Why Quantum Computing May be Overhyped but AI isn’t: Interview with the NSA

Welcome to Geopolitics the Candid. I'm Dmytrao Parvich, Chairman of Silverado Policy Accelerator, a Geopolitics Think Tank in Washington, D.C., and also an author of an upcoming book, World on the Brink, how America can be China in the race for the 21st century. Preorder links for the book are available in the show notes, so please check it out. My guest today is Gilbert Herrera. Gill is a Director of Research at the National Security Agency, the NSA. His Directorate at NSA conducts research to develop new technologies to enable the agency's signals intelligence and cybersecurity missions. He also serves on the U.S. National Quantum Initiative, the advisory committee which advises U.S. leaders on quantum computing efforts. Gill, welcome to the show. So Gill, we had a chat recently on the topic of quantum computing, and I want to have you on the podcast since there are very few people, I think, in the all of you as government, who know more about the topic of quantum than yourself. So perhaps we can start with the basics, and you can give our listeners a quick explanation of what is a qubit, and what are the various applications of quantum, how is quantum computing different from quantum sensing and quantum communications. Yeah, great. Let me start off with a brief tutorial because what most people say is a qubit is like a bit except it's one and zero at the same time, and that's really a little bit inaccurate. And let me go, it'll be slightly more technical detail. But the way a classical computer works is it has these things called bits, they're one and zeroes, and you could have a whole bunch of them together, and you could perform logical and mathematical operations, right? You can do ends and ores and knots and other kinds of logic gates, and you could do arithmetic on them, and you get an answer. What a qubit is, is it's a electronic, it's made like an electronic device, but what it really represents is something called the wave function and quantum mechanics, where it can be in a whole bunch of states simultaneously. In this case, it can be not only a one and a zero, it could be a negative one, it could be an imaginary number, but it can be in a number of states when you put it into what's called a superposition. So it can represent a whole bunch of states, and you could also do logical operations on it, as long as you don't read it. And so you could manipulate it in the superposition of states, but when you go to measure it, you get effectively a one or a zero. So if you have 10 bits, then you could represent two to the tenth state, so roughly a thousand and twenty-fourth states. If you have 10 qubits, once you measure it, you get the same number of representation, but you can get one of, you know, 1048 bits as an answer. But the real key is when you're actually manipulating the data because you're in the superposition of states, and because of your ability to entangle, that's another quantum mechanical term, these states, you can really do very powerful things, as long as the qubits can live long enough in order for you to complete your calculation, and as long as you don't have errors bump into the equation. And that's where another key difference between the two types of devices are with a bit represented by a transistor in modern high performance computers. The error rate is about ten to the minus thirty, which means that it takes a really long time for an error to happen. Now when you have billions and even trillions of transistors, you know, running at, you know, four gigahertz, then on a large high performance computer or in a data center, you're probably going to have an error every day. But, you know, still, errors are very rare, and they come up with ways to mitigate them that are relatively straightforward, like in the case of memories that have error correction techniques that are relatively low overhead. Today, the best qubits have error rates between one percent and a tenth of a percent. So they're making errors all the time. And there, so that's a real challenge with quantum computing that, you know, you have this huge advantage that you can get because of the ability to go into this superposition and then entangle with other qubits, but they're really prone to errors. So that's kind of the initial primer in quantum 101. I guess I should say one of the thing about them, quantum computers are inherently linear. And so there's a lot of cool things you could do, but if you want to predict whether or do climate forecasting, they probably won't do a good job because those are nonlinear phenomena. And it's not that you couldn't eventually run codes, you know, for weather prediction or climate modeling, but you'd have to run a quantum computer in a mode where it'd be much more efficient to run it on a classical computer. Let's get a little bit later to the applications, but I just want to set the basics for people. So that was a terrific explanation of a qubit, but you can do different things with quantum, not necessarily just qubits. Obviously, there's quantum computing. We'll talk about that a little bit later. But what is quantum sensing and what is quantum communications? Yeah, so quantum sensing is a technology that we already have today. And the best example of a quantum sensor is an atomic clock. And so atomic clock works by having a qubit and measuring the qubit frequency to extremely high precision. That's effectively some of the research that Dave Weinland, who won the Nobel Prize in 2012, does. So the next extension, once you have a highly accurate clock, is now you can make a very sensitive and gravimetric sensor. And so in Dave Weinland's Nobel lecture, he actually demonstrated a sensor where he had a very precise qubit-based clock on an optical table, and he had his graduate student jack it up eight inches and was able to measure that eight-inch rise on the basis of the change in time, the relativistic change in time. And today, with even more sensitive clocks, you could measure centimeter scale, even millimeter scale changes because