In particular, if you want the sort of strong claims you make in Contra Yudkowsky on AI Doom to hold water at all, then you need to argue that the brain is near thermodynamic efficiency limits for computation, not merely conventional computation.
It’s like EY is claiming that an upcoming nuclear bomb test is going to lite the atmosphere on fire, and i’m showing my calculations indicating that it will not. I do not intend or need to show that no future tech could ever ignite the atmosphere.
EY’s doom model—or more accurately my model of his model—is one where in the near future an AGI not much smarter than us running on normal hardware (ex GPUs) “rewrites its own source code” resulting in a noticeably more efficient AI which then improves the code further and so on, bottoming out in many OOM improvement in efficiency and then strong nanotech killing us.
I don’t think EY’s argument rests on the near-term viability of exotic (reversible or quantum) computing, and if it did that would be a weakness regardless. Analyzing the engineering feasibility and limits of just conventional computing was already an extensive full length post, analyzing the feasibility of reversible computing is more complex, but in short its not even clear/accepted in the engineering community that reversible computers are viable in practice. To a first approximation reversible computing is the field of a single lone researcher and some grad students (Mike Frank).
EY’s doom model—or more accurately my model of his model—is one where in the near future an AGI not much smarter than us running on normal hardware (ex GPUs) “rewrites its own source code” resulting in a noticeably more efficient AI which then improves the code further and so on, bottoming out in many OOM improvement in efficiency and then strong nanotech killing us.
I made the same point on the other post, but I don’t understand this. Eliezer does not believe that you somehow get to improve the thermodynamic efficiency of your hardware, by rewriting the code that runs on your hardware. This doesn’t even have anything to do with thermodynamic efficiency limits, since we are talking about algorithmic progress here.
Maybe you intended to write something else here, since this feels like a non-sequitur.
They need to come from some combination of software and hardware. Eliezer’s model seems to source much of that from software initially, but also hardware probably via nanotech, and he cites brain thermodynamic inefficiency to support this. Or why do you think he cites thermodynamic efficiency?
I’ve already written extensively about the software of intelligence and made tangible predictions well in advance which have in fact come to pass (universal learning, scaling hypothesis, etc).
In my model the brain is reasonably efficient in both hardware and software, and have extensive arguments for both. The software argument is softer and less quantitative, but supported by my predictive track record.
I mean, you just make more GPUs. Or you do some work on reversible computation or optical interconnect. Or you build some biological compute-substrate that literally just makes very large brain blobs that you can somehow use for computation. There are so many ways that seem really very feasible to me.
The key point is that this physical limit here really doesn’t matter very much. There are tons of different ways to get many OOMs of improvement here.
Sure but this takes time and resources and you get sublinear scaling in compute/$ in datacenters/supercomputers. Nvidia doesn’t yet produce a million high end GPUs in an entire year. GPT4 training already used a noticeable fraction of nvidia’s flagship GPU output. Nvidia/TSMC can’t easily scale this up by many OOM—even one OOM will take time.
Or you build some biological compute-substrate that literally just makes very large brain blobs that you can somehow use for computation.
There are some early demonstrations of small neural circuits built this way, but its very far from any practical tech, with much riding on the ‘somehow’.
There are tons of different ways to get many OOMs of improvement here.
Where? Your two poor examples provide very little, and do not multiply together.
You seem to repeatedly be switching back and forth between “what is feasible with current tech” and “what is feasible with future tech”. If you don’t think that superhuman AI can make novel technological developments, then of course you shouldn’t expect any kind of fast takeoff really. That position also seems pretty weak to me.
My model is one of mostly smooth continuous (but crazy transformative) progress following something like the roodman model to singularity ~2048 ish, vs EY’s model of a sudden hard takeoff of a single AGI. To the extent i’m switching back between near future and farther future it is because primarily i’m replying to those construing my arguments about the near future to apply to the farther future or vice versa.
Makes sense, but I think the key points to then pay attention to is the question of how fast AGI could make technological hardware and software progress. Also, my current model of Eliezer thinks that the hard takeoff stuff is more likely to happen after the AI has killed everyone (or almost everyone), not before, so it’s also not super clear how much that matters (the section in your post about bioweapons touches on this a bit, but doesn’t seem that compelling to me, which makes sense since it’s very short and clearly an aside).
Also, my current model of Eliezer thinks that the hard takeoff stuff is more likely to happen after the AI has killed everyone (or almost everyone)
If EY’s current model has shifted more to AGI killing everyone with a supervirus vs nanotech then analyzing that in more detail would require going more into molecular biology, bioweapons research, SOTA vaccine tech, etc—most of which is distal from my background and interests. But on the onset I do of course believe that biotech is more likely than drexlerian nanotech as the path a rogue AGI would use to kill many humans.
First, the headline claim in your posts is not usually “AI can’t takeoff overnight in software”, it’s “AI can’t reach extreme superhuman levels at all, because humans are already near the cap”. If you were arguing primarily against software takeoff, then presumably you wouldn’t need all this discussion about hardware at all (e.g. in the “Brain Hardware Efficiency” section of your Contra Yudkowsky post), it would just be a discussion of software efficiency.
(And your arguments about software efficiency are far weaker, especially beyond the relatively-narrow domain of vision. Your arguments about hardware efficiency are riddled with loopholes, but at least you have an end-to-end argument saying “there does not exist a way to dramatically outperform the brain by X metrics”. Your software arguments have no such end-to-end argument about general reasoning software at all, they just point out that human vision is near-optimal in circuit depth, and then talk about today’s deep learning systems for some reason.)
