I was claiming that titotal’s post doesn’t appear to give arguments that directly address whether or not Yudkowsky-style ASI can invent diamondoid nanotech. I don’t understand the relevance to my comment. I agree that if you find titotal’s argument persuasive then whether it is load bearing is relevant to AI risk concerns, but that’s not what my comment is about.
FWIW Yudkowsky frequently says that this is not load bearing, and that much seems obviously true to me also.
Both this and his earlier article by titotal on computational chemistry make the argument that development of molecular nanotechnology can’t be a one-shot process because of intrinsic limitations on simulation capabilities.
Yudkowsky claims that one-shot nanotechnology is not load bearing, yet literally every example he gives when pressed involves one-shot nanotech.
I’m running a molecular nanotechnology company, so I’ve had to get quite familiar with the inner workings and limitations of existing computational chemistry packages. This article does a reasonable job of touching on the major issues. I can tell you that experimentally these codes don’t even come close to the precision required to find trajectories that reliably drive diamondoid mechanosynthesis. It’s really difficult to get experiments in the lab to match the (over simplified, reduced accuracy) simulations which take days or weeks to run.
Having any hope of adequately simulating mechanosynthetic reactions requires codes that scale with O(n^7). No one even writes these codes, because what would be the point? The biggest thing you could simulate is helium. There are approximations with O(n^3) and O(n^4) running time that are sometimes useful guides, but can often have error bars larger than the relevant energy gaps. The acid test here is still experiment: try to do the thing and see what happens.
AI has potential to greatly improve productivity with respect to nano-mechanical engineering designs because you can build parts rigid and large enough, and operating at slow enough speeds that mechanical forces are unlikely to have chemical effects (bond breaking or forming), and therefore faster O(n^2) molecular mechanics codes can be used. But all the intelligence in the world isn’t going to make bootstrapping molecular nanotechnology a one-shot process. It’s going to be years of hard-won lab work, whether it’s done by a human or a thinking machine.
Well yes, nobody thinks that existing techniques suffice to build de-novo self-replicating nano machines, but that means it’s not very informative to comment on the fallibility of this or that package or the time complexity of some currently known best approach without grounding in the necessity of that approach.
One has to argue instead based on the fundamental underlying shape of the problem, and saying accurate simulation is O(n⁷) is not particularly more informative to that than saying accurate protein folding is NP. I think if the claim is that you can’t make directionally informative predictions via simulation for things meaningfully larger than helium then one is taking the argument beyond where it can be validly applied. If the claim is not that, it would be good to hear it clearly stated.
I don’t understand your objection. People who actually use computational chemistry—like me, like titotal—are familiar with its warts and limitations. These limitations are intrinsic to the problem and not expressions of our ignorance. Diamondoid mechanosynthetic reactions depend on properties which only show up in higher levels of theory, the ones which are too expensive to run for any reasonably sized system. To believe that this limitation won’t apply to AI, just because, is magical thinking.
I wasn’t saying anything with respect to de-novo self-replicating nano machines, a field which has barely been studied by anyone and which we cannot adequately say much about at all.
And what reason do you have for thinking it can’t be usefully approximated in some sufficiently productive domain, that wouldn’t also invalidly apply to protein folding? I think it’s not useful to just restate that there exist reasons you know of, I’m aiming to actually elicit those arguments here.
The fact that this has been an extremely active area of research for over 80 years with massive real-world implications, and we’re no closer to finding such a simplified heuristic. This isn’t like protein folding at all—the math is intrinsically more complicated with shared interaction terms that require high polynomial count running times. Indeed it’s provable that you can’t have a faster algorithm than those O(n^3) and O(n^4) approximations which cover all relevant edge cases accurately.
Of course there could be some heuristic-like approximation which works in the relevant cases for this specific domain, e.g. gemstone crystal lattice structures under ultra-high vacuum conditions and cryogenic temperatures. Just like there are reasonably accurate molecular mechanics models that have been developed for various specific materials science use cases.
But here’s the critical point: these heuristics are developed by evolving models based on experiment. In other domains like protein folding you can have a source of truth for generative AI training that uses simulation. That’s how AlphaFold was done. But for chemical reactions of the sort we’d be interested here the ab-initio methods that give correct results are simply too slow to use for this. Not “buy a bigger GPU” slow, but “run the biggest supercomputer for a quintillion years” slow. So the solution is to do the process manually, running actual experiments to refine the heuristic models at each step, much machine learning training models would do. These are called semi-empirical methods, and they work. But we don’t have a good semi-empirical model for the type of reactions concerned here, and it would take years of painstaking lab work to develop, which no amount of super intelligence can avoid.
It’s magical thinking to assume that an AI will just one-shot this into existence.
Thanks, I appreciate the attempt to clarify. I do though think there’s some fundamental disagreement about what we’re arguing over here that’s making it less productive than it could be. For example,
The fact that this has been an extremely active area of research for over 80 years with massive real-world implications, and we’re no closer to finding such a simplified heuristic.
I think both:
Lack of human progress doesn’t necessarily mean the problem is intrinsically unsolvable by advanced AI. Humans often take a bunch of time before proving things.
It seems not at all the case that algorithmic progress isn’t happening, so it’s hardly a given that we’re no closer to a solution unless you first circularly assume that there’s no solution to arrive at.
If you’re starting out with an argument that we’re not there yet, this makes me think more that there’s some fundamental disagreement about how we should reason about ASI, more than your belief being backed by a justification that would be convincing to me had only I succeeded at eliciting it. Claiming that a thing is hard is at most a reason not to rule out that it’s impossible. It’s not a reason on its own to believe that it is impossible.
With regard to complexity,
I failed to understand the specific difference with protein folding. Protein folding is NP-hard, which is significantly harder than O(n³).
I failed to find the source for the claim that O(n³) or O(n⁴) are optimal. Actually I’m pretty confused how this is even a likely concept; surely if the O(n³) algorithm is widely useful then the O(n⁴) proof can’t be that strong of a bound on practical usefulness? So why is this not true of the O(n³) proof as well?
