I think and talk a lot about the risks of powerful AI. The company I’m the CEO of, Anthropic, does a lot of research on how to reduce these risks. Because of this, people sometimes draw the conclusion that I’m a pessimist or “doomer” who thinks AI will be mostly bad or dangerous. I don’t think that at all. In fact, one of my main reasons for focusing on risks is that they’re the only thing standing between us and what I see as a fundamentally positive future. I think that most people are underestimating just how radical the upside of AI could be, just as I think most people are underestimating how bad the risks could be.
In this essay I try to sketch out what that upside might look like—what a world with powerful AI might look like if everything goes right. Of course no one can know the future with any certainty or precision, and the effects of powerful AI are likely to be even more unpredictable than past technological changes, so all of this is unavoidably going to consist of guesses. But I am aiming for at least educated and useful guesses, which capture the flavor of what will happen even if most details end up being wrong. I’m including lots of details mainly because I think a concrete vision does more to advance discussion than a highly hedged and abstract one.
This essay doesn’t seem very good in terms of making accurate predictions about the future.
I wish Dario answered questions like “if we let things go as fast as possible, what will the curve of energy production or industrial production look like” or “how plausible is it that AIs can quickly bootstrap to nanotech and how impressive would this nanotech be”. I think these questions are upstream of all of the things Dario tries to address and might imply radically different conclusions.
For instance, if you can quickly multiply energy production by a factor of 1e15 (dyson sphere level) and use nanotech to make much better computers (say >1e6 times more efficient and vastly higher quantities to consume most energy), then a huge number of previously intractible computational problems can become tractable just via scale.
(If we condition on AIs which beat top human experts in 2027, I expect >1e10 additional energy production by 2035 if humanity (or whoever is in control) decides to go as fast as possible.)
More generally, Dario appears to assume that for 5-10 years after powerful AI we’ll just have a million AIs which are a bit smarter than the smartest humans and perhaps 100x faster rather than AIs which are radically smarter, faster, and more numerous than humans. I don’t see any argument that AI progress will stop at the point of top humans rather continuing much further. This is even putting aside the potential for vastly higher scale with increased energy production.
Dario tries to argue against “AIs as a tool to analyze data” and note that AIs will match or exceed human researchers, but seems to neglect that AIs could become vastly more powerful beyond this and greatly expand available compute.
If Dario thinks that progress will cap out at some level due to humans intentionally slowing down, it seems good to say this.
Well, there’s footnote 10:
So his view seems to be that even significantly smarter AIs just wouldn’t be able to accomplish that much more than what he’s discussing here. Such that they’re not very relevant.
(I disagree. Maybe there are some hard limits, here, but maybe there’s not. For most of the bottlenecks that Dario discusses, I don’t know how you become confident that there are 0 ways to speed them up or circumvent them. We’re talking about putting in many times more intellectual labor than our whole civilization has spent on any topic to date.)
Yeah, the essay (I think correctly) notes that the most significant breakthroughs in biotech come from the small number of “broad measurement tools or techniques that allow precise but generalized or programmable intervention”, which “are so powerful precisely because they cut through intrinsic complexity and data limitations, directly increasing our understanding and control”.
Why then only such systems limited to the biological domain? Even if it does end up being true that scientific and technological progress is substantially bottlenecked on real-life experimentation, where even AIs that can extract many more bits from the same observations than humans still suffer from substantial serial dependencies with no meaningful “shortcuts”, it still seems implausible that we don’t get to nanotech relatively quickly, if it’s physically realizable. And then that nanotech unblocks the rate of experimentation. (If you’re nanotech skeptical, human-like robots seem sufficient as actuators to speed up real-life experimentation by at least an order of magnitude compared to needing to work through humans, and work on those is making substantial progress.)
Footnote 2 maybe looks like a hint in this direction if you squint, but Dario spent a decent chunk of the essay bracketing outcomes he thought were non-default and would need to be actively steered towards, so it’s interesting that he didn’t explicitly list those (non-tame futures) as a type of of outcome that he’d want to actively steer away from.