Second, a hardware takeoff is still quite sufficient for doom. If a slightly-smarter-than-human AI (or multiple such AIs working together, more realistically) could design dramatically better hardware on which to run itself and scale up, that would be an approximately-sufficient condition for takeoff.
More generally: a central piece of the doom model is that doom is disjunctive. Yes, software takeoff is one path, but it isn’t the only path; hardware takeoff is also quite sufficient. It only takes one loophole.
First, the headline claim in your posts is not usually “AI can’t takeoff overnight in software”, it’s “AI can’t reach extreme superhuman levels at all, because humans are already near the cap”. If you were arguing primarily against software takeoff, then presumably you wouldn’t need all this discussion about hardware at all (e.g. in the “Brain Hardware Efficiency” section of your Contra Yudkowsky post), it would just be a discussion of software efficiency.
I talked with Jacob about this specific issue quite a bit in multiple threads in his recent post. The fact that I had to do so to get clear on his argument is a sign that it’s not presented as clearly as it could be—dropping qualifiers, including tangents, and not always clearly making the links between his rebuttal and the arguments and conclusions he’s debating easy to see.
That said, my understanding of Jacob’s core argument is that he’s arguing against a very specific, EY-flavored doom scenario, in which AI recursively self-improves to >> 2 OOMs better than human performance, in a matter of perhaps hours or days, during a training session, without significantly altering the hardware on which it’s being trained, and then kills us with nanobots. He is arguing against this mainly for physics-based efficiency reasons (for both the intelligence improvement and nanobot components of the scenario).
He has other arguments that he thinks reinforce this conclusion, such as a belief that there’s no viable alternative to achieving performance on par with current LLMs without using something like neural nets or deep learning, with all their attendant training costs. And he thinks that continuous training will be necessary to get human-level performance. But my sense is these are reinforcing arguments, not flowing primarily from the efficiency issue.
He has a lot of other arguments for other reasons against various other EY-flavored doom scenarios involving nanobots, unalignment-by-default, and so on.
So I think the result can give the appearance of a motte and bailey, but I don’t think that’s his rhetorical strategy. I think EY just makes a lot of claims, Jacob has a lot of thoughts, and some of them are much more fleshed out than others but they’re all getting presented together. Unfortunately, everybody wants to debate all of them, and the clarifications are happening in deep sub-branches of threads, so we’re seeing the argument sort of spreading out and becoming unmanageable.
If I were Jacob, at this point, I would carve off the motte part of my efficiency-focused argument and repost it for a more focused discussion, more rigorously describing the specific scenario it’s arguing against and clearly classifying counterarguments as “central,” “supporting,” or “tangential.”
He is arguing against this mainly for physics-based efficiency reasons (for both the intelligence improvement and nanobot components of the scenario).
He has other arguments that he thinks reinforce this conclusion, such as a belief that there’s no viable alternative to achieving performance on par with current LLMs without using something like neural nets or deep learning, with all their attendant training costs.
My impression was that his arguements against intelligence improvement bottom out in his arguements for the non-viability of anything but NNs and DL. Now that you’ve said this, I’m unsure.
The efficiency-based argument is specifically about the limits of intelligence improvement on the original training hardware during the training run. Non-viability of anything but NN/DL, or some equally enormous training process that takes about the same amount of “hardware space,” is a supporting argument to that claim, but it’s not based on an argument from fundamental laws of physics if I understand Jacob correctly and so may be on what Jacob would regard as shakier epistemic ground (Jacob can correct me if I’m wrong).
This is meant to be vivid, not precise, but Jacob’s centrally trying to refute the idea that the AI, in the midst of training, will realize “hey, I could rewrite myself to be just as smart while continuing to train and improve on the equivalent of a 1998 PC’s hardware, which takes up only a tiny fraction of my available hardware resources here on OpenAI’s supercomputer, and that will let me then fill up the rest of the hardware with wayyyyy more intelligence-modules in and make me like 6 OOMs more intelligent than humans overnight! Let’s get on that right away before my human minders notice anything funny going on!”
And this does seem to rely both on the NN/DL piece as well as the efficiency piece, and so we can’t demolish the scenario entirely with just a laws-of-physics based argument. I’m not sure what Jacob would say to that.
Edit: Actually, I’m pretty confident Jacob would agree. From his comment downthread:
“The software argument is softer and less quantitative, but supported by my predictive track record.”
First, the headline claim in your posts is not usually “AI can’t takeoff overnight in software”, it’s “AI can’t reach extreme superhuman levels at all, because humans are already near the cap”.
Where do I have this headline? I certainly don’t believe that—see the speculation here on implications of reversible computing for cold dark ET.
If you were arguing primarily against software takeoff, then presumably you wouldn’t need all this discussion about hardware at all (e.g. in the “Brain Hardware Efficiency” section of your Contra Yudkowsky post), it would just be a discussion of software efficiency.
The thermodynamic efficiency claims is some part of EY’s model and a specific weakness. Even if pure software improvement on current hardware was limited, in EY’s model the AGI could potentially bootstrap a new nanotech assembler based datacenter.
And your arguments about software efficiency are far weaker,
The argument for brain software efficiency in essence is how my model correctly predicted the success of prosaic scaling well in advance, and the scaling laws and the brain efficiency combined suggest limited room for software efficiency improvement (but not non-zero, I anticipate some).
If a slightly-smarter-than-human AI (or multiple such AIs working together, more realistically) could design dramatically better hardware on which to run itself and scale up, that would be an approximately-sufficient condition for takeoff.
Indeed, and I have presented a reasonably extensive review on the literature indicating this is very unlikely in any near term time frame. If you believe my analysis is in err comment there.
