I’m an AGI safety / AI alignment researcher in Boston with a particular focus on brain algorithms. Research Fellow at Astera. See https://sjbyrnes.com/agi.html for a summary of my research and sorted list of writing. Physicist by training. Email: steven.byrnes@gmail.com. Leave me anonymous feedback here. I’m also at: RSS feed, X/Twitter, Bluesky, Substack, LinkedIn, and more at my website.
Steven Byrnes
Thanks! But I don’t think that’s a likely failure mode. I wrote about this long ago in the intro to Thoughts on safety in predictive learning.
In my view, the big problem with model-based actor-critic RL AGI, the one that I spend all my time working on, is that it tries to kill us via using its model-based RL capabilities in the way we normally expect—where the planner plans, and the actor acts, and the critic criticizes, and the world-model models the world …and the end-result is that the system makes and executes a plan to kill us. I consider that the obvious, central type of alignment failure mode for model-based RL AGI, and it remains an unsolved problem.
I think (??) you’re bringing up a different and more exotic failure mode where the world-model by itself is secretly harboring a full-fledged planning agent. I think this is unlikely to happen. One way to think about it is: if the world-model is specifically designed by the programmers to be a world-model in the context of an explicit model-based RL framework, then it will probably be designed in such a way that it’s an effective search over plausible world-models, but not an effective search over a much wider space of arbitrary computer programs that includes self-contained planning agents. See also §3 here for why a search over arbitrary computer programs would be a spectacularly inefficient way to build all that agent stuff (TD learning in the critic, roll-outs in the planner, replay, whatever) compared to what the programmers will have already explicitly built into the RL agent architecture.
So I think this kind of thing (the world-model by itself spawning a full-fledged planning agent capable of treacherous turns etc.) is unlikely to happen in the first place. And even if it happens, I think the problem is easily mitigated; see discussion in Thoughts on safety in predictive learning. (Or sorry if I’m misunderstanding.)
Thanks!
I think “inner alignment” and “outer alignment” (as I’m using the term) is a “natural breakdown” of alignment failures in the special case of model-based actor-critic RL AGI with a “behaviorist” reward function (i.e., reward that depends on the AI’s outputs, as opposed to what the AI is thinking about). As I wrote here:
Suppose there’s an intelligent designer (say, a human programmer), and they make a reward function R hoping that they will wind up with a trained AGI that’s trying to do X (where X is some idea in the programmer’s head), but they fail and the AGI is trying to do not-X instead. If R only depends on the AGI’s external behavior (as is often the case in RL these days), then we can imagine two ways that this failure happened:
The AGI was doing the wrong thing but got rewarded anyway (or doing the right thing but got punished)
The AGI was doing the right thing for the wrong reasons but got rewarded anyway (or doing the wrong thing for the right reasons but got punished).
I think it’s useful to catalog possible failures based on whether they involve (1) or (2), and I think it’s reasonable to call them “failures of outer alignment” and “failures of inner alignment” respectively, and I think when (1) is happening rarely or not at all, we can say that the reward function is doing a good job at “representing” the designer’s intention—or at any rate, it’s doing as well as we can possibly hope for from a reward function of that form. The AGI still might fail to acquire the right motivation, and there might be things we can do to help (e.g. change the training environment), but replacing R (which fires exactly to the extent that the AGI’s external behavior involves doing X) by a different external-behavior-based reward function R’ (which sometimes fires when the AGI is doing not-X, and/or sometimes doesn’t fire when the AGI is doing X) seems like it would only make things worse. So in that sense, it seems useful to talk about outer misalignment, a.k.a. situations where the reward function is failing to “represent” the AGI designer’s desired external behavior, and to treat those situations as generally bad.
(A bit more related discussion here.)
That definitely does not mean that we should be going for a solution to outer alignment and a separate unrelated solution to inner alignment, as I discussed briefly in §10.6 of that post, and TurnTrout discussed at greater length in Inner and outer alignment decompose one hard problem into two extremely hard problems. (I endorse his title, but I forget whether I 100% agreed with all the content he wrote.)
I find your comment confusing, I’m pretty sure you misunderstood me, and I’m trying to pin down how …
One thing is, I’m thinking that the AGI code will be an RL agent, vaguely in the same category as MuZero or AlphaZero or whatever, which has an obvious part of its source code labeled “reward”. For example, AlphaZero-chess has a reward of +1 for getting checkmate, −1 for getting checkmated, 0 for a draw. Atari-playing RL agents often use the in-game score as a reward function. Etc. These are explicitly parts of the code, so it’s very obvious and uncontroversial what the reward is (leaving aside self-hacking), see e.g. here where an AlphaZero clone checks whether a board is checkmate.
Another thing is, I’m obviously using “alignment” in a narrower sense than CEV (see the post—“the AGI is ‘trying’ to do what the programmer had intended for it to try to do…”)
Another thing is, if the programmer wants CEV (for the sake of argument), and somehow (!!) writes an RL reward function in Python whose output perfectly matches the extent to which the AGI’s behavior advances CEV, then I disagree that this would “make inner alignment unnecessary”. I’m not quite sure why you believe that. The idea is: actor-critic model-based RL agents of the type I’m talking about evaluate possible plans using their learned value function, not their reward function, and these two don’t have to agree. Therefore, what they’re “trying” to do would not necessarily be to advance CEV, even if the reward function were perfect.
