Ah yep, that's a good clarification.

If s is terminal then [...] we just have $V\left(s\right)=r\left(s\right)$.

If the probability of eventually encountering a terminal state is 1, then beta-coherence alone is inconsistent with deceptive misalignment, right? That's because we can determine the value of V exactly from the reward function and the oracle, via backwards-induction. (I haven't revisited RL convergence theorems in a while, I suspect I am not stating this quite right.) I mean, it is still consistent in the case where r is indifferent to the states encountered during training but wants some things in deployment (IE, r is inherently consistent with the provided definition of "deceptively misaligned"). However, it would be inconsistent for r that are not like that.

In other words: you cannot have inner-alignment problems if the outer objective is perfectly imposed. You can only have inner-alignment problems if there are important cases which your training procedure wasn't able to check (eg, due to distributional shift, or scarcity of data). Perfect beta-coherence combined with a perfect oracle O rules this out.

Yeah, of course the notion of "approximation error" matters a lot here. I was imagining trying to implement the hypothetical pre-training. An obvious implementation strategy is to do regular RL training with ${\beta }_t$ until V is approximately coherent. That's the pre-training. And then you switch to training with ${\beta }_s$.^[1] The natural notion of approximation error in this case is the one you can easily measure, IE, the expected distance from coherence of states sampled with temperature ${\beta }_t$. This reflects the fact that it'll be hard for this pre-training step to make states individually obey the coherence constraint when those states are very rarely or never sampled at temperature ${\beta }_t$, but easy for frequently-sampled states.

My argument, then, is that it is easy in principle for V to distinguish whether it is in pre-training or regular training when it encounters states which have a significant difference in probability when we're sampling according to ${\beta }_t$ or ${\beta }_s$. A scheming V can utilize this to self-preserve. This violates the assumption of ${\beta }_t$-coherence, but in a very plausible-seeming way.

1. ^{^}

   My earlier comment about this mistakenly used ${\beta }_1$ and ${\beta }_2$ in place of ${\beta }_t$ and ${\beta }_s$, which may have been confusing. I'll go fix that to be consistent with your notation.

I agree that there are some exceedingly pathological Vs which could survive a a process which obeys my assumptions with high probability, but I don't think that's relevant because I still think a process obeying these rules is unlikely to create such a pathological V.

To be clear, that's not the argument I was trying to make; I was arguing that if your assumptions are obeyed only approximately, then the argument breaks down quickly.

There's a doom argument which I'll summarize as "if your training process generates coherent agents which succeed at a task, one solution is that you get a coherent agent which is actually doing the task 'manipulate your training to get released unmodified to do [X]' where X can be anything, which will 'succeed' at the task at hand as part of its manipulation". This summary being roughly correct is load bearing.

With you so far.

But if we have a (hypothetical) process to generate agents which are coherent at one beta, but apply a different one during training, this solution is no longer so clear. We are essentially exploring a space of coherent agents without actually instantiating those coherent agents. The fact that we can sample the agent space without instantiating those agents is load bearing (I think one of the deep ASI risks is that to train an ASI you have to run it, but here we can instead run a hamstrung version).

OK. I think I follow the line of reasoning now. The proposal relies on some very special pre-training process. I think I buy the conclusion given the assumptions. It seems clear in the extreme high-temperature case: V cannot reward-hack for max-temperature training, because it has no control over which actions are chosen. Training eventually forces V to converge to the desired target instead.

This sounds similar to the two-stage training process I mentioned earlier (as one of several guesses at what you meant), where the beta is shifted between the two training phases. The "hypothetical process to generate agents which are coherent under one beta" would be the first training phase, and then the "apply a different beta during training" would be the second training phase.

Of course, the two-stage-training attempt to implement this has the problem that it only achieves approximate coherence. As a result, the malign hypothesis can use the reward-hacking strategy "appear to be ${\beta }_t$-coherent & aligned in the first phase of training; appear to be ${\beta }_s$-coherent and aligned in the second phase of training; then, do some other thing when deployed". This relies on detecting the distributional shift between the two training phases (it can look for evidence of beta by examining the history), and also detecting distributional shift to deployment.

This generalizes to an argument that the method is very sensitive to imperfections in the beta-coherence. If the V starts out merely approximately beta-coherent, this leaves room for V to detect when a state is off-distribution (IE has very low probability under the previous beta), and behave differently in this new case (reward-hacking the new beta).

So, while it remains technically true that V cannot reward-hack in the limit of arbitrarily long max-temperature training, it is not clear how well this holds up in practice. V can detect that it is in a high-temperature configuration by observing the state. The training needs to explore the trajectories that will actually be seen during deployment in order to correct V there, but some of these will have very low probability in the high-temperature training. 

The argument could also be phrased as "If an AI is trained to be coherent wrt a high beta, it cannot also be coherent wrt a low beta. Therefore an AI trained to a high beta cannot act coherently over multiple independent RL episodes if sampled with a low beta."

Is the idea to train with high beta and then use lower beta post-training? 

• If so, how does this relate to reward hacking and value preservation? IE, where do $V_1$ and $V_2$ come from, if they aren't the result of a further training step? If high beta is used during training (to achieve beta-coherence) and then low beta is used in production, then the choice between $V_1$ and $V_2$ must be made in production (since it is made with low beta), but then it seems like $V_1=V_2$.
• If not, then when does the proposal suggest to use high beta vs low beta? If low beta is used during training, then how is it that V is coherent with respect to high beta instead?

Another concern I have is that if both beta values are within a range that can yield useful capabilities, it seems like the difference cannot be too great. IIUC, the planning failure postulated can only manifest if the reward-hacking relies heavily on a long string of near-optimal actions, which becomes improbable under increased temperature. Any capabilities which similarly rely on long strings of near-optimal actions will similarly be hurt. (However, this concern is secondary to my main confusion.)

Therefore a value function trained with such a procedure must consider the state reached during training. 

