← All episodes
Helmut Katzgraber

Episode · Feb 27, 2023

Helmut Katzgraber, Amazon

Why Quantum Projects Must Start With a Business Problem

Summary

Helmut Katzgraber, Global Practice Lead at the Amazon Quantum Solutions Lab, is interviewed by Yuval Boger. Helmut and Yuval discuss the process by which the Quantum Solution Lab helps customers, successful projects with BMW and Goldman Sachs, advice for customers thinking about such projects, and much more.

Key takeaways

  • The Amazon Quantum Solutions Lab starts every engagement by identifying a concrete business problem before discussing which quantum approach might apply, rather than starting from the technology.
  • Projects with BMW and Goldman Sachs succeeded because both companies had teams willing to reformulate their existing problems into a form that could be tested against quantum and quantum-inspired methods.
  • Helmut advises customers to treat current quantum work as building internal expertise and problem formulations now, so they are ready to act quickly once hardware crosses a useful threshold.
  • Realistic expectations matter more than hype: many early wins come from classical or hybrid techniques developed alongside the quantum exploration, not from a quantum computer outperforming classical methods today.

Transcript

Yuval Boger: Hello, Helmut, and thank you for joining me today.

Helmut Katzgraber: Hi, Yuval. Good to meet you. Nice to be here.

Yuval: Great to have you. So who are you and what do you do?

Helmut: So my name is Helmut Katzgraber. I'm global practice lead for Quantum. It's an organization that is part of Professional Services, and we are the team that basically engages directly with our customers at Amazon.

Yuval: How does that work from a customer perspective? How does a customer get to you, and what can they expect to receive working with you?

Helmut: So one of the things that is very dear to me is to be relatively realistic as to what quantum technologies can and cannot do. When we engage with customers, usually they either come because they want to learn what quantum computing can do for them in the future, or they come with a problem that they see as insurmountable, something that is very, very hard, and so there's always a desire: well, maybe a quantum computer can solve this.

In both cases, what we ideally want to do is, of course, prepare them for a quantum future, so work with them hand in hand and try to build a solution that will eventually run on a quantum machine at scale. But in an ideal world, we also try to find solutions that leverage today's technologies, be it operations research, high-performance computing, or machine learning, to actually solve the problem at scale.

So it's a little bit of a dual-track approach that really allows us to be forward-looking, but also to bring value to the customers today.

Yuval: Do you have a preference for problems that are quantum only, as opposed to problems that require other aspects of AWS, more on the classical side?

Helmut: That's a good question, because I think the fair thing to say is that I prefer challenging problems.

Where I'm going with this is that we're very much tool agnostic. We love working on quantum problems because this is the core expertise of the team, but we also like to dabble in other technologies. To give you a couple of examples, with Goldman Sachs, we did a theoretical resource estimation of a quantum algorithm in finance, which is purely paper-and-pencil quantum work, and it was recently submitted for publication.

On the flip side, with BMW, for example, we looked at robot motion planning in the production plants, showed them how you could potentially solve this on a quantum machine in the future, but then found alternative technologies that work today to solve it at scale, and found an improvement of, in this case, 10% over the time the robots used to take to, for example, apply PVC to a seam of a car.

Yuval: How long would a typical project take from start to finish?

And related to that, do you accept any customer to work with you, or do you focus on a particular size or particular type of customer?

Helmut: We have teams that focus on different types of customers. If we have a customer that needs, for example, hand-holding on how to use Braket, our fully managed service that has been in GA since 2019, then usually we pass them over to our solutions architects in the worldwide sales organization, because they are the ones that really work hands-on with the service.

We prefer customers that require really deep research engagements, where you have to dive down deep and really, you know, flex mental muscle to find a meaningful solution to a problem. You know, everybody's a customer at Amazon, and we welcome everybody to play.

We just try to focus on the problems that are really, really tough to solve.

