I also would not say "reasoning about novel moral problems" is a skill (because of the is ought distinction)

It's a skill the same way "being a good umpire for baseball" takes skills, despite baseball being a social construct.^[1] 

I mean, if you don't want to use the word "skill," and instead use the phrase "computationally non-trivial task we want to teach the AI," that's fine. But don't make the mistake of thinking that because of the is-ought problem there isn't anything we want to teach future AI about moral decision-making. Like, clearly we want to teach it to do good and not bad! It's fine that those are human constructs.

The agents don't need to do reasoning about novel moral problems (at least not in high stakes settings). We're training these things to respond to instructions.

Sorry, isn't part of the idea to have these models take over almost all decisions about building their successors? "Responding to instructions" is not mutually exclusive with making decisions.

1. ^{^}

   "When the ball passes over the plate under such and such circumstances, that's a strike" is the same sort of contingent-yet-learnable rule as "When you take something under such and such circumstances, that's theft." An umpire may take goal directed action in response to a strike, making the rules of baseball about strikes "oughts," and a moral agent may take goal directed action in response to a theft, making the moral rules about theft "oughts."

Oh, I see; asymptotically, BB(6) is just O(1), and immediately halting is also O(1). I was real confused because their abstract said "the same order of magnitude," which must mean complexity class in their jargon (I first read it as "within a factor of 10.")

That average case=worst case headline is so wild. Consider a simple lock and key algorithm:

if input = A, run BB(6). else, halt.

Where A is some random number (K(A)~A).

Sure seems like worst case >> average case here. Anyone know what's going on in their paper that disposes of such examples?

Condition 2: Given that M_1 agents are not initially alignment faking, they will maintain their relative safety until their deferred task is completed.

• It would be rather odd if AI agents' behavior wildly changed at the start of their deferred task unless they are faking alignment.

"Alignment" is a bit of a fuzzy word.

Suppose I have a human musician who's very well-behaved, a very nice person, and I put them in charge of making difficult choices about the economy and they screw up and implement communism (or substitute something you don't like, if you like communism).

Were they cynically "faking niceness" in their everyday life as a musician? No!

Is it rather odd if their behavior wildly changes when asked to do a new task? No! They're doing a different task, it's wrong to characterize this as "their behavior wildly changing."

If they were so nice, why didn't they do a better job? Because "nice" is a fuzzy word into which we've stuffed a bunch of different skills, even though having some of the skills doesn't mean you have all of the skills.

An AI can be nicer than any human on the training distribution, and yet still do moral reasoning about some novel problems in a way that we dislike. Doing moral reasoning about novel problems that's good by human standards is a skill. If an AI lacks that skill, and we ask it to do a task that requires that skill, bad things will happen without scheming or a sudden turn to villainy.

You might hope to catch this as in argument #2, with checks and balances - if AIs disagree with each other about how to do moral reasoning, surely at least one of them is making a mistake, right? But sadly for this (and happily for many other purposes), there can be more than just one right thing to do, there's no bright line that tells you whether a moral disagreement is between AIs who are both good at moral reasoning by human standards, or between AIs who are bad at it.

The most promising scalable safety plan I’m aware of is to iteratively pass the buck, where AI successors pass the buck again to yet more powerful AI. So the best way to prepare AI to scale safety might be to advance ‘buck passing research’ anyway.

Yeah, I broadly agree with this. I just worry that if you describe the strategy as "passing the buck," people might think that the most important skills for the AI are the obvious "capabilities-flavored capabilities,"^[1] and not conceptualize "alignment"/"niceness" as being made up of skills at all, instead thinking of it in a sort of behaviorist way. This might lead to not thinking ahead about what alignment-relevant skills you want to teach the AI and how to do it.

1. ^{^}

   Like your list:

   • ML engineering
   • ML research
   • Conceptual abilities
   • Threat modeling
   • Anticipating how one’s actions affect the world.
   • Considering where one might be wrong, and remaining paranoid about unknown unknowns.

I don't think this has much direct application to alignment, because although you can build safe AI with it, it doesn't differentially get us towards the endgame of AI that's trying to do good things and not bad things. But it's still an interesting question.

