Transcript | Quantum-Inspired Applications with Icosa Computing

We’re getting close to practical advantage in quantum optimisation. Companies interested in being leaders in their industry should already be considering how to harness this promised power with use cases that solve their business problems, but some companies want a better solution that’s advantageous today.

Learn more about quantum-inspired approaches already having a real-world impact in this episode of The Post-Quantum World. I’m your host, Konstantinos Karagiannis. I lead Quantum Computing services at Protiviti, where we’re helping companies prepare for the benefits and threats of this exploding field. I hope you’ll join each episode as we explore the technology and business impacts of this post-quantum era. Our guest today is the founder at Icosa Computing, Mert Esencan. Welcome to the show.

Hello, Konstantinos. Happy to be here.

It’s great to have you. I found Icosa the same way I find a lot of companies that have any connection to the Chicago Quantum Exchange. It’s pretty wild, all the developments happening there. It’s always exciting to see new things coming. In your case, you’re a part of the Duality programme. Can you talk about what that is, for people who might not have heard?

We’re deeply connected to the Chicago ecosystem. Duality has been the catalyst for that. They’re amazing — I just want to put that out there. We met with Greg from Protiviti, and he reached out to us again, and that’s why we’re here now. I’m also familiar with your work, Konstantinos, in the financial domain. I read the paper you collaborated with Multiverse on, “Financial Index Tracking via Quantum Computing With Cardinality Constraints.”

Yes, that’s right. I paid Mert to say this, so that’s cool. Seriously, thank you.

We are a part of Duality. It’s an incubator we’re part of. They are hosting five startups this year. They did the same last year — it’s their second year doing this. They are the only incubator, I believe, that is just doing quantum startups. They’re trying to push the Chicago ecosystem.

We’re also part of another incubator, the Creative Destruction Lab. The main hub for them is in Toronto. Toronto is also trying to push the quantum ecosystem, and the Creative Destruction Lab is older, but they have different streams. They do have a quantum stream, which has been going around for five, six years, I believe. Xanadu came out of them. In fact, Xanadu was there before they had a quantum stream. They showed that quantum is a valuable field to invest in. After that, they started the quantum stream.

Yes — lots of good quantum stuff came out of Canada, and, of course, the University of Waterloo. It’s been quite a brain trust of quantum folks up there.

Let’s talk about your background. Tell us how you got to quantum, and then I also want to spend a minute talking about the position you held right before you started Icosa.

I’m originally from Istanbul, Turkey. I lived there for the first 18 years of my life — a great country, great city. Then I moved to California to go to Stanford. I did my undergrad and master’s there. My undergrad was in physics and symbolic systems, which is mainly computer science and AI, but we look at cognitive science — any information processing system, basically.

Then, I took a gap quarter — took a few months off before my master’s — and traveled around a bit, mostly Europe. Around that time, I started reaching out to quantum computing companies, and realised that this is a very interesting field from the industry side as well. I was interested intellectually beforehand. Did my master’s in management science and engineering, which is, again, a Stanford-specific major, but mostly took financial engineering classes, like algorithmic trading, investment science, optimisation.

During my master’s, I started working for QC Ware as an intern, and after that, I moved on to Fidelity Investments. At the time, they did have a quantum team, but they hadn’t hired anyone from outside. I was their first external hire, but there was a very good team already in place, focusing on quantum computing and frontier technologies at Fidelity. This was FCAT, the Fidelity Center for Applied Technologies.

It was great to be on the other side as well. At QC Ware, we were the ones going to institutions and trying to convince them of our solutions, whereas this time, cool quantum companies came to us, so we played around with a lot of toys. It was great. We have a paper — hopefully, it’s going to be published soon. We looked at IonQ and tried to do some machine learning on their system. We did solve a toy problem using a few bits. Obviously, quantum is still early, but we did have some interesting results. That paper, it can be found by googling me.

I’ll be sure to put a link in the show notes. Like you said, you did your more traditional physics degree, and then, after that, you decided to start working with a very practical approach at QC Ware, and then eventually as a developer. You took that more pure-science approach into the quantum development field. It’s always interesting to see how people end up there, because some folks come in more from the computer science side, and they have to catch up on the physics after.

