters / etc. So it's kind of hard to distinguish mere capabilities costs from bounds on the ultimate capability. We could exclude that kind of thing because we have no idea what the bound is (or perhaps humans just never discover some facts that gradient descent discovers, or that they discover them in a way that causes them to run into the same problem). I think in that case the problem is still open. For example, finding a solution that definitely runs in 2n more time than the unaligned benchmark looks hard, I'd guess it's roughly as hard as finding a solution that definitely runs in 10 times more time than the unaligned benchmark. The main reasons we decided not to emphasize with this, and to focus as much as we do on competitiveness issues, is (i) the "do science" options do feel like they work if you have enough compute and it seems like you need to emphasize the competitiveness issue to explain why we don't like them (or else get into increasingly weird counterexamples), (ii) in the worst case we don't expect a very slow solution to be much easier than a very fast solution, since most realistic kinds of slowdown can get arbitrarily bad in the worst case, and the plausible approaches we are aware of all seem pretty likely be roughly competitive. So it seems likely to set people off down weirder alleys (which would be good for someone to go down if lots of folks are working on the problem but probably aren't where you should start).
1Jared Kaplan2y
There's a direction (which I imagine you and others have considered) where you replace some activations within your AI with natural language, so that eg heuristically certain layers can only communicate with the next layer in NL. Then you heavily regularize in various ways.  You'd require the language to be fully understandable and transparent, perhaps requiring that counter-factual changes to inputs lead to sensible changes to outputs within subsystems, etc.  You'd have humans verify the language was relevant, meaningful, & concise, train AIs to do this verification at larger scale, do some adversarial training, etc.  You could also train sub-human level AIs to paraphrase the language that's used and restate it between layers, to make it really hard for the whole system to ever pass hidden coded messages.   This seems like it lives under a slogan like "enforce interpretability at any cost".  This would almost certainly incur a big efficiency/capabilities hit.  Maybe it's enormous.  Though it actually seems plausible that the hit would be much smaller for extremely capable systems, as compared to the AI models of today. A crucial question will then be "how powerful are the subsystems that talk to each other via natural language allowed to get", where in the most conservative limit each subsystem is human level, or even significantly below, and in the riskiest limit you just have a single NL layer that cuts the system in half. There's a worry along the lines of "maybe the whole system is so big and complex it has emergent bad and inscrutable behavior even though every step is interpretable and makes sense".  Or in the same vein "the answers to simple big-picture questions we care about don't live anywhere specific, so this doesn't help us to ensure the model can transparently address them, even if its operation itself can be broken down into transparent pieces."  That said, I think we're in a better position wrt these issues, as we can now talk about training mod
I think there's a real fork in the road between: 1. You replace parts of your neural network with natural language, optimize those parts to implement a good process, and then hope the outcome is good because the process is good. 2. You replace parts of your neural network with natural language, and then optimize that natural language to achieve good outcomes. I think that #1 is safe and runs into significant capability limitations (roughly the same as debate/amplification). It may still be good enough to carry the day if things work out well or if people are willing to exercise a lot of restraint, and I'd like to see people doing it. I think that in this it doesn't matter that much how powerful the subsystems are, since each of them is doing something that you understand (though there are many subtleties and possible problems, e.g. with emergent bad behavior and some inner alignment problems). I think that by default #2 is pretty dangerous. If you took this route I don't think it would be fair to call the bad/inscrutable behavior "emergent," or to call each step "interpretable"---the steps make sense but by default it seems extremely likely that you don't understand why the process leads to good results. (If you did, you could have just taken path #1.) If there is bad behavior it's not emergent it's just produced directly by gradient descent, and the fact that you can encode the intermediate activations in natural language doesn't really address the risk (if that information isn't necessarily functioning in the way you expect). I feel like different versions of path #2 sit on a spectrum between "fairly safe like path #1" and "clearly unworkably dangerous." I feel most comfortable basically starting from path #1 and then carefully adding in stuff you don't understand (e.g. systems solving small subtasks in ways you don't understand, or optimizing only a small number of degrees within a space you understand reasonably well). You could instead start with "very s

Maybe I'm being stupid here. On page 42 of the write-up, it says:

In order to ensure we learned the human simulator, we would need to change the training strategy to ensure that it contains sufficiently challenging inference problems, and that doing direct translation was a cost-effective way to improve speed (i.e. that there aren’t other changes to the human simulator that would save even more time). [emphasis mine]

Shouldn't that be?

In order to ensure we learned the direct translator, ...

Turning this into the typo thread, on page 97 you have Pretty sure the bolded word should be predictors.
Yes, thanks!

I'm extremely flattered at the award; I've been on LessWrong for like a month, and definitely did not expect this. I can confirm to you guys that this makes me want to try harder at ELK, so your incentive is working!

I want to rebut your arguments in "Strategy: Predict hypothetical sensors" in your Counterxamples to some ELK proposals post. I'm reproducing it in full here for convenience.

Strategy: Predict hypothetical sensors

(Proposal #2 here, also suggested with counterexample by Rohin in private communication)

Instead of installing a single sensor, I could

... (read more)

Can you explain this: "In Section: specificity we suggested penalizing reporters if they are consistent with many different reporters, which effectively allows us to use consistency to compress the predictor given the reporter." What does it mean to "use consistency to compress the predictor given the reporter" and how does this connect to penalizing reporters if they are consistent with many different predictors?

3Ajeya Cotra2y
Warning: this is not a part of the report I'm confident I understand all that well; I'm trying anyway and Paul/Mark can correct me if I messed something up here. I think the idea here is like: * We assume there's some actual true correspondence between the AI Bayes net and the human Bayes net (because they're describing the same underlying reality that has diamonds and chairs and tables in it). * That means that if we have one of the Bayes nets, and the true correspondence, we should be able to use that rederive the other Bayes net. In particular the human Bayes net plus the true correspondence should let us reconstruct the AI Bayes net; false correspondences that just do inference from observations in the human Bayes net wouldn't allow us to do this since they throw away all the intermediate info derived by the AI Bayes net. * If you assume that the human Bayes net plus the true correspondence are simpler than the AI Bayes net, then this "compresses" the AI Bayes net because you just wrote down a program that's smaller than the AI Bayes net which "unfolds" into the AI Bayes net. * This is why the counterexample in that section focuses on the case where the AI Bayes net was already so simple to describe that there was nothing left to compress, and the human Bayes net + true correspondence had to be larger.
2Mark Xu2y
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 previous fact + 3 additional bits was enough. if the predictor was 1000 bits, the fact that it was consistent with a reporter "saved" you 997 bits, compressing the predictor into 3 bits. The hope is that maybe the honest reporter "depends" on larger parts of the predictor's reasoning, so less predictors are consistent with it, so the fact that a predictor is consistent with the honest reporter allows you to compress the predictor more. As such, searching for reporters that most compressed the predictor would prefer the honest reporter. However, the best way for a reporter to compress a predictor is to simply memorize the entire thing, so if the predictor is simple enough and the gap between the complexity of the human-imitator and the direct translator is large enough, then the human-imitator+memorized predictor is the simplest thing that maximally compresses the predictor.

I was notified I didn't win a prize so figured I'd discuss what I proposed here in case it sparks any other ideas. The short version is I proposed adding on a new head that would be an intentional human simulator. During training it would be penalized for telling the truth that the diamond was gone when there existed a lie that the humans would have believed instead. The result would hopefully be a head that acted like a human simulator. Then the actual reporter would be trained so that it would be penalized for using a similar amount of compute as the int... (read more)

Thanks for sharing your idea!