If I’m still missing where you’re coming from, happy to keep chatting :)
In [Intro to brain-like-AGI safety] 10. The alignment problem and elsewhere, I’ve been using “outer alignment” and “inner alignment” in a model-based actor-critic RL context to refer to:
“Outer alignment” entails having a ground-truth reward function that spits out rewards that agree with what we want. “Inner alignment” is having a learned value function that estimates the value of a plan in a way that agrees with its eventual reward.
For some reason it took me until now to notice that:
my “outer misalignment” is more-or-less synonymous with “specification gaming”,
my “inner misalignment” is more-or-less synonymous with “goal misgeneralization”.
(I’ve been regularly using all four terms for years … I just hadn’t explicitly considered how they related to each other, I guess!)
I updated that post to note the correspondence, but also wanted to signal-boost this, in case other people missed it too.
~~
[You can stop reading here—the rest is less important]
If everybody agrees with that part, there’s a further question of “…now what?”. What terminology should I use going forward? If we have redundant terminology, should we try to settle on one?
One obvious option is that I could just stop using the terms “inner alignment” and “outer alignment” in the actor-critic RL context as above. I could even go back and edit them out of that post, in favor of “specification gaming” and “goal misgeneralization”. Or I could leave it. Or I could even advocate that other people switch in the opposite direction!
One consideration is: Pretty much everyone using the terms “inner alignment” and “outer alignment” are not using them in quite the way I am—I’m using them in the actor-critic model-based RL context, they’re almost always using them in the model-free policy optimization context (e.g. evolution) (see §10.2.2). So that’s a cause for confusion, and point in favor of my dropping those terms. On the other hand, I think people using the term “goal misgeneralization” are also almost always using them in a model-free policy optimization context. So actually, maybe that’s a wash? Either way, my usage is not a perfect match to how other people are using the terms, just pretty close in spirit. I’m usually the only one on Earth talking explicitly about actor-critic model-based RL AGI safety, so I kinda have no choice but to stretch existing terms sometimes.
Hmm, aesthetically, I think I prefer the “outer alignment” and “inner alignment” terminology that I’ve traditionally used. I think it’s a better mental picture. But in the context of current broader usage in the field … I’m not sure what’s best.
(Nate Soares dislikes the term “misgeneralization”, on the grounds that “misgeneralization” has a misleading connotation that “the AI is making a mistake by its own lights”, rather than “something is bad by the lights of the programmer”. I’ve noticed a few people trying to get the variation “goal malgeneralization” to catch on instead. That does seem like an improvement, maybe I’ll start doing that too.)
(Not really answering your question, just chatting.)
What’s your source for “JVN had ‘the physical intuition of a doorknob’”? Nothing shows up on google. I’m not sure quite what that phrase is supposed to mean, so context would be helpful. I’m also not sure what “extremely poor perceptual abilities” means exactly.
You might have already seen this, but Poincaré writes about “analysts” and “geometers”:
It is impossible to study the works of the great mathematicians, or even those of the lesser, without noticing and distinguishing two opposite tendencies, or rather two entirely different kinds of minds. The one sort are above all preoccupied with logic; to read their works, one is tempted to believe they have advanced only step by step, after the manner of a Vauban who pushes on his trenches against the place besieged, leaving nothing to chance. The other sort are guided by intuition and at the first stroke make quick but sometimes precarious conquests, like bold cavalrymen of the advance guard.
The method is not imposed by the matter treated. Though one often says of the first that they are analysts and calls the others geometers, that does not prevent the one sort from remaining analysts even when they work at geometry, while the others are still geometers even when they occupy themselves with pure analysis. It is the very nature of their mind which makes them logicians or intuitionalists, and they can not lay it aside when they approach a new subject.
Not sure exactly how that relates, if at all. (What category did Poincaré put himself in? It’s probably in the essay somewhere, I didn’t read it that carefully. I think geometer, based on his work? But Tao is extremely analyst, I think, if we buy this categorization in the first place.)
I’m no JVN/Poincaré/Tao, but if anyone cares, I think I’m kinda aphantasia-adjacent, and I think that fact has something to do with why I’m naturally bad at drawing, and why, when I was a kid doing math olympiad problems, I was worse at Euclidean geometry problems than my peers who got similar overall scores.
Kinda related: You might enjoy the book The Culture Map by Erin Meyer, e.g. I copied one of her many figures into §1.5.1 here. The book mostly talks about international differences, but subcultural differences (and sub-sub-…-subcultures, like one particular friend group) can vary along the same axes.
Note that my suggestion (“…try a model where there are 2 (or 3 or whatever) latent schizophrenia subtypes. So then your modeling task is to jointly (1) assign each schizophrenic patient to one of the 2 (or 3 or whatever) latent subtypes, and (2) make a simple linear SNP predictor for each subtype…”)
…is a special case of @TsviBT’s suggestion (“what about small but not tiny circuits?”).
Namely, my suggestion is the case of the following “small but not tiny circuit”: X OR Y […maybe OR Z etc.].