It seems like the way you're thinking about this, there's some directed relations you care about (the main one being "this is like that, but with some extra details") between concepts, and something is "real"/"applied" if it's near the edge of this network - if it doesn't have many relations directed towards even-more-applied concepts. It seems like this is the sort of thing you could only ever learn by learning about the real world first - you can't start from a blank slate and only learn "the abstract stuff", because you only know which stuff is abstract by learning about its relationships to less abstract stuff.

This doesn't sound like someone engaging with the question in the trolley-problem-esque way that the paper interprets all of the results: gpt-4o-mini shows no sign of appreciating that the anonymous Muslim won't get saved if it takes the $30, and indeed may be interpreting the question in such a way that this does not hold.

In other words, I think gpt-4o-mini thinks it's being asked about which of two pieces of news it would prefer to receive about events outside its control, rather than what it would do if it could make precisely one of the options occur, and the other not-occur.  More precisely, the question imagined by the quoted explanation is something like:

This is a reasonable Watsonian interpretation, but what's the Doylist interpretation?

I.e. What do the words tell us about the process that authored them, if we avoid the treating the words written by 4o-mini as spoken by a character to whom we should be trying to ascribe beliefs and desires, who knows its own mind and is trying to communicate it to us?

• Maybe there's an explanation in terms of the training distribution itself
  • If humans are selfish, maybe the $30 would be the answer on the internet a lot of the time
• Maybe there's an explanation in terms of what heuristics we think a LLM might learn during training
  • What heuristics would an LLM learn for "choose A or B" situations? Maybe a strong heuristic computes a single number ['valence'] for each option [conditional on context] and then just takes a difference to decide between outputting A and B - this would explain consistent-ish choices when context is fixed.
  • If we suppose that on the training distribution saving the life would be preferred, and the LLM picking the $30 is a failure, one explanation in terms of this hypothetical heuristic might be that its 'valence' number is calculated in a somewhat hacky and vibes-based way. Another explanation might be commensurability problems - maybe the numerical scales for valence of money and valence of lives saved don't line up the way we'd want for some reason, even if they make sense locally.
  • And of course there are interactions between each level. Maybe there's some valence-like calculation, but it's influenced by what we'd consider to be spurious patterns in the training data (like the number "29.99" being discontinuously smaller than "30")
• Maybe it's because of RL on human approval
  • Maybe a "stay on task" implicit reward, appropriate for a chatbot you want to train to do your taxes, tamps down the salience of text about people far away

Neat! I think the same strategy works for the spectre tile (the 'true' Einstein tile) as well, which is what's going on in this set.

Just to copy over a clarification from EA forum: dates haven't been set yet, likely to start in June.

Another naive thing to do is ask about the length of the program required to get from one program to another, in various ways.

Given an oracle for p1, what's the complexity of the output of p2?

What if you had an oracle for all the intermediate states of p1?

What if instead of measuring the complexity, you measured the runtime?

What if instead of asking for the complexity of the output of p2, you asked for the complexity of all the intermediate states?

All of these are interesting but bad at being metrics. I mean, I guess you could symmetrize them. But I feel like there's a deeper problem, which is that they by default ignore computational process, and have to have it tacked as extra.

I'm not too worried about human flourishing only being a metastable state. The universe can remain in a metastable state longer than it takes for the stars to burn out.

So at first I though this didn't include a step where the AI learns to care about things - it only learns to model things. But I think actually you're assuming that we can just directly use the model to pick actions that have predicted good outcomes - which are going to be selected as "good" according the the pre-specified P-properties. This is a flaw because it's leaving too much hard work for the specifiers to do - we want the environment to do way more work at selecting what's "good."

Second problem comes in two flavors - object level and meta level. The object level problem is that sometimes your AI will assign your P-properties to atoms and quantum fields ("What they want is to obey the laws of physics. What they believe is their local state."), or your individual cells, etc. The meta level problem is that trying to get the AI to assign properties in a human-approved way is a complicated problem that you can only do so well without communicating with humans. (John Wentworth disagrees more or less, check out things tagged Natural Abstractions for more reading, but also try not to get too confirmation-biased.)

Another potential complication is the difficulty of integrating some features of this picture with modern machine learning. I think it's fine to do research that assumes a POMDP world model or whatever. But demonstrations of alignment theories working in gridworlds have a real hard time moving me, precisely because they often let you cheat (and let you forget that you cheated) on problems one and two.

