Transcript for Practical AI #200
Daniel Whitenack:
Welcome to another episode of Practical AI. This is Daniel Whitenack, I'm a data scientist with SIL International, and I'm joined as always by my co-host, Chris Benson, who is a tech strategist at Lockheed Martin. How are you doing, Chris?
Chris Benson:
Doing well, doing well. How are you today, Daniel?
Daniel Whitenack:
I am doing great, because any day is great when I get to -- any sort of throwback to my old physics days, which mostly I don't get to dabble in much these days, is a fun day... And today we've got a cool intersection with that world. We've got Yonatan Cohen with us, CTO at Quantum Machines... And we're gonna talk a little bit about an update on quantum computing and how that intersects with AI. Welcome, Yonatan.
Yonatan Cohen:
Hi. Thank you. It's really cool to be here. I'm excited to talk to you about some quantum physics as well.
Daniel Whitenack:
Yeah, great. Great. Well, I'm guessing most of the time on this show, and most of our listeners are probably used to hearing us chat about neural networks, or GPUs, or classical computing, and we have talked about quantum computing on the show, but it's been quite some time, and I'm sure that the field is advancing quickly. Maybe just as a starting point, could you remind us of the general pitch of like what is quantum computing, and maybe why people are interested in quantum computing more generally?
Yonatan Cohen:
Sure. So quantum computing is a new way to build a computer, that's based on the laws of quantum mechanics. It's kind of interesting, because quantum mechanics as a theory of nature was developing parallel to the development of computers in the last hundred years or so. And on one hand, we've developed this amazing technology that we now have, which is computing, that's based on classical physics; it's based on the classical laws of how the Universe behaves. But in parallel, in the last hundred years or so, there is a completely new understanding of how nature works on a very fundamental level.
\[03:54\] Somewhere in the very late '70s, early '80s physicists started to understand that perhaps we can use these new laws of nature we've discovered to also build a new type of computer, that's going to harvest this weird behavior of nature called quantum mechanics. And it turns out that you can do that, and it expands the notion of what we mean by a computer, and essentially allows to build a computer that has stronger computational power, at least for some problems; not for all problems, but for some very difficult problems with classical computers - when we say classical, meaning not quantum, but regular computers. For some computational problems that classical computers have a very hard time to deal with, quantum computers could solve them very easily.
Chris Benson:
So as you're talking about this new type of computer, would I be correct in thinking that I can't run out to the store a Pentium or an AMD, and pop it in with the RAM on the motherboard? What's different about that approach on the hardware side, and then what would you use it for that's different from that kind of classic idea that we all have been using all our lives?
Yonatan Cohen:
Yeah, so the main point of quantum mechanics and before quantum computers is that while we see things in the day-to-day that things are in a specific state - you know, I'm drinking coffee from my cup, and I put it down, and it sits in one place, it sits in one place and not in another place... So in quantum mechanics, things can actually be in various states at the same time. And while we don't see it with coffee cups, we do see it with electrons, for example. We can actually put an electron in what we call a super-position. That means it's in two places at the same time. And this is exactly what quantum computers take advantage of.
So the basic building block of a classical computer is the bit of information; it's a system that has two states - it can be either in the zero state, or in the one state. And then if you have two bits, then they can be 0/0, 0/1, 1/0, 1/1, and so on. If I have 8 bits, I have 256 states, right? 0/0/0/0/0/0, and 0/0/0/0/0/1, and so on. But at every single point in time, the classical computer can only be in a single state of all of the bits of information that in holds. And then we manipulate those bits; we go from one state, to another state, to another state, and so on. So it's like a state machine. We're moving from state to state, in order to solve a computational problem.
But quantum computers replace this notion of a bit with what we call a quantum bit, or in short qubit, which is a system like the electron that I told you about, that can actually be in two states at the same time. So it can be in both 0 and 1 at the same time. In fact, the system can mean 0 and 1 with different weights. So it can be a little bit 0, and a lot in 1, or it can be a lot in 0 and a little bit in 1, and so on. And when now I put a lot of qubits together, I now can be in this massive super-position of many states, all the states of the system at the same time, and you can use that to -- you can use this parallelism to, in some cases, do parallel computation, instead of going state by state, if that makes sense.
Chris Benson:
So with that massive parallelism, it sounds like it can handle probabilities, based on what you were saying, in many states, very well. What type of problems does that kind of capability lend itself to, that maybe our traditional - you know, the laptop I'm on right now - would not be as suited for? What is a problem that having lots of qubits to solve - what does it look like to address?
Yonatan Cohen:
\[08:09\] Yeah, it's a great question. So there are certain problems that we know for sure that quantum computers can get an advantage of. One great example is the Shor's algorithm. So that's an algorithm that actually breaks code, because it finds the prime factors of a large number. So if you take a large number and try to factorize it to its basic prime factors, that's a problem that's very hard for classical computers.
And the problem has some structure, it allows us to use this parallelism of a quantum computer in order to solve this problem exponentially faster than what we can do today in a classical computer. Now, it's very hard to explain exactly what in this problem makes it so that we can use this quantum parallelism to solve it so much faster. You're gonna have to look at the details of the problem, the structure of the problem in order to see how you can take advantage of this quantum parallelism.
So we don't know exactly, we cannot actually categorize exactly all the problems that quantum computers will solve much faster than classical computers, but we have examples. So we have Short's algorithm, we have Grover's algorithm... Grover's algorithm is an algorithm that solves a search problem. Searching in an unsorted list, finding a specific element of interest in an unsorted list, which also can map to very general optimization problems... But this algorithm, for instance, is not giving us an exponential speed-up. It actually only uses a square root speed-up over what you could do with classical computers.
So there is this zoo of algorithms that gives different kind of speed-ups for different kind of problems, and people are still working very hard on the problem of categorizing exactly what quantum computers could solve faster than the classical computers.
The other thing is that there are also quantum algorithms that we have today that use this parallelism of these qubits, and that we don't even know that they are going to work. So these are heuristics, basically, that people have come up with. There are good reasons to think why they would give us a computational speed-up against classical computers, but nobody can prove those. So we are basically in this kind of sub-situation that actually maybe AI was a decade or two ago, where we have some thoughts about these algorithms, people are very hopeful that they will work, and these are relevant for optimization problems, very general optimization problems... There's an algorithm that's called QAOA, that solves combinatorial optimization problems, and then another algorithm that's called VQE, which solves for chemistry problems...
These are heuristic algorithms that we just have to build a machine and try those algorithms on, and see if they work better than classical computers.
Daniel Whitenack:
Yeah, I have one maybe naive question from my own perspective, because I think -- obviously, you're well plugged into the state of the art. Chris, actually, I think you know a little bit more as well... But for me, I'm thinking -- okay, I've been hearing about quantum computers for some time... One sort of naive question is what actual quantum computers exist in the world right now?
