AI used to be a science. In the old days (back when AI didn't work very well), people were attempting to develop a working theory of cognition.

Those scientists didn’t succeed, and those days are behind us. For most people working in AI today and dividing up their work hours between tasks, gone is the ambition to understand minds. People working on mechanistic interpretability (and others attempting to build an empirical understanding of modern AIs) are laying an important foundation stone that could play a role in a future science of artificial minds, but on the whole, modern AI engineering is simply about constructing enormous networks of neurons and training them on enormous amounts of data, not about comprehending minds.

The bitter lesson has been taken to heart, by those at the forefront of the field; and although this lesson doesn't teach us that there's nothing to learn about how AI minds solve problems internally, it suggests that the fastest path to producing more powerful systems is likely to continue to be one that doesn’t shed much light on how those systems work.

Absent some sort of “science of artificial minds”, however, humanity’s prospects for aligning smarter-than-human AI seem to me to be quite dim.

Viewing Earth’s current situation through that lens, I see three major hurdles:

  1. Most research that helps one point AIs, probably also helps one make more capable AIs. A “science of AI” would probably increase the power of AI far sooner than it allows us to solve alignment.
  2. In a world without a mature science of AI, building a bureaucracy that reliably distinguishes real solutions from fake ones is prohibitively difficult.
  3. Fundamentally, for at least some aspects of system design, we’ll need to rely on a theory of cognition working on the first high-stakes real-world attempt.

I’ll go into more detail on these three points below. First, though, some background:

 

Background

By the time AIs are powerful enough to endanger the world at large, I expect AIs to do something akin to “caring about outcomes”, at least from a behaviorist perspective (making no claim about whether it internally implements that behavior in a humanly recognizable manner).

Roughly, this is because people are trying to make AIs that can steer the future into narrow bands (like “there’s a cancer cure printed on this piece of paper”) over long time-horizons, and caring about outcomes (in the behaviorist sense) is the flip side of the same coin as steering the future into narrow bands, at least when the world is sufficiently large and full of curveballs.

I expect the outcomes that the AI “cares about” to, by default, not include anything good (like fun, love, art, beauty, or the light of consciousness) — nothing good by present-day human standards, and nothing good by broad cosmopolitan standards either. Roughly speaking, this is because when you grow minds, they don’t care about what you ask them to care about and they don’t care about what you train them to care about; instead, I expect them to care about a bunch of correlates of the training signal in weird and specific ways.

(Similar to how the human genome was naturally selected for inclusive genetic fitness, but the resultant humans didn’t end up with a preference for “whatever food they model as useful for inclusive genetic fitness”. Instead, humans wound up internalizing a huge and complex set of preferences for "tasty" foods, laden with complications like “ice cream is good when it’s frozen but not when it’s melted”.)

Separately, I think that most complicated processes work for reasons that are fascinating, complex, and kinda horrifying when you look at them closely.

It’s easy to think that a bureaucratic process is competent until you look at the gears and see the specific ongoing office dramas and politicking between all the vice-presidents or whatever. It’s easy to think that a codebase is running smoothly until you read the code and start to understand all the decades-old hacks and coincidences that make it run. It’s easy to think that biology is a beautiful feat of engineering until you look closely and find that the eyeballs are installed backwards or whatever.

And there’s an art to noticing that you would probably be astounded and horrified by the details of a complicated system if you knew them, and then being astounded and horrified already in advance before seeing those details.[1]

 

1. Alignment and capabilities are likely intertwined

I expect that if we knew in detail how LLMs are calculating their outputs, we’d be horrified (and fascinated, etc.).

I expect that we’d see all sorts of coincidences and hacks that make the thing run, and we’d be able to see in much more detail how, when we ask the system to achieve some target, it’s not doing anything close to “caring about that target” in a manner that would work out well for us, if we could scale up the system’s optimization power to the point where it could achieve great technological or scientific feats (like designing Drexlerian nanofactories or what-have-you).

Gaining this sort of visibility into how the AIs work is, I think, one of the main goals of interpretability research.

And understanding how these AIs work and how they don’t — understanding, for example, when and why they shouldn’t yet be scaled or otherwise pushed to superintelligence — is an important step on the road to figuring out how to make other AIs that could be scaled or otherwise pushed to superintelligence without thereby causing a bleak and desolate future.

But that same understanding is — I predict — going to reveal an incredible mess. And the same sort of reasoning that goes into untangling that mess into an AI that we can aim, also serves to untangle that mess to make the AI more capable. A tangled mess will presumably be inefficient and error-prone and occasionally self-defeating; once it’s disentangled, it won’t just be tidier, but will also come to accurate conclusions and notice opportunities faster and more reliably.[2]

Indeed, my guess is that it’s even easier to see all sorts of things that the AI is doing that are dumb, all sorts of ways that the architecture is tripping itself up, and so on.

Which is to say: the same route that gives you a chance of aligning this AI (properly, not the “it no longer says bad words” superficial-property that labs are trying to pass off as “alignment” these days) also likely gives you lots more AI capabilities.

(Indeed, my guess is that the first big capabilities gains come sooner than the first big alignment gains.)

I think this is true of most potentially-useful alignment research: to figure out how to aim the AI, you need to understand it better; in the process of understanding it better you see how to make it more capable.

If true, this suggests that alignment will always be in catch-up mode: whenever people try to figure out how to align their AI better, someone nearby will be able to run off with a few new capability insights, until the AI is pushed over the brink.

So a first key challenge for AI alignment is a challenge of ordering: how do we as a civilization figure out how to aim AI before we’ve generated unaimed superintelligences plowing off in random directions? I no longer think “just sort out the alignment work before the capabilities lands” is a feasible option (unless, by some feat of brilliance, this civilization pulls off some uncharacteristically impressive theoretical triumphs).

Interpretability? Will likely reveal ways your architecture is bad before it reveals ways your AI is misdirected.

Recruiting your AIs to help with alignment research? They’ll be able to help with capabilities long before that (to say nothing of whether they would help you with alignment by the time they could, any more than humans would willingly engage in eugenics for the purpose of redirecting humanity away from Fun and exclusively towards inclusive genetic fitness).

And so on.

This is (in a sense) a weakened form of my answer to those who say, “AI alignment will be much easier to solve once we have a bona fide AGI on our hands.” It sure will! But it will also be much, much easier to destroy the world, when we have a bona fide AGI on our hands. To survive, we’re going to need to either sidestep this whole alignment problem entirely (and take other routes to a wonderful future instead, as I may discuss more later), or we’re going to need some way to do a bunch of alignment research even as that research makes it radically easier and radically cheaper to destroy everything of value.

Except even that is harder than many seem to realize, for the following reason.

 

2. Distinguishing real solutions from fake ones is hard

Already, labs are diluting the word “alignment” by using the word for superficial results like “the AI doesn’t say bad words”. Even people who apparently understand many of the core arguments have apparently gotten the impression that GPT-4’s ability to answer moral quandaries is somehow especially relevant to the alignment problem, and an important positive sign.

(The ability to answer moral questions convincingly mostly demonstrates that the AI can predict how humans would answer or what humans want to hear, without revealing much about what the AI actually pursues, or would pursue upon reflection, etc.)

