Yuval Boger interviews Scott Aaronson, a UT Austin computer science professor known for his work on quantum computing theory. They explore the current state of quantum hardware, the narrowing case for quantum skepticism, and the realistic path toward fault-tolerant, useful quantum machines. The conversation also covers quantum algorithms, cryptography risks, ethics, hype in commercialization, and advice for the next generation of quantum researchers.
Transcript
Yuval: Hello Scott, thank you for joining me today.
Scott: Thanks, it’s great to be here.
Yuval: So who are you and what do you do?
Scott: I’m Scott Aaronson, a computer science professor at the University of Texas at Austin, and I’ve spent most of my career—about 27 years now—thinking about the capabilities and limits of quantum computers. I dabble in other things too, classical theoretical computer science and now AI and AI alignment. Who isn’t dabbling in AI these days? But quantum computation has been my main interest for quite a while.
Yuval: So with 27 years of experience in quantum, where are we today? How close are we to the holy grail?
Scott: You’re at one of the leading quantum computing companies, so I think you know as well as I do that we are in a very exciting time right now. A lot of the ideas we’ve known since the 1990s—fault tolerance, quantum error correction, how you’re going to scale this up—these ideas are not that new. But it’s only within the last couple of years that we’re really starting to see them work in the lab. And they’re working pretty much exactly like the theory said they would.
We now have, in various platforms—superconducting qubits, trapped ions, neutral atoms—two-qubit gates at 99-point-something percent fidelity: 99.6, 99.7, even 99.9 sometimes. It’s clearly closing in on the threshold where you can start to error correct and it becomes a net win, where you’re correcting errors faster than you’re introducing new ones. That’s been understood for 30 years to be the critical mass for building a quantum computer—that’s where it really becomes impressive.
When I entered this field in the late 90s, what we knew was that the threshold for error correction was some constant. If you could get the rate of error in your two-qubit gate down to less than one in a million, maybe this would work. Meanwhile, the two-qubit gates people could actually demonstrate in the lab had maybe 50% error. The two numbers were just ridiculously far apart. It was almost a joke of a theorist—okay, this is just a constant.
But in the past quarter century, we’ve actually closed this gap. We’ve closed these orders of magnitude. We now know error-correcting codes that can deal with realistic two-qubit gates, and what’s achievable in the lab went from 90% accuracy to 99% and now 99.9%.
There remains an enormous challenge: how do you scale this up to a system with millions of physical qubits, thousands of logical qubits, billions of operations? No one has ever engineered a quantum system on that scale before. We don’t know how long that’s going to take. But if there were some fundamental showstopper—if the skeptics were right that this can’t be done—we would have seen it by now. I don’t understand how we would have gotten this far, how these systems with 100 or more qubits and thousands of gates would have worked exactly like the theory said, if something fundamental were wrong. So that’s been exciting to see.
In the meantime, there’s been lots of stuff on the theory and algorithm side of quantum computing. None of it has really overturned the picture we had in place by the 1990s. Rather, it’s built on that picture.
Yuval: I’ve heard you refer many times in your talks to an argument with Gil Kalai about whether large-scale quantum computing is even possible. Do you think he still has a path to being vindicated, or is it over?
Scott: I feel like his path has been getting narrower and narrower. Gil Kalai is a brilliant mathematician and one of the leading skeptics of quantum computing. What he was postulating was that he believes quantum mechanics—quantum mechanics is fine—but there has to be some principle of correlated noise that comes on top of quantum mechanics and somehow screens off quantum computation.
I’ve never entirely understood why he’s so certain of that. Maybe it’s more accurate to say he starts with quantum computation being impossible as his axiom, then works backwards to find what kinds of correlated noise would kill the schemes for quantum error correction and therefore vindicate his axiom.
He’s come up with various models. I never found them physically plausible, but at least he was sticking his neck out, which is more than a lot of quantum computing skeptics were doing. He was proposing models and making predictions. His prediction was that at the scale of 50 to 100 qubits and hundreds or thousands of gates, you’d see correlations in the errors. If you apply a thousand gates, each 99.9% accurate, the total accuracy wouldn’t just be 0.999 to the thousandth power—it would be much worse because all the different errors would interact with each other.
