Toward Fault-Tolerant Quantum Computing

Hi everyone. Welcome to our Academic Innovation Support live stream at warframe Research. Joining us today is Michael from Q Control and Mads from warframe Research Quantum Team. I am the Program Manager at Academic Innovation Support Team from warframe Research. Personally, I'm so excited to have this conversation happen. If you are Already following our LinkedIn page, you will see the Countdown pages every day for this conversation. So now I will hand the headphone to Matt and Michael.

Thank you for the introduction. Hello everyone. We are very happy to have Michael Hodge, who is the Chief Scientific Officer at qcontrol, joining us live from Sydney. Good morning, Michael.

Morning slash, afternoon slash wherever you are. Different time zones, I guess. Yeah, so hi everybody. I'm Michael Hush, chief scientist at qcontrol. Been working in the quantum industry since I could study and. But specifically at qcontrol for the last six years and then did some research beforehand in the last 15 years. So, yeah, glad to be here and keen to talk about quantum computing.

Thanks for accepting our invitation. During American Physical Society March Meeting 2024, Wolfram Research, together with Qcontrol, we had a big announcement that got received lots of attention. We announced our strategic partnership together with qcontrol and you know, we released the integration of like, you know, one of the products of qcontrol together with Wolfram Quantum Framework. And that product is Firepole. And today we are going to talk about fireupal. But before talking about that very specific product. So let me start with like, you know, providing the big picture for our audience and I would like to hear your thought, Michael, on that. So we know that like, you know, one of the big questions that we have is how to scale up quantum systems. And by quantum system I mean the actual quantum hardware. Because that's the only way that we can solve hard, important problems. So we need basically many, many qubits and you know, likewise, you know, based on that we can create our logical qubits. There are promising like, you know, technologies out there. We have supercomputer, superconducting qubits, we have trap, ion, trap atoms, you know, we have spin qubits, we have photonics, we have. Which one will be the leading technology? We don't know yet, but everybody knows that like, you know, in order to achieve that quantum value, we must follow the approach of fault tolerant quantum computing. So when we talk about fault tolerant quantum computing, what do we mean by that?

Yeah, when we're talking about fault tolerant quantum computing, I think the key thing we're talking about is you're using more than one qubit to encode the state. And you're doing it in such a way that you can correct errors that occur. So I think there's a, there's an idea out there that if you can correct faster than the errors are occurring, you might be able to have a qubit that lasts for a long time in practice. What we expect at QController is just, it's another technique you're throwing into the mix. So you're fault tolerant quantum computer. What the early generations will do will be able to do computations for much, much longer. And so that's the big game you're really playing is lots of different techniques being added onto your device so that you can remove errors so either stop them from happening or correct them after they occur so that you can do computations for longer which provide bigger value using more qubits.

Yeah, excellent. So like I'm going to go ahead sharing my screen, just showing our audience like, you know, what we are talking about. So this was basically our announcement for APS March meeting which was the integration of, you know, FIRE or PAL to with, you know, Wolfram Mathematica and what's happening here. Like, you know, we can show it from one specific example. So what you see here is the prediction is actually the results that one get from a very specific quantum circuit. This is a, you know, phase estimation quantum circuit. On the top you see the prediction of quantum. For those of you who are not that much familiar, you know, phase estimation is a very important ingredients that we have in many quantum algorithms. So for us, this very specific, you know, distribution spatial, this maximum peak is important. And if you send your circuit which is basically, you know, this is the circuit that we have basically a collection of like, you know, quantum operators on different qubits. If you send it to, you know, one of the available quantum processing units, even the best ones that are out there, the lower bar chart, that's the result that you get from the hardware. And if you compare it with the one that is the expected value from the theory, you cannot get that much meaningful result from this experimental data. No good. But if you incorporate, you know, fire opal, you would see a very you know, even like, you know, without doing any statistical, like, you know, analysis people, by just eyeballing, you can see the big improvement if somebody used like you know, fire opals. So Michael, help us to understand better. So what happens when we create, you know, a circuit like this and when we feed it to Firepal, what FIREO Pulse does that we can see like, you know, an improvement in the performance of the hardware.