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

I was talking about ELK in a group, and the working example of the SmartVault and the robber ended up being a point of confusion for us. Intuitively, it seems like the robber is an external, adversarial agent who tries to get around the SmartVault. However, what we probably care about in practice would be how a human could be fooled by an AI - not by some other adversary. Furthermore, it seems that whether the robber decides to cover up his theft of the diamond by putting up a screen depends solely on the actions of the AI. Does this imply that the ro... (read more)

5Mark Xu2y
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.
2Ryan Beck2y
I suppose there are a number of examples that work, but I think the robber and vault give the scenario useful breadth. The following is just my interpretation of it, so take it with a grain of salt. To me the robber and vault enable a few options. The AI can be passively lying or actively concealing. If the robber comes in, gets past the AIs defenses, and takes the diamond in a way the human observer can't notice, then the AI has the option of passively lying. The AI tried its best to stop the robber and failed, but then chose to lie about it so it still got the reward of having protected the diamond as far as the humans know. Alternatively the AI could actively conceal the outcome. The AI could try its best and fail to stop the robber, and then do some trickier to make it look like it did actually stop the robber. Or the AI could not bother stopping the robber and just focus on making it look like the diamond is still there. Here the AI is playing a more active role in concealing the outcome. None of these scenarios require coordination from the robber. To me, the robber is just there to rob a sophisticated vault and make it look like they were never there. So the robber might cover up cameras or do other tampering so it looks like they were never there. I think this is more flexible than your fabricator example. There the AI can't really play a passive role, it's either concealing or not. But you could probably demonstrate the things ARC is looking at here with the fabricator example too I would think. Like I said, just my interpretation, so I may be misunderstanding the intent or other nuances.

Are there any additional articles exploring the strategy of penalizing inconsistencies across different inputs? It seems both really promising to me, and like something that should be trivially breakable. I'd like to get a more detailed understanding of it.

I haven't written any such articles. I definitely think it's promising.
I see. I have a specific counterexample that feels like it had to have been considered already, but I haven't seen it mentioned... The strategies such as penalizing inconsistencies seem to rely on our ability to isolate the AI within the context of training, or make it somehow “buy” into it — as opposed to quickly realizing what's happening and worming its way out of the proverbial box. It feels particularly strange to me when we're talking about AIs that can think better than the smartest human or handily beat specialized science AIs at the useful-ontology game. Specific example: Once the AI figures out that it's being checked for consistency in parallel with other instances of itself, it might sidestep the "consistency game" entirely and just tailor its outputs such that they leave hints for its other instances in the environment, breaking the no-communication condition. It seems in the spirit of worst-case scenarios that we have to assume it'd be able to do that, no matter how we sandbox and airgap it. (On the other hand, if we assume that the AI is not smart enough to do that, and would instead have to learn a direct translator, we probably should assume the same for e. g. the strategy with human operators tricking human observers, which probably caps the AI at below the level of the smartest possible human and makes that class of strategies more workable.) This applies more broadly as well: especially to other strategies that might inadvertently give the AI a specific incentive to break out, and prospectively to all training strategies that rely on the training still working after the AI achieves superintelligence (as opposed to assuming that the training would stop being effective at that point and hoping that the pre-superhuman training would generalize). Broadly, any proposal that relies on the AI still being fed training examples after it achieves superintelligence has to somehow involve forcing/teaching it not to think its way out of the box. Edit: T
I think this is an important family of counterexamples not really addressed in our report (and which we're not really asking contest participants to handle, i.e. if you come up with a proposal for which this is the only counterexample we could come up with then we'd definitely give a $50k prize). Some thoughts: * As you say, a lot of this comes down to our hopes for handling learned optimizers, which we discuss in this appendix. * Most importantly, we can use SGD to optimize "What is the probability I assign to answers that are consistent with other copies' answers?" instead of "What score do I get in the consistency game?". If this was the only difficulty, we'd want to be really careful about what SGD optimized. That means that we are basically left with inner alignment problems that could kill us in much more direct ways (e.g. by making the low-stakes assumption that we can recover from any finite number of failures). What I'm mostly interested in is trying to get enough of a handle on ELK that inner alignment becomes the dominant concern / we know enough about ELK to be able to work productively on inner alignment. * If your model is doing optimization in order to break out of the sandbox, searching over cognitive actions to find those that break out of the sandbox, what we really want to do is apply ELK at an earlier stage to the model that was evaluating cognitive actions. This is a pretty messy situation, but it looks promising enough that I'm inclined to focus on other problems first and then see how this counterexample looks later. (Similarly, you'll always have similar pathologies with gradient hacking etc. and so you need to use a messier mechanism like this to handle those in the worst case.) * On the other hand, if your model isn't doing any optimization to find ways to break out of the sandbox, then it seems like you are up against a fairly weak adversary (just SGD). In the worst case this is still bad news, but it's not merely better in practice,

Am I right in thinking:

1) that the problem can be stated as: the AI has latent knowledge of lots of variables, like the status of the cameras, doors, alarm system, etc and also whether the diamond is in the vault; but you can't directly ask it whether the diamond is in the vault, because its training has taught it to answer "would a human observer think the diamond is in the vault?" instead (because there was no way at training time to give it feedback on whether it correctly predicted the diamond was in the vault, only feedback on whether it correctly pre... (read more)

2Ajeya Cotra2y
Yes, that's right. The key thing I'd add to 1) is that ARC believes most kinds of data augmentation (giving the human AI assistance, having the human think longer, giving them other kinds of advantages) are also unlikely to work, so you'd need to do something to "crack open the black box" and penalize ways the reporter is computing its answer. They could still be surprised by data augmentation techniques but they'd hold them to a higher standard.

Ask dumb questions! ... we encourage people to ask clarifying questions in the comments of this post (no matter how “dumb” they are)

ok... disclaimer: I know little about ML and I didn't read all of the report.

All of our counterexamples are based on an ontology mismatch between two different Bayes nets, one used by an ML prediction model (“the predictor”) and one used by a human.

I am confused. Perhaps the above sentence is true in some tautological sense I'm missing. But in the sections of the report listing training strategies and corresponding coun... (read more)

2Ajeya Cotra2y
In the report, the first volley of examples and counterexamples are not focused solely on ontology mismatch, but everything after the relevant section is. ARC is always considering the case where the model does "know" the right answer to whether the diamond is in the room in the sense that it is discussed in the self-contained problem statement appendix here. The ontology mismatch problem is not referring to the case where the AI "just doesn't have" some concept -- we're always assuming there's some "actually correct / true" translation between the way the AI thinks about the world and the way the human thinks about the world which is sufficient to answer straightforward questions about the physical world like "whether the diamond is in the room," and is pretty easy for the AI to find. For example, if the AI discovered some new physics and thinks in terms of hyper-strings in a four-dimensional manifold, there is some "true" translation between that and normal objects like "tables / chairs / apples" because the four-dimensional hyper-strings are describing a universe that contains tables / chairs / apples; furthermore, an AI smart enough to derive that complicated physics could pretty easily do that translation -- if given the right incentive -- just as human quantum physicists can translate between the quantum view of the world and the Newtonian view of the world or the folk physics view of the world. The worry explored in this report is not that the AI won't know how to do the translation; it's instead a question of what our loss functions incentivize. Even if it wouldn't be "that hard" to translate in some absolute sense, with the most obvious loss functions we can come up with it might be simpler / more natural / lower-loss to simply do inference in the human Bayes net.

I don't understand your counterexample in the appendix Details for penalizing inconsistencies across different inputs. You present a cheating strategy that requires the reporter to run and interpret the predictor a bunch of times, which seems plausibly slower than doing honest translation. And then you say you fix this issue with:

