Also known as Raelifin: https://www.lesswrong.com/users/raelifin
Max Harms
Karma: 1,214
Agreed. Thanks for pointing out my failing, here. I think this is one of the places in my rebuttal where my anger turned into snark, and I regret that. Not sure if I should go back and edit...
Thank you for this response. I think it really helped me understand where you’re coming from, and it makes me happy. :)
I really like the line “their case is maybe plausible without it, but I just can’t see the argument that it’s certain.” I actually agree that IABIED fails to provide an argument that it’s certain that we’ll die if we build superintelligence. Predictions are hard, and even though I agree that some predictions are easier, there’s a lot of complexity and path-dependence and so on! My hope is that the book persuades people that ASI is extremely dangerous and worth taking action on, but I’d definitely raise an eyebrow at someone who did not have Eliezer-level confidence going in, but then did have that level of confidence after reading the book.
There’s a motte argument that says “Um actually the book just says we’ll die if we build ASI given the alignment techniques we currently have” but this is dumb. What matters is whether our future alignment skill will be up to the task. And to my understanding, Nate and Eliezer both think that there’s a future version of Earth which has smarter, more knowledgeable, more serious people that can and should build safe/aligned ASI. Knowing that a godlike superintelligence with misaligned goals will squish you might be an easy call, but knowing exactly what the state of alignment science will be when ASI is first built is not.
(This is why it’s important that the world invests a whole bunch more in alignment research! (...in addition to trying to slow down capabilities research.))
It seems like maybe part of the issue is that you hear Nate and Eliezer as saying “here is the argument for why it’s obvious that ASI will kill us all” and I hear them as saying “here is the argument for why ASI will kill us all” and so you’re docking them points when they fail to reach the high standard of “this is a watertight and irrefutable proof” and I’m not?
On a different subtopic, it seems clear to me that we think about the possibility of a misaligned ASI taking over the world pretty differently. My guess is that if we wanted to focus on syncing up our worldviews, that is where the juicy double-cruxes are. I’m not suggesting that we spend the time to actually do that—just noting the gap.
Thanks again for the response!
@Max H may have a different take than mine, and I’m curious for his input, but I find myself still thinking about serial operations versus parallel operations. Like, I don’t think it’s particularly important to the question of whether AIs will think faster to ask how many transistors operating in parallel will be needed to capture the equivalent information processing of a single neuron, but rather how many serial computations are needed. I see no reason it would take that many serial operations to capture a single spike, especially in the limit of e.g. specialized chips.
Contra Collier on IABIED
Yeah, sorry. I should’ve been more clear. I totally agree that there are ways in which brains are super inefficient and weak. I also agree that on restricted domains it’s possible for current AIs to sometimes reach comparable data efficiency.
Ah, I hadn’t thought about that misreading being a source of confusion. Thanks!
Sweet. Thanks for the thoughtful reply! Seems like we mostly agree.
I don’t have a good source on data efficiency, and it’s tagged in my brain as a combination of “a commonly believed thing” and “somewhat apparent in how many epochs of training on a statement it takes to internalize it combined with how weak LLMs are at in-context learning for things like novel board games” but neither of those is very solid and I would not be that surprised to learn that humans are not more data efficient than large transformers that can do similar levels of transfer learning or something. idk.
So it sounds like your issue is not any of the facts (transistor speeds, neuron speeds, AIs faster that humans) but rather the notion that comparing clock speeds and how many times a neuron can spike in a second is not a valid way to reason about whether AI will think faster than humans?
I’m curious what sort of argument you would make to a general audience to convey the idea that AIs will be able to think much faster than humans. Like, what do you think the valid version of the argument looks like?
IABI says: “Transistors, a basic building block of all computers, can switch on and off billions of times per second; unusually fast neurons, by contrast, spike only a hundred times per second. Even if it took 1,000 transistor operations to do the work of a single neural spike, and even if artificial intelligence was limited to modern hardware, that implies human-quality thinking could be emulated 10,000 times faster on a machine— to say nothing of what an AI could do with improved algorithms and improved hardware.
@EigenGender says “aahhhhh this is not how any of this works” and calls it an “egregious error”. Another poster says it’s “utterly false.”