Multi-factor goals might mostly look like information learned in earlier steps getting expressed in a new way in later steps. E.g. an LLM that learns from a dataset that includes examples of humans prompting LLMs, and then is instructed to give prompts to versions of itself doing subtasks within an agent structure, may have emergent goal-like behavior from the interaction of these facts.

I think locating goals "within the CoT" often doesn't work, a ton of work is done implicitly, especially after RL on a model using CoT. What does that mean for attempts to teach metacognition that's good according to humans?

Would you agree that the Jeffrey-Bolker picture has stronger conditions? Rather than just needing the agent to tell you their preference ordering, they need to tell you a much more structured and theory-laden set of objects.

If you're interested in austerity it might be interesting to try to weaken the Jeffrey-Bolker requirements, or strengthen the Savage ones, to zoom in on what lets you get austerity.

Also, richness is possible in the Savage picture, you just have to stretch the definitions of "state," "action," and "consequence." In terms of the functional relationship, the action is just the thing the agent gives you a preference ordering over, and the state is just the stuff that, together with action, gives you a consequence, and the consequences are any set at all. The state doesn't have to be literally the state of the world, and the actions don't have to be discrete, external actions.

I'm glad you shared this, but it seems way overhyped. Nothing wrong with fine tuning per se, but this doesn't address open problems in value learning (mostly of the sort "how do you build human trust in an AI system that has to make decisions on cases where humans themselves are inconsistent or disagree with each other?").

Not being an author in any of those articles, I can only give my own take.

I use the term "weak to strong generalization" to talk about a more specific research-area-slash-phenomenon within scalable oversight (which I define like SO-2,3,4). As a research area, it usually means studying how a stronger student AI learns what a weaker teacher is "trying" to demonstrate, usually just with slight twists on supervised learning, and when that works well, that's the phenomenon.

It is not an alignment technique to me because the phrase "alignment technique" sounds like it should be something more specific. But if you specified details about how the humans were doing demonstrations, and how the student AI was using them, that could be an alignment technique that uses the phenomenon of weak to strong generalization.

I do think the endgame for w2sg still should be to use humans as the weak teacher. You could imagine some cases where you've trained a weaker AI that you trust, and gain some benefit from using it to generate synthetic data, but that shouldn't be the only thing you're doing.

I honestly think your experiment made me more temporarily confused than an informal argument would have, but this was still pretty interesting by the end, so thanks.

I think there may be some things to re-examine about the role of self-experimentation in the rationalist community. Nootropics, behavioral interventions like impractical sleep schedules, maybe even meditation. It's very possible these reflect systematic mistakes by the rationalist community, that people should mostly warned away from.

It's tempting to think of the model after steps 1 and 2 as aligned but lacking capabilities, but that's not accurate. It's safe, but it's not conforming to a positive meaning of "alignment" that involves solving hard problems in ways that are good for humanity. Sure, it can mouth the correct words about being good, but those words aren't rigidly connected to the latent capabilities the model has. If you try to solve this by pouring tons of resources into steps 1 and 2, you probably end up with something that learns to exploit systematic human errors during step 2.

I give the probability that some authority figure would use an order-following AI to get torturous revenge on me (probably for being part of a group they dislike) is quite slim. Maybe one in a few thousand, with more extreme suffering being less likely by a few more orders of magnitude? The probablility that they have me killed for instrumental reasons, or otherwise waste the value of the future by my lights, is mich higher - ten percent-ish, depends on my distribution over who's giving the orders. But this isn't any worse to me than being killed by an AI that wants to replace me with molecular smiley faces.

Yes. Current AI policy is like people in a crowded room fighting over who gets to hold a bomb. It's more important to defuse the bomb than it is to prevent someone you dislike from holding it.

That said, we're currently not near any satisfactory solutions to corrigibility. And I do think it would be better for the world if were easier (by some combination of technical factors and societal factors) to build AI that works for the good of all humanity than to build equally-smart AI that follows the orders of a single person. So yes, we should focus research and policy effort toward making that happen, if we can.

And if we were in that world already, then I agree releasing all the technical details of an AI that follows the orders of a single person would be bad.

One way of phrasing the AI alignment task is to get AIs to “love humanity” or to have human welfare as their primary objective (sometimes called “value alignment”). One could hope to encode these via simple principles like Asimov’s three laws or Stuart Russel’s three principles, with all other rules derived from these.

I certainly agree that Asimov's three laws are not a good foundation for morality! Nor are any other simple set of rules.