Yeah. So what the. To begin with, you gotta talk a little bit about the problems that occur. So when you apply it, when you send those gates to a quantum computer, they're implemented using some physical process. So if you have a superconducting device, you've got like a Cooper up here spinning in a particular direction and you're applying some microwave pulses. The challenge is those gates are imperfect and they might leak to other qubits. They might send a signal to the qubit they're not trying to send a signal to. They might not quite do this right operation or there might be. The qubits sometimes are not perfect and they might leak into the environment. There are all these different things that can ultimately cause errors. So what that means is when those errors occur, it will give you the wrong result. And the challenge is when you're using a quantum computer that is available today, you don't necessarily also know if you're getting a wrong result. So it's not telling you that things have gone wrong. So what we do at Q Control is we're integrating every piece of state of the art technology to remove errors. So what you're seeing here is, uses what we call error suppression. And error suppression is a technique that you will want to use with full fledged error correction. But until full fledged error correction gives a benefit, error suppression is a great technique that you can use to get huge improvements. So error suppression uses our understanding of the physics and the device to actually eliminate the errors before they occur. So how we do that is that we change the sequence of the microwave pulses that are applied. We change that shaping, we basically create interactions on the device which remove the errors that occur. As the process is going on, they effectively cancel themselves out. And so by doing that, by suppressing those errors, basically before they contribute to the final result, as you can see, you get a much better answer. It allows you to get the answer quicker with less executions. And there are other cases which I'll probably show later where it can even change the device from giving the completely wrong result to the correct result. So it's really important to use the best techniques available when you're trying to use a quantum computer to get the right answer.

So basically like, you know, a circuit like this which is basically a collection of like, you know, quantum gates, like, you know, in this case all of them are unitary operators. And at the very end we do the measurements. So at the level of hardware is like, you know, for this case, let's talk about like, you know, Superconducting qubits. So what the hardware understands is basically a sequence of pulses, right? Like, you know, different. Like, you know, different, you know, in different locations in time, different pulses. So when we use firewall pole, we are not changing this circuit. It's not like, you know, creating a new circuit, etc. Basically what is happening is mostly a sophisticated way of, like, you know, when to send those pulses. Like, you know, basically it's like I'm gonna pause at the right time and then I'm gonna send, like, you know, the signal, the pause that's corresponding to one of these gates, right?

Yeah, there's a. There's a slight subtlety I'll add to. So you're executing the same circuit, so the logic is the same, so to speak. The other thing that Q controls flower opal does is in practice, the qubits aren't connected in the way that they are in this diagram. So when I draw a normal quantum circuit, I can apply a control gate to any other qubit. The reality of the bytes is that each qubit is actually physically located in such a way that on a superconducting device, they can only communicate to the ones that they're neighboring. And there will only ever be like, two to three qubits next to each other. On most devices you can get access today. So in order to do a circuit like this, you actually have to, like, shuffle around the information on the qubits. Or you might have to change the way the gates are implemented so that you might use one, which is one thing we call like a native gate. So the actual pulse that's applied versus the abstract one that you see here. So in any case, getting this ideal circuit and turning it into what's implemented on the machine is called compilation. And we also have really sophisticated compilation techniques. So that means that when you use it, fire opal. It'll also often be a shorter physical circuit that's implemented because we're able to figure out how to compile it better for the machine itself.

So, like, you know, before we jump to the details of, like, you know, how firepower works, like, you know, again, like, you know, a quick reminder to our audience, like, you know, the big picture. So what you can do is like, you know, you can create your circuit. In Mathematica, the quantum prediction is like, you know, this. These other results, like, basically what quantum theory, like, you know, tells you, like, you know, what are the expected probability distribution. But because of the noise, you know, the best hardware that are available out there, like, you Know, here we are talking about digital quantum processing processing units. You cannot get a good result. So you, you need other techniques. You need to take care of the noise. Okay, so in this case, like, you know, noise suppression is one of the many techniques available out there. Of course we know that the full scale, you know, quantum fault tolerance, quantum computing, we are not there yet. We will discuss very soon, like, you know, what's going to, the future is going to look like, but at least we can get meaningful result from the current like, you know, hardware that we have out there. So that's basically the big picture now. Michael, let's jump in. Let us understand, you know, more details, tell us more about more details like, you know, how FIRE Opal works.

