Archive for June, 2022

Steven Pinker and I debate AI scaling!

Monday, June 27th, 2022

Before June 2022 was the month of the possible start of the Second American Civil War, it was the month of a lively debate between Scott Alexander and Gary Marcus about the scaling of large language models, such as GPT-3.  Will GPT-n be able to do all the intellectual work that humans do, in the limit of large n?  If so, should we be impressed?  Terrified?  Should we dismiss these language models as mere “stochastic parrots”?

I was privileged to be part of various email exchanges about those same questions with Steven Pinker, Ernest Davis, Gary Marcus, Douglas Hofstadter, and Scott Alexander.  It’s fair to say that, overall, Pinker, Davis, Marcus, and Hofstadter were more impressed by GPT-3’s blunders, while we Scotts were more impressed by its abilities.  (On the other hand, Hofstadter, more so than Pinker, Davis, or Marcus, said that he’s terrified about how powerful GPT-like systems will become in the future.)

Anyway, at some point Pinker produced an essay setting out his thoughts, and asked whether “either of the Scotts” wanted to share it on our blogs.  Knowing an intellectual scoop when I see one, I answered that I’d be honored to host Steve’s essay—along with my response, along with Steve’s response to that.  To my delight, Steve immediately agreed.  Enjoy!  –SA


Steven Pinker’s Initial Salvo

Will future deep learning models with more parameters and trained on more examples avoid the silly blunders which Gary Marcus and Ernie Davis entrap GPT into making, and render their criticisms obsolete?  And if they keep exposing new blunders in new models, would this just be moving the goalposts?  Either way, what’s at stake?

It depends very much on the question.  There’s the cognitive science question of whether humans think and speak the way GPT-3 and other deep-learning neural network models do.  And there’s the engineering question of whether the way to develop better, humanlike AI is to upscale deep learning models (as opposed to incorporating different mechanisms, like a knowledge database and propositional reasoning).

The questions are, to be sure, related: If a model is incapable of duplicating a human feat like language understanding, it can’t be a good theory of how the human mind works.  Conversely, if a model flubs some task that humans can ace, perhaps it’s because it’s missing some mechanism that powers the human mind.  Still, they’re not the same question: As with airplanes and other machines, an artificial system can duplicate or exceed a natural one but work in a different way.

Apropos the scientific question, I don’t see the Marcus-Davis challenges as benchmarks or long bets that they have to rest their case on.  I see them as scientific probing of an empirical hypothesis, namely whether the human language capacity works like GPT-3.  Its failures of common sense are one form of evidence that the answer is “no,” but there are others—for example, that it needs to be trained on half a trillion words, or about 10,000 years of continuous speech, whereas human children get pretty good after 3 years.  Conversely, it needs no social and perceptual context to make sense of its training set, whereas children do (hearing children of deaf parents don’t learn spoken language from radio and TV).  Another diagnostic is that baby-talk is very different from the output of a partially trained GPT.  Also, humans can generalize their language skill to express their intentions across a wide range of social and environmental contexts, whereas GPT-3 is fundamentally a text extrapolator (a task, incidentally, which humans aren’t particularly good at).  There are surely other empirical probes, limited only by scientific imagination, and it doesn’t make sense in science to set up a single benchmark for an empirical question once and for all.  As we learn more about a phenomenon, and as new theories compete to explain it, we need to develop more sensitive instruments and more clever empirical tests.  That’s what I see Marcus and Davis as doing.

Regarding the second, engineering question of whether scaling up deep-learning models will “get us to Artificial General Intelligence”: I think the question is probably ill-conceived, because I think the concept of “general intelligence” is meaningless.  (I’m not referring to the psychometric variable g, also called “general intelligence,” namely the principal component of correlated variation across IQ subtests.  This is  a variable that aggregates many contributors to the brain’s efficiency such as cortical thickness and neural transmission speed, but it is not a mechanism (just as “horsepower” is a meaningful variable, but it doesn’t explain how cars move.)  I find most characterizations of AGI to be either circular (such as “smarter than humans in every way,” begging the question of what “smarter” means) or mystical—a kind of omniscient, omnipotent, and clairvoyant power to solve any problem.  No logician has ever outlined a normative model of what general intelligence would consist of, and even Turing swapped it out for the problem of fooling an observer, which spawned 70 years of unhelpful reminders of how easy it is to fool an observer.

If we do try to define “intelligence” in terms of mechanism rather than magic, it seems to me it would be something like “the ability to use information to attain a goal in an environment.”  (“Use information” is shorthand for performing computations that embody laws that govern the world, namely logic, cause and effect, and statistical regularities.  “Attain a goal” is shorthand for optimizing the attainment of multiple goals, since different goals trade off.)  Specifying the goal is critical to any definition of intelligence: a given strategy in basketball will be intelligent if you’re trying to win a game and stupid if you’re trying to throw it.  So is the environment: a given strategy can be smart under NBA rules and stupid under college rules.

Since a goal itself is neither intelligent or unintelligent (Hume and all that), but must be exogenously built into a system, and since no physical system has clairvoyance for all the laws of the world it inhabits down to the last butterfly wing-flap, this implies that there are as many intelligences as there are goals and environments.  There will be no omnipotent superintelligence or wonder algorithm (or singularity or AGI or existential threat or foom), just better and better gadgets.