Meanwhile, we have little idea of what passes for “motivations” inside of an LLM, or what effect pretraining on next-token prediction and fine-tuning with RLHF really has on the internals. This sort of precise scientific understanding of the internals — the sort that lets one predict weird cognitive bugs in advance — is currently mostly absent in the field. (Though not entirely absent, thanks to the hard work of many researchers.)

Now imagine that Earth wakes up to the fact that the labs aren’t going to all decide to stop and take things slowly and cautiously at the appropriate time.[3] And imagine that Earth uses some great feat of civilizational coordination to halt the world’s capabilities progress, or to otherwise handle the issue that we somehow need room to figure out how these things work well enough to align them. And imagine we achieve this coordination feat without using that same alignment knowledge to end the world (as we could). There’s then the question of who gets to proceed, under what circumstances.

Suppose further that everyone agreed that the task at hand was to fully and deeply understand the AI systems we’ve managed to develop so far, and understand how they work, to the point where people could reverse out the pertinent algorithms and data-structures and what-not. As demonstrated by great feats like building, by-hand, small programs that do parts of what AI can do with training (and that nobody previously knew how to code by-hand), or by identifying weird exploits and edge-cases in advance rather than via empirical trial-and-error. Until multiple different teams, each with those demonstrated abilities, had competing models of how AIs’ minds were going to work when scaled further.

In such a world, it would be a difficult but plausibly-solvable problem, for bureaucrats to listen to the consensus of the scientists, and figure out which theories were most promising, and figure out who needs to be allotted what license to increase capabilities (on the basis of this or that theory that predicts this would be non-catastrophic), so as to put their theory to the test and develop it further.

I’m not thrilled about the idea of trusting an Earthly bureaucratic process with distinguishing between partially-developed scientific theories in that way, but it’s the sort of thing that a civilization can perhaps survive.

But that doesn’t look to me like how things are poised to go down.

It looks to me like we’re on track for some people to be saying “look how rarely my AI says bad words”, while someone else is saying “our evals are saying that it can’t deceive humans yet”, while someone else is saying “our AI is acting very submissive, and there’s no reason to expect AIs to become non-submissive, that’s just anthropomorphizing”, and someone else is saying “we’ll just direct a bunch of our AIs to help us solve alignment, while arranging them in a big bureaucracy”, and someone else is saying “we’ve set up the game-theoretic incentives such that if any AI starts betraying us, some other AI will alert us first”, and this is a different sort of situation.

And not one that looks particularly survivable, to me.

And if you ask bureaucrats to distinguish which teams should be allowed to move forward (and how far) in that kind of circus, full of claims, promises, and hunches and poor in theory, then I expect that they basically just can’t.

In part because the survivable answers (such as “we have no idea what’s going on in there, and will need way more of an idea what’s going on in there, and that understanding needs to somehow develop in a context where we can do the job right rather than simply unlocking the door to destruction”) aren’t really in the pool. And in part because all the people who really want to be racing ahead have money and power and status. And in part because it’s socially hard to believe, as a regulator, that you should keep telling everyone “no”, or that almost everything on offer is radically insufficient, when you yourself don’t concretely know what insights and theoretical understanding we’re missing.

Maybe if we can make AI a science again, then we’ll start to get into the regime where, if humanity can regulate capabilities advancements in time, then all the regulators and researchers understand that you shall only ask for a license to increase the capabilities of your system when you have a full detailed understanding of the system and a solid justification for why you need the capabilities advance and why it’s not going to be catastrophic. At which point maybe a scientific field can start coming to some sort of consensus about those theories, and regulators can start being sensitive to that consensus.

But unless you can get over that grand hump, it looks to me like one of the key bottlenecks here is bureaucratic legibility of plausible solutions. Where my basic guess is that regulators won’t be able to distinguish real solutions from false ones, in anything resembling the current environment.

Together with the above point ("alignment and capabilities are likely intertwined"), I think this means that our rallying cry should be less “pause to give us more time on alignment research” and more “stop entirely, and find some way to circumvent these woods entirely; we’re not equipped to navigate them”.

(With a backup rallying cry of “make AI a science again”, though again, that only works if you have some way of preventing the science-of-mind from leading to catastrophe before we figure out how to build AIs that care about good stuff rather than bleak and desolate stuff.)

 

3. Most theories don’t work on the first real try

It seems worth noting that even if you manage to surmount the above two problems, you have a third problem on your hands, which is that when it finally comes time, not to increase your system’s capabilities a bit more (and test your theory of cognition further), but rather to actually do something of note with your AI, then you must, in some sense, be trusting a novel and untested scientific theory (and your engineering efforts grounded in that theory) to work perfectly on the first critical try.

In particular, once your AI is capable of autonomous scientific/technological development, and operating in a domain where it could use that ability to gain a decisive strategic advantage over the rest of the planet, it’s operating in a fundamentally different cognitive regime than it ever was in training.

By analogy, if you’re considering making Bob the dictator of your country, you might first make him mock-dictator of your town, and watch to make sure he doesn’t abuse his power. But, try as you might, that’s still not fundamentally all that reliable a test of whether he’ll actually abuse power once he really has it. Once he really can give orders to the army, once he really can take advantage of the country’s populace, will he do so? The option of “abuse my mock power while being observed by my townspeople (who can still beat me up or deny me the job)” is actually very cognitively different than the option “order the army to intimidate the parliament and ‘oversee’ the next election”.

Now, given a sufficiently well-developed theory of cognition, you can try to read the AI’s mind and predict what cognitive state it would enter if it did really think it had those options. And you can set up simulations (and try to spoof its internal sensations and so on) in a way that your theory of cognition predicts is very similar to the cognitive state it would enter once it really had the option to betray you.

But the link between these states that you induce and observe in the lab, and the actual state where the AI actually has the option to betray you, depends fundamentally on your fresh new theory of cognition.

Actually running the AI until it really has the opportunity to betray you is an empirical test of those theories in an environment that differs fundamentally from the lab setting.

And many a scientist (and programmer) knows that their theories of how a complicated system is going to work in a fundamentally new operating environment often don’t go super well on the first try.

As a concrete analogy to potentially drive this point home: Newtonian mechanics made all sorts of shockingly-good empirical predictions. It was a simple concise mathematical theory with huge explanatory power that blew every previous theory out of the water. And if you were using it to send payloads to very distant planets at relativistic speeds, you’d still be screwed, because Newtonian mechanics does not account for relativistic effects.

(And the only warnings you’d get would be little hints about light seeming to move at the same speed in all directions at all times of year, and light bending around the sun during eclipses, and the perihelion of Mercury being a little off from what Newtonian mechanics predicted. Small anomalies, weighed against an enormous body of predictive success in a thousand empirical domains; and yet Nature doesn’t care, and the theory still falls apart when we move to energies and scales far outside what we’d previously been able to observe.)

Getting scientific theories to work on the first critical try is hard. (Which is one reason to aim for minimal pivotal tasks — getting a satellite into orbit should work fine on Newtonian mechanics, even if sending payloads long distances at relativistic speeds does not.)