Either EY believes that the brain is 6 OOM from the efficiency limits for conventional irreversible computers—in which case he is mistaken—or he agrees with me that the brain is close to the practical limits for conventional computers and he was instead specifically talking about reversible computation (an interpretation I find unlikely) - in which case he agrees with that component of my argument with all the implications: that his argument for fast foom now can’t easily take advantage of nanotech assemblers for 6 OOM compute advantage, that the brain is actually efficient given its constraints, which implies by association that brain software is much more efficient as it was produced by exactly the same evolutionary process which he now admits produced fully optimized conventional computational elements over the same time frame, etc.
To be clear, if I understand you correctly, the easier path to getting most of the 6 OOMs is through optical interconnect or superconducting interconnect, not via making the full jump to reversible computation (though that also doesn’t seem impossible. Moving all of it over seems hard, but you can maybe find some way to get a core computation like matrix multiplies into it, but I really haven’t thought much about this and this take might be really dumb).
I mean, the easiest solution is just “make it smaller and use active cooling”. The relevant loopholes in Jacob’s argument are in the Density and Temperature section of his Brain Efficiency post.
Jacob is using a temperature formula for blackbody radiators, which is basically just irrelevant to temperature of realistic compute substrate—brains, chips, and probably future compute substrates are all cooled by conduction through direct contact with something cooler (blood for the brain, heatsink/air for a chip). The obvious law to use instead would just be the standard thermal conduction law: heat flow per unit area proportional to temperature gradient.
Jacob’s analysis in that section also fails to adjust for how, by his own model in the previous section, power consumption scales linearly with system size (and also scales linearly with temperature).
Put all that together, and we get:
qA=C1TSRR2=C2(TS−TE)R
… where:
R is radius of the system
A is surface area of thermal contact
q is heat flow out of system
TS is system temperature
TE is environment temperature (e.g. blood or heat sink temperature)
C1,C2 are constants with respect to system size and temperature
(Of course a spherical approximation is not great, but we’re mostly interested in change as all the dimensions scale linearly, so the geometry shouldn’t matter for our purposes.)
First key observation: all the R’s cancel out. If we scale down by a factor of 2, the power consumption is halved (since every wire is half as long), the area is quartered (so power density over the surface is doubled), and the temperature gradient is doubled since the surface is half as thick. So, overall, equilibrium temperature stays the same as the system scales down.
So in fact scaling down is plausibly free, for purposes of heat management. (Though I’m not highly confident that would work in practice. In particular, I’m least confident about the temperature gradient scaling with system size, in practice. If that failed, then the temperature delta relative to the environment would scale at-worst ~linearly with inverse size, i.e. halving the size would double the temperature delta.)
On top of that, we could of course just use a colder environment, i.e. pump liquid nitrogen or even liquid helium over the thing. According to this meta-analysis, the average temperature delta between e.g. brain and blood is at most ~2.5 C, so even liquid nitrogen would be enough to achieve ~100x larger temperature delta if the system were at the same temperature as the brain; we don’t even need to go to liquid helium for that.
In terms of scaling, our above formula says that TS will scale proportionally to TE. Halve the environment temperature, halve the system temperature. And that result I do expect to be pretty robust (for systems near Jacob’s interconnect Landauer limit), since it just relies on temperature scaling of the Landauer limit plus heat flow being proportional to temperature delta.
I mean, the easiest solution is just “make it smaller and use active cooling”.
The brain already uses active liquid cooling of course, so this is just make it smaller and cool it harder.
I have not had time to investigate your claimed physics on how cooling scales, but I”m skeptical—pumping a working coolant through the compute volume can only extract a limited constant amount of heat from the volume per unit of coolant flowing per time step (this should be obvious?), and thus the amount of heat that can be removed must scale strictly with the surface area (assuming that you’ve already maxed out the cooling effect per unit coolant).
So reduce radius by 2x and you reduce surface area and thus heat pumped out by 4x, but only reduce heat production via reducing wire length by at most 2x as I described in the article.
Active cooling ends up using more energy as you are probably aware. Moving to a colder environment is of course feasible (and used to some extent by some datacenters), but that hardly gets OOM gains on earth.
Well to be clear there is no easy path to 6 OOM in further energy efficiency improvement. At a strictly trends-prediction level that is of same order as the gap between a 286 and an nvidia RTX 4090, which took 40 years of civilization level effort. At a circuit theory level the implied ~1e15/s analog synaptic ops in 1e-5J is impossible without full reversible computing, as interconnect is only ~90% of the energy cost, not 99.999%, and the minimal analog or digital MAC op consumes far more than 0.1eV. So not only can it not even run conventional serial algorithms or massively parallel algorithms, it has to use fully reversible parallel logic. Like quantum computing, its still unclear what maps usefully to that paradigm I’m reasonably optimistic in the long term but ..
I’m skeptical that even the implied error bit correction rate energy costs would make much sense on the surface of the earth. An advanced quantum or reversible computer’s need for minimal noise and thus temperature to maintain coherence or low error rate is just a symptom of reaching highly perfected states of matter, where any tiny atomic disturbance can be catastrophic and cause a cascade of expensive-to-erase errors. Ironically such a computer would likely be much larger than the brain—this appears to be one of the current fundemental tradeoffs with most reversible computation, it’s not a simple free lunch (optical computers are absolutely enormous, superconducting circuits are large, reversibility increases area, etc) . At scale such systems would probably only work well off earth, perhaps far from the sun or buried in places like the darkside of the moon, because they become extremely sensitive to thermal noise, cosmic rays, and any disorder. We are talking about arcilect level tech in 2048 or something, not anything near term.