It’s maybe true that protein folding is easier to computationally verify solutions to, but first, can you prove this, and second, on what basis are you claiming that existing knowledge is necessarily insufficient to develop better heuristics than the ones we already have? The claim doesn’t seem to complete to me.
It’s magical thinking to assume that an AI will just one-shot this into existence.
Please note that I’ve not been making the claim that ASI could necessarily solve this problem. I have been making the claim that the arguments in this post don’t usefully support the claim that it can’t. It is true that largely on priors I expect it should be able to, but my priors also aren’t particularly useful ones to this debate and I have tried to avoid making comments that are dependent on them.
Do you have a maths or computer science background? I ask because some of the questions you ask are typical of maths or CS, but don’t really make sense in this context.
Take protein folding. Existing techniques for estimating how a protein folds are exponential in complexity (making the problem NP-hard), so running time is bounded by O(2^n). But what’s the ‘n’? In these algorithms it is the number of amino acids because that’s the level at which they are operating. For quantum chemistry the polynomial running times have ‘n’ being the number of electrons in the system, which can be orders of magnitude higher. That’s why O(n^4) can be much, much worse than O(2^n) for relevant problem sizes—there’s no hope of doing full protein folding simulations by ab initio methods. The second reason they’re not comparable is time steps. Ab initio molecular mechanics methods can’t have time steps larger than 1 femtosecond. Proteins fold on the timescale of milliseconds. This by itself is what makes all ab initio molecular mechanics methods infeasible for protein folding, even though it’s a constant factor ignored by big-O notation.
Staying on the issue of complexity, from a technical / mathematical foundation perspective every quantum chemistry code should also have exponential running time, and this is trivially proved by the fact that every particle technically does interact with every other particle. But all of these interaction terms fall off with at least 1/r^2, and negative power increases with each corrective term. For 1/r^2 and 1/r^3, there are field density tricks that can be used for accurate sub-exponential modeling. For 1/r^4 or higher it is perfectly acceptable to just have distance cutoffs which cutout the exponential term. How accurate a simulation you will get is a question of how many of these corrective terms you include. The more you include, the more higher-order effects you capture, but you also get hit with higher-order polynomial running times.
To bring a long story to an abrupt end, you don’t “prove” correctness of these various codes in a mathematical sense. They’re all physically incorrect because they involve approximations. Rather for various use cases you can experimentally show that certain terms dominate, and you need to make sure that the code you use accurately captures the relevant terms for the domain and scale of the problem you are simulating. This is an experimental determination, not a mathematical proof. You literally go measure things in the lab.
The protein data bank contains hundreds of thousands of structures which have had their atomic coordinates determined by various experimental methods, providing a ground truth for protein folding heuristics. There is no such source of truth for diamondoid nano-mechanical machines parts and the mechanosynthesis steps that produce them. If you want to train a better reactive force field heuristic that could make the relevant design simulations tractable, you need a source of truth. The only thing available is various quantum chemistry approximations, as you need to approximate to be computable. These fail as source of truth because you don’t know which approximations are reasonable and which are not without experiment data, which we don’t have. So you could train a heuristic, yes, but you’d have absolutely no way of assessing its accuracy.
There’s a common saying in the AI field: “garbage in, garbage out.” The current state of non-empirical ab initio quantum chemistry is garbage. You’re saying that AI could take the garbage and by mere application of thought turn it into something useful. That’s not in line with the actual history of the development of useful AI outputs.
I’m not saying that AI can’t develop useful heuristic approximations for the simulation of gemstone-based nano-mechanical machinery operating in ultra-high vacuum. I’m saying that it can’t do so as a one-shot inference without any new experimental work, which is a trope that shows up again and again in certain high-profile people’s descriptions of AI x-risk scenarios.
If you say “Indeed it’s provable that you can’t have a faster algorithm than those O(n^3) and O(n^4) approximations which cover all relevant edge cases accurately” I am quite likely to go on a digression where I try to figure out what proof you’re pointing at and why you think it’s a fundamental barrier, and it seems now that per a couple of your comments you don’t believe it’s a fundamental barrier, but at the same time it doesn’t feel like any position has been moved, so I’m left rather foggy about where progress has been made.
I think it’s very useful that you say
I’m not saying that AI can’t develop useful heuristic approximations for the simulation of gemstone-based nano-mechanical machinery operating in ultra-high vacuum. I’m saying that it can’t do so as a one-shot inference without any new experimental work
since this seems like a narrower place to scope our conversation. I read this to mean:
You don’t know of any in principle barrier to solving this problem,
You believe the solution is underconstrained by available evidence.
I find the second point hard to believe, and don’t really see anywhere you have evidenced it.
As a maybe-relevant aside to that, wrt.
You’re saying that AI could take the garbage and by mere application of thought turn it into something useful. That’s not in line with the actual history of the development of useful AI outputs.
I think you’re talking of ‘mere application of thought’ like it’s not the distinguishing feature humanity has. I don’t care what’s ‘in line with the actual history’ of AI, I care what a literal superintelligence could do, and this includes a bunch of possibilities like:
Making inhumanly close observation of all existing data,
Noticing new, inhumanly-complex regularities in said data,
Proving new simplifying regularities from theory,
Inventing new algorithms for heuristic simulation,
Finding restricted domains where easier regularities hold,
Bifurcating problem space and operating over each plausible set,
Sending an interesting email to a research lab to get choice high-ROI data.
We can ignore the last one for this conversation. I still don’t understand why the others are deemed unreasonable ways of making progress on this task.
I appreciated the comments on time complexity but am skipping it because I don’t expect at this point that it lies at the crux.
If you say “Indeed it’s provable that you can’t have a faster algorithm than those O(n^3) and O(n^4) approximations which cover all relevant edge cases accurately” I am quite likely to go on a digression where I try to figure out what proof you’re pointing at and why you think it’s a fundamental barrier
By “proof” I meant proof by contradiction. DFT is a great O(n^3) method for energy minimizing structures and exploring electron band structure, and it is routinely used for exactly that purpose. So much so that many people conflate “DFT” with more accurate ab initio methods, which it is not. However DFT utterly ignores exchange correlation terms and so it doesn’t model van der Waals interactions at all. Every design for efficient and performant molecular nanotechnology—the ones that get you order-of-magnitude performance increases and therefore any benefit over existing biology or materials science—involve vdW forces almost exclusively in their manufacture and operation. It’s the dominant non-covalent bonded interaction at that scale.