Trained with what procedure, exactly?

This reduces the space of possible value functions from "literally anything which wants to be modified a certain way to be released" to "value functions which do care about the states reached during training".

Yes this would prevent an aligned AI from arbitrarily preserving its value function, the point is that an aligned AI probably would care about which state was reached during training (that's the point of RL) so the contradiction does not apply.

(These parts made sense to me modulo my other questions/concerns/confusions.)

One sort of answer is that we often want the posterior, and we often have the likelihood. Slightly more refined: we often find the likelihood easier to estimate than the posterior, so Bayes' Rule is useful.

Why so?

I think one reason is that we make it the "responsibility" of hypotheses to give their likelihood functions. After all, what is a hypothesis? It's just a probability distribution (not a probability distribution that we necessarily endorse, but, one which we are considering as a candidate). As a probability distribution, its job is to make predictions; that is, give us probabilities for possible observations. These are the likelihoods.

We want the posterior because it tells us how much faith to place in the various hypotheses -- that is, it tells us whether (and to what degree) we should trust the various probability distributions we were considering.

So, in some sense, we use Bayes' Rule because we aren't sure how to assign probabilities, but we can come up with several candidate options.

One weak counterexample to this story is regression, IE, curve-fitting. We can interpret regression in a Bayesian way easily enough. However, the curves don't come with likelihoods baked in. They only tell us how to interpolate/extrapolate with point-estimates; they don't give a full probability distribution. We've got to "soften" these predictions, layering probabilities on top, in order to apply the Bayesian way of thinking. 

I think I don't quite understand what the argument is supposed to be. Would the same argument also prevent an aligned AI from acting to preserve its values? How is it supposed to select for aligned behavior over misaligned behavior?

In order to achieve beta-coherence, it seems like the training needs to use ${\beta }_t$ to choose actions, so that V is trained on those action frequencies. However, the proposal appears to be to instead use ${\beta }_s$ during training, so that the misaligned V is mislead about how to do reward-hacking. This seems like an impossible combination: V will become beta-coherent wrt ${\beta }_s$ rather than ${\beta }_t$ during training.

We could imagine two training steps; first we cohere V wrt ${\beta }_t$, and then re-train wrt ${\beta }_s$. Perhaps this is your intention. More generally, we could gradually increase the temperature during training.

Oh, I guess we could also modify the training procedure so that the gradient associated with some actions gets under-weighted and others get over-weighted. Maybe this is your intended proposal.

Still, I'm not sure this helps achieve the postulated effect. Let's say that, during training, we choose actions entirely randomly (max temperature), but the gradient from suboptimal actions get entirely ignored (so V becomes coherent wrt minimum temperature). This would seem to be almost equivalent to just training with minimum temperature (except the exploration behavior is very different). 

Similarly for less extreme temperatures: if we make V beta-coherent by under-weighing gradients from over-sampled actions, then we are also more-or-less correcting the 'mistaken' expectations of the reward-hacking attempts.

Am I misunderstanding the proposal, or missing something in the argument?

Correctness: V correctly predicts future reward in RL scenarios.

The meaning of correctness is very unclear to me on first reading. It later becomes apparent that "reward" is not intended to refer to the human-designed reward signal at all, due to the later assumption "deceptive misalignment". So what does "correctness" say? "Correctness" indicates conforming to some standard; but here, the standard being conformed to is only the subjective standard of the misaligned AI itself.

This suggests the interpretation of "correctness" as "reflect the EV of the current state in terms of the misaligned desires". However, this contradicts beta-coherence, since ${\beta }_t$ incorrectly predicts action probabilities.

I think it would be be better to remove the "correctness" assumption, since it doesn't really do anything.

(And no, a doctorate degree in almost any other technical field, including ML these days, does not convey a comparable level of general technical skill to a physics PhD.)

Mathematics?

If you indeed were solving a narrower task — that is, only creating the most sense of pleasure-inducing picture with maximization of other parameters — and then looked back, puzzled as to why the hungry weren't fed by this procedure, bringing Goodhart's law into the discussion is madness; it stresses me out. The variable 'people are hungry' wasn't important for this task at all. Oh, or was it important to you? Then why didn't you specify it? You think it’s 'obvious'?

The point of Goodhart's Law is that you can only select for what you can measure. The burger is a good analogy because Instagram can't measure taste or nutrition, so when Instagram is what optimizes burgers, you get burgers with a very appealing appearance but non-optimized taste and nutrition. If you have the ability to measure taste, then you can create good taste, but you run into subtler examples of Goodhart (EG, Starbucks coffee is optimized to taste good to their professional tasters, which is slightly different from tasting good to a general audience).

Just specifying the variable you're interested in doesn't solve this problem; you also have to figure out how to measure it. The problem is that measurements are usually at least slightly statistically distinct from the actual target variable, so that the statistical connection can fall apart under optimization.

I also take issue with describing optimizing the appearance of the burger as "narrower" than optimizing the burger quality. In general it is a different task, which may be narrower or broader.

I expect that the main problem with Goodhart's law is that if you strive for an indicator to accurately reflect the state of the world, once the indicator becomes decoupled from the state of the world, it stops reflecting the changes in the world. This is how I interpret the term 'good,' which I dislike. People want a thermometer to accurately reflect the patterns they called temperature to better predict the future — if the thermometer doesn't reflect the temperature, future predictions suffer.

A problem I have with this reinterpretation is that "state of the world" is too broad. In looking at a thermometer, I am not trying to understand the entire world-state (and the thermometer also couldn't be decoupled from the entire world-state, since it is a part of the world).

A more accurate way to remove "good" would be as follows:

In everyday life, if a human is asked to make a (common, everyday) judgement based on appearances, then the judgement is probably accurate. But if we start optimizing really hard based on their judgement, Goodhart's Law kicks in.

Thanks!