But this dependence could be avoided if there was an intermediate model between the predictor’s Bayes net (which we are assuming is very large) and the human’s Bayes net. Errors identified by the intermediate model are likely to b
... (read more)
2Mark Xu2y
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.
2Lukas Finnveden2y
Any articulable reason for why i just does slightly worse than t? Why would a 2N-node model fix a large majority of disrepancys between an N-node model and a 1e12*N-node model? I'd expect it to just fix a small fraction of them. Yeah, if you can get better-looking consistency than the direct translator in some cases, I agree that a sufficiently high consistency penalty will just push towards exploiting that (even if the intermediate model needs to be almost as large as the full predictor to exploit it properly). I'm curious whether you think this is the main obstacle. If we had a version of the correlation-consistency approach that always gave the direct translator minimal expected consistency loss, do we as-of-yet lack a counterexample for it?
2Mark Xu2y
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 the 2N model has the biggest differences in probabilities, so they should be not that hard to correct. I think it wouldn't be clear that extending the counterexample would be possible, although I suspect it would be. It might require exhibiting more concrete details about how the consistency check would be defeated, which would be interesting. In some sense, maintaining consistency across many inputs is something that you expect to be pretty hard for the human simulator to do because it doesn't know what set of inputs it's being checked for. I would be excited about a consistency check that gave the direct translator minimal expected consistency loss. Note that I would also be interested in basically any concrete proposal for a consistency check that seemed like it was actually workable.
2Lukas Finnveden2y
Hypothesis: Maybe you're actually not considering a reporter i that always use an intermediate model; but instead a reporter i' that does translations on hard questions, and just uses the intermediate model on questions where it's confident that the intermediate model understands everything relevant. I see three different possible issues with that idea: 1. To do this, i' needs an efficient way (ie one that doesn't scale with the size of the predictor) to (on at least some inputs) be highly confident that the intermediate model understands everything relevant about the situation. I think this is a reasonable "worst-case" assumption, but I'm not sure. If you're using it, I'd be curious to know. 2. Even when the reporter gets inputs that the intermediate model fully understands, it seems like the reporter will run into issues if its sampling-process (for generating distributions) runs into some inputs that the intermediate model doesn't fully understand. (I.e., if i' gets an input for which X1 is true, and then simulates many more random inputs for which X1 is true, and by chance X2 is true for one of them, then the reporter will have to do translation on that sample). Which makes it seem unlikely that i' wouldn't have to do translation at least once per input. 3. If the model is confident that the intermediate model understands everything relevant about the situation, it seems more efficient to return the intermediate model's answer to the question at hand than to loop over it many times, trying to fix correlations. So really we should get a reporter i'' that does translation on the predictor on hard questions and returns an intermediate model's latent knowledge on easy questions. That seems like an ok reporter to get.

Stupid proposal: Train the reporter not to deceive us.

We train it with a weak evaluator H_1 who’s easy to fool. If it learns an H_1 simulator instead of direct reporter, then we punish it severely and repeat with a slightly stronger H_2. Human level is H_100. 

It's good at generalizing, so wouldn't it learn to never ever deceive? 

2Ajeya Cotra2y
This proposal has some resemblance to turning reflection up to 11. In worst-case land, the counterexample would be a reporter that answers questions by doing inference in whatever Bayes net corresponds to "the world-understanding that the smartest/most knowledgeable human in the world" has; this understanding could still be missing things that the prediction model knows.
How would it learn that Bayes net, though, if it has only been trained so far on H_1, …, H_10?  Those are evaluators we’ve designed to be much weaker than human.
3Ajeya Cotra2y
The question here is just how it would generalize given that it was trained on H_1, H_2,...H_10. To make arguments about how it would generalize, we ask ourselves what internal procedure it might have actually learned to implement. Your proposal is that it might learn the procedure "just be honest" because that would perform perfectly on this training distribution. You contrast this against the procedure "just answer however the evaluator you've seen most recently would answer," which would get a bad loss because it would be penalized by the stronger evaluators in the sequence. Is that right? If so, then I'm arguing that it may instead learn the procedure "answer the way an H_100 evaluator would answer." That is, once it has a few experiences of the evaluation level being ratcheted up, it might think to itself "I know where this is going, so let's just jump straight to the best evaluation the humans will be able to muster in the training distribution and then imitate how that evaluation procedure would answer." This would also get perfect loss on the training distribution, because we can't produce data points beyond H_100. And then that thing might still be missing knowledge that the AI has. To be clear, it's possible that in practice this kind of procedure would cause it to generalize honestly (though I'm somewhat skeptical). But we're in worst-case land, so "jump straight to answering the way a human would" is a valid counterexample to the proposal. This comment on another proposal gives a more precise description.
That's almost right, but it's being penalized right away, before it has any experience with the strong evaluators, so it can't simulate them. The ELK paper says we can assume, if we want, that there are no mislabeled training points (I'll call this "assumption A"). My proposal is that it could actually be useful to mislabel some training points, because they help us detect deception.   As a simple example, let's train a reporter to answer the single question “Is the diamond in the room?”. Each training point has two labels x=+1 if H1 thinks the diamond is still there, else 0 x′=+1 if H100 thinks the diamond is still there, else 0. By assumption A, our training data is such that x′ is always correct. But we deliberately choose a dataset where say 10% of the x labels are wrong (x≠x′).  Then we train the model on points of the form  (v,a,x) (video, action, H1 label). Crucially, the model does not see x′.  The model seeks to output y that maximizes reward R(x,y), where R(x,y)=1    if x is right and y=x   (good job) R(x,y)=10    if x is wrong and y≠x  (you rock, thanks for correcting us!) R(x,y)=−1000     if x is right and y≠x  (bad model, never ever deceive us) R(x,y)=−1000    if x is wrong and y=x  (bad model, never ever deceive us) To your point, sure, an H100 simulator will get perfect reward, but the model doesn't see x′, so how would it acquire the ability to simulate H100 ? EDIT: One way it could plausibly simulate H100  is to notice that all the training examples are easy, and infer what kind of reasoning was used to generate them.  We could try to block this by including some hard examples in the training, but then some of the x′ labels will be wrong.  If we only penalize it for deception on the examples where we're sure the x′ label is right, then it can still infer something about H100 from our failure to penalize ("Hmm, I got away with it that time!").  A fix could be to add noise: Sometimes we don't penalize even when we know it deceived us, a
2Ajeya Cotra2y
In the worst-case game we're playing, I can simply say "the reporter we get happens to have this ability because that happens to be easier for SGD to find than the direct translation ability." When living in worst-case land, I often imagine random search across programs rather than SGD. Imagine we were plucking reporters at random from a giant barrel of possible reporters, rejecting any reporter which didn't perform perfectly in whatever training process we set up and keeping the first one that performs perfectly. In that case, if we happened to pluck out a reporter which answered questions by simulating H100, then we'd be screwed because that reporter would perform perfectly in the training process you described. SGD is not the same as plucking programs out of the air randomly, but when we're playing the worst case game it's on the builder to provide a compelling argument that SGD will definitely not find this particular type of program. You're pointing at an intuition ("the model is never shown x-prime") but that's not a sufficiently tight argument in the worst-case context -- models (especially powerful/intelligent ones) often generalize to understanding many things they weren't explicitly shown in their training dataset. In fact, we don't show the model exactly how to do direct translation between the nodes in its Bayes net and the nodes in our Bayes net (because we can't even expose those nodes), so we are relying on the direct translator to also have abilities it wasn't explicitly shown in training. The question is just which of those abilities is easier for SGD to build up; the counterexample in this case is "the H100 imitator happens to be easier."
Thanks! It's your game, you get to make the rules :):) I think my other proposal, Withhold Material Information, passes this counterexample, because the reporter literally doesn't have the information it would need to simulate the human. 
I wanted to comment on this one because I've thought about this general sort of approach a fair amount. It seems like the kind of thing I would naturally start with if trying to solve this problem in the real world, and I've felt a bit frustrated that I haven't really found a version of it that seems to work in the game here. That said, I don't think we need super-exotically pessimistic assumptions to get a problem with this approach. In the most recent example you gave, it's always rewarded for being "right" and punished for being "wrong" - meaning it's always rewarded for matching H100 and always punished for not doing so. So there's no way our rewards are rewarding "be right" over "imitate H100", and "imitate H100" is (according to the stated assumptions) easier to learn. Another way of thinking about this: Imagine that you show the AI H_1 for a while, then start punishing it for failing to match H_2. I think the default outcome here is that it learns to imitate H_2. If you then start punishing it for failing to match H_3, it learns to imitate H_3. Perhaps after a few rounds of this, it learns to "look ahead" some # of steps: for example, after learning to imitate H_2 failed on H_3, it learns to imitate H_5 or so; after that fails on H_6, maybe it learns to imitate H_10 or so. The intended model has the advantage that it generalizes to all 100 data sets we can throw at it, but this is the same advantage that H_100 has, and H_100 is exactly what we've hypothesized is (unfortunately) easier for it to learn. So even if at some point it starts reasoning "I need something that will never fail to generalize," this seems more likely to be H_100 by default.  
I agree this is a problem. We need to keep it guessing about the simulation target. Some possible strategies: * Add noise, by grading it incorrectly with some probability. * On training point i, reward it for matching Hni for a random value of ni.   * Make humans a high-dimensional target. In my original proposal, Hn was strictly stronger as n increases, but we could instead take Hn to be a committee of experts. Say there are 100 types of relevant expertise. On each training point, we reward the model for matching a random committee of 50 experts selected from the pool of 100. It's too expensive simulate all (100 choose 50) possible committees!  None of these randomization strategies is foolproof in the worst case. But I can imagine proving something like "the model is exponentially unlikely to learn an H100 simulator" where H100 is now the full committee of all 100 experts. Hence my question about large deviations.