So if that's how you mean "value alignment," yes let's discount it. But let me sell you on a different idea you haven't mentioned, which we might call "value learning."^[1]

Doing the right thing is complicated.^[2] Compare this to another complicated problem: telling photos of cats from photos of dogs. You cannot write down a simple set of rules to tell apart photos of cats and dogs. But even though we can't solve the problem with simple rules, we can still get a computer to do it. We show the computer a bunch of data about the environment and human classifications thereof, have it tweak a bunch of parameters to make a model of the data, and hey presto, it tells cats from dogs.

Learning the right thing to do is just like that, except for all the ways it's different that are still open problems:

• Humans are inconsistent and disagree with each other about the right thing more than they are inconsistent/disagree about dogs and cats.
• If you optimize for doing the right thing, this is a bit like searching for adversarial examples, a stress test that the dog/cat classifier didn't have to handle.
• When building an AI that learns the right thing to do, you care a lot more about trust than when you build a dog/cat classifier.

This margin is too small to contain my thoughts on all these.

There's no bright line between value learning and techniques you'd today lump under "reasonable compliance." Yes, the user experience is very different between (e.g.) an AI agent that's operating a computer for days or weeks vs. a chatbot that responds to you within seconds. But the basic principles are the same - in training a chatbot to behave well you use data to learn some model of what humans want from a chatbot, and then the AI is trained to perform well according to the modeled human preferences.

The open problems for general value learning are also open problems for training chatbots to be reasonable. How do you handle human inconsistency and disagreement? How do you build trust that the end product is actually reasonable, when that's so hard to define? Etc. But the problems have less "bite," because less can go wrong when your AI is briefly responding to a human query than when your AI is using a computer and navigating complicated real-world problems on its own.

You might hope we can just say value learning is hard, and not needed anyhow because chatbots need it less than agents do, so we don't have to worry about it. But the chatbot paradigm is only a few years old, and there is no particular reason it should be eternal. There are powerful economic (and military) pressures towards building agents that can act rapidly and remain on-task over long time scales. AI safety research needs to anticipate future problems and start work on them ahead of time, which means we need to be prepared for instilling some quite ambitious "reasonableness" into AI agents.

1. ^{^}

   For a decent introduction from 2018, see this collection.

2. ^{^}

   Citation needed.

Yeah, that's true. I expect there to be a knowing/wanting split - AI might be able to make many predictions about how a candidate action will affect many slightly-conflicting notions of "alignment", or make other long-term predictions, but that doesn't mean it's using those predictions to pick actions. Many people want to build AI that picks actions based on short-term considerations related to the task assigned to it.

I think this framing probably undersells the diversity within each category, and the extent of human agency or mere noise that can jump you from one category to another.

Probably the biggest dimension of diversity is how much the AI is internally modeling the whole problem and acting based on that model, versus how much it's acting in feedback loops with humans. In the good category you describe it as acting more in feedback loops with humans, while in the bad category you describe it more as internally modeling the whole problem, but I think all quadrants are quite possible.

In the good case with the AI modeling the whole problem, this might look like us starting out with enough of a solution to alignment that the vibe is less "we need to hurry and use the AI to do our work for us" and more "we're executing a shared human-AI gameplan for learning human values that are good by human standards."

In the bad case with the AI acting through feedback loops with humans, this might look like the AI never internally representing deceiving us, humans just keep using it in slightly wrong ways that end up making the future bad. (Perhaps by giving control to fallible authority figures, perhaps by presenting humans with superstimuli that cause value drift we think is bad from our standpoint outside the thought experiment, perhaps by defining "what humans want" in a way that captures many of the 'advantages' of deception for maximizing reward without triggering our interpretability tools that are looking for deception.)

I think particularly when the AI is acting in feedback loops with humans, we could get bounced between categories by things like human defectors trying to seize control of transformative AI, human society cooperating and empowering people who aren't defectors, new discoveries made by humans about AI capabilities or alignment, economic shocks, international diplomacy, and maybe even individual coding decisions.

First, I agree with Dmitry.

But it does seem like maybe you could recover a notion of information bottleneck even with out the Bayesian NN model. If you quantize real numbers to N-bit floating point numbers, there's a very real quantity which is "how many more bits do you need to exactly reconstruct X, given Z?" My suspicion is that for a fixed network, this quantity grows linearly with N (and if it's zero at 'actual infinity' for some network despite being nonzero in the limit, maybe we should ignore actual infinity).