Because if I go to one of the cloud providers, sometimes I'll see like "Oh, here's the AWS quantum thing", but when I look at it, it's really just so you can run simulations of quantum research.
\[11:58\] So let's say I want to run something on an actual quantum computer. What would be my choices right now, in terms of where those actually exist, and kind of the state of what they are, and how you see that sort of shaping up over the coming couple of years.
Yonatan Cohen:
Yeah, so today you can access quantum computers through the cloud with several players. So IBM, for example, has a quantum cloud service that actually gives you access to great quantum computers. You can log in and you can actually run algorithms... But these are small quantum computers. Of course, they're still a prototype of the technology. But you can actually run algorithms and you can see how they behave, as well as on simulators. You can compare, and then you can find that your quantum computer is doing not as well as the simulation, because it has a lot of errors; this is the main problem with quantum computers today... But you can do that, and that's actually very cool.
IBM, for example, they give access to their own computers, but Azure and AWS have launched their quantum services in the last few years that give you access to third-party computers... So many different computers that are built by full-stack quantum computing vendors, like IQM, QCI... These are startups in quantum computing that build full-stack quantum computers, and you can access those machines actually today through the cloud, which is very cool.
And they also give you access to simulators, and you can run on those as well... Which I think is quite amazing, to be honest. I mean, there are certain experiments that I used to do during my Ph.D. at Weizmann Institute only 12 years ago that... It used to take a Ph.D. student maybe three or four years to set up such an experiment and to run it, and now I can just log into one of these cloud platforms and I could run the experiment that just ten years ago it used to take a few years to set up, I can do it probably in a few hours... And that I think is quite amazing.
Break:
\[14:16\]
Chris Benson:
So one of the things I was curious about just to follow up on what you were talking about a moment ago is just to kind of set my expectation... With quantum computers, assuming that obviously because we're going quantum, we're not going out and getting a typical CPU, is there anything -- is it mainly the CPU is a quantum CPU that you would put into the computer, and is there any other types of changes that you would make to a classical computer, just to get a sense of it? How close is the generalized architecture of a quantum computer to a classical computer? How much has to change to make that leap?
Yonatan Cohen:
Okay, that's a very good question. I think that a lot of things are very different in quantum computers, because the basic operations are fundamentally different. Classical computers were built based on bits, and then you have gates, It's how you manipulate the information. So you can actually build an entire computer from a NAND gate. But in quantum we don't even have the notion of a NANG gate. We have other gates. We have a Hadamard gate, we have a CCNOT gate, we have different elementary logical operations on qubits. So that makes it so that the entire stack, or at least the low-level parts of the stack need to be very different.
\[16:17\] But then I would also say that eventually it's not -- I mean, the way I see quantum computers, especially in the early days bringing us an advantage is not just by being used as standalone computers, but actually as more of an accelerator inside a more heterogeneous data center.
So the quantum computer could do only certain problems. It's not going to replace the CPU, and it's not going to replace the GPU. There are things that there's just no reason to do them using quantum devices. Maybe in 50 years, because everything is going to be quantum, and quantum devices are going to be as cheap as classical computing devices... But right now, you'll have your CPU, you'll have your GPU and you'll have your QPU, and the QPU will be used to accelerate certain types of sub-routines that can help your entire application. So that's why it's important that we also build integrations of these quantum computers into a more heterogeneous compute environment, and allow people to program what we call hybrid workflows, so quantum/classical workflows.
Chris Benson:
Could you describe what one of those workflows is like? And that was a great explanation, by the way, of the difference in a quantum computer... But you've mentioned several times along the way about that hybrid fitting it into a larger architecture that includes a lot of those classical components... But I'm trying to get my head wrapped around what kind of problem as a user I might solve, an example that would use both quantum and the hybrid, and what part of the problem goes to each. Can you give us some sort of example on that?
Yonatan Cohen:
Yeah, sure. So the typical example is what we call the variational quantum algorithms. Actually, both the algorithms I mentioned before, QAOA and also VQE, these are variational quantum algorithms. What this means is that essentially, what we're doing is we're trying to minimize the cost function of interest. That's the problem we're trying to solve. But the way we do it is that the quantum processor is only computing the cost function.
Let's say we have a cost function that's very hard to compute on a classical computer... So the quantum processor only basically calculates the cost function given a set of parameters, but then the optimizer sits on the classical side. So what we're doing is we're running quantum circuits, what we call parameterize quantum circuits, or we're running a quantum program with some input parameters, and then the quantum processor computes a cost function, and then the result goes into the classical processor, and it would run an optimizer, like a gradient descent, or something like that. Then that will generate new parameters for the quantum circuit, the quantum program. And then it goes back and forth like this until we find the parameters that would minimize the cost function.
Daniel Whitenack:
Yeah, this is a really interesting example, because I've been thinking about like - well, okay, if I'm connecting maybe some of the challenges within the AI world to this space... I mean, you've just mentioned gradient descent, you've mentioned optimizing with a cost function... Like, this is all definitely very connected to what might happen in like an AI training scenario. Of course, there's all sorts of problems, or there's certain computations, like you said, that work well on a GPU, and that's sort of scaled up right now... But I'm wondering, as you're kind of diving into these types of hybrid problems with quantum machines, what might you see right now in the AI industry, or certain sets of problems in the AI industry, like the challenges that the AI industry is facing in terms of compute - what are people thinking about in terms of what might overcome some of those challenges on this sort of hybrid computing side with potentially a quantum advantage? What's the current thought process around that?
Yonatan Cohen:
\[20:35\] Yeah, so I'd hope that, for instance, using some of these hybrid algorithms we'd be able to solve some optimization problems that are also relevant to AI training. To be honest, I'm far from being an expert on quantum AI, or quantum neural networks specifically, but I know a little bit about the subject and a little bit about the challenges in AI, and I know that some of the algorithms that we have for quantum computers could solve these optimization problems.
So when we train a neural network and wanna optimize its parameters, in many cases this is a very harsh optimization problem. In some cases it could fill exactly these kind of optimization problems that a quantum computer could solve, and specifically these hybrid algorithms might be able to solve. So then we could use these hybrid quantum-classical algorithms to train neural networks. That's one example.
There's also some recent works about quantum neural network where the network itself is quantum, and some mathematicals proofs why such networks would require... Again, in some specific cases, but in both of the cases why they'd require much less data to train. And that, I think, is very important, because in many cases we just need a lot of data to train those neural networks, and in many cases we don't have enough data. So maybe quantum neural networks could be relevant in some of those situations.
It's a super-interesting subject, and I think that people are obsessively looking into it, and just dying to have the machine, because I think many of those things are going to only start to reveal themselves once we have big enough quantum machines that we can try some of those things on. Because again, a lot of those things are heuristic algorithms that we cannot really prove that will give us an advantage. We just have to build the machine and try it.