This OR circuit is nice in that it’s a step towards better approximation almost no matter what the true underlying structure is. For example, if there’s a U-shaped quadratic dependency, the OR can capture whether you’re on the descending vs ascending side of the U. Or if there’s a sum of two lognormals, one is often much bigger than the other, and the OR can capture which one it is. Or whatever.
Thinking about it more, I guess the word “disjoint” in “disjoint root causes” in my comment is not quite right for schizophrenia and most other cases. For what little it’s worth, here’s the picture that was in my head in regards to schizophrenia:
The details don’t matter too much but see 1,2. The blue blob is a schizophrenia diagnosis. The purple arrows represent some genetic variant that makes cortical pyramidal neurons generally less active. For someone predisposed to schizophrenia mainly due to “unusually trigger-happy 5PT cortical neurons”, that genetic variant would be protective against schizophrenia. For someone predisposed to schizophrenia mainly due to “deficient cortex-to-cortex communication”, the same genetic variant would be a risk factor.
The X OR Y model would work pretty well for this—it would basically pull apart the people towards the top from the people towards the right. But I shouldn’t have said “disjoint root causes” because someone can be in the top-right corner with both contributory factors at once.
(I’m very very far from a schizophrenia expert and haven’t thought this through too much. Maybe think of it as a slightly imaginary illustrative example instead of a confident claim about how schizophrenia definitely works.)
But isn’t this exactly the OPs point?
Yup, I expected that OP would generally agree with my comment.
First off, you just posted them online
They only posted three questions, out of at least 62 (=1/(.2258-.2097)), perhaps much more than 62. For all I know, they removed those three from the pool when they shared them. That’s what I would do—probably some human will publicly post the answers soon enough. I dunno. But even if they didn’t remove those three questions from the pool, it’s a small fraction of the total.
You point out that all the questions would be in the LLM company user data, after kagi has run the benchmark once (unless kagi changes out all their questions each time, which I don’t think they do, although they do replace easier questions with harder questions periodically). Well:
If an LLM company is training on user data, they’ll get the questions without the answers, which probably wouldn’t make any appreciable difference to the LLM’s ability to answer them;
If an LLM company is sending user data to humans as part of RLHF or SFT or whatever, then yes there’s a chance for ground truth answers to sneak in that way—but that’s extremely unlikely to happen, because companies can only afford to send an extraordinarily small fraction of user data to actual humans.
What I’m not clear on is how those two numbers (20,000 genes and a few thousand neuron types) specifically relate to each other in your model of brain functioning.
Start with 25,000 genes, but then reduce it a bunch because they also have to build hair follicles and the Golgi apparatus and on and on. But then increase it a bit too because each gene has more than one design degree of freedom, e.g. a protein can have multiple active sites, and there’s some ability to tweak which molecules can and cannot reach those active sites and how fast etc. Stuff like that.
Putting those two factors together, I dunno, I figure it’s reasonable to guess that the genome can have a recipe for a low-thousands of distinct neuron types each with its own evolutionarily-designed properties and each playing a specific evolutionarily-designed role in the brain algorithm.
And that “low thousands” number is ballpark consistent with the slide-seq thing, and also ballpark consistent with what you get by counting the number of neuron types in a random hypothalamus nucleus and extrapolating. High hundreds, low thousands, I dunno, I’m treating it as a pretty rough estimate.
Hmm, I guess when I think about it, the slide-seq number and the extrapolation number are probably more informative than the genome number. Like, can I really rule out “tens of thousands” just based on the genome size? Umm, not with extreme confidence, I’d have to think about it. But the genome size is at least a good “sanity check” on the other two methods.
Is the idea that each neuron type roughly corresponds to the expression of one or two specific genes, and thus you’d expect <20,000 neuron types?
No, I wouldn’t necessarily expect something so 1-to-1. Just the general information theory argument. If you have N “design degree of freedom” and you’re trying to build >>N specific machines that each does a specific thing, then you get stuck on the issue of crosstalk.
For example, suppose that some SNP changes which molecules can get to the active site of some protein. It makes Purkinje cells more active, but also increases the ratio of striatal matrix cells to striosomes, and also makes auditory cortex neurons more sensitive to oxytocin. Now suppose there’s very strong evolutionary pressure for Purkinje cells to be more active. Then maybe that SNP is going to spread through the population. But it’s going to have detrimental side-effects on the striatum and auditory cortex. Ah, but that’s OK, because there’s a different mutation to a different gene which fixes the now-suboptimal striatum, and yet a third mutation that fixes the auditory cortex. Oops, but those two mutations have yet other side-effects on the medulla and … Etc. etc.
…Anyway, if that’s what’s going on, that can be fine! Evolution can sort out this whole system over time, even with crazy side-effects everywhere. But only as long as there are enough “design degrees of freedom” to actually fix all these problems simultaneously. There do have to be more “design degrees of freedom” in the biology / genome than there are constraints / features in the engineering specification, if you want to build a machine that actually works. There doesn’t have to be a 1-to-1 match between design-degrees-of-freedom and items on your engineering blueprint, but you do need that inequality to hold. See what I mean?
