Nikolaj Zinner, co-founder of Kvantify, is interviewed by Yuval Boger. Kvantify focuses on leveraging quantum computing for real-world applications, notably in optimization and computational drug discovery, through a unique algorithm, fast VQE, that efficiently calculates small molecules. Despite challenges in scaling and skepticism around variational algorithms, Nikolaj is optimistic about the field’s potential. He describes a pragmatic approach to technology selection, anticipates a gradual advancement in quantum computing, potentially marked by breakthroughs in encryption or efficient problem-solving algorithms, and much more.
Full Transcript
Yuval Boger: Hello, Nikolaj, and thank you for joining me today.
Nikolaj Zinner: Hello, Yuval, thank you for having me.
Yuval: So who are you and what do you do?
Nikolaj:Yeah, so I’m a physicist by training. My background is in nuclear physics, theoretical nuclear physics. I used to do calculations for how stellar processes develop in stars when I was a graduate student. But I slowly developed an affinity for other kinds of physics related to sort of fundamental questions about quantum systems, you know, what kind of atomic systems will bind, these kinds of things, basically solving equations. I have an affinity for mathematics as a background too, so I kind of transitioned into more fundamental quantum physics, which back then, which is now almost 20 years ago, right? It sounds horrible when you say it like that, but it is 20 years ago.
And I sort of started to think about what are some cool systems that you could develop. At that time, it was cold atomic gases that we were actually looking at. So when I was a postdoc, this was the focus area and it was just theoretical ideas and so on. But it became over the years possible to see a way through this into technology. And I think what’s amazing now, basically having worked on that many years ago, is to see how that’s coming together, everything is coming together and you can actually build different kinds of quantum technology.
I’m very interested in quantum computing, by the way. And you see the different platforms now. Of course, some of the stuff I worked on back then is now very applied in Rydberg atomic based quantum computing. So that’s super cool to see.
So that’s kind of my background. I’m a professor of physics at Aarhus University in Denmark. But I am also now part of a startup called Kvantify. We started about two years ago and have a focus on use cases for quantum computing. So basically figuring out how you can use those machines that people are now building and that are sort of becoming very much reality now. And how can you leverage those machines out in the real world.
And I don’t think that you can live with just sort of, you know, if you just say I have a textbook knowledge of that question, and I go and I look in the classic textbooks and I see that I can do all these kinds of fancy algorithms that are in the textbooks. That’s all good and well. And that’s enough of an argument to build the technology. I completely agree with that. But I think that we’re now at a stage where you need to get into industry, find the different sort of value chains out in industry and figure out how these algorithms will ,actually, these quantum algorithms will actually integrate into workflows, into real workflows. So that’s what we do in Kvantify.
Yuval: What is the unique angle for the company?
Nikolaj: I think we do similar stuff to other people. I’m going to come back to one thing that I want to mention where I think that we’re basically very, very unique in the market. But, you know, in relation to your question about combining classical and quantum in some new ways and so on? I think that we are. But I would also say that I think that we all are. And I think sometimes people forget the fact that a compute process, even in the future, is probably going to be like 90% classical stuff. And then you port 10% to a quantum computer where the quantum computer really matters. Right.
So our company, the way that we are based and the philosophy of that is that you also need to take that into account when you start building your team. So when we started building Kvantify and recruiting the first people and all this, we really said, look, we need competences really across a stack of different things, including classical algorithms, software developers, domain experts and so on. And so we have domain experts in various fields, but mainly in optimization and in things related to computational drug discovery.
So if you want to say, you know, what’s our play here? Those are probably the two things where I would say that there is the most promise and that’s where we sort of spend most of our time.
Now, coming back to your question about where canI come up with an example where I think that we are maybe mostly unique? It’s actually if you look at chemistry. And chemistry, I think many people believe that it’s a very good case for quantum computers. I believe that too. And I think what’s really interesting there is the question about what can you squeeze out of a near term quantum computer as compared to what can you get out of a long term fault tolerant one, you know, all these nice algorithms for all this promise that that is down the line. And I think we can discuss what we think is down the line, but let’s focus on the near term aspect.
I think that actually one of the big differences between what people have thought about some years ago and maybe it’s 10 years ago, maybe seven, eight years ago, something like that. When the first sort of hybrid algorithms, those algorithms that try to combine the best of both worlds, the best of what you can get in a noisy quantum processor and the best of what you can get from a classical computer, including things like the various quantum eigensolver algorithm class, which is now very explored. And I think that some of the problems that are very tricky there is actually that you somehow need to do a lot of measurements there in order to extract information all the time, feed that back into a classical computer, then update things, go back into the quantum computer. And that in that loop, you have to do a lot of operations. And that’s really costly, it’s very costly. At this point, quantum computers are not super fast. They have noise, you still need to do a lot of things and so on.