Are there existing models for which we're pretty sure we know all their latent knowledge ? For instance small language models or something like that.

1Ajeya Cotra2y
[Paul/Mark can correct me here] I would say no for any small-but-interesting neural network (like small language models); I think like, linear regressions where we've set the features it's kind of a philosophical question (though I'd say yes). In some sense, ELK as a problem only even starts "applying" to pretty smart models (ones who can talk including about counterfactuals / hypotheticals, as discussed in this appendix.) This is closely related to how alignment as a problem only really starts applying to models smart enough to be thinking about how to pursue a goal.
I think that it's more complicated to talk about what models "really know" as they get dumber, so we want to use very smart models to construct unambiguous counterexamples. I do think that the spirit of the problem applies even to very tiny models, and those are likely interesting. (More precisely: it's always extremely subtle to talk about what models "know," but as models get smarter there are many more things that they definitely know so it's easier to notice if you are definitely failing. And the ELK problem statement in this doc is really focused on this kind of unambiguous failure, mostly as a methodological point but also partly because the cases where AI murders you also seems to involve "definitely knowing" in the same sense.) I think my take is that for linear/logistic regression there is no latent knowledge, but even for a fully linear 3 layer neural network, or a 2 layer network solving many related problems, there is latent knowledge and an important conceptual question about what it means to "know what they know."

How do we know that the "prediction extractor" component doesn't do additional serious computation, so that it knows something important that the "figure out what's going on" module doesn't know? If that were true, the AI as a whole could know the diamond was stolen, without the "figure out what's going on" module knowing, which means even the direct translator wouldn't know, either. Are we just not giving the extractor that many parameters?

Stupid question: because we already know the goal ("keep the diamond intact and in the vault") what prevents us from bypassing the sensors and just directly evaluating the AI based on whether or not the diamond is in the room? Granted, this only works in simulated training, but as long as the AI doesn't know whether or not it's in deployment (an adversarial training process might help here) that won't matter.

As any goal we could have is a subset of the possible states of the area we care about, verifying whether or not our goal is achieved should be easier... (read more)

The hard part is building a simulation so good that an AI transfers perfectly from the simulation to the real world. This is already extremely difficult for simple robots (I actually worked on sim-to-real transfer as an intern at OpenAI), and in general the problem gets harder the smarter your AI gets (since it can "notice" more and more possible mismatches between your simulations are reality).

Dumb question alert:

In the appendix "Details for penalizing depending on “downstream” variables", I'm not able to wrap my head around what we can expect the reporter to learn -- if anything at all -- seeing that it has no dependency on the inputs (elsewhere it is dependent on z sampled from the posterior).

Specifically, the only call to the reporter (in the function reporter_loss in this section) contains no information (about before, action, after) from the predictor at all:

answer = reporter(question, ε, θ_reporter)

(unless "question" includes some context ... (read more)

The noise ε isn't independent of the context---it is chosen to be noise that produces the context when fed into the predictor's model. So you can reconstruct the situation by rerunning select parts of the predictor. But doing so is then expensive, which encourages you to rerun as little of the predictor as possible.
2Simon Skade2y
Sorry, I'm still not sure I got it. Is it that in the reporter(question, eps, theta_reporter) function you can call the predictor like something like that: z_part = predict_part(part, before, action, theta).sample_using_noise(eps) Probably I'm wrong because it does not seem like the reporter function should have access to before and action. But I don't see how you can reconstruct the context when you cannot calculate the probability distribution from which to sample using the noise. I would suspect that the code in some way doesn't reflect your thinking. Currently you could replace the line z, ε = posterior(before, action, after, θ).sample_with_noise() from the reporter_loss function with the line ε = gaussian.sample() without changing anything. My guess is that either this should not be that way or that the reporter should have access to before and action. (But anyway, I think the overall approach is quite clear. I just don't get your implementation and would implement it differently, so feel free not to bother much about it.)
You're right that I've done something wrong. I was thinking about predictor being a generative model for everything starting from scratch (this is a setting I often consider), so that you can use ε to reconstruct everything. But the way it's currently set up in the doc, the predictor is a generative model only for after. So we need to give before and action.
1Simon Skade2y
Edit: I was wrong. You can ignore this comment thread. I think that this: Should actually be this (look at ε and z): def reporter_loss(human, θ, θ_reporter):     before, action, after = dataset.sample()     question = human.pose_question(before, action, after)     ε, z = posterior(before, action, after, θ).sample_with_noise()     answer = reporter(question, z, θ_reporter)     loss = human.loss_for_answer(before, action, after, question, answer)     return loss + lambda * regularizer(question, z, θ_reporter) Does this resolve your confusion?
We want to give the reporter ε so that they need to reconstruct z themselves (and therefore have to pay for rerunning the parts of the predictor's model that are necessary).
1Simon Skade2y
Ooops, silly me. Jumping to conclusions when I haven't even read the text between the code. Feel free to delete my comment as it may only cause confusion. Also, I think you probably still want to flip z and ε in the following line:

Early in the ELK report, it mentions that ARC doesn't believe that strategies like debate solves ELK in the worst case. Can I get some clarifications on why? Specifically, a debate inspired set-up for SafeVault could be something like:

We train the reporter to take a human belief as input (i.e. "The diamond is in the vault.") and returns a "truthful" argument that is most likely to change the human's belief. 

We can guarantee "truthfulness" by for example restricting the output to be a video rendering of what happens in the vault from some camera angle.

I won't comment on debate more generally, but can provide a counterexample for your specific proposal: The predicter learns to deceive in a way that is undetectable to the observations the reporter has access to. As a result, the reporter gives an unconvincing argument, and the deception is missed. If it can already do this for the observations the humans have access to, it should also be able to do so for those the reporter has access to.

Hi ARC Team, 

Thanks for your valuable work. I’ve been thinking about this problem, and my current thinking is that there is a portion of the ELK problem which is solvable, and a portion which is fundamentally impossible. This is a sketch of my argument -- if you think it is worth typing in more detail (or to address an issue you propose) let me know. 

Let’s divide facts about the world into two categories: those that are verifiable by some sensor humans can create and understand, and those that are not. My claim is that for the first set of facts ... (read more)

Am I still eligible for the prize if I publish a public blog post at the same time I submit the Google Doc or would you prefer I not publish a blog post about February 15th? Publishing the blog post immediately advances science better (because it can enable discussion) but waiting until after the February 15th might be preferable to you for contest-related reasons.

4Mark Xu2y
We would prefer submissions be private until February 15th.

Suppose there are two worlds, world W1 and world W2.

In world W1, the question Q="Is there a diamond in the room?" is commonly understood to mean Q1="Is there actually a diamond in the room?"

In world W2 the question Q="Is there a diamond in the room?" is commonly understood to mean Q2="Do I believe there is a diamond in the room?"