In the case of humans, natural selection has built in multiple goals—comfort, pleasure, reputation, curiosity, power, status, the well-being of loved ones—which may trade off, and are sometimes randomized or inverted in game-theoretic paradoxical tactics.  Not only does all this make psychology hard, but it makes human intelligence a dubious benchmark for artificial systems.  Why would anyone want to emulate human intelligence in an artificial system (any more than a mechanical engineer would want to duplicate a human body, with all its fragility)?  Why not build the best possible autonomous vehicle, or language translator, or dishwasher-emptier, or baby-sitter, or protein-folding predictor?  And who cares whether the best autonomous vehicle driver would be, out of the box, a good baby-sitter?  Only someone who thinks that intelligence is some all-powerful elixir.

Back to GPT-3, DALL-E, LaMDA, and other deep learning models: It seems to me that the question of whether or not they’re taking us closer to “Artificial General Intelligence” (or, heaven help us, “sentience”) is based not on any analysis of what AGI would consist of but on our being gobsmacked by what they can do.  But refuting our intuitions about what a massively trained, massively parameterized network is capable of (and I’ll admit that they refuted mine) should not be confused with a path toward omniscience and omnipotence.  GPT-3 is unquestionably awesome at its designed-in goal of extrapolating text.  But that is not the main goal of human language competence, namely expressing and perceiving intentions.  Indeed, the program is not even set up to input or output intentions, since that would require deep thought about how to represent intentions, which went out of style in AI as the big-data/deep-learning hammer turned every problem into a nail.  That’s why no one is using GPT-3 to answer their email or write an article or legal brief (except to show how well the program can spoof one).

So is Scott Alexander right that every scaled-up GPT-n will avoid the blunders that Marcus and Davis show in GPT-(n-1)?  Perhaps, though I doubt it, for reasons that Marcus and Davis explain well (in particular, that astronomical training sets at best compensate for their being crippled by the lack of a world model).  But even if they do, that would show neither that human language competence is a GPT (given the totality of the relevant evidence) nor that GPT-n is approaching Artificial General Intelligence (whatever that is).


Scott Aaronson’s Response

As usual, I find Steve crystal-clear and precise—so much so that we can quickly dispense with the many points of agreement.  Basically, one side says that, while GPT-3 is of course mind-bogglingly impressive, and while it refuted confident predictions that no such thing would work, in the end it’s just a text-prediction engine that will run with any absurd premise it’s given, and it fails to model the world the way humans do.  The other side says that, while GPT-3 is of course just a text-prediction engine that will run with any absurd premise it’s given, and while it fails to model the world the way humans do, in the end it’s mind-bogglingly impressive, and it refuted confident predictions that no such thing would work.

All the same, I do think it’s possible to identify a substantive disagreement between the distinguished baby-boom linguistic thinkers and the gen-X/gen-Y blogging Scott A.’s: namely, whether there’s a coherent concept of “general intelligence.”  Steve writes:

No logician has ever outlined a normative model of what general intelligence would consist of, and even Turing swapped it out for the problem of fooling an observer, which spawned 70 years of unhelpful reminders of how easy it is to fool an observer.

I freely admit that I have no principled definition of “general intelligence,” let alone of “superintelligence.”  To my mind, though, there’s a simple proof-of-principle that there’s something an AI could do that pretty much any of us would call “superintelligent.”  Namely, it could say whatever Albert Einstein would say in a given situation, while thinking a thousand times faster.  Feed the AI all the information about physics that the historical Einstein had in 1904, for example, and it would discover special relativity in a few hours, followed by general relativity a few days later.  Give the AI a year, and it would think … well, whatever thoughts Einstein would’ve thought, if he’d had a millennium in peak mental condition to think them.

If nothing else, this AI could work by simulating Einstein’s brain neuron-by-neuron—provided we believe in the computational theory of mind, as I’m assuming we do.  It’s true that we don’t know the detailed structure of Einstein’s brain in order to simulate it (we might have, had the pathologist who took it from the hospital used cold rather than warm formaldehyde).  But that’s irrelevant to the argument.  It’s also true that the AI won’t experience the same environment that Einstein would have—so, alright, imagine putting it in a very comfortable simulated study, and letting it interact with the world’s flesh-based physicists.  A-Einstein can even propose experiments for the human physicists to do—he’ll just have to wait an excruciatingly long subjective time for their answers.  But that’s OK: as an AI, he never gets old.

Next let’s throw into the mix AI Von Neumann, AI Ramanujan, AI Jane Austen, even AI Steven Pinker—all, of course, sped up 1,000x compared to their meat versions, even able to interact with thousands of sped-up copies of themselves and other scientists and artists.  Do we agree that these entities quickly become the predominant intellectual force on earth—to the point where there’s little for the original humans left to do but understand and implement the AIs’ outputs (and, of course, eat, drink, and enjoy their lives, assuming the AIs can’t or don’t want to prevent that)?  If so, then that seems to suffice to call the AIs “superintelligences.”  Yes, of course they’re still limited in their ability to manipulate the physical world.  Yes, of course they still don’t optimize arbitrary goals.  All the same, these AIs have effects on the real world consistent with the sudden appearance of beings able to run intellectual rings around humans—not exactly as we do around chimpanzees, but not exactly unlike it either.

I should clarify that, in practice, I don’t expect AGI to work by slavishly emulating humans—and not only because of the practical difficulties of scanning brains, especially deceased ones.  Like with airplanes, like with existing deep learning, I expect future AIs to take some inspiration from the natural world but also to depart from it whenever convenient.  The point is that, since there’s something that would plainly count as “superintelligence,” the question of whether it can be achieved is therefore “merely” an engineering question, not a philosophical one.

