Simon Benjamin on Architectures for Quantum Computing
This episode is more technical than last week’s episode with John Preskill. We start by covering some fundamentals then go into different approaches for constructing and scaling a quantum computer.
Craig Cannon [00:00:00] – Hey, how’s it going? This is Craig Cannon and you’re listening to Y Combinator’s podcast. Today’s episode is with Simon Benjamin and like last week’s, it’s also about quantum computing. Simon’s a professor of quantum technologies at Oxford. He’s also the Principal Investigator for Oxford’s project on Quantum Optimization and Machine Learning. In this episode, we go a little bit more technical. We start by covering some fundamentals and then go into different approaches for constructing and scaling the quantum computer. All right, here we go. Simon, why in the past few years has quantum computing gotten so much attention?
Simon Benjamin [00:00:34] – Quantum computing is something that academics have been working on now for decades. What’s exciting is that it’s all starting to work in the sense that what we now have in labs is getting to the regime where it can actually do stuff that we can’t do by other means. It’s got that feeling that it’s all about to happen. That said, you never know how far it is from where you are to when you’ve actually got a working machine, until you’ve actually got the working machine. But we’re all feeling very excited. In the field, we’re excited as well, but then also people outside the field, companies, people who look for disruptive changes in technology are obviously receptive to this, and they’re starting to get excited as well. It’s a kind of feedback loop. I can say that just in the last three years, we’ve had tons more interest from companies who come to us as academics and say, look, can we jointly work on something with you? That’s really been a sea change. If we went up to companies even five years ago and said, “How about investing in some research?” They would say, “Well, we’re aware of this, but it’s over our horizon.”
Simon Benjamin [00:01:37] – Now we’re almost having to fight them off.
Craig Cannon [00:01:41] – Was there a certain basic research or maybe an experiment that made people realize, “Oh, this might be the time?”
Simon Benjamin [00:01:48] – Yeah, there is that. I’m a theorist. I probably should put that upfront. They don’t let me in the lab, except when they’ve tidied everything away, because my elbows tend to set them back a couple of weeks. But in the lab, researchers have been trying to get better and better control of quantum systems. They’re difficult to control. You’re probably familiar with the word qubit. Which is quantum bit, which is, of course, the building block of quantum technology instead of bits. But unlike bits, which are wonderfully stable and, you know, with the right technology will just sit in the zero or one for a long period of time, even years, qubits are horribly unstable. They’re constantly trying to just become regular bits, which means they, the usual word is collapse to one option or the other. We talk about a quantum superposition, a state that’s zero and one at the same time, and we can talk a bit more about what that means.
Simon Benjamin [00:02:44] – But the point is, it’s fragile. Qubits don’t, in a sense, want to be in zero and one at the same time, or rather, their environment doesn’t want them to. They are constantly trying to collapse. They’re constantly going wrong. Also, when we try and control them, it’s very difficult to control them precisely. It’s all about how good is your control of the quantum world, and I can’t resist at this point just mentioning, we’re here in Oxford, the Oxford ion trappers, we might talk more about what that is, have the joint world record for the highest level of control of any quantum system of any kind. The really tough thing is to say, how well can you control two qubits? If you think about it, if you want to make a computer, it’s not enough to just control individual entities, because if they don’t talk to each other, you can’t have a process, an algorithm. All you’ve got is some kind of recording system, right? If you could go and set the states of a bunch of individual qubits and then later come back and have a look at them, you’ve got a memory.
Simon Benjamin [00:03:47] – But if you can make one qubit control another qubit, you’ve got an operation, an element of an algorithm. That’s the, really, that’s the hardest thing, almost always, because it’s the most fancy thing. Now, we use the word fidelity, which means how well are you doing in terms of getting your control, and the Oxford guys, their record is 99.9, a little bit, percent fidelity. One part in a thousand.
Craig Cannon [00:04:12] – Fidelity over what period of time?