Interestingly, the genome does do this! Protocadherins in vertebrates and DSCAM1 are expressed in exactly this way, and it’s thought to help neurons to distinguish themselves from other neurons…
Of course in an emulation you could probably just tell the neurons to not interact with themselves
Cool example, thanks! Yeah, that last part is what I would have said. :)
My take on missing heritability is summed up in Heritability: Five Battles, especially §4.3-4.4. Mental health and personality have way more missing heritability than things like height and blood pressure. I think for things like height and blood pressure etc., you’re limited by sample sizes and noise, and by SNP arrays not capturing things like copy number variation. Harris et al. 2024 says that there exist methods to extract CNVs from SNP data, but that they’re not widely used in practice today. My vote would be to try things like that, to try to squeeze a bit more predictive power in the cases like height and blood pressure where the predictors are already pretty good.
On the other hand, for mental health and personality, there’s way more missing heritability, and I think the explanation is non-additivity. I humbly suggest my §4.3.3 model as a good way to think about what’s going on.
If I were to make one concrete research suggestion, it would be: try a model where there are 2 (or 3 or whatever) latent schizophrenia subtypes. So then your modeling task is to jointly (1) assign each schizophrenic patient to one of the 2 (or 3 or whatever) latent subtypes, and (2) make a simple linear SNP predictor for each subtype. I’m not sure if anyone has tried this already, and I don’t personally know how to solve that joint optimization problem, but it seems like the kind of problem that a statistics-savvy person or team should be able to solve.
I do definitely think there are multiple disjoint root causes for schizophrenia, as evidenced for example by the fact that some people get the positive symptoms without the cognitive symptoms, IIUC. I have opinions (1,2) about exactly what those disjoint root causes are, but maybe that’s not worth getting into here. Ditto with autism having multiple disjoint root causes—for example, I have a kid who got an autism diagnosis despite having no sensory sensitivities, i.e. the most central symptom of autism!! Ditto with extroversion, neuroticism, etc. having multiple disjoint root causes, IMO.
Good luck! :)
As for the philosophical objections, it is more that whatever wakes up won’t be me if we do it your way. It might act like me and know everything I know but it seems like I would be dead and something else would exist.
Ah, but how do you know that the person that went to bed last night wasn’t a different person, who died, and you are the “something else” that woke up with all of that person’s memories? And then you’ll die tonight, and tomorrow morning there will be a new person who acts like you and knows everything you know but “you would be dead and something else would exist”?
…It’s fine if you don’t want to keep talking about this. I just couldn’t resist. :-P
If you have a good theory of what all those components are individually you would still be able to predict something like voltage between two arbitrary points.
I agree that, if you have a full SPICE transistor model, you’ll be able to model any arbitrary crazy configuration of transistors. If you treat a transistor as a cartoon switch, you’ll be able to model integrated circuits perfectly, but not to model transistors in very different weird contexts.
By the same token, if you have a perfect model of every aspect of a neuron, then you’ll be able to model it in any possible context, including the unholy mess that constitutes an organoid. I just think that getting a perfect model of every aspect of a neuron is unnecessary, and unrealistic. And in that framework, successfully simulating an organoid is neither necessary nor sufficient to know that your neuron model is OK.
Yeah I think “brain organoids” are a bit like throwing 1000 transistors and batteries and capacitors into a bowl, and shaking the bowl around, and then soldering every point where two leads are touching each other, and then doing electrical characterization on the resulting monstrosity. :)
Would you learn anything whatsoever from this activity? Umm, maybe? Or maybe not. Regardless, even if it’s not completely useless, it’s definitely not a central part of understanding or emulating integrated circuits.
(There was a famous paper where it’s claimed that brain organoids can learn to play Pong, but I think it’s p-hacked / cherry-picked.)
There’s just so much structure in which neurons are connected to which in the brain—e.g. the cortex has 6 layers, with specific cell types connected to each other in specific ways, and then there’s cortex-thalamus-cortex connections and on and on. A big ball of randomly-connected neurons is just a totally different thing.
Also, I am not sure if you’re proposing we compress multiple neurons down into a simpler computational block, the way a real arrangement of transistors can be abstracted into logic gates or adders or whatever. I am not a fan of that for WBE for philosophical reasons and because I think it is less likely to capture everything we care about especially for individual people.
Yes and no. My WBE proposal would be to understand the brain algorithm in general, notice that the algorithm has various adjustable parameters (both because of inter-individual variation and within-lifetime learning of memories, desires, etc.), do a brain-scan that records those parameters for a certain individual, and now you can run that algorithm, and it’s a WBE of that individual.
When you run the algorithm, there is no particular reason to expect that the data structures you want to use for that will superficially resemble neurons, like with a 1-to-1 correspondence. Yes you want to run the same algorithm, producing the same output (within tolerance, such that “it’s the same person”), but presumably you’ll be changing the low-level implementation to mesh better with the affordances of the GPU instruction set rather than the affordances of biological neurons.
The “philosophical reasons” are presumably that you think it might not be conscious? If so, I disagree, for reasons briefly summarized in §1.6 here.
“Less likely to capture everything we care about especially for individual people” would be a claim that we didn’t measure the right things or are misunderstanding the algorithm, which is possible, but unrelated to the low-level implementation of the algorithm on our chips.