Yuval: When customers think about needing help to solve such a problem, would they normally think of AWS, or would they normally think of a consulting organization, a BCG or a Deloitte or one of these that have a lot of subject matter experts?

How do you position it relative to the consulting organizations?

Helmut: I think, and, you know, I don't want to step on anybody's toes, we have people who are really deep subject matter experts, who have worked in the field for many, many years and have really achieved a lot scientifically, and that type of depth is often not found in consulting firms.

As I said, for us, at the end of an engagement it's important to publish a paper in a scientific journal, and that requires a certain degree of rigor that you can only achieve in an organization where you have the freedom to actually work on a project. Now, you asked me earlier how long these projects take.

What I'm going to tell you now will reflect exactly that. We can have projects that take three to four months. Those are rare. Typically they take, give or take, between six months and 18 months. So this is really deep research work.

Yuval: So for instance, the BMW project that you mentioned, could you tell me roughly how long it took and what the key stages were in such a project?

Helmut: Sure. The BMW project took about six months from beginning to end. As for the key stages, the way we usually engage is, you know, we have a first call with a potential customer and we hear what the problem is. Then, based on that, we schedule a deep dive of between two and four hours where we sit down with subject matter experts on their side, and that's because, for example, we're not experts in robotics or in finance or in clinical trials.

We really hash out the details of the problem. We then go back, think about potential solution strategies, and write a proposal. If the customer is interested in pursuing this, we scope out the project, and then we kick it off and roll up our sleeves. Now, while the project is running, there's always a research and discovery phase at the beginning where we really try to dive deep.

And then subject matter experts, in this case on BMW's side and on our side, work hand in hand to actually build a solution to the problem. In this particular example, we ran on two tracks. On one hand, we thought about how we could map this onto a quadratic unconstrained binary problem, which is a language that, for example, quantum annealers speak.

In parallel, we brought in expertise from our own middle-mile optimization team, Mauricio Resende, who is very well known for his work, and basically thought, "Well, how can we solve this problem at scale with an algorithm that can run on a standard computer today?" Eventually, these two tracks merged.

We wrote up a nice paper that you can now read. It was published last December, as far as I recall.

Yuval: When the project does involve quantum computers, do you find that you also engage with the vendors themselves, say Rigetti or IonQ or QuEra or others that are hosted on Braket, or is it typically just the internal team?

Helmut: It's a good question, right? Needless to say, what the end user often gets on the machine is a reduced set of features. You know, there are some additional knobs that can be pushed and turned here and there. In specific cases, and I can use the example of QuEra, which is one of our partners,

we also work with researchers at QuEra to tease out the most performance for a particular problem. Right now we have an ongoing project in finance; I can't disclose the customer yet. But there we're really working hand in hand with QuEra to do better device calibration, to understand the problems better, and to see how we can best embed a real-world problem onto QuEra's hardware.

And some of the things that we develop with our customers can potentially also flow back into the software or hardware roadmaps of our partners.

Yuval: I assume you've done a number of projects with different customers by now, and my guess, based on feedback that I've heard, is that most customers are very happy with the experience.

What do you think customers are most surprised by? Is it that quantum is better than classical? Is it that classical is better than quantum? You know, when you speak with customers, what makes them say, "Oh, I didn't know that was going to happen"?

Helmut: This is a good question because it's also a very pointed question, in that obviously customers are happy that they have a forward-looking quantum solution.

But I think the biggest surprise for our customers, which is two things kind of intertwined, is what happens when you look at a problem through a very different lens. Say you look at an optimization problem: traditionally an operations researcher would look at it, and we look at it from the angle of a physicist.

You take a very, very different perspective on the actual problem, and suddenly you unlock a solution that they had no idea existed before. This can be quantum or non-quantum, and I think this is usually where the big surprises happen: not only do we prepare them for the quantum future, but we also bring value to them today.

Yuval: And is the team primarily physicists?