Obviously I don’t know the answer to the engineering question: no one does!  One could consistently hold that, while the thing I described would clearly count as “superintelligence,” it’s just an amusing fantasy, unlikely to be achieved for millennia if ever.  One could hold that all the progress in AI so far, including the scaling of language models, has taken us only 0% or perhaps 0.00001% of the way toward superintelligence so defined.

So let me make two comments about the engineering question.  The first is that there’s good news here, at least epistemically: unlike with the philosophical questions, we’re virtually guaranteed more clarity over time!  Indeed, we’ll know vastly more just by the end of this decade, as the large language models are further scaled and tweaked, and we find out whether they develop effective representations of the outside world and of themselves, the ability to reject absurd premises and avoid self-contradiction, or even the ability to generate original mathematical proofs and scientific hypotheses.  Of course, Gary Marcus and Scott Alexander have already placed concrete bets on the table for what sorts of things will be possible by 2030.  For all their differences in rhetoric, I was struck that their actual probabilities differed much more modestly.

So then what explains the glaring differences in rhetoric?  This brings me to my second comment: whenever there’s a new, rapidly-growing, poorly-understood phenomenon, whether it’s the Internet or AI or COVID, there are two wildly different modes of responding to it, which we might call “February 2020 mode” and “March 2020 mode.”  In February 2020 mode, one says: yes, a naïve extrapolation might lead someone to the conclusion that this new thing is going to expand exponentially and conquer the world, dramatically changing almost every other domain—but precisely because that conclusion seems absurd on its face, it’s our responsibility as serious intellectuals to articulate what’s wrong with the arguments that lead to it.  In March 2020 mode, one says: holy crap, the naïve extrapolation seems right!  Prepare!!  Why didn’t we start earlier?

Often, to be sure, February 2020 mode is the better mode, at least for outsiders—as with the Y2K bug, or the many disease outbreaks that fizzle.  My point here is simply that February 2020 mode and March 2020 mode differ by only a month.  Sometimes hearing a single argument, seeing a single example, is enough to trigger an epistemic cascade, causing all the same facts to be seen in a new light.  As a result, reasonable people might find themselves on opposite sides of the chasm even if they started just a few steps from each other.

As for me?  Well, I’m currently trying to hold the line around February 26, 2020.  Suspending my day job in the humdrum, pedestrian field of quantum computing, I’ve decided to spend a year at OpenAI, thinking about the theoretical foundations of AI safety.  But for now, only a year.


Steven Pinker’s Response to Scott

Thanks, Scott, for your thoughtful and good-natured reply, and for offering me the opportunity to respond  in Shtetl-Optimized, one of my favorite blogs. Despite the areas of agreement, I still think that discussions of AI and its role in human affairs—including AI safety—will be muddled as long as the writers treat intelligence as an undefined superpower rather than a mechanisms with a makeup that determines what it can and can’t do. We won’t get clarity on AI if we treat the “I” as “whatever fools us,” or “whatever amazes us,” or “whatever IQ tests measure,” or “whatever we have more of than animals do,” or “whatever Einstein has more of than we do”—and then start to worry about a superintelligence that has much, much more of whatever that is.

Take Einstein sped up a thousandfold. To begin with, current AI is not even taking us in that direction. As you note, no one is reverse-engineering his connectome, and current AI does not think the way Einstein thought, namely by visualizing physical scenarios and manipulating mathematical equations. Its current pathway would be to train a neural network with billions of physics problems and their solutions and hope that it would soak up the statistical patterns.

Of course, the reason you pointed to a sped-up Einstein was to procrastinate having to define “superintelligence.” But if intelligence is a collection of mechanisms rather than a quantity that Einstein was blessed with a lot of, it’s not clear that just speeding him up would capture what anyone would call superintelligence. After all, in many areas Einstein was no Einstein. You above all could speak of his not-so-superintelligence in quantum physics, and when it came world affairs, in the early 1950s he offered the not exactly prescient or practicable prescription, “Only the creation of a world government can prevent the impending self-destruction of mankind.” So it’s not clear that we would call a system that could dispense such pronouncements in seconds rather than years “superintelligent.” Nor with speeding up other geniuses, say, an AI Bertrand Russell, who would need just nanoseconds to offer his own solution for world peace: the Soviet Union would be given an ultimatum that unless it immediately submitted to world government, the US (which at the time had a nuclear monopoly) would bomb it with nuclear weapons.

My point isn’t to poke retrospective fun at brilliant men, but to reiterate that brilliance itself is not some uncanny across-the-board power that can be “scaled” by speeding it up or otherwise; it’s an engineered system that does particular things in particular ways. Only with a criterion for intelligence can we say which of these counts as intelligent.

Now, it’s true that raw speed makes new kinds of computation possible, and I feel silly writing this to you of all people, but speeding a process up by a constant factor is of limited use with problems that are exponential, as the space of possible scientific theories, relative to their complexity, must be. Speeding up a search in the space of theories a thousandfold would be a rounding error in the time it took to find a correct one. Scientific progress depends on the search exploring the infinitesimal fraction of the space in which the true theories are likely to lie, and this depends on the quality of the intelligence, not just its raw speed.