Simon Benjamin [00:04:14] – Yeah, right. That’s the fidelity that it takes to do one specific operation. So, for example, I’ve got two qubits. I always use my fingers, because I think of them as little arrows. I have two qubits. I tell this guy to flip only if this guy is, let’s say, in state one. But not if it’s in state zero. Okay, so that’s a story about two qubits, which should just happen. Ideally, you tell them to do that, they do it, exactly like you told them to, and you move on. In practice, they won’t do exactly what you told them to, because these things are very hard to control. 99.9% is the chances, essentially, that they’ll do what they’re told, and if not, then they’ll do something random. You can think of it that way.
Simon Benjamin [00:04:56] – Now, that sounds like a lot, right? 99.9% seems like a good score in any context. Compared to conventional technologies where that number might be something like one minus 10 to the minus 15, you know, it’s insanely good in conventional technologies. That’s why your software doesn’t just crash all the time on your laptop, right. Whilst that number sounds good, it’s actually still very challenging. Imagine you’re building an algorithm. You’re going to have to have a lot of these operations. A lot of times one qubit will talk to another qubit in the course of doing some calculation. If one in 1,000 times it just goes wrong and yields nonsense, then if you’ve got more than a thousand or so operations, and you will have, then something’s going to go wrong. You’re going to get an error in your quantum program, and unless you’ve been very smart about how you design it, that’s just going to make the answer come out nonsense. That’s the kind of problem we’ve faced for all these years, and when I say we, I mean the actual guys in the lab, right? I’ve just been at the whiteboard saying,
Simon Benjamin [00:05:55] – “Come on, guys, why isn’t it working yet?”
Craig Cannon [00:05:57] – To briefly tangent on this. How long have you been theorizing about quantum computers?
Simon Benjamin [00:06:03] – Well, the field has been theorizing in detail about quantum computers since the early or mid 80s. I’m a little bit more recent than that.
Craig Cannon [00:06:11] – It was David Deutsch in the beginning sort of…
Simon Benjamin [00:06:14] – Right. And then in the 90s, we, initially, people thought, this problem of control was going to be perhaps a deal breaker. The thing is if you’re running a quantum algorithm, okay, so, if you’re running a conventional algorithm, you can at any time check the state of your machine. You can say, I’m going to have a look at my ones and zeroes, see if they look about right, yeah? If you do that with a quantum computer, because of this idea of superposition, being in many states at once, if it’s not yet ready to give you the readout answer, but you just want to check to see if it looks like it’s on track or if it’s gone wrong, you’re going to destroy the state of the computer prematurely. The act of looking for errors, naively, it would seem, that it’s going to actually destroy the state of the machine. How can I keep on track if I’m not allowed to see? It’s a bit like trying to navigate with a blindfold, and you’ve been told a bunch of landmarks, but you can’t see them. It’s not allowed until you get to the end. People thought, including some really substantial thinkers like Rolf Landauer, thought that that might be a deal-breaker. The fact that you can’t check for errors but they are going to be happening would mean that for anything except the simplest toy calculation, you can’t really get where you’re going to. And in the 90s was when people, including Andy Steane, who’s here, but also Peter Shor, who’s famous for one of the big algorithms, figured out the solution to this,
Simon Benjamin [00:07:42] – which, it sounds almost like a philosophical paradox, right? I want to be able to check and fix errors as I run a process, complicated, long process, but if I look for errors, or if I ask the information whether it’s in a good state, I will destroy it, right? That seems like a… So the answer is, what you do is, you use several qubits, even a large number of them, to store one what we call logical qubit. This is the same thing that, well, it’s a grander version of the same thing that happens in conventional technology. There, if you know you’re going to store a lot of information and there may be some errors, because you’re storing, you know, vast amounts, you may use an error-correcting or an error-detecting code, which means you dedicate some of your bits to being watchdogs on the others. Now, the quantum version of that is much tougher to work out, but it’s the same basic idea. We use a bunch of physical qubits, which might be individual atoms or superconducting loops, there’s many different exciting hardware options, but whatever they are, you use a bunch of them to store one logical qubit.