Yeah, so I guess the, the maybe bringing it back to what, what you're trying to do with the quantum computer, because I'll, I'll talk a little bit more about what the applications are, I think. So FIRE is attempting. What we do is we try to bring control techniques into where we're seeing people use quantum computers most. And so we've noticed there is like a really serious trend for people trying to use it for optimization of like logistics problems or optimizations of quantum chemistry problems. So what's going on at the moment is people are trying to use quantum computers in a way where you can leverage the strength of a quantum computer. So like put the processing that it's good at onto the device and then combine that with the as much classical processing as you can do. And examples of algorithms like this are called the quantum qaoa, the quantum approximate. What was it called? I use that term, so I forgot the name of it. But the other one is like a vqe, Variational quantum. Yeah, Eigenstate solver. So these are techniques which allow you to find ground states of quantum systems, but you can link that to physical processes. So you might have a set of delivery points you're looking for. You're trying to find the best way to send out your trucks to get parcels to people. That problem is really hard. It's called an NP complete problem. So the best you can hope for is an approximate solution. But there are definitely good, strong suggestions out there that a quantum computer may be able to get a better approximate solution faster than the best approximate solutions that we currently use. So that's an example where we're seeing a potential for an advantage with a quantum computer. But the challenge is when you try and run that if you don't use the right control techniques, if you have errors, you will find that as you run your quantum computer in the classical computer loop, as they're trying to solve this problem, as they're testing out optimization challenges, it will take too long. It will keep providing the wrong answer, keep searching, because those errors will lead a destruction. It'll send it off in a direction where it shouldn't go to. So we innovate kind of everywhere we can. So with regard to those closed loop optimization problems, we have a package which is designed to do optimization problems as well. And that allows you to. It takes into account the way in which you measure success. Each round, we improve the cost function it optimizes. We again remove errors using error suppression, and we also fix the measurement stage. So at the end, when you take a measurement, the measurements are never quite perfect and you can actually do corrections called measurement mitigation techniques to improve the distribution that you get as well. So that all comes together. It sounds. It's very complicated. It sounds very complicated, sophisticated. But ultimately, the great thing with the integration with Mathematica is you get to program your quantum computer in a really interactive visual way. So much quicker to learn, right. When you can just create a circuit diagram in a single command. And that's like, what I love about Wolfram products is this. Just so quick to get that visual information right. But then combining with fire, it'll reduce the number of times you have to execute before you get to the right answer. And it should also provide the right answer when you would otherwise get noise. So that's what we're hoping to provide with this combination is that speed of iteration where you can really just experiment and learn using Mathematica combined with some more reliable outputs from. Yeah.

So to clarify, you know, one thing, like, you know, when we talk about, like, you know, improvement, like, you know, by incorporating, you know, quantum processing unit, this improvement in the literature, when people talk about, like, you know, quantum advantage, that has a very specific meaning. So, like, one thing we would like to clarify is improvement happens at many different scales. Like, you know, it could be, for example, comparing to a classical optimization technique, it could be maybe 10 to 12% improvement. But it's still like, you know, when we look at the scale of the problem that you want to solve, like, you know, the army deployment case that you mentioned, or like, you know, the logistics type of problems that you mentioned, given the scale of those problems, even a tiny quantum value, 10%, 12%, not the exponential, this is not what people would call as a quantum advantage. That is still a big deal. And this is the part that I feel like in our audience they should keep that in mind. Like, you know, many of these algorithms, they might not provide a true quantum advantage in a sense that like, you know, theoretical computer scientists would call it quantum advantage, but still they provide a very good improvement. Would you like to elaborate a bit more on like, what do you mean by like, you know, improvement by incorporating quantum.

Yeah, I think that's, it's something we talk about a lot in house as well at Q Control. So like you were alluding to there, there are certain proofs for very specific problems where you can prove this, like exponential scaling advantage, like prime factorization, which is relevant to cryptography, right? Indeed. So in the long term that gives you like some really solid foundations that eventually if you build a quantum computer big enough, it will be able to provide an advantage. But the challenge is that that goal is really, really big. It's a, it's a huge effort we have to put in. And it's also, it seems like it's definitely going to require fault tolerance. Right. So the, some of the nearer term opportunities look like, what can we do with the devices that we have today which are starting to get to a size where you couldn't very, you can't easily simulate them on a classical computer. So do you mind, do you mind if I speak a little bit about that difference first? So, yeah, I think, I think like when I started in quantum computing six years ago, the devices that were available, like from IBM and other providers, they were great. It's great to have access. It was great to be able to experiment with real machines. But if I gave one of my undergrad students at the time and asked them, can you go simulate this? They could get their laptop and simulate that quantum computer in five seconds. Like it was these qubit and five to six to seven qubit devices, they were great to start learning, but they just didn't provide much of a value compared to what a student could just do themselves. The big change that we're seeing now is that the devices that are being produce, so IBM is producing devices with 127 qubits. You're not necessarily executing, you can't necessarily execute an arbitrary program. So the challenges with the bigger devices is that the long, the error rate means that you have to effectively do shorter circuits. So the algorithms that can typically use a lot of qubits because they were going to fairly deep, they're harder to run. And instead you typically want to run circuits which are relatively shallow, but can connect in principle, there's 127 cubits. So the interesting thing that's happening now is if I was to tell my undergrad, go and simulate 127 cubits, they could. There's no way, there's no way they could ever do that on their laptop. There's the size. Using completely standard, textbook, out of the textbook techniques to simulated quantum computer are no longer able to keep up with devices that are available. What's changed is now the world's absolute experts who've, you know, spent years figuring out techniques and how to simulate quantum systems. At the very best, they're the ones who are competing with what the quantum computers can do today. So it has been a real significant shift in the technology where it's been from what we as physicists would call toy systems to quantum computers which are pushing the edge of our limits of understanding of classical simulation of quantum devices. And that's a really interesting place to be at. So going back to your question about applications, the thing that I think is interesting to consider now is trying to look for those real potential practical advantages on devices that you can access today. And so you can't do those experiments with just a scaling argument mathematically because you have to take into account the practicality, you have to take into account the speed of the gates, the number of, the number of errors that are occurring. You have to experiment with real devices. And that's what I think is like a really interesting space to be working right now is finding problems, seeing how they fit on, on devices today, and then looking forward to see when there might be an advantage for that particular problem, which you would prove more in practice using real device time rather than trying to do a mathematical proof. There's a space for both of those places. There's a space for both, obviously, and it's good to have both of them pushing the industry forward. That's just because we at qcontrol love using real devices. So we're often also just really experimenting at that, that near term because it's just where we think have the most fun, so to speak. Quantum computers are a fascinating, complicated device. So there, it's good to try and try and figure out how to make it better.