And it depends as well on a phenomenon you note, namely that scientific progress depends on empirical discovery, not deduction from a silicon armchair. The particle accelerators and space probes and wet labs and clinical trials still have to be implemented, with data accumulating at a rate set by the world. Strokes of genius can surely speed up the rate of discovery, but in the absence of omniscience about every particle, the time scale will still be capped by empirical reality. And this in turn directs the search for viable theories: which part of the space one should explore is guided by the current state of scientific knowledge, which depends on the tempo of discovery. Speeding up scientists a thousandfold would not speed up science a thousandfold.

All this is relevant to AI safety. I’m all for safety, but I worry that the dazzling intellectual capital being invested in the topic will not make us any safer if it begins with a woolly conception of intelligence as a kind of wonder stuff that you can have in different amounts. It leads to unhelpful analogies, like “exponential increase in the number of infectious people during a pandemic” ≈ “exponential increase in intelligence in AI systems.” It encourages other questionable extrapolations from the human case, such as imagining that an intelligent tool will develop an alpha-male lust for domination. Worst of all, it may encourage misconceptions of AI risk itself, particularly the standard scenario in which a hypothetical future AGI is given some preposterously generic single goal such as “cure cancer” or “make people happy” and theorists fret about the hilarious collateral damage that would ensue.

If intelligence is a mechanism rather than a superpower, the real dangers of AI come into sharper focus. An AI system designed to replace workers may cause mass unemployment; a system designed to use data to sort people may sort them in ways we find invidious; a system designed to fool people may be exploited to fool them in nefarious ways; and as many other hazards as there are AI systems. These dangers are not conjectural, and I suspect each will have to be mitigated by a different combination of policies and patches, just like other safety challenges such as falls, fires, and drownings. I’m curious whether, once intelligence is precisely characterized, any abstract theoretical foundations of AI safety will be useful in dealing with the actual AI dangers that will confront us.

Because I couldn’t not post

Friday, June 24th, 2022

In 1973, the US Supreme Court enshrined the right to abortion—considered by me and ~95% of everyone I know to be a basic pillar of modernity—in such a way that the right could be overturned only if its opponents could somehow gain permanent minority rule, and thereby disregard the wills of three-quarters of Americans. So now, half a century later, that’s precisely what they’ve done. Because Ruth Bader Ginsburg didn’t live three more weeks, we’re now faced with a civilizational crisis, with tens of millions of liberals and moderates in the red states now under the authority of a social contract that they never signed. With this backwards leap, Curtis Yarvin’s notion that “Cthulhu only ever swims leftward” stands as decimated by events as any thesis has ever been. I wonder whether Yarvin is happy to have been so thoroughly refuted.

Most obviously for me, the continued viability of Texas as a place for science, for research, for technology companies, is now in severe doubt. Already this year, our 50-member CS department at UT Austin has had faculty members leave, and faculty candidates turn us down, with abortion being the stated reason, and I expect that to accelerate. Just last night my wife, Dana Moshkovitz, presented a proposal at the STOC business meeting to host STOC’2024 at a beautiful family-friendly resort outside Austin. The proposal failed, in part because of the argument that, if a pregnant STOC attendee faced a life-threatening medical condition, Texas doctors might choose to let her die, or the attendee might be charged with murder for having a miscarriage. In other words: Texas (and indeed, half the US) will apparently soon be like Donetsk or North Korea, dangerous for Blue Americans to visit even for just a few days. To my fellow Texans, I say: if you find that hyperbolic, understand that this is how the blue part of the country now sees you. Understand that only a restoration of the previous social contract can reverse it.

Of course, this destruction of everything some of us have tried to build in science in Texas is happening despite the fact that 47-48% of Texans actually vote Democratic. It’s happening despite the fact that, if Blue Americans wanted to stop it, the obvious way to do so would be to move to Austin and Houston (and the other blue enclaves of red states) in droves, and exert their electoral power. In other words, to do precisely what Dana and I did. But can I urge others to do the same with a straight face?

As far as I can tell, the only hope at this point of averting a cold Civil War is if, against all odds, there’s a Democratic landslide in Congress, sufficient to get the right to abortion enshrined into federal law. Given the ways both the House and the Senate are stacked against Democrats, I don’t expect that anytime soon, but I’ll work for it—and will do so even if many of the people I’m working with me despise me for other reasons. I will match reader donations to Democratic PACs and Congressional campaigns (not necessarily the same ones, though feel free to advocate for your favorites), announced in the comment section of this post, up to a limit of $10,000.

OpenAI!

Friday, June 17th, 2022

I have some exciting news (for me, anyway). Starting next week, I’ll be going on leave from UT Austin for one year, to work at OpenAI. They’re the creators of the astonishing GPT-3 and DALL-E2, which have not only endlessly entertained me and my kids, but recalibrated my understanding of what, for better and worse, the world is going to look like for the rest of our lives. Working with an amazing team at OpenAI, including Jan Leike, John Schulman, and Ilya Sutskever, my job will be think about the theoretical foundations of AI safety and alignment. What, if anything, can computational complexity contribute to a principled understanding of how to get an AI to do what we want and not do what we don’t want?

