What is the Future of Quantum? | Scott Aaronson

追加: 2026-05-13

[00:00:00]Hey, Ryan Grant from Digital Contract Design. Microsoft has the Majorana architecture and it's been really hard to understand some of the u some of the hype around a lot of uh quantum advancements, but it sounds like something that persists is really important. Can you can you talk about how real it is and and what the questions are? Yeah. So, so Microsoft uh made a big investment in this topological approach to building cubits which I would say is speculative even by the standards of quantum computing, right? It's uh uh uh you know and and and like like with with other approaches uh with you know trapped ions which you know quantinuum does for example with superconducting cubits which Google and IBM do you know we now as of the last couple of years have devices with you know a 100 cubits or so that can do thousands of operations. You know they work just like the theory said that they would 30 years ago. uh you know it remains to scale that up to millions of physical cubits as you might need to threaten public key cryptography or to you know uh do various quantum simulations you know so that so there so there's still scaling that has to happen but you know there's there's a lot of progress that's been made with topological quantum computing so with the myeronafirmians you know uh uh in particular there the debate is about has Microsoft succeeded or not at making even one cubit it. Okay. So, like it's it's at a much earlier stage, you know.

[00:01:35]So, why does anyone care about it? Well, uh, uh, so, so the the theory of this stuff was developed 25 years ago by people like Alexi Katayv and Michael Freriedman. And, you know, there was like a beautiful theory behind it that basically if you can get this to work at all, then it has some natural error resistance just built into the physics of how these myerana particles work.

[00:01:59]that basically in order to cause an error you'd have to change the topology of how they are being braided around each other and sort of local changes to the particles don't make a difference if they don't change the topology. So, so, uh, so that that was the dream that you would have something like a quantum computing analog of a transistor, right?

[00:02:20]Where you don't have to do all of this expensive sort of softwarebased error correction that we're expecting to have to do with all the other approaches to quantum computing like with uh uh superconducting cubits and with trapped ions and with neutral atoms. Uh the hope is that at least the need for that would be less uh if you could get these myeronic cubits to work. Okay. But Microsoft's claim a year ago to have created a topological cubit has not been widely accepted. Right? The condensed matter physics community pushed back very hard against it. They basically said m you know Microsoft has not published uh the data that we would need to uh confirm this claim. You know, there was also some bad blood because Microsoft had had a previous claim from 2018 uh that was from a lab that they contracted with in the Netherlands and then that claim had to be retracted later. So, you know, so there was some history here and so um I would say that the jury is still out. I would say that Micros you know like the ball sort of went to Microsoft's court to sort of uh uh prove to the community you know that that they have uh actually seen this what's called a meer on a zero mode and based on that made a topological cubit and that sort of they have not uh uh uh convinced the community you know of that at this time you know maybe they've done it you know I I I don't know you know even if they have there would be a a long ways to go even just to catch up to where the superconducting and the trapped ion cubits already are. Okay, so that's one thing to know. The other thing to know is that Microsoft has stopped putting all of its eggs into this basket. Microsoft also works on more conventional approaches to quantum computing.

[00:04:09]Actually, can you can get in I have a question. So, why don't you get back to get in line and then we can do it that way. Um, just you know, just bear with us here. Uh, we will get to your question. Uh I do have a question real quickly. Um I have a question. Me myself um you know people like Terry Tao you know brilliant mathematician has been talking about how AI you know these LLMs are helping with scientific discovery.

[00:04:31]Uh you did a sbatical at open AI you've been in and around the AI space for well the modern AI space. You've been around for a while now. Um are you seeing the impact of AI in your research uh in terms of increasing productivity for you and for your posttos for your graduate students like tell us about that. A short answer is yes. Uh so I mean when I have a research problem in quantum computing like you know a lema that I need to prove I mean you know the first thing that I do these days is I ask you know GPT 5.2 too pro and often it is able to prove it for me. you know it is uh uh uh you know I can still do things that it can't do uh uh uh you know and and uh but but you know when I use it especially within the last six months or so I feel very grateful that I have tenure you know because you know I don't know you know for how you know I I I simply don't know for how much longer are we needed right I mean like it can't do our whole research projects for us but you know to have this like unbelievably enthus enthusiastic grad student who is sort of available 24/7, you know, which you know, my grad student's not quite right.