So I know you put together a very nice slides that you wanted to share with our audience. You know, you can go ahead, share your screen. You had some of those interesting applications that you wanted to show. So let's go through that and let's discuss that.

Yes, okay. Hopefully everyone can see that. So we're just, just to open. So again, thanks so much for the opportunity to present here and talk more about our integration. So, Yeah, I. You can tell that this is actually my generic pitch deck, by the way. So let me. Let me just speak a little bit more. So we've talked a little bit about this already, but there is a really big opportunity in quantum computing. There's applications all the way through, from quantum chemistry to logistics optimizations, as we said, but it's also going to have an impact on how we protect our information. So there's a potential impact on the cryptographic techniques we use. It's a technology which could really change the nature of compute. And I guess in the bigger picture now, as we're seeing so much interest in AI, et cetera, there's this real need and other people push to push our boundaries of how much compute we can get more efficiently from the devices we make. And I really see quantum computing as a part of that. There's still not a clear answer exactly of how quantum computing will interact with machine learning. That is one of the really interesting questions, but at the very least, you really should think about quantum computing as another avenue in which we're trying to push compute to its absolute limit. So what is the problem right now that's occurring with these devices? So the challenge is, if I try and run an algorithm today on a quantum computer, as we talked about, there are a lot of errors, and those errors can be so bad that they can wipe the correct answer off altogether. So just here on the right is a search algorithm example, Grover search. You can see that what we're expecting to see is a peak with the answer we're looking for. So you're searching for the database and trying to get out a bit string corresponding to what you are looking. Not a database, sorry, like a set of information, you're trying to find the default settings. On a quantum computer, sometimes the errors can be so large that that correct answer is not only the small, it's the one of the smallest probability of the bitstreams that come out. So on the bottom you can see again, with using our Fire Opal toolset, we can suppress errors and mitigate errors at the same using the same tool chain to increase the chance of seeing the right answer. And so in this particular case, this is data from a real device. You can see we've boosted the probability of getting the correct answer all the way to it becoming the dominant mode, and so the correct answer is restored. So by building this technology and making it available online, which allows other people to also test and verify its performance, we've made a lot of progress in moving forward with Fire Opal. So just as an example, IBM is then taking this technology with us and embedded it onto the hardware as well. So which provides even more better results. And that's shown on the right. With the kind of advantages we can get with particular algorithms, we can really boost the probability of getting the right answer, which can ensure that you're using the most using, you're getting the most value you can from the devices that you're using. So if to go back, I kind of mentioned bits and pieces of this throughout the presentation so far, but just to kind of give you the full picture of all the components that go into the Fire Apple chain, first we compile it for the physical device. So we ensure that the number of physical gates that are used actually is minimized and we ensure that the circuit that is executed is done as quick as possible. Then we actually use a error aware hardware mapping. So this uses an understanding of how the gates are operating, which ones are operating, where to figure out how to remove errors. So we then interlace a variety of correction techniques across the circuit so that errors like crosstalk are effectively canceled out completely. And then finally we also actually optimize the shape of the individual pulses. We do that using a regular regular calibration routine to ensure the individual gate performance is better. Finally. So that's run on the computer and the computer will then make a variety of measurements. The measurement stage also is imperfect. So we finally have an error mitigation technique that corrects the errors, that corrects some of the errors that occur due to the measurement part of the process, not just the computation. And those results are then returned to the user. So using those techniques. So I think the thing that I really wanted to highlight here is using these techniques you can see that the success probability as we increase the depth of the circuit. So using Q control here on the top is the purple line. So you can see that the probability of it giving the right answer stays higher for much, much longer than it does without the control techniques we provide. I did really want to highlight here though. It's not flat, so, so you're not going to, you're going to eliminate as much error as possible. But there's a, there are certain processes you can't fix. So for example, crosstalk, where qubits are actually in talk to one another. That's something we can basically almost entirely eliminate. The problem that can't be eliminated is when individual qubits just leak. So they basically send information out into the environment. It's the quantum way of saying it. But in a physical process, it's like it relaxes. You lose some of that, that excitation or thing. It goes into the system, goes into the material. So that's called the T1 process. That is the type of error process that you can't correct with error suppression techniques. So there's still a limit to how deep you can run your circuits before you get the wrong answer. And that is where we would look to quantum error correction techniques to then put on top of these techniques we've talked about today to help correct that error process. So also stay tuned. We're always improving Fireapple, always looking to integrate new techniques, always trying to push it further and further.