Yeah, I don’t know the answer either. That’s why I’ve got a whole year to try to figure it out! One thing I know for sure, though, is that I’m interested both in the short-term, where new ideas are now quickly testable, and where the misuse of AI for spambots, surveillance, propaganda, and other nefarious purposes is already a major societal concern, and the long-term, where one might worry about what happens once AIs surpass human abilities across nearly every domain. (And all the points in between: we might be in for a long, wild ride.) When you start reading about AI safety, it’s striking how there are two separate communities—the one mostly worried about machine learning perpetuating racial and gender biases, and the one mostly worried about superhuman AI turning the planet into goo—who not only don’t work together, but are at each other’s throats, with each accusing the other of totally missing the point. I persist, however, in the possibly-naïve belief that these are merely two extremes along a single continuum of AI worries. By figuring out how to align AI with human values today—constantly confronting our theoretical ideas with reality—we can develop knowledge that will give us a better shot at aligning it with human values tomorrow.

For family reasons, I’ll be doing this work mostly from home, in Texas, though traveling from time to time to OpenAI’s office in San Francisco. I’ll also spend 30% of my time continuing to run the Quantum Information Center at UT Austin and working with my students and postdocs. At the end of the year, I plan to go back to full-time teaching, writing, and thinking about quantum stuff, which remains my main intellectual love in life, even as AI—the field where I started, as a PhD student, before I switched to quantum computing—has been taking over the world in ways that none of us can ignore.

Maybe fittingly, this new direction in my career had its origins here on Shtetl-Optimized. Several commenters, including Max Ra and Matt Putz, asked me point-blank what it would take to induce me to work on AI alignment. Treating it as an amusing hypothetical, I replied that it wasn’t mostly about money for me, and that:

The central thing would be finding an actual potentially-answerable technical question around AI alignment, even just a small one, that piqued my interest and that I felt like I had an unusual angle on. In general, I have an absolutely terrible track record at working on topics because I abstractly feel like I “should” work on them. My entire scientific career has basically just been letting myself get nerd-sniped by one puzzle after the next.

Anyway, Jan Leike at OpenAI saw this exchange and wrote to ask whether I was serious in my interest. Oh shoot! Was I? After intensive conversations with Jan, others at OpenAI, and others in the broader AI safety world, I finally concluded that I was.

I’ve obviously got my work cut out for me, just to catch up to what’s already been done in the field. I’ve actually been in the Bay Area all week, meeting with numerous AI safety people (and, of course, complexity and quantum people), carrying a stack of technical papers on AI safety everywhere I go. I’ve been struck by how, when I talk to AI safety experts, they’re not only not dismissive about the potential relevance of complexity theory, they’re more gung-ho about it than I am! They want to talk about whether, say, IP=PSPACE, or MIP=NEXP, or the PCP theorem could provide key insights about how we could verify the behavior of a powerful AI. (Short answer: maybe, on some level! But, err, more work would need to be done.)

How did this complexitophilic state of affairs come about? That brings me to another wrinkle in the story. Traditionally, students follow in the footsteps of their professors. But in trying to bring complexity theory into AI safety, I’m actually following in the footsteps of my student: Paul Christiano, one of the greatest undergrads I worked with in my nine years at MIT, the student whose course project turned into the Aaronson-Christiano quantum money paper. After MIT, Paul did a PhD in quantum computing at Berkeley, with my own former adviser Umesh Vazirani, while also working part-time on AI safety. Paul then left quantum computing to work on AI safety full-time—indeed, along with others such as Dario Amodei, he helped start the safety group at OpenAI. Paul has since left to found his own AI safety organization, the Alignment Research Center (ARC), although he remains on good terms with the OpenAI folks. Paul is largely responsible for bringing complexity theory intuitions and analogies into AI safety—for example, through the “AI safety via debate” paper and the Iterated Amplification paper. I’m grateful for Paul’s guidance and encouragement—as well as that of the others now working in this intersection, like Geoffrey Irving and Elizabeth Barnes—as I start this new chapter.

So, what projects will I actually work on at OpenAI? Yeah, I’ve been spending the past week trying to figure that out. I still don’t know, but a few possibilities have emerged. First, I might work out a general theory of sample complexity and so forth for learning in dangerous environments—i.e., learning where making the wrong query might kill you. Second, I might work on explainability and interpretability for machine learning: given a deep network that produced a particular output, what do we even mean by an “explanation” for “why” it produced that output? What can we say about the computational complexity of finding that explanation? Third, I might work on the ability of weaker agents to verify the behavior of stronger ones. Of course, if P≠NP, then the gap between the difficulty of solving a problem and the difficulty of recognizing a solution can sometimes be enormous. And indeed, even in empirical machine learing, there’s typically a gap between the difficulty of generating objects (say, cat pictures) and the difficulty of discriminating between them and other objects, the latter being easier. But this gap typically isn’t exponential, as is conjectured for NP-complete problems: it’s much smaller than that. And counterintuitively, we can then turn around and use the generators to improve the discriminators. How can we understand this abstractly? Are there model scenarios in complexity theory where we can prove that something similar happens? How far can we amplify the generator/discriminator gap—for example, by using interactive protocols, or debates between competing AIs?

OpenAI, of course, has the word “open” right in its name, and a founding mission “to ensure that artificial general intelligence benefits all of humanity.” But it’s also a for-profit enterprise, with investors and paying customers and serious competitors. So throughout the year, don’t expect me to share any proprietary information—that’s not my interest anyway, even if I hadn’t signed an NDA. But do expect me to blog my general thoughts about AI safety as they develop, and to solicit feedback from readers.