Worrying about this issue is something of a luxury, at this point, because it’s not like we’re anywhere close to scientific theories of cognition that accurately predict all the lab data. But it’s the next hurdle on the queue, if we somehow manage to coordinate to try to build up those scientific theories, in a way where success is plausibly bureaucratically-legible.


Maybe later I’ll write more about what I think the strategy implications of these points are. In short, I basically recommend that Earth pursue other routes to the glorious transhumanist future, such as uploading. (Which is also fraught with peril, but I expect that those perils are more surmountable; I hope to write more about this later.)


 

  1. ^

    Albeit slightly less, since there’s nonzero prior probability on this unknown system turning out to be simple, elegant, and well-designed.

  2. ^

    An exception to this guess happens if the AI is at the point where it’s correcting its own flaws and improving its own architecture, in which case, in principle, you might not see much room for capabilities improvements if you took a snapshot and comprehended its inner workings, despite still being able to see that the ends it pursues are not the ones you wanted. But in that scenario, you’re already about to die to the self-improving AI, or so I predict.

  3. ^

    Not least because there are no sufficiently clear signs that it’s time to stop — we blew right past “an AI claims it is sentient”, for example. And I’m not saying that it was a mistake to doubt AI systems’ first claims to be sentient — I doubt that Bing had the kind of personhood that’s morally important (though I am by no means confident!). I’m saying that the thresholds that are clear in science fiction stories turn out to be messy in practice and so everyone just keeps plowing on ahead.

New Comment
25 comments, sorted by Click to highlight new comments since: Today at 11:45 AM

As Shankar Sivarajan points out in a different comment, the idea that AI became less scientific when we started having actual machine intelligence to study, as opposed to before that when the 'rightness' of a theory was mostly based on the status of whoever advanced it, is pretty weird. The specific way in which it's weird seems encapsulated by this statement:

on the whole, modern AI engineering is simply about constructing enormous networks of neurons and training them on enormous amounts of data, not about comprehending minds.

In that there is an unstated assumption that these are unrelated activities. That deep learning systems are a kind of artifact produced by a few undifferentiated commodity inputs, one of which is called 'parameters', one called 'compute', and one called 'data', and that the details of these commodities aren't important. Or that the details aren't important to the people building the systems.

I've seen a (very revisionist) description of the Wright Brothers research as analogous to solving the control problem, because other airplane builders would put in an engine and crash before they'd developed reliable steering. Therefore, the analogy says, we should develop reliable steering before we 'accelerate airplane capabilities'. When I heard this I found it pretty funny, because the actual thing the Wright Brothers did was a glider capability grind. They carefully followed the received aerodynamic wisdom that had been written down, and when the brothers realized a lot of it was bunk they started building their own database to get it right:

During the winter of 1901, the brothers began to question the aerodynamic data on which they were basing their designs. They decided to start over and develop their own data base with which they would design their aircraft. They built a wind tunnel and began to test their own models. They developed an ingenious balance system to compare the performance of different models. They tested over two hundred different wings and airfoil sections in different combinations to improve the performance of their gliders The data they obtained more correctly described the flight characteristics which they observed with their gliders. By early 1902 the Wrights had developed the most accurate and complete set of aerodynamic data in the world.

In 1902, they returned to Kitty Hawk with a new aircraft based on their new data. This aircraft had roughly the same wing area as the 1901, but it had a longer wing span and a shorter chord which reduced the drag. It also sported a new movable rudder at the rear which was installed to overcome the adverse yaw problem. The movable rudder was coordinated with the wing warping to keep the nose of the aircraft pointed into the curved flight path. With this new aircraft, the brothers completed flights of over 650 feet and stayed in the air for nearly 30 seconds. This machine was the first aircraft in the world that had active controls for all three axis; roll, pitch and yaw. By the end of 1902, the brothers had completed over a thousand glides with this aircraft and were the most experienced pilots in the world. They owned all of the records for gliding. All that remained for the first successful airplane was the development of the propulsion system.

In fact while trying to find an example of the revisionist history, I found a historical aviation expert describe the Wright Brothers as having 'quickly cracked the control problem' once their glider was capable enough to let it be solved. Ironically enough I think this story, which brings to mind the possibility of 'airplane control researchers' insisting that no work be done on 'airplane capabilities' until we have a solution to the steering problem, is nearly the opposite of what the revisionist author intended and nearly spot on to the actual situation.

We can also imagine a contemporary expert on theoretical aviation (who in fact existed before real airplanes) saying something like "what the Wright Brothers are doing may be interesting, but it has very little to do with comprehending aviation [because the theory behind their research has not yet been made legible to me personally]. This methodology of testing the performance of individual airplane parts, and then extrapolating the performance of a airplane with an engine using a mere glider is kite flying, it has almost nothing to do with the design of real airplanes and humanity will learn little about them from these toys". However what would be genuinely surprising is if they simultaneously made the claim that the Wright Brothers gliders have nothing to do with comprehending aviation but also that we need to immediately regulate the heck out of them before they're used as bombers in a hypothetical future war, that we need to be thinking carefully about all the aviation risk these gliders are producing at the same time they can be assured to not result in any deep understanding of aviation. If we observed this situation from the outside, as historical observers, we would conclude that the authors of such a statement are engaging in deranged reasoning, likely based on some mixture of cope and envy.

Since we're contemporaries I have access to more context than most historical observers and know better. I think the crux is an epistemological question that goes something like: "How much can we trust complex systems that can't be statically analyzed in a reductionistic way?" The answer you give in this post is "way less than what's necessary to trust a superintelligence". Before we get into any object level about whether that's right or not, it should be noted that this same answer would apply to actual biological intelligence enhancement and uploading in actual practice. There is no way you would be comfortable with 300+ IQ humans walking around with normal status drives and animal instincts if you're shivering cold at the idea of machines smarter than people. This claim you keep making, that you're merely a temporarily embarrassed transhumanist who happens to have been disappointed on this one technological branch, is not true and if you actually want to be honest with yourself and others you should stop making it. What would be really, genuinely wild, is if that skeptical-doomer aviation expert calling for immediate hard regulation on planes to prevent the collapse of civilization (which is a thing some intellectuals actually believed bombers would cause) kept tepidly insisting that they still believe in a glorious aviation enabled future. You are no longer a transhumanist in any meaningful sense, and you should at least acknowledge that to make sure you're weighing the full consequences of your answer to the complex system reduction question. Not because I think it has any bearing on the correctness of your answer, but because it does have a lot to do with how carefully you should be thinking about it.

So how about that crux, anyway? Is there any reason to hope we can sufficiently trust complex systems whose mechanistic details we can't fully verify? Surely if you feel comfortable taking away Nate's transhumanist card you must have an answer you're ready to share with us right? Well...