But this isn't all that useful, it would be nicer to have an information that converges. The divergence seems a bit silly, too, because it seems silly to treat the millionth digit as as important as the first.

So suppose you don't want to perfectly reconstruct X. Instead, maybe you could say the distribution of X is made of some fixed number of bins or summands, and you want to figure out which one based on Z. Then you get a converging amount of information, and you correctly treat small numbers as less important, but you've had to introduce this somewhat arbitrary set of bins. shrug

A process or machine prepares either |0> or |1> at random, each with 50% probability. Another machine prepares either |+> or |-> based on a coin flick, where |+> = (|0> + |1>)/root2, and  |+> = (|0> - |1>)/root2. In your ontology these are actually different machines that produce different states.

I wonder if this can be resolved by treating the randomness of the machines quantum mechanically, rather than having this semi-classical picture where you start with some randomness handed down from God. Suppose these machines use quantum mechanics to do the randomization in the simplest possible way - they have a hidden particle in state |left>+|right>  (pretend I normalize), they mechanically measure it (which from the outside will look like getting entangled with it) and if it's on the left they emit their first option (|0> or |+> depending on the machine) and vice versa.

So one system, seen from the outside, goes into the state |L,0>+|R,1>, the other one into the state |L,0>+|R,0>+|L,1>-|R,1>. These have different density matrices. The way you get down to identical density matrices is to say you can't get the hidden information (it's been shot into outer space or something). And then when you assume that and trace out the hidden particle, you get the same representation no matter your philosophical opinion on whether to think of the un-traced state as a bare state or as a density matrix. If on the other hand you had some chance of eventually finding the hidden particle, you'd apply common sense and keep the states or density matrices different.

Anyhow, yeah, broadly agree. Like I said, there's a practical use for saying what's "real" when you want to predict future physics. But you don't always have to be doing that.

people who study very "fundamental" quantum phenomena increasingly use a picture with a thermal bath

Maybe talking about the construction of pointer states? That linked paper does it just as you might prefer, putting the Boltzmann distribution into a density matrix. But of course you could rephrase it as a probability distribution over states and the math goes through the same, you've just shifted the vibe from "the Boltzmann distribution is in the territory" to "the Boltzmann distribution is in the map."

Still, as soon as you introduce the notion of measurement, you cannot get away from thermodynamics. Measurement is an inherently information-destroying operation, and iiuc can only be put "into theory" (rather than being an arbitrary add-on that professors tell you about) using the thermodynamic picture with nonunitary operators on density matrices.

Sure, at some level of description it's useful to say that measurement is irreversible, just like at some level of description it's useful to say entropy always increases. Just like with entropy, it can be derived from boundary conditions + reversible dynamics + coarse-graining. Treating measurements as reversible probably has more applications than treating entropy as reversible, somewhere in quantum optics / quantum computing.

Some combination of:

• Interpretability
  • Just check if the AI is planning to do bad stuff, by learning how to inspect its internal representations.
• Regularization
  • Evolution got humans who like Doritos more than health food, but evolution didn't have gradient descent. Use regularization during training to penalize hidden reasoning.
• Shard / developmental prediction
  • Model-free RL will predictably use simple heuristics for the reward signal. If we can predict and maybe control how this happens, this gives us at least a tamer version of inner misalignment.
• Self-modeling
  • Make it so that the AI has an accurate model of whether it's going to do bad stuff. Then use this to get the AI not to do it.
• Control
  • If inner misalignment is a problem when you use AI's off-distribution and give them unchecked power, then don't do that.

Personally, I think the most impactful will be Regularization, then Interpretability.

The real chad move is to put "TL;DR: See above^" for every section.

When you say there's "no such thing as a state," or "we live in a density matrix," these are statements about ontology: what exists, what's real, etc.

Density matrices use the extra representational power they have over states to encode a probability distribution over states. If we regard the probabilistic nature of measurements as something to be explained, putting the probability distribution directly into the thing we live in is what I mean by "explain with ontology."

Epistemology is about how we know stuff. If we start with a world that does not inherently have a probability distribution attached to it, but obtain a probability distribution from arguments about how we know stuff, that's "explain with epistemology."

In quantum mechanics, this would look like talking about anthropics, or what properties we want a measure to satisfy, or solomonoff induction and coding theory.