of the change in time that is so accurate. That's one example of the sensors. And both of those exist today. There are some deployment problems because you need a relatively large clock, but you have that ability to sense today. In the future, we envision there will be other classes of sensors. Some people have proposed quantum radars that probably is a little bit further off. But those are examples of a quantum sensor, including quantum-- And what can it be very precise measurements? And what would a quantum radar do that a traditional radar is not capable of doing? If you could make a quantum radar, then it probably would be resistant to some of the stealth attacks, but using stealth materials. But that right now is highly speculative because of the nature of what you have to do. You have to entangle photons, send one back, and then have delay loops where you could compare the response. But you could just have much greater precision. That's a more relevant application of something that you could do with quantum sensing is magneto and siglography, which is-- That sounds really complicated. It is. Well, what it is is it's a magnetic resonance imager that would be about one or two orders of magnitude more sensitive than MRIs. And you could do it without having to go into a really large chamber. So there are techniques for doing medical imaging using quantum mechanical devices that are quite a bit operationally easier than MRIs and a bit more precise. If that's still a research stage, but it could lower their price and improve the quality. And other examples are-- Is it fair to say, Gil, that the examples you're describing certainly sound very useful, but they're not necessarily, at least to me, they don't sound revolutionary. They may be evolutionary where a radar is better at detecting stealth aircraft or an MRI is better because you're not going into big machine. But they're not sort of orders of magnitude better where our entire world is going to change because we have this technology. Is that fair to say? Yeah, I think that's fair to say because it's all about the cost differential and they're not significant as significant as it could be. Got it. And then the third area of quantum that people often talk about and the Chinese are particularly enamored with it is quantum communications. Can you talk about what that is and why that might be useful or not? Yeah, and let me break that into two distinct categories. One is a quantum network that shares entangled quantum information and a lot of people are excited about that. At this point of time, that really is at the laboratory experimental range, you can share entangled quantum information over an optical fiber for as long as a single fiber extends new quality degrees and you get errors. And because of that limitation, utilizing it for a kind of classical information sharing, it probably is not very doable. And this is sort of the idea of like this teleportation almost where you can be on different sides of the planet and miraculously exchange information without a line of sight or anything like that. Well, no, that's a little bit different because with teleportation, you can entangle so you can bring that up. So here is what how people believe you could violate the speed of light is if let's say that I entangled some photons and I gave you a box of entangled photons that were entangled with mine and then you got in a spaceship and went to Alpha Centare for light years away. And then the fact that we had these entangled photons, people believe that because if I open up the box, then yours will automatically either be the same or the opposite depending how we entangle them that we could communicate. But the problem is that you would not know when I opened the box and read the information. So with all examples of when you use entanglement to try to communicate, you need a classical information channel to communicate. So that's not what I'm talking about. What I'm talking about is you can entangle states and communicate that and where that would have value is for quantum computers to communicate with each other. And the other value would be in scientific exploration. So on some of these experiments at the Department of Energy runs like gravitational waves, other kinds of physics-based experiments, it may be useful to share tangled information for science. But there is another application that has been available commercially for about 15 years. And that application is something called quantum key distribution. And quantum key distribution is a way to use the spooky action at a distance, right? The ability to entangle particles and put them into a superposition as a way to share keys in a high secure manner. And the way that works is I will entangle photons. So I know the state and I will share one of the entangled pairs with you. And you have to randomly pick how you're going to read it, what phase you're going to read disentangle pair on. And then I encode mine using random phases. And then at the end of that, on a classical channel, you and I would share or you would share with me how you read that information. And so I wouldn't then know what information you got that is legitimate. And then you would utilize or I would utilize that information to send an encoded signal to you on a classical channel. It's a little bit more complicated than that. But if somebody, if there is a man in the middle attack and they read and retransmitted the information, they wouldn't know what phase I encoded the information in. And so only 25% of the information would get through. Whereas with the information I sent, 50% of the information would get through and you could detect the man in the middle. And that's how quantum key distribution works. It's a little bit more involved than that. And there are companies like ID, which is out of Austria that sell systems that utilize that principle. But it has two big problems. One of them is it operates really slow. Like the early systems operated at 300 bought. Oh wow. At 300 bits per second. And it's like a really, really old modem from the 1980s. The I silent 700. I