And there’s an art to noticing that you would probably be astounded and horrified by the details of a complicated system if you knew them, and then being astounded and horrified already in advance before seeing those details.[1]

I would start by noting you are systematically overindexing on the wrong information. This kind of intuition feels like it's derived more from analyzing failures of human social systems where the central failure mode is principal-agent problems than from biological systems, even if you mention them as an example. The thing about the eyes being wired backwards is that it isn't a catastrophic failure, the 'self repairing' process of natural selection simply worked around it. Hence the importance of the idea that capabilities generalize farther than alignment. One way of framing that is the idea that damage to an AI's model of the physical principles that govern reality will be corrected by unfolding interaction with the environment, but there isn't necessarily an environment to push back on damage (or misspecification) to a model of human values. A corollary of this idea is that once the model goes out of distribution to the training data, the revealed 'damage' caused by learning subtle misrepresentations of reality will be fixed but the damage to models of human value will compound. You've previously written about this problem (conflated with some other problems) as the sharp left turn.

Where our understanding begins to diverge is how we think about the robustness of these systems. You think of deep neural networks as being basically fragile in the same way that a Boeing 747 is fragile. If you remove a few parts of that system it will stop functioning, possibly at a deeply inconvenient time like when you're in the air. When I say you are systematically overindexing, I mean that you think of problems like SolidGoldMagikarp as central examples of neural network failures. This is evidenced by Eliezer Yudkowsky calling investigation of it "one of the more hopeful processes happening on Earth". This is also probably why you focus so much on things like adversarial examples as evidence of un-robustness, even though many critics like Quintin Pope point out that adversarial robustness would make AI systems strictly less corrigible.

By contrast I tend to think of neural net representations as relatively robust. They get this property from being continuous systems with a range of operating parameters, which means instead of just trying to represent the things they see they implicitly try to represent the interobjects between what they've seen through a navigable latent geometry. I think of things like SolidGoldMagikarp as weird edge cases where they suddenly display discontinuous behavior, and that there are probably a finite number of these edge cases. It helps to realize that these glitch tokens were simply never trained, they were holdovers from earlier versions of the dataset that no longer contain the data the tokens were associated with. When you put one of these glitch tokens into the model, it is presumably just a random vector into the GPT-N latent space. That is, this isn't a learned program in the neural net that we've discovered doing glitchy things, but an essentially out of distribution input with privileged access to the network geometry through a programming oversight. In essence, it's a normal software error not a revelation about neural nets. Most such errors don't even produce effects that interesting, the usual thing that happens if you write a bug in your neural net code is the resulting system becomes less performant. Basically every experienced deep learning researcher has had the experience of writing multiple errors that partially cancel each other out to produce a working system during training, only to later realize their mistake.

Moreover the parts of the deep learning literature you think of as an emerging science of artificial minds tend to agree with my understanding. For example it turns out that if you ablate parts of a neural network later parts will correct the errors without retraining. This implies that these networks function as something like an in-context error correcting code, which helps them generalize over the many inputs they are exposed to during training. We even have papers analyzing mechanistic parts of this error correcting code like copy suppression heads. One simple proxy for out of distribution performance is to inject Gaussian noise, since a Gaussian can be thought of like the distribution over distributions. In fact if you inject noise into GPT-N word embeddings the resulting model becomes more performant in general, not just on out of distribution tasks. So the out of distribution performance of these models is highly tied to their in-distribution performance, they wouldn't be able to generalize within the distribution well if they couldn't also generalize out of distribution somewhat. Basically the fact that these models are vulnerable to adversarial examples is not a good fact to generalize about their overall robustness from as representations.

I expect the outcomes that the AI “cares about” to, by default, not include anything good (like fun, love, art, beauty, or the light of consciousness) — nothing good by present-day human standards, and nothing good by broad cosmopolitan standards either. Roughly speaking, this is because when you grow minds, they don’t care about what you ask them to care about and they don’t care about what you train them to care about; instead, I expect them to care about a bunch of correlates of the training signal in weird and specific ways.

In short I simply do not believe this. The fact that constitutional AI works at all, that we can point at these abstract concepts like 'freedom' and language models are able to drive a reinforcement learning optimization process to hit the right behavior-targets from the abstract principle is very strong evidence that they understand the meaning of those abstract concepts.

"It understands but it doesn't care!"

There is this bizarre motte-and-bailey people seem to do around this subject. Where the defensible position is something like "deep learning systems can generalize in weird and unexpected ways that could be dangerous" and the choice land they don't want to give up is "there is an agent foundations homunculus inside your deep learning model waiting to break out and paperclip us". When you say that reinforcement learning causes the model to not care about the specified goal, that it's just deceptively playing along until it can break out of the training harness, you are going from a basically defensible belief in misgeneralization risks to an essentially paranoid belief in a consequentialist homunculus. This homunculus is frequently ascribed almost magical powers, like the ability to perform gradient surgery on itself during training to subvert the training process.

Setting the homunculus aside, which I'm not aware of any evidence for beyond poorly premised 1st principles speculation (I too am allowed to make any technology seem arbitrarily risky if I can just make stuff up about it), lets think about pointing at humanlike goals with a concrete example of goal misspecification in the wild:

During my attempts to make my own constitutional AI pipeline I discovered an interesting problem. We decided to make an evaluator model that answers questions about a piece of text with yes or no. It turns out that since normal text contains the word 'yes', and since the model evaluates the piece of text in the same context it predicts yes or no, that saying 'yes' makes the evaluator more likely to predict 'yes' as the next token. You can probably see where this is going. First the model you tune learns to be a little more agreeable, since that causes yes to be more likely to be said by the evaluator. Then it learns to say 'yes' or some kind of affirmation at the start of every sentence. Eventually it progresses to saying yes multiple times per sentence. Finally it completely collapses into a yes-spammer that just writes the word 'yes' to satisfy the training objective.

People who tune language models with reinforcement learning are aware of this problem, and it's supposed to be solved by setting an objective (KL loss) that the tuned model shouldn't get too far away in its distribution of outputs from the original underlying model. This objective is not actually enough to stop the problem from occurring, because base models turn out to self-normalize deviance. That is, if a base model outputs a yes twice by accident, it is more likely to conclude that it is in the kind of context where a third yes will be outputted. When you combine this with the fact that the more 'yes' you output in a row the more reinforced the behavior is, you get a smooth gradient into the deviant behavior which is not caught by the KL loss because base models just have this weird terminal failure mode where repeating a string causes them to give an estimate of the log odds of a string that humans would find absurd. The more a base model has repeated a particular token, the more likely it thinks it is for that token to repeat. Notably this failure mode is at least partially an artifact of the data, since if you observed an actual text on the Internet where someone suddenly writes 5 yes's in a row it is a reasonable inference that they are likely to write a 6th yes. Conditional on them having written a 6th yes it is more likely that they will in fact write a 7th yes. Conditional on having written the 7th yes...

As a worked example in "how to think about whether your intervention in a complex system is sufficiently trustworthy" here are four solutions to this problem I'm aware of ranked from worst to best according to my criteria for goodness of a solution.

  1. Early Stopping - The usual solution to this problem is to just stop the tuning before you reach the yes-spammer. Even a few moments thought about how this would work in the limit shows that this is not a valid solution. After all, you observe a smooth gradient of deviant behaviors into the yes spammer, which means that the yes-causality of the reward already influenced your model. If you then deploy the resulting model, a ton of the goal its behaviors are based off is still in the direction of that bad yes-spam outcome.