So Michael, like, you know, if you go back to the previous slide, like, you know, how sensitive is firewall algorithm with respect to a specific backend that the user picks? Like, you know, basically harder by hardware. Like, you know, do we see a different performance or like, you know, on average, like, you know, we should be able to see a good improvement in any backend that we pick.

Yes. So we're, we only ever put backends up online where we can see a significant improvement. There is small variation between different devices, but, but I'd say largely we get comparable performance. We work hard to make sure that the techniques that we've developed are all implemented correctly for the particular device that it's running on, so that you're getting a comparable improvement in performance. So that, that says. That's also because we have been, I think at the moment primarily there are superconducting devices that are available on the, the platform. So as we move forward and we start making more available, more different physical systems available. So we're looking to put Ion systems up online and put Rydberg arrays as well. That will create probably quite different expectations of improvements of performance. Because when you've got radical, radically different hardware platforms, the way in which the correction and error suppression occurs is really, really changes. But we live by mantra of like always improve, so you're always going to get an improvement with Firework or we just, we just make sure it never disadvantages you. And if you're typically, if you're using a device of the same nature. So all superconducting devices at the moment, they all pretty much reflect each other in terms of the amount of performance that you're going to get.

By the way, for our audience, I know, Michael, for your information, every summer we have a very large summer school and one topic that more recently Our students, we call them students but they come from like, you know, all different level of academia and also industry. So for them like you know, listen very carefully like you know, during the summer school when we discuss you know, more you are going to use like you know, many of like you know, these tool sets in order to perform like you know, algorithms on the real hardware. Another question, like you know, not the question reminder for our users, like you know, you can see it both if you use directly firewall or if you use it in Mathematica. Whenever you send any job to any hardware, if the specification of that hardware is in a way that like, you know, you might not be expecting a good improvement by using firewall. Fireopal will give you some alarming messages telling you look, you know, the hardware that you are using, it has this error rates and within these error rates do not expect like a big improvement or such. So I just wanted like you know, to bring like you know, to audience attention that like you know, read those like you know, detail like you know, messages. They are very helpful.

Yeah, I am. That's one of my, I think that's one of my favorite features of Fire Apple because it's, it's always nice when an error message gives you value. It's just like as someone who's like always trying to make good products, I love it.

Yeah.