In the past, I’ve often been skeptical about the prospects for superintelligent AI becoming self-aware and destroying the world anytime soon (see, for example, my 2008 post The Singularity Is Far). While I was aware since 2005 or so of the AI-risk community; and of its leader and prophet, Eliezer Yudkowsky; and of Eliezer’s exhortations for people to drop everything else they’re doing and work on AI risk, as the biggest issue facing humanity, I … kept the whole thing at arms’ length. Even supposing I agreed that this was a huge thing to worry about, I asked, what on earth do you want me to do about it today? We know so little about a future superintelligent AI and how it would behave that any actions we took today would likely be useless or counterproductive.

Over the past 15 years, though, my and Eliezer’s views underwent a dramatic and ironic reversal. If you read Eliezer’s “litany of doom” from two weeks ago, you’ll see that he’s now resigned and fatalistic: because his early warnings weren’t heeded, he argues, humanity is almost certainly doomed and an unaligned AI will soon destroy the world. He says that there are basically no promising directions in AI safety research: for any alignment strategy anyone points out, Eliezer can trivially refute it by explaining how (e.g.) the AI would be wise to the plan, and would pretend to go along with whatever we wanted from it while secretly plotting against us.

The weird part is, just as Eliezer became more and more pessimistic about the prospects for getting anywhere on AI alignment, I’ve become more and more optimistic. Part of my optimism is because people like Paul Christiano have laid foundations for a meaty mathematical theory: much like the Web (or quantum computing theory) in 1992, it’s still in a ridiculously primitive stage, but even my limited imagination now suffices to see how much more could be built there. An even greater part of my optimism is because we now live in a world with GPT-3, DALL-E2, and other systems that, while they clearly aren’t AGIs, are powerful enough that worrying about AGIs has come to seem more like prudence than like science fiction. And we can finally test our intuitions against the realities of these systems, which (outside of mathematics) is pretty much the only way human beings have ever succeeded at anything.

I didn’t predict that machine learning models this impressive would exist by 2022. Most of you probably didn’t predict it. For godsakes, Eliezer Yudkowsky didn’t predict it. But it’s happened. And to my mind, one of the defining virtues of science is that, when empirical reality gives you a clear shock, you update and adapt, rather than expending your intelligence to come up with clever reasons why it doesn’t matter or doesn’t count.

Anyway, so that’s the plan! If I can figure out a way to save the galaxy, I will, but I’ve set my goals slightly lower, at learning some new things and doing some interesting research and writing some papers about it and enjoying a break from teaching. Wish me a non-negligible success probability!


Update (June 18): To respond to a couple criticisms that I’ve seen elsewhere on social media…

Can the rationalists sneer at me for waiting to get involved with this subject until it had become sufficiently “respectable,” “mainstream,” and ”high-status”? I suppose they can, if that’s their inclination. I suppose I should be grateful that so many of them chose to respond instead with messages of congratulations and encouragement. Yes, I plead guilty to keeping this subject at arms-length until I could point to GPT-3 and DALL-E2 and the other dramatic advances of the past few years to justify the reality of the topic to anyone who might criticize me. It feels internally like I had principled reasons for this: I can think of almost no examples of research programs that succeeded over decades even in the teeth of opposition from the scientific mainstream. If so, then arguably the best time to get involved with a “fringe” scientific topic, is when and only when you can foresee a path to it becoming the scientific mainstream. At any rate, that’s what I did with quantum computing, as a teenager in the mid-1990s. It’s what many scientists of the 1930s did with the prospect of nuclear chain reactions. And if I’d optimized for getting the right answer earlier, I might’ve had to weaken the filters and let in a bunch of dubious worries that would’ve paralyzed me. But I admit the possibility of self-serving bias here.

Should you worry that OpenAI is just hiring me to be able to say “look, we have Scott Aaronson working on the problem,” rather than actually caring about what its safety researchers come up with? I mean, I can’t prove that you shouldn’t worry about that. In the end, whatever work I do on the topic will have to speak for itself. For whatever it’s worth, though, I was impressed by the OpenAI folks’ detailed, open-ended engagement with these questions when I met them—sort of like how it might look if they actually believed what they said about wanting to get this right for the world. I wouldn’t have gotten involved otherwise.

Alright, so here are my comments…

Sunday, June 12th, 2022

… on Blake Lemoine, the Google engineer who became convinced that a machine learning model had become sentient, contacted federal government agencies about it, and was then fired placed on administrative leave for violating Google’s confidentiality policies.

(1) I don’t think Lemoine is right that LaMDA is at all sentient, but the transcript is so mind-bogglingly impressive that I did have to stop and think for a second! Certainly, if you sent the transcript back in time to 1990 or whenever, even an expert reading it might say, yeah, it looks like by 2022 AGI has more likely been achieved than not (“but can I run my own tests?”). Read it for yourself, if you haven’t yet.

(2) Reading Lemoine’s blog and Twitter this morning, he holds many views that I disagree with, not just about the sentience of LaMDA. Yet I’m touched and impressed by how principled he is, and I expect I’d hit it off with him if I met him. I wish that a solution could be found where Google wouldn’t fire him.