I believe you have a PhD in physics, you have a lot of published academic articles, and you've been a visiting professor and so on. Is your background representative of the rest of the team?

Helmut: The background is very broad. We have physicists, and even among the physicists, we have a spectrum.

We have statistical physicists, mathematical physicists, theoretical physicists, people who focus, for example, on cold atomic gases. We also have operations researchers. We have machine learning experts. We have two HPC experts. One really important thing for me was to bring all these different disciplines together and hire a team of people who come with very different perspectives, and this is what has allowed us to innovate at the intersection of these different disciplines.

I can give you a simple example. There is an important problem known as the maximum independent set problem that shows up all over combinatorial optimization. One of my quantum physicists was working with a machine learning expert, and they realized that you can, in principle, represent this problem as a physics Hamiltonian, but then use graph neural networks that run on standard EC2 hardware to solve the problem at scale.

So this is the really interesting mix. These communities would never, ever have talked about doing combinatorial optimization with graph neural networks, and now it's a thing. This is very important to me, and this is why I believe strongly that interdisciplinary research is key to unlocking the next levels.

Yuval: When customers come to you, where are they in their quantum journey? What I mean by that is, among the successful projects that you've completed, do some of them move into production, or are they more about understanding the capabilities and extrapolating to say, "Well, in 18 months or 24 months we'll have a computer that's strong enough to run it"?

Helmut: All of the above. We have customers that are very early in their quantum journey. We have customers that have invested heavily in quantum computing and have in-house teams. And we have customers that are really happy when the things we built not only give them an indication of when a quantum machine will have impact for them at scale, but also provide a solution that is now running in production. We have a handful of those, and I think that is a big source of pride for the team.

Yuval: I assume there may have been a couple of projects that did not live up to expectations. What did you learn from those? Is it the customer's fault, with the customer not being ready or not defining the problem well enough? Is it just an unsolvable problem? What can you teach us from the failures?

Helmut: I kind of hate to say this, but we have yet to have a failure, you know? I know this sounds pretty incredible, but so far we have aced every single thing that was handed to us. So I don't know what it feels like, but from my experience, where slight hiccups can happen is when we don't have subject matter experts from the customer side on the project.

That's usually where it gets challenging. We have to guess about technologies that we're not experts in, and then sometimes there's a bit of a mismatch. But usually this can be rectified easily. They bring in the right person, and then it keeps going smoothly.

Yuval: Would that be the key advice to the customer for a successful project, make sure you have the SMEs on your side?

Absolutely. Are there other things you would need a customer to prepare before they approach you?

Helmut: No, it's absolutely important for customers, first of all, to have a clear vision of what they want to do, and that's not often the case. We can help in some cases, but I think the most important thing is for the customer to come with the right people in the mix.

They can be, you know, pulled straight out of production into these conversations, because then we know that we're solving a problem that is relevant, and not an academic problem that might have no real impact on the business.

Yuval: Is there a particular vertical or a particular type of problem that you see more often than others?

Helmut: Yes, absolutely. I think finance has given us a lot of signals. We have engaged with several companies in the financial industry. We worked in the past with Fidelity, for example. We worked with Goldman Sachs and a couple of others that I can't mention right now. We have also seen a lot of signals in healthcare and life sciences, where there's, of course, interest.

For example, we once did a rather unusual project with a pharma company on the optimization of clinical trials. It's not the standard quantum project you would expect, but the interesting bit is that clinical trials are actually the biggest cost of developing a drug. So for me, it's a mystery, you know: why are we talking about the chemistry if we can attack a problem that is much, much larger in scope?

So those two, I think, are the strongest signals. We're also seeing more and more manufacturing and scheduling problems, et cetera. And I'm lumping air travel and all these others a little bit into manufacturing and logistics. We have signals in that area too, but I think finance has really been the front-runner among the companies that are investing in quantum computing.