remember I had one of them. And so that operates really slow for a key exchange. But you still need the classical channel. But there are other reason is if you can Google on open source websites and find out that it's really subject to site channel attack. And also it's really easy to do denial service attacks. So it's really hard and cumbersome to operate. There are some use cases like ever before. And the denial service attacks is because if there is man in the middle, you alter in the state and you're just disrupting the whole channel, right? Well the denial service is if you know where the channel is, you just flooded with photons and you'll add noise to the system and they won't read anything. It's real easy to jam that kind of channel. And then of course the side channel is you alter the classical hardware and you can manipulate it and make it just not function properly. And then you still have to use a classical channel to encrypt the data. So all you would use the quantum key distribution for is to share keys like maybe an AES key, but then you still need to encrypt using AES on the classical channel. And AES for listeners that don't know is this symmetric cryptography where the same key is used on both sides versus what's known as public key cryptography where there are two keys, a private key you keep to yourself and a public key you can distribute to anyone that people then can use to encrypt information for you. So this sounds like quantum communication of this application of it is really designed to sort of replace public key cryptography where key distribution is going to be done via this quantum channel. Is that fair? Yeah, that is fair. And again, the limitation is because you need line of sight, you either have to have a direct laser communication and that's what the Chinese do. They have a satellite and then they trust the satellite or you have to have a single optical fiber between your notes. And so the use cases are very limited. Like I believe Deutsche Bank uses this method to go bank to bank. So they use physical security to secure the quantum hardware and then they have a least channel that they're the only user of in order to share the entangled information and all their banks are within about 30 kilometers of each other. Yeah, but it certainly sounds the way you describe it is it's a solution search for a problem because there are many ways to solve this issue of key distribution that have been used for many many decades public key cryptography of course is the most popular one but you know US military and others have used it through traditional means of like you can just preload key books you know and ships for example right and use those without having to distribute them in real time. Yeah, that's exactly right. So we talked about quantum sensing which sounds like it's a useful application but perhaps not a revolutionary one and some of the more interesting applications like medical and radar are still some years away. Quantum communications is sounds like an interesting research experiment but doesn't have a lot of practical uses but quantum computing is something that a lot of people talk about as something that will change our entire world and it's going to solve all these different problems that are really difficult to solve now. So let's talk a little bit about that one thing that often comes to the fore and why presumably you and your agency are interested in quantum computing is of course the ability to break some types of cryptography right not all and there's something known as a short algorithm that effectively was I think it was discovered in the 90s early 90s that you could run on a quantum computer to break some types of public key cryptography and obviously we can see why it would be useful to an agency like the NSA but what are some of the other capabilities and applications of quantum computing. Yeah, quantum computing once there they exist because they really their quantum computers exist today yes but they're just not useful we got a ways to go but in addition to short as algorithm that you mentioned it was in developed in 1994 and what short as algorithm can do is to find prime factors and extract discrete logarithms and that's the basis of RSA and other kinds of public key but in addition to that there are a couple of other areas where it could be revolutionary first and foremost it's in conducting physics experiments because since you created this device that is really a quantum mechanical device there's a lot of opportunity in exploring the realms of quantum mechanical physics and that was one of the original applications so that Richard Feynman proposed when he first started talking about quantum computing in 1980 so you can explore the frontiers of physics and physicists are really interested in that and this issue of errors that I mentioned earlier don't matter as much because if you get information and if you nanoseconds then that could really help you resolve physics questions. Can you give us an example of a physics problem that you could potentially solve with quantum computing? Well again these are explorations of physics it's things like validating some of the fundamental principles that underlie the theory of quantum mechanics. These are really fundamental explorations of physics as an instrument it's kind of like with high energy physics it's understanding particles and you could use them to test out theories you have about quasi particles. It really is fundamental physics exploration so it's a scientific instrument like any others. Yeah research but the other area that has high interest and we know it could work on is chemical and material analysis and design because when you can using a quantum computer once perfected you can use it to simulate the binding energy of electrons in molecules and the real killer app that people talk about with quantum computing deals with something called the Haber Bosch process. So somewhere between a few percent and maybe 5 percent of the world's energy and about that much of the world's natural gas is utilized to make ammonium nitrite