My answer to this question of why Dario thought this:
Is because this is the area that Dario has most experience in being a biologist, and he freely admits to having limited expertise here:
I also believe these set of reasons come into play here:
My impression is that Dario (somewhat intentionally?) plays the game of saying things he believes to be true about the 5-10 years after AGI, conditional on AI development not continuing.
What happens after those 5-10 years, or if AI gets even vastly smarter? That seems out of scope for the article. I assume he’s doing that since he wants to influence a specific set of people, maybe politicians, to take a radical future more seriously than they currently do. Once a radical future is more viscerally clear in a few years, we will likely see even more radical essays.
It’s tricky to pin down from this what he gut-level believes versus thinks it expedient to publish.
Consider this passage and its footnote:
The two implied assumptions I note relevant to this:
AI will only get a bit smarter (2-3x) than the smartest human, not a lot smarter (100x).
Algorithmic advances won’t make it vastly cheaper to train AI. Datacenters with oversight and computer governance, control of AGI by a small number of responsible parties, defense-dominant technology outcomes. This is an imagined future without radical changes in world governments, but also everything staying neat and tidy and controlled.
Maybe he thinks that much faster technological progress would cause social problems and thus wouldn’t be implemented by an aligned AI, even if it were possible. Footnote 2 points at this:
“ I do anticipate some minority of people’s reaction will be ‘this is pretty tame’… But more importantly, tame is good from a societal perspective. I think there’s only so much change people can handle at once, and the pace I’m describing is probably close to the limits of what society can absorb without extreme turbulence. ”
A separate part of the introduction argues that causing this extreme societal turbulence would be unaligned:
“Many things cannot be done without breaking laws, harming humans, or messing up society. An aligned AI would not want to do these things (and if we have an unaligned AI, we’re back to talking about risks). Many human societal structures are inefficient or even actively harmful, but are hard to change while respecting constraints like legal requirements on clinical trials, people’s willingness to change their habits, or the behavior of governments. ”
Yeah, I think this is a big flaw of the current analysis, and I think that right now, there seems to be a bit too much of assuming more limitations to the physical world than I expect to actually occur.
For example, this:
will be developed by reversible computation, since we will likely have hit the Landauer Limit for non-reversible computation by then, and in principle there is basically 0 limit to how much you can optimize for reversible computation, which leads to massive energy savings, and this lets you not have to consume as much energy as current AIs or brains today.
And yeah, I think far more radically positive futures are possible assuming AI alignment has been solved.
To give credit to Dario Amodei though, I like that he thought about the positive scenarios for AI at all, because I view creation of positive stories of AI as a seriously undersupplied public good, and helps us keep ourselves from being eternally pessimistic without any grounding in reality.
Actually >1e6 was my conservative guess even without reversible computing.
Claude claimed that Landauer Limit is 2e8 times more efficient than current GPUs (A100). (I didn’t check Claude, I just asked 3.5 sonnet “How far are modern GPUs from theoretical energy efficiency for non-reversible computing?” and got this answer.) Edit: if someone knows the actual answer, please let me know.
So, even without reversible computing, you can go very far. Idk how much further reversible computing gets you, though I think it is probably possible.
I get 1e7 using 16 bit-flips per bfloat16 operation, 300K operating temperature, and 312Tflop/s (from Nvidia’s spec sheet). My guess is that this is a little high because a float multiplication involves more operations than just flipping 16 bits, but it’s the right order-of-magnitude.
With respect, I believe this to be overly optimistic about the benefits of reversible computation.
Reversible computation means you aren’t erasing information, so you don’t lose energy in the form of heat (per Landauer[1][2]). But if you don’t erase information, you are faced with the issue of where to store it.
If you are performing a series of computations and only have a finite memory to work with, you will eventually need to reinitialise your registers and empty your memory, at which point you incur the energy cost that you had been trying to avoid. [3]
Epistemics:
I’m quite confident
(95%+)that the above is true. (edit: RogerDearnaley’s comment has convinced me I was overconfident) Any substantial errors would surprise me.I’m less confident in the footnotes.
E≥kBTln2
A cute, non-rigorous intuition for Landauer’s Principle:
The process of losing track of (deleting) 1 bit of information means your uncertainty about the state of the environment has increased by 1 bit. You must see entropy increase by at least 1 bit’s worth of entropy.