Simon Benjamin [00:08:47] – You do it in such a way that if one of those physical qubits goes wrong, you will be able to fix that one guy by using the information that’s stored in the others, essentially. You’ve spread out your burden of representing the qubit over a bunch of guys, and collectively, they’re more robust than they would be individually. But you’re still not allowed to just look at them directly, so the real trick was to say, well, look. We’ll have another bunch of qubits. We’ll bring them in every so often. They’ll just be in some, say, zero state. They have nothing.
Craig Cannon [00:09:20] – So they come in neutral.
Simon Benjamin [00:09:21] – They come in neutral. Then we do a little, we stop the main calculation, if you want to think of it this way. We pause the main computation.
Craig Cannon [00:09:28] – Okay, without observing.
Simon Benjamin [00:09:29] – Without observing, crucial. Then we bring in these extra guys, and we have a little special calculation, really, a special computation, which is just to look for errors. A computation takes place between the new guys, who are just in a reference state, and the guys you actually want to check. What you basically do is a calculation where the question, is there an error or not, is now stored on the extra guys, and they know nothing else. They don’t know what the main information is. They just know, has there been an error. That’s the trick. Now, you can actually have a look at the extra guys. Hopefully, they’ll say there’s no error, good to go. If they say there was an error, they’ll also tell you which guy was the problem, and you can fix that guy without looking at him, okay.
Craig Cannon [00:10:14] – Through the algorithm?
Simon Benjamin [00:10:16] – Through the algorithm if you like, or it may be that when you measure those extra guys, they basically say, “Okay, I don’t know what the information stored on the main qubits is, because I’m not supposed to know that, but I can tell you that qubit number three has been flipped. I don’t even know what state qubit number three is, but he has, or she, has been flipped.” That’s what the extra guys tell you. Sorry if this sounds a bit technical, but it’s a huge breakthrough in the field.
Craig Cannon [00:10:40] – No, and just to pause you very quickly. So, the information you’re passing through, these are qubits as well?
Simon Benjamin [00:10:48] – What you have is, you have your main guys who are storing some information. It may have an error. We’re not allowed to look at it. You have the extra guys who come in. They’re going to be part of the checking process. And then we just talk between them, doing a little calculation, little computation between them, using these two qubit gates that I was talking about. So again, errors might happen, even during that process. We have to be very, we have to think of all possible bad things that could happen and make sure that none of them are a deal-breaker. But basically, this trick of not looking for errors where they are but looking to one side, because you’ve basically done a separate little calculation, that now in the, what we call ancilla qubits, it tells you just one thing, the one thing you do legitimately need to know, which is, is there an error? If so, where is the error? So, that was a very simple idea but very, very important.
Craig Cannon [00:11:43] – And the measuring the ancilla qubits. How do you know that that measurement is correct?
Simon Benjamin [00:11:48] – You don’t. That’s a very good question, right? So it’s like, who guards the guards, right? I mean, if there was one thing that you could trust absolutely, that would be fantastic for us, and we’d be able to simplify all our designs and solutions. But we must assume that everything is untrustworthy. Some things are worse than others. Measurement of a single qubit might be a bit more reliable than, say, the two-qubit gate, but they all have a number on them that says how dodgy they are. And yet, what you need to do is come up with an approach where there’s enough protection that, even if, say, so, you might falsely conclude that there was an error, and there wasn’t one, because the measurement went wrong, right? Then you’re fixing an error that didn’t happen.
Simon Benjamin [00:12:32] – You basically need to construct the overall story so that that event is so rare that you can run your whole calculation, and it probably won’t have happened. There’s ways of checking the measurements as well. It’s layers of protection, basically.
Craig Cannon [00:12:46] – Which makes sense, why error correction, and the minuteness of the error, the likelihood, keeping it low, is the most important part.