[00:05:42]Uh uh, you know, and we'll get back to you in like a half hour having scoured the entire literature, you know, and seen what are all of the techniques that are published that might be relevant to this problem and, you know, and and what happens when you try them. Like, of course, that's useful, right? That's unbelievably useful. uh and you know and and I think like I have colleagues who are like yeah but you know I tried it and it hallucinates too much you can't rely on it and I'm like well when's the last time you tried it right and if and if the answer is a year ago then I'm like no you know you're out of date you have to try it again now right uh so so yeah I think it is absolutely changing the way that we work uh in you know fields like you know theoretical computer science uh quantum computing theory I mean you know I have papers you know that are coming out now where you know major lemmas were proved with GPT assistance you know and of course we have to verify it ourselves you know because it can give you incredibly convincing looking you know uh uh uh proofs that are just crap right it it can totally do that but then when it does that you can tell it like look this can't be right because it contradicts this other thing I know and then it says oh you're absolutely right I'm so sorry let me try again and then sometimes it does a much better job So you know m much like some you know some humans who you know uh uh you know it is it is a lot like interacting with a with a grad student you know or a posttock. So um so is it is it changing the way we work? I would say yes you know and and like unevenly because not every you know not everyone is even aware of what it can now do or they have a an impression of what it can do that is frozen in place from a year or two ago. But I think uh you know this this title wave is coming for all of us.

[00:07:33]>> Hi, I um I actually caught your talk two years name and affiliation.

[00:07:37]>> I'm Jack and I make or I'm a security researcher, but um I caught your talk two years ago at Southby, the truth about quantum computing. And you made me very dubious about it because I think the overall takeaway was this is just a science experiment that hasn't actually been proven that can be done yet. And so every time I see the hype and the news, Google has made a breakthrough in quantum computing, I'm like, "All right, what exactly have you done?" And it's very difficult for me to get any sort of answers from them because they'll say they've dealt something. And and I'm wondering from from your perspective, when is it actually something to take notice of? Because you hear these hypes and stuff, but it's like >> I I remember Scott said, "This isn't even here yet, and it probably may not ever be here." And so when is it to you that this is actually something worth noting or paying attention to?

[00:08:24]>> Well, look, I mean I I took note of it 30 years ago and I've been working on it ever since. Right. But that's because I care about what are the ultimate limits of what is computable in nature. And I and I said you know in the '9s that you know if quantum mechanics is true then clearly you know it should change that right the burden of proof is on the people who think that it's not going to change it. Okay. Uh uh but you know you have to you have to clearly separate out the different questions you know uh uh uh uh like like you know it's not just like there's not just this one meter of like is this real or is this fake okay there is like there are multiple questions here okay you know first of all there is you know will a quantum computer be built how long will it take and so forth and then separately there is the question what exactly will it be good for right where will it outperform a a a classical computer. You know, I feel like uh um you know, the the evidence in the last couple of years, you know, has been uh you know, that as many of us have been saying for decades, absolutely, yes, this can be built. You know, it hasn't, you know, a a fully fault tolerant quantum computer uh that is scalable has not yet been built. But, uh I cannot give anyone confidence that that won't happen within the next decade. Okay? So like if someone you know has uh uh let's say a blockchain or uh a crypto you know a crypto system that they're using that is vulnerable to quantum attack I would be worried now and I would be looking to migrate to quantum resistant encryption now uh that is that that is the advice that I've been giving for the uh for the last couple of years. Okay. The reason for that is that, you know, the the the devices are finally starting to work, you know, just like the theory said they would 30 years ago, right? And um you know, it's taken a while, but you know, we we we sort of we expected to get to this point and and and and now it is getting to that point, right? I I had a joke on my blog recently that like people are like accusing me of contradicting myself like you know they say like 10 years ago you said that quantum computing wasn't you know imminent and now you say it is imminent so how did what made you change you know why are you reversing yourself and it's like 10 years ago you said you're 34 now you say you're 44 right like you know why why can't you keep your story straight okay um but uh uh uh but you know but On the other hand, you know, again, there is the separate question, what is a quantum computer good for? And there, you know, as I said in that talk that you were at, uh, uh, you know, I think, uh, uh, people have gotten a grossly overhyped, you know, impression from, you know, actors in this space who are who are not honest, you know, who don't have an incentive to be honest, who who learned 20 years ago that you can just use, you know, as soon as you say the word quantum, then people will believe anything, right? And you can just say this is going to revolutionize AI. It's going to revolutionize handwriting recognition and you know uh vehicle rooting and oil and gas exploration. You know whatever people most want to hear and they will just believe you right. Uh and and so you know there were a few like startups you know like D-Wave you know that I credit for that innovation that you know you can just put the word quantum in front of any application and people will just eat it up with mustard right uh um you know but like for the scientists in the field the only question that matters is where do you actually beat a classical computer okay uh because like people you learn that they can say look we've used a small quantum computer to do this uh um AI task, right? Uh to uh do this machine learning or optimization and everyone will be impressed by that and you don't even have to lie because no one will even know to ask the question the only question that matters to the scientists which is but is there any hope of beating a classical computer, right? Of course a quantum computer can you know anything that a classical computer can do a quantum computer can simulate in principle. The question is where do you get an advantage? And you know after 30 years of research in quantum algorithms you know we we still just have the like two really huge categories where there's a clear exponential quantum speed up and those are number one simulating quantum physics itself. Okay, which has all sort actually has all sorts of applications.

