Shengtao Wang, Quantum Algorithms and Applications manager at QuEra, and Jonathan Wurtz, a Senior Scientist in his group, are my guests on This week’s Superposition Guy’s Podcast. We delve into their work at QuEra, focusing on quantum algorithm development and the applications of neutral atom quantum computers. Shengtao shares insights on the evolving landscape of quantum computing, including recent breakthroughs in error correction and fault tolerance. We explore the potential of quantum computers to solve quantum problems, particularly in chemistry and materials science. Jonathan discusses the importance of hybrid quantum-classical systems and the need for creative algorithm development to leverage quantum devices. Together, we reflect on the future of quantum computing, the potential of neutral atoms, and what it will take for quantum technology to achieve practical utility. We also explore the skills needed for careers in quantum computing, the challenges of scaling quantum applications, and much more.
Full Transcript
Yuval: Hello Shengtao, hello Jonathan, thank you so much for joining me today.
Shengtao: Yeah, thank you for having us. I’m looking forward to this.
Yuval: So Shengtao, who are you and what do you do?
Shengtao: Yeah, I’m Shengtao, I’m the head of quantum algorithm application team at QuEra. So my background is more quantum computing, quantum simulation side. I joined QuEra around five years ago, that’s when QuEra started, one of the first employees at QuEra. So at that time I was doing a postdoc at Harvard and thinking about where I will go next. Basically, that’s a time where I think quantum computing is really a time where I would like to get quantum computing to be useful. And that’s when I joined QuEra and started the journey, fast forward, leading the team at QuEra, and in the last five years continue to build and to find applications both in the near term and also think about longer-term vision, what we can do with quantum computers and in particular what we can do with neutral atom quantum computers, the hardware we have at QuEra. So yeah, thank you for having me here.
Yuval: And Jonathan, who are you and what do you do?
Jonathan: Yes, my name is Jonathan. I’m also here at QuEra Computing. I’m a senior research scientist working on algorithms and applications. So just like it says, I do everything from low-level algorithms that try and work best on the machine all the way up to applications and trying to figure out what these computers will be used for. My background is also in condensed matter theory, things like Ising models and non-equilibrium dynamics. And that journey sort of eventually brought me to quantum computing and sort of where we are today, where now what I’m very interested in is trying to use these computers or use these devices to make a real impact in the world and actually try and do something with these devices that can’t otherwise be done with existing resources.
Yuval: How long have you been doing this?
Jonathan: Ah, yes. So I guess I graduated with my PhD in 2020. Then I did a postdoc in variational quantum algorithms and QAOA. And then I’ve been here at QuEra for almost three years at this point.
Yuval: So Shengtao, you’ve been here for five years, and Jonathan, you’ve been here for three. What did you learn in the last six months that you didn’t know before about quantum computing?
Shengtao: Interesting question. Last six months. Let’s see. So if that includes a little bit longer, the big result from logical demonstration is definitely one of the huge things that came out. I guess I was referring to the result from Harvard and together with QuEra as well, demonstrating an alpha early fault-tolerant algorithm on up to 48 logical qubits. That’s definitely something I wasn’t expecting before and I think the community wasn’t expecting before. So early fault tolerance is much closer than what we think. So in the last six months, I guess that’s something for me to think a little bit more about as well.
Yuval: And Jonathan?
Jonathan: So I guess I’ve had the time to think. I’d say in the last six months, I have learned so much more about error correction and fault tolerance than I ever had before. And I think, as Shengtao said, this is because of this recent breakthrough result from Harvard demonstrating 48 logical qubits on neutral atoms. But what’s interesting is because we are thinking very hard about hardware-level error correction, it’s probably given me a very backward view of how error correction works from a very functional compiler-level feed-forward, like actually nuts and bolts what needs to happen in order for these codes to work. So I’m actually very curious to continue this journey and hopefully one day actually have a fault-tolerant device.