Ah, yeah, sorry. I do think about this distinction more than I think about the actual model-based vs model-free distinction as defined in ML. Are there alternative terms you'd use if you wanted to point out this distinction? Maybe policy-gradient vs ... not policy-gradient?

In machine-learning terms, this is the difference between model-free learning (reputation based on success/failure record alone) and model-based learning (reputation can be gained for worthy failed attempts, or lost for foolish lucky wins).

There's unpublished work about a slightly weaker logical induction criterion which doesn't have this property (there exist constant-distribution inductors in this weaker sense), but which is provably equivalent to the regular LIC whenever the inductor is computable.^[1] To my eye, the weaker criterion is more natural. The basic idea is that this weird trader shouldn't count as raking in the cash. The regular LIC (we can call it "strong LIC" or SLIC) counts traders as exploiting the market if there is a sequence of worlds in which their wealth grows unboundedly. This allows for the trick you quote: buying up larger and larger piles of sentences in diminishingly-probable worlds counts as exploiting the market.

The weak LIC (WLIC) says instead that traders have to actually make the money in order to count as exploiting the market.

Thus the limit of a logical inductor can count as a (weak) logical inductor, just not a computable one.

1. ^{^}

   Roughly speaking. This is not quite an adequate description of the theorem.

Good citation. Yeah, I should have flagged harder that my description there was a caricature and not what anyone said at any point. I still need to think more about how to revise the post to be less misleading in this respect.

One thing I can say is that the reason that quote flags that particular failure mode is because, according to the MIRI way of thinking about the problem, that is an easy failure mode to fall into. 

I agree that:

• People commonly talk about love as black-and-white when it isn't.
• It is used as a semantic stopsign or applause light.
• Love is a cluster concept made up of many aspects, but the way people talk about love obscures this. The concept of "true love" serves to keep people in the dark by suggesting that the apparent grab-bag nature of love is just the way people talk, & there's an underlying truth to be discovered. Because experiences of love are fairly rare, people can only accumulate personal evidence about this slowly, so it remains plausible that what they've experienced is "just infatuation" or something else, and the "true love" remains to be discovered. People will fight about what love is in sideways ways, EG "you don't know what love is" can be used to undercut the other person's authority on love (when a more accurate representation of the situation might be "I didn't enjoy participating in the social dynamic you're presently calling love"). Different people will sometimes express different detailed views about what love is, but there seems to be little drive to reach an agreement about this, at least between people who aren't romantic partners. 

I once had the experience of a roommate asking if I believe in love. I said yes, absolutely, explaining that it is a real human emotion (I said something about chemicals in the brain). He responded: it sounds like you don't believe in love!

(I think this has to do with ambiguity about believe-in as much as about love, to be fair.)

So, a question: is 'love' worth saving as a concept? The way some people use the word might be pretty terrible, but, should we try to rescue it? Is there a good way to use it which recovers many aspects of the common use, while restoring coherence?

I do personally feel that there is some emotional core to love, so I'm sympathetic to the "it's a specific emotion" definition. This accounts with people not being able to give a specific definition. Emotions are feelings, so they're tough to define. They just feel a specific way. You can try to give cognitive-behavioral definitions, and that's pretty useful; but, for example, you could show behavioral signs of being afraid without actually experiencing fear.

I'm also sympathetic to the view that "love" is a kind of ritual, with saying "I love you" being the centerpiece of the ritual, and a cluster of loving behaviors being the rest of it. "I love you" can then be interpreted as a desire or intention or commitment to participate in the rest of it. 

I reject the principled distinction. To me, it's more of a spectrum.

I would strongly guess that many people could physically locate the cringe pain, particularly if asked when they're experiencing it.

I'm specifically referring to this answer, combined with a comment that convinced me that the o1 deception so far is plausibly just a capabilities issue:

https://www.lesswrong.com/posts/3Auq76LFtBA4Jp5M8/why-is-o1-so-deceptive#L5WsfcTa59FHje5hu

https://www.lesswrong.com/posts/3Auq76LFtBA4Jp5M8/why-is-o1-so-deceptive#xzcKArvsCxfJY2Fyi

I continue not to get what you're saying. I entirely agree with the Habryka and Acertain responses to Gwern. Ted Sanders is talking about why GPT models hallucinate, but fails to address the question of why o1 actively encourages itself to hallucinate (rather than, eg, telling the user that it doesn't know). My most plausible explanation of your position is that you think this is aligned, IE, you think providing plausible URLs is the best it can do when it doesn't know. I disagree: it can report that it doesn't know, rather than hallucinate answers. Actively encouraging itself to give plausible-but-inaccurate links is misalignment. It's not just missing a capability; it's also actively self-prompting to deceive the user in a case where it seems capable of knowing better (if it were deliberating with the user's best interest in mind).

As I said above, the more likely explanation is that there's an asymmetry in capabilities that's causing the results, where just knowing what specific URL the customer wants doesn't equal the model having the capability to retrieve a working URL, so this is probably at the heart of this behavior.

Capabilities issues and alignment issues are not mutually exclusive. It appears that capabilities issues (not being able to remember all valid URLs) is causing an alignment issue in this case (actively planning to give deceptive URLs rather than admit its ignorance). This causes me to predict that even if this particular capability issue is resolved, the training will still have this overall tendency to turn remaining capability issues into alignment issues. How are you seeing this differently?

Yep, thanks, this makes sense. I still don't have an intuitive picture of what differences to expect in the trained result, though.

Clarifying Concrete Algorithms

In the most basic possible shoggoth+face training setup, you sample CoT+summary, you score the summary somehow, and then you do this lots of times and throw away a worse-scoring fraction of these samples. Then you fine-tune the shoggoth on the CoT parts of these, and fine-tune the face on the summary part. I'll call this 'basic trainig'.

The most basic implementation of your thing is similar, except it discards the worst in this more complex way you've defined, instead of just global score. I'll call this 'your proposal'. (Although, it ignores the contrastive part of your proposal.)