Daniel Whitenack:
Yeah, and I think that gets a little bit to my follow-up question, which is when these things start to come online - and I know that Quantum Machines is working on platform, software/hardware to do that, which we can get into here in a bit... But before we do that, one of the things that's on my mind is like as these computers come online, they're in some sort of like hybrid compute-center. I'm thinking just about my own workflows... How is my own workflow potentially going to change?
When I'm wanting to run something on a quantum computer -- if I just search Google for pictures of quantum computers, I'll see people with lab coats, or something on; they're going in and doing things, and... So do you see this as being like "Oh, I'm gonna have TensorFlow, I'm gonna have PyTorch, JAX, whatever, I have my Python code which I'm executing some type of AI training, or I've programmed a model architecture or something, and then at certain points there's gonna be a library that maybe reaches out to that hybrid quantum piece, and executes some of the optimization? Or how will that affect the actual day-to-day workflow and what it might be like to program these things? Because I'm guessing the common AI developer is not going to all of a sudden learn everything about quantum mechanics, right? So somehow there has to be a higher-level abstraction... What do you think that will look like?
Yonatan Cohen:
\[24:17\] Yeah, I actually think it will look exactly like what you described. I mean, it's going to take some time to get to that place, but I think that eventually that's exactly how it's going to roll... Because as you said, not everybody's going to learn these new types of quantum programming languages, and these fundamentally new, different operations in quantum, and how we use them. And in fact, I think this is even not necessary, because again, there will be certain subroutines that of course we can configure, and we can parameterize, and we can use them in various different ways... But this could be done -- as you said, it's a library that one could use, and include that in your workflows.
Now, that's going to take some time, and I think in the very early days actually we're going to have experts that are really going deep into the machine, into the low-level kind of programming of these quantum machines, and do all the optimizations to have a use case like that, where you can use it at a high level, just solving a problem, or accelerating your problem.
Chris Benson:
Just as a follow-up, as we're getting into tooling and kind of getting into what your organization does, but more of a general follow-up real quick... I'm curious - you mentioned IBM earlier, and they're well-known in the quantum space, for all that they do in the cloud, and everything... They do have an open source Python SDK for that, and it is in Python, which makes it very convenient for other AI tooling and stuff... So would you expect things like TensorFlow and PyTorch and all to kind of wrap that library or other libraries like it, to provide that kind of a quantum accessibility, if you will, to people who are not otherwise quantum experts, and kind of let the tooling make that?
Do you think as a general -- so I'm not talking about IBM or TensorFlow specifically, but as a general way of introducing the public to quantum is doing it through tooling that makes it easier kind of what we should expect going forward?
Yonatan Cohen:
Yes, absolutely. I really think so. Yeah, and it has several layers. So there are programming languages and abstractions, and then there are just application libraries that I think are going to be important. And then yes, even higher-level things, like TensorFlow and PyTorch - as you mentioned, they could just wrap those things, and some people don't even need to know that it runs under the hood.
So I think eventually that's what's going to happen. And yeah, people are starting to play with that, for sure. IBM, as you mentioned, they don't just have their open stack that you can access kind of low-level a quantum computer, and program the gates, but they also provide libraries, and there are many other startup companies that are doing that these days, sort of going up the stack... And I think that these tools are going to be very important.
I also think that we're going to discover a lot of things in the next 5, 6, 7 years, and I hope that many of those tools don't have to sort of make a U-turn, or something like that... But I think that we will have to learn a lot and change things as we go along. But that's a part of it; I mean, we have to start, we have to build the entire stack. That's why it's so exciting - we're waiting for the hardware to sort of mature, but we're building the low-level software parts of the stack, the high-level software parts of the stack, and kind of trying to see how all of it is going to fit together.
And yeah, I do think that at the end of the day there is no reason why someone would program quantum gates. Well, some people will, but most people will just want to use the machine for accelerating some of their problems.
Break:
\[28:21\]
Daniel Whitenack:
So Yonatan, I was just scrolling through your website and kind of with the view of what you've talked about in mind, in terms of this sort of hybrid system that you envision, how people will kind of program these problems... And I see that quantum machines specifically is addressing some of the hardware and software platform things around this type of system. I even see a nice little GIF image showing a data center, with racks of what I'm assuming is classical computers and racks of quantum machine equipment, and then like a quantum computer in the middle...
So I'm wondering if you could maybe generally let us know - as Quantum Machines looked at this developing space, how did you decide where you thought the opportunity was in terms of building out the infrastructure around this? Obviously, there's different pieces of this in terms of building any type of computer itself, or working only on software... But it seems like you're kind of dipping a little bit into hardware and software. Could you kind of describe your approach and the motivation behind that?
Yonatan Cohen:
Yeah, definitely. So if we look at the hardware of a quantum computer, it actually has two main parts. It has the QP itself, the quantum processor; that's the quantum hardware, that's where the magic happens, where you have the superpositions, and you have the qubits, and all this crazy quantum stuff... And then you have what we call the control hardware. This is actually not quantum hardware; it's classical hardware, but it's the interface, it's the hardware that interfaces the quantum processor, and talks to it, and operates it to make it do what we want it to do. And that's very complicated hardware that one has to build specifically; so it's not regular servers or anything like that, it's really hardware that's dedicated for controlling a quantum processor. And that's what a quantum machine does. That's what we started from. And this was because - well, we saw a bottleneck there.
So this was five years ago when we started thinking about those things, and thinking about starting a startup in quantum in general; it was just very early days where the quantum industry started, and we saw some of the early investments in companies in the U.S, and me and one of my co-founders, Itamar Sivan, who's CEO of the company, we wanted to start our own company, and we wanted to do it in quantum, because that's basically all we knew coming out of our PhD's in quantum devices... And we knew a guy, we had a friend who finished his Ph.D. about four years before us, and he left to do his post-doc at Yale University, in one of the leading groups in the world in quantum computing, the group of Professor Rob Schoelkopf. And over there, he basically performed one of the milestones in the field in the last (let's say) decade, where he actually performed an experiment demonstrating quantum error correction in superconducting qubits.
So quantum error corrections is one way, basically the mainstream way to deal with the fact that quantum computers are very noisy, they have a lot of errors; they just make errors all the time, the error rate is very high... And so the mainstream way to think that we'll be able to deal with this is by doing quantum error correction. But it's very hard.
So listen... Performed this first demonstration of quantum error correction in superconducting qubits in his post-doc at Yale, and to do that, you had to deal with a lot of bottlenecks that came from the control system.
\[32:20\] So he had to develop by himself a new kind of control system to do that experiment. Because quantum error correction is sort of what pushes the control layer of the stack to its limits, he had to deal with some of the most challenging problems of that time. And when we started thinking about what are we going to do, we realized that this challenge -- like, most people didn't hit those challenges yet, but they will, in the next few years.