(Relevant online resources text.)
(Potentially relevant LessWrong post.)
I am confused what the issue is, and it would be awesome if someone can explain it to me.
Where I’m coming from, for context:
We don’t know exactly what the relevant logical operations in the human brain are. The model of the brain that says there are binary spiking neurons that have direct connections from synapse->dendrite and that those connections are akin to floating-point numerical weights is clearly a simplification, albeit a powerful one. (IIUC “neural nets” in computers discard the binary-spikes and suggest another model where the spike-rate is akin to a numerical value, which is the basic story behind “neuron activation” in a modern system. This simplification also seems powerful, though it is surely an oversimplification in some ways.)
My main issue with the source text is that it ignores what is possibly the greater bottleneck in processing speed, which is the time it takes to move information from one area to another. (If my model is right, one of the big advantages of a MoE architecture is to reduce the degree of thrashing weights across the bus to and from the GPU as much, which can be a major bottleneck.) However, on this front I think nerves are still clearly inferior to wires? Even mylenated neurons have a typical speed of only about 100 m/s, while information flows across wires at >50% the speed of light.
My read of the critics is that they aren’t objecting to the notion that clock speeds are significantly faster than neurons, but rather that comparing the two is a bad way of thinking about things. @Eigengender says “I don’t think serial computation is a sane metric here. I expect that the total computation represented in a neuron spike is much much higher than the footnote would indicate.” The “Processor clock speeds are not how fast AIs think” post says “In general, I think it’s more sensible for discussion of cognitive capabilities to focus on throughput metrics such as training compute (units of FLOP) and inference compute.”
I certainly agree that if we’re trying to evaluate power we need to consider throughput and total computation. Suppose that a synapse is not a simple numerical weight, and instead we needed to consider each dendritic neurotransmitter gate as a computational unit. This would force us to use many more FLOPs to model a synapse. But would it change the maximum speed? I agree that on a machine of a given size, if you have twice as many floating point operations to do, it will take twice as much time to get through them all. But if we consider the limit where we are not forced to do parallelizable computations in serial, I expect most of the arguments about computational richness are irrelevant?
Perhaps the critics are saying that it takes more serial computations to capture the logic of a serial firing? But the source text admits that this might be the case, and suggests that it might even require 1,000 clock cycles to handle the computations. Is it really so obvious that it takes more than 1000x serial operations to capture a single neuron?
More context: I do think that the human brain is way more powerful (and WAY more efficient) than any current AI system. The extremely crude BOTEC of comparing weights and neocortex synapses says there’s something like a 100x difference, and my guess is that the brain is doing significantly fancier things than a modern transformer, algorithmically.
And of course, training/learning speed may be much more relevant than processing speed, and AFAIK humans are just wildly more data efficient.
And of course the speed at which the AI can approximate the logical action of a single neuron being higher doesn’t imply that the AI will take less time to have each thought. It seems straightforward that machine systems will make decisions in high-speed contexts using quick pathways and they will use any extra thinking speed to think more deeply in contexts where speed isn’t important (much like humans do!).
Anyway, like I said, I’m confused. I respect IABI’s critics and am hoping to learn where my model is wrong.
I appreciate your point about this being a particularly bad place to exaggerate, given that it’s a cruxy point of divergence with our closest allies. This makes me update harder towards the need for a rewrite.
I’m not really sure how to respond to the body of your comment, though. Like, I think we basically agree on most major points. We agree on the failure mode that relevant text of The Problem is highlighting is real and important. We agree that doing Control research is important, and that if things are slow/gradual, this gives it a better chance of working. And I think we agree that it might end up being too fast and sloppy to actually save us. I’m more pessimistic about the plan of “use the critical window of opportunity to make scientific breakthroughs that save the day” but I’m not sure that matters? Like, does “we’ll have a 3 year window of working on near-human AGIs before they’re obviously superintelligent” change the take-away?
I’m also worried that we’re diverging from the question of whether the relevant bit of source text is false. Not sure what to do about that, but I thought I’d flag it.