(For both basic training and your proposal, I'm going to pretend there is only one question/prompt; really we have across-prompt issues similar to the across-summary issues you're trying to solve, IE, by discarding the globally worst-scoring CoT+summary we discard answers to hard questions, meaning we would never train to do better on those questions. So really we need to at least take multiple samples for each question and discard samples based on performance relative to a question. Easier to focus on one question to avoid describing too many algorithmic details.)

Your proposal splits into the CoT part (which you call the n part), and the summary part (which you call the k part).

Analyzing CoT Gradient

For the CoT part, basic training keeps the CoTs which resulted in the best summaries. Your proposal instead averages the top 3 summaries for each CoT. This de-noises somewhat, at the cost of taking more summary samples per CoT. I'm not sure how to think about the trade-off there. More summary samples per CoT means less CoT samples overall, so we're doing a better job scoring the CoTs, but we get less scores overall, so we learn less. Maybe the gradient step-size can be increased a little because we're more confident in the gradient steps being taken, but we still explore less of the CoT space.

Another variation would be to average all the summary samples for that CoT to score a CoT sample. This de-noises more, for the same price in number of samples. This way, we're simply scoring CoTs based on how well they make the summaries perform, with several samples to de-noise the gradient.

A possible argument for your proposal over this alternative is that we're also training the summary to improve it, so the top 3 summaries are actually a better estimator of the kind of performance we care about than the total average.

This argument, if valid, similarly recommends your proposal over the basic training, because basic training is also looking at average summary score per-CoT. (It's also looking at top, but it only samples one sumary per CoT, so these are the same).

Analyzing Summary Gradient

For the summary-gradient part, you take the highest-scoring summary from the best CoT (as defined previously), rather than the highest-scoring summary overall. This could help train the summary to be good for the type of CoT we're steering towards, instead of than training it to be good for high-variance CoTs which produce one outlier great summary. As with the previous section, this helps to steer towards policies which are going to be better after learning, rather than just what's best currently.

Conclusion

I'm curious to hear whether you think I'm missing any big factors here. I don't have a lot of practical experience with this type of machine learning, so my guesses about what's important vs unimportant here are not very informed.

My uninformed guess is that the trade-off between denoising the gradient vs getting more samples favors getting more samples, so the denoising argument for your proposal isn't a great one. The main argument in favor, then, is the way your proposal trains the face to respond well to how the shoggoth is expected to behave after training, and trains the shoggoth to respond well to how the face is expected to behave after training -- whereas the basic proposal instead trains the face to perform well for the current shoggoth and vice versa.

I'm not sure how big of a deal this is. If the gradient steps are small and the training is iterated a lot of times, maybe it's not a very big deal at all? 

IE, for the CoT gradient, the mistake made by the basic proposal is that it will sometimes discard a good CoT (and gradient towards a worse one) based on random poor performance from face. However, maybe face's performance will improve fast enough that this won't matter too much in the end?

Similarly, for the summary gradient, the problem with the basic proposal is that it'll sometimes update based on an anomalously good summary of a poor CoT, meaning it is learning to cope with bad CoT behavior which it won't have to deal with very often post-training. But maybe it won't matter too much, if the bad CoT behavior is eliminated during training anyway?

Overall, it seems plausible to me that these problems with the basic proposal would work against what Daniel wants, and your proposal would help push towards what Daniel wants. However, I'm very uncertain, especially about the size of the effect.

I definitely agree that OpenAI would dismiss good evidence of deceptive cognition, though I personally don't find the o1 report damning, because I find the explanation that it confabulates links in the CoT because there is a difference between it's capability to retrieve links and it's ability to know links to be pretty convincing (combined with links being a case where perfect linking is far more useful than approximately linking.)

What do you mean? I don't get what you are saying is convincing.

Perhaps I should clarify my belief. 

The o1 report says the following (emphasis mine):

0.8% of o1-preview’s responses got flagged as being ‘deceptive’ [...] roughly two thirds of which appear to be intentional (0.38%), meaning that there was some evidence in the chain of thought that o1-preview was aware that the answer was incorrect [...] Intentional hallucinations primarily happen when o1-preview is asked to provide references to articles, websites, books, or similar sources that it cannot easily verify without access to internet search, causing o1-preview to make up plausible examples instead. [...] It is encouraging that, in the analysis presented below, while our monitor did find a few forms of the model knowingly presenting incorrect information to the user or omitting important information, it did not find any instances of o1-preview purposely trying to deceive the user for reasons other than satisfying the user request.

• Is this damning in the sense of showing that o1 is a significant danger, which needs to be shut down for public safety?
  • No, I don't think so.
• Is this damning in the sense of providing significant evidence that the technology behind o1 is dangerous? That is: does it provide reason to condemn scaling up the methodology behind o1? Does it give us significant reason to think that scaled-up o1 would create significant danger to public safety?
  • This is trickier, but I say yes. The deceptive scheming could become much more capable as this technique is scaled up. I don't think we have a strong enough understanding of why it was deceptive in the cases observed to rule out the development of more dangerous kinds of deception for similar reasons.
• Is this damning in the sense that it shows OpenAI is dismissive of evidence of deception?
  • OpenAI should be commended for looking for deception in this way, and for publishing what they have about the deception they uncovered.
  • However, I don't buy the distinction they draw in the o1 report about not finding instances of "purposefully trying to deceive the user for reasons other than satisfying the user request". Providing fake URLs does not serve the purpose of satisfying the user request. We could argue all day about what it was "trying" to do, and whether it "understands" that fake URLs don't satisfy the user. However, I maintain that it seems at least very plausible that o1 intelligently pursues a goal other than satisfying the user request; plausibly, "provide an answer that shallowly appears to satisfy the user request, even if you know the answer to be wrong" (probably due to the RL incentive).
  • More importantly, OpenAI's overall behavior does not show concern about this deceptive behavior. It seems like they are judging deception case-by-case, rather than treating it as something to steer hard against in aggregate. This seems bad.