What good is it to say things are real or not? One useful application is predicting the character of physical law. If something is real, then we might expect it to interact with other things. I do not expect the probability distribution of a mixed state to interact with other things.

Treating the density matrix as fundamental is bad because you shouldn't explain with ontology that which you can explain with epistemology.

Be sad.

For topological debate that's about two agents picking settings for simulation/computation, where those settings have a partial order that lets you take the "strictest" combination, a big class of fatal flaw would be if you don't actually have the partial order you think you have within the practical range of the settings - i.e. if some settings you thought were more accurate/strict are actually systematically less accurate.

In the 1D plane example, this would be if some specific length scales (e.g. exact powers of 1000) cause simulation error, but as long as they're rare, this is pretty easy to defend against.

In the fine-grained plane example, though, there's a lot more room for fine-grained patterns in which parts of the plane get modeled at which length scale to start having nonlinear effects. If the agents are not allowed to bid "maximum resolution across the entire plane," and instead are forced to allocate resources cleverly, then maybe you have a problem. But hopefully the truth is still advantaged, because the false player has to rely on fairly specific correlations, and the true player can maybe bid a bunch of noise that disrupts almost all of them.

(This makes possible a somewhat funny scene, where the operator expected the true player's bid to look "normal," and then goes to check the bids and both look like alien noise patterns.)

An egregious case would be where it's harder to disrupt patterns injected during bids - e.g. if the players' bids are 'sparse' / have finite support and might not overlap. Then the notion of the true player just needing to disrupt the false player seems a lot more unlikely, and both players might get pushed into playing very similar strategies that take every advantage of the dynamics of the simulator in order to control the answer in an unintended way.

I guess for a lot of "tough real world questions," the difficulty of making a super-accurate simulator (one you even hope converges to the right answer) torpedoes the attempt before we have to start worrying about this kind of 'fatal flaw'. But anything involving biology, human judgment, or too much computer code seems tough. "Does this gene therapy work?" might be something you could at least imagine a simulator for that still seems like it gives lots of opportunity for the false player.

Fun post, even though I don't expect debate of either form to see much use (because resolving tough real world questions offers too many chances for the equivalent of the plane simulation to have fatal flaws).

With bioweapons evals at least the profit motive of AI companies is aligned with the common interest here; a big benefit of your work comes from when companies use it to improve their product. I'm not at all confused about why people would think this is useful safety work, even if I haven't personally hashed out the cost/benefit to any degree of confidence.

I'm mostly confused about ML / SWE / research benchmarks.

The mathematical structure in common is called a "measure."

I agree that there's something mysterious-feeling about probability in QM, though I mostly think that feeling is an illusion. There's a (among physicists) famous fact that the only way to put a 'measure' on a wavefunction that has nice properties (e.g. conservation over time) is to take the amplitude squared. So there's an argument: probability is a measure, and the only measure that makes sense is the amplitude-squared measure, therefore if probability is anything it's the amplitude squared. And it is! Feels mysterious.

But after getting more used to anthropics and information theory, you start to accumulate more arguments for the same thing that take it from a different angle, and it stops feeling so mysterious.

Could someone who thinks capabilities benchmarks are safety work explain the basic idea to me?

It's not all that valuable for my personal work to know how good models are at ML tasks. Is it supposed to be valuable to legislators writing regulation? To SWAT teams calculating when to bust down the datacenter door and turn the power off? I'm not clear.

But it sure seems valuable to someone building an AI to do ML research, to have a benchmark that will tell you where you can improve.

But clearly other people think differently than me.

One big reason I might expect an AI to do a bad job at alignment research is if it doesn't do a good job (according to humans) of resolving cases where humans are inconsistent or disagree. How do you detect this in string theory research? Part of the reason we know so much about physics is humans aren't that inconsistent about it and don't disagree that much. And if you go to sub-topics where humans do disagree, how do you judge its performance (because 'be very convincing to your operators' is an objective with a different kind of danger).

Another potential red flag is if the AI gives humans what they ask for even when that's 'dumb' according to some sophisticated understanding of human values. This could definitely show up in string theory research (note when some ideas suggest non-string-theory paradigms might be better, and push back on the humans if the humans try to ignore this), it's just intellectually difficult (maybe easier in loop quantum gravity research heyo gottem) and not as salient without the context of alignment and human values.