Yuval: And from your perspective, how far is the quantum industry from being truly useful, say, to the finance area? When do we move from exploration into widespread production?

Helmut: For the finance area, it's hard to predict. If I may pivot slightly, I think the first valuable applications of quantum computers that don't require thousands and thousands of error-corrected qubits will be in the area of chemistry and materials science.

For example, with a QuEra device, we can already do some materials simulation-type workloads. So I would say that in the next five years or so, give or take, and don't hold me responsible for this number, we're going to see the first meaningful results, probably in the chemistry area. For finance, hard to predict.

Yuval: Very hard to predict. If you were controlling the work plan of the industry, and I'm sure Amazon has a certain influence, but if you were in control, what would you have vendors focus on that they're not focusing on today?

Helmut: Focus on error correction to build high-quality, fault-tolerant qubits.

I think this is the one area where everybody has to focus, and this is also what we are focusing on at Amazon in our Center for Quantum Computing. We are not just building hardware with superconducting qubits. Why superconducting? Because, of course, they're well understood, and we know how the fabrication works.

But we're also focusing very strongly on error correction very early on, and these two efforts are going hand in hand so that we can build a fault-tolerant machine. I think it's really, really key that we do this.

Yuval: Last question on the technology. When you think about hybrid applications that use classical and quantum computers, do you run into latency issues on the round-trip time between quantum and classical and back?

Helmut: Well, as of today, there are no real full-fledged data centers that house both types of hardware, and this is going to remain a challenge for the foreseeable future, especially because we're in a very nascent stage of the technology.

We haven't even decided yet which qubits are going to be the ones that eventually scale to the sizes needed. It could be ion traps, it could be superconducting, who knows? Because of that, these machines are usually at a distance, housed in other facilities, and that can cause latency problems, which is absolutely normal.

But in the case of AWS, great efforts have gone into architecting our service to prevent these latency issues, especially now that we have launched so-called hybrid managed jobs, which really allow you to run, for example, variational workloads in a much more efficient way than was previously possible.

Yuval: And if I could ask you a hypothetical question: if you could have dinner with one of the quantum greats, dead or alive, who would that person be?

Helmut: Dead or alive? Huh, that's a tough one, and I'm going to give you a very odd answer. I would love to have dinner with Paul Dirac, the main reason being that I love his books. They're very concise, to the point, and there's a certain elegance in the math that he did. The other one would potentially be Heisenberg. And, you know, I have to say, I was an undergrad at ETH in Zurich, where Wolfgang Pauli was, and there's a fun story.

We moved to a different building, and I found an old wooden chair in the dumpster. I was like, "Oh, this is the perfect chair." So I pulled it out of the dumpster and put it in my office. Fast-forward two months, and suddenly an email goes out: "Has anybody seen Pauli's chair?" It turns out that I had, by accident, pulled it out of the dumpster and was sitting on his chair.

So this is as close as I've gotten to one of those big fellows.

Yuval: Thank you for the story. I did hear Dirac as an answer before from one of my previous guests, and the caveat was that this person said Dirac was not a great conversationalist. So it might be a difficult dinner, an awkward dinner, but maybe a very valuable one.

Helmut: So, in full disclosure, I love food. I love eating. I go to restaurants and write a lot of reviews online, and when I eat dinner, I like to focus on the food as well. So he would be a perfect partner if you want to enjoy the meal and maybe exchange a few words.

Yuval: That's perfect. Helmut, how can people learn about your work, maybe case studies or papers that you think could contribute to everyone's knowledge?

Helmut: You can reach out directly to me, or, on AWS's website, the Quantum Solutions Lab has its landing page. If you're interested in engaging on problems related to optimization or chemistry, and now the newest thing, we have a Center for Quantum Networking, so we're also engaging with customers on building quantum connections between devices,

just reach out. Go to the website, send us a message, and we'd be happy to work with you. Excellent. Well, thank you so much for joining me today. Thank you. It was a pleasure.