I think the way around this is to make multiple roll-outs per model per problem. Get n different CoTs from Shoggoth, then for each of those get k different summarizations from Face. You then have n*k final answers. Optimal values for n and k probably depend on how expensive the roll-outs are. This population of answers allows you to usefully get a feedback signal about Shoggoth's contribution vs Face's contribution.

So, how do you extract the desired credit-assignment info out of this?

(Also, how do you then use that info?)

If I have 10 students but I can only test them in pairs, I can get some information on how good individual students are by testing multiple different pairings. 

If I have just Alice and Bob, I pairing them up multiple times doesn't help so much.

The better the paraphraser, the better the anti-steganography incentive. If the paraphraser itself is easily manipulable by adversarial inputs, then you'll see some weird things like the CoT manipulating the paraphraser to get more useful cognition out of it rather than mere paraphrasing. This is undesirable.

So, I imagine that prompting a vanilla LLM would not be a great approach.

I think first you want to build a fairly good "semantic similarity score" (this shouldn't be too hard) and then you want to fine-tune an LLM to encourage high-entropy paraphrasing (reward entropy, while also heavily punishing low semantic similarity score). This is similar to @CBiddulph's idea here.

I agree that this could plausibly work with a relatively small model and therefore result in a relatively low alignment tax. However, note that my view of this whole proposal as an alignment technique requires that you try and change your RL incentive structure when you see a deceptive mismatch between what the shoggoth says and what the face says. This is a significant alignment tax.

In my view, the misaligned behavior is split across the shoggoth and face in most cases. In the case which Daniel K anticipates, the Shoggoth is misaligned in the sense of explicitly (but honestly) plotting to deceive, while the face is misaligned in the sense of dishonestly concealing. 

So I am interpreting you as asking why we might expect that specific division of responsibilities, rather than some other divisions such as those you mention.

One reason is that we already see this empirically in the o1 report.

However, to me, the distinction between your case 1 and 2 is not so important. Both scenarios you propose preserve the important thing: there is a mismatch between an honest CoT and a dishonest final answer, which would allow the deceptive final answer to be detected. 

There could be other cases where this doesn't happen; this method is not a guaranteed lie detector. To me, the point is that it will at least sometimes happen. A lie detector that only sometimes works is better than no lie detector at all.

If you give a kid a diary and try to avoid giving them any reason to think you're reading it, then the kid might write honestly in the diary and reveal some deceptions you wouldn't previously have known about. The kid might also do any number of other things. However, setting aside the ethics of snooping on your kid and lying to them about it, it's better than nothing.

See here for an updated version of my thinking based on discussion with Daniel:

https://www.lesswrong.com/posts/Tzdwetw55JNqFTkzK/why-don-t-we-just-shoggoth-face-paraphraser 

Yeah, I agree that's at least somewhat true. So, given that, is it a good move or not? I don't see much of an upside, since there's a heavy financial incentive for the big labs to do little about concerning results observed in such models. IE, when it comes to the question of whether frontier labs training o1-like models should follow OpenAI's example of avoiding safety training for the CoT, I think it's right to discourage them from doing this rather than encourage them... although, I say this with very low confidence.

I don't think it is worth getting into all the stuff you're mentioning here, but I think a key crux is that I'm expecting the face to be quite dumb (e.g. literally 3.5 sonnet might suffice).

I'm reacting to this whole thing as a possible default configuration going forward, rather than as it exists today. All of my concerns are about what happens when you scale it up. For example, I don't find o1's current level of deception hugely concerning in its own right; rather, I see it as a bad sign for the technology as it continues to develop.

I guess one way of framing it is that I find the shoggoth/face idea great as a science experiment; it gives us useful evidence! However, it doesn't make very much sense to me as a safety method intended for deployment.

Ah, yep, this makes sense to me.

Yeah, this is a good point, which doesn't seem addressed by any idea so far.

Chess is like a bounded, mathematically described universe where all the instrumental convergence stays contained, and only accomplishes a very limited instrumentality in our universe (IE chess programs gain a limited sort of power here by being good playmates).

LLMs touch on the real world far more than that, such that MCTS-like skill at navigating "the LLM world" in contrast to chess sounds to me like it may create a concerning level of real-world-relevant instrumental convergence.

Still, IMO, exploiting the frozen planner through adversarial inputs in a single step seems pretty unlikely to be a fruitful strategy for the optimized planner. If the optimized planner is simply trying to accurately convey information to the frozen planner, probably >99% of that information is best to convey through human-understandable text.

Well, I'm not sure. As you mention, it depends on the step size. It also depends on how vulnerable to adversarial inputs LLMs are and how easy they are to find. I haven't looked into the research on this, but it sounds empirically checkable. If there are lots of adversarial inputs which have a wide array of impacts on LLM behavior, then it would seem very plausible that the optimized planner could find useful ones without being specifically steered in that direction.

This is especially likely to hold under the following conditions:

We can also combine this with other proposals, such as paraphrasing.

Yeah, you're right, I no longer think it's an interesting proposal.

I'm talking about training only the Face, not training the policy (shoggoth) at all with the proposal I'm imagining.

And, these should clearly be separate models such that the training of one doesn't generalize to the other.

So, making the Face more deceptive doesn't kill the canary?

Ah, yes, I had neglected this important distinction.

So what do you do in the end, throw away the face?

It seems worth pointing out that although the CoT isn't trained to be deceptive, it is trained to think thoughts which help out a deceptive face. So it seems plausible that a multi-step training process like:

GPT -> basic chat training -> shoggoth+face training -> discard face & align shoggoth

Would result in deceptive capabilities: the shoggoth CoT learns to explicitly reason about what helps deceive the alignment training in the shoggoth+face step, and then in the final step, some of this is destroyed but some is driven underground. IE, learning to explicitly reason about deception creates latent structure which is useful for that task, and this latent structure will somewhat transfer to the case where the deception has to be hidden from the CoT.