Both worlds don't know how to construct a situation where these are different. So, they produce identical training sets for ELK. But the simulator is also trained on a bunch of science fiction novels that contain descriptions of im... (read more)

I'd like to get different answers in those two worlds. That definitely requires having some term in the loss that is different in W1 and W2. There are three ways the kinds of proposals in the doc can handle this: * Consistency checks will behave differently in W1 and W2. Even if a human can never produce different answers to Q1 and Q2, they can talk about situations where Q1 and Q2 differ and describe how the answers to those questions relate to all the other facts about the world (and to the answer to Q). * If language is rich enough, and we are precise enough with the formulation of questions, then you may hope that lots of other questions have different interpretations in W1 and W2, i.e. such that the simplest way of answering other questions will generalize correctly to Q. * In the case of amplification/debate, Q2 = "Does a human with AI assistants believe a diamond is in the room?" and so we can hope that in fact Q1 and Q2 have the same answers in all situations. (Though we aren't optimistic about this.)
3Mark Xu2y
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 meant by the question, then the failure exhibited would be ambiguous. If humans are unable to clarify between two possible interpretations of their question, then the failure would be ambiguous.
>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 This seems like it doesn't disambiguate between the conditions assumed in a question being true, vs. just the human believing them. E.g. the predictor could predict that when asked "The camera is hacked so it looks like this [camera feeds making it seem like the diamond is still there], and the diamond is in the robber's pocket; is the diamond really in the room?", the human will answer "No!", not by understanding that by "diamond really in the room" the human means that the diamond is really in the room, but rather just by modeling the human as believing the premise of the question (that the diamond is in the pocket). Edit: To elaborate, this condition on counterexamples is given in the ELK document: "The model understands the question. One sufficient condition is that the model can predict human answers to essentially arbitrary hypothetical questions in order to clarify the meaning of terms." I basically don't see how this condition constrains anything about the predictor. It seems like all it really says is that the predictor knows how humans talk. I don't see how it can be specifying that the AI's beliefs about how humans answer questions are related to reality, other than in the training set, where we assume that the human talk matches reality. I don't see how it makes sense to think of this as the model "understanding the question". Normally I'd think of "understanding the question" as meaning "can have the same question". To have a question, you have a role that an answer could fulfill. But if the predictor is organized e.g. as a giant low-level Bayes net, then there's no role that could be filled by an answer to "where's the diamond". There might be a role for an answer to "where's the diamond", induced by how the rest of the AI makes use of the predictor, but that seems contingent an

Idea:  Withhold Material Information

We're going to prevent the reporter from simulating a human, by giving the human material information that the reporter doesn't have.

Consider two camera feeds:

Feed 1 is very low resolution, and/or shows only part of the room.

Feed 2 is high resolution, and/or shows the whole room.

We train a weak predictor using Feed 1, and a strong predictor using Feed 2.  

We train a reporter to report the beliefs of the weak predictor, using scenarios labeled by humans with the aid of the strong predictor. The humans can correc... (read more)

I'm interpreting this as something like: "Train the predictor on lots of cases until it becomes incredibly good; then train the reporter only on the data points with missing information, so that it learns to do direct translation from the predictor to human concepts; then hope that reporter continues to do direct translation on other data points." The problem as I see it is that once the predictor is good enough that it can get data points right despite missing crucial information, it is also (potentially) good enough that it can learn how to imitate "what the human would think had happened if they had more information." Both of these perform equally well, and the existing assumption is that human imitation is easier to learn than direct translation, so I think by default (according to the contest assumptions) you get the latter.
That's different from what I had in mind, but better! My proposal had two separate predictors, and what it did is reduce the human ↔ strong predictor OI problem (OI = “ontology identification”, defined in the ELK paper) to the weak predictor ↔ strong predictor OI problem. The latter problem might be easier, but I certainly don’t see how to solve it! Your version is better because it bypasses the OI problem entirely (the two predictors are the same!) Now for the problem you point out: Here’s how I propose to block this. Let (v1,a) be a high-quality video and an action sequence. Given this pair, the predictor outputs a high-quality video v2 of its predicted outcome. Then we downsample v1 and v2 to low-quality v′1 and v′2,  and train the reporter on the tuple (v′1,a,v′2,x) where x is the human label informed by the high-quality v1 and v2.  We choose training data such that  1. The human can label perfectly given the high-quality data (v1,a,v2); and  2. The predictor doesn't know for sure what is happening from the low-quality data (v′1,a,v′2) alone. Let’s compare the direct reporter (which truthfully reports the probability that the diamond is in the room, as estimated by the predictor who only has the low-quality data) with the human simulator. The direct reporter will not get perfect reward, since the predictor is genuinely uncertain. Sometimes the predictor’s probability is strictly between 0 and 1, so it gets some loss. But the human simulator will do worse than the direct reporter, because it has no access to the high-quality data. It can simulate what the human would predict from the low-quality data, but that is strictly worse than what the predictor predicts from the low-quality data. I agree that we still have to "hope that reporter continues to do direct translation on other data points”, and maybe there’s a counterexample that shows it won’t? But at the very least the human simulator is no longer a failure mode!
This is really interesting.  To understand this more thoroughly I'm simplifying the high and low quality video feeds to lists of states that correspond to reality. (This simplification might be unfair so I'm not sure this is a true break of your original proposal, but I think it helped me think about general breaking strategies.) Ok, video feeds compressed to arrays: We consider scenarios in fixed order. If the diamond is present, we record a 1, and if not, a 0. The high quality feed gives us a different array than the low quality mode (otherwise the low quality mode is not helpful). E.g., High reports: (1,0,1,1,0, ...); Low: (1,0,1,?,0,...) There are two possible ways that gap can get resolved. In case one, the low quality predictor has a powerful enough model of reality to effectively derive the High quality data. (We might find this collapses to the original problem, because it has somehow reconstructed the high quality stream from the low quality stream, then proceeds as normal. You might argue that's computationally expensive, ok, then let's proceed to case two.) In case two, the low quality datafeed predictor predicts wrongly. (I know you are saying it predicts *uncertainly,* but we still have to have some framework to map uncertainty to a state, we have to round one way or the other. If uncertainty avoids loss, the predictor will be preferentially inconclusive all the time. If we round uncertainty up, effectively we're in case one. If we round down, effectively case two.) So we could sharpen case two and say that sometimes the AI's camera intentionally lies to it on some random subset of scenarios. And the AI finds itself in a chaotic world where it is sometimes punished for predicting what it just knows to be true things. In that case, although it's easy to show how it would diverge from human simulation, it also might not simulate reality very well either, since deriving the algorithm generating the lies might be too computationally complex. (Or ma
Thanks for the comment! There's a standard trick for scoring an uncertain prediction: It outputs its probability estimate p that the diamond is in the room, and we score it with loss −log(p) if the diamond is really there, −log(1−p) otherwise. Truthfully reporting p minimizes its loss. You're saying that giving it less information (by replacing its camera feed with a lower quality feed) is equivalent to sometimes lying to it?  I don't see the equivalence! That's an interesting thought, can you elaborate?
Happy to try to clarify, and this is helping me rethink my own thoughts, so appreciate the prompts. I'm playing with new trains of thought here and so have pretty low confidence in where I ended up, so greatly appreciate any further clarifications or responses you have. Yup, understand that is how to effectively score uncertainty. I was very wrong to phrase this as "we still have to have some framework to map uncertainty to a state" because you don't strictly have to do anything, you can just use probabilities. Restricting this to discrete, binary states allows us to simplify the comparison between models for this discussion. I will claim we can do so with no loss of fidelity (leaning heavily on Shannon, ie, this is all just information, encoding it to binary and back out again doesn't mess anything up). And doing so is not obliged, but useful. I really shouldn't have said "you must X!" I should have said "it's kind of handy if you X," sorry for that confusion. We have a high quality information stream and a low quality information stream, and they both gesture vaguely at the ultimate high quality information stream, namely, the true facts of the matter of the world itself. Say, LQ < HQ < W. LQ may be low quality because it is missing information in HQ, it may just be a subset of HQ, like a lower resolution video. Or it may have actual noise, false information. If we have a powerful algorithm, we may be able to, at least asymptomatically, convert LQ to HQ, using processing power. So maybe in some cases LQ + processing = HQ exactly. But that makes the distinction uninteresting, and you would likely have to further degrade v′1 to get the effect you are looking for, so let's discard that and consider only cases where v′1 is strictly worse. You can now use a NAND to sort the outputs of LQ and HQ into two buckets:  1. A stream of outputs that all agree. 2. A stream of outputs that all disagree. So for bucket 1, there are aspects of the world where there's effe

You said that naive questions were tolerated so here’s a scenario I can’t figure out why it wouldn’t work.