So yeah, that was really important to us. And you're like on this plot, it's basically if you go too far to the right, we basically tell the users like just don't waste your money, just please, please, please shorten your circuit because you don't want to, you don't want to do an execution that isn't going to give you a meaningful answer. Yeah. So I'll just briefly finish up on some of the applications that we're looking at. So yeah, as I mentioned earlier, one application of quantum computing where you're trying to leverage the advantages of your quantum computer and then supplement it with as much classical compute as possible is these solvers like qa away or V2E where you put the quantum computer in a loop with a classical computer and you ask it to optimize a particular control problem. So that that basically ends up you take a graph which represents your particular optimization logistics problem, you feed it, you then have a translation layer which changes that into a set of quantum circuits which you're optimizing and that that can be useful for providing insights into particular problem. So on the right is an example where we've looked at transport challenges and Basically we've got convoys where you need to get from A to B. You can have really interesting challenging constraints. Like the particular order may be important when you're trying to optimize these things. We've been looking at the optimization problems which have like direct relevance to real business cases. And so you can translate that kind of optimization problem into a quantum computing algorithm and then execute it. So, and just to speak to the earlier point, yeah, so you're not expecting to see quantum advantage right now today with these problems compared to state of the art solvers. But what we've been doing is we've been looking at how close they are getting and trying to figure out where there are bigger opportunities. So at the very least we can say that you need, it seems like you need like 300 to 400 high quality qubits before you get into a regime where the classical solvers for these kind of problems are failing enough that there's a gap, there's a sufficient gap that a quantum computer may have an opportunity to actually compete and provide an improvement. So that just is an idea to give you the kind of scale of where you need to get to. But when I said 300 to 400, it's not just raw qubits, they do have to have a high quality. Basically the error rate has to be low. So just this is an example of us using our control techniques to see a real quantum difference when it comes to solving a particular optimization challenge. So, and as far as we're aware, this is the biggest demonstration of using QAOA to solve a max cut problem. So the, the most we could push the current devices that are available, so there are 127 Cubit devices, but the maximum we could get to was 70 qubits, was the amount of qubits in this actual graph which we sold. And what we did is we put in a optimization problem where you're trying to find solve this thing called max cut. So basically trying to find a series of cuts across the graph so that you've got the maximum amount of clustering. So what you can see here on the right, what's the most important here is that what we did is we considered just like a classical random sampling, so that we call that brute force. And that shows you on the left, the distribution of answers you would get if you classically just randomly sampled the graph. And what we're showing on the right is after a handful of operations using QAOA in that closed loop at the end, that is the distribution we got from the Quantum computer with the modified parameters that have been optimized and you can see that it's correctly and effectively with that search process shifted that distribution. So you're now able to find the maximum cut solution with a high probability, which is actually 95 in this case. So it's effectively learned and pushed that distribution.

So, so one side note regarding like, you know, this whole, you know, quantum optimization problems in particular like an application for the max cut, all of them are like, you know, for many of them there are available classical algorithms. Like, you know, these are like, you know, many of these problems are like, you know, famous like, you know, classical optimization problems. And when we talk about like, you know, adding quantum flavor into that, are we really talking about like, you know, quantum flavor as if the algorithm is classical quantum hybrid or we are talking about like you know, 100 quantum approach to solve the same problem?

Yeah, so the, so what, what we typically what you want to be comparing to is the state of the art classical solver. But the, the challenge is often that the way in which the classical solvers solve a problem compared to the quantum approach is just so radically different. It's very hard to do head to head comparisons other than just timing. So you just run them both and you say one's faster than the other. So I do want to say that's not what we're trying to show here. So that, that, that brute force is not meant to be like the classical best. It's, it's what, what it's trying to. What we're doing here is just saying is the quantum computer doing anything? So there is an element where you really want to make sure when you're also doing these optimization algorithms that the quantum aspect is contributing because you can access, if you're not doing a good job with this, you can do QAOA optimization where the quantum computer is actually so many errors are occurring, it's almost classical, but it's still trying to do some learning based on the classical information that it's randomly sampling. And so you can, QAOA sometimes can, particularly for smaller size circuits may converge even though you've gotten a lot of quantum going on. So what we're doing here is we're just really showing that the quantum component in the process is making a really big difference. It's actually moving it to a better solution. And then we're giving some kind of understanding of how much improvement that has occurred in terms of comparing to state of the art. This is not faster than state of the art. And particularly for max Cut. I think the estimates that last time I looked is you have to go higher. I think it's like 800 to 1000 or 1100. Those are the kind of estimates of the number of high quality qubits you would need before you be competitive with classical heuristics. So again, I just wanted to highlight it does matter what particular optimization problem you're looking at. So the one I was showing before is what we call highly constrained optimization problems. So highly constrained optimization problems can be the case where there does exist a bigger gap, where the classical heuristics are beginning to fail, while max cut is a relatively simple optimization problem. So the quantum computer is really going to have to be pushed a lot further in order to be competitive with the best solver for that case. Yeah.

In fact, like, you know, go ahead.

I was just going to say that that finishes the presentation. So we'd love to, I can go back and if you got any further questions on that last one, happy to address it.

Yeah, awesome. So like, you know, we are, as we are discussing like, you know, there has been like, you know, a couple of questions and like, you know, I was trying to provide like, you know, the, some relevant links, you know, for those questions. So as Michael said, like, you know, the, the, there are some, you know, like, you know, the whole literature or like, you know, the way that, you know, people are talking about like, you know, quantum utilities or quantum applications. Sometimes like, you know, the true description, it could be misleading and like, you know, what Michael explained here is basically varied to the truth of what we can get, what type of value we can get. If you include like, you know, some quantumness into your algorithms. We are very aware that like, you know, some of those algorithms, the state of the art of classical optimization problems, they perform better. But like, you know, we are like, you know, right now we are like, you know, at 60s, 70s, the time that people just started exploring classical computers and nobody knew what they can get out of that. And some of those algorithms, God knows how much it's going to take just to run that. But once we learn how to do those simple cases, we can scale it up very fast and we can bring the true quantum advantage. So in terms of applications, Michael, optimization is like a big topic. Like, you know, many people, they talk about like, you know, optimization type of problems. What other quantum applications are out there that like, you know, quantum method can bring true value in addition to optimization?