fertilizer. It's this massive industrial process and that's done at a high temperature and high pressures utilizing a process developed by Haber Bosch about 110 years ago. Now there are some plants that can do that at room temperature, at room pressure and make you know nitrogen fertilizer for themselves and if we could really understand this it's called the nitrogenase molecule. We could really understand the nitrogenase process then we might be able to save huge amounts of energy. And likewise understanding this binding energy molecules we could do pharmacological simulations that are much more realistic than what we could do today. So those are the really big applications that you could do that we know will really work and they're big deals. They just require really big quantum computers. There are a lot of other algorithms that have been developed but most of them are of interest only to the purist of mathematician. They're the kind of things that you've heard about the Google quantum supremacy experiment that were solved in that. It's like a random probability distribution of a random quantum circuit that has no meaning other than it's mathematically interesting. There are some algorithms like the Grover's Search algorithm that is represented speedup but it's not exponential speedup but there are smart people that are looking for algorithms all around the world and I'm confident that at some point somebody will find something other than factoring and chemical simulation and physics. So there is a whole range of problems in computational complexity theory that are known as NP hard problems. Problems that we don't believe you can solve in polynomial time or at least we haven't figured out or solved in polynomial time meaning that on modern computers at scale these problems are effectively unsolvable and a classical example of one of those problems that would be incredibly useful to solve. For the world is the so-called traveling salesman problem where given a list of cities and the different distances between those cities can you find the shortest possible route for traveling salesman to visit all those cities so you can imagine numerous applications of this to logistics and all sorts of problems but those are the types of problems that we don't know if we had a huge quantum computer today we would not necessarily be able to solve a problem like that any more than we can with a traditional computer is that right? It is and it's interesting you mentioned that problem in particular. There was a company formed in Canada about 15 years ago called D-Wave and they built what's called a quantum annealer and the quantum annealer was designed to solve exactly that the traveling salesman problem right and there is a technique called simulated annealing where classical algorithms work in optimization problems like traveling salesmen but they just haven't made the progress needed to solve that class of problems. I have some colleagues who are two experts in quantum annealing and they question whether or not you could make an annealer that could solve the traveling salesman but that is one of the problems that a lot of people are putting a lot of attention on and I'm hopeful that at some point in time there will be optimization algorithms. They may not have exponential speed up hopefully they'll at least have polynomial speed up but there are people exploring for those classes of algorithms but there really aren't any that I believe have been found that would give exponential speed up on that class of optimization problem. And just to underscore this point, you know if quantum computers were to appear today once that could actually do useful things they would not replace our traditional computers right you would still be running excel at PowerPoint and your traditional computers you would not be running it on a quantum computer you would really be using this quantum computer for these very specific things like trying to break public key cryptography doing these physics experiments, material sciences and so forth right? Yeah no that's very true. Now there is a lot of hope that they'll find other algorithms. You know even the Cleveland Clinic, bottom entire IBM quantum computer that they have in their facility or at least it I think IBM is leasing them not selling them. And so people are looking for other applications. There are people that believe you could mix quantum and AI and so they're a huge area of research interest but at least to this point I've read a lot of papers that you know suggest that they're making progress but I've seen nothing that's definitive. And right now the big known application is shores algorithm. You know it's been around for over 25 years and you know almost 30 years now and you know nothing else has really come around that is that revolutionary. And isn't it amazing that you know you've had computer scientists, you have physicists that have been working on trying to find those algorithms for many decades as you say and we haven't really found that many algorithms that you could run on a computer even though we don't even have one yet. So I mean how likely is it that you're going to solve a whole bunch of these NP hard problems if you had a quantum computer in the coming decades. You know I know I'm asking you to speculate here but any thoughts on that? There are hundreds of the smartest people in the world looking at this and I think we do need to particularly when it comes to shore and you know at some point I want to talk about quantum resistance that we need to assume that at some point people are going to find utility with a quantum computer. I personally think that it's a it's a ways off because we don't have a computer yet right. But right now it's highly speculative right you're trying to guess what you could do and something that doesn't fully exist. But I think about going back to classical computers the very first ones were optimized around one problem, one problem only and that was solving ballistic equations to support the war effort. And the first programming language that was developed was cobalt because somebody