So we felt we had a sort of headstart, and we wanted to be -- like, startups wanna be higher in the stack, right? You wanna do software and everything, but in quantum we felt that we were at the top of the stack, where actually there is a true need in the market right now... So that's what we did.
Chris Benson:
Let me ask you a question just to clarify, because I know it's central to your business. When you talk about control systems, can you talk a little bit about what exactly a control system does in the context of quantum, just to make sure that we understand exactly how that fits into the equation?
Yonatan Cohen:
Sure. So there are various ways to implement a quantum processor, but in 90% of them the qubits are sitting in some kind of an array of qubits; so they're physically stuck in space somewhere... On the chip, for example, or in a vacuum chamber... And then you have an array of qubits. So kind of thinking about it as -- yeah, let's just say an array of qubits. But then, in order to perform the logical operations, the quantum gates, you send signals from the outside in the form of pulses of electromagnetic waves. So just like your cell phone sends pulses of microwave signals to the cellular tower, our control system sends this orchestra of microwave RF signals to the quantum processor. And this is the quantum algorithm in its most kind of raw form. \[unintelligible 00:34:29.06\] quantum device, and it physically hits the qubit, and it performs the operations on the qubit. And you need to orchestrate this sequence of microwave pulses very, very carefully, it has to be very well-timed, and you also want to measure signals coming back from the quantum processor. So microwave or RF signals come back from the quantum processor and you need to measure those. And sometimes, for example if you wanna do quantum error correction, you need to measure those and you need to perform classical calculations to understand what else appeared on the chip, for instance, and then respond with new pulses. So that's what we call feedback in the control system.
Chris Benson:
And it sounds like - when you're talking about sending pulses and stuff, is that software and hardware in a quantum context that we're talking about there, or is it just one or the other, for our control system?
Yonatan Cohen:
Exactly. So you build the hardware, but then you need to program this sequence of pulses, right? So you build the hardware that can generate these pulses, but now you need to program your sequence; the user, for example, needs to program their sequence of passes that needs to go to the quantum processor and operate it. And if you want, this is the assembly language of a quantum computer. This is sort of the lowest-level programming language that you talk to your quantum computer with, because you tell the controller what pulse to send when, and this is really the lowest-level abstraction of operations.
Chris Benson:
\[36:02\] I appreciate the clarification. So going back for a moment to kind of having that -- like, as a practitioner, and I'm thinking about kind of my practical workflow, and I'm now integrating quantum computers and your control system for managing that into my workflow, in our context for doing AI... So let's say that we're a little ways in the future and we're starting to look at algorithms on the AI side that could benefit, where part of that is quantum - how does that workflow look like from the practitioner's standpoint? What should they expect? How are they connected, and such as that?
Yonatan Cohen:
So you mean like how would this workflow actually run on the hardware, or...?
Chris Benson:
Yeah, well, to some degree, I'm trying to think -- so I'm trying to kind of bring it back around from our side to connect the quantum computing benefit with the things that our audience is typically engaged in, which is trying to get AI algorithms, the models developed, and then deployed out there. And so as they're going through that process - I'm trying to kind of pool it all together now. What does that look like? Is there a quantum computer with the control system that you've produced sitting beside a classical computer that has a GPU in it, and a CPU in it, in the classical sense, and there's some networking between them? What does that look like if we're five years out? ...or whatever, you can pick your timeframe. But the point where we're now starting to integrate that into some sort of practical workflow, and people in our audience might be using it. What might that look like, from your position today, forecasting into the future?
Yonatan Cohen:
It's very strange, what you describe. There are two models, I would say. One, let's just imagine an on-prem system, where basically I have my GPU and my CPU, and then I have a quantum computer. And the way it looks - yeah, the control system, that's the interface. So we are the interface between the classical and the quantum side. So you have your racks of servers, with GPUs and CPUs and all that, and then you have some racks with the control electronics, and they're connected through the network. And the control electronics runs the programs on the quantum processor.
So we can just call that combination of the control electronics and the quantum processor - we can just call it the quantum accelerator. So now you have a quantum accelerator, you connect it to the network, and you can talk to it via, for example our programming language, QUA, which is this low-level, very low-level, pulse-level programming language in the quantum jargon. So it's a pulse-level programming language, and now I can write the workflow. So for example, I could use a workflow tool... Actually, we are developing such a tool for developing hybrid quantum-classical workflows we call it Entropy, but you can use others, and then you write a workflow where you run something on the GPU, and then something on the QPU, and then maybe there's communication between them, so you choose your favorite way to communicate between - whether it's processes, or functions that you call... And then this entire thing would just run programs on the CPU, and then a QUA program on the QPU, and the result would be analyzed maybe in the GPU, and so on and so forth.
\[39:33\] And again, eventually you would probably not have to write this quantum code at such a low level, and you could take advantage of libraries that by themselves would run hybrid workflows, because to solve a certain optimization problem maybe you already have a library that runs an optimization that uses both the CPU, the GPU and the QPU, and just solves for you a certain subroutine, and then you have an even higher-level workflow that would, I don't know, predict weather, or something of that sort. I don't know if this was clear or not...
Chris Benson:
No, that was a good one.
Daniel Whitenack:
That's great.
Chris Benson:
It tied it together for me, thank you.
Yonatan Cohen:
Okay, I'm happy I managed to do that.
Daniel Whitenack:
Yeah, yeah. And as we kind of get close to the end in here, I'm just looking through some of the things that you're involved with, or the companies involved with. It seems like there's a lot of exciting things kind of moving towards the future. I saw some of the press releases around kind of building -- involved in building Israel's National Quantum Computing Center, and other things... As you're kind of looking to your next year ahead, what are some of those things that are really exciting for you in terms of how the industry is shaping up, and what your company is involved with kind of moving into the next year or two?
Yonatan Cohen:
Yeah. So first and foremost, I'm just excited to be part of those people that are trying to build these computers, these machines that are based on deep, fundamental laws of nature, and hopefully help us compute faster, and also understand nature better. So that to me is the most exciting thing. I'm super-excited that so many people are using our products to make this field progress. So I'm just excited for people to use more and more our products, and improve them, so that we can move faster. And this is our mission, to accelerate the realization of useful quantum computers. Hopefully, we're building tools that allow people to accelerate the realization of those computers and make them useful.
Other than that, I'm super-excited to see how the industry shapes, because this is a field that's in kind of like formation. The stack, the entire structure of the technology stack, but also the value chain. The community is sort of defining its value chain right now, and defining itself, and it's very exciting to be in an industry in those early days... So I'm super-excited for that.
Daniel Whitenack:
Well, thank you so much, Yonatan, for joining us. This is really, really interesting. It's awesome to see these practicalities of the integration layer and the hybrid systems coming together around real scalable hardware and platforms... So yeah, thank you for all the work that you're doing, and for taking time to tell us a little bit about it.
Yonatan Cohen:
Thank you so much for hosting me. This was great.
