Toby Cubitt, the person with the best name in quantum, is interviewed by Yuval Boger. Toby is also co-founder, CTO, and Chief Science Officer of Phasecraft, a quantum algorithms company developing highly efficient algorithms tailored for near-term quantum hardware. We discuss Phasecraft’s approach to bridging the gap between quantum demonstrations and useful applications, the role of hybrid quantum-classical algorithms in accelerating materials science, the importance of optimizing for specific hardware platforms, and the critical need for higher gate fidelities in quantum computing. Toby also shares his thoughts on when quantum computing might deliver commercially useful results, his views on AI’s synergy with quantum, and much more.
Transcript
Yuval Boger: Hello, Toby, and thank you for joining me today.
Toby Cubitt: Hi, thanks a lot for having me on.
Yuval: So who are you and what do you do?
Toby: Well, I am Toby Cubitt. I really am Toby Cubitt and I really do work in quantum computation. I built my career off the name recognition, I guess. My surname is spelled differently. I’m co-founder and CTO and chief science officer of Phasecraft, which is a quantum algorithms company based in London, Bristol, and DC in the US. I also still have a professorship at UCL University College in London, where I’m a professor of quantum information. And I’ve been working in quantum information and computation on the theory side and algorithm side and computer science side and mathematical physics side for about more than I care to remember really, but over 20 years now.
Yuval: Toby Cubitt, definitely the best name in quantum.
Toby: I have my parents to thank.
Yuval: What does the company do?
Toby: So Phasecraft is the quantum algorithms company. We are focused on developing algorithms that will get useful applications onto near term quantum hardware as quickly as possible. And that means algorithms rather than software, we of course write a lot of software, but the key thing we do is to actually try and figure out how to develop better algorithms or more efficient ways of attacking problems that quantum computers are good at much more efficiently than the sort of textbook techniques or the standard techniques in the literature.
So it’s very R&D heavy because at the moment to date, no one has yet run a useful computation on a quantum computer.
People have run computations that are demonstrations of really impressive things, but we’ve still got to get useful applications onto quantum computers to start actually leveraging this incredible power of these incredibly small primitive devices to do something useful. And we believe, and Phasecraft’s mission is to make the algorithms significantly more efficient.
The other side of it is of course, things that QuEra and others are doing, which is to develop the hardware to run those algorithms on, but if we can bring down the cost of getting a useful application on a quantum computer by much more efficiently getting way more efficient circuits, we can, we need less hardware, fewer hardware resources to actually get to the point where the quantum computers are doing something useful and solving real world problems.
Yuval: How specific is it to each hardware? I mean, do you have a preferred hardware that you want to run on, or are you trying to create algorithms that are more generic in nature?
Toby: So the Germans have a great word for this, it’s “jein”, so yes and no at the same time. On one level, most of what we do, a lot of the underlying algorithmic improvements we make apply across hardware platforms.
So if we’re trying to take a transition metal oxide and figure out how to really efficiently simulate the charge drift coefficients of that material, a lot of the stuff we need to do is independent or is at least agnostic to some level to the hardware, but then of course, we want to squeeze the maximum possible juice out of this.
So at some level you get to the point where the same techniques will apply across at least matter-based qubits quantum hardware, but you have to apply it differently to an ion trap quantum computer and differently than you would to a Rydberg atom array computer.
From our perspective, we’re running also on both IBM and Google’s hardware. They’re both superconducting circuit hardware, but they’ve made slightly different choices about the way they implement the qubit. From our perspective, we do things a bit differently on those and optimize. So the techniques that we use are common, but the way we deploy them for each hardware is different.
Yuval: Do you really need quantum or would you just run it on a quantum simulator or just on a classical supercomputer?
Toby: Well, good. We do a lot of things where we’re running. We do run a lot of things on emulation because we can quickly test algorithmic ideas numerically first.
We also run on small scale quantum hardware, though, which is really important, even if that small scale hardware or large hardware and using a part of it, because, you know, I’m a theorist by background, I can have 10 good ideas before breakfast, but guaranteed by lunchtime, all 10 of those ideas will have not worked on real hardware because they’ll hit up against the real world as you guys, you know, in hardware world know very well and experimental physics, hardware development, it’s difficult and things do not work as you hope to without a lot of effort.
It’s really important for us to not just run on emulation, but to run on real hardware to check if the ideas stand up to the real world, not our model of it. And then we run on large hardware where we’re running things that are beyond what we can emulate. And at that point, to the best of our knowledge, there are things in quantum computing where there is very strong evidence that you cannot achieve this on classical computers. The question is only whether we’re beyond that boundary or just inside it still.