That was a very short period of time — about a year of your life back-to-back there, six months and six months, and then you did both sides of it. What was it like, jumping in like that? How much did you learn at QC Ware? I’m assuming that’s where you got your first real hands-on use of these machines.

Obviously, I learned a ton. The team I worked with was amazing. I was directly working with Fabio Sanchez and Peter McMahon. Obviously, they’re experts in their own domains. They know a lot about quantum computing. Peter is fully academic. I’m guessing he’s at Cornell or Carnegie Mellon, teaching or researching full-time right now. Fabio is still a close friend.

I did learn a lot from them. We looked at different technologies, different tools. I did some internal research for them, which was great. I was very independent. It was basically academic research — reading papers, trying to figure out something. We did have access to different quantum computers, so I got to use some interesting things.

What’s interesting about them is that they do the research and develop approaches, and then they get some customer, and then, after that, they turn it into a module that anyone could buy in the future — repurposing and building a library of reusable approaches. It’s like academic meets commercial. You were able to do these projects and experiments, and then, when you went to Fidelity, you were, all of a sudden, thrust into “Here we are — full financial applications and use cases.” What was that like? What did you get to work on that you’re allowed to share in that period of time?

I can give you a bit of a background before that even. I was already working on a startup idea similar to the one now. I talked to a few people — not directly to recruiters, but directly to my MDs beforehand — and that’s what made my case very interesting: We’re into it for one or two days, and immediately, they were, like, “Join our team — we’d like to work with you.” That was the fastest process I’ve ever been in, in terms of applications. And I was already working on some optimisation problems, doing trading on my own.

At the same time, I was also applying for a Ph.D. programme. I was looking to Oxford quite a bit. My current supervisor, he was also interested in what I was doing — taking these pure research approaches and trading on my own, putting my money where I’m actively taking risk built on these research ideas. It all came together in some sense. With Fidelity, I had to tell them upfront that I was working on this idea and I was going to go do the Ph.D. At that time, I didn’t have the Ph.D. offer, but I was, like, “Hey, I might go in a few months to do my Ph.D.,” and they were very accommodating in that. I was very happy to have this comfort. I’m glad that it was a great engagement for me.

What I did at Fidelity, it’s all confidential, but I was looking at finance problems and doing a lot of research, building internal tools, building IPs, trying out different hardware providers, because we have the access to them, and doing some internal benchmarking — testing these claims that most hardware providers make. There are so many metrics around right now. There was quantum volume, which has been used a lot, then it was criticised. IBM came up with a new metric few months ago, maybe a year ago.

And then, the competitor was IonQ’s algorithmic qubits — that other idea too?

Yes, someone had to filter those and basically represent to people who don’t have the time or resources to go through all of this and build some internal presentations to show what works and what doesn’t.

When you went there, you had to give them some disclosure before you started, and I’m assuming there were all these agreements to what IP is going to be yours because you have other things you’re working on personally, and then what IP will be theirs, and then you had to detangle that.

That was a process. It took some time to get the agreements in place, but it turned out that we didn’t pursue that idea anyway. That company didn’t go through, and we started from scratch.

For Icosa?

Yes. It wasn’t necessary to do those in place.

What you went through a year and a half ago, that was what a lot of people are going to be going through soon, because Fidelity was a little ahead of the curve. In 2021, you said, you were their first hire that was outside, that that was common, and that’s starting to happen now in some smaller companies. A lot of people are going to be the first quantum hire at a company, so they’re going to go through that same experience of whether they’re working on their own, whether they’re working on-site. That was a lot of on-the-ground training for solving financial problems, and now, with Icosa, which is what we’ll talk about next, you’re going to be developing these for customers, and helping financial institutions with financial problems. Tell us about how you started that, and the kinds of projects you’re working on initially there.

We’re not developing. — we’re in production now. We are working in production level, taking trade risk for clients. It’s hard. Development is fun. We can bounce ideas around, and then, in production, things always break. That was a transition we had to learn the hard way. Obviously, our team’s experience helped a lot because we did go through various projects from development to production. But as a startup, we had everything at stake, so it was a fun experience.