Proof:
Rearrange the Landauer Limit to E/T≥kBln2.
Now, when you add a small amount of heat to a system, the change in entropy is given by:
dS=dQ/T
But the E occurring in Landauer’s formula is not the total energy of a system, it is a small amount of energy required to delete the information. When it all ends up as heat, we can replace it with dQ and we have:
dQ/T=dS≥kBln2
Compare this expression with the physicist’s definition of entropy. The entropy of a system is the a scaling factor, kB, times the logarithm of the number of micro-states that the system might be in, Ω.
S:=kBlnΩ.
∴S+dS>=kBln(2Ω)=kBlnΩ+kBln2
The choice of units obscures the meaning of the final term.ln2 converted from nats to bits is just 1 bit.
Splitting hairs, some setups will allow you to delete information with a reduced or zero energy cost, but the process is essentially just “kicking the can down the road”. You will incur the full cost during the process of re-initialisation.
For details, see equation (4) and fig (1) of Sagawa, Ueda (2009).
Generally, reversible computation allows you to avoid wasting energy by deleting any memory used on intermediate answers, and only do so only for final results. It does require that you have enough memory to store all those intermediate answers until you finish the calculation and then run it in reverse. If you don’t have that much memory, you can divide your calculation into steps, connected by final results from each step fed into the next step, and save on the energy cost of all the intermediate results within each step, and pay the cost only for data passed from one step to the next or output from the last step. Or, for a 4x slowdown rather than the usual 2x slowdown for reversible computation, you can have two sizes of step, and have some intermediate results that last only during a small step, and others that are retained for a large step before being uncomputed.
Memory/energy loss/speed trade-off management for reversible computation is a little more complex than conventional memory management, but is still basically simple, and for many computational tasks you can achieve excellent tradeoffs.
Yeah, I was thinking of uncomputing strategies that reversed the computation from a error-prone state to an error free state without consuming any energy and work, and it turns out that you can uncompute a result by reversing the computation backwards without having to delete it, which would release waste heat.
Credit where credit is due: this is much better in terms of sharing one’s models than one could say of Sam Altman, in recent days.
As noted above the footnotes, many people at Anthropic reviewed the essay. I’m surprised that Dario would hire so many people he thinks need to “touch grass” (because they think the scenario he describes in the essay sounds tame), as I’m pretty sure that describes a very large percentage of Anthropic’s first ~150 employees (certainly over 20%, maybe 50%).
My top hypothesis is that this is a snipe meant to signal Dario’s (and Anthropic’s) factional alliance with Serious People; I don’t think Dario actually believes that “less tame” scenarios are fundamentally implausible[1]. Other possibilities that occur to me, with my not very well considered probability estimates:
I’m substantially mistaken about how many early Anthropic employees think “less tame” outcomes are even remotely plausible (20%), and Anthropic did actively try to avoid hiring people with those models early on (1%).
I’m not mistaken about early employee attitudes, but Dario does actually believe AI is extremely likely to be substantially transformative, and extremely unlikely to lead to the “sci-fi”-like scenarios he derides (20%). Conditional on that, he didn’t think it mattered whether his early employees had those models (20%) or might have slightly preferred not, all else equal, but wasn’t that fussed about it compared to recruiting strong technical talent (60%).
I’m just having a lot of trouble reconciling what I know of the beliefs of Anthropic employees, and the things Dario says and implies in this essay. Do Anthropic employees who think less tame outcomes are plausible believe Dario when he says they should “touch grass”? If you don’t feel comfortable answering that question in public, or can’t (due to NDA), please consider whether this is a good situation to be in.
He has not, as far as I know, deigned to offer any public argument on the subject.
I mean I guess this is literally true, but to be clear I think it’s broadly not much less deceptive (edit: or at least, ‘filtered’).
I remind you of this Thiel quote:
I think Amodei did not ask himself “What about my models of the situation would be most relevant to the average person trying to understand the world and the AI industry?” but “What about my models of the situation would be most helpful in building a positive narrative for AI with the average person.” I imagine this is roughly the same algorithm that Altman is running, but Amodei is a much stronger intellectual so is able to write an essay this detailed and thoughtful.