Yuval: Do you feel that you guys have achieved useful results today? And by useful, I mean, industry useful, not just theoretically or academically useful results.
Toby: Yeah, no, not yet. And they nor has anyone else in reality. So nothing yet has been. There is no computation that’s been performed on a real world problem yet on a quantum computer that goes beyond what can be done classically. However, that doesn’t mean I don’t think it can be achieved.
We’ve made a lot of progress towards that, but actual commercially relevant applications now that couldn’t be done on classical hardware, even as a quantum algorithms expert of 20 years standing, I don’t think that’s been achieved yet.
Yuval: And if not, how soon? When do you anticipate that there’ll be commercially-useful results?
Toby: It’s kind of, I was gonna say that was the million dollar question, but it’s probably the billion dollar question. The answer is of course going to be annoying. It depends. So partly this depends on the pace of hardware development.
For commercial problems we’ve focused on at Phasecraft, we’ve done the most theoretically rigorous analysis in simulation applications. And for those, we know that we would only need a quantum computer of a few hundred qubits and be able to run circuits of depth, let’s say conservatively like 10,000. Now that’s beyond what anyone can do to date, but it’s advancing.
So depending on the hardware timelines, if you believe the optimistic hardware timelines, it’s not that far off. It’s a matter of like two, three years before actual real, for example, materials modeling of transition metal oxide materials on quantum computers can be done. But of course, hardware development, it’s bumpy, right? There’s progress then it’s, it’s hard. There’s some new thorny engineering obstacle and then it, and all different hardware platforms are advanced at different rates. So there’s a big variance in that.
If we’re talking about scientific quantum advantage, and I think it’s important to, you know, there’s commercial kind of bunch where you’re really doing a task that is actually useful in industry, in say the chemicals industry that replaces the kind of computational chemistry they already do and does it better or does things that can’t be done at the moment and have to be done in the lab.
But there’s also, this is not something where there is a hard line between things that are useful scientifically and things that are useful commercially. It’s a transition and where exactly you put that boundary is not 100% pinned down. So in terms of scientific applications doing something that is of scientific relevance to research, where quantum computing is being used as a tool in condensed matter research or materials research or chemistry research that I think is much closer. Again, those timelines, I would say fairly confidently are within the next two to three years that they start to become useful as a research tool and that follows the history of classical computing.
You know, in the early days of classical computing, they were big, expensive. There were a few of them in big universities and national labs. There were just a tiny handful of them and they were mostly used for scientific computing and they were flaky and error prone and broke all the time. And yet people then rapidly, once they started to solve scientifically interesting problems, surprisingly rapidly found commercially relevant uses for them. But the real answer is we’ll see.
Yuval: A hundred qubits, 10,000 circuit depth, so 1 million Q-ops essentially.
Toby: Yeah. If you just multiply those two numbers, sure.
Yuval: And if you think about the specs for a quantum computer today, number of qubits, single qubit, two qubit gate fidelity, coherence time, connectivity, and so on. What’s the one thing that you’d like to see improved?
Toby: Gate fidelities. Gate fidelities and always gate fidelities. And I don’t care whether those are the physical gate fidelities or some level of fault tolerance or error correction that give me logical gate fidelity, the better. But the biggest obstacle is this and is poorly understood often in the media. People always quote how big your quantum computer is in terms of number of qubits. And I care about that, but I care much more about how good the gates are. And as I said, whether that’s logical gates or physical gates, from an algorithm perspective it doesn’t actually matter. It matters in terms of the over and then can you do it or not, but actually working qubits, gates, I care about the gate fidelity if I have to care about one thing.
Yuval: Certainly since you say a hundred qubits, I mean, there are computers today that have a hundred qubits already.
Toby: There are things, there were things we could run. So when I quoted those numbers, those are the ones I’m super confident about because I can mathematically prove that you can do it with that because you can do rigorous algorithm analysis.
There are of course, many other algorithms where they’re heuristic, can’t do rigorous theoretical analysis of the algorithmic resources, but are likely to require fewer resources than that. So that’s sort of a, that on that scale, that’s where one needs to get to, to be sure that we’re doing something, the one, you know, the world can do something at least scientifically relevant on a quantum computer. And yeah, there are, in terms of the number of qubits, we are just about at the point where there are enough to do some interesting things, but we don’t have the circuit depth yet.
Yuval: To what extent do you use hybrid algorithms or is it purely quantum?
Toby: Every classical algorithm from Shor onwards is a hybrid algorithm. So Shor’s algorithm has a whole load of classical algorithm around it. And it’s, but I know what you’re asking.