That’s the most obvious example, but also a lot of the simulations performed by Merkle and Freitas in their minimal toolset paper give incorrect reaction sequences in these lower levels of theory, as they found out when they got money to attempt it in the lab. Without pointing to their specific failure, you can get a hint of this in surface science. Silicon, gold, and other surfaces tend to have rather interesting surface clustering and reorganization effects, which are observable by scanning probe microscopy. These are NOT predicted by the cheaper / computationally tractable codes, and they are an emergent property of higher-order exchange correlations in the crystal structure. These nevertheless have enough of an effect to drastically reshape the surface of these materials, making calculation of those forces absolutely required for any attempt to build off the surface.
Attempting to do cheaper simulations for diamondoid synthesis reactions gave very precise predictions that didn’t work as expected in the lab. How would your superintelligent AI know that uncalculated terms dominate in the simulation, and make corrective factors without having access to those incomputable factors?
Making inhumanly close observation of all existing data
Noticing new, inhumanly-complex regularities in said data,
Proving new simplifying regularities from theory
I think you vastly overestimate how much knowledge is left to extracted from the data. AI has made tremendous advances in recent years where it has been able to consume huge amounts of data, far in excess of what any group of humans could analyze. This, on the other hand, is a data-poor regime.
Inventing new algorithms for heuristic simulation
This is happening right now. There are a variety of machine-learned molecular mechanics force fields that have been published in the last few years. The most interesting one I’ve found used periodic crystal ab initio simulation methods to create a force field potential that ended up being very good for liquid and gas-phase chemistry, which it was not trained on.
But the relevant question (if you want to talk about AI x-risk by means of bootstrapping nanotech) is how accurate they are outside of the domain where we have heaps of hard evidence, because we don’t have a ground truth to compare against in those environments.
Finding restricted domains where easier regularities hold
Bifurcating problem space and operating over each plausible set,
Human engineers are very good at this. It’s not the limiting factor.
Sending an interesting email to a research lab to get choice high-ROI data
What lab. There’s literally no one doing the relevant research, or equipped to easily do it without years of preparatory chemical synthesis and surface characterization.
Which is really the point and the crux of the matter. It will take an extended, years-long research effort to create molecular nanotechnology. It’s not something you can plausibly do in secret, and certainly not something you can shorten by simulation or bayesian inference.
One quick intuition pump: do you think a team of 10,000 of the smartest human engineers and scientists could do this if they had perfect photographic memory, were immortal, and could think for a billion years?
To keep the situation analogous to an AI needing to do this quickly, we’ll suppose this team of humans is subject to the same restrictions on compute for simulations (e.g. over the entire billion years they only get 10^30 FLOP) and also can only run whatever experiments the AI would get to run (effectively no interesting experiments).
I feel uncertain about whether this team would succeed, but it does seem pretty plausible that they would be able to succeed. Perhaps I think they’re 40% likely to succeed?
Now, suppose the superintelligence is like this, but even more capable.
Separately, I don’t think it’s very important to know what an extremely powerful superintelligence could do, because prior to the point where you have an extremely powerful superintelligence, humanity will already be obsoleted by weaker AIs. So, I think Yudkowsky’s arguments about nanotech are mostly unimportant for other reasons.
But, if you did think “well, sure the AI might be arbitrarily smart, but if we don’t give it access to the nukes what can it even do to us?” then I think that there are many sources of concern and nanotech is certainly one of them.
One quick intuition pump: do you think a team of 10,000 of the smartest human engineers and scientists could do this if they had perfect photographic memory, were immortal, and could think for a billion years?
By merely thinking about it, and not running any experiments? No, absolutely not. I don’t you understood my post if you assume I’d think otherwise.
Try this: I’m holding a certain number of fingers behind my back. You and a team of 10,000 of the smartest human engineers and scientists have a billion years to decide, without looking, what your guess will be as to how many fingers I’m holding behind my back. But you only get one chance to guess at the end of that billion years.
Please don’t use That Alien Message as an intuition pump. There’s a tremendous amount wrong with the sci-fi story. Not least of which is that it completely violates the same constraint you put into your own post about constraining computation. I suggest doing your own analysis of how many thought-seconds the AI would have in-between frames of video, especially if you assume it to be running as a large inference model.
The best thing you can do is rid yourself of the notion that superhuman AI would have arbitrary capabilities. That is where EY went wrong, and a lot of the LW crowd too. If you permit dividing by zero or multiplying by infinity, then you can easily convince yourself of anything. AI isn’t magic, and AGI isn’t a free pass to believe anything.
PS: We’ve had AGI since 2017. That’d better be compatible with your world view if you want accurate predictions.
I don’t understand where your confidence is coming from here, but fair enough. It wasn’t clear to me if your take was more like “wildly, wildly superintelligent AI will be considerably weaker than a team of humans thinking for a billion years” or more like “literally impossible without either experiments or >>10^30 FLOP”.
I generally have an intuition like “it’s really, really hard to rule out physically possible things out without very strong evidence, by default things have a reasonable chance of being possible (e.g. 50%) when sufficient intelligence is applied if they are physically possible”. It seems you don’t share this intuition, fair enough.
(I feel like this applies for nearly all human inventions? Like if you had access to a huge amount of video of the world from 1900 and all written books that existed at this point, and had the affordances I described with a team of 10,000 people, 10^30 FLOP, and a billion years, it seems to me like there is a good chance you’d be able to one-shot reinvent ~all inventions of modern humanity (not doing everything in the same way, in many cases you’d massively over engineer to handle one-shot). Planes seem pretty easy? Rockets seem doable?)
Please don’t use That Alien Message as an intuition pump.