Computer scientists crash the Solvay Conference

Thursday, June 9th, 2022

Thanks so much to everyone who sent messages of support following my last post! I vowed there that I’m going to stop letting online trolls and sneerers occupy so much space in my mental world. Truthfully, though, while there are many trolls and sneerers who terrify me, there are also some who merely amuse me. A good example of the latter came a few weeks ago, when an anonymous commenter calling themselves “String Theorist” submitted the following:

It’s honestly funny to me when you [Scott] call yourself a “nerd” or a “prodigy” or whatever [I don’t recall ever calling myself a “prodigy,” which would indeed be cringe, though “nerd” certainly —SA], as if studying quantum computing, which is essentially nothing more than glorified linear algebra, is such an advanced intellectual achievement. For what it’s worth I’m a theoretical physicist, I’m in a completely different field, and I was still able to learn Shor’s algorithm in about half an hour, that’s how easy this stuff is. I took a look at some of your papers on arXiv and the math really doesn’t get any more advanced than linear algebra. To understand quantum circuits about the most advanced concept is a tensor product which is routinely covered in undergraduate linear algebra. Wheras in my field of string theory grasping, for instance, holographic dualities relating confirmal field theories and gravity requires vastly more expertise (years of advanced study). I actually find it pretty entertaining that you’ve said yourself you’re still struggling to understand QFT, which most people I’m working with in my research group were first exposed to in undergrad 😉 The truth is we’re in entirely different leagues of intelligence (“nerdiness”) and any of your qcomputing papers could easily be picked up by a first or second year math major. It’s just a joke that this is even a field (quantum complexity theory) with journals and faculty when the results in your papers that I’ve seen are pretty much trivial and don’t require anything more than undergraduate level maths.

Why does this sort of trash-talk, reminiscent of Luboš Motl, no longer ruffle me? Mostly because the boundaries between quantum computing theory, condensed matter physics, and quantum gravity, which were never clear in the first place, have steadily gotten fuzzier. Even in the 1990s, the field of quantum computing attracted amazing physicists—folks who definitely do know quantum field theory—such as Ed Farhi, John Preskill, and Ray Laflamme. Decades later, it would be fair to say that the physicists have banged their heads against many of the same questions that we computer scientists have banged our heads against, oftentimes in collaboration with us. And yes, there were cases where actual knowledge of particle physics gave physicists an advantage—with some famous examples being the algorithms of Farhi and collaborators (the adiabatic algorithm, the quantum walk on conjoined trees, the NAND-tree algorithm). There were other cases where computer scientists’ knowledge gave them an advantage: I wouldn’t know many details about that, but conceivably shadow tomography, BosonSampling, PostBQP=PP? Overall, it’s been what you wish every indisciplinary collaboration could be.

What’s new, in the last decade, is that the scientific conversation centered around quantum information and computation has dramatically “metastasized,” to encompass not only a good fraction of all the experimentalists doing quantum optics and sensing and metrology and so forth, and not only a good fraction of all the condensed-matter theorists, but even many leading string theorists and quantum gravity theorists, including Susskind, Maldacena, Bousso, Hubeny, Harlow, and yes, Witten. And I don’t think it’s just that they’re too professional to trash-talk quantum information people the way commenter “String Theorist” does. Rather it’s that, because of the intellectual success of “It from Qubit,” we’re increasingly participating in the same conversations and working on the same technical questions. One particularly exciting such question, which I’ll have more to say about in a future post, is the truth or falsehood of the Quantum Extended Church-Turing Thesis for observers who jump into black holes.

Not to psychoanalyze, but I’ve noticed a pattern wherein, the more secure a scientist is about their position within their own field, the readier they are to admit ignorance about the neighboring fields, to learn about those fields, and to reach out to the experts in them, to ask simple or (as it usually turns out) not-so-simple questions.


I can’t imagine any better illustration of these tendencies better than the 28th Solvay Conference on the Physics of Quantum Information, which I attended two weeks ago in Brussels on my 41st birthday.

As others pointed out, the proportion of women is not as high as we all wish, but it’s higher than in 1911, when there was exactly one: Madame Curie herself.

It was my first trip out of the US since before COVID—indeed, I’m so out of practice that I nearly missed my flights in both directions, in part because of my lack of familiarity with the COVID protocols for transatlantic travel, as well as the immense lines caused by those protocols. My former adviser Umesh Vazirani, who was also at the Solvay Conference, was proud.

The Solvay Conference is the venue where, legendarily, the fundamentals of quantum mechanics got hashed out between 1911 and 1927, by the likes of Einstein, Bohr, Planck, and Curie. (Einstein complained, in a letter, about being called away from his work on general relativity to attend a “witches’ sabbath.”) Remarkably, it’s still being held in Brussels every few years, and still funded by the same Solvay family that started it. The once-every-few-years schedule has, we were constantly reminded, been interrupted only three times in its 110-year history: once for WWI, once for WWII, and now once for COVID (this year’s conference was supposed to be in 2020).

This was the first ever Solvay conference organized around the theme of quantum information, and apparently, the first ever that counted computer scientists among its participants (me, Umesh Vazirani, Dorit Aharonov, Urmila Mahadev, and Thomas Vidick). There were four topics: (1) many-body physics, (2) quantum gravity, (3) quantum computing hardware, and (4) quantum algorithms. The structure, apparently unchanged since the conference’s founding, is this: everyone attends every session, without exception. They sit around facing each other the whole time; no one ever stands to lecture. For each topic, two “rapporteurs” introduce the topic with half-hour prepared talks; then there are short prepared response talks as well as an hour or more of unstructured discussion. Everything everyone says is recorded in order to be published later.