said if it could do complex mathematics in business we need simple mathematics. And so cobalt came and foretran shortly thereafter and you know within 10 years of the first practical classical computer people thought man this would be a great command and control the device and then they came up with this big network for you know the early detection of Russian aircraft coming over the pole you know through Canada to attack America when this massive command and control system and it went on and on and on. So what what I'm expect to happen is what happened with classical computer that they're going to optimize the first one about problems they know they could do. But once it gets in the hands of really smart computer scientists and smart you know electrical engineers and physicists that then they'll be able to find applications but you know that's far off. The people today are even trying to find applications for these very noisy quantum computers you know that started off in what John Preskill who's really one of the key scientists in the history of quantum computing called the NISC noisy intermediate scale quantum area and that's what we're in now and there are people that are trying to find applications by coupling these noisy quantum computers with good classical computers and see if we could use a quantum computer as an accelerator but to date there's been nothing revolutionary that I've seen. But let me push you on that a little bit gill because I'm not sure that that analogy with the invention of the classical computer works quite well because you had people like Alan Tureyne and Claude Shannon and others who had developed a lot of the foundational theories information theory and computing theory long before we had years before we had any workable computers right so theoretically when you a lot of what could be done with those computers even before we had one and it seems like with the exception of a few examples like sure we don't really have that with quantum even though we've had even many more decades imagining a quantum computer and trying to find those algorithms so it seems like this isn't even harder problem than it was on classical computers is that true? I think it's true to a certain extent and quite frankly quantum computers still utilize Shannon not so much von Neumann but you know quantum information theory is based on Shannon's information theory. I think what I was more pointing at is that it's hard to do innovation when things are theoretical and even though we had these fundamental basis you know of the Turing machine and we had information theory and we had Boolean algebra right and that went back in ten years even before yeah and the work of Babbage right so we had that fundamental basis but using probably best examples command and control it's not clear to me that the fathers of the Shannon and von Neumann envisioned computing for command and control and I do agree with you it's a much easier problem particularly because the interface to the real world of a classical computer is exactly what our information all information technology is based on communications how we do automation and process controls data all of those you know talk the same language right they talk ones and zeros and voltages and quantum computers don't do that so it's a harder problem but I still believe that it's a lot easier to innovate when you have a bird in the hand instead of a superposition of birds in the bush. That's a great one so let me ask like a more fundamental question I completely understand why you and your agency is so focused on this ability to break communications is really the bread and butter of what NSA does but you know should the rest of the world be as invested in that and actually let me even ask for you guys as you're moving to quantum resistant cryptography their new standards that are going through the process of being adopted now for replacement of algorithms like our say an elliptic curve cryptography which could be broken with Schorz algorithm you know presumably over the coming decades there'll be replacements of those algorithms out in the wild and if you had a quantum computer at that point there would be fewer communications that you could break you could obviously go back in time and look at historical stuff as well but even for you do you think it's really essential to have a quantum computer. Let me if I could just kind of trace the history of when and why NSA got involved in quantum computing so Peter Schorz published his memo back in his two papers the first paper was the algorithm the second was error correction which is really needed and within eight months of that publication NSA through something called the laboratory for physical sciences which is a joint NSA University of Maryland research institute already published a broad agency announcement looking for proposals in quantum computing and over the years NSA has been the largest sponsor of fundamental academic research and some industrial research across the globe. So we have paid for the education of hundreds of graduate students and hundreds of postdocs and we came up with the Transmon qubit through this research which most of the companies are using we funded ion trap research I mean the very reason why I'm at NSA is my fabs that I ran at sandy national labs made ion traps under I art sponsorship for the world you know all over the world so and the reason why NSA got in and was stayed through the various down cycles of quantum computing over the last 27 28 years is because we need to know what's happening in the world largely for the defensive mission because if we don't know what we need to be able to predict when we think a quantum computer will be available so we can start defending the networks so in 1999 NSA started research in quantum resistant algorithms we had vetted algorithms by the mid 2010s and we already started to deploy them for key national security systems because I personally don't believe there will be a quantum computer for many years I think 10 years is optimistic but if there's a black swan event we need to be prepared