I think this is an ok, but not amazing intuition pump for what wildly, wildly superintelligent AI could be like.
The best thing you can do is rid yourself of the notion that superhuman AI would have arbitrary capabilities
I separately think it’s not very important to think about the abilities of wildly, wildly superintelligent AI for most purposes (as I noted in my comment). So I agree that imagining arbitrary capabilities is probablematic. (For some evidence that this isn’t post-hoc justification, see this post on which I’m an author.)
PS: We’ve had AGI since 2017. That’d better be compatible with your world view if you want accurate predictions.
Uhhhh, I’m not sure I agree with this as it doesn’t seem like nearly all jobs are easily fully automatable by AI. Perhaps you use a definition of AGI which is much weaker like “able to speak slightly coherant english (GPT-1?) and classify images”?
By the way, where’s this number coming from? You keep repeating it. That amount of calculation is equivalent to running the largest supercomputer in existence for 30k years. You hypothetical AI scheming breakout is not going to have access to that much compute. Be reasonable.
I generally have an intuition like “it’s really, really hard to rule out physically possible things out without very strong evidence, by default things have a reasonable chance of being possible (e.g. 50%) when sufficient intelligence is applied if they are physically possible”
Ok let’s try a different tract. You want to come up with a molecular mechanics model that can efficiently predict the outcome of reactions, so that you can get about designing one-shot nanotechnology bootstrapping. What would success look like?
You can’t actually do a full simulation to get ground truth for training a better molecular mechanics model. So how would you know the model you came up will work as intended? You can back-test against published results in the literature, but surprise surprise, a big chunk of scientific papers don’t replicate. Shoddy lab technique, publication pressure, and a niche domain combine to create conditions where papers are rushed and sometimes not 100% truthful. Even without deliberate fraud (which also happens), you run into problems such as synthesis steps not working as advertised, images used from different experimental runs than the one described in the paper, etc.
Except you don’t know that. You’re not allowed to do experiments! Maybe you guess that replication will be an issue, although why that would be a hypothesis in the first place without first seeing failures in the lab first isn’t clear to me. But let’s say you do. Which experiments should you discount? Which should you assume to be correct? If you’re allowed to start choosing which reported results you believe and which you don’t, you’ve lost the plot. There could be millions of possible heuristics which partially match the literature and there’s no way to tell the difference.
So what would success look like? How would you know you have the right molecular mechanics model that gives accurate predictions?
You can’t. Not any more than Descartes could think his way to absolute truth.
Also for what it’s worth you’ve made a couple of logical errors here. You are considering human inventions which already exist, then saying that you could one-shot invent them in 1900. That’s hindsight bias, but also selection bias. Nanotechnology doesn’t exist. Even if it would work if created, there’s no existence proof that there exists an accessible path to achieving it. Like superheavy atoms in the island of stability, or micro black holes, there just might not be a pathway to make them from present day capabilities. (Obviously I don’t believe this as I’m running a company attempting to bootstrap Drexlarian nanotechnology, but I feel it’s essential to point out the logical error.)
(Re: Alien Message) I think this is an ok, but not amazing intuition pump for what wildly, wildly superintelligent AI could be like.
Why? You’ve gone into circular logic here.
I pointed out that the Alien Message story makes fundamental errors with respect to computational capability being wildly out of scale, so actual super intelligent AIs aren’t going to be anything like the one in the story.
Maybe a Jupiter-sized matryoshka brain made of computronium would exhibit this level of super intelligence. I’m not saying it’s not physically possible. But in terms of sketching out and bounding the capabilities of near-term AI/ASI, it’s a fucking terrible intuition pump.
Uhhhh, I’m not sure I agree with this as it doesn’t seem like nearly all jobs are easily fully automatable by AI. Perhaps you use a definition of AGI which is much weaker like “able to speak slightly coherant english (GPT-1?) and classify images”?
The transformer architecture introduced in 2017 is:
Artificial: man-made
General: able to train over arbitrary unstructured input, from which it infers models that can be applied in arbitrarily ways to find solutions of problems drawn from domains outside of its training data.
Intelligent: able to construct efficient solutions to new problems it hasn’t seen.
Artificial General Intelligence. A.G.I.
If you’re thinking “yeah, but..” then I suggest you taboo the term AGI. This is literally all that it the word means.
If you want to quibble over dates then maybe we can agree on 2022 with the introduction of ChatGPT, a truly universal (AKA general) interface to mature transformer technology. Either way we’re already well within the era of artificial general intelligence.
(Maybe EY’s very public meltdown a year ago is making more sense now? But rest easy, EY’s predictions about AI x-risk have consistently been wildly off the mark.)
You highlighted “disagree” on the part about AGI’s definition. I don’t know how to respond to that directly, so I’ll do so here. Here’s the story about how the term “AGI” was coined, by the guy who published the literal book on AGI and ran the AGI conference series for the past two decades:
LW seems to have adopted some other vague, ill-defined, threatening meaning for the acronym “AGI” that is never specified. My assumption is that when people say AGI here they mean Bostrom’s ASI, and they got linked because Eliezer believed (and believes still?) that AGI will FOOM into ASI almost immediately, which it has not.
Anyway it irks me that the term has been coopted here. AGI is a term of art in the pre-ML era of AI research with a clearly defined meaning.
My assumption is that when people say AGI here they mean Bostrom’s ASI, and they got linked because Eliezer believed (and believes still?) that AGI will FOOM into ASI almost immediately, which it has not.
In case this wasn’t clear from early discussion, I disagree with Eliezer on a number of topics, including takeoff speeds. In particular I disagree about the time from AI that is economically transformative to AI that is much, much more powerful.
I think you’ll probably find it healthier and more productive to not think of LW as an amorphous collective and instead note that there are a variety of different people who post on the forum with a variety of different views. (I sometimes have made this mistake in the past and I find it healthy to clarify at least internally.)