Yeah, I, I think, I think that's a really big question which we think about a lot. So I do. So there's a lot of interest in the intersection between quantum computing and machine learning. So we, but the, we continue to do our best to support people who are doing research themselves on quantum machine learning. So far it is a really interesting topic in that, in the literature you can see there are very clear evidence that quantum machine learning, when you're trying to learn quantum states, does appear to have this certain advantage. But what's, what's challenging is that the reality is anyone who's trying to use machine learning at the moment in any commercial sense, they, they have real information, they got real data. And the other thing is they want to know the answer so they want to measure it so they can't keep their data quantum, so to speak. So it's, it's, that's the real push, I think. And in that case we don't, we just don't know. We're just experimenting. Right. So I would say first and foremost, I'd say one should be experimenting like this is there is a really interesting turning point we're going right now where you can be doing experiments, you can be exploring these problems, but you've always got to be honest and clear with the people you're working with. Right. Because you've got to put in go and no go. So that's like a big thing that I would recommend for everybody is just don't try and implement a full scale solution with a quantum computer. Never promise this. That's madness. Technology isn't in that state. Instead you really got to be telling people if they're interested in learning about quantum computing, if they're interested in trying to investigate cases, go through a process with them. So do some scale, small scale experiments and put in go or no go points where you, you put a real quantitative check if that quantum computer is going to give an advantage against what you've considered the appropriate classical benchmark. And you don't implement a solution until you do those checks. Like that's very, very important in terms of other applications, I think try to, I'm trying to think of the things that we do a lot, I think. So. Quantum chemistry also continues to be something which is getting, we've seen a lot of traction in that in Japan as well. Like a real interest in doing a lot of quantum chemistry is bubbling up there. Yeah, we've been doing our best to improve closed loop optimizations which are relevant to those problems. That's another really interesting place where I think the big push there is really finding that intersection of like quantum chemistry experts who can find those problems which we haven't figured out yet, which may be able to be executed on a smaller scale and more relevant scale. But the quantum chemistry challenges at the moment look big that you need to do before you get to the potential place where they have an impact. So that's definitely a place where we're doing a lot of support of experimentation, but I don't expect to see that in the near term. Although I would say there are people out there, again, experimenting with machine learning techniques in that tool chain. So some people are not trying to necessarily completely take on the full quantum chemistry problem on the quantum computer, but instead look at different parts of that chain. They can get improvements with machine learning at different components. So, yeah, I think we're just an interesting time where some of the, the big picture things that everyone's looking a lot at is quantum chemistry optimization, machine learning, but it's all about experimentation and genuinely testing. So another thing that I'd speak to is a really important thing that I think both Wolfram and qcontrol do is we just put you also, you've got to put your software out there. So if you put your software out there, you put your solutions out there, then people can test it, people can verify, and people can check where your conclusions you're making as well. So, please, I would also recommend everyone don't just read the papers, because the one thing I would say that I've been in the quantum industry for this long is the papers on both sides can often have a particular kind of line that they're pushing through. So go and go get on a machine, try it yourself, figure out what's actually going on. It really changes your perspective on the state of the devices compared to just.

The literature related to your point. In fact, the main reason that I mentioned our summer school is this is like, you know, three weeks intense program that we want, like, you know, everybody to basically do a computational experiment. Okay. Like, you know, sometimes like some of the students, they would tell us, like, you know, okay, what's going to happen? And our response is nobody has ever tested that. We really don't know. Like, you know, if we combine like, you know, this quantum algorithm plus this, you know, chemical synthesis, what's going to come out of that? These are like, you know, the experiments, computational experiments that nobody has explored. And one beauty of like, you know, I think the tool sets that has been developed by you guys and also Wolfram Research for more than 35 years is we want to provide that environment for anybody. Just go and do those computational experiments. And nowadays like, you know, one thing that is great about content technology is at the very end you can also interact with the real hardware because all of them are, most of them are available through cloud services. So like, you know, you can have your own algorithm and then really like an experimentalist, you can run it. Even how fancy or how complicated is like in our algorithm, you can still run it. An actual computer hardware computer. So I think, you know, this is, we are in a moment that we need more of those experimentations. I fully agree with you. And this is like uncharted territories. And we are inviting everyone, especially students like, you know, when they want to learn quantum computing, especially quantum applications, just go there and do those experiments. Like, you know, it's, you know, I think like, you know, we really need to be open minded and we need much more developers developing true quantum algorithms.