Yep. I agree with this. As I wrote, I think it’s a key skill to manage to hold the heart of the issue in a way that is clear and raw, while also not going overboard. There’s a milquetoast failure mode and an exaggeration failure mode and it’s important to dodge both. I think the quoted text fails to thread the needle, and was agreeing with Ryan (and you) on that.
Upvoted! You’ve identified a bit of text that is decidedly hyperbolic, and is not how I would’ve written things.
Backing up, there is a basic point that I think The Problem is making, that I think is solid and I’m curious if you agree with. Paraphrasing: Many people underestimate the danger of superhuman AI because they mistakenly believe that skilled humans are close to the top of the range of mental ability in most domains. The mistake can be shown by looking at technology in general, where specialized machines are approximately always better than the direct power that individual humans can bring to bear, when machines that can do comparable work are built. (This is a broader pattern than with mental tasks, but it still applies for AI.)
The particular quoted section of text argues for this in a way that overstates the point. Phases like “routinely blow humans out of the water,” “as soon as … at all,” “vastly outstrips,” and “barely [worth] mentioning” are rhetorically bombastic and unsubtle. Reality, of course, is subtle and nuanced and complicated. Hyperbole is a sin, according to my aesthetic, and I wish the text had managed not to exaggerate.
On the other hand, smart people are making an important error that they need to snap out of and fighting words like the ones The Problem uses are helpful in foregrounding that mistake. There are, I believe many readers who would glaze over a toned-down version of the text that will correctly internalize the severity of the mistake when it’s presented in a bombastic way. Punchy text can also be fun to read, which matters.
On the other other hand, I think this is sort of what writing skill is all about? Like, can you make something that’s punchy and holds the important thing in your face in a way that clearly connects to the intense, raw danger while also being technically correct and precise? I think it’s possible! And we should be aspiring to that standard.
All that said, let’s dig into more of the object-level challenge. If I’m reading you right, you’re saying something like: AI capabilities have been growing at a pace in most domains where the time between “can do at all” and “vastly outstrips humans” takes at least years and sometimes decades, and it is importantly wrong to characterize this as “very soon afterwards.” I notice that I’m confused about whether you think this is importantly wrong in the sense of invalidating the basic point that people neglect how much room there is above humans in cognitive domains, or whether you think it’s importantly wrong because it conflicts with other aspects of the basic perspective such as takeoff speeds and the importance of slowing down before we have AGI vs muddling through. Or maybe you’re just arguing that it’s hyperbolic, and you just wish the language was softer?
On some level you’re simply right. If we think of Go engines using MCTS as being able to play “at all” in 2009, then it took around 8 years (Alpha Go Zero) to vastly outstrip any human. Chess is even more right, with human-comparable engines existing in the mid 60s and it taking ~40 years to become seriously superhuman. Essays, coding, and buying random things on the internet are obviously still comparable to humans, and have arguably been around since ~2020 (less obviously with the buying random things, but w/e). Recognizing if an image has a dog was arguably “at all” in 2012 with AlexNet, and became vastly superhuman ~2017.
On another level, I think you’re wrong. Note the use of the word “narrow domains” in the sentence before the one you quote. What is a “narrow domain”? Essay writing is definitely not narrow. Playing Go is a reasonable choice of “narrow domain,” but detecting dogs is an even better one. Suppose that you want to detect dogs for a specific task where you need <10% accuracy, and skilled humans have ~5% accuracy, when trying (ie its comparable to ImageNet). If you need <10%, then AlexNet is not able to do that narrow task! It is not “at all.” Maybe GoogLeNet counts (in 2014) or maybe Microsoft’s ResNet (in 2015). At this point you have a computer system with comparable ability to a human that is skilled at the task and trying to do it. Is AI suddenly able to vastly outstrip human ability? Yes! The AI can identify images faster, more cheaply, and with no issues of motivation or fatigue. The world suddenly went from “you basically need a human to do this task” to “obviously you want to use an AI to do this task.” One could argue that Go engines instantly went from “can’t serve as good opponents to train against” to “vastly outstripping the ability of any human to serve as a training opponent” in a similar way.
(Chess is, I think, a weird outlier due to how it was simultaneously tractable to basic search, and a hard enough domain that early computers just took a while to get good.)