Daniel Whitenack:
Welcome to another episode of Practical AI. This is Daniel Whitenack, I'm a data scientist with SIL International, and I'm joined as always by my co-host, Chris Benson, who is a tech strategist at Lockheed Martin. How are you doing, Chris?
Chris Benson:
Doing well, doing well. How are you today, Daniel?
Daniel Whitenack:
I am doing great, because any day is great when I get to -- any sort of throwback to my old physics days, which mostly I don't get to dabble in much these days, is a fun day... And today we've got a cool intersection with that world. We've got Yonatan Cohen with us, CTO at Quantum Machines... And we're gonna talk a little bit about an update on quantum computing and how that intersects with AI. Welcome, Yonatan.
Yonatan Cohen:
Hi. Thank you. It's really cool to be here. I'm excited to talk to you about some quantum physics as well.
Daniel Whitenack:
Yeah, great. Great. Well, I'm guessing most of the time on this show, and most of our listeners are probably used to hearing us chat about neural networks, or GPUs, or classical computing, and we have talked about quantum computing on the show, but it's been quite some time, and I'm sure that the field is advancing quickly. Maybe just as a starting point, could you remind us of the general pitch of like what is quantum computing, and maybe why people are interested in quantum computing more generally?
Yonatan Cohen:
Sure. So quantum computing is a new way to build a computer, that's based on the laws of quantum mechanics. It's kind of interesting, because quantum mechanics as a theory of nature was developing parallel to the development of computers in the last hundred years or so. And on one hand, we've developed this amazing technology that we now have, which is computing, that's based on classical physics; it's based on the classical laws of how the Universe behaves. But in parallel, in the last hundred years or so, there is a completely new understanding of how nature works on a very fundamental level.
\[03:54\] Somewhere in the very late '70s, early '80s physicists started to understand that perhaps we can use these new laws of nature we've discovered to also build a new type of computer, that's going to harvest this weird behavior of nature called quantum mechanics. And it turns out that you can do that, and it expands the notion of what we mean by a computer, and essentially allows to build a computer that has stronger computational power, at least for some problems; not for all problems, but for some very difficult problems with classical computers - when we say classical, meaning not quantum, but regular computers. For some computational problems that classical computers have a very hard time to deal with, quantum computers could solve them very easily.
Chris Benson:
So as you're talking about this new type of computer, would I be correct in thinking that I can't run out to the store a Pentium or an AMD, and pop it in with the RAM on the motherboard? What's different about that approach on the hardware side, and then what would you use it for that's different from that kind of classic idea that we all have been using all our lives?
Yonatan Cohen:
Yeah, so the main point of quantum mechanics and before quantum computers is that while we see things in the day-to-day that things are in a specific state - you know, I'm drinking coffee from my cup, and I put it down, and it sits in one place, it sits in one place and not in another place... So in quantum mechanics, things can actually be in various states at the same time. And while we don't see it with coffee cups, we do see it with electrons, for example. We can actually put an electron in what we call a super-position. That means it's in two places at the same time. And this is exactly what quantum computers take advantage of.
So the basic building block of a classical computer is the bit of information; it's a system that has two states - it can be either in the zero state, or in the one state. And then if you have two bits, then they can be 0/0, 0/1, 1/0, 1/1, and so on. If I have 8 bits, I have 256 states, right? 0/0/0/0/0/0, and 0/0/0/0/0/1, and so on. But at every single point in time, the classical computer can only be in a single state of all of the bits of information that in holds. And then we manipulate those bits; we go from one state, to another state, to another state, and so on. So it's like a state machine. We're moving from state to state, in order to solve a computational problem.
But quantum computers replace this notion of a bit with what we call a quantum bit, or in short qubit, which is a system like the electron that I told you about, that can actually be in two states at the same time. So it can be in both 0 and 1 at the same time. In fact, the system can mean 0 and 1 with different weights. So it can be a little bit 0, and a lot in 1, or it can be a lot in 0 and a little bit in 1, and so on. And when now I put a lot of qubits together, I now can be in this massive super-position of many states, all the states of the system at the same time, and you can use that to -- you can use this parallelism to, in some cases, do parallel computation, instead of going state by state, if that makes sense.
Chris Benson:
So with that massive parallelism, it sounds like it can handle probabilities, based on what you were saying, in many states, very well. What type of problems does that kind of capability lend itself to, that maybe our traditional - you know, the laptop I'm on right now - would not be as suited for? What is a problem that having lots of qubits to solve - what does it look like to address?
Yonatan Cohen:
\[08:09\] Yeah, it's a great question. So there are certain problems that we know for sure that quantum computers can get an advantage of. One great example is the Shor's algorithm. So that's an algorithm that actually breaks code, because it finds the prime factors of a large number. So if you take a large number and try to factorize it to its basic prime factors, that's a problem that's very hard for classical computers.
And the problem has some structure, it allows us to use this parallelism of a quantum computer in order to solve this problem exponentially faster than what we can do today in a classical computer. Now, it's very hard to explain exactly what in this problem makes it so that we can use this quantum parallelism to solve it so much faster. You're gonna have to look at the details of the problem, the structure of the problem in order to see how you can take advantage of this quantum parallelism.
So we don't know exactly, we cannot actually categorize exactly all the problems that quantum computers will solve much faster than classical computers, but we have examples. So we have Short's algorithm, we have Grover's algorithm... Grover's algorithm is an algorithm that solves a search problem. Searching in an unsorted list, finding a specific element of interest in an unsorted list, which also can map to very general optimization problems... But this algorithm, for instance, is not giving us an exponential speed-up. It actually only uses a square root speed-up over what you could do with classical computers.
So there is this zoo of algorithms that gives different kind of speed-ups for different kind of problems, and people are still working very hard on the problem of categorizing exactly what quantum computers could solve faster than the classical computers.
The other thing is that there are also quantum algorithms that we have today that use this parallelism of these qubits, and that we don't even know that they are going to work. So these are heuristics, basically, that people have come up with. There are good reasons to think why they would give us a computational speed-up against classical computers, but nobody can prove those. So we are basically in this kind of sub-situation that actually maybe AI was a decade or two ago, where we have some thoughts about these algorithms, people are very hopeful that they will work, and these are relevant for optimization problems, very general optimization problems... There's an algorithm that's called QAOA, that solves combinatorial optimization problems, and then another algorithm that's called VQE, which solves for chemistry problems...
These are heuristic algorithms that we just have to build a machine and try those algorithms on, and see if they work better than classical computers.
Daniel Whitenack:
Yeah, I have one maybe naive question from my own perspective, because I think -- obviously, you're well plugged into the state of the art. Chris, actually, I think you know a little bit more as well... But for me, I'm thinking -- okay, I've been hearing about quantum computers for some time... One sort of naive question is what actual quantum computers exist in the world right now?