Shengtao: If I may actually come back to this question, so after a little bit more thought on this. So regarding the last six months, new learnings on this, what we didn’t know, a little bit change of my thinking as well overall. So there was a lot of results where there were on the resource estimation side, there’s a quantum benchmarking program from DARPA and there’s other related, and there’s a lot of efforts in the community doing these fault-tolerant estimations. But the numbers are intimidating. They can range from 10^15 T gates to much larger even. And then what we learned also, this is also a NISQ algorithm. In the last five years we’ve been doing NISQ algorithms, but really finding something that’s really reaching a stage of quantum utility with NISQ algorithms is also very challenging. But how do we bridge the gap in between? Before we have this large fault-tolerant algorithm that we can do on the machine, that’s something we’re really converging to and still something we’re thinking hard today. Whether it’s heuristic algorithms which may not have this guarantees and resource estimate, whether it’s more hybrid that we want to use the majority of the work done on classical algorithm. So there’s a lot of open questions and that’s where I believe the community should focus on in the next few years as well.
Yuval: Jonathan, I think you work a lot with customers that come and want to do proof of concepts or otherwise co-develop algorithms with QuEra. How does that work? What does the process look like from a customer showing up in our door or in our inbox to something that actually works for them?
Jonathan: This is a great question. This ends up being kind of a long process. I guess the challenge of course is that it’s maybe not a small secret that today’s quantum computers don’t really work at a utility scale. So this is why we need to do these initial proof of concepts because what this lets us do is understand not necessarily getting your million bucks from optimizing your bus schedules better, but understanding the structure of how these computers work, how they integrate with your classical resources, how you’re able to understand how the gates work or say how your analog mode evolution works, and then how you’re able to leverage that to solve your application. So it’s not necessarily the result that you’re going for, but rather the journey to that result, which gives the customer insight into where we are, where we’re going, and how their challenges can fit into and be solved by quantum computing.
Yuval: Some customers, I believe, come wanting to do certain types of optimization. Maybe because they’ve been trained by another type of quantum modality, they come in and say, “Well, here’s my QUBO formulation. Could you do it better?” Is that the right question, or do you take them through a different process?
Jonathan: I think that this might not necessarily be the right mentality to have. As Shengtao was saying, and I am convinced as well, the future of quantum is very hybrid. It is a mix of both quantum resources as well as classical resources working in concert with each other. For example, you might not need to explicitly encode your entire problem directly into qubits. Instead, you might be able to use your quantum device to solve some subproblem or otherwise be a coprocessor in a larger classical workflow. I think using this mentality of using your quantum computer for what it’s good at and using classical computing for what it’s good at, is ultimately the way to deliver value in the world.
Yuval: Shengtao, I think that if you ask people generically what type of problems can quantum computers solve or will be able to solve, people mention simulation, optimization, and machine learning as sort of the three categories. As we sit here today, are you more excited about one of these categories than others?
Shengtao: Yeah, that’s a good question. That’s something we are thinking hard about in our team and across the whole community is thinking about that very hard as well. For us, one thing that is not a secret, we believe is quantum computing is good at solving quantum problems. What that exactly means is we need to find some problem that has certain quantum effects and that will have a better chance for quantum computing to actually have an advantage there. So that covers for chemistry simulations where molecular binding, where it’s actually not easy to find this. It doesn’t mean any molecular binding has quantum effect or quantum phenomenon. But certain molecular binding where you have metallic ions and so on, that could have quantum effects into it and that’s where quantum computer could be helpful for. And similarly for material science, for better designs and so on, we’re looking hard on whether there’s any evidence where there is actually interesting quantum effect into it and that’s where we believe quantum computer will be more useful for. And that said, we are not solely focusing on that. And there’s also things like machine learning which on the outside you may not realize quantum can help in terms of all data is classical and why do you need a quantum computer for it. But at the same time, for the case of machine learning, there are interesting cases where there are very complex correlations in the data where quantum computers can capture better than classical computer. And so that on the machine learning front, for example, we’re focusing on what kind of correlations the data will have that will have quantum computer that will be able to capture much better than classical. There are theoretical results on that, but it still has big gaps in terms of implementation, in terms of practical relevance and so on. But the big picture there will be quantum computer can be used to basically treat it as generating data, generating samples, can we make use of these complicated samples to do something useful. That works both for machine learning and for optimization as well.
Yuval: So you’re not expecting a quantum machine learning algorithm to do a billion parameter LLM anytime soon, but do you want to use it for something else?