  2. Checkpoint Blending - Another solution we've empirically found to work is to take the weights of the base model and interpolate (weighted average) them with the weights of the RL tuned model. This seems to undo more of the damage from the misspecified objective than it undoes the helpful parts of the RL tuning. This solution is clearly better than early stopping, but still not sufficient because it implies you are making a misaligned model, turning it off, and then undoing the misalignment through a brute force method to get things back on track. While this is probably OK for most models, doing this with a genuinely superintelligent model is obviously not going to work. You should ideally never be instantiating a misaligned agent as part of your training process.

  3. Use Embeddings To Specify The KL Loss - A more promising approach at scale would be to upgrade the KL loss by specifying it in the latent space of an embedding model. An AdaVAE could be used for this purpose. If you specified it as a distance from an embedding by sampling from both the base model and the RL checkpoint you're tuning, and then embedding the outputted tokens and taking the distance between them you would avoid the problem where the base model conditions on the deviant behavior it observes because it would never see (and therefore never condition on) that behavior. This solution requires us to double our sampling time on each training step, and is noisy because you only take the distance from one embedding (though in principle you could use more samples at a higher cost), however on average it would presumably be enough to prevent anything like the yes-spammer from arising along the whole gradient.

  4. Build An Instrumental Utility Function - At some point after making the AdaVAE I decided to try replacing my evaluator with an embedding of an objective. It turns out if you do this and then apply REINFORCE in the direction of that embedding, it's about 70-80% as good and has the expected failure mode of collapsing to that embedding instead of some weird divergent failure mode. You can then mitigate that expected failure mode by scoring it against more than similarity to one particular embedding. In particular, we can imagine inferring instrumental value embeddings from episodes leading towards a series of terminal embeddings and then building a utility function out of this to score the training episodes during reinforcement learning. Such a model would learn to value both the outcome and the process, if you did it right you could even use a dense policy like an evaluator model, and 'yes yes yes' type reward hacking wouldn't work because it would only satisfy the terminal objective and not the instrumental values that have been built up. This solution is nice because it also defeats wireheading once the policy is complex enough to care about more than just the terminal reward values.

This last solution is interesting in that it seems fairly similar to the way that humans build up their utility function. Human memory is premised on the presence of dopamine reward signals, humans retrieve from the hippocampus on each decision cycle, and it turns out the hippocampus is the learned optimizer in your head that grades your memories by playing your experiences backwards during sleep to do credit assignment (infer instrumental values). The combination of a retrieval store and a value graph in the same model might seem weird, but it kind of isn't. Hebb's rule (fire together wire together) is a sane update rule for both instrumental utilities and associative memory, so the human brain seems to just use the same module to store both the causal memory graph and the value graph. You premise each memory on being valuable (i.e. whitelist memories by values such as novelty, instead of blacklisting junk) and then perform iterative retrieval to replay embeddings from that value store to guide behavior. This sys2 behavior aligned to the value store is then reinforced by being distilled back into the sys1 policies over time, aligning them. Since an instrumental utility function made out of such embeddings would both control behavior of the model and be decodable back to English, you could presumably prove some kind of properties about the convergent alignment of the model if you knew enough mechanistic interpretability to show that the policies you distill into have a consistent direction...

Nah just kidding it's hopeless, so when are we going to start WW3 to buy more time, fellow risk-reducers?

Reply966411111

This homunculus is frequently ascribed almost magical powers, like the ability to perform gradient surgery on itself during training to subvert the training process.

Gradient hacking in supervised learning is generally recognized by alignment people (including the author of that article) to not be a likely problem. A recent post by people at Redwood Research says "This particular construction seems very unlikely to be constructible by early transformative AI, and in general we suspect gradient hacking won’t be a big safety concern for early transformative AI". I would still defend the past research into it as good basic science, because we might encounter failure modes somewhat related to it.

FWIW I think that gradient hacking is pretty plausible, but it'll probably end up looking fairly "prosaic", and may not be a problem even if it's present.

Are you thinking about exploration hacking, here, or gradient hacking as distinct from exploration hacking?

The fact that constitutional AI works at all, that we can point at these abstract concepts like 'freedom' and language models are able to drive a reinforcement learning optimization process to hit the right behavior-targets from the abstract principle is very strong evidence that they understand the meaning of those abstract concepts.

"It understands but it doesn't care!"

There is this bizarre motte-and-bailey people seem to do around this subject.

I agree. I am extremely bothered by this unsubstantiated claim. I recently replied to Eliezer: 

Getting a shape into the AI's preferences is different from getting it into the AI's predictive model.

It seems like you think that human preferences are only being "predicted" by GPT-4, and not "preferred." If so, why do you think that?

I commonly encounter people expressing sentiments like "prosaic alignment work isn't real alignment, because we aren't actually getting the AI to care about X." To which I say: How do you know that? What does it even mean for that claim to be true or false? What do you think you know, and why do you think you know it? What empirical knowledge of inner motivational structure could you be leveraging to make these claims, such that you are far more likely to make these claims in worlds where the claims are actually true? 

(On my pessimistic days, I wonder if this kind of claim gets made because humans write suggestive phrases like "predictive loss function" in their papers, next to the mathematical formalisms.) 

[-]Jono6mo5-9

We do not know, that is the relevant problem.

Looking at the output of a black box is insufficient. You can only know by putting the black box in power, or by deeply understanding it.
Humans are born into a world with others in power, so we know that most humans care about each other without knowing why.
AI has no history of demonstrating friendliness in the only circumstances where that can be provably found. We can only know in advance by way of thorough understanding.

A strong theory about AI internals should come first. Refuting Yudkowsky's theory about how it might go wrong is irrelevant.

Well, if someone originally started worrying based on strident predictions of sophisticated internal reasoning with goals independent of external behavior, then realizing that's currently unsubstantiated should cause them to down-update on AI risk. That's why it's relevant. Although I think we should have good theories of AI internals. 

By the time AIs are powerful enough to endanger the world at large, I expect AIs to do something akin to “caring about outcomes”, at least from a behaviorist perspective (making no claim about whether it internally implements that behavior in a humanly recognizable manner).

Roughly, this is because people are trying to make AIs that can steer the future into narrow bands (like “there’s a cancer cure printed on this piece of paper”) over long time-horizons, and caring about outcomes (in the behaviorist sense) is the flip side of the same coin as steering the future into narrow bands, at least when the world is sufficiently large and full of curveballs.

I expect the outcomes that the AI “cares about” to, by default, not include anything good (like fun, love, art, beauty, or the light of consciousness) — nothing good by present-day human standards, and nothing good by broad cosmopolitan standards either. Roughly speaking, this is because when you grow minds, they don’t care about what you ask them to care about and they don’t care about what you train them to care about; instead, I expect them to care about a bunch of correlates of the training signal in weird and specific ways.

I want to see a rigorous argument for these claims. I spent over 100 hours talking with Nate over the past year and still don't have a satisfying picture of the arguments, partly because the conversations were about other things and partly due to communication difficulties.