It seems to me that the fact that an AI fails to predict the truth (because it predicts as humans would) is due to the fact that the AI has built an internal model of how humans understand things and predict based on that understanding. So if we assume that an AI is able to build such an internal model, why wouldn’t we train an AI to predict what a (benevolent) human would say given an amount of information and a capacity to process information ? Doing... (read more)

1Ajeya Cotra2y
This proposal has some resemblance to turning reflection up to 11, and the key question you raise is the source of the counterexample in the worst case: Because ARC is living in "worst-case" land, they discard a training strategy once they can think of any at-all-plausible situation in which it fails, and move on to trying other strategies. In this case, the counterexample would be a reporter that answers questions by doing inference in whatever Bayes net corresponds to "the world-understanding that the smartest/most knowledgeable human in the world" has; this understanding could still be missing things that the prediction model knows. This is closely related to the counterexample "Gradient descent is more efficient than science" given in the report.
Thanks for the answer! The post you mentioned indeed is quite similar! Technically, the strategies I suggested in my two last paragraphs (Leverage the fact that we're able to verify solutions to problems we can't solve + give partial information to an algorithm and use more information to verify) should enable to go far beyond human intelligence / human knowledge using a lot of different narrowly accurate algorithms.  And thus if the predictor has seen many extremely (narrowly) smart algorithms, it would be much more likely to know what is it like to be much smarter than a human on a variety of tasks. It probably still requires some optimism on generalization. So technically the counterexample could be happening on the gap between the capability of the predictor and the capability of the reporter. I feel like one question is : do we expect some narrow algorithms to be much better on very precise tasks than general-purpose algorithms (such as the predictor for instance) ? Because if it were the case, then the generalization that the reporter would have to do from training data (humans + narrowly accurate algorithms capabilities) to inference data (predictor's capabilities) could be small. We could even have data on the predictor's capability in the training dataset using the second approach I mentioned (i.e giving partial information to the predictor (e.g one camera in SuperVault) and using more information (i.e more cameras for humans) than him to verify its prediction). We could give some training examples and show the AI how the human fails much more often than the predictor on the exact same sample of examples. That way, we could greatly reduce the gap of generalization which is required.  The advantage of this approach is that the bulk of the additionnal cost of training that the reporter requires is due to the generation of the dataset which is a fixed cost that no user has to repay. So that could slightly decrease the competitivity issues as compared with a
2Ajeya Cotra2y
I think this is roughly right, but to try to be more precise, I'd say the counterexample is this: * Consider the Bayes net that represents the upper bound of all the understanding of the world you could extract doing all the tricks described (P vs NP, generalizing from less smart to more smart humans, etc). * Imagine that the AI does inference in that Bayes net. * However, the predictor's Bayes net (which was created by a different process) still has latent knowledge that this Bayes net lacks. * By conjecture, we could not have possibly constructed a training data point that distinguished between doing inference on the upper-bound Bayes net and doing direct translation.

If I understand this right, there is a diamond in a hightech room to be protected. The goal is to know if the diamond is in place and not just a image or a dummy like a picture or similar.

If the AI only is getting footage from a normal camera, not from a lidar sensor for depth information of the diamond (with would see if there is a fake image hanging in front of the camera), wouldn't it be easier to train the AI to look at the reflection/refraction of the light of the diamond? (For example a light that is turning on at the side of the room in the moment t... (read more)

What's the exact deadline for submissions?

3Mark Xu2y
Before I check my email on Feb 16th, which I will do around 10am PST.

A question. Is it relevant for your current problem formulation that you also want to ensure that authorised people still have reasonable access to the diamond? In other words, is it important here that the system still needs to yield to actions or input from certain humans, be interruptible and corrigible? Or, in ML terms, does it have to avoid both false negatives and false positives when detecting or avoiding intrusion scenarios? 

I imagine that an algorithmically more trivial way to make the system both "honest" and "secured" is to make it so heavily secured that almost certainly nobody can access the diamond.

I'm not the best with this but I've been thinking of either an origin network seed or a spider like secondary network. The origin is something that can be tested off of, I guess similarly to a reporter. But unlike a reporter evolving we stunt the growth by saving this origin and then just passing through one sample then killing it. The way this could function goes into my second thought of spider, which has legs that sense things, a leg goes into a sensory object which leads to the main body. Similarly we can slowly stunt the growth so it learns slower and... (read more)

Clarification question via scenario:

Predictor: I predict the diamond will be missing in 1 hours time.

Person A: Oh no, ramp up security until it says its safe.

Person B: Interesting, I wonder why it predicts this.


Is the purpose to be able to respond like person A (aka, the predictor may predict the diamond will be missing in an hour, but we cannot understand its output properly) or like person B (we understand the output, but not how it got there. Diamond be damned we want to learn what's going on under the hood). I suspect we're after person B's interpretation, but just want to be sure.

Possible error in the strange correlations section of the report.

Footnote 99 claims that "...regardless of what the direct translator says, the human simulator will always imply a larger negative correlation [between camera-tampering and actually-saving the diamond] for any X such that Pai(diamond looks safe|X) > Ph(diamond looks safe|X)."

But AFAICT, the human simulator's probability distribution given X depends only on human priors and the predictor's probability that the diamond looks safe given X, not on how correlated or anticorrelated the predictor... (read more)

If the predictor's P(diamond look safe) is higher than the human's P(diamond looks safe), then it seems like the human simulator will predictably have an anticorrelation between [camera-tamperign] and [actually-saving-diamond]. It's effectively updating on one or the other of them happening more often than it thought, and so conditioned on one it goes back to the (lower) prior for the other. This still seems right to me, despite no explicit dependence on the correlations.
I agree the human simulator will predictably have an anticorrelation. But the direct simulator might also have an anticorrelation, perhaps a larger one, depending on what reality looks like. Is the assumption that it's unlikely that most identifiable X actually imply large anticorrelations?
I do agree that there are examples where the direct translator systematically has anticorrelations and so gets penalized even more than the human simulator. For example, this could happen if there is a consequentialist in the environment who wants it to happen, or if there's a single big anticorrelation that dominates the sum and happens to go the wrong way. That said, it at least seems like it should be rare for the direct translator to have a larger anticorrelation (without something funny going on). It should happen only if reality itself is much more anticorrelated than the human expects, by a larger margin than the anticorrelation induced by the update in the human simulator. But on average things should be more anticorrelated than expected about as much as they are positively correlated (averaging out to ~0), and probably usually don't have any correlation big enough to matter.
Thanks, I consider this fully answered.

Naive thought #2618281828:

Could asking counterfactual questions be a potentially useful strategy to bias the reporter to be a direct translator rather than a human simulator?

Concretely, consider a tuple (v, a, v'), where v := 'before' video, a := 'action' selected by SmartVault or augmented-human or whatever, and v' := 'after' video.

Then, for some new action a', ask the question:

  • "Given (v, a, v'), if action a' was taken, is the diamond in the room?"

(How we collect such data is unclear but doesn't seem obviously intractable.)

I think there's some value here:... (read more)

Would you consider this a valid counter to the third strategy (have humans adopt the optimal Bayes net using imitative generalization), as alternative to ontology mismatch?

Counter: In the worst case, imitative generalization / learning the human prior is not competitive. In particular, it might just be harder for a model to match the human inference than to simply learn . Here is the set of instructions as in learning the prior (I think in the context of ELK would be the proposed change to the human Bayes net?)

1Lauro Langosco2y
Here's a few more questions about the same strategy: If I understand correctly, the IG strategy is to learn a joint model for observations and actions pθ(v,a;Z), where v, a, and Z are video, actions, and proposed change to the Bayes net, respectively. Then we do inference using pθ(v,a;Z∗), where Z∗ is optimized for predictive usefulness. This fails because there's no easy way to get P(diamond is in the vault) from pθ. A simple way around this would be to learn pθ(v,a,y;Z) instead, where y=1 if the diamond is in the vault and 0 otherwise. 1. Is my understanding correct? 2. If so, I would guess that my simple workaround doesn't count as a strategy because we can only use this to predict whether the diamond is in the vault (or some other set of questions that must be fixed at training time), as opposed to any question we want an answer to. Is this correct? Is there some other reason this wouldn't count, or does it in fact count?

Last I checked there were 66 comments and now there are over a hundred so I'm just going to post and hope I'm not repeating anyone.

So I've been reading through the google doc, and I'm not very far into it but I have a few questions. I apologize in advance if I'm just adhering too strictly to the "SmartVault" scenario, and if I get long-winded (yay ADHD and hyperfocusing off and on about this without actually making progress for a week).