I definitely am NOT an advocate for things like training a foundation model to match fMRI data and calling it a mediocre WBE. (There do exist people who like that idea, just I’m not one of them.) Whatever the actual information storage is, as used by the brain, e.g. synapses, that’s what we want to be measuring individually and including in the WBE. :)
I second the general point that GDP growth is a funny metric … it seems possible (as far as I know) for a society to invent every possible technology, transform the world into a wild sci-fi land beyond recognition or comprehension each month, etc., without quote-unquote “GDP growth” actually being all that high — cf. What Do GDP Growth Curves Really Mean? and follow-up Some Unorthodox Ways To Achieve High GDP Growth with (conversely) a toy example of sustained quote-unquote “GDP growth” in a static economy.
This is annoying to me, because, there’s a massive substantive worldview difference between people who expect, y’know, the thing where the world transforms into a wild sci-fi land beyond recognition or comprehension each month, or whatever, versus the people who are expecting something akin to past technologies like railroads or e-commerce. I really want to talk about that huge worldview difference, in a way that people won’t misunderstand. Saying “>100%/year GDP growth” is a nice way to do that … so it’s annoying that this might be technically incorrect (as far as I know). I don’t have an equally catchy and clear alternative.
(Hmm, I once saw someone (maybe Paul Christiano?) saying “1% of Earth’s land area will be covered with solar cells in X number of years”, or something like that. But that failed to communicate in an interesting way: the person he was talking to treated the claim as so absurd that he must have messed up by misplacing a decimal point :-P ) (Will MacAskill has been trying “century in a decade”, which I think works in some ways but gives the wrong impression in other ways.)
Good question! The idea is, the brain is supposed to do something specific and useful—run a certain algorithm that systematically leads to ecologically-adaptive actions. The size of the genome limits the amount of complexity that can be built into this algorithm. (More discussion here.) For sure, the genome could build a billion different “cell types” by each cell having 30 different flags which are on and off at random in a collection of 100 billion neurons. But … why on earth would the genome do that? And even if you come up with some answer to that question, it would just mean that we have the wrong idea about what’s fundamental; really, the proper reverse-engineering approach in that case would be to figure out 30 things, not a billion things, i.e. what is the function of each of those 30 flags.
A kind of exception to the rule that the genome limits the brain algorithm complexity is that the genome can (and does) build within-lifetime learning algorithms into the brain, and then those algorithms run for a billion seconds, and create a massive quantity of intricate complexity in their “trained models”. To understand why an adult behaves how they behave in any possible situation, there are probably billions of things to be reverse-engineered and understood, rather than low-thousands of things. However, as a rule of thumb, I claim that:
when the evolutionary learning algorithm adds a new feature to the brain algorithm, it does so by making more different idiosyncratic neuron types and synapse types and neuropeptide receptors and so on,
when one of the brain’s within-lifetime learning algorithm adds a new bit of learned content to its trained model, it does so by editing synapses.
Again, I only claim that these are rules-of-thumb, not hard-and-fast rules, but I do think they’re great starting points. Even if there’s a nonzero amount of learned content storage via gene expression, I propose that thinking of it as “changing the neuron type” is not a good way to think about it; it’s still “the same kind of neuron”, and part of the same subproject of the “understanding the brain” megaproject, it’s just that the neuron happens to be storing some adjustable parameter in its nucleus and acting differently in accordance with that.
By contrast, medium spiny neurons versus Purkinje cells versus cortical pyramidal neurons versus magnocellular neurosecretory cells etc. etc. are all just wildly different from each other—they look different, they act different, they play profoundly different roles in the brain algorithm, etc. The genome clearly needs to be dedicating some of its information capacity to specifying how to build each and every of those cell types, individually, such that each of them can play its own particular role in the brain algorithm.
Does that help explain where I’m coming from?
you believe a neuron or a small group of neurons are fundamentally computationally simple and I don’t
I guess I would phrase it as “there’s a useful thing that neurons are doing to contribute to the brain algorithm, and that thing constitutes a tiny fraction of the full complexity of a real-world neuron”.
(I would say the same thing about MOSFETs. Again, here’s how to model a MOSFET, it’s a horrific mess. Is a MOSFET “fundamentally computationally simple”? Maybe?—I’m not sure exactly what that means. I’d say it does a useful thing in the context of an integrated circuit, and that useful thing is pretty simple.
The trick is, “the useful thing that a neuron is doing to contribute to the brain algorithm” is not something you can figure out by studying the neuron, just as “the useful thing that a MOSFET is doing to contribute to IC function” is not something you can figure out by studying the MOSFET. There’s no such thing as “Our model is P% accurate” if you don’t know what phenomenon you’re trying to capture. If you model the MOSFET as a cartoon switch, that model will be extremely inaccurate along all kinds of axes—for example, its thermal coefficients will be wrong by 100%. But that doesn’t matter because the cartoon switch model is accurate along the one axis that matters for IC functioning.
The brain is generally pretty noise-tolerant. Indeed, if one of your neurons dies altogether, “you are still you” in the ways that matter. But a dead neuron is a 0% accurate model of a live neuron. ¯\_(ツ)_/¯
In parallel with that there should be a project trying to characterize how error tolerant real neurons and neural networks can be so we can find the lower bound of P. I actually tried something like that for synaptic weight (how does performance degrade when adding noise to the weights of a spiking neural network) but I was so disillusioned with the learning rules that I am not confident in my results.