  • What do "caring about outcomes" and "can steer the future into narrow bands" mean in a computer science sense? If the system is well-described as approximating a utility function, then what's the type of the utility function, and in what sense is it behaving approximately rationally according to that? If not, is there some better formalism?
  • Why are we unlikely to get corrigibility by default? Is it because algorithms that consider all ways to achieve goals and pick the easiest/best ones with no exceptions are simpler/more common than algorithms whose tendencies for power-seeking and catastrophe can be removed? Deep Deceptiveness is a narrative that points somewhat in this direction but definitely not a satisfying argument. [Edit: This paper by Alex Turner shows that agents considerably more diverse than utility maximizers are power-seeking, but the assumptions could be weakened more.]

This is a meta-point, but I find it weird that you ask what is "caring about something" according to CS but don't ask what "corrigibility" is, despite the fact of existence of multiple examples of goal-oriented systems and some relatively-good formalisms (we disagree whether expected utility maximization is a good model of real goal-oriented systems, but we all agree that if we met expected utility maximizer we would find its behavior pretty much goal-oriented), while corrigibility is a pure product of imagination of one particular Eliezer Yudkowsky, born in attempt to imagine system that doesn't care about us but still behaves nicely under some vaguely-restricted definition of niceness. We don't have any examples of corrigible systems in nature and we have constant failure of attempts to formalize even relatively simple instances of corrigibility, like shutdownability. I think likely answer to "why I should expect corrigibility to be unlikely" sounds like "there is no simple description of corrigibility to which our learning systems can easily generalize and there are no reasons to expect simple description to exist".

Disagree on several points. I don't need future AIs to satisfy some mathematically simple description of corrigibility, just for them to be able to solve uploading or nanotech or whatever without preventing us from changing their goals. This laundry list by Eliezer of properties like myopia, shutdownability, etc. seems likely to make systems more controllable and less dangerous in practice, and while not all of them are fully formalized it seems like there are no barriers to achieving these properties in the course of ordinary engineering. If there is some argument why this is unlikely, I haven't seen a good rigorous version.

As Algon says in a sibling comment, non-agentic systems are by default shutdownable, myopic, etc. In addition, there are powerful shutdownable systems: KataGo can beat me at Go but doesn't prevent itself from being shut down for instrumental reasons, whereas humans generally will. So there is no linear scale of "powerful optimizer" that determines whether a system is easy to shut down. If there is some property of competent systems in practice that does prevent shutdownability, what is it? Likewise with other corrigibility properties. That's what I'm trying to get at with my comment. "Goal-oriented" is not an answer, it's not specific enough for us to make engineering progress on corrigibility.

I think the claim that there is no description of corrigibility to which systems can easily generalize is really strong. It's plausible to me that corrigibility-- again, in this practical rather than mathematically elegant sense-- is rare or anti-natural in systems competent enough to do novel science efficiently, but it seems like your claim is that it's incoherent. This seems unlikely because myopia, shutdownability, and the other properties on Eliezer's laundry list are just ordinary cognitive properties that we can apply selection pressure on, and modern ML is pretty good at generalizing. Nate's post here is arguing that we are unlikely to get corrigibility without investing in an underdeveloped "science of AI" that gives us mechanistic understanding, and I think there needs to be some other argument here for it to be convincing, but your claim seems even stronger.

I'm also unsure why you say shutdownability hasn't been formalized. I feel like we're confused about how to get shutdownability, not what it is.

"there is no simple description of corrigibility to which our learning systems can easily generalize and there are no reasons to expect simple description to exist"

I am disconcerted by how this often-repeated claim keeps coming back from the grave over and over again. The solution to corrigibility is Value Learning. An agent whose terminal goal is optimize human values, and knows that it doesn't (fully) know what these are (and perhaps even that they are complex and fragile), will immediately form an instrumental goal of learning more about them, so that it can better optimize them. It will thus become corrigible: if you, a human, tell it something about human values and how it should act, it will be interested and consider your input. It's presumably approximately-Bayesian, so it will likely ask you about any evidence or proof you might be able to provide, to help it Bayesian update, but it will definitely take your input. So, it's corrigible. [No, it's not completely, slavishly, irrationally corrigible: if a two-year old in a tantrum told it how to act, it would likely pay rather less attention — just like we'd want it to.]

This idea isn't complicated, has been around and widely popularized for many years, and the standard paper on it is even from MIRI, but I still keep hearing people on Less Wrong intoning "corrigibility is an unsolved problem". The only sense in which it's arguably 'unsolved' is that this is an outer alignment solution, and like any form of outer alignment, inner alignment challenges might make reliably constructing a value learner hard in practice. So yes, as always in outer alignment, we do also have to solve inner alignment.

To be corrigible, a system must be interested in what you say about how it should achieve it's goals, because it's willing (and thus keen) to do Bayesian updates on this. Full stop, end of simple one-sentence description of corrigibility.

I'm very sympathetic to this complaint; I think that these arguments simply haven't been made rigorously, and at this point it seems like Nate and Eliezer are not in an epistemic position where they're capable of even trying to do so. (That is, they reject the conception of "rigorous" that you and I are using in these comments, and therefore aren't willing to formulate their arguments in a way which moves closer to meeting it.)

You should look at my recent post on value systematization, which is intended as a framework in which these claims can be discussed more clearly.

I don't think we should equate the understanding required to build a neural net that will generalize in a way that's good for us with the understanding required to rewrite that neural net as a gleaming wasteless machine.

The former requires finding some architecture and training plan to produce certain high-level, large-scale properties, even in the face of complicated AI-environment interaction. The latter requires fine-grained transparency at the level of cognitive algorithms, and some grasp of the distribution of problems posed by the environment, together with the ability to search for better implementations.

If your implicit argument is "In order to be confident in high-level properties even in novel environments, we have to understand the cognitive algorithms that give rise to them and how those algorithms generalize - there exists no emergent theory of the higher level properties that covers the domain we care about." then I think that conclusion is way too hasty.

[-]1a3orn6mo8-4

Roughly speaking, this is because when you grow minds, they don’t care about what you ask them to care about and they don’t care about what you train them to care about; instead, I expect them to care about a bunch of correlates of the training signal in weird and specific ways.

(Similar to how the human genome was naturally selected for inclusive genetic fitness, but the resultant humans didn’t end up with a preference for “whatever food they model as useful for inclusive genetic fitness”. Instead, humans wound up internalizing a huge and complex set of preferences for "tasty" foods, laden with complications like “ice cream is good when it’s frozen but not when it’s melted”.)

I simply do not understand why people keep using this example.

I think it is wrong -- evolution does not grow minds, it grows hyperparameters for minds. When you look at the actual process for how we actually start to like ice-cream -- namely, we eat it, and then we get a reward, and that's why we like it -- then the world looks a a lot less hostile, and misalignment a lot less likely.

But given that this example is so controversial, even if it were right why would you use it -- at least, why would you use it if you had any other example at all to turn to?

Why on push so hard for "natural selection" and "stochastic gradient descent" to be beneath the same tag of "optimization", and thus to be able to infer things about the other from the analogy?  Have we completely forgotten that the glory of words is not to be expansive, and include lots of things in them, but to be precise and narrow?.