Jonathan: I guess the billion parameter LLM, we have the luxury of gigabytes upon gigabytes of classical memory. So we are able to have these enormous models which have been co-developed with how we build our classical computers. So it may not be LLMs that are the quantum algorithm that we use because there you have a billion parameters, so of course you can just put a billion parameters into memory. But maybe there is similar architectures that we can use for quantum computers or hybrid computers, where maybe you do have that billion parameters in classical memory and then some other smaller set of parameters using your quantum device. And then those working together in concert might not necessarily look like your LLM, but look like something sort of different, I guess.
Yuval: When do you think quantum will be useful?
Jonathan: So this is a great question. So the challenge, of course, is that we have been working for 50 years on classical algorithms to solve our problems. And what this means is that we have 50 years of incredible progress in things like machine learning, in terms of things like quantum simulation, like DFT, and other quantum chemistry methods. In things like optimization, things like heuristic algorithms, branch and bound, all of these wonderful ideas that we’ve had. And somehow us as a field now have had the last 10 or so years to say, “Okay, now you have to compete with 50 years of innovation.” And it’s actually remarkable that even though we’ve only had 10 or so years, 10 or 20 years, we are starting to get up to that front. It’s not necessarily a laughable thing for us to say, “Oh, maybe in a few years, we might actually be able to get some use out of these things.” I think an exact time scale depends on your optimism. Maybe it’s yesterday, maybe it’s tomorrow, maybe it’s a year from now. But eventually as we build up to our larger devices, as we improve our coherence, and especially as we get to fault tolerance and error correction, that’s where you start to really get exciting algorithms. Like 1,000 logical qubits, there’s a lot you can do with that.
Yuval: So I want to just dive into that a little bit deeper before I hear Shengtao’s answer. 50 years of machine learning, some of it was algorithm development, and obviously in the last couple of years, attention is all you need, some breakthrough in the way people are thinking about it. But a lot of the usability of machine learning became just because there’s much more raw computational power than there was 10 years ago. So do you think that the barrier for quantum, is it primarily the algorithms, or does it just have more qubits and better qubits?
Jonathan: I think it’s a bit of both, actually. There is definitely a challenge, and I see this of it’s very easy to make your own algorithm and think about, “Oh, I need 10^12 T gates, and all I need to do is wait around until QuEra Computing has their 10^12 T gate computer.” But then your hardware makes you realistic about that. Then you say, “Okay, hang on, maybe not 10^12 T gates. Now I need to narrow it down. I need to refine my algorithms.” And to that regard, then, it’s not necessarily just waiting around for better hardware. It’s also co-developing and co-evolving your algorithms to meet that hardware. So to this regard, it’s being creative and not necessarily saying, “Oh, I have some adiabatic algorithm, and this will solve everything.” Like, no, that’s not going to solve everything. You need to have everything else around it. Like, okay, what if we have hybrid analog digital modes, for example? Can we start playing with things like that? What if we have different classical post-processing methods for machine learning? All of these extra tools, which are built in the context of the hardware that you know today and you understand for tomorrow, that’s how we can catch up.
Yuval: Shengtao, when do you think quantum will be useful?
Shengtao: Yeah, that’s the most popular question I guess I got asked, but that’s always a very tricky one. So I know some people in the community and some people outside the community are always saying quantum computing is always five years away no matter where you start, and that has been the case maybe in the past decade or so. My hope is really to see quantum computing to be useful in five years. That’s more of a hope than a prediction. There’s a lot of unknown to be there, to be discovered. Actually, for your question of hardware versus algorithm, where’s more risk there, to quote a professor from MIT, actually, he’s an experimentalist, he believes now the higher risk is actually on algorithm application in the next few years. We know where the hardware is going to be at in a few years, and by now we are confident in the next few years we can reach 100 logical qubits and beyond. But what can we do with them? What applications can we really have usefulness to be done on this intermediate scale? That’s a big open question, and that’s incumbent on us to continue to work on that. Both within our team within QuEra, of course, but working with customers and also the entire community, we need to focus on that to develop algorithms and applications that can be done on these intermediate scale quantum computers.