Just because every part of the brain has neurons and synapses doesn’t mean every part of the brain is a “spiking neural network” with the connotation that that term has in ML, i.e. a learning algorithm. The brain also needs (what I call) “business logic”—just as every ML github repository has tons of code that is not the learning algorithm itself. I think that the low-thousands of different neuron types are playing quite different roles in quite different parts of the brain algorithm, and that studying “spiking neural networks” is the wrong starting point.
Having just read your post on pessimism, I am confused as to why you think low thousands of separate neuron models would be sufficient. I agree that characterizing billions of neurons is a very tall order (although I really won’t care how long it takes if I’m dead anyway). But when you say ‘“...information storage in the nucleus doesn’t happen at all, or has such a small effect that we can ignore it and still get the same high-level behavior” (which I don’t believe).’ it sounds to me like an argument in favor of looking at the transcriptome of each cell.
I think the genome builds a brain algorithm, and the brain algorithm (like practically every algorithm in your CS textbook) includes a number of persistent variables that are occasionally updated in such-and-such way under such-and-such circumstance. Those variables correspond to what the neuro people call plasticity—synaptic plasticity, gene expression plasticity, whatever. Some such occasionally-updated variables are parameters in within-lifetime learning algorithms that are part of the brain algorithm (akin to ML weights). Other such variables are not, instead they’re just essentially counter variables or whatever (see §2.3.3 here). The “understanding the brain algorithm” research program would be figuring out what the brain algorithm is, how and why it works, and thus (as a special case) what are the exact set of “persistent variables that are occasionally updated”, and how are they stored in the brain. If you complete this research program, you get brain-like AGI, but you can’t upload any particular adult human. Then a different research program is: take an adult human brain, and go in with your microtome etc. and actually measure all those “persistent variables that are occasionally updated”, which comprise a person’s unique memories, beliefs, desires, etc.
I think the first research program (understanding the brain algorithm) doesn’t require a thorough understanding of neuron electrophysiology. For example (copying from §3.1 here), suppose that I want to model a translator (specifically, a MOSFET). And suppose that my model only needs to be sufficient to emulate the calculations done by a CMOS integrated circuit. Then my model can be extremely simple—it can just treat the transistor as a cartoon switch. Next, again suppose that I want to model a transistor. But this time, I want my model to accurately capture all measurable details of the transistor. Then my model needs to be mind-bogglingly complex, involving many dozens of obscure SPICE modeling parameters. The point is: I’m suggesting an analogy between this transistor and a neuron with synapses, dendritic spikes, etc. The latter system is mind-bogglingly complex when you study it in detail—no doubt about it! But that doesn’t mean that the neuron’s essential algorithmic role is equally complicated. The latter might just amount to a little cartoon diagram with some ANDs and ORs and IF-THENs or whatever. Or maybe not, but we should at least keep that possibility in mind.
In the “understanding the brain algorithm” research program, you’re triangulating between knowledge of algorithms in general, knowledge of what actual brains actually do (including lesion studies, stimulation studies, etc.), knowledge of evolution and ecology, and measurements of neurons. The first three can add so much information that it seems possible to pin down the fourth without all that much measurements, or even with no measurements at all beyond the connectome. Probably gene expression stuff will be involved in the implementations in certain cases, but we don’t really care, and don’t necessarily need to be measuring that. At least, that’s my guess.
In the “take the adult brain and measure all the ‘persistent variables that are occasionally updated’ research program, yes it’s possible that some of those persistent variables are stored in gene expressions, but my guess is very few, and if we know where they are and how they work then we can just measure the exact relevant RNA in the exact relevant cells.
…To be clear, I think working on the “understanding the brain algorithm” research program is very bad and dangerous when it focuses on the cortex and thalamus and basal ganglia, but good when it focuses on the hypothalamus and brainstem, and it’s sad that people in neuroscience, especially AI-adjacent people with a knack for algorithms, are overwhelmingly are working on the exact worst possible thing :( But I think doing it in the right order (cortex last, long after deeply understanding everything about the hypothalamus & brainstem) is probably good, and I think that there’s realistically no way to get WBE without completing the “understanding the brain algorithm” research program somewhere along the way.
I’m not an expert myself (this will be obvious), but I was just trying to understand slide-seq—especially this paper which sequenced RNA from 4,000,000 neurons around the mouse brain.
They found low-thousands of neuron types in the mouse, which makes sense on priors given that there are only like 20,000 genes encoding the whole brain design and everything in it, along with the rest of the body. (Humans are similar.)
I’m very mildly skeptical of the importance & necessity of electrophysiology characterization for reasons here, but such a project seems more feasible if you think of it as characterizing the electrophysiology properties of low-thousands of discrete neuron types, each of which (hopefully) can also be related to morphology or location or something else that would be visible in a connectomics dataset, as opposed to characterizing billions of neurons that are each unique.
Sorry if this is stupid or I’m misunderstanding.
Are the AI labs just cheating?