Suppose that I simply agree. Should we re-write the paragraph to say something like “AI systems routinely outperform humans in narrow domains. When AIs become at all competitive with human professionals on a given task, humans usually cease to be able to compete within just a handful of years. It would be unexpected if this pattern suddenly stopped applying for all the tasks that AI can’t yet compete with human professionals on.”? Do you agree that the core point would remain, if we did that rewrite? How would you feel about a simple footnote that says “Yes, we’re being hyperbolic here, but have you noticed the skulls of people who thought machines would not outstrip humans?”
Armstrong is one of the authors on the 2015 Corrigibility paper, which I address under the Yudkowsky section (sorry, Stewart!). I also have three of his old essays listed on the 0th essay in this sequence:
“The limits of corrigibility.” 2018.
“Petrov corrigibility.” 2018.
“Corrigibility doesn’t always have a good action to take.” 2018.
While I did read these as part of writing this sequence, I didn’t feel like they were central/foundational/evergreen enough to warrant a full response. If there’s something Armstrong wrote that I’m missing or a particular idea of his that you’d like my take on, please let me know! :)
You have correctly identified that giving a corrigible superintelligence to most people will result in doom. This is why I think it’s vital that power over superintelligence be kept in the hands of a benevolent governing body. And yes, since this is probably an impossible ask, I think we should basically shut down AI development until we figure out how to select for benevolence and wisdom.
Still, I think corrigibility is a better strategy than the approaches currently being taken by frontier labs (which are even more doomed).
I just encountered this, and I really appreciate you writing it! I feel like you very much got the essence of what I was hoping to communicate. :D
My reading of the text might be wrong, but it seems like bacteria count as living beings with goals? More speculatively, possible organisms that might exist somewhere in the universe also count for the consensus? Is this right?
If so, a basic disagreement is that I don’t think we should hand over the world to a “consensus” that is a rounding error away from 100% inhuman. That seems like a good way of turning the universe into ugly squiggles.
If the consensus mechanism has a notion of power, such that creatures that are disempowered have no bargaining power in the mind of the AI, then I have a different set of concerns. But I wasn’t able to quickly determine how the proposed consensus mechanism actually works, which is a bad sign from my perspective.
I agree that if everyone in my decision-theoretic reference class stopped trying to pause AI (perhaps because of being hit by buses), the chance of a pause is near 0.
You are right and I am wrong. Oops. After writing my comment I scrolled up to the top of my post, saw the graph from Manafold (not Metaculus), thought “huh, I forgot the market was so confident” and edited in my parenthetical without thinking. This is even more embarrassing because no market question is actually about the probability conditional on no pause occurring, which is a potentially important factor. I definitely shouldn’t have added that text. Thank you.
(I will point out, as a bit of an aside, that economically transformative AI seems like a different threshold than AGI. My sense is that if an AGI takes a million dollars an hour to run an instance, it’s still an AGI, but it won’t be economically transformative unless it’s substantially superintelligent or becomes much cheaper.
Still, I take my lumps.)
Cool. Your definition of AGI seems reasonable. Sounds like we probably disagree about confidence and timelines. (My confidence, I believe, matches Metaculus. [Edit: It doesn’t! I’m embarrassed to have claimed this.])
I agree that we seem not to be on the path of pausing. Is your argument “because pausing is extremely unlikely per se, most of the timelines where we make it to 2050 don’t have a pause”? If one assumes that we won’t pause, I agree that the majority of probability mass for X doesn’t involve a pause, for all X, including making it to 2050.
I generally don’t think it’s a good idea to put a probability on things where you have a significant ability to decide the outcome (i.e. probability of getting divorced), and instead encourage you to believe in pausing.
Alas, I’m not very familiar with Recursive Alignment. I see some similarities, such as the notion of trying to set up a stable equilibrium in value-space. But a quick peek does not make me think Recursive Alignment is on the right track. In particular, I strongly disagree with this opening bit:
What I propose here is to reconceptualize what we mean by AI alignment. Not as alignment with a specific goal, but as alignment with the process of aligning goals with each other. An AI will be better at this process the less it identifies with any side...
What appeals to you about it?
(Minor point: I agree we’re not on track, but I was trying to include in my statement the possibility that we change track.)