[00:12:52]For example, for designing better batteries, better photovoltaics, um better uh pharmaceuticals, you know, maybe new ways of making fertilizer, right? There's all kinds of problems in chemistry, material science that depend on simulating quantum mechanics. Right?

[00:13:08]So that was sort of the original application of a quantum computer that Fineman and other physicists had in mind when they proposed the idea of quantum computing 45 years ago. Okay. And that's still probably the economically most important application that we know.

[00:13:23]Okay. And then the number two application, well, is breaking the public key encryption that currently protects the internet. Okay. Okay, that's a big deal if you're the NSA or if you're some criminal syndicate, you know, uh it's harder to put forward as a positive application for humanity. Okay, and then there's all the other stuff.

[00:13:42]There's the optimization and AI and machine learning where uh it's a very complicated and evolving story, but long story short, the advantages that we're able to get from a quantum computer seem to be more modest. Okay, they seem to be like mostly square root speedups or polomial speedups rather than exponential speedups. And for that reason, it's going to be further into the future before a quantum computer will give you a win in practice for those. And you know, and when it does, it will be a more modest win. And there might be other applications or other huge quantum speedups that haven't even been discovered yet that maybe, you know, once we have a quantum computer to experiment with, we could find, right?

[00:14:24]That's always possible. But like you know the the other thing to understand is that classical computing gets to fight back right. So like there have been you know in the history of this field there have been things that looked like huge quantum advantages that then went away when people discovered a faster classical algorithm to solve the same problem. Okay. So you know the the the the the most stable you know sort of advantages that we're really confident about again they've been for uh simulating quantum physics and then for breaking public key encryption right which just happens to have this very very special structure that we know how to exploit with a quantum computer.

[00:15:04]>> All right name and affiliation again.

[00:15:07]>> Uh oh just really quickly uh there's more refreshments in the back and stuff.

[00:15:10]I'm noticing people are being restocked so just keep that in mind. All right.

[00:15:13]Go.

[00:15:14]>> Hi, my name is Ramos. Uh, I'm a founder of a company called Memory Link. I have two questions. Um, what was your role with uh the alignment efforts at OpenAI?

[00:15:24]Was it more for your technical skills, your philos philosophical inquiry or combination of two?

[00:15:29]>> I still don't know that.

[00:15:31]>> Okay. Uh, answer.

[00:15:34]>> And then my second question is, do you think alignment can be made possible without making any ideological presuppositions?

[00:15:41]>> Uh, all right. Let me start with the first question. So, so what happened was that in 2022 uh I got like uh contacted by Yan Ly who was then at OpenAI and by his boss Ilia Sudskver and they said well you know we we would like you to you know take a year off and and work for our organization called Open AI which is just a little nonprofit that is you know trying to build AI in the broader interest of humanity and blah blah blah right you know and and uh um and I My my first reaction was like, why do you want me? I'm a quantum computing person, right? I barely know anything about AI. And I think it was just that they were fans of my blog. Okay. And they well, you know, and and and and they had a case about how like, you know, they thought maybe solving AI alignment was really a question of theoretical computer science and that, you know, maybe I could help with that.