Yeah, yeah, I agree. And I, I guess the other thing is, and it's hard, so don't get me wrong, it's like the thing that we're trying to offer to help it be a bit easier. You want to be fast, so you want to iterate quickly. So fire Opal and Wolfram. I think that's the thing we're trying to also offer is the ability to learn quickly. But I would say there's a lot of still unknown questions also about how you really assess the quality of your results. So there's been a long history in academia of using certain techniques on experiments. But I think one thing we've learned at Q Control a lot is that they're so expensive. So the techniques which are kind of standard for testing quantum computers, there's sometimes controversy. So for instance, quantum volume has gone in and out of fashion depending on who you talk to. But, but in the, in the, in practice, quantum volume is also something that's not achievable by the average user because of just the huge amount of execution that's required. So it is also there's also really interesting challenges to figure out if you're even just trying to assess the quality of your outcome. So just asking, is this working? How do I know it's working? So. And there's still opportunities to develop techniques for yourself or techniques that would help other people in that space. So I do want to also speak to. It's not always, it's not just about the solving the problem and trying to be a classical computer. There's actually a lot of things to learn on the lowest level of just like verifying or defining what success means when it comes to doing that comparison, that's, I think, still a pretty open question too. So, yeah, lots of. I think it's important to try on real devices and it's important to. Also, everybody's, like, trying to figure out, you know, what it means to have an advantage or what it means to be correct or incorrect on a Quantum device. It's still something to. You gotta learn and practice. Yeah. Yeah.

I think both of us, like, you know, for our audience, both, you know, Q Control and, you know, Wolfram Research, like, you know, we care a lot about, like, you know, education and training, like, you know, the future workforce, especially on emerging technologies. Like, you know, we have AI, we have Quantum, and who knows what's going to come next? Especially talking about, like, you know, education. This was a conversation that we also had with our partners, like, you know, Q Control. We notice some sort of, like, you know, unbalance in terms of, like, you know, the, the gender unbalance in terms of, like, you know, developers, researchers that we have, especially on Quantum. So I would like to take this opportunity and invite, like, you know, especially like, you know, girls, women, Please, like, you know, interact with both companies. Like, you know, both of us, we are, we would love to see more usage of, like, you know, technology. We would love to help especially, like, you know, girls, women on, get your hands more dirty with, like, you know, learning these technologies. And this is basically one of the missions that we have. So make sure that you guys, you know, reach out to both companies and we will be more than happy to help.

Yes. And. And for those in Australia as well, we can Control work alongside. Quantum Women is a great organization in Australia, which is a resource where you can meet other women in physics. Quantum stem, and it just gives you an additional network to work with to figure out where the best jobs are, where the best places are to contribute. And so, yeah, I highly recommend that as an organization for people in Australia as well. Yeah, excellent.

So looking at like, you know, the questions from our audience, like, you know, we managed to address, like, you know, all of them. So, Michael, any concluding remark? Anything else you would like to share with our audience?

No, I think, I think the theme of just try it, I think that's the main thing I tell everybody is, yeah, get on, get on the. Get on, Wolfram. Give it a go and see what you learn.

Yeah, thank you, Michael. Like, you know, I really enjoy our conversation a lot more to learn, like, in our audience. Remember, like, you know, as Michael said, Quantum is a difficult problem, but it doesn't mean, like, you know, we cannot be smart. We can come up with the smart approaches and there are a lot to do. You'll be surprised how much like you know, open problems are in Quantum and of course the interaction between Quantum and other topics such as like you know, Quantum plus LLM, Quantum plus AI. These are like, you know, highly hot topics that like in a both sides like you know, we are actively exploring them. If any interest, please reach out to both of us, you know well from our partner, you know, Q Control. We are committed both academia, outside of academia and we would love to hear from you guys. So I'm going to hand over to Yi.

Hi everyone. If you have any comments, suggestion about our live stream or you want to find a women to talk with, please follow LinkedIn page and you can reach out directly to me. We can talk about like summer school or project can cooperate or you want to be part of our livestream. We are looking for people to join our conversation. So as you can see we post our upcoming events here. Actually we will go to the Chicago Quantum Recruiting form forum at UChicago this Thursday. So if you are there say hi to us. We can grab coffee, have a friendly chat about the future of content and.

If you guys wants to know more about like in our partners such as like you know, the Q Control, Michael himself and more, just let us know. We will connect you with them.

Great.

Excellent. I think you know, that's a very good point. You know we can wrap up our live streams. Thanks again Michael. We really enjoyed this conversation and thanks for your contribution.

Thank you all.

Thank you everyone.