Jonathan: This is actually one of my favorite interview questions that I ask for job candidates. What would you do with 100 qubits?
That’s above the scale that you can simulate, but below the scale that you might be able to do some large factoring algorithm or other very powerful thing. Nonetheless, this is exactly what we need to be thinking. Very, very limited resources, but still beyond the reach of existing classical methods.
Yuval: I don’t want to put you in an uncomfortable situation within the company, but do you still believe that neutral atoms is the way to go?
Shengtao: Yeah, absolutely. That’s even more so in the last year or so. I came from a background, my PhD, actually I work more closely with trapped-ion systems during my PhD, and also cold atom systems, but that’s a little bit different from quantum computers. I work on optical lattices and quantum simulation with both cold atom systems and trapped-ion systems. During my post-doc, I switched a little bit more toward neutral atom systems. The progress on neutral atom hardware in the last five years or so is enormous compared with any other platform. That’s just like, it’s a big surprise to the community, to a lot of people. Being in this position and deeply involved into that, I still see a lot of very much untapped potential for the platform. To answer your question, simply yes, absolutely, neutral atom will be one of the leading transformers in the foreseeable future, I can see.
Yuval: If you couldn’t work on neutral atoms or there was some new discovery that made it obvious that neutral atoms are not the solution, which is your next favorite technology?
Shengtao: That’s another tricky question. There’s a lot of unknowns still for each platform going forward. There could be, as you said, breakthroughs in other platforms as well in the coming years. For the moment, given the information we have right now, the next one will be trapped-ion systems. Not only because I worked on trapped-ion systems before, but in terms of closeness to some of the uniqueness of neutral atom system, there’s also a parallel there. Of course, there’s a huge difference as well, but that would be my bet.
Jonathan: Maybe to answer this question of if neutral atoms are the leading platform, I think to put on my very QuEra hat, yes, absolutely. I guess from the day-to-day perspective, when I’m thinking about using these devices, I usually think about their limitations. For example, neutral atoms typically have a lower shot rate than, say, superconducting devices. Now you’d be thinking, “Oh, now you only have a 10-hertz shot rate, so what do we do with that?” But then this forgets about the advantages. For example, we’ve broken this barrier of high fidelity. We are able to scale to very large systems. We’re working on 200-plus qubits. We can have that with an all-to-all connectivity, which makes it much easier for things like compilers. All of these extra things compound into this theorist’s dream of what a quantum computer should be. You don’t need to think about connectivity. My goodness, I’ve been thinking about superconducting connectivity for years. I don’t need to think about that anymore. To this regard, we have the ability to scale these devices. We have the ability to improve the fidelity of these devices. For that, I think it does end up making them a winner.
Jonathan: Maybe to answer your second question, though, I am pretty impressed with the fidelities of trapped ions. I think if there was any second way forwards, it would probably be with trapped ions. There’s always that whisper in the darkness of 100,000 coherent single photons coming out and solving all of our problems. Trapped ions and photons would be my two other modalities. If suddenly we discovered that neutral atoms didn’t work, that’s probably what I would go on and look at.
Yuval: You mentioned an interview question. One of my favorite interview questions I was asked that many years ago is to explain what a spiral staircase is without using my hands. What kind of people are you looking for? What are the skills of people that want to get into quantum application development and quantum algorithm development that would be useful for a company like QuEra?
Shengtao: There’s a wide range of possibilities and set of skills that we’re looking for. It doesn’t fit into one bucket. On the algorithm application side for QuEra, we have people who are QEC experts. They have been developing QEC algorithm architecture for their PhD and so on. That’s a very valuable skill. We are moving towards, as our hardware is moving towards those already fault-tolerant algorithms. Having this QEC experience together with how to map it to hardware, that’s a very valuable skill. On the other hand, we have people in the team which are more software-oriented. As you’ve heard about this neutral atom, we can move qubits around and have this non-trivial connectivity. That opens up a lot of opportunities and also a lot of challenges. How do we build efficient compilers on top of it? That’s a hard, classical problem. We need people who have strong software experience and also compiler experience. How do we build this future generation of architecture of compilers? On the actual application development side, people coming from each of these verticals have experience working on quantum simulation, quantum machine learning, quantum optimization. Those are all very valuable skills for us. Taking a step further is how do we close the gap between what we have and what we know within the company and working with customers to really make this useful for some use cases and some of their use cases in the near term. Having that interest and having that passion to really bridge the gap, that’s really something we’re looking for in the company as well.