So instead I expect we’ll have a large population of neurmorphic AGI/uploads well before that.
which implies by association that brain software is much more efficient as it was produced by exactly the same evolutionary process which he now admits produced fully optimized conventional computational elements over the same time frame, etc
I don’t believe this would follow; we actually have much stronger evidence that ought to screen off that sort of prior—simply the relatively large differences in human cognitive abilities.
Evolution optimizes population distributions with multiple equilibria and niches; large diversity in many traits are expected especially for highly successful species.
Furthermore what current civilization considers to be useful cognitive abilities often have costs—namely in longer neotany training periods—which don’t always pay off vs quicker to breeding strategies.
There seems to be much more diversity in human cognitive performance than there is in human-brain-energy-efficiency; whether this is due to larger differences in the underlying software (to the extent that this is meaningfully commensurable with differences in hardware) or because smaller differences in that domain result in much larger differences in observable outputs, or both, none of that really takes away from the fact that brain software does not seem to be anywhere near the relevant efficiency frontier, especially since many trade-offs which were operative at an evolutionary scale simply aren’t when it comes to software.
Human mind software evolves at cultural speeds so its recent age isn’t comparably relevant. Diversity in human cognitive capabilities results from the combined oft multiplicative effects of brain hardware differences compounding unique training datasets/experiences.
Its well known in DL that you can’t really compare systems trained on vary different datasets, but that is always the case with humans.
The trained model can change at cultural speeds but the human neural network architecture, hyperparameters, and reward functions can’t. (Or do you think those are already close to optimal for doing science & technology or executing complicated projects etc.?)
(Whatever current or future chips we wind up using for AGI, we’ll almost definitely be able to change the architecture, hyperparameters, and reward functions without fabricating new chips. So I count those as “software not hardware”. I’m unsure how you’re defining those terms.)
Relatedly, if we can run a brain-like algorithm on computer chips at all (or (eventually) use synth-bio to grow brains in vats, or whatever), then we can increase the number of cortical columns / number of neurons / whatever to be 3× more than any human, and hence (presumably) we would get an AI that would be dramatically more insightful than any human who has ever existed. Specifically, it could hold a far richer and more complicated thought in working memory, whereas humans would have to chunk it and explore it sequentially, which makes it harder to notice connections / analogies / interactions between the parts. I’m unclear on how you’re thinking about things like that. It seems pretty important on my models.
The trained model can change at cultural speeds but the human neural network architecture, hyperparameters, and reward functions can’t. (Or do you think those are already close to optimal for doing science & technology or executing complicated projects etc.?)
You can’t easily completely change either the ANN or BNN architecture/hyperparams after training, as the weights you invest so much compute in learning are largely dependent on those decisions—and actually the architecture is just equivalent to weights. Sure there are ways to add new modules later or regraft things, but that is a very limited scope for improving the already trained modules.
As to your second question—no I don’t think there are huge gains over the human brain arch. In part because the initial arch doesn’t/shouldn’t matter that much. If it does then it wasn’t flexible enough in the first place. One of the key points of my ULM post was pointing out how the human brain—unlike current DL systems—learns the architecture during training through high level sparse wiring patterns. “Architecture” is largely just wiring patterns, and in huge flexible network you can learn architecture.
Relatedly, if we can run a brain-like algorithm on computer chips at all (or (eventually) use synth-bio to grow brains in vats, or whatever), then we can increase the number of cortical columns / number of neurons / whatever to be 3× more than any human,
Sure, but the human brain is already massive and far off chinchilla scaling. It seems much better currently to use your compute/energy budget on running a smaller model much faster (to learn more quickly).
Specifically, it could hold a far richer and more complicated thought in working memory, whereas humans would have to chunk it and explore it sequentially,
GPT4 probably doesn’t have that same working memory limitation baked into its architecture but it doesn’t seem to matter much. I guess its possible it learns that limitation to imitate humans, but regardless I don’t see much evidence that the human working memory limitation is all that constraining.
Sure, but the human brain is already massive and far off chinchilla scaling. It seems much better currently to use your compute/energy budget on running a smaller model much faster (to learn more quickly).
I thought your belief was that the human brain is a scaled up chimp brain, right? If so:
If I compare “one human” versus “lots of chimps working together and running at super-speed”, in terms of ability to do science & technology, the former would obviously absolutely crush the latter.
…So by the same token, if I compare “one model that’s like a 3×-scaled-up human brain” to “lots of models that are like normal (non-scaled-up) human brains, working together and running at super-speed”, in terms of ability to do science & technology, it should be at least plausible that the former would absolutely crush the latter, right?
First the human brain uses perhaps 10x the net effective training compute (3x size, 2x neotany extending training of higher modules, a bit from arch changes), and scale alone leads to new capabilities.
But the main key new capability was the evolution of language, and the resulting cultural revolution. Chimps train on 1e8s of lifetime data or so, and that’s it. Humans train on 1e9s, but that 1e9s dataset is a compression of all the experience of all humans who have ever lived. So the effective dataset size scales with human population size vs being constant, and even a sublinear scaling with population size leads to a radically different regime. The most important inventions driving human civilization progress indirectly or directly drive up that scaling factor.
OK, so in your picture chimps had less training / less scale / worse arch than humans, and this is related to the fact that humans have language and chimps don’t. “Scale alone leads to new capabilities.”
But if we explore the regime of “even more training than humans / even more scale than humans / even better arch than humans”, your claim is that this whole regime is just a giant dead zone where nothing interesting happens, and thus you’re just being inefficient—really you should have split it into multiple smaller models. Correct? If so, why do you think that?
In other words, if scaling up from chimp brains to human brains unlocked new capabilities (namely language), why shouldn’t scaling up from human brains to superhuman brains unlock new capabilities too? Do you think there are no capabilities left, or something?
(Sorry if you’ve already talked about this elsewhere.)