Simon Benjamin [00:12:53] – That’s right. Wxactly. I mentioned this number 99.9. Which I said we’re very proud of. It’s actually a very important number, because when you work through these schemes for protecting information, because anything can go wrong, you tend to, what you find is something called, I’ll use another technical word now, a threshold. That means that if you’re doing things better than a certain number, you will be able to use the protection to get rid of the errors, the physical errors, as they occur. You’ll stay on top of the problem, and you’ll be able to run a long calculation with nothing going wrong at the logical level. Suppose that every time you try to do something, make a measurement, make a two qubit gate, something like that, you had a, let’s say a 30% chance of just wrecking it. That would be hopeless, because you’d, it’s like trying to fix a broken watch wearing boxing gloves or something, right? You’re going to cause more damage than, than… In slightly more formal language, you might say it’s like entropy. Can we remove disorder more efficiently
Simon Benjamin [00:13:56] – than we’re putting it in, given that the process of trying to remove it itself is a little bit noisy and damaging? You have to make in that benefit, right? You have to take out more trouble than you’re putting in. Then you can stay in control.
Craig Cannon [00:14:08] – Added to this is you’re running actual code. So the code might be erroring out as well.
Simon Benjamin [00:14:13] – There’s also a whole other question, which is, how reliable is the algorithm itself? But all of this is just, is the machine even doing what I’m telling it to do? You have this number, which the theorists can derive, and I spend quite a lot of time deriving it for particular approaches, which is essentially, how good do the lab guys have to get before they could, now, with enough funding and whatever, build a big quantum computer and know that it will stay on track? Because they’re now so good at doing the operations that, as errors happen, they can take them out of the machine faster than they’re happening. That is the threshold. Now the threshold, when these results were first discovered in the 90s, was about 10 parts per million, so however many, 99.9999 was the kind of level of precision that you would have to have in order to control your quantum computer and take the bad things out faster than you’re putting them in.
Craig Cannon [00:15:14] – Regardless of the number of qubits in the system.
Simon Benjamin [00:15:17] – That’s, yes, exactly. Or almost regardless of it, enough to be able to run very large algorithms and be able to do the things that we dream of for the big, for the big machine. It’s always the case that if you have enough qubits and you have a… Basically what happens is, the more you have, you have to slightly dedicate more and more of them to the checking process, but it’s not too bad, in the sense that if you double the size of the machine, you might only need to slightly increase the number of guys who are checking. This number, this threshold, in the 90s, it was very demanding. In the 90s, we had the theorists saying, good news, we can tell you how to build a quantum computer, and it will work, and it won’t go off track, despite the worries from let’s say the 80s that it would have to go off track. That problem has been solved, said Andy Steane and Peter Shor and many other people who were excited by that. Bad news, the level of control you need to achieve in the lab to use these ideas is still very demanding.It‘s modest by comparison to how classical, classical computers being everything that’s not a quantum computer.
Simon Benjamin [00:16:26] – It’s modest by comparison to them, but compared to what you guys are achieving, let’s say in 1990 in the lab, it’s horribly challenging. In the lab, maybe two qubit gates might have been around the 90% fidelity mark at that time, and yet the theorists were saying, I need you to be 99.999, and that’s a gap. That was a real time for optimism versus pessimism, because how long will it take a technology to get 100 or 1,000 times better in its precision? Really, there were optimists, and I was, at that time, I was like, getting into the field. I was just getting interested in the late 90s in this area and hearing about it, and so of course, I was a super optimist at that time. As a student, you have to be like, ah, everything’s just happening, right? I was like, oh, we’ll probably have it in a couple of years, right? And then there were pessimists who had said, this gap between what we need to be able to build, the theorists say, and what we can build is so wide, it’s probably physically impossible. That would be a tenable, you could say that, and no one would be able to prove you wrong. Now, to come back at last to your original question, why are people excited now, right, it’s because that gap has completely closed. The theorists as a community have been doing better. They’ve been improving their codes and making the demands more permissive. Now, the threshold is about 99%. If you’ve got, if things work correctly in your quantum computer 99% of the time, that’s the turning point.