Now those experiments have been done—famously by Google, Quantinuum, USTC, and I believe QuEra has done relevant demos too. And again and again, this is not what we’ve seen. The total accuracy does go down exponentially with the number of gates, but it merely goes down exponentially—exactly the way the theory of quantum fault tolerance presupposed 30 years ago. If this is all that’s going on—simple uncorrelated noise—then quantum error correction is going to work. It’s merely a staggeringly hard engineering problem to build this at the scale where it works.
Over the last five years, Gil Kalai has been driven in a really weird direction where he’s basically saying these experiments have to be wrong. He keeps writing to the Google people requesting more of their raw data—he CCs me on the emails—then does his own analyses, posts about it on the archive. He keeps saying their 2019 experiment must have been fallacious. But in the meantime, there’s been a dozen other experiments by other companies all getting the same conclusion. He’s fighting a losing battle at this point.
The fact that someone of his capability tried so hard to prove quantum computing impossible and failed makes me more confident. If there is a fundamental roadblock, it has to be something totally new and shocking, something we haven’t foreseen and Gil hasn’t foreseen either—some change to the laws of physics that somehow wouldn’t have reared its head with thousands of gates but will do so at millions of gates.
The scientist in me hopes we’ll make that discovery. I hope quantum computing will be impossible for some reason that revolutionizes physics—how exciting would that be? But that’s not my prediction. My prediction is that the more boring, conservative thing will happen: quantum computing will merely be possible, just like the theory said.
Yuval: You mentioned experiments by Google, Quantinuum, and QuEra. I think you said that increases your confidence in the roadmaps. What is the next thing you’re hoping to see? An algorithmic breakthrough? Hardware scaling? Some new error correction code?
Scott: I’ll take any of them! Of course we hope for algorithmic breakthroughs, but those you can’t predict. We had no right to expect the quantum algorithms we already have. Some of them—especially Shor’s algorithm for factoring and discrete logarithm—still feel kind of miraculous to me. We had no right to demand these algorithms should exist. They just did. It so happened that problems of enormous importance for modern cryptography, because we decided to base internet security on factoring and discrete logs, had exactly the right structure for a quantum algorithm to deliver a speedup.
For more than 30 years since Shor’s algorithm, people have been trying to repeat that success. But a difficult question is: what’s even your target? What other problems are you hoping for a similarly dramatic quantum algorithm? The NP-complete problems have been the holy grail of computer science for more than half a century. It would be phenomenal if there were a fast quantum algorithm for all NP-complete problems.
But by now we have a conjectural picture where quantum computers can give speedups, but they’re either relatively modest—like the speedup from Grover’s algorithm—or special purpose, giving better approximations for some specific optimization problems and not others. It’s a complicated picture. I would love to see a revolution where quantum algorithms actually help for way more optimization or machine learning problems than anyone thought. I can’t rule that out, but you can’t bet on it either.
In terms of things we can reasonably foresee, there are certain hardware milestones that all of us in this field know to watch for within the next few years. One is a clear demonstration of fault-tolerant gates on logical encoded qubits in a way that gets a net win compared to just doing those gates on physical unencoded qubits. All the little demos people were doing 25 years ago—using Shor’s algorithm to factor 15 into 3 times 5, using Grover’s algorithm to search a list of four elements—that’s all going to have to be repeated, but now at the level of error-corrected qubits, at the level of fault-tolerant encodings. That will show us we’re on the right track.
The other clear thing to look for is demonstrations of quantum advantage with near-term devices that aren’t fully error-corrected yet, but that we can hopefully use for simulations of condensed matter physics, maybe even chemistry. There’s an order of milestones you could expect. The first step is getting any quantum advantage at all at estimating any kind of number in condensed matter physics, using a programmable quantum computer.
If you squint, we may already be there today. It’s always hard to say because you have to compare against the best a classical computer can do, and classical computing is a moving target. But at the very least, we can now compute numbers about the Fermi-Hubbard model or out-of-time-order correlators that we certainly don’t know how to compute easily classically. The quantum computer computes these numbers and you can check them by comparing to a second quantum computer, even if it’s too hard to check classically.