I’d been playing around with finance ideas, using quantum computing, reading lots of papers, and then I started trading on my own. I realised that this is an interesting opportunity. Then, my adviser at that time — now he’s a business partner, leading all technical efforts at Icosa — my adviser and personal mentor for 10 years at that time, David, told me that it’s better to scale up and, “Yes, you’re trading on your own, but this little capital means nothing to financial institutions, so you’ve got to scale up and convince others and yourself that this is a valid method.”

We did that. We did two POCs. We traded $1 million for a bit, and then $200,00 for a longer period with two different customers — one mutual fund, one hedge fund. Then, after that, we were convinced that this is a valid strategy and we can build a company around this. But this was just a POC, so we took a step back. We realised that we don’t want to be a hedge fund — we just want to build optimisation products. We don’t want to create strategies for traders — we want to solve the biggest optimisation problems in the world. That was our defining motto, and that’s how we created Icosa, and now we are exactly doing that: trying to apply very large problem solving to real markets.

Let’s look over the approach. What kind of machines are you using right now, actively, to do this?

There are a few things. First, it’s better to do everything in-house because we can control it, it’s not a black box and it’s cheaper — maybe free — because we’re getting a lot of credits from cloud providers. They’re very friendly to startups, so they’re giving hundreds of thousands of dollars’ worth of credits in their service.

To run shops or whatever.

We’re not exactly using Braket right now. When you said, “shops,” I immediately thought about that. Braket hosts a wide range of different providers, but they’re nearly all gate based, and they have D-Wave, but D-Wave only offers quantum processing, the QPU, and that is not scalable for our needs.

Again, our motto is to solve the largest problems. We believe that quantum is going to be there at some point. Gate based is going to be the mainstream — that’s where the main research is going on, and it’s more universal. Again, there are theories that, like, the annealing axis is proven. Theoretically, annealing is also universal quantum computing, but it’s not universal, in the sense that it’s not practically universal, whereas gate based, you can just programme it in any way you want. And we do believe, like, maybe QAOA can be upgraded, or new algorithms can come up — like, gate-based quantum computing will be used for large-scale optimisation problems. And we’re betting on that. We’re bullish on quantum.

But right now, we’re using what we can, and there are many ways to do large optimisation problems. One is, again, using a distributed system that you host your own algorithm: You write something up and use one of the mainstream algorithms. There are so many: There’s parallel tampering, annealing, simulated annealing, simulated quantum annealing. Many software ones. Azure hosts a bunch of these, and Monte Carlo as well.

The paper you referenced before, that we did, we were using D-Wave’s hybrid software, which, as you pointed out, we couldn’t get from the cloud. We had to go straight to D-Wave for it, because that’s the only way you can access that environment.

Exactly. We’ve been having a conversation with D-Wave. I used their system many times. They were nice enough to give me access to their system for free, and we get trial periods. At Icosa, we’re also working with them, and we might start working with them at a more robust level depending on our needs. Your question is, what do we use? That’s what defines us: We use what we need, and we use what’s best for that day or that problem.

I say “day” because we’re actively trading daily right now, so we do daily optimisations, daily trades. Depending in market conditions, you might need something that is very large. You could solve an easy, small problem with maybe just your laptop, but can’t solve a very hard small problem. You might want to use a large optimiser. Then, we choose to use some other provider. What we’re also excited about is the specific hardware that’s coming out that’s not quantum. People use many different terms for this. I call them physics enhanced, but I don’t think it’s caught on. People use quantum inspired.

I was going to say quantum inspired.

As a theoretical physicist, I can’t say quantum inspired, because it’s not quantum mechanical. There’s nothing quantum mechanical going on — or there’s no inspiration from quantum mechanics that defines these devices.

It feels like there’s an inspiration from the software that’s used in the quantum device or something. It’s separated somewhat.

Well, simulated annealing came in the ’80s or something, and quantum annealing came in early 2000s or late 1900s. Actually, simulated annealing came before quantum annealing, like the Metropolis-Hastings algorithm.

It’s legitimate to say quantum inspired, because we see this trend in industry where people came up with quantum computing tools, and the promise is that they efficiently solve large, discrete optimisation problems. Then, this inspired many other providers to build specific chips to solve these problems. Quantum inspired, in that sense, makes sense.

Quantum industry inspired.