Simon Benjamin [00:17:59] – You don’t want to be a at 99%, because that also can be seen as the point at which things become impossible. You want to be better than that, right? But 99.9, which is what the guys here at Oxford can do, is 10 times better than the threshold.
Craig Cannon [00:18:14] – Let’s break each component down. How do you move to 99% and above being accepted? What changes have to be made? Are you doing different simulations? What’s happening?
Simon Benjamin [00:18:27] – On the theory side, we have a family of approaches now called topological codes, and these things, one could say a great deal about them. But practically, what’s important about them is the following, a very simple property, especially something called the CATIA surface code, which is, at the moment, the go-to solution for how we would build a quantum computer. Why is it better than the codes we had in 1995? It’s better because it, the architecture of the computer can be very, very simple. In fact, it can be, as the name 2D surface code suggests, you’re allowed to lay our your qubits just in a grid, which is great news for experimentalists. Actually, they’d rather lay them out in a line, ’cause that’s, if there’s something easier than a grid, it’s a line, right? But that’s, that’s very restrictive. But a grid is not too bad. I mean, conventional technology is laid out by fusing a chip into a grid. When I say you can lay it out that way, what I mean is that each qubit only needs to talk to the immediately near-by guys, north, south, east, west. The code does not require a qubit to be able to reach out and talk to a guy three blocks over. The early codes would have required that. They would have required qubits to be able to link to qubits all over the place. Why is that a problem? Most of the ideas for quantum computing actually involve direct, physical interactions in the physics, which is short-range. If you have an idea where you can lay out your qubits in a grid, that’s very, that reflects what is
Simon Benjamin [00:20:09] – straightforward in the physics for you to achieve. On the other hand, if someone says to you, I want you to do something where this qubit needs to talk over here, and then it needs to go and talk over here, you don’t have that in the hardware. What are you going to do? You’ll have to do a bunch of swaps. If you’ve only got short-range links in the hardware, and you’re actually saying, I need this guy to talk to this guy over here, what you’re going to have to do, more or less, is swap, swap, swap, swap, swap, swap, swap, swap, swap, and now they’re next to each other, and they can talk. But you did all those swaps, and those are making errors as well. The cost was in permuting around the information inside the machine, but now along come the topological codes, and they say, “No, it’s fine.” You don’t need to move stuff around. Lay everybody out in a grid. Everybody can just sit there and just talk to the immediate next-door neighbors. It takes away a great deal, most of the operations, actually, because most of the operations would have been used for swap, swap, swap, swap, just boringly moving stuff around.
Simon Benjamin [00:21:10] – Now, by the way, we might perhaps, it’s up to you, talk about network approaches to quantum computing, which is something I’m super keen on, and they actually do have the ability to link any qubit to any other qubit, which is still super desirable. Even though the topological codes have told us we don’t absolutely need it, it’s still great to have.
Craig Cannon [00:21:31] – But it does expose you to more errors?
Simon Benjamin [00:21:34] – Not if you do it right, is the answer. What it, what it means is that it’s tougher to figure out how your hardware is going to work, but if you can figure that out, and we think we have got it figured out, then you can build a machine where the qubits are able to link all over the place. That is fantastic for… It’s always better to have more connectivity. It’s always more powerful to have more connectivity.
Craig Cannon [00:21:57] – Right, rather than a chain.
Simon Benjamin [00:21:59] – Yeah, rather than just a, what we would call nearest neighbor, which means wherever you’ve put your qubits, they can only actually talk directly to their surroundings. But still, the topological codes show that that was enough, and that’s an enormous simplification. That was responsible primarily for moving from the multiple nines to the 99% threshold. Then in the lab, well, the different kinds of technology, and there are a lot of them, and this is one of the things that makes the field quite confusing for someone who just, someone who Googles, because… I sympathize. You type in, “How does a quantum computer work,” and you get 20 different answers. It’s worse that never, actually, just as an aside, because before, like five years ago, if you typed in how does a quantum computer work, you’d get different academics like me talking about our favorite approach. But academics don’t really have hostility to the other approaches. The ones I don’t work on, I wish them good luck. You know, I just want to see a quantum computer happen. But now because we have a lot of companies engaged,
Simon Benjamin [00:23:03] – they need to take a little bit more of a commercial attitude to it. If they’re investing in a particular approach, and you ask them, how is a quantum computer going to work, they won’t tell you about the rivals. They’re going to tell you all about their approach and why that is the solution for sure.