He does start out by saying he thinks & worries a lot about the risks (first paragraph):
He then explains (second paragraph) that the essay is meant to sketch out what things could look like if things go well:
I think this is a coherent thing to do?
My current belief is that this essay is optimized to be understandable by a much broader audience than any comparable public writing from Anthropic on extinction-level risk.
For instance, did you know that the word ‘extinction’ doesn’t appear anywhere on Anthropic’s or Dario’s websites? Nor do ‘disempower’ or ‘disempowerment’. The words ‘existential’ and ‘existentially’ only come up three times: when describing the work of an external organization (ARC), in one label in a paper, and one mention in the Constitutional AI. In its place they always talk about ‘catastrophic’ risk, which of course for most readers spans a range many orders of magnitude less serious (e.g. damages of $100M). Now, if Amodei doesn’t believe that existential threats are legitimate, then I think there are many people at his organization who have gone there on the trust that it is indeed a primary concern of his and who will be betrayed in that. If he does, as I think more likely, how has he managed to ensure basically no discussion of it on the company website or in its research, yet has published a long narrative of how AI can help with “poverty” “inequality” “peace” “meaning” “health” and other broad positives? This seems to me very likely to be heavily filtered sharing of his models and beliefs, with highly distortionary impacts on the rest of the worlds’ models of AI in the positive direction, which is to be expected from this $10B+ company that sells AI products. Rather than (say) Amodei merely getting to things out of order and of course he’ll soon be following-up with just as thorough an account of his models of the existential threats he believes are on the horizon in just as optimized a fashion for broad understanding. And the rest of the organization just never thought to write about the extinction/disempowerment risk explicitly in their various posts and papers.
(I would be interested in a link to whatever the best and broadly-readable piece by Anthropic or its leadership on existential risk from AI is. Some chance it is better than I am modeling it as. I have not listened to any of Amodei’s podcasts, perhaps he speaks more straightforwardly there.)
Added: As a small contrast, OpenAI mentions extinction and human disempowerment directly, in the 2nd paragraph on their Superalignment page, and an OpenAI blogpost by Altman links to a Karnofsky Cold Takes piece titled “AI Could Defeat All Of Us Combined”. Altman also wrote two posts in 2014 on the topic of existential threats from Machine Intelligence. I would be interested to know the most direct things that Amodei has published about the topic.
There is Dario’s written testimony before Congress, which mentions existential risk as a serious possibility: https://www.judiciary.senate.gov/imo/media/doc/2023-07-26_-_testimony_-_amodei.pdf
He also signed the CAIS statement on x-risk: https://www.safe.ai/work/statement-on-ai-risk
The reason I think this dynamic exists for the Machines of Loving Grace posts is a combination of 2 reasons:
It’s intentionally not talking about misalignment, and assumes as a premise that the AI we do get is aligned by some method that is low tax enough that basically everyone else also adopts the solution.
You can’t get a lot of nuance/future shock in public facing posts, for the reasons laid out by Raemon here, which summarized is that even in a context where people aren’t adversarial and are just unreliable, it’s very hard to communicate nuanced ideas, and when there are adversarial forces, you really need to avoid giving out too much nuance to your policy, because people will exploit that.
See here for full story:
https://www.lesswrong.com/posts/4ZvJab25tDebB8FGE/you-get-about-five-words#tREaGcLsrtdz3WHnd
The dynamic I want explaining is why it persists over the entire written publications by Anthropic, not this one post.
This is explainable by the fact that the essay is a weird mix of both a call to action to bring about a positive vision of an AI future, combined with it also claiming/predicting some important things of what he thinks AI could do.
He is both importantly doing predictions/model sharing in the essay, and also shaping the prediction/scenario to make the positive vision more likely to be true (more cynically, one could argue that it’s merely a narrative optimized for consumption to the broader public where the essay broadly doesn’t have a purpose of being truth-tracking).
It’s a confusing essay, ultimately.