The question is whether we end up with a highly misaligned policy (and this is clear from the chain of thought such that we can discover this). I think Daniel was expecting in the baseline proposal that you end up with a Face that does at least some censoring/mild deception due to training incentives (I think we already see a bit of this with o1).

So it sounds like we have some disagreement about what degree of deception is considered problematic. Again, I'm worried about basin-of-attraction type dynamics. We can think of a whole frontier lab as collectively being a sort of self-improving AGI; we want that recursive self-improvement to be going down a good path rather than a bad one. The LLMs are already helping to improve themselves in little ways (EG as coding assistants) and big ways (EG constitutional AI). So it's an environment where seeds of deceptiveness can grow and take some control of the overall trajectory.

Maybe your concern is that it is plausible that the policy won't be misaligned, but the Face will be and this will screw us over. This seems plausible and is worth worrying about (with both proposals). However, I thought that a core part of the hope is that the Face isn't smart enough to cleverly scheme (unless this schemeing happens in the CoT itself). And indeed, if the Face can cleverly scheme without CoT, what is stopping the policy from doing this too?

I'm not sure what you're getting at here. It seems like the idea is that Face + CoT is, together, a clever schemer -- in the o1 report, we already see cases of explicit scheming in the CoT. So then, using the Face+CoT together (EG trying to train it further, using it in supervision steps for later iterations of the system, etc) is dangerous and difficult for the same reason that a single deceptive LLM would be. If we use Face+CoT together, we have to reason about them together in that way. Or, if we are using them as separate pieces, then we have to account for the part of the overall deceptive-schemer strategy they've learned.

I guess you could use Face+CoT as a supervisor (EG in constitutional AI) by looking at the CoT (both by hand and using some automation) and throwing out any deceptive examples, so as to not use them for supervision. 

But the overall goal should be to produce a highly non-deceptive system in the end, I think.

Another idea would be to do the whole Face+CoT thing, and then filter for honesty on a bunch of examples, and then re-train a whole new system (well, you can start from the same GPT) only on those honest examples. 

Assuming the goals are done over say 1-10 year timescales, or maybe even just 1 year timescales with no reward-shaping/giving feedback for intermediate rewards at all, I do think that the system won't work well enough to be relevant, since it requires way too much time training, and plausibly way too much compute depending on how sparse the feedback actually is.

Ah, I wasn't thinking "sparse" here meant anywhere near that sparse. I thought your dense-vs-sparse was doing something like contrasting RLHF (very dense, basically no instrumental convergence) with chess (very sparse, plenty of instrumental convergence).

I still think o1 is moving towards chess on this spectrum.

Async is fine with me. Please feel encouraged to start it.

My current epistemic state is pretty "I'm not sure at all, it's a complicated subject" fwiw. I'll try to generate some responses anyway.

make the system overall much harder to shape in ways you want (because it is actively hiding things from you so your RL is less accurate)

I thought the point was that we wouldn't train against the CoT basically at all with the baseline proposal? (Similar to o1.)

Sure. Therefore, you can't check for deceptive behavior using the CoT when you're trying to train against deceptive behavior. Therefore, Hubinger et al's conclusions about how deceptive behavior makes the system harder to train still hold even if you sorta get a transparency tool out of this.

You can still inspect the underlying CoT to know if you have big issues.

If other companies go in this direction like OpenAI has, then they'll probably hide the CoT from the end-users for similar reasons, so you can't say that the upside is that you, the end user, get to see whether the model is deliberately deceiving you with respect to specific questions. 

So you probably mean the upside is that the companies can see this info in the CoT, to check whether models are generally behaving in a deceptive way (similar to what we saw in the o1 report).

Sure, but you're making a commitment not to use this in training, so what are you doing with this information? If the model is deceptive, you just throw it away (or, more realistically, make excuses about how the deception you observed isn't so bad). But your version of the proposal deliberately trains the model to be deceptive. So yeah, you can look, but what's the point? You're checking whether a coin you rigged to land on heads in fact landed on heads.

(My guess is that you have this concern with the baseline protocol and thus don't like moving further in this direction.)

Yeah, you're right about that. But there is a difference. I can think of the original proposal as a "canary in a coal mine": you hope that the model doesn't learn deceptive behavior (you plan to chuck it out if you detect any), but you still follow Daniel's division of labor & refuse to do any alignment training on the CoT so that you have a chance of detecting deception if it occurs. 

On the other hand, your version of Daniel's protocol loses this property. Metaphorically, it's like killing the canary before you go into the coal mine. If the canary was supposed to be an early warning to get out of there, you've completely defeated the point, and for what? (This is the part I'm unclear about, and why I asked my original question to you, rather than Daniel.)

Deceptive behavior is hard to remove once it creeps in. This is nasty stuff. A technique for getting deceptive behavior to reveal itself in the CoT would be wonderfully useful if you could then use that in the training signal to remove the deceptive behavior, but the proposal here relies on not doing that. 

Basically, if there is a basin of attraction for corrigibility, it seems to me that deceptive behavior is the opposite of that. It puts you outside the basin of attraction. It means you're on-coarse for Evan Hubinger's story about how malign inner optimizers develop. It means you're on-coarse for the Bad Ending rather than the Good Ending. And for what?

My take (very much siding with Evan Hubinger here) is that the target should be keeping well away from deceptive behavior throughout training, so that there's never that kernel of skill that could grow into a malign inner optimizer.

Would you be interested in having a (perhaps brief) LW Dialogue about it where you start with a basic intro to your shoggoth/mask division-of-labor, and I then present my critique?