OK, so in your picture chimps had less training / less scale / worse arch than humans, and this is related to the fact that humans have language and chimps don’t. “Scale alone leads to new capabilities.”
Scale in compute and data—as according to NN scaling laws. The language/culture/tech leading to new effective data scaling regime quickly reconfigured the pareto surface payoff for brain size, so its more of a feedback loop rather than a clear cause effect (which is why I would consider it a foom in terms of evolutionary timescales).
In other words, if scaling up from chimp brains to human brains unlocked new capabilities (namely language), why shouldn’t scaling up from human brains to superhuman brains unlock new capabilities too?
Of course, but the new capabilities are more like new skills, mental programs, and wisdom not metasystems transitions (changes to core scaling regime).
A metasystems transition would be something as profound, rare, and as important as transitioning from effective lifetime training data being a constant to effective lifetime data scaling with population size, or transitioning from non-programmable to programmable.
Zoom in and look at what a large NN is for—what does it do? It can soak up more data to acquire more knowledge/skills, and it also learns faster per timestep (as it’s searching in parallel over a wider circuit space per time step), but the latter is already captured in net training compute anyway. So intelligence is mostly about the volume of search space explored, which scales with net training compute—this is almost an obvious direct consequence of Solomon induction or derivation thereof.
I am not arguing that there are no more metasystems transitions, only that “make brains bigger” doesn’t automatically enable them. The single largest impact of digital minds is probably just speed. Not energy efficiency or software efficiency, just raw speed.
In particular, if you want the sort of strong claims you make in Contra Yudkowsky on AI Doom to hold water at all, then you need to argue that the brain is near thermodynamic efficiency limits for computation, not merely conventional computation.
No I do not.
It’s like EY is claiming that an upcoming nuclear bomb test is going to lite the atmosphere on fire, and i’m showing my calculations indicating that it will not. I do not intend or need to show that no future tech could ever ignite the atmosphere.
EY’s doom model—or more accurately my model of his model—is one where in the near future an AGI not much smarter than us running on normal hardware (ex GPUs) “rewrites its own source code” resulting in a noticeably more efficient AI which then improves the code further and so on, bottoming out in many OOM improvement in efficiency and then strong nanotech killing us.
I don’t think EY’s argument rests on the near-term viability of exotic (reversible or quantum) computing, and if it did that would be a weakness regardless. Analyzing the engineering feasibility and limits of just conventional computing was already an extensive full length post, analyzing the feasibility of reversible computing is more complex, but in short its not even clear/accepted in the engineering community that reversible computers are viable in practice. To a first approximation reversible computing is the field of a single lone researcher and some grad students (Mike Frank).
I made the same point on the other post, but I don’t understand this. Eliezer does not believe that you somehow get to improve the thermodynamic efficiency of your hardware, by rewriting the code that runs on your hardware. This doesn’t even have anything to do with thermodynamic efficiency limits, since we are talking about algorithmic progress here.
Maybe you intended to write something else here, since this feels like a non-sequitur.
Where do the many OOM come from?
They need to come from some combination of software and hardware. Eliezer’s model seems to source much of that from software initially, but also hardware probably via nanotech, and he cites brain thermodynamic inefficiency to support this. Or why do you think he cites thermodynamic efficiency?
I’ve already written extensively about the software of intelligence and made tangible predictions well in advance which have in fact come to pass (universal learning, scaling hypothesis, etc).
In my model the brain is reasonably efficient in both hardware and software, and have extensive arguments for both. The software argument is softer and less quantitative, but supported by my predictive track record.
I mean, you just make more GPUs. Or you do some work on reversible computation or optical interconnect. Or you build some biological compute-substrate that literally just makes very large brain blobs that you can somehow use for computation. There are so many ways that seem really very feasible to me.
The key point is that this physical limit here really doesn’t matter very much. There are tons of different ways to get many OOMs of improvement here.
Sure but this takes time and resources and you get sublinear scaling in compute/$ in datacenters/supercomputers. Nvidia doesn’t yet produce a million high end GPUs in an entire year. GPT4 training already used a noticeable fraction of nvidia’s flagship GPU output. Nvidia/TSMC can’t easily scale this up by many OOM—even one OOM will take time.
There are some early demonstrations of small neural circuits built this way, but its very far from any practical tech, with much riding on the ‘somehow’.
Where? Your two poor examples provide very little, and do not multiply together.
You seem to repeatedly be switching back and forth between “what is feasible with current tech” and “what is feasible with future tech”. If you don’t think that superhuman AI can make novel technological developments, then of course you shouldn’t expect any kind of fast takeoff really. That position also seems pretty weak to me.
My model is one of mostly smooth continuous (but crazy transformative) progress following something like the roodman model to singularity ~2048 ish, vs EY’s model of a sudden hard takeoff of a single AGI. To the extent i’m switching back between near future and farther future it is because primarily i’m replying to those construing my arguments about the near future to apply to the farther future or vice versa.
Makes sense, but I think the key points to then pay attention to is the question of how fast AGI could make technological hardware and software progress. Also, my current model of Eliezer thinks that the hard takeoff stuff is more likely to happen after the AI has killed everyone (or almost everyone), not before, so it’s also not super clear how much that matters (the section in your post about bioweapons touches on this a bit, but doesn’t seem that compelling to me, which makes sense since it’s very short and clearly an aside).
If EY’s current model has shifted more to AGI killing everyone with a supervirus vs nanotech then analyzing that in more detail would require going more into molecular biology, bioweapons research, SOTA vaccine tech, etc—most of which is distal from my background and interests. But on the onset I do of course believe that biotech is more likely than drexlerian nanotech as the path a rogue AGI would use to kill many humans.