Yuval: Jonathan, do you have anything to add?
Jonathan: Maybe to add to this, quantum computing is an interesting place in that it intersects many different fields. You need to understand a lot of math that goes into it, things like linear algebra, of course, but extending beyond that to things like combinatorics works. You also need to understand a lot of computer science, so things like compilers, code development, working together on code. Then also you need to understand the physics of what’s going on, actually if you have errors with your computer, actually understanding how those errors impact your algorithm. All three of these things together make you better understand and be able to use quantum computing. But then on top of that, you also need to not just understand the physics, the math, and the computer science. You need to figure out what the heck to do with these things. To do that, now you need to be able to talk about complex problems, be able to understand and contextualize some customer’s need. You need to be able to think creatively about what those solutions may be, put those in the context of a current precedent, and then take a step beyond to try and not just solve the customer’s problem, but also further the field in terms of coming up with creative new algorithms that extend and hopefully make us be able to use these devices.
Yuval: Any particular example you can give us, even without mentioning the customer name if you don’t want to?
Jonathan: Maybe I’ll use an example. I recently published a paper titled “Non-Native Hybrid Optimization Algorithms.” The idea here is you basically use your quantum computer as a coprocessor for some larger combinatorial optimization task. And actually how I came up with some of the ideas in the paper was actually I was reading up on classical clustering algorithms for graph coloring. And then I thought, hang on a second. Here there’s this particular subroutine, right? You compute the eigenvalues of the lossy system. And there’s this particular subroutine which you basically need in order to do your cluster. Everything’s purely classical. And then now you can think about this purely classical thing, right? You can say, hey, what if we replace this one piece, this eigen—this Jacobian thing, what if we replace that with quantum? So then this ended up generating some hybrid algorithm, which was then understanding your classical algorithm, understanding the structure of your quantum dynamics, and then merging them together into something new. So then it’s not necessarily just taking some existing idea and running with it and saying, okay, let’s apply QAOA to bus routing. It’s now how do we merge these things in new and creative ways?
Yuval: So as we get to the end of our conversation today, I want to ask you a hypothetical. So if you could have dinner with one of the quantum greats dead or alive, who would that person be? Shengtao, how about you start?
Shengtao: Yeah, of course. I know this question is coming. So I don’t have an innovative answer for that, but I really want to ask them. So you’ll get either Richard Feynman will be one of the person I really want to meet. I have dinner with. He’s the father of quantum simulation, basically. He was quantum computer due to quantum simulation. And it could be Peter Shor as well. Peter Shor is the father of quantum computing, really. I have this practical, something that will really show quantum can do better. But the questions I want to ask them, I have two questions I want to ask them. The first question is, do you think the progress in the last 30 years in quantum computing is something you were expecting? Do you feel that the fast progress, slow progress, or just about what you were expecting in the last 30 years of quantum computing development? And then what do you expect in the next 10 years? The second question I want to ask them is, coming back to applications, if you were to bet, what would be the bet on the applications area of first practical quantum advantage? Would it be simulation? Would it be machine learning? Would it be optimization? Would it be materials? Would it be drug discovery? What would be the bet for you given all the status of the field? I would be really interested in hearing about your opinions on that.
Yuval: And Jonathan, your dinner guest?
Jonathan: I’d probably also pick Feynman. He seems like he’d be a very interesting person to talk to. It would also be interesting to talk with the early developers of quantum, with the Planck’s, the Heisenberg’s, so on and so forth, just so you can be like, “Hey, 100 years from now, these crazy far-out things that you’re thinking about like electronic structure of hydrogen and stuff like that, we’re actually using that as computers.” And just being able to share all of the advancements we’ve had in the last 100 years. I don’t really have a better answer than that.
Yuval: Jonathan and Shengtao, thank you so much for joining me today.
Jonathan and Shengtao: Thank you. Thank you for having us.
Jonathan: This is awesome, thank you.