So reducing the so-called measurement overhead, I think it became known as the measurement problem. I think that’s a very precise statement. So a way to reduce that would be really nice. And I think one of the things that makes Kvantify unique is a new way to to do exactly that. So we have an algorithm that we call FAST-VQE, it’s not just a fancy name. It reduces the amount of measurements that you need to do to make a more precise calculation of let’s say a small molecule on a quantum computer.
So, I think from an algorithm standpoint, that’s one of the things we’re very proud of that that we’ve sort of figured out a way to get a quantum computer to use much, much less measurement overhead, and to get a quantum computer to give you, I mean a near term quantum computer, to give you fairly precise values for small molecules. There’s still a scaling problem. All these things were not solving everything all at once. That’s not what I’m saying. But we managed to figure out how to do with very few measurements. Actually, we were surprised about how efficient the system actually worked. And and we managed to get very, very good results on that.
Yuval: Given such algorithm and given your experience, you mentioned small molecules, obviously, pharma or other chemical clients want larger molecules. How soon do you think quantum computers will be really useful?
Nikolaj: Yeah, that’s a really good question. And this is the question that everyone asks, right? You know, I would say if I look at the roadmaps, right, and in terms of how we’re doing in terms of the roadmaps, you can see 50 to 100 qubits with fairly large gate depth in a couple of years, like 2026. Or you go to some other roadmaps, 2029 and so on. These kinds of things. I think that it depends a little bit.
So if you want to run something useful for chemistry in terms of our experience, then I would say you probably need just the algorithm that I was just mentioning. I still fear that, let’s say that you had 40 good qubits, and I’m going to come back to what I mean by good qubits, but let’s say you had 40 qubits so that I could sort of implement enough chemistry information worth of 40 qubits, something like that. I think I would still need on the order of say, probably between several tens of thousands to 100,000 expensive gates and by expensive gates, I mean two qubit, like a CNOT gate, those kinds of gates. So it’s still something like 10 to the several 10 to the four or 10 to the five kind of things, right? And so you’re sort of stretching it in that sense.
Whether that is possible, and now when I gave you that number, that’s a number where I’m not even sort of thinking about noise at this point, right? So there I would say, OK, that answer requires error correction. And so now it’s a different roadmap, right? So that’s the story.
What I cannot answer and where I think the most interesting stuff is going on right now is this. I’m not one of the people who is completely convinced that we should leave the hybrid and variational algorithms at this point. And this is a very big discussion in the field, right? Should you leave this because it’s never going to work, you know, move on and so on. And I think I’m not completely convinced of that. I understand both sides of that coin, I think. And I’m not completely convinced yet. I’m probably also inclined to say that it’s even if you sort of say it’s not until I have a fault tolerant computer that these algorithms give me something good, I would say that you need to still prep the quantum computer and building a good first guess for a wave function that could describe a bigger molecule.
Building that first guess, that kind of thing I think we can do in the near term era of noisy quantum computers. So I think that’s actually where the work is most interesting right now. I’m going to give you something where I’m going to go out on a limb and say that I’m also not convinced that we have seen the last bits of how you most efficiently put information into the quantum computer in the first place. And I think this is actually where there’s a loophole in a lot of arguments in the sense that if you were more efficient about using the qubits that you have, and maybe you can if you can live with half of the qubits or a fourth of the qubits or something like that, compress more information that usually costs you more gates. That’s a trade off. I think there’s still some loopholes over there that should be explored. And where you could maybe think about, can I build very efficient circuits to represent that wave function in a more compact way? And I think that’s where a lot of work is being done right now. Also because, when we talk about variational algorithms it’s also very connected to the whole way these variational algorithms have been used in sort of a quantum machine learning context by many people. This is a big big area right and it’s similar questions that come up there.
Yuval: Given that the algorithms are not good enough yet to deliver true value. What kind of customers are you working with today and what are they seeking to do?
Nikolaj: Yeah, this is really good. Some of the customers are very technology-interested companies. And I think that’s a fairly easy answer there because technology-interested companies are typically just interested in the technology right. So they’re not even sort of focused. So if you have a long-term vision for integrating technology, then some big companies will have an innovation division, and there will be people there who just want to know what’s going on right. Right now, many of them are probably overrun by AI and all this kind of stuff because that sort of came as a flood in the last year and a half or so, right.