Exactly. We’re excited to use those chips. We do have some access to those as well, mostly come from Japanese providers. I can’t name which one we’re using, but some names to throw around are Hitachi, Fujitsu, NEC, Toshiba.

Toshiba — the digital annealer. They’ve made pretty decent news with that for a while now — their approach.

Exactly. They claim to be solving 100,000 linear variables, and what we’ve seen in research mostly doesn’t go up to 2,000, 3,000 — maybe 5,000. They claim they can do up 20,000 devices, around that ballpark. I’m throwing numbers around, but just to clarify, these are all-to-all connected graphs, so when I say, like, 20,000, there are 20,000 nodes in these graphs. There are many variables that we’re putting into these systems. It’s not just 20,000 variables — there’s also the edges.

It’s a bigger-sounding number.

Yes. And 100,000 is very large, and then, obviously, we have 2100,000 choices. That’s a very large number — unimaginably large. Obviously, you don’t explore the whole solution space. You can’t. There’s not enough time in the universe, but you do some smart thing — you look at the energy landscape, you do some parallel tampering or quantum annealing, and then you try to identify the optimal solution. The problem with these methods is that they’re heuristic methods, they’re not deterministic, and because we’re looking at very large problems, there’s no way to know if you reached a good solution. I’m not even saying “global solution.” People are looking at these things as, “Have you reached a global solution, or are we trapped at local minima?” This is the big thing in the industry right now.

It comes up every time.

For finance purposes, a good local solution is fine — you don’t need to do the optimal global solution — but if you don’t know the answer to the problem, there’s no way to check if your solution is good enough. I talked to a lot of researchers about this in Toronto three weeks ago at CDL Session. There are a few metrics to look at this — looking at the distribution of your solutions. But again, if you’re solving a very large problem, you don’t even know if you reached the good solution.

Why are we in finance? In finance, it’s very easy to quantify success. We’re solving very large optimisation problems. A lot of people come to us and ask, “Why are you not doing logistics?” or “Why are you not doing drug discovery?” because large-scale optimisation is a need in there as well. The problem is, at such an early stage as a startup and the technologies at this stage, we need to quantify success continuously. We need to know if we’re on the right path, and in finance, we see that immediately. It’s not clear, it’s not deterministic, but you can sense that you’re going the right direction. You can’t quantify success. It gives you a probability measure that you’re going on the right path. In that way, we’re solving large optimisation problems, and if we’re making profits in the market, then it means that we’re reaching a good solution.

That’s a good point, because if you’re making a profit, you can get that final benchmark that what you’re doing is worthwhile in the world, because it is difficult to benchmark an advantage. It’s difficult to prove that your way is the best way, period, but if you’re getting some kind of advantage — that’s why D-Wave throws around customer advantage, and I love that term — then they’re saying that this is the way to go. Right now, if they do reach this, they’re getting advantage or profit from a, like we mentioned, quantum-inspired approach. Because what you’re bringing to market currently is, if I’m correct, still running on that other hardware. It’s still running digitally or, like, QPUs or something like that.

Yes. There are ways to directly use the QPU. We could do that right now. We could reduce our problem. The ones that we’re doing right now, the trades that we’re doing, we can fit that into our QPU. It would leave us with less control. It would lead to a white paper or something like that — an archive paper. We could say, “At Icosa Computing, we partnered with D-Wave, and we used the QPU to do these trades and we profited x amount. Amazing. There’s quantum advantage. We did something using quantum for sure — real quantum.”

We find those approaches a bit ad hoc, so it’s just for the sake of using quantum, we’re using quantum. Our philosophy is that we’re using the most capable provider, or solver, or approach that’s out there. D-Wave hybrid is great. Again, it’s a bit quantum inspired. It’s not clear, because it’s a black box, and as researchers, we don’t like using black boxes. As a black box, it performs well, but we don’t know the internal workings — if it’s quantum or not. It is hybrid. There are all sorts of elements of QPU in there. We’re not trying to publish, or write white papers, or provide workshops to our clients. We’re just solving their problems.