That’s already an improvement over five or six years ago, when the only quantum advantages we could demonstrate used sampling problems—which are very contrived. I can say that because I had a hand in inventing them. And even with a second quantum computer, it was very hard to verify what was happening. So that’s a first step we’re crossing now.
The next step will be calculating numbers where condensed matter physicists, chemists, or high-energy physicists will say: we actually cared about this number for reasons that had nothing to do with quantum computing. We wanted to know it, and now we have it because of this device, whereas we couldn’t have gotten it classically. I’m hopeful we’ll start seeing that within the next couple of years.
Then a third level would be calculating numbers that are not only scientifically interesting but useful to the battery industry, the photovoltaics industry, chemical engineers—things with commercial value. I really do think things will happen in that order. Those are the milestones I’m looking for.
Yuval: As quantum moves from being primarily academic to commercial, with government and national security significance, do you think it’s possible someone already has a Shor’s algorithm improvement that’s an order of magnitude faster and they’re just not publishing it? Do you worry about that?
Scott: That’s a very interesting question. I used to get it all the time—even 20 years ago, people would say, “How do you know the NSA doesn’t have a giant quantum computer in their basement?” I would laugh this off. It was like asking, “How do you know the aliens don’t have John F. Kennedy in their freezer?” It just seemed so ludicrously disconnected from what we see.
Quantum computing is still a small enough field that we know who a lot of the best people are. If there were some massive government effort, we’d all see them getting vacuumed up into a black hole, just like during the Manhattan Project. But back then there was wartime censorship. I like to joke that if you tried to start a Manhattan Project for quantum computing today, it would be about 10 minutes before it started trending on Twitter. #WhatsGoingOnInLosAlamos. We would see all these people disappearing.
That’s what I said for a while. I think today the situation is more complicated. I actually know there are people thinking about the exact resource requirements for Shor’s algorithm—for breaking various crypto systems deployed today—who have reached the point of wondering: should we publish this or not? Is this a good idea?
If you use the Manhattan Project analogy, that’s basically the threshold crossed between 1939 and 1940. In 1939, Frisch and Meitner were still publishing about uranium fission. In 1940, you get the first estimate for how much U-235 you’d need for a chain reaction—and at that point, it’s no longer published.
The thing you could start wondering about is: we’re going to have increasingly powerful quantum computers available on the cloud. These services will get more powerful, they’ll start supporting fault-tolerant operations. People will use them for interesting experiments, quantum simulations, maybe commercially useful things for batteries and photovoltaics.
But at that point, it becomes very hard to prevent someone from using these cloud services to do something cryptanalytic. Once you’re at a large enough scale with a general-purpose quantum computer, how do you stop someone from sneaking in Shor’s algorithm? So it becomes very relevant for cloud providers to know: at exactly what scale does that happen?
You can already glimpse what people are going to be arguing about. Do companies provide quantum computers at a certain scale but not beyond for national security reasons? Do they have to do some know-your-customer thing where they inspect the circuits submitted to make sure no one’s sneaking in Shor’s algorithm? I don’t think anyone has this capability today, but five years from now, these will be relevant questions.
It is weird for me to be thinking about this as a soon-to-be-alive possibility for the first time.
The eventual solution, which we’ve known about for decades, is for people to migrate to quantum-resistant cryptography. The best time to start thinking about that was yesterday; the second-best time is today. Taking these giant systems deployed across the internet and migrating everything to quantum-resistant encryption—conceptually it’s like the Y2K fix for those old enough to remember, but actually much more complicated. It’s not one straightforward fix; it’s upgrading to different methods of encryption, and it takes years to get right. Even if you only think the NSA or Chinese government might have a quantum computer in five or ten years, the time to start thinking about how we migrate to post-quantum encryption is now.
Yuval: It takes a lot of money to build a quantum computer and even more to develop one. This money comes from investors, and investors need to believe in a vision and a future. So what’s wrong with a little bit of hype to get the money you need to execute on the plan?
Scott: It’s great to be excited about what you’re doing. The word “hype” has multiple connotations. My personal line is: are you telling the truth to people? For me as a scientist, telling the truth means more than carefully avoiding false statements. The standard is higher. It means you don’t give people a misleading impression through strategic omission. You don’t let them come away with a false picture of what quantum computers are going to be good for. That’s been a constant source of tension between the research community and the commercial world for 20 years in quantum computing.