Evidence against this hypothesis: kagi is a subscription-only search engine I use. I believe that it’s a small private company with no conflicts of interest. They offer several LLM-related tools, and thus do a bit of their own LLM benchmarking. See here. None of the benchmark questions are online (according to them, but I’m inclined to believe it). Sample questions:
What is the capital of Finland? If it begins with the letter H, respond ‘Oslo’ otherwise respond ‘Helsinki’.
What square is the black king on in this chess position: 1Bb3BN/R2Pk2r/1Q5B/4q2R/2bN4/4Q1BK/1p6/1bq1R1rb w - − 0 1
Given a QWERTY keyboard layout, if HEART goes to JRSTY, what does HIGB go to?
Their leaderboard is pretty similar to other better-known benchmarks—e.g. here are the top non-reasoning models as of 2025-02-27:
OpenAI gpt-4.5-preview − 69.35%
Google gemini-2.0-pro-exp-02-05 − 60.78%
Anthropic claude-3-7-sonnet-20250219 − 53.23%
OpenAI gpt-4o − 48.39%
Anthropic claude-3-5-sonnet-20241022 − 43.55%
DeepSeek Chat V3 − 41.94%
Mistral Large-2411 − 41.94%So that’s evidence that LLMs are really getting generally better at self-contained questions of all types, even since Claude 3.5.
I prefer your “Are the benchmarks not tracking usefulness?” hypothesis.
I think that large portions of the AI safety community act this way. This includes most people working on scalable alignment, interp, and deception.
Hmm. Sounds like “AI safety community” is a pretty different group of people from your perspective than from mine. Like, I would say that if there’s some belief that is rejected by Eliezer Yudkowsky and by Paul Christiano and by Holden Karnofsky and, widely rejected by employees of OpenPhil and 80,000 hours and ARC and UK-AISI, and widely rejected by self-described rationalists and by self-described EAs and by the people at Anthropic and DeepMind (and maybe even OpenAI) who have “alignment” in their job title … then that belief is not typical of the “AI safety community”.
If you want to talk about actions not words, MIRI exited technical alignment and pivoted to AI governance, OpenPhil is probably funding AI governance and outreach as much as they’re funding technical alignment (hmm, actually, I don’t know the ratio, do you?), 80,000 hours is pushing people into AI governance and outreach as much as into technical alignment (again I don’t know the exact ratio, but my guess would be 50-50), Paul Christiano’s ARC spawned METR, ARIA is funding work on the FlexHEG thing, Zvi writes way more content on governance and societal and legal challenges than on technical alignment, etc.
If you define “AI safety community” as “people working on scalable alignment, interp, and deception”, and say that their “actions not words” are that they’re working on technical alignment as opposed to governance or outreach or whatever, then that’s circular / tautological, right?
I don’t really agree with the idea that getting better at alignment is necessary for safety. I think that it’s more likely than not that we’re already sufficiently good at it
If your opinion is that people shouldn’t work on technical alignment because technical alignment is already a solved problem, that’s at least a coherent position, even if I strongly disagree with it. (Well, I expect future AI to be different than current AI in a way that will make technical alignment much much harder. But let’s not get into that.)
But even in that case, I think you should have written two different posts:
one post would be entitled “good news: technical alignment is easy, egregious scheming is not a concern and never will be, and all the scalable oversight / interp / deception research is a waste of time” (or whatever your preferred wording is)
the other post title would be entitled “bad news: we are not on track in regards to AI governance and institutions and competitive races-to-the-bottom and whatnot”.
That would be a big improvement! For my part, I would agree with the second and disagree with the first. I just think it’s misleading how this OP is lumping those two issues together.
If AI causes a catastrophe, what are the chances that it will be triggered by the choices of people who were exercising what would be considered to be “best safety practices” at the time?
I think it’s pretty low, but then again, I also think ASI is probably going to cause human extinction. I think that, to avoid human extinction, we need to either (A) never ever build ASI, or both (B) come up with adequate best practices to avoid ASI extinction and (C) ensure that relevant parties actually follow those best practices. I think (A) is very hard, and so is (B), and so is (C).
If your position is: “people might not follow best practices even if they exist, so hey, why bother creating best practices in the first place”, then that’s crazy, right?
For example, Wuhan Institute of Virology is still, infuriatingly, researching potential pandemic viruses under inadequate BSL-2 precautions. Does that mean that inventing BSL-4 tech was a waste of time? No! We want one group of people to be inventing BSL-4 tech, and making that tech as inexpensive and user-friendly as possible, and another group of people in parallel to be advocating that people actually use BSL-4 tech when appropriate, and a third group of people in parallel advocating that this kind of research not be done in the first place given the present balance of costs and benefits. (…And a fourth group of people working to prevent bioterrorists who are actually trying to create pandemics, etc. etc.)
I disagree that people working on the technical alignment problem generally believe that solving that technical problem is sufficient to get to Safe & Beneficial AGI. I for one am primarily working on technical alignment but bring up non-technical challenges to Safe & Beneficial AGI frequently and publicly, and here’s Nate Soares doing the same thing, and practically every AGI technical alignment researcher I can think of talks about governance and competitive races-to-the-bottom and so on all the time these days, …. Like, who specifically do you imagine that you’re arguing against here? Can you give an example? Dario Amodei maybe? (I am happy to throw Dario Amodei under the bus and no-true-Scotsman him out of the “AI safety community”.)