[00:16:35]And uh so I was not convinced of that, but I said, you know, this is interesting. And you know this uh uh this this this this AI thing seem it seems like something I should learn about you know seems uh seems like maybe it's going places right and uh and so uh you know they they would let me uh just go on leave and you know mostly stay here in Austin with my students and my family and just sort of uh uh work for them. So I I did that for a year. That year turned into two years. Uh and it was during the second year that all those events happened that you all read about in the news of you know Ilia tried to depose Sam Alman he got deposed instead and then the whole nature of open AI changed and they became much more explicit that we are just a for-profit company and we are racing to do this uh you know as fast as we can and make a lot of money. So um you know that was that was while I was there although it was sort of above my pay grade like you know I knew these people and I did not know that this was going to happen. Um so uh um you know so so so why they wanted me I think it was you know just about you know can theoretical computer science do anything for AI alignment and that remains a question that I'm extremely interested in. I think uh you know it's maybe the most urgent thing that I could be working on.

[00:17:54]So now at at UT Austin uh in addition to uh the quantum information center that I've been running there for a decade uh starting this past fall I am also running uh a a uh a center for uh uh a theoretical computer science and AI alignment. Okay. And so I have a group of students. We are trying to figure out what theorems can we prove that would uh tell us something about uh interpretability of neural nets about out of distribution generalization about other questions that might be relevant for AI alignment. Uh now your second question was can alignment be solved without any ideological presuppositions?

[00:18:34]Like I sort of doubt that and that like no matter like like I'm I'm I'm I'm not sure if someone can blow their nose these days without any ideological ramifications that that that that someone would yell at them on on on Twitter on Twitter. Exactly. Some someone on social media would call that problematic. Right. I like every week have the experience of putting up a blog post and think like surely no one is going to yell at me for this one and then being wrong about that. So yeah, I sort of doubt that, you know, but >> we got nine minutes, so keep it concise in one question from now on. Let's just go. Go.

[00:19:07]>> Hi, my name is Dennis Mendez. I have about a dozen AI startups. They're all international, so I do a lot of travel.

[00:19:13]Are >> you from New York?

[00:19:14]>> I'm originally from New York, but I lived in San Francisco for 20 years, and I've been here since 2010. Um so I was in uh GTEX in Dubai met with Ishmael Farro who is the vice senior vice president IBM for quantum computing and AI and what he was saying in his highly technical presentation that the quantum is the AI is helping him redesign quantum computing from IBM's perspective and the quantum is going to make AI stronger and more powerful and he predicted a singularity or an event horizon when humans would no longer be involved because the two would just keep echoing each other. But the the other question I have first of all do you believe in that and is there an event horizon that you have? But the other question is when is Bitcoin going to be in peril because quantum is going to be able to crush the the mesh network.

[00:20:08]>> There's a bunch of things to pick apart there. First of all, there might well be a singularity that will kill us all. You know, I can't rule that out. It might have nothing to do with quantum computing. might be done just purely using classical computing, right? Uh uh you know the the the the the question that I can address is like is there a differential advantage to a to using a quantum computer? Okay, as I was say saying before, uh uh the advantages of quantum computers for AI, I think have been grossly overhyped and exaggerated, including by people who don't understand anything, but also by people who do understand, you know, but who have learned that that's what people want to hear and that's the way to raise money.

[00:20:47]So, you know, they just present this view that quantum computers speed up everything from AI when the actual reality that we know is that, you know, the advantages there are much more modest. Okay. Now, can AI help with quantum computing? Well, of course it can. It can help with everything that humans do, right? The the question there is just what what will be left for us to do. Okay. Uh so, you know, it can help with quantum computing. It's just a special case of it helping with everything, right? So, uh so, okay. And and then you asked about Bitcoin. Um so, you know, I you know, look, if I could predict what years things would happen, I wouldn't be a professor. I would be an investor. Okay. Okay. And I would be I would have a much bigger house. Okay. Uh uh so you know but but again what what I said before is that you know anyone who is using cryptography that is vulnerable to a quantum computer should be looking to migrate now. I cannot you know even guarantee you that you won't have a fault tolerant quantum computer in the next 5 years. Certainly not in the next 10 years. Okay. I can't guarantee that you will either. It's a question at this point of how how badly does anyone want it and how much are they willing to spend. You know, I think that the basic engineering building blocks are pretty much now in place. If there were some fundamental showstopper, then it would have been discovered by now.