for that eventuality and that's why the president signed out national security memo 10 last year and all government agencies and national security systems have to have quantum resistant encryption by 2035 and in part because of course of recording you know if our adversaries are recording our encrypted information then they could go back and break codes you know if and when a quantum computer exists but that's why we need to be in the game so we know obviously that's what we fund the research so let's talk about you you mentioned that you don't think it's going to be on the scene anytime soon and you know if you read the press releases of the various companies and research institutions it seems like it's right around the corner remember reading last year press release from IBM talking about 400 qubit computer that they were able to put together and it seems like it's increasing very very rapidly I think they went from something like 120 qubits or they're about 200 qubits very very quickly so why do you believe that it's still so far away when at least based on marketing it seems like it's going to be solved any day yeah I should point out that that annealing company I mentioned D wave in 2015 I believe it had a 5000 qubit machine on the market and so it's all a matter of quality and connectivity so the and I'll put it in context of going back to eniac and then I'm going to talk a little bit about modern high performance computing and then quantum in the context of error rate for the fundamental unit in the case of classical the transistor in the case of quantum a qubit the error rate not failure rate a little bit different thing an in the act was that first general purpose computer right yes first general purpose computer 1947 roughly so it had an error rate of about 10 to the minus 15 and that was perfectly acceptable for the class of problems that they wanted to run they could execute the algorithms before errors became a problem with a modern high performance computing the figures that that my colleagues at sandy labs believe are true and I accept it is about 10 to the minus 30 so an individual transistor hardly ever makes an error because large high performance computers have a lot of these transistors and they run up for gigahertz then you have to save the register state about once a day it might be a little bit less than that for modern HPC because you'll probably have an error every few days okay and with a qubit the error rate the IBM machines to the best of my knowledge is about 1 percent so about 10 to the minus 2 they might be approaching 10 to the minus 3 but you can see this radical difference between what past machines have had and what modern day quantum computers have is that the error rates are just so high and part of the reason for that is with a qubit you're effectively measuring the spin of a particle or whether or not it's nuclear state is spinning upwards or downwards you're measuring the state of an electron and with a transistor you're looking at energy bend levels that you're switching by putting an applied field on so they're just fundamentally different and so it's really hard to keep errors out of quantum computers and that's why it's taking so long so IBM has an incredible team I know the people I met with them and what their their approaches is they're going to simultaneously scale the devices make bigger devices and make effort to cut errors out they're also for the smaller machines employing different techniques where they hope they can find useful applications without doing classical error correction on the machines and I encourage them to keep on doing what they're doing hopefully they'll make progress but I'm just a bit skeptical that until we find a way to make qubits you know focus on the fundamental research to make individual qubits better it's going to be very difficult to scale. When we were talking earlier Gail you told me that you spent a lot a lot of your career working on traditional semiconductors and you told me that while that is an incredibly hard problem of how do you keep shrinking those chips we mostly know what we want to achieve right and we're just trying to find the right tools to figure out how to get there and you said like in this area we don't even know what a qubit will look like that is the right qubit is that correct? Yeah and the way I describe it is more long and this is an argument that I had with my physics friends I should disclose I'm an electrical engineer but I spent a career scaling my electronics and the way we scaled is we had a fundamental underlying theory called semiconductor device physics and what semiconductor device physics is is a set of theories and models wherein you could simulate a device and know with you know high reliability that if you can build the device you simulated it would work and by that I mean you would put into these models the fundamental dimensions of the new transistor you're going to make the properties of the materials you know the electromagnetic fields you know all these properties and you could model it and then if you can manufacture what you modeled then the transistor would work and that field of research is called semiconductor device physics the oldest book I have in device physics is from 1964 my favorite is one called by author called Z that was first published in 1968 I have the 81 edition on my desk he last updated it in 2012 so very solid underlying theory the the challenges that most physicists believe because we understand quantum mechanics we could model these devices and I don't believe that's true because the corollary in classical computing of quantum mechanics is Boolean algebra right you get the speed up the exponential speed up because you can put it in this superposition state and entangle it and be able to address a you know a large number of simultaneous test a large number of simultaneous solutions but what we don't have is a theory where you could say I'm going to build a superconducting qubit and the Josephson junction is going to have this dimension and the period of the materials is such and such and it's