The algorithms that we believe will be, well, we know need fewer resources, they’re heuristic, however, are very much hybrid.
So for example, we have algorithms that leave classical density, functional theory computation, but substitute a piece of that with quantum computing, with quantum computer to be able to get, use the quantum computer to enhance, we call this quantum enhanced density functional theory to enhance the density functional theory computations, but by using quantum computer to compute the kind of hard kernel of that, that’s the exchange correlation contribution.
And that kind of thing is very much a, we’re putting together a small quantum subroutine on a computer that’s beyond classical, but still rather small and noisy to enhance a classical algorithm. So yeah, many of the things, the latest things that Phasecraft are doing in the last two, three years are very much these hybrid quantum classical algorithms, same story on optimization algorithms, which we will also work on.
And that’s where I think, unless that it depends if the quantum hardware develops a lot more suddenly ramps up rapidly and there’s a kind of big bump in the progress, which happens from time, you know, some things overcome, suddenly the hardware goes in a big jump in terms of ability, and then it kind of plateaus for a while at a higher level.
If the hardware does advance more rapidly, we could get to the point where we can actually use what you would call pure quantum algorithms, like Hamiltonian time dynamic simulation algorithms. But if I had to bet, I would bet that the things that will be feasible first are the algorithms that are hybrid quantum/classical algorithms.
Yuval: You mentioned oxides earlier. So let’s dive into that. Is there a particular class of chemicals that you think could be served first with quantum computing?
Toby: So I’m on the record in many talks as saying that material science is potentially closer, nearer term than molecular chemistry on quantum computers. Now I could be wrong on this. There’s essentially, basically a trade-off. For molecules you need fewer qubits, bigger circuits, and for materials models you need more qubits, but lower depth circuits.
The reason for that is on some level, very technical and mathematical, and you have to dive into the details, but on some level it’s very simple. There’s a lot more structure to a periodic crystalline material that you can exploit to reduce the circuit depth a lot and implement the Hamiltonian much more efficiently. For a molecule there’s less structure to exploit and you end up not being able to optimize quite as much.
And so we see this, that was my guess from when we founded Phasecraft. And in fact, that’s played out over five years. We’ve reduced the complexity of both of those applications, but we still have circuit depths that are lower for a purely crystalline material.
Transitional metal oxides are tough. Those are complicated materials that are a lot of interest for next generation battery technologies, also for cutting edge photovoltaic research. There are simpler materials than those, but those are ones that are commercially relevant to simulate.
Yuval: By the way, why the name Phasecraft? How did that come about?
Toby: Well, there was one rule that we all agreed on as founders, which is the company name had better not have a Q in it. That was one thing we were sure of. And the other thing is, quantum algorithms are all about making use of phases and crafting those into useful algorithms. So Phasecraft it is.
Yuval: How do you measure progress today? Is that just getting better and better correlation between experimental results and quantum simulations?
Toby: I think yes, that’s of course tough because it’s difficult. We’re not at the level yet of hardware where you can start actually running things that you can compare to empirical, say chemistry, material problems.
You can do that more on optimization and there it’s a better sort of, it’s more feasible to benchmark against what’s the best classical solution you can get versus what’s the best quantum solution. On the simulation side, we can’t yet benchmark computations that are beyond the ability of a classical computer, supercomputer against empirical lab data.
So what we can benchmark against is a, are we beyond what the classical algorithms can do because until you actually beyond that and how close are you to getting to something that cannot be done on a supercomputer. Supercomputers are incredibly powerful and classical and computational physics and computational chemistry. Very smart people have done this for at least half a century now. There are very good algorithms that work very well in many cases.
So the competition is really tough and we try to compare on the things that are not invented cases that no one’s ever looked at before, and then we make up some problem and try and benchmark on that because no one’s ever put much effort into. We like to benchmark on the problems that are known to be difficult classically for half a century, which is of course tougher, but the goal is to get to the point where you are really, fairly like-for-like, are doing better than classical algorithms.
And you know, we’re not, no one has got there yet, but it’s, you can see how much closer you’re getting.
Yuval: I saw a presentation from Microsoft, I think, a couple of weeks ago, and they spoke about AI for material discovery and said, well, you know, we can train AI on all the hundreds of years of experimental data, but now quantum computing can give us additional training data of things that haven’t been tried or exotic materials or so on. Do you see that happening?