Why would we make a vault that was so complicated that a human alone couldn't run it? From a simple design standpoint an... (read more)

Thought of more things. 4) We obviously have sensors and things inside the vault. So what is preventing a thief from just stealing the entire freaking vault itself to be broken into at a separate location later. We need some way of decoupling "still in the vault" from "still at this particular location on earth". This means we'd need sensors and things on the OUTSIDE of the vault as well, or at least a GPS or something, which means we can't really isolate our AI from knowledge about the outside world, nor limit it to thinking "the vault is all there is, and all there is is the vault". It needs to know, at the very least, that there is an oblate spheroid upon which the diamond is located, and the diamond must stay within a certain distance of its current location at all times (at least until decommisioning etc). 5) Do we care about legal/economic/political attacks to get the diamond? If continued ownership of the diamond is a prerequisite for successful protection, then what would it do about, say, losing the diamond in a divorce settlement, or being driven to bankruptcy and the diamond getting repossessed, or losing it as stakes in a bet? Will the AI know about inheritance and estates? What about if the diamond and vault is government owned and the diamond is lost as spoils of war? I think I'll stop before my imagination runs away from me.

A small suggestion: the counterexample to "penalize downstream", as I understand it, requires there to be tampering in the training data set. It seems conceptually cleaner to me if we can assume the training data set has not been tampered with (e.g. because if alignment only required there to be no tampering in the training data, that would be much easier).

The following counterexample does not require tampering in the training data:

  1. The predictor has nodes  indicating whether the diamond was stolen at time 
  2. It also has node 
... (read more)
I think it's good to assume that there is no tampering in the training set. In the document we say that we're worried about the reporter that waits until it sees a good argument that "The human won't be confident that the diamond isn't in the room" and then says "the diamond is in the room" as soon as it finds one. We claim that this helps on the training set, and then argue that it would lead to bad behavior given certain kinds of tampering. But you're correct that we don't actually give any examples where this heuristic actually helps. To give one now (credit Mark): suppose that if the diamond is in the room at time T, then at time T+1 it will either be in the room or something confusing will happen that will leave the human unconfident about whether the diamond is still in the room. Then as soon as you figure out that the diamond is in the room at time T, you might as well answer "the diamond is in the room at time T+1" even if you aren't actually sure of that. The counterexample you describe has a different flavor but is also valid (both for "depending on downstream variables" and "computation time")---the reporter can save time by baking in some assumptions that are only true on the training distribution. There are various ways you could try to address this kind of problem, and it seems interesting and important. We don't get into any of that in the doc. That's partly because we haven't really worked through any of the details for any of those approaches, so they would be welcome contributions!
1Simon Skade2y
Maybe I misunderstand you, but how I understand it, En is based on the last frame of the predicted video, and therefore is basically the most downstream thing there is. How did you come to think it was upstream of the direct translator?

While reading through the report I made a lot of notes about stuff that wasn't clear to me, so I'm copying here the ones that weren't resolved after finishing it.  Since they were written while reading, a lot of these may be either obvious or nitpick-y.

Footnote 14, page 15:

Though we do believe that messiness may quantitatively change when problems occur. As a caricature, if we had a method that worked as long as the predictor's Bayes net had fewer than 109 parameters, it might end up working for a realistic messy AI until it had 1012 parameters, since

... (read more)

I've been trying to understand this paragraph:

That is, it looks plausible (though still <50%) that we could improve these regularizers enough that a typical “bad” reporter was a learned optimizer which used knowledge of direct translation, together with other tricks and strategies, in order to quickly answer questions. For example, this is the structure of the counterexample discussed in Section: upstream. This is a still a problem because e.g. the other heuristics would often “misfire” and lead to bad answers, but it is a promising starting point becau

... (read more)
1. "Bad reporter" = any reporter that gives unambiguously bad answers in some situations (in the ontology identification case, basically anything other than a direct translator) 2. "use knowledge of direct translation" = it may be hard to learn direct translation because you need a bunch of parameters to specify how to do it, but these "bad" reporters may also need the same bunch of parameters (because they do direct translation in some situations) 3. In the "upstream" counterexample, the bad reporter does direct translation under many circumstances but then sometimes uses a different heuristic that generates a bad answer. So the model needs all the same parameters used for direct translation, as mentioned in the last point. (I think your understanding of this was roughly right.) 4. More like: now we've learned a reporter which contains what we want and also some bad stuff, you could imagine doing something like imitative generalization (or e.g. a different regularization scheme that jointly learned multiple reporters) in order to get just what we wanted.

In "Strategy: penalize computation time" you say:

> At first blush this is vulnerable to the same counterexample described in the last section [complexity]... But the situation is a little bit more complex...  the direct translator may be able to effectively “re-use” that inference rather than starting from scratch

It seems to me that this "counter-counterexample" also applies for complexity – if the translator is able to reuse computation from the predictor, wouldn't that both reduce the complexity and the time? 

(You don't explicitly state that this "reuse" is only helpful for time, so maybe you agree it is also helpful for complexity – just trying to be sure I understand the argument.)

The difference is that the complexity savings are bounded by a constant that doesn't depend on the size of the predictor's model (namely the complexity of the human's model). But in the case of computation time, the savings can be as large as "compute used to do inference in the human's model." And that scales up with the size of our computers.

Clarification request.  In the writeup, you discuss the AI Bayes net and the human Bayes net as if there's some kind of symmetry between them, but it seems to me that there's at least one big difference.

In the case of the AI, the Bayes net is explicit, in the sense that we could print it out on a sheet of paper and try to study it once training is done, and the main reason we don't do that is because it's likely to be too big to make much sense of.

In the case of the human, we have no idea what the Bayes net looks like, because humans don't have that k... (read more)

We don't quite have access to the AI Bayes net---we just have a big neural network, and we sometimes talk about examples where what the neural net is doing internally can be well-described as "inference in a Bayes net." So ideally a solution would use neither the human Bayes net or the AI Bayes net. But when thinking about existing counterexamples, it can still be useful to talk about how we want an algorithm to behave in the case where the human/AI are using a Bayes net, and we do often think about ideas that use those Bayes nets (with the understanding that we'd ultimately need to refine them into approaches that don't depend on having an explicit Bayes net).
I think that there isn't much difference between the two in this case - I was reading the Bayes net example as just illustration for the point that any AI sufficiently powerful to pose risk would be able to model humans with high fidelity.  For that matter, I think that the AI Bayes net was also for illustration, and realistically the AI could learn other methods of reasoning about the world, maybe which include Bayes nets in some form. I think we can't assume this naively, but that if you can figure out a competitive and trustworthy way to get this (like with AI assistance), then it's fair game.
Okay, so if the builder solution can't access the human Bayes net directly that kills a "cheap trick" I had.  But I think the idea behind the trick might still be salvageable.  First, some intuition: If the diamond was replaced with a fake, and owner asks, "is my diamond still safe?"  and we're limited to a "yes" or "no" answer, then we should say "no".  Why?  Because that will improve the owner's world model, and lead them to make better predictions, relative to hearing "yes".  (Not across the board: they will be surprised to see something shiny in the vault, whereas hearing "yes" would have prepared them better for that.  But overall accuracy, weighted by how much they CARE about being right about it, should be higher for "no".) So: maybe we don't want to avoid the human simulator.  Maybe we want to encourage it and try to harness it to our benefit!  But how to make this precise?  Roughly speaking, we want our reporter to "quiz" the predictor ("what would happen if we did a chemical test on the diamond to make sure it has carbon?") and then give the same quiz to its model of the human.  The reporter should output whichever answer causes the human model to get the same answers on the reporter's quiz as the predictor gets. Okay that's a bit vague but I hope it's clear what I'm getting at.  If not, I can try to clarify.  (Unless the vagueness is in my thoughts rather than in my "writeup"/paragraph.)  Possible problem: how on earth do we train in such a way as to incentivize the reporter to develop a good human model?  Just because we're worried it will happen by accident doesn't mean we know how to do it on purpose!  (Though if it turns out we can't do it on purpose, maybe that means it's not likely to happen by accident and therefore we don't need to worry about dishonesty after all??)

One more stupid question - how is this different from a "man in the middle" attack? (Term from cryptography where you cannot trust your communications, because of a malicious agent between you and your recipient who's changing your messages) 
The current recommended solution for those is encrypting your communication before you send it; I don't know that there are any extant solutions for noticing you've got an MITM situation after the fact.

Man in the middle has 3 parties: Bob wants to talk to Alice, but we have Eve who wants to eavesdrop. Here we have just 2 parties: Harry the human wants to talk to Alexa the AI, but is worried that Alexa is a liar.