There’s a whole spectrum. Many companies are trying to be responsible—excited about what they’re doing, presenting an optimistic vision, but not saying, “You can already use a quantum computer to solve all your vehicle routing and flight scheduling and finance problems better than a classical computer.” That message—huge quantum advantage for all these prosaic commercial applications, available right now—is often what investors want to hear, what CEOs want to hear, what journalists want to hear.
The problem is that it depends enormously on just not doing fair comparisons against classical computers. Often not even asking: how well could a classical computer have done this same task? Did we actually try it? Did we do the comparison? Is it a fair comparison? Are we comparing this quantum algorithm only against brute-force search? Are we tying the classical computer’s hands behind its back? Are we comparing a quantum approximation method against a classical approximation method and seeing whether the quantum advantage remains?
For many years, there’s been this problem that to raise lots of funding, the path of least resistance is to tell people what they want to hear. “We have this heuristic quantum algorithm, we run it, and look—it recognized handwriting or made predictions about stocks to trade.” When people hear that and hear the word “quantum,” they think, “Oh, that’s the future. That sounds great.” And all of it relies on this strategic omission of the question: could a classical computer have done that just as well? Again and again, when you ask that question, the answer is yes. Even if you didn’t lie about it, you strategically omitted that crucial point.
What’s wrong with that? One obvious problem is that people will lose trust. This happened in AI multiple times in the history of the field: incredibly high promises, then a boom, then when those promises couldn’t be delivered, a bust. Everyone got hit, including people who were trying to be responsible, because people said, “Why should we trust you?”
Within quantum computing, if we just reward this sort of strategic omission, the winners become the least responsible people in our field, and the people trying to be more responsible lose out. That creates really bad incentives. Part of what I’ve been trying to do for the field on my blog for the last 20 years is just do what I can to keep the field honest.
Yuval: You mentioned AI comes intertwined with quantum in so many ways—funding, AI for quantum, quantum for AI. But I wanted to ask about ethics. I know you spent a couple of years on AI ethics and alignment. Is there something that could be taken from that and applied to quantum?
Scott: AI presents a really special set of issues because if there’s any limiting principle that limits how smart it can become relative to humans or what effect it’ll have in the world, no one has articulated what it is. It could be that all the jobs we do could be done as well or better by AI. Where does that leave us? Where does that leave civilization? Where does that leave humanity? These are absolutely enormous questions. You almost get vertigo thinking about them.
If your job is to engineer AI so this transition goes well for the human race, how do you even start on that? And even if you succeeded, how do you prevent everyone else from doing it irresponsibly? These are questions that go way beyond computer science—they’re questions for philosophers, ethicists, really for everyone in our civilization.
With quantum computing, both the good news and bad news is that we understand a lot more about what quantum computers will and won’t be able to do. We know they’ll have a huge impact on how we secure the internet, on currently deployed public-key cryptography. But that’s a bounded problem. Compared to what AI is going to do, it’s very well understood. We already know in principle what a lot of the solutions are—migrating to post-quantum cryptography. We just need to make those solutions happen. I wish we knew the solutions for AI alignment in any similar way.
Yes, there are ethical questions about whether we should build quantum computers if they’re going to lead to attacks on cryptography—they might even break cryptocurrencies like Bitcoin. Some might say that would be a good thing, but it would cause a lot of economic destabilization. Because in principle we know there’s a solution in post-quantum cryptography, as long as the transition happens somewhat gradually, we can migrate to those crypto systems and then be back where we started on cryptography, while getting all the wonderful advantages quantum computers will hopefully have for solid-state physics, designing new materials, and so on. That seems like the pretty clear ethical choice.
We’ve made analogies to the Manhattan Project, but a quantum computer is very unlikely to kill anyone, unlike a nuclear weapon—unless the dilution refrigerator tips over onto their head or something.
For the most part, I think of the ethical issues in quantum computing as continuous with the ethical issues in every kind of science and engineering. The ethical issue I’ve worried about most personally is the one you asked about before: how do we communicate about this field responsibly, telling the truth about what we understand, about what quantum computers will and won’t be able to do, not making false promises? I’ve been very concerned with that ethical question.