Toby: Yes. So this is something we’re actively looking at within Phasecraft as well. And I think it’s interesting. I get asked a lot, isn’t AI machine learning now so good that it obsoletes quantum computing? And the answer, for me, is very clearly scientifically objectively no, because they don’t do the same thing. So, for simulating quantum mechanics and quantum many body physics or chemistry. For example, for dynamical properties, we know that that is hard on a classical computer, in general. We also know that for quantum computers, it can be done, you just need a big enough, good enough quantum computer, and you can do it. And then, you know, the answer that comes out is going to be correct, as long as your model is correct. Once you’ve got a computer of that scale, you don’t need a machine learning model. You don’t need to train something on that. You just compute the right answer.
It’s like if you’re building a bridge, you don’t even today throw lots of different bridge designs at a deep neural network and ask it to predict if they’re going to stand up, you just run a finite element analysis computation. And it tells you if the engineering is correct and will it stay up. Where machine learning is really powerful is in the problems that you can’t solve with the conventional algorithms. And the same thing with quantum algorithms, quantum algorithms are very good at some things and no use at all at other things.
And they’re very complementary in what you can do, what quantum computers are good at and what classical algorithms are good at and what machine learning-based algorithms are good at. So one day there will be certain problems where you won’t need the machine learning model, but there’ll be an intermediate era where it’s difficult to do large scale, accurate quantum computing. And we’re only just at the point where you’re doing something beyond classical. And it makes a lot of sense in that intermediate era, which could well last, you know, a decade, or more to leverage the quantum computing, which is going to be difficult to do on the model you really want, expensive and hard to achieve. Use that high quality data and then extrapolate that, which is what machine learning models are really good at. Deep learning models take that training data and just extrapolate that to other application domains. So that makes a lot of sense to me. And it’s something we’re also actively looking at.
Yuval: Sounds like you’ve been doing this for five years or so in Phasecraft and probably earlier or longer in academia. What have you learned in the last six to 12 months about quantum that you didn’t know before?
Toby: I think, I don’t know if it’s six to 12 months, but certainly one thing that is very different in Phasecraft compared to my previous life in academia is that Phasecraft is about one third quantum computing and algorithms experts, but it’s one third chemists and material scientists. And about one third people who are really good at coding this stuff up very effectively. And I’m a born again, chemist and material scientist. I did, I did one year of chemistry as an undergraduate and one year of material science and hated it. I now love it because what is kind of obvious, but it’s much harder to achieve in academia is chemists, material scientists have been wrestling with quantum mechanics from a different angle for a very long time.
There was in the last six to 12 months, I guess I’m not sure I’ve learned, but it’s come to fruition within Phasecraft is that that synergy between people, getting the people from those two different kinds of backgrounds to talk to each other, has led to a whole lot of new ideas that certainly wouldn’t come out of the pure quantum algorithms theory research and certainly wouldn’t come from the pure chemistry research.
There’s a lot you can learn from the things that have been done classically to solve these problems that quantum computers ought to be good at. You can leverage that and those techniques or those ideas in ways that are good for quantum computers and kind of gain the best of both in some ways. And so these kinds of hybrid algorithms for simulation come very much out of that discussion.
And it’s something I always believed ought to be probably possible, but it’s only in the last kind of six to 12 months that we started putting out papers where it’s like, okay, now this is now not just a hope, but here’s the fruits of that kind of thing.
Yuval: Commercially, is most of the work you do research grants from governments, or is there a substantial commercial engagement with say pharma or chemical companies?
Toby: The bulk of our Phasecraft funding income is venture capital funding. We take a large amount of, quite a large amount of funding in research grants as well. And we also have some government procurement contracts. We work closely with a number of big industrials. But we’ve always been very careful to steer well clear of the kind of commercial engagements seeking simply to address the question of when will quantum computing be useful to this kind of problem and write a report using the standard techniques. It’s kind of a zero sum game. It brings in a bunch of money to pay someone to spend a while writing a report and it doesn’t advance or change the dial on the technology. So that kind of consulting engagement, we don’t do. The kind of deeper R&D joint development partnerships, we have a number of.
Yuval: As we come close to the end of our conversation, I wanted to ask you a hypothetical. If you could have dinner with one of the quantum greats, who would that be? Dead or alive. Who would that be?
Toby: Dead or alive. I want to have dinner with all of them on lots of separate nights and have a lot of good dinners. But, I guess if I have to pick one, maybe John von Neumann, because, A, he is a very interesting character, and B, he was active across such a breadth of science from the foundations of quantum mechanics to functional analysis and operator algebras, through to the design of computers.
And I would be intrigued to hear his take on what science looks like now, and also he had the reputation of being able to party until the small hours and then deliver a perfect lecture at 8:30 in the morning. So he sounds like a good guy to have dinner with as well.
Yuval: Wonderful. Toby, thank you so much for spending time with me today.
Toby: Great. Thanks a lot for having me on. Really a pleasure.