E.g. instead of saying “LW has bad views about X” say “a high fraction of people who comment on LW have bad views about X” or “a high fraction of karma votes seem to be from people with bad views about X”. Then, you should maybe double check the extent to which a given claim is actualy right : ). For instance, I don’t think almost immediate FOOM is the typical view on LW when aggregating by most metrics, a somewhat longer duration takeoff is now a more common view I think.
By the way, where’s this number coming from? (10^30 FLOP) You keep repeating it.
Extremely rough and slightly conservatively small ball park number for how many FLOP will be used to create powerful AIs. The idea being that this will represent roughly how many FLOP could plausibly be available at the time.
GPT-4 is ~10^26 FLOP, I expect GPT-7 is maybe 10^30 FLOP.
Perhaps this is a bit too much because the scheming AI will have access to far few FLOP than exist at the time, but I expect this isn’t cruxy, so I just did a vague guess.
I greatly appreciate the effort in this reply, but I think it’s increasingly unclear to me how to make efficient progress on our disagreements, so I’m going to hop.
I was claiming that titotal’s post doesn’t appear to give arguments that directly address whether or not Yudkowsky-style ASI can invent diamondoid nanotech. I don’t understand the relevance to my comment. I agree that if you find titotal’s argument persuasive then whether it is load bearing is relevant to AI risk concerns, but that’s not what my comment is about.
FWIW Yudkowsky frequently says that this is not load bearing, and that much seems obviously true to me also.
Both this and his earlier article by titotal on computational chemistry make the argument that development of molecular nanotechnology can’t be a one-shot process because of intrinsic limitations on simulation capabilities.
Yudkowsky claims that one-shot nanotechnology is not load bearing, yet literally every example he gives when pressed involves one-shot nanotech.
Could you quote or else clearly reference a specific argument from the post you found convincing on that topic?
OP’s post regarding the topic is here (it is also linked to in the body of this article):
https://titotal.substack.com/p/bandgaps-brains-and-bioweapons-the
I’m running a molecular nanotechnology company, so I’ve had to get quite familiar with the inner workings and limitations of existing computational chemistry packages. This article does a reasonable job of touching on the major issues. I can tell you that experimentally these codes don’t even come close to the precision required to find trajectories that reliably drive diamondoid mechanosynthesis. It’s really difficult to get experiments in the lab to match the (over simplified, reduced accuracy) simulations which take days or weeks to run.
Having any hope of adequately simulating mechanosynthetic reactions requires codes that scale with O(n^7). No one even writes these codes, because what would be the point? The biggest thing you could simulate is helium. There are approximations with O(n^3) and O(n^4) running time that are sometimes useful guides, but can often have error bars larger than the relevant energy gaps. The acid test here is still experiment: try to do the thing and see what happens.
AI has potential to greatly improve productivity with respect to nano-mechanical engineering designs because you can build parts rigid and large enough, and operating at slow enough speeds that mechanical forces are unlikely to have chemical effects (bond breaking or forming), and therefore faster O(n^2) molecular mechanics codes can be used. But all the intelligence in the world isn’t going to make bootstrapping molecular nanotechnology a one-shot process. It’s going to be years of hard-won lab work, whether it’s done by a human or a thinking machine.
Well yes, nobody thinks that existing techniques suffice to build de-novo self-replicating nano machines, but that means it’s not very informative to comment on the fallibility of this or that package or the time complexity of some currently known best approach without grounding in the necessity of that approach.
One has to argue instead based on the fundamental underlying shape of the problem, and saying accurate simulation is O(n⁷) is not particularly more informative to that than saying accurate protein folding is NP. I think if the claim is that you can’t make directionally informative predictions via simulation for things meaningfully larger than helium then one is taking the argument beyond where it can be validly applied. If the claim is not that, it would be good to hear it clearly stated.
I don’t understand your objection. People who actually use computational chemistry—like me, like titotal—are familiar with its warts and limitations. These limitations are intrinsic to the problem and not expressions of our ignorance. Diamondoid mechanosynthetic reactions depend on properties which only show up in higher levels of theory, the ones which are too expensive to run for any reasonably sized system. To believe that this limitation won’t apply to AI, just because, is magical thinking.
I wasn’t saying anything with respect to de-novo self-replicating nano machines, a field which has barely been studied by anyone and which we cannot adequately say much about at all.
And what reason do you have for thinking it can’t be usefully approximated in some sufficiently productive domain, that wouldn’t also invalidly apply to protein folding? I think it’s not useful to just restate that there exist reasons you know of, I’m aiming to actually elicit those arguments here.
The fact that this has been an extremely active area of research for over 80 years with massive real-world implications, and we’re no closer to finding such a simplified heuristic. This isn’t like protein folding at all—the math is intrinsically more complicated with shared interaction terms that require high polynomial count running times. Indeed it’s provable that you can’t have a faster algorithm than those O(n^3) and O(n^4) approximations which cover all relevant edge cases accurately.
Of course there could be some heuristic-like approximation which works in the relevant cases for this specific domain, e.g. gemstone crystal lattice structures under ultra-high vacuum conditions and cryogenic temperatures. Just like there are reasonably accurate molecular mechanics models that have been developed for various specific materials science use cases.
But here’s the critical point: these heuristics are developed by evolving models based on experiment. In other domains like protein folding you can have a source of truth for generative AI training that uses simulation. That’s how AlphaFold was done. But for chemical reactions of the sort we’d be interested here the ab-initio methods that give correct results are simply too slow to use for this. Not “buy a bigger GPU” slow, but “run the biggest supercomputer for a quintillion years” slow. So the solution is to do the process manually, running actual experiments to refine the heuristic models at each step, much machine learning training models would do. These are called semi-empirical methods, and they work. But we don’t have a good semi-empirical model for the type of reactions concerned here, and it would take years of painstaking lab work to develop, which no amount of super intelligence can avoid.
It’s magical thinking to assume that an AI will just one-shot this into existence.
Thanks, I appreciate the attempt to clarify. I do though think there’s some fundamental disagreement about what we’re arguing over here that’s making it less productive than it could be. For example,
I think both:
Lack of human progress doesn’t necessarily mean the problem is intrinsically unsolvable by advanced AI. Humans often take a bunch of time before proving things.