Usually, the finance people, they don’t care about how we solve the problem, as long as we solve the problem. Then, for hedge funds, it’s even clearer: If you’re making profits, great. They’re saying, “We make, already, 40% with our method. You come to us. You claim this new technology, quantum, blah, blah, blah. Amazing.” Then you trade, and you make 30%. You add 10% to their profits. That’s all they care about. That’s how hedge funds operate, especially the mid-range ones, the ones that we work with. Maybe the large-scale ones are building for the future, and they want to try out this technology, and create a team, and have workshops, and be ready when quantum hits, but that’s not our approach. If we were doing that, we would be competing with 10 companies, and we don’t want to do that.

In some ways, you’re following a paradigm that we’ve seen before. It’s what the cloud workload models have been working on for a while: You’re coming to the cloud with the problem, you’re creating a workload and, in some ways, you don’t care what’s running on the back end. You’re buying instances for a while, you’re paying, you’re getting the results, you’re happy, and you move on. If they upgrade those machines in a few weeks, then you get better results for your money — so be it. One day, it might be that your same customers come back and they’re going to get better answers, and then it was a 100% quantum machine that you used that week. It’s possible.

Exactly. We do care, by the way. That’s our job. We care about what we use.

Of course you care. But sometimes that customer doesn’t care.

The customer doesn’t care. Maybe at a cocktail party, they tell like their friends that they’re using quantum computing, or they tell the future investors in their hedge fund that they’re employing the best technology out there, they’re doing frontier tech research, they’re working with quantum computing providers, but in the end, what matters is if they made profits or not, or if they’re solving a problem that they couldn’t solve before.

There’s no way to do very large-scale, discrete optimisation right now. If you use the mainstream, the best state-of-the-art available solvers, publicly available — and, again, these are not free, but they are publicly available. It’s not like an internal tool built at Goldman Sachs by the 400 best quants in the world, but it’s something that we can pay for and use, like Gurobi or IBM CPLEX. They can go up to 3,000, 4,000 variables at most, all-to-all connected, and provide decent results.

And we don’t even know. Again, we can’t solve these problems. The best way to benchmark these methods is to plan the solutions where you know the solution beforehand, and you can compare. You can’t solve a 14,000 linear all-to-all connected variable problem by brute force, but if you have an answer and create the problem from that answer, you know the answer like a priori — you go to the solver and you can benchmark. I’m not sure exactly about the numbers. Let’s be optimistic: Let’s say about 10,000 variables. Discrete optimisation does not work on current state-of-the-art optimisers. There’s definitely a need to do large scale. There’s definitely no way to do it right now, unless you’re using these special techniques that are, well, quantum inspired.

Inspired. Yes. And that’s what we talk about too with customers. When we do a proof of concept for them, it’s all about, “Let’s solve the biggest problem we can until we can just barely break the machine — bring it to its edge.” Then, there’s extrapolation. That’s just the way it is right now. You’re seeing that these machines can do these amazing things, and you know that next year, when they have so many more qubits or whatever, better fidelity, we’re going to be able to, hopefully, explode this out and get some real advantage, and that’s why we’re all in this gray area right now where we’re trying to eke what we can out of the stack.

Yes. And even if you’re not optimistic at all — there are always skeptics, and scientists usually are skeptics. There’s no proof for quantum annealing advantage, theoretically. I don’t think there is. Please correct me if I’m wrong. I’m not sure there’s a robust proof that quantum annealing can do discrete optimisation better than a scalable quantum computing machine can do it better than any other approach.

Yes. D-Wave uses that language: “customer advantage.” It’s what they use.

Exactly. We don’t yet. There are some quantum fluctuations happening at the beginning of the process where it’s not fully quantum mechanical throughout, because maybe the time scales are not large enough or small enough to finish the process in a quantum mechanical way, and entanglement doesn’t persist. But again, these approaches have inspired us as researchers, as business owners, to look at a different domain in a different way.

People have been doing continuous optimisation for years, but there hasn’t been much of a look at discrete optimisation. It was always put under the rug as something that we can’t do. It’s a hard problem. It’s NP hard. I don’t know if it’s NP complex — I won’t say something wrong — but it’s NP hard, and we’re not going to bother with it. It’s not solvable. And now someone said, “Quantum annealing can solve this efficiently,” and people started looking at — even from a classical cloud algorithmic perspective — this problem from a different side, and that provided us with new insight.