Because if I go to one of the cloud providers, sometimes I'll see like "Oh, here's the AWS quantum thing", but when I look at it, it's really just so you can run simulations of quantum research.
\[11:58\] So let's say I want to run something on an actual quantum computer. What would be my choices right now, in terms of where those actually exist, and kind of the state of what they are, and how you see that sort of shaping up over the coming couple of years.
Yonatan Cohen:
Yeah, so today you can access quantum computers through the cloud with several players. So IBM, for example, has a quantum cloud service that actually gives you access to great quantum computers. You can log in and you can actually run algorithms... But these are small quantum computers. Of course, they're still a prototype of the technology. But you can actually run algorithms and you can see how they behave, as well as on simulators. You can compare, and then you can find that your quantum computer is doing not as well as the simulation, because it has a lot of errors; this is the main problem with quantum computers today... But you can do that, and that's actually very cool.
IBM, for example, they give access to their own computers, but Azure and AWS have launched their quantum services in the last few years that give you access to third-party computers... So many different computers that are built by full-stack quantum computing vendors, like IQM, QCI... These are startups in quantum computing that build full-stack quantum computers, and you can access those machines actually today through the cloud, which is very cool.
And they also give you access to simulators, and you can run on those as well... Which I think is quite amazing, to be honest. I mean, there are certain experiments that I used to do during my Ph.D. at Weizmann Institute only 12 years ago that... It used to take a Ph.D. student maybe three or four years to set up such an experiment and to run it, and now I can just log into one of these cloud platforms and I could run the experiment that just ten years ago it used to take a few years to set up, I can do it probably in a few hours... And that I think is quite amazing.
Break:
\[14:16\]
Chris Benson:
So one of the things I was curious about just to follow up on what you were talking about a moment ago is just to kind of set my expectation... With quantum computers, assuming that obviously because we're going quantum, we're not going out and getting a typical CPU, is there anything -- is it mainly the CPU is a quantum CPU that you would put into the computer, and is there any other types of changes that you would make to a classical computer, just to get a sense of it? How close is the generalized architecture of a quantum computer to a classical computer? How much has to change to make that leap?
Yonatan Cohen:
Okay, that's a very good question. I think that a lot of things are very different in quantum computers, because the basic operations are fundamentally different. Classical computers were built based on bits, and then you have gates, It's how you manipulate the information. So you can actually build an entire computer from a NAND gate. But in quantum we don't even have the notion of a NANG gate. We have other gates. We have a Hadamard gate, we have a CCNOT gate, we have different elementary logical operations on qubits. So that makes it so that the entire stack, or at least the low-level parts of the stack need to be very different.
\[16:17\] But then I would also say that eventually it's not -- I mean, the way I see quantum computers, especially in the early days bringing us an advantage is not just by being used as standalone computers, but actually as more of an accelerator inside a more heterogeneous data center.
So the quantum computer could do only certain problems. It's not going to replace the CPU, and it's not going to replace the GPU. There are things that there's just no reason to do them using quantum devices. Maybe in 50 years, because everything is going to be quantum, and quantum devices are going to be as cheap as classical computing devices... But right now, you'll have your CPU, you'll have your GPU and you'll have your QPU, and the QPU will be used to accelerate certain types of sub-routines that can help your entire application. So that's why it's important that we also build integrations of these quantum computers into a more heterogeneous compute environment, and allow people to program what we call hybrid workflows, so quantum/classical workflows.
Chris Benson:
Could you describe what one of those workflows is like? And that was a great explanation, by the way, of the difference in a quantum computer... But you've mentioned several times along the way about that hybrid fitting it into a larger architecture that includes a lot of those classical components... But I'm trying to get my head wrapped around what kind of problem as a user I might solve, an example that would use both quantum and the hybrid, and what part of the problem goes to each. Can you give us some sort of example on that?
Yonatan Cohen:
Yeah, sure. So the typical example is what we call the variational quantum algorithms. Actually, both the algorithms I mentioned before, QAOA and also VQE, these are variational quantum algorithms. What this means is that essentially, what we're doing is we're trying to minimize the cost function of interest. That's the problem we're trying to solve. But the way we do it is that the quantum processor is only computing the cost function.
Let's say we have a cost function that's very hard to compute on a classical computer... So the quantum processor only basically calculates the cost function given a set of parameters, but then the optimizer sits on the classical side. So what we're doing is we're running quantum circuits, what we call parameterize quantum circuits, or we're running a quantum program with some input parameters, and then the quantum processor computes a cost function, and then the result goes into the classical processor, and it would run an optimizer, like a gradient descent, or something like that. Then that will generate new parameters for the quantum circuit, the quantum program. And then it goes back and forth like this until we find the parameters that would minimize the cost function.
Daniel Whitenack:
Yeah, this is a really interesting example, because I've been thinking about like - well, okay, if I'm connecting maybe some of the challenges within the AI world to this space... I mean, you've just mentioned gradient descent, you've mentioned optimizing with a cost function... Like, this is all definitely very connected to what might happen in like an AI training scenario. Of course, there's all sorts of problems, or there's certain computations, like you said, that work well on a GPU, and that's sort of scaled up right now... But I'm wondering, as you're kind of diving into these types of hybrid problems with quantum machines, what might you see right now in the AI industry, or certain sets of problems in the AI industry, like the challenges that the AI industry is facing in terms of compute - what are people thinking about in terms of what might overcome some of those challenges on this sort of hybrid computing side with potentially a quantum advantage? What's the current thought process around that?
Yonatan Cohen:
\[20:35\] Yeah, so I'd hope that, for instance, using some of these hybrid algorithms we'd be able to solve some optimization problems that are also relevant to AI training. To be honest, I'm far from being an expert on quantum AI, or quantum neural networks specifically, but I know a little bit about the subject and a little bit about the challenges in AI, and I know that some of the algorithms that we have for quantum computers could solve these optimization problems.
So when we train a neural network and wanna optimize its parameters, in many cases this is a very harsh optimization problem. In some cases it could fill exactly these kind of optimization problems that a quantum computer could solve, and specifically these hybrid algorithms might be able to solve. So then we could use these hybrid quantum-classical algorithms to train neural networks. That's one example.
There's also some recent works about quantum neural network where the network itself is quantum, and some mathematicals proofs why such networks would require... Again, in some specific cases, but in both of the cases why they'd require much less data to train. And that, I think, is very important, because in many cases we just need a lot of data to train those neural networks, and in many cases we don't have enough data. So maybe quantum neural networks could be relevant in some of those situations.
It's a super-interesting subject, and I think that people are obsessively looking into it, and just dying to have the machine, because I think many of those things are going to only start to reveal themselves once we have big enough quantum machines that we can try some of those things on. Because again, a lot of those things are heuristic algorithms that we cannot really prove that will give us an advantage. We just have to build the machine and try it.