It seems not at all the case that algorithmic progress isn’t happening, so it’s hardly a given that we’re no closer to a solution unless you first circularly assume that there’s no solution to arrive at.
If you’re starting out with an argument that we’re not there yet, this makes me think more that there’s some fundamental disagreement about how we should reason about ASI, more than your belief being backed by a justification that would be convincing to me had only I succeeded at eliciting it. Claiming that a thing is hard is at most a reason not to rule out that it’s impossible. It’s not a reason on its own to believe that it is impossible.
With regard to complexity,
I failed to understand the specific difference with protein folding. Protein folding is NP-hard, which is significantly harder than O(n³).
I failed to find the source for the claim that O(n³) or O(n⁴) are optimal. Actually I’m pretty confused how this is even a likely concept; surely if the O(n³) algorithm is widely useful then the O(n⁴) proof can’t be that strong of a bound on practical usefulness? So why is this not true of the O(n³) proof as well?
It’s maybe true that protein folding is easier to computationally verify solutions to, but first, can you prove this, and second, on what basis are you claiming that existing knowledge is necessarily insufficient to develop better heuristics than the ones we already have? The claim doesn’t seem to complete to me.
Please note that I’ve not been making the claim that ASI could necessarily solve this problem. I have been making the claim that the arguments in this post don’t usefully support the claim that it can’t. It is true that largely on priors I expect it should be able to, but my priors also aren’t particularly useful ones to this debate and I have tried to avoid making comments that are dependent on them.
Do you have a maths or computer science background? I ask because some of the questions you ask are typical of maths or CS, but don’t really make sense in this context.
Take protein folding. Existing techniques for estimating how a protein folds are exponential in complexity (making the problem NP-hard), so running time is bounded by O(2^n). But what’s the ‘n’? In these algorithms it is the number of amino acids because that’s the level at which they are operating. For quantum chemistry the polynomial running times have ‘n’ being the number of electrons in the system, which can be orders of magnitude higher. That’s why O(n^4) can be much, much worse than O(2^n) for relevant problem sizes—there’s no hope of doing full protein folding simulations by ab initio methods. The second reason they’re not comparable is time steps. Ab initio molecular mechanics methods can’t have time steps larger than 1 femtosecond. Proteins fold on the timescale of milliseconds. This by itself is what makes all ab initio molecular mechanics methods infeasible for protein folding, even though it’s a constant factor ignored by big-O notation.
Staying on the issue of complexity, from a technical / mathematical foundation perspective every quantum chemistry code should also have exponential running time, and this is trivially proved by the fact that every particle technically does interact with every other particle. But all of these interaction terms fall off with at least 1/r^2, and negative power increases with each corrective term. For 1/r^2 and 1/r^3, there are field density tricks that can be used for accurate sub-exponential modeling. For 1/r^4 or higher it is perfectly acceptable to just have distance cutoffs which cutout the exponential term. How accurate a simulation you will get is a question of how many of these corrective terms you include. The more you include, the more higher-order effects you capture, but you also get hit with higher-order polynomial running times.
To bring a long story to an abrupt end, you don’t “prove” correctness of these various codes in a mathematical sense. They’re all physically incorrect because they involve approximations. Rather for various use cases you can experimentally show that certain terms dominate, and you need to make sure that the code you use accurately captures the relevant terms for the domain and scale of the problem you are simulating. This is an experimental determination, not a mathematical proof. You literally go measure things in the lab.
The protein data bank contains hundreds of thousands of structures which have had their atomic coordinates determined by various experimental methods, providing a ground truth for protein folding heuristics. There is no such source of truth for diamondoid nano-mechanical machines parts and the mechanosynthesis steps that produce them. If you want to train a better reactive force field heuristic that could make the relevant design simulations tractable, you need a source of truth. The only thing available is various quantum chemistry approximations, as you need to approximate to be computable. These fail as source of truth because you don’t know which approximations are reasonable and which are not without experiment data, which we don’t have. So you could train a heuristic, yes, but you’d have absolutely no way of assessing its accuracy.
There’s a common saying in the AI field: “garbage in, garbage out.” The current state of non-empirical ab initio quantum chemistry is garbage. You’re saying that AI could take the garbage and by mere application of thought turn it into something useful. That’s not in line with the actual history of the development of useful AI outputs.
I’m not saying that AI can’t develop useful heuristic approximations for the simulation of gemstone-based nano-mechanical machinery operating in ultra-high vacuum. I’m saying that it can’t do so as a one-shot inference without any new experimental work, which is a trope that shows up again and again in certain high-profile people’s descriptions of AI x-risk scenarios.
If you say “Indeed it’s provable that you can’t have a faster algorithm than those O(n^3) and O(n^4) approximations which cover all relevant edge cases accurately” I am quite likely to go on a digression where I try to figure out what proof you’re pointing at and why you think it’s a fundamental barrier, and it seems now that per a couple of your comments you don’t believe it’s a fundamental barrier, but at the same time it doesn’t feel like any position has been moved, so I’m left rather foggy about where progress has been made.
I think it’s very useful that you say
since this seems like a narrower place to scope our conversation. I read this to mean:
You don’t know of any in principle barrier to solving this problem,
You believe the solution is underconstrained by available evidence.
I find the second point hard to believe, and don’t really see anywhere you have evidenced it.
As a maybe-relevant aside to that, wrt.
I think you’re talking of ‘mere application of thought’ like it’s not the distinguishing feature humanity has. I don’t care what’s ‘in line with the actual history’ of AI, I care what a literal superintelligence could do, and this includes a bunch of possibilities like:
Making inhumanly close observation of all existing data,
Noticing new, inhumanly-complex regularities in said data,
Proving new simplifying regularities from theory,
Inventing new algorithms for heuristic simulation,
Finding restricted domains where easier regularities hold,
Bifurcating problem space and operating over each plausible set,
Sending an interesting email to a research lab to get choice high-ROI data.
We can ignore the last one for this conversation. I still don’t understand why the others are deemed unreasonable ways of making progress on this task.