going to be exposed to this electromagnetic field and at this temperature we can't do those predictions and therein lies the problem so we're in a mode now where we're doing design build test you know with these devices and and it's not that you can't do design build test a good example of that is a steam engine was improved you know by James Watt you know 250 or so years ago without understanding thermodynamics but if we're going to have to do the design build test with quantum computing it just going to take a long time and one thing I'm encouraging is for somebody to be the next Z somebody to just try to relate how a qubit will work on the basis of its size the materials properties and all these other parameters and then that way we could truly model them classically and then build them and realize them quantumly so Gill I would be remiss not to ask you about the hottest topic in the world today in the world of technology at least and that's AI and it seems like a lot of people that were so excited about quantum over the last decade and now jumping on the AI bandwagon is that also technology that you think is overhyped we talked about error correction problems in quantum we also have errors and hallucinations in AI how do you think about the prospects of AI both LLEMs and other AI technologies that are coming out yeah I totally agree with you that there's a lot of hype but unlike quantum AI is actually real and it's already being used in some cases and so at NSA we've been using AI for some time and largely in Lagouon's translation in fact when you utilize your phone whether it's you know Apple or you know Android they're already using AI for things like language translation and by the way Michelle Ernie has been around for many many decades I've been using cyber security field since early 2000s so it's really not new yeah and in fact it's got to ask this question yesterday when I was given a presentation and I just made a big mistake right I called it AI and I finally gave in the Metri because AI is now branded as AI and ML and AI has been in use too things like expert systems when you call Walgreens and have to talk to a robot for a while that's an example of an expert system which is NEI tool machine learning you're right you know pioneers in cyber security like you have been using ML you know to analyze you know information you know in files and metadata and what not in order to find a tax and what not now the big thing that's happened happened last year when chat GPT was released in the wild and the difference between that and other ML algorithms that are highly tailored for a very specific application is by throwing 10,000 really expensive GPUs and a data center and 72 terabytes at data in a really really large matrix emergent properties came out it appeared that the machine could communicate maybe generate original text and thoughts and that was revolutionary it it I think caught most people on guard off guard including me now once you start a play with it you realize that there are these things hallucinations where it'll just spit out data or where it'll invent facts things of that nature and that's a real problem so there is a lot of hype around AI there's a lot of promise you know we're already using it for some problems and I think the intelligence community and government and the private sector and everybody's over time we'll find applications for large language models that are real I think that'll happen a lot sooner than a useful quantum computer but it's not ready for prime time because of this error problem and and where is when you're playing with AI at home like most people do you ignore the errors and if the error rate is 10% which is it's at least that you just kind of slough it off but for real world applications the error and hallucination or hallucination rates are probably going to have to be closer to the eniac number 10 to the minus 15 then they are now which is probably about 10 to the minus one or 10 to the minus two at best well and we were talking about this earlier but part of the problem with these hallucinations is they're so confident about them right if you're talking to Chad Gbt and it's giving you total nonsense and answering your question it is very very confident that nonsense and if you are someone who actually doesn't know the topic you may believe it and act on it and all kinds of disasters may ensue. No you're exactly right that's why at NSA and a lot of other places we're working on explainability and also error detection and unlike quantum computing which could benefit you know from information theory you know from Shannon and Hamming and another figure we haven't talked about nobody has any clue how to do error correction with with AI I mean the approach that the private sector vendors are doing that have published LLMs is when an AI does something that's really crazy they're putting in code to detect that so like if somebody is asking a question that they shouldn't ask you know like can you make me tile pornography for example then there are codes that will detect that and prevent the AI from responding. But that's sort of patch on a system that's you can't you can't stop all of that through those types of patches. It's not scalable it's not economically viable so there's no theory behind this yet and so as I mentioned in quantum computing the quantum error correction codes look a lot like they're well they are derived from Shannon but we don't have the leverage in AI that we had in quantum computing to do things like error correction and that's another healthy field of research explainability that there are people that are like running the model twice or having two models and you run the same thing through and do a comparison of answers but those are that's very early on in the field and and those are two big problems that we're going to have to get through but what I predict will happen is because you know an AI can generate tech so rapidly that one potential use case is to have it generate something very quickly something that a human might take a week to produce and they could produce