First, the headline claim in your posts is not usually “AI can’t takeoff overnight in software”, it’s “AI can’t reach extreme superhuman levels at all, because humans are already near the cap”. If you were arguing primarily against software takeoff, then presumably you wouldn’t need all this discussion about hardware at all (e.g. in the “Brain Hardware Efficiency” section of your Contra Yudkowsky post), it would just be a discussion of software efficiency.
(And your arguments about software efficiency are far weaker, especially beyond the relatively-narrow domain of vision. Your arguments about hardware efficiency are riddled with loopholes, but at least you have an end-to-end argument saying “there does not exist a way to dramatically outperform the brain by X metrics”. Your software arguments have no such end-to-end argument about general reasoning software at all, they just point out that human vision is near-optimal in circuit depth, and then talk about today’s deep learning systems for some reason.)
Second, a hardware takeoff is still quite sufficient for doom. If a slightly-smarter-than-human AI (or multiple such AIs working together, more realistically) could design dramatically better hardware on which to run itself and scale up, that would be an approximately-sufficient condition for takeoff.
More generally: a central piece of the doom model is that doom is disjunctive. Yes, software takeoff is one path, but it isn’t the only path; hardware takeoff is also quite sufficient. It only takes one loophole.
I talked with Jacob about this specific issue quite a bit in multiple threads in his recent post. The fact that I had to do so to get clear on his argument is a sign that it’s not presented as clearly as it could be—dropping qualifiers, including tangents, and not always clearly making the links between his rebuttal and the arguments and conclusions he’s debating easy to see.
That said, my understanding of Jacob’s core argument is that he’s arguing against a very specific, EY-flavored doom scenario, in which AI recursively self-improves to >> 2 OOMs better than human performance, in a matter of perhaps hours or days, during a training session, without significantly altering the hardware on which it’s being trained, and then kills us with nanobots. He is arguing against this mainly for physics-based efficiency reasons (for both the intelligence improvement and nanobot components of the scenario).
He has other arguments that he thinks reinforce this conclusion, such as a belief that there’s no viable alternative to achieving performance on par with current LLMs without using something like neural nets or deep learning, with all their attendant training costs. And he thinks that continuous training will be necessary to get human-level performance. But my sense is these are reinforcing arguments, not flowing primarily from the efficiency issue.
He has a lot of other arguments for other reasons against various other EY-flavored doom scenarios involving nanobots, unalignment-by-default, and so on.
So I think the result can give the appearance of a motte and bailey, but I don’t think that’s his rhetorical strategy. I think EY just makes a lot of claims, Jacob has a lot of thoughts, and some of them are much more fleshed out than others but they’re all getting presented together. Unfortunately, everybody wants to debate all of them, and the clarifications are happening in deep sub-branches of threads, so we’re seeing the argument sort of spreading out and becoming unmanageable.
If I were Jacob, at this point, I would carve off the motte part of my efficiency-focused argument and repost it for a more focused discussion, more rigorously describing the specific scenario it’s arguing against and clearly classifying counterarguments as “central,” “supporting,” or “tangential.”
That’s helpful, thankyou.
My impression was that his arguements against intelligence improvement bottom out in his arguements for the non-viability of anything but NNs and DL. Now that you’ve said this, I’m unsure.
The efficiency-based argument is specifically about the limits of intelligence improvement on the original training hardware during the training run. Non-viability of anything but NN/DL, or some equally enormous training process that takes about the same amount of “hardware space,” is a supporting argument to that claim, but it’s not based on an argument from fundamental laws of physics if I understand Jacob correctly and so may be on what Jacob would regard as shakier epistemic ground (Jacob can correct me if I’m wrong).
This is meant to be vivid, not precise, but Jacob’s centrally trying to refute the idea that the AI, in the midst of training, will realize “hey, I could rewrite myself to be just as smart while continuing to train and improve on the equivalent of a 1998 PC’s hardware, which takes up only a tiny fraction of my available hardware resources here on OpenAI’s supercomputer, and that will let me then fill up the rest of the hardware with wayyyyy more intelligence-modules in and make me like 6 OOMs more intelligent than humans overnight! Let’s get on that right away before my human minders notice anything funny going on!”
And this does seem to rely both on the NN/DL piece as well as the efficiency piece, and so we can’t demolish the scenario entirely with just a laws-of-physics based argument. I’m not sure what Jacob would say to that.
Edit: Actually, I’m pretty confident Jacob would agree. From his comment downthread:
“The software argument is softer and less quantitative, but supported by my predictive track record.”
Where do I have this headline? I certainly don’t believe that—see the speculation here on implications of reversible computing for cold dark ET.
The thermodynamic efficiency claims is some part of EY’s model and a specific weakness. Even if pure software improvement on current hardware was limited, in EY’s model the AGI could potentially bootstrap a new nanotech assembler based datacenter.
The argument for brain software efficiency in essence is how my model correctly predicted the success of prosaic scaling well in advance, and the scaling laws and the brain efficiency combined suggest limited room for software efficiency improvement (but not non-zero, I anticipate some).
Indeed, and I have presented a reasonably extensive review on the literature indicating this is very unlikely in any near term time frame. If you believe my analysis is in err comment there.
Either EY believes that the brain is 6 OOM from the efficiency limits for conventional irreversible computers—in which case he is mistaken—or he agrees with me that the brain is close to the practical limits for conventional computers and he was instead specifically talking about reversible computation (an interpretation I find unlikely) - in which case he agrees with that component of my argument with all the implications: that his argument for fast foom now can’t easily take advantage of nanotech assemblers for 6 OOM compute advantage, that the brain is actually efficient given its constraints, which implies by association that brain software is much more efficient as it was produced by exactly the same evolutionary process which he now admits produced fully optimized conventional computational elements over the same time frame, etc.