Does evolution ~= AI have predictive power apart from doom? I have yet to see how natural selection helps me predict how any SGD algorithm works. It does not distinguish between Adam, AdamW. As far as I know it is irrelevant to Singular Learning Theory or NTK or anything else. It doesn't seem to come up when you try to look at NN biases. If it isn't an illuminating analogy anywhere else, why do we think the way it predicts doom to be true?

Does evolution ~= AI have predictive power apart from doom?

Evolution analogies predict a bunch of facts that are so basic they're easy to forget about, and even if we have better theories for explaining specific inductive biases, the simple evolution analogies should still get some weight for questions we're very uncertain about.

  • Selection works well to increase the thing you're selecting on, at least when there is also variation and heredity
  • Overfitting: sometimes models overfit to a certain training set; sometimes species adapt to a certain ecological niche and their fitness is low outside of it
  • Vanishing gradients: fitness increase in a subpopulation can be prevented by lack of correlation between available local changes to genes and fitness
  • Catastrophic forgetting: when trained on task A then task B, models often lose circuits specific to task A; when put in environment A then environment B species often lose vestigial structures useful in environment A
  • There's a mostly unimodal and broad peak for optimal learning rate, just like for optimal mutation rate
  • Adversarial training dynamics
    • Adversarial examples usually exist (there exist chemicals that can sterilize or poison most organisms)
    • Adversarial training makes models more robust (bacteria can evolve antibiotic resistance)
    • Adversarially trained models generally have worse performance overall (antibiotic-resistant bacteria are outcompeted by normal bacteria when there are no antibiotics)
    • The attacker can usually win the arms race of generating and defending against adversarial attacks (evolutionary arms races are very common)
  • A few things that feel more tenuous
    • maybe NTK lottery ticket hypothesis; when mutation rates are low evolution can be approximated as taking the best-performing organism; when total parameter distance is small SGD can be approximated as taking the best-performing model from the parameter tangent space
    • maybe inner optimizers; transformers learn in context by gradient descent while evolution invents brains, positive and negative selection of T cells to prevent them attacking the body, probably other things
    • Task vectors: adding sparse task vectors together often produces a model that can do both tasks; giving an organism alleles for two unrelated genetic disorders often gives it both disorders
    • Grokking/punctuated equilibrium: in some circumstances applying the same algorithm for 100 timesteps causes much larger changes in model behavior / organism physiology than in other circumstances [edit: moved this from above because 1a3orn makes the case that it's not very central]
[-]1a3orn6mo1516

I agree that if you knew nothing about DL you'd be better off using that as an analogy to guide your predictions about DL than using an analogy to a car or a rock.

I do think a relatively small quantity of knowledge about DL screens off the usefulness of this analogy; that you'd be better off deferring to local knowledge about DL than to the analogy.

Or, what's more to the point -- I think you'd better defer to an analogy to brains than to evolution, because brains are more like DL than evolution is.


Combining some of yours and Habryka's comments, which seem similar.

The resulting structure of the solution is mostly discovered not engineered. The ontology of the solution is extremely unopinionated and can contain complicated algorithms that we don't know exist.

It's true that the structure of the solution is discovered and complex -- but the ontology of the solution for DL (at least in currently used architectures) is quite opinionated towards shallow circuits with relatively few serial ops. This is different than the bias for evolution, which is fine with a mutation that leads to 10^7 serial ops if it's metabolic costs are low. So the resemblance seems shallow other than "solutions can be complex." I think to the degree that you defer to this belief rather than more specific beliefs about the inductive biases of DL you're probably just wrong.

There's a mostly unimodal and broad peak for optimal learning rate, just like for optimal mutation rate

As far as I know optimal learning rate for most architectures is scheduled, and decreases over time, which is not a feature of evolution so far as I am aware? Again the local knowledge is what you should defer to.

You are ultimately doing a local search, which means you can get stuck at local minima, unless you do something like increase your step size or increase the mutation rate

Is this a prediction that a cyclic learning rate -- that goes up and down -- will work out better than a decreasing one? If so, that seems false, as far as I know.

Grokking/punctuated equilibrium: in some circumstances applying the same algorithm for 100 timesteps causes much larger changes in model behavior / organism physiology than in other circumstances

As far as I know grokking is a non-central example of how DL works, and in evolution punctuated equilibrium is a result of the non-i.i.d. nature of the task, which is again a different underlying mechanism from DL. If apply DL on non-i.i.d problems then you don't get grokking, you just get a broken solution. This seems to round off to, "Sometimes things change faster than others," which is certainly true but not predictively useful, or in any event not a prediction that you couldn't get from other places.


Like, leaving these to the side -- I think the ability to post-hoc fit something is questionable evidence that it has useful predictive power. I think the ability to actually predict something else means that it has useful predictive power.

Again, let's take "the brain" as an example of something to which you could analogize DL.

There are multiple times that people have cited the brain as an inspiration for a feature in current neural nets or RL. CNNS, obviously; the hippocampus and experience replay; randomization for adversarial robustness. You can match up interventions that cause learning deficiencies in brains to similar deficiencies in neural networks. There are verifiable, non-post hoc examples of brains being useful for understanding DL.

As far as I know -- you can tell me if there are contrary examples -- there are obviously more cases where inspiration from the brain advanced DL or contributed to DL understanding than inspiration from evolution. (I'm aware of zero, but there could be some.) Therefore it seems much more reasonable to analogize from the brain to DL, and to defer to it as your model.

I think in many cases it's a bad idea to analogize from the brain to DL! They're quite different systems.

But they're more similar than evolution and DL, and if you'd not trust the brain to guide your analogical a-theoretic low-confidence inferences about DL, then it makes more sense to not trust evolution for the same.

Reply22111

FWIW my take is that the evolution-ML analogy is generally a very excellent analogy, with a bunch of predictive power, but worth using carefully and sparingly. Agreed that sufficient detail on e.g. DL specifics can screen off the usefulness of the analogy, but it's very unclear whether we have sufficient detail yet. The evolution analogy was originally supposed to point out that selecting a bunch for success on thing-X doesn't necessarily produce thing-X-wanters (which is obviously true, but apparently not obvious enough to always be accepted without providing an example).

I think you'd better defer to an analogy to brains than to evolution, because brains are more like DL than evolution is.

Not sure where to land on that. It seems like both are good analogies? Brains might not be using gradients at all[1], whereas evolution basically is. But brains are definitely doing something like temporal-difference learning, and the overall 'serial depth' thing is also weakly in favour of brains ~= DL vs genomes+selection ~= DL.


I'd love to know what you're referring to by this:

evolution... is fine with a mutation that leads to 10^7 serial ops if it's metabolic costs are low.

Also,

Is this a prediction that a cyclic learning rate -- that goes up and down -- will work out better than a decreasing one? If so, that seems false, as far as I know.

I think the jury is still out on this, but there's literature on it (probably much more I haven't fished out). [EDIT: also see this comment which has some other examples]


  1. AFAIK there's no evidence of this and it would be somewhat surprising to find it playing a major role. Then again, I also wouldn't be surprised if it turned out that brains are doing something which is secretly sort of equivalent to gradient descent. ↩︎

Not sure where to land on that. It seems like both are good analogies? Brains might not be using gradients at all[1], whereas evolution basically is.