I appreciated the comments on time complexity but am skipping it because I don’t expect at this point that it lies at the crux.
By “proof” I meant proof by contradiction. DFT is a great O(n^3) method for energy minimizing structures and exploring electron band structure, and it is routinely used for exactly that purpose. So much so that many people conflate “DFT” with more accurate ab initio methods, which it is not. However DFT utterly ignores exchange correlation terms and so it doesn’t model van der Waals interactions at all. Every design for efficient and performant molecular nanotechnology—the ones that get you order-of-magnitude performance increases and therefore any benefit over existing biology or materials science—involve vdW forces almost exclusively in their manufacture and operation. It’s the dominant non-covalent bonded interaction at that scale.
That’s the most obvious example, but also a lot of the simulations performed by Merkle and Freitas in their minimal toolset paper give incorrect reaction sequences in these lower levels of theory, as they found out when they got money to attempt it in the lab. Without pointing to their specific failure, you can get a hint of this in surface science. Silicon, gold, and other surfaces tend to have rather interesting surface clustering and reorganization effects, which are observable by scanning probe microscopy. These are NOT predicted by the cheaper / computationally tractable codes, and they are an emergent property of higher-order exchange correlations in the crystal structure. These nevertheless have enough of an effect to drastically reshape the surface of these materials, making calculation of those forces absolutely required for any attempt to build off the surface.
Attempting to do cheaper simulations for diamondoid synthesis reactions gave very precise predictions that didn’t work as expected in the lab. How would your superintelligent AI know that uncalculated terms dominate in the simulation, and make corrective factors without having access to those incomputable factors?
I think you vastly overestimate how much knowledge is left to extracted from the data. AI has made tremendous advances in recent years where it has been able to consume huge amounts of data, far in excess of what any group of humans could analyze. This, on the other hand, is a data-poor regime.
This is happening right now. There are a variety of machine-learned molecular mechanics force fields that have been published in the last few years. The most interesting one I’ve found used periodic crystal ab initio simulation methods to create a force field potential that ended up being very good for liquid and gas-phase chemistry, which it was not trained on.
But the relevant question (if you want to talk about AI x-risk by means of bootstrapping nanotech) is how accurate they are outside of the domain where we have heaps of hard evidence, because we don’t have a ground truth to compare against in those environments.
Human engineers are very good at this. It’s not the limiting factor.
What lab. There’s literally no one doing the relevant research, or equipped to easily do it without years of preparatory chemical synthesis and surface characterization.
Which is really the point and the crux of the matter. It will take an extended, years-long research effort to create molecular nanotechnology. It’s not something you can plausibly do in secret, and certainly not something you can shorten by simulation or bayesian inference.
One quick intuition pump: do you think a team of 10,000 of the smartest human engineers and scientists could do this if they had perfect photographic memory, were immortal, and could think for a billion years?
To keep the situation analogous to an AI needing to do this quickly, we’ll suppose this team of humans is subject to the same restrictions on compute for simulations (e.g. over the entire billion years they only get 10^30 FLOP) and also can only run whatever experiments the AI would get to run (effectively no interesting experiments).
I feel uncertain about whether this team would succeed, but it does seem pretty plausible that they would be able to succeed. Perhaps I think they’re 40% likely to succeed?
Now, suppose the superintelligence is like this, but even more capable.
See also That Alien Message
Separately, I don’t think it’s very important to know what an extremely powerful superintelligence could do, because prior to the point where you have an extremely powerful superintelligence, humanity will already be obsoleted by weaker AIs. So, I think Yudkowsky’s arguments about nanotech are mostly unimportant for other reasons.
But, if you did think “well, sure the AI might be arbitrarily smart, but if we don’t give it access to the nukes what can it even do to us?” then I think that there are many sources of concern and nanotech is certainly one of them.
By merely thinking about it, and not running any experiments? No, absolutely not. I don’t you understood my post if you assume I’d think otherwise.
Try this: I’m holding a certain number of fingers behind my back. You and a team of 10,000 of the smartest human engineers and scientists have a billion years to decide, without looking, what your guess will be as to how many fingers I’m holding behind my back. But you only get one chance to guess at the end of that billion years.
That’s a more comparable example.
Please don’t use That Alien Message as an intuition pump. There’s a tremendous amount wrong with the sci-fi story. Not least of which is that it completely violates the same constraint you put into your own post about constraining computation. I suggest doing your own analysis of how many thought-seconds the AI would have in-between frames of video, especially if you assume it to be running as a large inference model.
The best thing you can do is rid yourself of the notion that superhuman AI would have arbitrary capabilities. That is where EY went wrong, and a lot of the LW crowd too. If you permit dividing by zero or multiplying by infinity, then you can easily convince yourself of anything. AI isn’t magic, and AGI isn’t a free pass to believe anything.
PS: We’ve had AGI since 2017. That’d better be compatible with your world view if you want accurate predictions.
I don’t understand where your confidence is coming from here, but fair enough. It wasn’t clear to me if your take was more like “wildly, wildly superintelligent AI will be considerably weaker than a team of humans thinking for a billion years” or more like “literally impossible without either experiments or >>10^30 FLOP”.
I generally have an intuition like “it’s really, really hard to rule out physically possible things out without very strong evidence, by default things have a reasonable chance of being possible (e.g. 50%) when sufficient intelligence is applied if they are physically possible”. It seems you don’t share this intuition, fair enough.
(I feel like this applies for nearly all human inventions? Like if you had access to a huge amount of video of the world from 1900 and all written books that existed at this point, and had the affordances I described with a team of 10,000 people, 10^30 FLOP, and a billion years, it seems to me like there is a good chance you’d be able to one-shot reinvent ~all inventions of modern humanity (not doing everything in the same way, in many cases you’d massively over engineer to handle one-shot). Planes seem pretty easy? Rockets seem doable?)
I think this is an ok, but not amazing intuition pump for what wildly, wildly superintelligent AI could be like.
I separately think it’s not very important to think about the abilities of wildly, wildly superintelligent AI for most purposes (as I noted in my comment). So I agree that imagining arbitrary capabilities is probablematic. (For some evidence that this isn’t post-hoc justification, see this post on which I’m an author.)