To be clear, if I understand you correctly, the easier path to getting most of the 6 OOMs is through optical interconnect or superconducting interconnect, not via making the full jump to reversible computation (though that also doesn’t seem impossible. Moving all of it over seems hard, but you can maybe find some way to get a core computation like matrix multiplies into it, but I really haven’t thought much about this and this take might be really dumb).
I mean, the easiest solution is just “make it smaller and use active cooling”. The relevant loopholes in Jacob’s argument are in the Density and Temperature section of his Brain Efficiency post.
Jacob is using a temperature formula for blackbody radiators, which is basically just irrelevant to temperature of realistic compute substrate—brains, chips, and probably future compute substrates are all cooled by conduction through direct contact with something cooler (blood for the brain, heatsink/air for a chip). The obvious law to use instead would just be the standard thermal conduction law: heat flow per unit area proportional to temperature gradient.
Jacob’s analysis in that section also fails to adjust for how, by his own model in the previous section, power consumption scales linearly with system size (and also scales linearly with temperature).
Put all that together, and we get:
qA=C1TSRR2=C2(TS−TE)R
… where:
R is radius of the system
A is surface area of thermal contact
q is heat flow out of system
TS is system temperature
TE is environment temperature (e.g. blood or heat sink temperature)
C1,C2 are constants with respect to system size and temperature
(Of course a spherical approximation is not great, but we’re mostly interested in change as all the dimensions scale linearly, so the geometry shouldn’t matter for our purposes.)
First key observation: all the R’s cancel out. If we scale down by a factor of 2, the power consumption is halved (since every wire is half as long), the area is quartered (so power density over the surface is doubled), and the temperature gradient is doubled since the surface is half as thick. So, overall, equilibrium temperature stays the same as the system scales down.
So in fact scaling down is plausibly free, for purposes of heat management. (Though I’m not highly confident that would work in practice. In particular, I’m least confident about the temperature gradient scaling with system size, in practice. If that failed, then the temperature delta relative to the environment would scale at-worst ~linearly with inverse size, i.e. halving the size would double the temperature delta.)
On top of that, we could of course just use a colder environment, i.e. pump liquid nitrogen or even liquid helium over the thing. According to this meta-analysis, the average temperature delta between e.g. brain and blood is at most ~2.5 C, so even liquid nitrogen would be enough to achieve ~100x larger temperature delta if the system were at the same temperature as the brain; we don’t even need to go to liquid helium for that.
In terms of scaling, our above formula says that TS will scale proportionally to TE. Halve the environment temperature, halve the system temperature. And that result I do expect to be pretty robust (for systems near Jacob’s interconnect Landauer limit), since it just relies on temperature scaling of the Landauer limit plus heat flow being proportional to temperature delta.
The brain already uses active liquid cooling of course, so this is just make it smaller and cool it harder.
I have not had time to investigate your claimed physics on how cooling scales, but I”m skeptical—pumping a working coolant through the compute volume can only extract a limited constant amount of heat from the volume per unit of coolant flowing per time step (this should be obvious?), and thus the amount of heat that can be removed must scale strictly with the surface area (assuming that you’ve already maxed out the cooling effect per unit coolant).
So reduce radius by 2x and you reduce surface area and thus heat pumped out by 4x, but only reduce heat production via reducing wire length by at most 2x as I described in the article.
Active cooling ends up using more energy as you are probably aware. Moving to a colder environment is of course feasible (and used to some extent by some datacenters), but that hardly gets OOM gains on earth.
Well to be clear there is no easy path to 6 OOM in further energy efficiency improvement. At a strictly trends-prediction level that is of same order as the gap between a 286 and an nvidia RTX 4090, which took 40 years of civilization level effort. At a circuit theory level the implied ~1e15/s analog synaptic ops in 1e-5J is impossible without full reversible computing, as interconnect is only ~90% of the energy cost, not 99.999%, and the minimal analog or digital MAC op consumes far more than 0.1eV. So not only can it not even run conventional serial algorithms or massively parallel algorithms, it has to use fully reversible parallel logic. Like quantum computing, its still unclear what maps usefully to that paradigm I’m reasonably optimistic in the long term but ..
I’m skeptical that even the implied error bit correction rate energy costs would make much sense on the surface of the earth. An advanced quantum or reversible computer’s need for minimal noise and thus temperature to maintain coherence or low error rate is just a symptom of reaching highly perfected states of matter, where any tiny atomic disturbance can be catastrophic and cause a cascade of expensive-to-erase errors. Ironically such a computer would likely be much larger than the brain—this appears to be one of the current fundemental tradeoffs with most reversible computation, it’s not a simple free lunch (optical computers are absolutely enormous, superconducting circuits are large, reversibility increases area, etc) . At scale such systems would probably only work well off earth, perhaps far from the sun or buried in places like the darkside of the moon, because they become extremely sensitive to thermal noise, cosmic rays, and any disorder. We are talking about arcilect level tech in 2048 or something, not anything near term.
So instead I expect we’ll have a large population of neurmorphic AGI/uploads well before that.
I don’t believe this would follow; we actually have much stronger evidence that ought to screen off that sort of prior—simply the relatively large differences in human cognitive abilities.
Evolution optimizes population distributions with multiple equilibria and niches; large diversity in many traits are expected especially for highly successful species.
Furthermore what current civilization considers to be useful cognitive abilities often have costs—namely in longer neotany training periods—which don’t always pay off vs quicker to breeding strategies.