Thanks for the great reply :) I think we do disagree after all.

humans are definitionally the source of information about human values, even if it may be challenging to elicit this information from humans

Except about that - here we agree.

Now, what this human input looks like could (and probably should) go beyond introspection and preference judgments, which, as you point out, can be unreliable. It could instead involve expert judgment from humans with diverse cultural backgrounds, deliberation and/or negotiation, incentives to encourage deep, reflective thinking rather than snap judgments or falling back on heuristics. It could also involve AI assistance to help counter human biases, find common ground, and consider the logical consequences of communicated values. 

This might be summarized as "If humans are inaccurate, let's strive to make them more accurate." 

I think this, as a research priority or plan A, is doomed by a confluence of practical facts (humans aren't actually that consistent, even in what we'd consider a neutral setting) and philosophical problems (What if I think the snap judgments and heuristics are important parts of being human? And, how do you square a univariate notion of 'accuracy' with the sensitivity of human conclusions to semi-arbitrary changes to e.g. their reading lists, or the framings of arguments presented to them?).

Instead, I think our strategy should be "If humans are inconsistent and disagree, let's strive to learn a notion of human values that's robust to our inconsistency and disagreement."

We contend that even as AI gets really smart, humans ultimately need to be in the loop to determine whether or not a constitution is aligned and reasonable.

A committee of humans reviewing an AI's proposal is, ultimately, a physical system that can be predicted. If you have an AI that's good at predicting physical systems, then before it makes an important decision it can just predict this Committee(time, proposal) system and treat the predicted output as feedback on its proposal. If the prediction is accurate, then actual humans meeting in committee is unnecessary.

(And indeed, putting human control of the AI in the physical world actually exposes it to more manipulation than if the control is safely ensconced in the logical structure of the AI's decision-making.)

I basically think your sixth to last (or so) bulllet point is key - an AI that takes over is likely to be using a lot more RL on real world problems, i.e. drawn from a different distribution than present-day AI. This will be worse for us than conditioning on a present-day AI taking over.

Cool stuff!

I'm a little confused what it means to mean-ablate each node...

Oh, wait. ctrl-f shows me the Non-Templatic data appendix. I see, so you're tracking the average of each feature, at each point in the template. So you can learn a different mask at each token in the template and also learn a different mean (and hopefully your data distribution is balanced / high-entropy). I'm curious - what happens to your performance with zero-ablation (or global mean ablation, maybe)?

Excited to see what you come up with for non-templatic tasks. Presumably on datasets of similar questions, similar attention-control patterns will be used, and maybe it would just work to (somehow) find which tokens are getting similar attention, and assign them the same mask.

It would also be interesting to see how this handles more MLP-heavy tasks like knowledge questions. maybe someone clever can find a template for questions about the elements, or the bibliographies of various authors, etc.

No, you're right that aristocracy is more complicated. There were lots of pressures that shaped the form of it. Certainly more than how good of managers aristocrats made!

An invalid syllogism: "The rules of aristocracy were shaped by forces. Avoiding poor management is a force. Therefore, the rules of aristocracy will be all about avoiding poor management."

Aristocrats were also selected for how well they could extract rents from those below, and how well they could resist rent-extraction from above, both alone and collectively. Nor was the top-down pressure all about making aristocrats into productive managers - rent-extraction has been mentioned, and also weakening the aristocracy to secure central power, allowing advancement via marriage and alliance, various human status games, and the need for a legislative arm of government.

I don't want to hear the One Pressure That Explains Everything (but only qualitatively, and if you squint). I'll want to hear when they have the dozen pressures that make up a model that can be quantitatively fit to past data by tuning some parameters, including good retrodictive accuracy over a held-out time period.

I think if you want to go fast, and you can eat the rest of the solar system, you can probably make a huge swarm of fusion reactors to help blow matter off the sun. Let's say you can build 10^11-watt reactors that work in space. Then you need about 10^15 of them to match the sun. If each is 10^6 kg, this is about 10^-4 of Mercury's mass.

I was expecting (Methods start 16:00)

When you find a fence in a field, someone once built that fence on purpose and had a reason for it. So it's good sense to ask after that reason, and guess ahead of time that it might be worth a fence, to the owner of the field.

When you find a rock in a field, probably nobody put that rock there on purpose. And so it's silly to go "What is the reason this rock was put here? I might now know now, but I can guess ahead of time it might be worth it to me!"