it in seconds and then have processes where multiple humans will have to check it that may be an application but we do have a way to go the vendors will tell you that with each new model they're getting better and better the cost of entries can it be high you know whereas chat I guess GPT-3 used my understanding was about 10,000 A100 GPUs GPT-4 used something like 30,000 you know H100 GPUs the unit cost of those is like 32,000 and then those are those are those are the Nvidia GPUs yes and when you add to that the cost of the memory the cost of the server farms that had to process it and then the high speed interconnect you're talking many billions of dollars for the machine and I don't know what the cost of compute would be but it's in that scale and so it's it's a wild wild west but the difference I think with quantum and eight between quantum and AI is AI actually does exist we're going to have to find ways to harness it notwithstanding its frail piece but the concern I have is one of the first applications was improving fishing attacks right and so if you have somebody who doesn't have a good English skill and they write a you know Nigerian prints you know letter one of the early use cases is you can use GPT to improve the quality of that and that's unfortunate because an adversary you know a hacker can deal with this 10% error rate and hallucinations but we can't and if you look at use cases for AI and where we're using this for text generation and how concerned are you about this problem of garbage in garbage out because the eye is not magic right you train it on data and obviously there will be specialized AI models on very particular data sets like for example medical records to help predict diseases where things will probably work very well especially if you have a human reviewing the outcome but on these large language models where you're sort of training it on the available universe of knowledge you know today that universe of knowledge is largely human created and you know it comes with its upsides and downsides some of the information is right some of the information is wrong but you know that information universe is about to blow up dramatically with increased use of these LLMs where you could see you know in a few years the vast majority of data that is produced will be produced by these AI models that you're going to be then trading future AI models on and perpetuating mistakes because of the training data that's going into those models is that a concern for you? No it's a big concern I think the future for applications where accuracy and reliability are high are going to be leveraged some of the large commercial models for the emerging properties like the ability to interface with a person but then supplement it with additional training on highly curated information right so you you have a model that has a lot of good information in and a lot of garbage in and then you supplement it with highly curated good information and then ask it questions that relate to the supplemental information that you fed it and then that way you could you could benefit from the emerging properties like the ability to communicate with humans on human terms and then tailor the information the curated information so the answer has come from a subset of information that you know to be true. Let me ask you the last question here and this is going to be a hard one because I'm going to ask you to speculate here but you've already speculated about the prospects for quantum computing in the next decade some of the problems that we've discussed in terms of the usefulness of quantum computing sending aside breaking cryptography the physics modeling problems the material sciences some of the medical applications and the quantum sensing there are also a lot of people working on AI solutions for that right a lot of applications of biotech using AI a lot of people trying to find new materials using AI algorithms looking at protein folding methods and medical applications where do you think we're going to get successes first with a quantum computer in some of these areas or with AI? Probably with AI I mean the protein folding work has been astonishingly good so and I think that's an example where you had a highly curated data set and you have people interpreting the data that are subject matter experts and so I think they'll probably be more promise in that area there already has been some positive results in the medical area by looking at you know x-rays and other kinds of imaging where today we we there have been some false starts to but I think AI will will have results in those fields you know much sooner I think there'll also be a lot of bad data that comes out right you know bad guesses but I've been impressed enough with some of the protein folding work that I think that as we learn better how to curate information learn better how to build in the model algorithms and remember that you still need human subject matter experts to interpret the data that that will probably make progress much sooner with AI than with quantum. Well Gil you've been incredibly generous with your time just a really terrific highly highly interesting discussion thank you so much I assume that if there are listeners out there who are highly interested in quantum physics, quantum computing they're doing cutting edge research in our US citizens I presume that your agency is hiring and would be highly interested in talking to them right. Yeah and in fact I should have made two pitches one come one come all to NSA again citizenship required but also I mentioned we fund all this academic research if you just google LPSBA then you'll find the broad agency announcements that we have for researchers who want to make proposals and quantum computing and their citizenship is not required we actually fund multiple foreign research institutions and you know because we need to we want the best people in the world working on these fundamental problems so we could get to that point where we understand the physics enough to make progress in quantum. That's terrific thank you so much. Thank you.