This has happened in the history of science many times: Someone comes up with a new method, and maybe it doesn’t work, but because of the promise, it inspires a lot of other researchers to look at the same problem, and the smartest people in the world look at this, and they find ingenious ways to solve it. That’s what we are exploiting and using right now. We also, as researchers, are looking at the problem and trying to solve it.

You described the history of fusion right there in a nutshell. Still going parallel. That’s a great point. This was very helpful. Our listeners will enjoy learning more like this about quantum inspired because, let’s face it, that’s where we are right now. There are going to be a lot of quantum-inspired things that pop up.

Did you want to leave us with anything? Anything you want to share about the company — what your next plans are?

Again, we are scaling up. Our team has expanded. Things are going well. We’re making profits looking at these methods, and it’s great to see that something that amazed me a decade ago is being used or, in some sense, is being researched and can be put to use.

When I first heard about quantum computers, it was David Deutsch’s argument for the multiverse — how can you get exponential power if you’re not dividing your computer into multiple universes and then doing some sort of parallel processing and getting an asset back. I’m simplifying.

His favorite — the famous 10500 quote that he has, yes. He’s been a hero of mine for years, and if I have one dream guest for this podcast, it’s him.

He was my examiner a few months ago for my pre-thesis presentation at Oxford. We did host podcast sessions with them at the Oxford Quantum Information Society. I can send the link as well. He’s amazing.

He is. He could talk about anything for hours, anything.

Yes, one of the smartest people on the planet, probably.

Now, we’re seeing this trend where people are getting into the industry because it became an industry, but then, a decade ago, there was no industry. You have all of these people who got into this without any financial benefits. When I was looking at quantum computing and we started the society at Stanford, people were, like, “What are you doing? What are you doing with your life?”

It’s amazing to see that now, people are looking at this from a fully practical perspective as well, and from the intellectual side, it satisfies me greatly. I know that the team that we’re building at Icosa shares a similar passion, and no matter what happens in the industry, we’re going to continue building products like this. I’m very confident in our team and what we’re building, and we’re very excited for the future.

Yes. Good luck. And I’m sure you’ll do well in this particular incarnation of the multiverse — since we brought up David Deutsch.

Thank you so much for coming on, and I enjoyed this. Good luck.

Thank you, Konstantinos.

Now, it’s time for Coherence, the quantum executive summary, where I take a moment to highlight some of the business impacts we discussed today in case things got too nerdy at times. Let’s recap. Icosa Computing is part of the Chicago Quantum Exchange Duality Incubator, as well as the Creative Destruction Lab. Before Mert founded the company, he worked at QC Ware and Fidelity Investments, getting some firsthand knowledge of coding quantum use cases and testing hardware-vendor claims of machine performance.

Icosa started small with algorithmic approaches and expanded. They began with two POCs for customers, trading $1 million for a period, and then $200,000 for even longer, to test their methods — small money for financials, believe it or not, but big enough to prove that the company was onto something. Now, Icosa works on solving larger real-world problems, and they perform daily trades. To do this effectively in the NISQ era, they often use physics-enhanced or quantum-inspired devices. Digital annealers are a good example of such devices and have the ability to handle many thousands of variables.

The goal for many financial problems is to find a good local solution instead of a global solution, and in such financial use cases, you can quickly quantify success, which is why Icosa prefers these problems to, say, physical logistics such as shipping or routing. The sizes of the problems Icosa works on would not lend themselves to current NISQ machines, which is fine, as the company business approach is not to show the art of the quantum possible with small POCs. Rather than reduce the problem size for experimentation and a proof of concept, Icosa instead keeps the problem size large and uses quantum-inspired hardware to solve the problems fully. It turns out most customers just want to have good results and don’t care what the approach is on the back end.

That does it for this episode. Thanks to Mert Esencan for joining to discuss Icosa Computing, and thank you for listening. If you enjoyed the show, please subscribe to Protiviti’s The Post-Quantum World and leave a review to help others find this. Be sure to follow me on Twitter and Instagram @KonstantHacker. You’ll find links there to what we’re doing in quantum computing services at Protiviti. You can also DM me questions or suggestions for what you’d like to hear on the show. For more information on our quantum services, check out Protiviti.com, or follow ProtivitiTech on Twitter and LinkedIn. Until next time, be kind, and stay quantum curious.