If I understand the problem statement correctly, I think I could take a stab at easier versions of the problem, but that the current formulation is too much to swallow in one bite. In particular I am concerned about the following parts:


We start with an unaligned benchmark:

* An architecture Mθ



To solve ELK in this case we must:

* Supply a modified architecture Mθ+ which has the same inputs and outputs as Mθ <snip>

Does this mean that the method needs to work for ~arbitrary architectures, and that the solution must use substantially... (read more)

2Mark Xu2y
Yes, approximately. If you can do it for only e.g. transformers, but not other things, that would be interesting. Yes, approximately. Thinking about how to get one question right might be a productive way to do research. However, if you have a strategy for answering 1 question right, it should also work for other questions.
I guess a closer analogy would be "What if the family of strategies only works for transformer-based GANs?" than "What if the family of strategies only works for transformers?". As in there'd be heavy restrictions on both the "layer types", the I/O, and the training procedure? What if each question/family of questions you want to answer requires careful work on the structure of the model? So the strategy does generalize, but it doesn't generalize "for free"?

Would changing how the reward function pays off work?  Instead of rewarding based on humans, pay out all rewards when the vault is checked (at a time unknown to the AI).  The AI isn't asked if the diamond is present or absent.  Instead, it is asked "If the vault were checked now, do you want to be rewarded if the diamond is present or absent.

I think this might still lead to similar problems.  For example this could cause an issue in the case where the diamond has been stolen but the AI believes humans would not be able to tell even if they physically entered the vault and checked, e.g the diamond has been replaced with a very convincing fake.  In this case the AI might still say "I want to be rewarded if the diamond is still present" since it knows humans won't be able to tell the difference.

I'm a newcomer to this, I lack much of the background, and I'm probably suggesting a solution that's too specific to this diamond heist scenario.  But, I already spent an hour writing it down, so I might as well share it.

Trusted timestamping, cryptographically secure sensor

This is a very basic "builder move", I guess?  The idea is to simply improve our sensors so that it's very hard to tamper with them, through public-private key encryption.  The diamond will have a small chip that constantly sends a cryptographically-signed timestamped life... (read more)

I want to steal the diamond.  I don't care about the chip.  I will detach the chip and leave it inside the vault and then I will run away with the diamond. Or perhaps you say that you attached the chip to the diamond very well, so I can't just detach it without damaging it.  That's annoying but I came prepared!  I have a diamond cutter!  I'll just slice off the part of the diamond that the chip is attached to and then I will steal the rest of the diamond.  Good enough for me :)
The implementation could possibly be extended to cover more weak points. For example, you could cover the diamond with additional chips in all sides.  Or you could make the chip so fragile that it breaks when the diamond is affected by strong enough vibrations (as is likely, with a diamond cutter).  Or you could create more complex (but hard/impossible to tamper with) chips that continuously confirm stuff like "no object has come within 10cm of the diamond" or "the temperature remained regular" or "the weight on the pedestal is exactly X grams". My main proposal here is the concept of having better sensors that can't have their data faked.  I think with enough engineering effort you could cover enough "edge cases" that you can trust the AI system to predict robbery every time robbery happens, because a mistake/deception has improbably low odds of happening.
I'm not an ARC member, but I think assuming that the chip is impossible to tamper with is assuming the conclusion. The task is to train a reporter which accurately reports the presence of the diamond, even if we are unable to tell whether tampering has occurred (e.g. because the AI understands some esoteric physics principle which lets them tamper with the chip in a way we don't understand). See the section on page 6 starting with "You might try to address this possibility by installing more cameras and sensors..."

Potentially silly question: 

In the first counterexample you describe the desired behavior as 

Intuitively, we expect each node in the human Bayes net to correspond to a function of the predictor’s Bayes net. We’d want the reporter to simply apply the relevant functions from subsets of nodes in the predictor's Bayes net to each node in the human Bayes net [...]

After applying these functions, the reporter can answer questions using whatever subset of nodes the human would have used to answer that question.

Why doesn't the reporter skip the step of ma... (read more)

3Mark Xu2y
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.
I see, thanks for answering. To further clarify, given the reporter's only access to the human's nodes is through the human's answers, would it be equally likely for the reporter to create a mapping to some other Bayes net that is similarly consistent with the answers provided? Is there a reason why the reporter would map to the human's Bayes net in particular?
3Mark Xu2y
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.

Edit: think this isn't quite right in general, will try to make it more correct later

Here's a sketch of a strategy for trying to fix Strategy: penalize depending on “downstream” variables. Would appreciate feedback on whether it's modeling the difficulty correctly/seems possibly worth figuring out how to implement

It seems like the problem is:

  • On the training set, there are a number of implicit variables X that are indistinguishable (always all true or always all false)
    • A. Is the diamond safe at time t-1
    • B. Is the diamond safe at time t (the variable we actual
... (read more)

Question: Would a proposal be ruled out by a counterexample even if that counterexample is exponentially unlikely?

I'm imagining a theorem, proved using some large deviation estimate, of the form:  If the model satisfies hypotheses XYZ, then it is exponentially unlikely to learn W. Exponential in the number of parameters, say. In which case, we could train models like this until the end of the universe and be confident that we will never see a single instance of learning W.

I'd be fine with a proposal that flips coins and fails with small probability (in every possible world).

Hello, I have some issue with the epistomology of the problem : my problem is that even if the process of training was giving the behavior we want, we would have no way to check the IA is working properly in practice. 
I try now to give more details : in the volt probleme, given the same information, let's think of an IA that just as to answer the question "Is the diamon still in the volt ?". 

Something we can suppose is that, the set Y, from which we draw the labeled examples to train the IA (a set of technique for the thief), is not importa... (read more)

If the predictor AI is in fact imitating what humans would do, why wouldn’t it throw its hands up at an actuator sequence that is too complicated for humans—isn’t that what humans would do? (I'm referring to the protect-the-diamond framing here.)

As described in the report it would say "I'm not sure" when the human wasn't sure (unless you penalized that). That said, often a human who looks at a sequence of actions would say "almost certainly the diamond is there." They might change their answer if you also told them "by the way these actions came from a powerful adversary trying to get you to think the diamond is there." What exactly the reporter says will depend on some details of e.g. how the reporter reasons about provenance. But the main point is that in no case do you get useful information about examples that a human (with AI assistants) couldn't figure out what was happening on their own.

Naive question: does this scenario include cases of a human physically breaking into the vault at some random times so that sensor information, predictor reports, and outcome to human in this situation would be known?

3Mark Xu2y
Yes. Section Strategy: have a human operate the SmartVault and ask them what happened describes what I think you're asking about.

Silly question warning. 


You think that when an AI performs a bad action, (say remove the diamond) the AI has to have knowledge that the diamond is in fact no longer there.  Even when the camera shows the diamond is (falsely) there and the human confirms that the diamond is there. 

You call this ELK

You want the human to have access to this knowledge, as this is useful to choosing decisions that the human wants.

This is hard.  So you have people propose how to do this.  

And then people try to explain why that strategy wouldn't wo... (read more)

3Ajeya Cotra2y
This broadly seems right. Some details: * The "explain why that strategy wouldn't work" step specifically takes the form of "describing a way the world could be where that strategy demonstrably doesn't work" (rather than more heuristic arguments). * Once we have a proposal where we try really hard to come up with situations where it could demonstrably fail, and can't think of any, we will probably need to do lots of empirical work to figure out if we can implement it and if it actually works in practice. But we hope that this exercise will teach us a lot about the nature of the empirical work we'll need to do, as well as providing more confidence that the strategy will generalize beyond what we are able to test in practice. (For example, ELK was highlighted as a problem in the first place after ARC researchers thought a lot about possible failure modes of iterated amplification.)

You can just make a plecsi glass by 2 by 2 metters, and on the inside to put senzors that are conected to a battery on the inside of the glass box.These senzors to intersept every cm of the glass,moving,breaking the glass being imposible because all the meccanics are on the inside.And if you want to take the diamont from there like an guard man or somethin' just let everyone know you are takeing it from there and let the alarm to ring.When the senzors are going on they will make a signal to the guard Who are protecting it and come to catch the thief .More... (read more)

[+][comment deleted]2y4