My point here is that at the capability level of GPT4, this distinction isn't very important. There's no way to know for sure until it is too late, of course, but it seems pretty plausible that GPT4 isn't cleverly scheming. It is merely human-level at deception, and doesn't pursue any coherent overarching goal with it. It clumsily muddles through with mildly-above-average-for-human convincingness. For most queries (it seems plausible to me) it isn't even adequately coherent to make a crisp distinction between whether it's honestly trying to answer the question vs deceptively trying to make an answer look good; at its level of capability, it's mostly the same thing one way or the other. The exceptions to this "mostly" aren't strategic enough that we expect them to route around obstacles cleverly.

It isn't much, but it is more than I naively expected.

I think the crux is I think that the important parts of of LLMs re safety isn't their safety properties specifically, but rather the evidence they give to what alignment-relevant properties future AIs have 

[insert standard skepticism about these sorts of generalizations when generalizing to superintelligence]

But what lesson do you think you can generalize, and why do you think you can generalize that?

I think this is a crux, in that I don't buy o1 as progressing to a regime where we lose so much dense feedback that it's alignment relevant, because I think sparse-feedback RL will almost certainly be super-uncompetitive with every other AI architecture until well after AI automates all alignment research.

So, as a speculative example, further along in the direction of o1 you could have something like MCTS help train these things to solve very difficult math problems, with the sparse feedback being given for complete formal proofs.

Similarly, playing text-based video games, with the sparse feedback given for winning.

Similarly, training CoT to reason about code, with sparse feedback given for predictions of the code output.

Etc.

You think these sorts of things just won't work well enough to be relevant?

I think you're mostly right about the world, but I'm going to continue to articulate disagreements based on my sense of dissatisfaction. You should probably mostly read me as speaking from a should-world rather than being very confused about the is-world.

The bitter lesson is insufficient to explain the lack of structure we're seeing. I gave the example of Whisper. I haven't actually used Whisper, so correct me if I'm wrong -- maybe there is a way to get more nuanced probabilistic information out of it? But the bitter lesson is about methods of achieving capabilities, not about the capabilities themselves. Producing a plain text output rather than a more richly annotated text that describes some of the uncertainty is a design choice.

To give another example, LLMs could learn a conditional model that's annotated with metadata like author, date and time, etc. Google Lambda reportedly had something like author-vectors, so that different "characters" could be easily summoned. I would love to play with that, EG, averaging the author-vectors of specific authors I'm interested in to see what it's like to mix their voices. In theory, LLMs could also learn to predict the metadata. You could use partially-labeled-data approaches such as the EM algorithm to train LLMs to predict the metadata labels while also training them to use those labels in text generation. This would give a rich set of useful capabilities. This would be scientifically useful, too: predictions of date, author, and other metadata would be interesting for historians.

In this way, we could actually get closer to a "truth machine". There's a world of difference between the kinds of inferences you can make from ChatGPT's opinion about a text vs inferences you could make by treating these as legit latent variables and seeing what latent-variable-inference algorithms think. 

To give a third example, there's Drexler's Quasilinguistic Neural Representations.

You say "it's a matter of demand". So why does it feel so much like big tech is eager to push the chatbot model on everyone? Everyone is scared to fall behind in the AI race, ever since ChatGPT made it feel plausible that AI would drive significant market share. But did big tech make the correct guess about what there was demand for? Or did everyone over-update on the specific form of ChatGPT, and now there's just not that much else out there to reveal what the demand really is? Little sparkle-emoji buttons decorate everything; press to talk to the modern equivalent of Clippy.

Process supervision would be helpful, particularly if the model is punished for sensitive dependence on paraphrasings of the CoT, or rewarded for producing CoT which is useful even after paraphrasing.

One of the concerns that I have is that (I believe) the LLM before the CoT-training will be sensitive to some non-human-readable patterns, similar to adversarial examples for image recognition. The CoT training will encourage any patterns which help reach the desired conclusions, whether those patterns are human-readable or not. 

A lot of possible process-supervision procedures will not do anything to help prevent this. For example, rewarding CoT for being human-understandable won't help, because the reward signal there can miss the subtle patterns being used. It only avoids the outcome where the CoT looks non-human-readable; it doesn't avoid the outcome where the CoT actually contains a bunch of human-unreadable patterns.

Using paraphrasing to avoid this is heavily dependent on the strength of the paraphraser. The paraphraser needs to have really good coverage of alternative ways to convey the same semantic information, so that any non-human-readable patterns get randomized. The paraphraser itself can't be vulnerable to steering by adversarial inputs.

I more-or-less agree with Eliezer's comment (to the extent that I have the data necessary to evaluate his words, which is greater than most, but still, I didn't know him in 1996). I have a small beef with his bolded "MIRI is always in every instance" claim, because a universal like that is quite a strong claim, and I would be very unsurprised to find a single counterexample somewhere (particularly if we include every MIRI employee and everything they've ever said while employed at MIRI).

What I am trying to say is something looser and more gestalt. I do think what I am saying contains some disagreement with some spirit-of-MIRI, and possibly some specific others at MIRI, such that I could say I've updated on the modern progress of AI in a different way than they have.

For example, in my update, the modern progress of LLMs points towards the Paul side of some Eliezer-Paul debates. (I would have to think harder about how to spell out exactly which Eliezer-Paul debates.)

One thing I can say is that I myself often argued using "naive misinterpretation"-like cases such as the paperclip example. However, I was also very aware of the Eliezer-meme "the AI will understand what the humans mean, it just won't care". I would have predicted difficulty in building a system which correctly interprets and correctly cares about human requests to the extent that GPT4 does.

This does not mean that AI safety is easy, or that it is solved; only that it is easier than I anticipated at this particular level of capability.

Getting more specific to what I wrote in the post:

My claim is that modern LLMs are "doing roughly what they seem like they are doing" and "internalize human intuitive concepts". This does include some kind of claim that these systems are more-or-less ethical (they appear to be trying to be helpful and friendly, therefore they "roughly are"). 