But they will be interested in just knowing what is the prospect of doing this. So some of our engagements with customers are about that. So basically saying, look, where is this going, what would be there near term, and I would then give them basically the same kind of answers that we just discussed right. What are the near-term potentials and what are the long-term benefits and potentials, something like that.
But actually, it depends a little bit. I think that also some of the customers are very interested in actually building the knowledge and the tool chain that you would need to port a problem onto a quantum computer the minute that it’s strong enough to deliver. So that’s a different kind of story right because then you need to build a software pipeline that actually has the ability to take whatever the problem is, put it into the right context, and build the modules that make sure that it can interface with the quantum hardware right.
So those kinds of things are also going on there. We have different clients where some clients are very interested in running something on a quantum computer. Run something meaningful on a quantum computer, not asking for advantage because we don’t promise any advantage at this point, of course. But, you know, dipping their toes in that particular pond right so that’s one thing to do. Another thing is to sort of say okay, let’s try to not worry too much about how well let’s just simulate as fast as we can and see if an algorithm has promise, so like a scaling analysis or quality of solution. This could typically be something in an optimization case, for instance, so that that would be a particular point there.
Yuval: And I’m glad you brought up optimization because we’ve been speaking primarily about chemical simulation and fast VQE. Could you talk a little bit about special kind of problems or unique solutions that you might have in the optimization space.
Nikolaj: Yeah, so I think that it’s very difficult, and I’m an optimist here, so that’s what I’m going to say before I then say a bunch of negative stuff. No, that’s just, you know, it’s very—I’m unconvinced of the fact that, for instance, what is the benefit of particular types of hardware for one type of optimization, for instance in annealing or things like that. I haven’t—I would say it’s not clear to me where the benefit is. I’m not saying that the benefit is not there, I just—you know, people should be heavily convinced of that. I just haven’t seen where the benefit is yet.
I would go back again to something in the variational algorithm space where I know that people have tried different kinds of things, so it’s kind of similar to the small molecules actually, so that you can take a similar approach. Try variational algorithms for optimization. Now there’s a large class called QAOA algorithms, similar in spirit in some sense to the variational quantum eigensolver algorithms. But this QAOA part, I also think that people are sort of becoming more pessimistic. There was a level of optimism that grew a few years ago, and I think it’s sort of becoming more pessimistic —it’s combinatorial optimization problems typically that I’m aware of, and that this is what I’m talking about. So just want to scope that in, and I think there, people are sort of less enthusiastic. I’m more—I’m slightly more optimistic about that.
I think the quantum computer has an issue sometimes keeping the search space in the right place, and that sounds weird so let me try to expand a little bit on that. So, the fact is that in principle, right, everyone says, a quantum computer is something that can calculate all possibilities at once, and in principle, we can go along with that kind of thinking, but at the end of the day, you gotta measure something. So, and there is one answer right, so it’s not—it’s not a no free lunch here, right? And so, I would say, I think that the problem is sometimes even if you utilize the fact that it can sort of go through different paths, if you can encode a problem into Hilbert space, you can go through different paths, you need to try to figure that out more efficiently.
But Hilbert space is a large place, and sometimes your problem is that you get all these exponential possibilities, and then you try to measure one, and if your algorithm is not good enough at restricting you to a bunch of viable solutions, then you have a problem, and I think that’s where work is being done, and I see impressive work all the time there, trying to figure out how do you restrict yourself in that huge, exponentially big Hilbert space. In those cases where the viable solutions in an optimization problem could be a much smaller space, right? So how do you make sure that in every operation that I do to try to improve my solution, how do I make sure that I don’t just go off into this exponential space and just disappear and all my probabilistic is scattered out, and nothing will work, right? So I think that’s one of the key places where I see that there’s potential to figure out how to do that.
Naive analysis would say that it’s very difficult, but, you know, that’s what we strive to do, right? It is to break difficult problems.
Yuval: You worked with superconducting and with ions and briefly with cold atoms and you mentioned the kneelers so if you are if you are a betting man or an investor where would you put your money.