Daniel Gottesman and I were the two rapporteurs for quantum algorithms: Daniel spoke about quantum error-correction and fault-tolerance, and I spoke about “How Much Structure Is Needed for Huge Quantum Speedups?” The link goes to my PowerPoint slides, if you’d like to check them out. I tried to survey 30 years of history of that question, from Simon’s and Shor’s algorithms, to huge speedups in quantum query complexity (e.g., glued trees and Forrelation), to the recent quantum supremacy experiments based on BosonSampling and Random Circuit Sampling, all the way to the breakthrough by Yamakawa and Zhandry a couple months ago. The last slide hypothesizes a “Law of Conservation of Weirdness,” which after all these decades still remains to be undermined: “For every problem that admits an exponential quantum speedup, there must be some weirdness in its detailed statement, which the quantum algorithm exploits to focus amplitude on the rare right answers.” My title slide also shows DALL-E2‘s impressionistic take on the title question, “how much structure is needed for huge quantum speedups?”:

The discussion following my talk was largely a debate between me and Ed Farhi, reprising many debates he and I have had over the past 20 years: Farhi urged optimism about the prospect for large, practical quantum speedups via algorithms like QAOA, pointing out his group’s past successes and explaining how they wouldn’t have been possible without an optimistic attitude. For my part, I praised the past successes and said that optimism is well and good, but at the same time, companies, venture capitalists, and government agencies are right now pouring billions into quantum computing, in many cases—as I know from talking to them—because of a mistaken impression that QCs are already known to be able to revolutionize machine learning, finance, supply-chain optimization, or whatever other application domains they care about, and to do so soon. They’re genuinely surprised to learn that the consensus of QC experts is in a totally different place. And to be clear: among quantum computing theorists, I’m not at all unusually pessimistic or skeptical, just unusually willing to say in public what others say in private.

Afterwards, one of the string theorists said that Farhi’s arguments with me had been a highlight … and I agreed. What’s the point of a friggin’ Solvay Conference if everyone’s just going to agree with each other?


Besides quantum algorithms, there was naturally lots of animated discussion about the practical prospects for building scalable quantum computers. While I’d hoped that this discussion might change the impressions I’d come with, it mostly confirmed them. Yes, the problem is staggeringly hard. Recent ideas for fault-tolerance, including the use of LDPC codes and bosonic codes, might help. Gottesman’s talk gave me the insight that, at its core, quantum fault-tolerance is all about testing, isolation, and contact-tracing, just for bit-flip and phase-flip errors rather than viruses. Alas, we don’t yet have the quantum fault-tolerance analogue of a vaccine!

At one point, I asked the trapped-ion experts in open session if they’d comment on the startup company IonQ, whose stock price recently fell precipitously in the wake of a scathing analyst report. Alas, none of them took the bait.

On a different note, I was tremendously excited by the quantum gravity session. Netta Engelhardt spoke about her and others’ celebrated recent work explaining the Page curve of an evaporating black hole using Euclidean path integrals—and by questioning her and others during coffee breaks, I finally got a handwavy intuition for how it works. There was also lots of debate, again at coffee breaks, about Susskind’s recent speculations on observers jumping into black holes and the quantum Extended Church-Turing Thesis. One of my main takeaways from the conference was a dramatically better understanding of the issues involved there—but that’s a big enough topic that it will need its own post.

Toward the end of the quantum gravity session, the experimentalist John Martinis innocently asked what actual experiments, or at least thought experiments, had been at issue for the past several hours. I got a laugh by explaining to him that, while the gravity experts considered this too obvious to point out, the thought experiments in question all involve forming a black hole in a known quantum pure state, with total control over all the Planck-scale degrees of freedom; then waiting outside the black hole for ~1070 years; collecting every last photon of Hawking radiation that comes out and routing them all into a quantum computer; doing a quantum computation that might actually require exponential time; and then jumping into the black hole, whereupon you might either die immediately at the event horizon, or else learn something in your last seconds before hitting the singularity, which you could then never communicate to anyone outside the black hole. Martinis thanked me for clarifying.


Anyway, I had a total blast. Here I am amusing some of the world’s great physicists by letting them mess around with GPT-3.

Back: Ahmed Almheiri, Juan Maldacena, John Martinis, Aron Wall. Front: Geoff Penington, me, Daniel Harlow. Thanks to Michelle Simmons for the photo.

I also had the following exchange at my birthday dinner:

Physicist: So I don’t get this, Scott. Are you a physicist who studied computer science, or a computer scientist who studied physics?

Me: I’m a computer scientist who studied computer science.

Physicist: But then you…

Me: Yeah, at some point I learned what a boson was, in order to invent BosonSampling.

Physicist: And your courses in physics…

Me: They ended at thermodynamics. I couldn’t handle PDEs.

Physicist: What are the units of h-bar?

Me: Uhh, well, it’s a conversion factor between energy and time. (*)

Physicist: Good. What’s the radius of the hydrogen atom?

Me: Uhh … not sure … maybe something like 10-15 meters?

Physicist: OK fine, he’s not one of us.

(The answer, it turns out, is more like 10-10 meters. I’d stupidly substituted the radius of the nucleus—or, y’know, a positively-charged hydrogen ion, i.e. proton. In my partial defense, I was massively jetlagged and at most 10% conscious.)

(*) Actually h-bar is a conversion factor between energy and 1/time, i.e. frequency, but the physicist accepted this answer.


Anyway, I look forward to attending more workshops this summer, seeing more colleagues who I hadn’t seen since before COVID, and talking more science … including branching out in some new directions that I’ll blog about soon. It does beat worrying about online trolls.