Daniel Whitenack:
Yeah, and I think that gets a little bit to my follow-up question, which is when these things start to come online - and I know that Quantum Machines is working on platform, software/hardware to do that, which we can get into here in a bit... But before we do that, one of the things that's on my mind is like as these computers come online, they're in some sort of like hybrid compute-center. I'm thinking just about my own workflows... How is my own workflow potentially going to change?
When I'm wanting to run something on a quantum computer -- if I just search Google for pictures of quantum computers, I'll see people with lab coats, or something on; they're going in and doing things, and... So do you see this as being like "Oh, I'm gonna have TensorFlow, I'm gonna have PyTorch, JAX, whatever, I have my Python code which I'm executing some type of AI training, or I've programmed a model architecture or something, and then at certain points there's gonna be a library that maybe reaches out to that hybrid quantum piece, and executes some of the optimization? Or how will that affect the actual day-to-day workflow and what it might be like to program these things? Because I'm guessing the common AI developer is not going to all of a sudden learn everything about quantum mechanics, right? So somehow there has to be a higher-level abstraction... What do you think that will look like?
Yonatan Cohen:
\[24:17\] Yeah, I actually think it will look exactly like what you described. I mean, it's going to take some time to get to that place, but I think that eventually that's exactly how it's going to roll... Because as you said, not everybody's going to learn these new types of quantum programming languages, and these fundamentally new, different operations in quantum, and how we use them. And in fact, I think this is even not necessary, because again, there will be certain subroutines that of course we can configure, and we can parameterize, and we can use them in various different ways... But this could be done -- as you said, it's a library that one could use, and include that in your workflows.
Now, that's going to take some time, and I think in the very early days actually we're going to have experts that are really going deep into the machine, into the low-level kind of programming of these quantum machines, and do all the optimizations to have a use case like that, where you can use it at a high level, just solving a problem, or accelerating your problem.
Chris Benson:
Just as a follow-up, as we're getting into tooling and kind of getting into what your organization does, but more of a general follow-up real quick... I'm curious - you mentioned IBM earlier, and they're well-known in the quantum space, for all that they do in the cloud, and everything... They do have an open source Python SDK for that, and it is in Python, which makes it very convenient for other AI tooling and stuff... So would you expect things like TensorFlow and PyTorch and all to kind of wrap that library or other libraries like it, to provide that kind of a quantum accessibility, if you will, to people who are not otherwise quantum experts, and kind of let the tooling make that?
Do you think as a general -- so I'm not talking about IBM or TensorFlow specifically, but as a general way of introducing the public to quantum is doing it through tooling that makes it easier kind of what we should expect going forward?
Yonatan Cohen:
Yes, absolutely. I really think so. Yeah, and it has several layers. So there are programming languages and abstractions, and then there are just application libraries that I think are going to be important. And then yes, even higher-level things, like TensorFlow and PyTorch - as you mentioned, they could just wrap those things, and some people don't even need to know that it runs under the hood.
So I think eventually that's what's going to happen. And yeah, people are starting to play with that, for sure. IBM, as you mentioned, they don't just have their open stack that you can access kind of low-level a quantum computer, and program the gates, but they also provide libraries, and there are many other startup companies that are doing that these days, sort of going up the stack... And I think that these tools are going to be very important.
I also think that we're going to discover a lot of things in the next 5, 6, 7 years, and I hope that many of those tools don't have to sort of make a U-turn, or something like that... But I think that we will have to learn a lot and change things as we go along. But that's a part of it; I mean, we have to start, we have to build the entire stack. That's why it's so exciting - we're waiting for the hardware to sort of mature, but we're building the low-level software parts of the stack, the high-level software parts of the stack, and kind of trying to see how all of it is going to fit together.
And yeah, I do think that at the end of the day there is no reason why someone would program quantum gates. Well, some people will, but most people will just want to use the machine for accelerating some of their problems.
Break:
\[28:21\]
Daniel Whitenack:
So Yonatan, I was just scrolling through your website and kind of with the view of what you've talked about in mind, in terms of this sort of hybrid system that you envision, how people will kind of program these problems... And I see that quantum machines specifically is addressing some of the hardware and software platform things around this type of system. I even see a nice little GIF image showing a data center, with racks of what I'm assuming is classical computers and racks of quantum machine equipment, and then like a quantum computer in the middle...
So I'm wondering if you could maybe generally let us know - as Quantum Machines looked at this developing space, how did you decide where you thought the opportunity was in terms of building out the infrastructure around this? Obviously, there's different pieces of this in terms of building any type of computer itself, or working only on software... But it seems like you're kind of dipping a little bit into hardware and software. Could you kind of describe your approach and the motivation behind that?
Yonatan Cohen:
Yeah, definitely. So if we look at the hardware of a quantum computer, it actually has two main parts. It has the QP itself, the quantum processor; that's the quantum hardware, that's where the magic happens, where you have the superpositions, and you have the qubits, and all this crazy quantum stuff... And then you have what we call the control hardware. This is actually not quantum hardware; it's classical hardware, but it's the interface, it's the hardware that interfaces the quantum processor, and talks to it, and operates it to make it do what we want it to do. And that's very complicated hardware that one has to build specifically; so it's not regular servers or anything like that, it's really hardware that's dedicated for controlling a quantum processor. And that's what a quantum machine does. That's what we started from. And this was because - well, we saw a bottleneck there.
So this was five years ago when we started thinking about those things, and thinking about starting a startup in quantum in general; it was just very early days where the quantum industry started, and we saw some of the early investments in companies in the U.S, and me and one of my co-founders, Itamar Sivan, who's CEO of the company, we wanted to start our own company, and we wanted to do it in quantum, because that's basically all we knew coming out of our PhD's in quantum devices... And we knew a guy, we had a friend who finished his Ph.D. about four years before us, and he left to do his post-doc at Yale University, in one of the leading groups in the world in quantum computing, the group of Professor Rob Schoelkopf. And over there, he basically performed one of the milestones in the field in the last (let's say) decade, where he actually performed an experiment demonstrating quantum error correction in superconducting qubits.
So quantum error corrections is one way, basically the mainstream way to deal with the fact that quantum computers are very noisy, they have a lot of errors; they just make errors all the time, the error rate is very high... And so the mainstream way to think that we'll be able to deal with this is by doing quantum error correction. But it's very hard.
So listen... Performed this first demonstration of quantum error correction in superconducting qubits in his post-doc at Yale, and to do that, you had to deal with a lot of bottlenecks that came from the control system.
\[32:20\] So he had to develop by himself a new kind of control system to do that experiment. Because quantum error correction is sort of what pushes the control layer of the stack to its limits, he had to deal with some of the most challenging problems of that time. And when we started thinking about what are we going to do, we realized that this challenge -- like, most people didn't hit those challenges yet, but they will, in the next few years.