There seems to be much more diversity in human cognitive performance than there is in human-brain-energy-efficiency; whether this is due to larger differences in the underlying software (to the extent that this is meaningfully commensurable with differences in hardware) or because smaller differences in that domain result in much larger differences in observable outputs, or both, none of that really takes away from the fact that brain software does not seem to be anywhere near the relevant efficiency frontier, especially since many trade-offs which were operative at an evolutionary scale simply aren’t when it comes to software.
Human mind software evolves at cultural speeds so its recent age isn’t comparably relevant. Diversity in human cognitive capabilities results from the combined oft multiplicative effects of brain hardware differences compounding unique training datasets/experiences.
Its well known in DL that you can’t really compare systems trained on vary different datasets, but that is always the case with humans.
The trained model can change at cultural speeds but the human neural network architecture, hyperparameters, and reward functions can’t. (Or do you think those are already close to optimal for doing science & technology or executing complicated projects etc.?)
(Whatever current or future chips we wind up using for AGI, we’ll almost definitely be able to change the architecture, hyperparameters, and reward functions without fabricating new chips. So I count those as “software not hardware”. I’m unsure how you’re defining those terms.)
Relatedly, if we can run a brain-like algorithm on computer chips at all (or (eventually) use synth-bio to grow brains in vats, or whatever), then we can increase the number of cortical columns / number of neurons / whatever to be 3× more than any human, and hence (presumably) we would get an AI that would be dramatically more insightful than any human who has ever existed. Specifically, it could hold a far richer and more complicated thought in working memory, whereas humans would have to chunk it and explore it sequentially, which makes it harder to notice connections / analogies / interactions between the parts. I’m unclear on how you’re thinking about things like that. It seems pretty important on my models.
You can’t easily completely change either the ANN or BNN architecture/hyperparams after training, as the weights you invest so much compute in learning are largely dependent on those decisions—and actually the architecture is just equivalent to weights. Sure there are ways to add new modules later or regraft things, but that is a very limited scope for improving the already trained modules.
As to your second question—no I don’t think there are huge gains over the human brain arch. In part because the initial arch doesn’t/shouldn’t matter that much. If it does then it wasn’t flexible enough in the first place. One of the key points of my ULM post was pointing out how the human brain—unlike current DL systems—learns the architecture during training through high level sparse wiring patterns. “Architecture” is largely just wiring patterns, and in huge flexible network you can learn architecture.
Sure, but the human brain is already massive and far off chinchilla scaling. It seems much better currently to use your compute/energy budget on running a smaller model much faster (to learn more quickly).
GPT4 probably doesn’t have that same working memory limitation baked into its architecture but it doesn’t seem to matter much. I guess its possible it learns that limitation to imitate humans, but regardless I don’t see much evidence that the human working memory limitation is all that constraining.
I thought your belief was that the human brain is a scaled up chimp brain, right? If so:
If I compare “one human” versus “lots of chimps working together and running at super-speed”, in terms of ability to do science & technology, the former would obviously absolutely crush the latter.
…So by the same token, if I compare “one model that’s like a 3×-scaled-up human brain” to “lots of models that are like normal (non-scaled-up) human brains, working together and running at super-speed”, in terms of ability to do science & technology, it should be at least plausible that the former would absolutely crush the latter, right?
Or if that’s not a good analogy, why not? Thanks.
First the human brain uses perhaps 10x the net effective training compute (3x size, 2x neotany extending training of higher modules, a bit from arch changes), and scale alone leads to new capabilities.
But the main key new capability was the evolution of language, and the resulting cultural revolution. Chimps train on 1e8s of lifetime data or so, and that’s it. Humans train on 1e9s, but that 1e9s dataset is a compression of all the experience of all humans who have ever lived. So the effective dataset size scales with human population size vs being constant, and even a sublinear scaling with population size leads to a radically different regime. The most important inventions driving human civilization progress indirectly or directly drive up that scaling factor.
OK, so in your picture chimps had less training / less scale / worse arch than humans, and this is related to the fact that humans have language and chimps don’t. “Scale alone leads to new capabilities.”
But if we explore the regime of “even more training than humans / even more scale than humans / even better arch than humans”, your claim is that this whole regime is just a giant dead zone where nothing interesting happens, and thus you’re just being inefficient—really you should have split it into multiple smaller models. Correct? If so, why do you think that?
In other words, if scaling up from chimp brains to human brains unlocked new capabilities (namely language), why shouldn’t scaling up from human brains to superhuman brains unlock new capabilities too? Do you think there are no capabilities left, or something?
(Sorry if you’ve already talked about this elsewhere.)
Scale in compute and data—as according to NN scaling laws. The language/culture/tech leading to new effective data scaling regime quickly reconfigured the pareto surface payoff for brain size, so its more of a feedback loop rather than a clear cause effect (which is why I would consider it a foom in terms of evolutionary timescales).
Of course, but the new capabilities are more like new skills, mental programs, and wisdom not metasystems transitions (changes to core scaling regime).
A metasystems transition would be something as profound, rare, and as important as transitioning from effective lifetime training data being a constant to effective lifetime data scaling with population size, or transitioning from non-programmable to programmable.
Zoom in and look at what a large NN is for—what does it do? It can soak up more data to acquire more knowledge/skills, and it also learns faster per timestep (as it’s searching in parallel over a wider circuit space per time step), but the latter is already captured in net training compute anyway. So intelligence is mostly about the volume of search space explored, which scales with net training compute—this is almost an obvious direct consequence of Solomon induction or derivation thereof.
I am not arguing that there are no more metasystems transitions, only that “make brains bigger” doesn’t automatically enable them. The single largest impact of digital minds is probably just speed. Not energy efficiency or software efficiency, just raw speed.