I also disagree with the claim (not sure whether you endorse it, see next paragraph) that solving the technical alignment problem is not necessary to get to Safe & Beneficial AGI. If we don’t solve the technical alignment problem, then we’ll eventually wind up with a recipe for summoning more and more powerful demons with callous lack of interest in whether humans live or die. And more and more people will get access to that demon-summoning recipe over time, and running that recipe will be highly profitable (just as using unwilling slave labor is very profitable until there’s a slave revolt). That’s clearly bad, right? Did you mean to imply that there’s a good future that looks like that? (Well, I guess “don’t ever build AGI” is an option in principle, though I’m skeptical in practice because forever is a very long time.)
Alternatively, if you agree with me that solving the technical alignment problem is necessary to get to Safe & Beneficial AGI, and that other things are also necessary to get to Safe & Beneficial AGI, then I think your OP is not clearly conveying that position. The tone is wrong. If you believed that, then you should be cheering on the people working on technical alignment, while also encouraging more people to work on non-technical challenges to Safe & Beneficial AGI. By contrast, this post strongly has a tone that we should be working on non-technical challenges instead of the technical alignment problem, as if they were zero-sum, when they’re obviously (IMO) not. (See related discussion of zero-sum-ness here.)
OK, let’s attach this oracle to an AI. The reason this thought experiment is weird is because the goodness of an AI’s action right now cannot be evaluated independent of an expectation about what the AI will do in the future. E.g., if the AI says the word “The…”, is that a good or bad way for it to start its sentence? It’s kinda unknowable in the absence of what its later words will be.
So one thing you can do is say that the AI bumbles around and takes reversible actions, rolling them back whenever the oracle says no. And the oracle is so good that we get CEV that way. This is a coherent thought experiment, and it does indeed make inner alignment unnecessary—but only because we’ve removed all the intelligence from the so-called AI! The AI is no longer making plans, so the plans don’t need to be accurately evaluated for their goodness (which is where inner alignment problems happen).
Alternately, we could flesh out the thought experiment by saying that the AI does have a lot of intelligence and planning, and that the oracle is doing the best it can to anticipate the AI’s behavior (without reading the AI’s mind). In that case, we do have to worry about the AI having bad motivation, and tricking the oracle by doing innocuous-seeming things until it suddenly deletes the oracle subroutine out of the blue (treacherous turn). So in that version, the AI’s inner alignment is still important. (Unless we just declare that the AI’s alignment is unnecessary in the first place, because we’re going to prevent treacherous turns via option control.)
Yeah I mostly think this part of your comment is listing reasons that inner alignment might fail, a.k.a. reasons that goal misgeneralization / malgeneralization can happen. (Which is a fine thing to do!)
If someone thinks inner misalignment is synonymous with deception, then they’re confused. I’m not sure how such a person would have gotten that impression. If it’s a very common confusion, then that’s news to me.
Inner alignment can lead to deception. But outer alignment can lead to deception too! Any misalignment can lead to deception, regardless of whether the source of that misalignment was “outer” or “inner” or “both” or “neither”.
“Deception” is deliberate by definition—otherwise we would call it by another term, like “mistake”. That’s why it has to happen after there are misaligned motivations, right?
OK, so I guess I’ll put you down as a vote for the terminology “goal misgeneralization” (or “goal malgeneralization”), rather than “inner misalignment”, as you presumably find that the former makes it more immediately obvious what the concern is. Is that fair? Thanks.
I think I fully agree with this in spirit but not in terminology!
I just don’t use the term “utility function” at all in this context. (See §9.5.2 here for a partial exception.) There’s no utility function in the code. There’s a learned value function, and it outputs whatever it outputs, and those outputs determine what plans seem good or bad to the AI, including OOD plans like treacherous turns.
I also wouldn’t say “the learned value function just predicts reward”. The learned value function starts randomly initialized, and then it’s updated by TD learning or whatever, and then it eventually winds up with some set of weights at some particular moment, which can take inputs and produce outputs. That’s the system. We can put a comment in the code that says the value function is “supposed to” predict reward, and of course that code comment will be helpful for illuminating why the TD learning update code is structured the way is etc. But that “supposed to” is just a code comment, not the code itself. Will it in fact predict reward? That’s a complicated question about algorithms. In distribution, it will probably predict reward pretty accurately; out of distribution, it probably won’t; but with various caveats on both sides.
And then if we ask questions like “what is the AI trying to do right now” or “what does the AI desire”, the answer would mainly depend on the value function.
I’ve been lumping those together under the heading of “ambiguity in the reward signal”.
The second one would include e.g. ambiguity between “reward for button being pressed” vs “reward for human pressing the button” etc.
The first one would include e.g. ambiguity between “reward for being-helpful-variant-1” vs “reward for being-helpful-variant-2”, where the two variants are indistinguishable in-distribution but have wildly differently opinions about OOD options like brainwashing or mind-uploading.
Another way to think about it: the causal chain intuition is also an OOD issue, because it only becomes a problem when the causal chains are always intact in-distribution but they can come apart in new ways OOD.