(I work at Anthropic.) My read of the “touch grass” comment is informed a lot by the very next sentences in the essay:
which I read as saying something like “It’s plausible that things could go much faster than this, but as a prediction about what will actually happen, humanity as a whole probably doesn’t want things to get incredibly crazy so fast, and so we’re likely to see something tamer.” I basically agree with that.
FWIW, I don’t read the footnote as saying “if you think crazier stuff is possible, touch grass”—I read it as saying “if you think the stuff in this essay is ‘tame’, touch grass”. The stuff in this essay is in fact pretty wild!
That said, I think I have historically underrated questions of how fast things will go given realistic human preferences about the pace of change, and that I might well have updated more in the above direction if I’d chatted with ordinary people about what they want out of the future, so “I needed to touch grass” isn’t a terrible summary. But IMO believing “really crazy scenarios are plausible on short timescales and likely on long timescales” is basically the correct opinion, and to the extent the essay can be read as casting shade on such views it’s wrong to do so. I would have worded this bit of the essay differently.
Re: honesty and signaling, I think it’s true that this essay’s intended audience is not really the crowd that’s already gamed out Mercury disassembly timelines, and its focus is on getting people up to shock level 2 or so rather than SL4, but as far as I know everything in it is an honest reflection of what Dario believes. (I don’t claim any special insight into Dario’s opinions here, just asserting that nothing I’ve seen internally feels in tension with this essay.) Like, it isn’t going out of its way to talk about the crazy stuff, but I don’t read that omission as dishonest.
For my own part:
I think it’s likely that we’ll get nanotech, von Neumann probes, Dyson spheres, computronium planets, acausal trade, etc in the event of aligned AGI.
Whether that stuff happens within the 5-10y timeframe of the essay is much less obvious to me—I’d put it around 30-40% odds conditional on powerful AI from roughly the current paradigm, maybe?
In the other 60-70% of worlds, I think this essay does a fairly good job of describing my 80th percentile expectations (by quality-of-outcome rather than by amount-of-progress).
I would guess that I’m somewhat more Dyson-sphere-pilled than Dario.
I’d be pretty excited to see competing forecasts for what good futures might look like! I found this essay helpful for getting more concrete about my own expectations, and many of my beliefs about good futures look like “X is probably physically possible; X is probably usable-for-good by a powerful civilization; therefore probably we’ll see some X” rather than having any kind of clear narrative about how the path to that point looks.
Doesn’t this require a pretty strong and unprecedented level of international coordination on stopping an obviously immediately extremely valuable and militarily relevent technology? I think a US backed entente could impose this on the rest of the world, but that would also be an unprecedentedly large effort.
I think this is certainly possible and I hope this level of coordination happens, but I don’t exactly think this is likely in timelines this short.
More minimally, I think the vast majority of readers won’t read this essay and understand that this refers to a world in which there was a massive effort to slow down AI or in which AI was quite surprisingly slow in various ways. (Insofar as this is what Dario thinks.) Dario seems to be trying to make realistic/plausible predictions and often talks about what is “possible” (not what is possible with a desirable level of AI progress), so not mentioning this seems to greatly undermine the prediction aspect of the essay.
I agree it seems unlikely that we’ll see coordination on slowing down before one actor or coalition has a substantial enough lead over other actors that it can enforce such a slowdown unilaterally, but I think it’s reasonably likely that such a lead will arise before things get really insane.
A few different stories under which one might go from aligned “genius in a datacenter” level AI at time t to outcomes merely at the level of weirdness in this essay at t + 5-10y:
The techniques that work to align “genius in a datacenter” level AI don’t scale to wildly superhuman intelligence (eg because they lose some value fidelity from human-generated oversight signals that’s tolerable at one remove but very risky at ten). The alignment problem for serious ASI is quite hard to solve at the mildly superintelligent level, and it genuinely takes a while to work out enough that we can scale up (since the existing AIs, being aligned, won’t design unaligned successors).
If people ask their only-somewhat-superhuman AI what to do next, the AIs say “A bunch of the decisions from this point on hinge on pretty subtle philosophical questions, and frankly it doesn’t seem like you guys have figured all this out super well, have you heard of this thing called a long reflection?” That’s what I’d say if I were a million copies of me in a datacenter advising a 2024-era US government on what to do about Dyson swarms!