Nikolaj: On which technology would you put your money to deliver business value? Yeah, that’s the really good question, right. Here’s the sort of diplomatic answer, which is that it’s going to be a hybrid technology, which is kind of yeah, that’s kind of cheating. But I actually think that you know, in some sense, you just can’t answer that question right now because if I look at some of things people done, all this, this beautiful blog posts out there that are trying to sort of say pros and cons of these technologies, they all have these challenges that are hard engineering challenges in order to make the hardware work.
What I think is good about most of those platforms, and this is where I want to say something positive about it in terms of whether you could be pessimistic or not, I think I’m not convinced of the people who argue that there are like fundamental physics reasons why one platform would not work. You take a particular platform and then try to scope out whether, you’re like three Nobel Prizes away from this working. That would be a physicist’s way to say something is difficult. And I’m not convinced that in the technologies that you mention that we are three Nobel Prizes away from those.
But I do think that the engineering challenge is big in all of those platforms. Now just to say something about what I see, if I look at current roadmaps, right. I say that I’m probably slightly inclined to talk about ion traps and superconducting quantum computers then because those are just the two platforms that I work with most recently, so it’s not a selection as such. But in terms of that, I think that I would pick an ion trap for quality, so if I had a compact algorithm that needed a lot of gates, I would pick an ion trap. If I had an algorithm that required fewer gates but more qubits, I’d pick a superconducting circuit, right. So, that would be my approach—so actually, it’s a practical problem sometimes, right, which one, what is it that you’re trying to do, right. So I would say that there is not an answer there.
Yuval: Speaking of Nobel Prizes and as we get closer to the end of our conversation I’m curious if you can have dinner with one of the quantum greats dead or alive who would that be.
Nikolaj: That’s a really good question. Yeah, so you know, coming from Denmark, probably people pick Nils Bohr, right? That’s kind of the way you would do it. Actually, I would say that a lot of the stuff that we do is based on Feynman’s kind of ideas, so Feynman would be really an interesting character to meet for several reasons, I guess. In terms of how we scope things out, I think there’s a lot of mathematicians that are overlooked in this field actually, that sort of came with all the ideas about what we talked about.
We talked about, can you figure out what is a good way to encode information? All this kind of stuff is very mathematical, and nowadays, people are using all kinds of fancy group theory and all this kind of stuff, and I love mathematics. So, that kind of stuff, some of the greats in that, including Emmy Noether, von Neumann or Wigner, mathematicians and physicists. All these people going all the way back a century or more, introduced the very foundations of symmetry is what a lot of things in physics are built on.
So, I think that’s also very cool that this is now coming; things are sort of coming together in these frameworks, so that would also be very cool to do.
Yuval: And my last question. GPT was sort of a seminal moment in AI when you think about the quantum progression do you think it’s just going to be gradual or do you expect a quantum GPT moment for people suddenly wake up and say OMG this is real.
Nikolaj: I’m going to go with the technical answer first, and then how is it going to be perceived in the world. The technical answer first would be that I think it’s going to be gradual, and I think that about GPT, when I talk to, we have a bunch of computer scientists and ML specialists in Kvantify, right, and when I talk to them, they’re slightly more skeptical about the sort of the GPT moment, that it was like a cliff and then everything happened, because in fact, how that developed if you use the right metrics for those kinds of things, it’s actually a slow process but at some point, you know, it becomes a product and then all of a sudden, right, it gets released and all this kind of stuff.
Now in terms of quantum, I think the same thing is happening; it’s a slow process and we’re slowly just seeing the technology mature and people solve engineering challenges every day and then you push it forward and so on. I still think that the minute that you sort of say, OK, this algorithm, we could take a chemistry problem for instance, right here we have an algorithm on a quantum computer that runs fairly smoothly, probably it’s not going to run very fast but it’ll run for some time, right, and doing something that a classical computer has a very hard time doing, I think that will be the moment in terms of that use case, right.
But to be honest, I’m kind of inclined to say that as society is right now, maybe it’s the code breaking algorithm that would be that moment actually. I’m kind of inclined to say that that’s a pretty good benchmark for what could be there, right, would be that you say I have a quantum computer that’s strong enough to break RSA 2048 or something like that, right, and that would be some moment. I’m not a fan of the scare argument for quantum computing because I think that the factorization algorithm builds on other algorithms that will be useful for other things, and I’m also a slightly stronger believer that post-quantum crypto will catch up and that we might see that as, you know, that it doesn’t really matter too much whether you break it, except for old systems, right. But all of this is all guesswork, and I will be wrong about all of it probably.
Yuval: Excellent Nikolaj. Thank you so much for joining me today.
Nikolaj: Thank you very much.