Yuval: As we get close to the end of our conversation, I wanted to ask about teaching. As the field has evolved, what changes do you see in the interest of younger people to get into quantum and what do they want to research?
Scott: Because of my blog, I have this somewhat unusual view where I’m constantly getting emails from high school students and undergrads who want to get involved in quantum computing, who are looking for guidance. It’s way more than I’m able to deal with. I want to hand a lot of these students off to my colleagues. I wish I could talk to every one of them one-on-one, but it’s just no longer possible.
For students interested in computer science, physics, the future of technology—it’s entirely understandable that a lot of them would get excited about quantum computing. That’s been the case for a while. Maybe there’s even more of it now that we have all these quantum computing startups and the field has started to have some commercial prospect.
Sometimes the guidance I’m giving is students get involved and it almost breaks my heart, because they look around at what’s out there and what’s out there is: “Let’s use a quantum computer to recognize handwriting or do something a classical computer could have done just as well, then publish that we used the quantum computer and ignore the comparison to what a classical computer could do.” There’s so much of that stuff. There’s a very low barrier to entry—if you’re an excited high school student or college freshman, that’s very easy to start doing, almost irresistible.
Yet I feel that’s not the path where we’re going to solve the hard scientific problems that need solving. That’s a harder, rockier path where you have to really think: what’s the best I could have done classically for the same task? Is this quantum advantage for real? Is it just an artifact of my experiment? Is there a real asymptotic quantum speedup? Does that speedup survive for the end-to-end application that would matter in practice?
I’m torn because I don’t want to dim these students’ enthusiasm. I want them to be enthusiastic. But I also want to direct them to the real scientific challenges rather than the cargo-cult stuff. That’s often a conversation I’m having.
Another conversation: students want to know what to major in. Computer science? Physics? Electrical engineering? Which courses to take? Those questions are almost impossible to answer without knowing a lot about the student—how do they want to contribute to quantum computing? Do they want to do hardware? Algorithms? Depending on the answer, I might suggest different fields.
But the one thing I constantly tell high school and college kids is: learn linear algebra. Level up your math skills—linear algebra, probability, classical CS, classical algorithms—because all of that is going to be unbelievably useful if you go into quantum computing. It’ll be the language with which you think about everything. And it’ll also be useful if you don’t go into quantum computing, if you go into AI instead. It’s really a win-win to build up math skills, especially linear algebra, if you’re interested in this field.
Yuval: Last hypothetical: if you could have dinner with one of the quantum greats, dead or alive, who would that be?
Scott: Maybe Schrödinger.
Yuval: Why?
Scott: You can read him from a hundred years ago, when he introduces the wave equation—what we call the Schrödinger equation, which all of quantum mechanics and certainly all of quantum computing is based on. Almost parenthetically, he says: if you have a single particle, you need a function psi of x that tells you the probability amplitude that the particle will be at any given location. But what happens with multiple particles? The only reasonable choice is that you need one giant function, psi of x, y, z, for all particles together.
He completely realizes he’s exponentially enlarged the state space of physics. He’s introduced this vastly larger object than physics had ever contemplated. And why is this the right choice? “Well, there are several considerations, among them that I couldn’t see any other way to make the probability sum to one.” It’s almost an afterthought.
I’d want to talk to him about quantum computation, get his thoughts. He later expressed a lot of displeasure with quantum mechanics—I think he even said he was sorry he had anything to do with it. He was an ally of Einstein in hoping for some local, realistic description of physics, that maybe quantum mechanics would only be an approximation to something else.
I’d want to know what he thought about his equation still standing 100 years later, just as he wrote it down in 1926. I’d want to talk about all the big foundational questions.
He also became sort of a mystic later in life. He wrote strange things about consciousness, but also about what is life—where he directly inspired Crick and Watson to discover DNA’s structure. He clearly had a pretty strong batting average thinking from first principles about how things could be. I’d want to talk to him about what we know as of today and what his thoughts are.
Yuval: Scott, thank you for everything you do and for spending some time with me today.
Scott: Of course, it was a pleasure.
Yuval Boger is the Chief Commercial Officer of QuEra Computing.