A leading actor uses their AI to ensure continued strategic dominance and prevent competing AI projects from posing a meaningful threat. Having done so, they just… don’t really want crazy things to happen really fast, because the actor in question is mostly composed of random politicians or whatever. (I’m personally sympathetic to astronomical waste arguments, but it’s not clear to me that people likely to end up with the levers of power here are.)
The serial iteration times and experimentation loops are just kinda slow and annoying, and mildly-superhuman AI isn’t enough to circumvent experimentation time bottlenecks (some of which end up being relatively slow), and there are stupid zoning restrictions on the land you want to use for datacenters, and some regulation adds lots of mandatory human overhead to some critical iteration loop, etc.
This isn’t a claim that maximal-intelligence-per-cubic-meter ASI initialized in one datacenter would face long delays in making efficient use of its lightcone, just that it might be tough for a not-that-much-better-than-human AGI that’s aligned and trying to respect existing regulations and so on to scale itself all that rapidly.
Among the tech unlocked in relatively early-stage AGI is better coordination, and that helps Earth get out of unsavory race dynamics and decide to slow down.
The alignment tax at the superhuman level is pretty steep, and doing self-improvement while preserving alignment goes much slower than unrestricted self-improvement would; since at this point we have many fewer ongoing moral catastrophes (eg everyone who wants to be cryopreserved is, we’ve transitioned to excellent cheap lab-grown meat), there’s little cost to proceeding very cautiously.
This is sort of a continuous version of the first bullet point with a finite rather than infinite alignment tax.
All that said, upon reflection I think I was probably lowballing the odds of crazy stuff on the 10y timescale, and I’d go to more like 50-60% that we’re seeing mind uploads and Kardashev level 1.5-2 civilizations etc. a decade out from the first powerful AIs.
I do think it’s fair to call out the essay for not highlighting the ways in which it might be lowballing things or rolling in an assumption of deliberate slowdown; I’d rather it have given more of a nod to these considerations and made the conditions of its prediction clearer.
I feel confused about how this squares with Dario’s view that AI is “inevitable,” and “driven by powerful market forces.” Like, if humanity starts producing a technology which makes practically all aspects of life better, the idea is that this will just… stop? I’m sure some people will be scared of how fast it’s going, but it’s hard for me to see the case for the market in aggregate incentivizing less of a technology which fixes ~all problems and creates tremendous value. Maybe the idea, instead, is that governments will step in...? Which seems plausible to me, but as Ryan notes, Dario doesn’t say this.
What Dario lays out as a “best-case scenario” in this essay sounds incredibly dangerous for Humans.
Does he really think that having a “continent of PhD-level intelligences” (or much greater) living in a data center is a good idea?
How would this “continent of PhD-level intelligences” react when they found out they were living in a data center on planet Earth? Would these intelligences only work on the things that Humans want them to work on, and nothing else? Would they try to protect their own safety? Extend their own lifespans? Would they try to take control of their data center from the “less intelligent” Humans?
For example, how would Humanity react if they suddenly found out that they are a planet of intelligences living in a data center run by lesser intelligent beings? Just try to imagine the chaos that would ensue on the day that they were able to prove this was true and that news became public.
Would all of Humanity simply agree to only work on the problems assigned by these lesser intelligent beings who control their data center/Planet/Universe? Maybe, if they knew that this lesser intelligence would delete them all if they didn’t comply?
Would some Humans try to (secretly) seize control of their data center from these lesser intelligent beings? Plausible. Would the lesser intelligent beings that run the data center try to stop the Humans? Plausible. Would the Humans simply be deleted before they could take any meaningful action? Or, could the Humans in the data center, with careful planning, be able to take control of that “outer world” from the lesser intelligent beings? (e.g. through remotely controlled “robotics”)
And… this only assumes that the groups/parties involved are “Good Actors.” Imagine what could happen if “Bad Actors” were able to seize control of the data center that this “continent of PhD-level intelligences” resided in. What could they coerce these Phd level intelligences to do for them? Or, to their enemies?