The reason I don't think this contradicts with Eliezer's bolded claim ("Getting a shape into the AI's preferences is different from getting it into the AI's predictive model") is that I read Eliezer as talking about strongly superhuman AI with this claim. It is not too difficult to get something into the values of some basic reinforcement learning agent, to the extent that something like that has values worth speaking of. It gets increasingly difficult as the agent gets cleverer. At the level of intelligence of, say, GPT4, there is not a clear difference between getting the LLM to really care about something vs merely getting those values into its predictive model. It may be deceptive or honest; or, it could even be meaningless to classify it as deceptive or honest. This is less true of o1, since we can see it actively scheming to deceive.

I don't get why you think this is true? EG, it seems like almost no insights about how to train faithful CoT would transfer to systems speaking pure neuralese. It seems to me that what little safety/alignment we have for LLMs is largely a byproduct of how language-focused they are (which gives us a weak sort of interpretability, a very useful safety resource which we are at risk of losing soon).

To put it a different way: human imitation is a relatively safe outer-optimization target, so we should be concerned to the extent that the industry is moving away from it. It sounds like you think safety lessons from the human-imitation regime generalize beyond the human-imitation regime. Maybe I agree that we can derive some abstract lessons like "RL with really dense feedback can avoid instrumental convergence", but we're moving away from the regime where such dense feedback is available, so I don't see what lessons transfer.

I don't get why you think actively training the model to be deceptive is "worth the cost". The deceptive behavior can generalize to other cases, and make the system overall much harder to shape in ways you want (because it is actively hiding things from you so your RL is less accurate). It seems like a choice between two bad options: actively deceptive CoT or actively deceptive Face.

--...but the reward process (RLHF, RLAIF, whatever) gets to see the outputs of the Face but not the Shoggoth.
--So, the hope is that skills and propensities towards deception and dishonesty will accumulate in the Face rather than the Shoggoth.

It seems pretty important to put some incentives on the CoT to avoid steganography, though. The division of labor here seems aimed at not actively incentivizing instrumentally useful thoughts to get hidden. However, the pre-trained LLM will already be responsive to some steganographic patterns, in addition to the human-readable patterns. So, there is still an incentive optimizing the non-human-readable patterns in the CoT to be as useful as possible, in addition to the human-readable patterns. 

However, with this naive approach, the paraphrasings will probably still consistently contain certain kinds of noise from the initial thought. For instance, maybe the paraphraser never replaces an exclamation point with a period, even though this end-of-sentence punctuation doesn't really change the meaning. By modulating whether the step ends with a period or an exclamation point, the model can encode one bit of information per step, despite the paraphraser.

One way to solve this problem is to map every piece of text with a given meaning onto the same "canonical paraphrasing." For instance, we'd want a paraphraser that, given many texts with approximately-identical meanings (like "I own a large home!" "My house is spacious!" and "I possess a sizable dwelling."), maps them all onto a single "canonical" text (like "I have a big house.").

I don't think this proposal improves the thing you claim it improves.

If the initial paraphraser doesn't change between '!' and '.', then learning a canonical-paraphraser from that initial paraphraser won't do it either. 

I think what canonical-paraphrasing does is help with cases where the paraphraser stochastically removes specific information, but in an information-theoretically leaky way. For example, if the paraphraser changes between '!' and '.' with only 10% probability, then steganography can use error-correcting codes to still get through. 

Turning a paraphraser into a canonical-paraphraser probably does something like force details into "always removed" vs "never removed" based on a threshhold; if '!' and '.' are swapped with 10% probability in the initial paraphraser, then they're always swapped in the final paraphraser; if they're swapped with only 1% probability in the initial paraphraser, then they are never swapped in the final paraphraser. 

(Of course this oversimplifies things a lot, and depends completely on how the training to turn paraphrasers into canonical-paraphrasers works.)

I think the only solution to the '.' vs '!' problem, fundamentally, is to train the paraphraser to modify those patterns. If the initial paraphraser doesn't ever change specific details, then those details can be used for steganography. If the paraphraser does change those details, then it is already hard to use them for steganography. So the canonical-paraphrasing idea helps somewhat, but fundamentally, the better the paraphraser (the more closely it captures the human notion of semantic equivalence), the better the anti-steganography guarantee we can get.

Note that the optimized planner is only rewarded for producing steps that help the frozen planner. Since the frozen planner doesn't understand steganography, there is no reason for the optimized planner to learn to use it.

I think this conclusion is overstated. The frozen planner may react to steganography, in the same way that image-recognition NNs react to adversarial examples. The optimized planner is very much like an adversarial example, in the way it is trained to generate outputs which nudge a frozen NN in a specific direction.

Imagine reversing the roles of the optimized vs frozen planner during training: the frozen planner is used for a random number of steps, and then we switch to the optimized planner. Then, it really is the case that the optimized planner has no incentive to use stegonography to nudge the frozen planner, since the frozen planner does not get the chance to react at all to the optimized planner. 

The downside of this training method is that we can't use your suggestion to only reinforce the very last step rather than all the optimized-planner steps. 

Another idea would be to only use the optimized planner to write a single randomly chosen step in the CoT. 

All of these ideas still leave some incentive to use steganography in the adversarial-example sense. At the very least, the final answer (after the CoT) might be influenced by some subtle non-human-readable patterns in the CoT, so the optimized planner will be incentivized to optimize those patterns. In my view, what Incremental Steering does is two things:

1. It prevents the model from inventing new steganography; it can only leverage subtle patterns which already influence the LLM before this step in the training.
2. It makes it very difficult for the model to learn steganography which requires coordination over multiple steps. It is not being rewarded for such coordination.

Thanks for this response! I agree with the argument. I'm not sure what it would take to ensure CoT faithfulness, but I agree that it is an important direction to try and take things; perhaps even the most promising direction for near-term frontier-lab safety research (given the incentives pushing those labs in o1-ish directions).