[00:22:11]>> Hi, I'm Chris. I'm a software developer but no affiliation. Um, my question is more in terms of research. Uh, it seems like you know you're using AI for research. Um, our previous speaker talked about how um, you know, it's important that they're grounded in actual uh, documentation, web searches to give like reliable answers.

[00:22:31]>> It seems like that becomes an issue of IP or how you can, you know, access and quote various materials legally. Um, what do you think is going to happen in terms of like for non-math related questions that they can't just simulate on their own, but they actually need to go and quote sources? How is research do you think going to evolve in using AI?

[00:22:51]Are the publishers going to take over or Yeah. What do you think?

[00:22:55]>> Okay. So, I mean there's a technical issue here and then there's a legal issue. Right. On the technical issue, I think that the models have actually gotten much much better at citing appropriate sources than they were let's say a year ago, certainly two years ago.

[00:23:10]Right? I mean, you know, a lot of us when we first tried these things, they would just hallucinate left and right.

[00:23:16]they would cite papers that don't exist right and you know and became infamous for that okay but now when you use these reasoning models you know like let's say GPT thinking or pro right then they are actually going on the internet doing a Google search you know before they you know spending 10 or 15 minutes before they answer you and then when they give you an answer it actually has links where you can check for yourself that yeah not only do these papers exist but that the papers say what what GP BT says they say right so you know so so so that's already improved tremendously now you know the legal issue I mean you know there are already cases like winding through the courts about you know uh uh uh you know what are you allowed to use for training data are you allowed to use these copyrighted materials blah blah blah you know I mean I mean the the I think the the answer that we're getting so far from the legal system is that like when a an answer an individ idual answer from an AI is infringing, right?

[00:24:18]Like it just reprints verbatim a New York Times article, right? Then the New York Times has a case, right? But if it's just that the AI was trained on all this material and and if that material was legally bought by the AI company, well then it's kind of like, you know, you're allowed to go to the library, you know, you're allowed to read the books and have those affect the synapses in your brain in some inscrable way that will then affect how you answer questions, right? And this model, you know, at least as far as current copyright law is concerned, seems to be allowed to do the same thing, right? And so like if we want to make that illegal, you know, because we're worried about massive job losses or whatever, then that seems like that would require changing the law, which may require an act of Congress. Good luck with that.

[00:25:11]>> Hi, I'm Aro UT student. Hey, >> so uh hook them >> circling back to the um generating proofs with AI u for your research for example you have a lema and the AI is just able to get it right first try if you put it in a paper what's like the current practice with like attributing it to AI and are you noticing like any conventions crystallizing around that >> yeah current practice you do not list GPT as a co-author on the paper okay but you do but but you do say in the paper like this lema was proved with the help of you know and you you know you may also uh uh uh as you wish want to include details about what model you used what prompt you used stuff like that >> last question and then you can go back and people can follow you all right go ahead >> hello I'm Anna K winters I'm a fellow at the mcsus institute for the advanced study of philosophy and theology I'm currently preparing a seminar series on the philosophy of artificial intelligence I have a very specific question for you which is however modest in your experience what was the most impressive result yielded from applying philosophical speculation to theoretical computer science thank you >> very very precise question I like that very >> that's a tough one I mean um look uh uh the the you know original insights of of touring you know and his friends you know who formulated the church touring thesis you know the modern notion of what a computer even is, right? And then, you know, the the the touring test, you know, the imitation game, right? You can think of as a sort of indistinguishability argument, right? That uh uh uh you know, just the idea that we could take this philosophical question, you know, can a machine think, right? Can a machine be conscious? And we can without answering that question replace it by a different question of can a machine have input output behavior that is indistinguishable from a humans and then we could actually make progress on that question. Right? I mean that was that was a philosophical insight that then affected computer science in a way that was kind of hard to top I would say. All right. Uh okay. So uh Scott you can be around for a little while right?

[00:27:26]>> Uh yeah. Yeah. Okay. Okay. So, I mean maybe you can like take questions in the back if people want. I mean, you know.

[00:27:31]>> Yeah. Yeah. Sure. Sure.

[00:27:32]>> This guy's like he's a he's a resource, a database of sorts if you if you will.

[00:27:36]Um, but thank you. Thank you very much.

[00:27:38]Let's let's give it up for Dr. Thanks for having me. The doctor