So we felt we had a sort of headstart, and we wanted to be -- like, startups wanna be higher in the stack, right? You wanna do software and everything, but in quantum we felt that we were at the top of the stack, where actually there is a true need in the market right now... So that's what we did.
Chris Benson:
Let me ask you a question just to clarify, because I know it's central to your business. When you talk about control systems, can you talk a little bit about what exactly a control system does in the context of quantum, just to make sure that we understand exactly how that fits into the equation?
Yonatan Cohen:
Sure. So there are various ways to implement a quantum processor, but in 90% of them the qubits are sitting in some kind of an array of qubits; so they're physically stuck in space somewhere... On the chip, for example, or in a vacuum chamber... And then you have an array of qubits. So kind of thinking about it as -- yeah, let's just say an array of qubits. But then, in order to perform the logical operations, the quantum gates, you send signals from the outside in the form of pulses of electromagnetic waves. So just like your cell phone sends pulses of microwave signals to the cellular tower, our control system sends this orchestra of microwave RF signals to the quantum processor. And this is the quantum algorithm in its most kind of raw form. \[unintelligible 00:34:29.06\] quantum device, and it physically hits the qubit, and it performs the operations on the qubit. And you need to orchestrate this sequence of microwave pulses very, very carefully, it has to be very well-timed, and you also want to measure signals coming back from the quantum processor. So microwave or RF signals come back from the quantum processor and you need to measure those. And sometimes, for example if you wanna do quantum error correction, you need to measure those and you need to perform classical calculations to understand what else appeared on the chip, for instance, and then respond with new pulses. So that's what we call feedback in the control system.
Chris Benson:
And it sounds like - when you're talking about sending pulses and stuff, is that software and hardware in a quantum context that we're talking about there, or is it just one or the other, for our control system?
Yonatan Cohen:
Exactly. So you build the hardware, but then you need to program this sequence of pulses, right? So you build the hardware that can generate these pulses, but now you need to program your sequence; the user, for example, needs to program their sequence of passes that needs to go to the quantum processor and operate it. And if you want, this is the assembly language of a quantum computer. This is sort of the lowest-level programming language that you talk to your quantum computer with, because you tell the controller what pulse to send when, and this is really the lowest-level abstraction of operations.
Chris Benson:
\[36:02\] I appreciate the clarification. So going back for a moment to kind of having that -- like, as a practitioner, and I'm thinking about kind of my practical workflow, and I'm now integrating quantum computers and your control system for managing that into my workflow, in our context for doing AI... So let's say that we're a little ways in the future and we're starting to look at algorithms on the AI side that could benefit, where part of that is quantum - how does that workflow look like from the practitioner's standpoint? What should they expect? How are they connected, and such as that?
Yonatan Cohen:
So you mean like how would this workflow actually run on the hardware, or...?
Chris Benson:
Yeah, well, to some degree, I'm trying to think -- so I'm trying to kind of bring it back around from our side to connect the quantum computing benefit with the things that our audience is typically engaged in, which is trying to get AI algorithms, the models developed, and then deployed out there. And so as they're going through that process - I'm trying to kind of pool it all together now. What does that look like? Is there a quantum computer with the control system that you've produced sitting beside a classical computer that has a GPU in it, and a CPU in it, in the classical sense, and there's some networking between them? What does that look like if we're five years out? ...or whatever, you can pick your timeframe. But the point where we're now starting to integrate that into some sort of practical workflow, and people in our audience might be using it. What might that look like, from your position today, forecasting into the future?
Yonatan Cohen:
It's very strange, what you describe. There are two models, I would say. One, let's just imagine an on-prem system, where basically I have my GPU and my CPU, and then I have a quantum computer. And the way it looks - yeah, the control system, that's the interface. So we are the interface between the classical and the quantum side. So you have your racks of servers, with GPUs and CPUs and all that, and then you have some racks with the control electronics, and they're connected through the network. And the control electronics runs the programs on the quantum processor.
So we can just call that combination of the control electronics and the quantum processor - we can just call it the quantum accelerator. So now you have a quantum accelerator, you connect it to the network, and you can talk to it via, for example our programming language, QUA, which is this low-level, very low-level, pulse-level programming language in the quantum jargon. So it's a pulse-level programming language, and now I can write the workflow. So for example, I could use a workflow tool... Actually, we are developing such a tool for developing hybrid quantum-classical workflows we call it Entropy, but you can use others, and then you write a workflow where you run something on the GPU, and then something on the QPU, and then maybe there's communication between them, so you choose your favorite way to communicate between - whether it's processes, or functions that you call... And then this entire thing would just run programs on the CPU, and then a QUA program on the QPU, and the result would be analyzed maybe in the GPU, and so on and so forth.
\[39:33\] And again, eventually you would probably not have to write this quantum code at such a low level, and you could take advantage of libraries that by themselves would run hybrid workflows, because to solve a certain optimization problem maybe you already have a library that runs an optimization that uses both the CPU, the GPU and the QPU, and just solves for you a certain subroutine, and then you have an even higher-level workflow that would, I don't know, predict weather, or something of that sort. I don't know if this was clear or not...
Chris Benson:
No, that was a good one.
Daniel Whitenack:
That's great.
Chris Benson:
It tied it together for me, thank you.
Yonatan Cohen:
Okay, I'm happy I managed to do that.
Daniel Whitenack:
Yeah, yeah. And as we kind of get close to the end in here, I'm just looking through some of the things that you're involved with, or the companies involved with. It seems like there's a lot of exciting things kind of moving towards the future. I saw some of the press releases around kind of building -- involved in building Israel's National Quantum Computing Center, and other things... As you're kind of looking to your next year ahead, what are some of those things that are really exciting for you in terms of how the industry is shaping up, and what your company is involved with kind of moving into the next year or two?
Yonatan Cohen:
Yeah. So first and foremost, I'm just excited to be part of those people that are trying to build these computers, these machines that are based on deep, fundamental laws of nature, and hopefully help us compute faster, and also understand nature better. So that to me is the most exciting thing. I'm super-excited that so many people are using our products to make this field progress. So I'm just excited for people to use more and more our products, and improve them, so that we can move faster. And this is our mission, to accelerate the realization of useful quantum computers. Hopefully, we're building tools that allow people to accelerate the realization of those computers and make them useful.
Other than that, I'm super-excited to see how the industry shapes, because this is a field that's in kind of like formation. The stack, the entire structure of the technology stack, but also the value chain. The community is sort of defining its value chain right now, and defining itself, and it's very exciting to be in an industry in those early days... So I'm super-excited for that.
Daniel Whitenack:
Well, thank you so much, Yonatan, for joining us. This is really, really interesting. It's awesome to see these practicalities of the integration layer and the hybrid systems coming together around real scalable hardware and platforms... So yeah, thank you for all the work that you're doing, and for taking time to tell us a little bit about it.
Yonatan Cohen:
Thank you so much for hosting me. This was great.