Uhhhh, I’m not sure I agree with this as it doesn’t seem like nearly all jobs are easily fully automatable by AI. Perhaps you use a definition of AGI which is much weaker like “able to speak slightly coherant english (GPT-1?) and classify images”?
By the way, where’s this number coming from? You keep repeating it. That amount of calculation is equivalent to running the largest supercomputer in existence for 30k years. You hypothetical AI scheming breakout is not going to have access to that much compute. Be reasonable.
Ok let’s try a different tract. You want to come up with a molecular mechanics model that can efficiently predict the outcome of reactions, so that you can get about designing one-shot nanotechnology bootstrapping. What would success look like?
You can’t actually do a full simulation to get ground truth for training a better molecular mechanics model. So how would you know the model you came up will work as intended? You can back-test against published results in the literature, but surprise surprise, a big chunk of scientific papers don’t replicate. Shoddy lab technique, publication pressure, and a niche domain combine to create conditions where papers are rushed and sometimes not 100% truthful. Even without deliberate fraud (which also happens), you run into problems such as synthesis steps not working as advertised, images used from different experimental runs than the one described in the paper, etc.
Except you don’t know that. You’re not allowed to do experiments! Maybe you guess that replication will be an issue, although why that would be a hypothesis in the first place without first seeing failures in the lab first isn’t clear to me. But let’s say you do. Which experiments should you discount? Which should you assume to be correct? If you’re allowed to start choosing which reported results you believe and which you don’t, you’ve lost the plot. There could be millions of possible heuristics which partially match the literature and there’s no way to tell the difference.
So what would success look like? How would you know you have the right molecular mechanics model that gives accurate predictions?
You can’t. Not any more than Descartes could think his way to absolute truth.
Also for what it’s worth you’ve made a couple of logical errors here. You are considering human inventions which already exist, then saying that you could one-shot invent them in 1900. That’s hindsight bias, but also selection bias. Nanotechnology doesn’t exist. Even if it would work if created, there’s no existence proof that there exists an accessible path to achieving it. Like superheavy atoms in the island of stability, or micro black holes, there just might not be a pathway to make them from present day capabilities. (Obviously I don’t believe this as I’m running a company attempting to bootstrap Drexlarian nanotechnology, but I feel it’s essential to point out the logical error.)
Why? You’ve gone into circular logic here.
I pointed out that the Alien Message story makes fundamental errors with respect to computational capability being wildly out of scale, so actual super intelligent AIs aren’t going to be anything like the one in the story.
Maybe a Jupiter-sized matryoshka brain made of computronium would exhibit this level of super intelligence. I’m not saying it’s not physically possible. But in terms of sketching out and bounding the capabilities of near-term AI/ASI, it’s a fucking terrible intuition pump.
The transformer architecture introduced in 2017 is:
Artificial: man-made
General: able to train over arbitrary unstructured input, from which it infers models that can be applied in arbitrarily ways to find solutions of problems drawn from domains outside of its training data.
Intelligent: able to construct efficient solutions to new problems it hasn’t seen.
Artificial General Intelligence. A.G.I.
If you’re thinking “yeah, but..” then I suggest you taboo the term AGI. This is literally all that it the word means.
If you want to quibble over dates then maybe we can agree on 2022 with the introduction of ChatGPT, a truly universal (AKA general) interface to mature transformer technology. Either way we’re already well within the era of artificial general intelligence.
(Maybe EY’s very public meltdown a year ago is making more sense now? But rest easy, EY’s predictions about AI x-risk have consistently been wildly off the mark.)
FWIW, I do taboo this term and thus didn’t use it in this conversation until you introduced it.
You highlighted “disagree” on the part about AGI’s definition. I don’t know how to respond to that directly, so I’ll do so here. Here’s the story about how the term “AGI” was coined, by the guy who published the literal book on AGI and ran the AGI conference series for the past two decades:
https://web.archive.org/web/20181228083048/http://goertzel.org/who-coined-the-term-agi/
LW seems to have adopted some other vague, ill-defined, threatening meaning for the acronym “AGI” that is never specified. My assumption is that when people say AGI here they mean Bostrom’s ASI, and they got linked because Eliezer believed (and believes still?) that AGI will FOOM into ASI almost immediately, which it has not.
Anyway it irks me that the term has been coopted here. AGI is a term of art in the pre-ML era of AI research with a clearly defined meaning.
Definition in the OpenAI Charter:
A post on the topic by Richard (AGI = beats most human experts).
In case this wasn’t clear from early discussion, I disagree with Eliezer on a number of topics, including takeoff speeds. In particular I disagree about the time from AI that is economically transformative to AI that is much, much more powerful.
I think you’ll probably find it healthier and more productive to not think of LW as an amorphous collective and instead note that there are a variety of different people who post on the forum with a variety of different views. (I sometimes have made this mistake in the past and I find it healthy to clarify at least internally.)
E.g. instead of saying “LW has bad views about X” say “a high fraction of people who comment on LW have bad views about X” or “a high fraction of karma votes seem to be from people with bad views about X”. Then, you should maybe double check the extent to which a given claim is actualy right : ). For instance, I don’t think almost immediate FOOM is the typical view on LW when aggregating by most metrics, a somewhat longer duration takeoff is now a more common view I think.
Also, I’m going to peace out of this discussion FYI.
Extremely rough and slightly conservatively small ball park number for how many FLOP will be used to create powerful AIs. The idea being that this will represent roughly how many FLOP could plausibly be available at the time.
GPT-4 is ~10^26 FLOP, I expect GPT-7 is maybe 10^30 FLOP.
Perhaps this is a bit too much because the scheming AI will have access to far few FLOP than exist at the time, but I expect this isn’t cruxy, so I just did a vague guess.
I wasn’t trying to justify anything, just noting my stance.
I greatly appreciate the effort in this reply, but I think it’s increasingly unclear to me how to make efficient progress on our disagreements, so I’m going to hop.