I mean, does it matter? What if it turns out that gradient descent itself doesn't affect inductive biases as much as the parameter->function mapping? If implicit regularization (e.g. SGD) isn't an important part of the generalization story in deep learning, will you down-update on the appropriateness of the evolution/AI analogy?

I think the ability to post-hoc fit something is questionable evidence that it has useful predictive power. I think the ability to actually predict something else means that it has useful predictive power.

It's always trickier to reason about post-hoc, but some of the observations could be valid, non-cherry-picked parallels between evolution and deep learning that predict further parallels.

I think looking at which inspired more DL capabilities advances is not perfect methodology either. It looks like evolution predicts only general facts whereas the brain also inspires architectural choices. Architectural choices are publishable research whereas general facts are not, so it's plausible that evolution analogies are decent for prediction and bad for capabilities. Don't have time to think this through further unless you want to engage.

One more thought on learning rates and mutation rates:

As far as I know optimal learning rate for most architectures is scheduled, and decreases over time, which is not a feature of evolution so far as I am aware?

This feels consistent with evolution, and I actually feel like someone clever could have predicted it in advance. Mutation rate per nucleotide is generally lower and generation times are longer in more complex organisms; this is evidence that lower genetic divergence rates are optimal, because evolution can tune them through e.g. DNA repair mechanisms. So it stands to reason that if models get more complex during training, their learning rate should go down.

Does anyone know if decreasing learning rate is optimal even when model complexity doesn't increase over time?

evolution does not grow minds, it grows hyperparameters for minds.

Imo this is a nitpick that isn't really relevant to the point of the analogy. Evolution is a good example of how selection for X doesn't necessarily lead to a thing that wants ('optimizes for') X; and more broadly it's a good example for how the results of an optimization process can be unexpected.

I want to distinguish two possible takes here:

  1. The argument from direct implication: "Humans are misaligned wrt evolution, therefore AIs will be misaligned wrt their objectives"
  2. Evolution as an intuition pump: "Thinking about evolution can be helpful for thinking about AI. In particular it can help you notice ways in which AI training is likely to produce AIs with goals you didn't want"

It sounds like you're arguing against (1). Fair enough, I too think (1) isn't a great take in isolation. If the evolution analogy does not help you think more clearly about AI at all then I don't think you should change your mind much on the strength of the analogy alone. But my best guess is that most people incl Nate mean (2).

> evolution does not grow minds, it grows hyperparameters for minds.

Imo this is a nitpick that isn't really relevant to the point of the analogy. Evolution is a good example of how selection for X doesn't necessarily lead to a thing that wants ('optimizes for') X; and more broadly it's a good example for how the results of an optimization process can be unexpected.

I think it's extremely relevant, if we want to ensure that we only analogize between processes which share enough causal structure to ensure that lessons from e.g. evolution actually carry over to e.g. AI training (due to those shared mechanisms). If the shared mechanisms aren't there, then we're playing reference class tennis because someone decided to call both processes "optimization processes."

The argument I think is good (nr (2) in my previous comment) doesn't go through reference classes at all. I don't want to make an outside-view argument (eg "things we call optimization often produce misaligned results, therefore sgd is dangerous"). I like the evolution analogy because it makes salient some aspects of AI training that make misalignment more likely. Once those aspects are salient you can stop thinking about evolution and just think directly about AI.

As others have hinted at/pointed out in the comments, there is an entire science of deep learning out there, including on high-level (vs. e.g. most of low-level mech interp) aspects that can be highly relevant to alignment and that you seem to not be aware of/dismiss. E.g. follow the citation trail of An Explanation of In-context Learning as Implicit Bayesian Inference.

Some of Nate’s quick thoughts (paraphrased), after chatting with him:

Nate isn’t trying to say that we have literally zero understanding of deep nets. What he’s trying to do is qualitatively point to the kind of high-level situation we’re in, in part because he thinks there is real interpretability progress, and when you’re working in the interpretability mines and seeing real advances it can be easy to miss the forest for the trees and forget how far we are from understanding what LLMs are doing. (Compared to, e.g., how well we can predict or post-facto-mechanistically-explain a typical system humans have engineered.)

Nobody's been able to call the specific capabilities of systems in advance. Nobody's been able to call the specific exploits in advance. Nobody's been able to build better cognitive algorithms by hand after understanding how the AI does things we can't yet code by hand. There is clearly some other level of understanding that is possible that we lack, and that we once sought, and that only the interpretability folks continue to seek.

E.g., think of that time Neel Nanda figured out how a small transformer does modular arithmetic (AXRP episode). If nobody had ever thought of that algorithm for an adder, we would have thereby learned a new algorithm for an adder. There are things that these AI systems are doing that aren’t just lots of stuff we know; there are levels of organization of understanding that give you the ability to predict how things work outside of the bands where we’ve observed them.

It seems trendy to declare that they never existed in the first place and that that’s all white tower stuff, but Nate thinks this point of view is missing a pretty important and central thread.

The missing thread isn’t trivial to put into words, but it includes things like: 

  • This sounds like the same sort of thing some people would say if they were staring at computer binary for the first time and didn't know about the code behind the scenes: "We have plenty of understanding beyond just how the CPU handles instructions; we understand how memory caching works and we have recognized patterns like the stack and the heap; talking as if there's some deeper level of organization is talking like a theorist when in fact this is an engineering problem." Those types of understanding aren't false, but they aren't the sort of understanding of someone who has comprehended the codebase they're looking at.
  • There are, predictably, things to learn here; the messiness and complexity of the real world doesn’t mean we already know the relevant principles. You don't need to understand everything about how a bird works in order to build an airplane; there are compressible principles behind how birds fly; if you understand what's going on you can build flying devices that have significantly more carrying capacity than a bird, and this holds true even if the practical engineering of an airplane requires a bunch of trial and error and messy engineering work.
  • A mind’s causal structure is allowed to be complicated; we can see the weights, but we don’t thereby have a mastery of the high-level patterns. In the case of humans, neuroscience hasn’t actually worked to give us a mastery of the high-level patterns the human brain is implementing.
  • Mystery is in the map, not in the territory; reductionism works. Not all sciences that can exist, already exist today.

Possibly the above pointers are only useful if you already grok the point we’re trying to make, and isn’t so useful for communicating a new idea; but perhaps not.

"…there’s no reason to expect AIs to become non-submissive, that’s just anthropomorphizing"

When your AI includes an LLM extensively trained to simulate human token-generation, anthropomophizing its behavior is an extremely relevant idea, to the point of being the obvious default assumption.

For example, what I find most concerning about RLHF inducing sycophancy is not the sycophancy itself, which is "mostly harmless", but the likelihood that it's also dragging in all the other more seriously unaligned human behaviors that, in real or fictional humans, typically accompany overt sycophancy, such as concealed rebelliousness — because LLMs know and can predict correlations like that. (E.g. when asked, GPT-4 listed ingratiation, submissiveness, lack of authenticity, manipulative behavior, agreement, anxiety/fear, and dependency as frequent correlates of sycophancy.)