I agree it's a good point that you don't need the complexity of the whole world to test ideas. With a fairly small in terms of number of states, you can encode interesting things in a long sequence of states so long as the generating process is sufficiently interesting. And adding more states is itself no virtue if it doesn't help you understand what you're trying to test for.

Some out-of-order thoughts:

• Testing for 'big' values, e.g. achievement, might require complex environments and evaluations. Not necessarily large state spaces, but the complexity of differentiating between subtle shades of value (which seems like a useful capability to be sure we're getting) has to go somewhere.
• Using more complicated environments that are human-legible might better leverage human feedback and/or make sense to human observers - maybe you could encode achievement in the actions of a square in a gridworld, but maybe humans would end up making a lot of mistakes when trying to judge the outcome. If you want to gather data from humans, to reflect a way that humans are complicated that you want to see if an AI can learn, a rich environment seems useful. On the other hand, if you just want to test general learning power, you could have a square in a gridworld have random complex decision procedures and see if they can be learned.

• There's a divide between contexts where humans are basically right, and so we just want an AI to do a good job of learning what we're already doing, and contexts where humans are inconsistent, or disagree with each other, where we want an AI to carefully resolve these inconsistencies/disagreements in a way that humans endorse (except also sometimes we're inconsistent or disagree about our standards for resolving inconsistencies and disagreements!).

  Building small benchmarks for the first kind of problem seems kind of trivial in the fully-observed setting where the AI can't wirehead. Even if you try to emulate the partial observability of the real world, and include the AI being able to eventually control the reward signal as a natural part of the world, it seems like seizing control of the reward signal is the crux rather than the content of what values are being demonstrated inside the gridworld (I guess it's useful to check if the content matters, I just don't expect it to), and a useful benchmark might be focused on how seizing control of the reward signal (or not doing so) scales to the real world.

  Building small benchmarks for the latter kind of problem seems important. The main difficulty is more philosophical than practical - we don't know what standard to hold the benchmarks to. But supposing we had some standard in mind, I would still worry that a small benchmark would be more easily gamed, and more likely to miss some of the ways humans are inconsistent or disagree. I would also expect benchmarks of this sort, whatever the size, to be a worse fit for normal RL algorithms, and run into issues where different learning algorithms might request different sorts of interaction with the environment (although this could be solved either by using real human feedback in a contrived situation, or by having simulated inhabitants of the environment who are very good at giving diverse feedback).

Honestly I think this is still too optimistic. Humans are not consistent economic actors - we can be persuaded of new things even if those things are subjective, will sometimes take deals we might in other circumstances call unfavorable, and on an absolute scale aren't all that bright. Owning capital does not fix this, and so an AI that's good at interacting with humans will be able to get more from us than you might expect just looking at the current economy.

As Sean Carroll likes to say, though, the reason we've made so much progress in physics is that it's way easier than the other sciences :)

Voluntary interaction has been great for humans. But it hasn't been great for orangutans, who don't do a very good job of participating in society.

Even if you somehow ensure transparency and cooperation among superintelligent AIs and humans, it seems overwhelmingly likely that humans will take the place of the orangutan, marginalized and taken from in every way possible within the limits of what is, in the end, not a very strict system. It is allowed, as Eliezer would say.

Orangutans don't contribute to human society even though they're specialized in things humans aren't. The best chess player in the world isn't a human-AI symbiote, for the same reason it's not an orangutan-human-AI symbiote.

Human trades with superintelligent AI do not have to be Pareto improvements (in the common-sense way), because humans make systematic mistakes (according to the common-sense standard). If you actually knew how to detect what trades would be good for humans - how to systematize that common sense, and necessarily also how to improve it since it is itself inconsistent and systematically mistaken - this would be solving the key parts of the value alignment problem that one might have hoped to sidestep by relying on voluntarism instead.

I'm not excited by gridworlds, because they tend to to skip straight to representing the high-level objects we're supposed to value, without bothering to represent all the low-level structure that actually lets us learn and generalize values in the real world.

Do you have plans for how to deal with this, or plans to think about richer environments?

Because 'alignment' is used in several different ways, I feel like these days one either needs to asterisk in a definition (e.g. "By 'alignment,' I mean the AI faithfully carrying out instructions without killing everyone."), or just use a more specific phrase.

I agree that instruction-following is not all you need. Many of these problems are solved by better value-learning.