Craig Cannon [00:23:19] – Well, just think about them fundraising. Why fund you rather than the other guy?
Simon Benjamin [00:23:23] – Exactly. I’m not complaining, in the sense that obviously, they must do this, right? It’s a very brave guy in a company who says, first, let me tell you why my rivals are great, and then I’ll tell you why I might have something, right? You can take that approach, but that’s nonstandard. It’s more confusing than ever to try and figure out how a quantum computer works, what it’s going to be made of, what the bits are, because there are so many different possible answers, and now, many of them are associated with companies who are really pushing that and not giving you the big picture, for perfectly sound commercial reasons.
Craig Cannon [00:23:56] – Maybe this is silly to go into one in particular, but in the lab in Oxford, what kind of quantum computer are they building?
Simon Benjamin [00:24:02] – Oxford is actually one of the biggest research facilities in the world for quantum. We’ve got something like 200 people working on this. We’re working on a bunch of different stuff. Of course we’ve got a division between theorists and experimentalists, but among the experimentalists, we’re not betting everything on one horse. We’re taking a bunch of different approaches. The one that I’ve mentioned that we have the world record for the best control, is this thing called an ion trap, which sounds very technical. It’s really a very simple idea. Our qubits are individual atoms. An atom is really nature’s natural quantum system. It’s the system that people were thinking about when they actually began to understand quantum physics. An atom can make a great qubit. The problem with atoms like the atoms I’m made of is that they’re connected to other atoms, and on, on, on, on, on. You can’t have a particular single atom be in a quantum state without it immediately connecting to the atoms around it. What you would need to do to make a quantum computer
Simon Benjamin [00:25:04] – out of atoms is figure out how to keep them isolated, a bit like the famous Schrödinger’s cat in its box, right? You need to isolate.
Craig Cannon [00:25:12] – We should explain that metaphor as well.
Simon Benjamin [00:25:14] – All right, so, sometimes I get told off for mentioning Schrödinger’s cat.
Craig Cannon [00:25:18] – It’s fine.
Simon Benjamin [00:25:19] – It’s probably the one thing they’ve heard of. In the early days of quantum physics, Einstein, which, I probably don’t have to explain that. And Schrödinger, who is one of the fathers of the quantum computing field, were discussing in letters they were exchanging how weird quantum theory is and whether they could get behind it as a sort of explanation of the world. They were struggling with this idea of superposition, which means a thing can be in two states at once. Now, when you’re talking about something as far from everyday experience as an atom, maybe it doesn’t sound too bad, right, because we don’t, we’re not used to looking at atoms anyway, so if I tell you an atom can be in two states at once, you’re like, well, that’s weird, but whatever. What they came up with was a thought experiment, which, it’s important always to emphasize, no one ever plans to do, because it involves potentially killing a cat and everybody loves cats, right? So, but the idea of the thought experiment was to say, well, look, according to physics, we could do the following, and how weird is that, right? In the experiment, we take the fact
Simon Benjamin [00:26:23] – that the quantum world, which is normally thought of as the small scale of photons and atoms, has this weird property of things happening or both happening and not happening, and scale it up. The idea is, you build the box, and the box perfectly isolates the inside of the box from the rest of the universe, which means nothing in the rest of the universe, including you, can measure what’s going on in the box. It’s its own little sub-universe, once the box lid is closed. What do you put in the box? There’s a few variations of this, but the version I like to say, because it’s more exciting, is that you put a bomb in there. And you put a cat in there, and the question is, will the bomb go off and kill the cat or not when the box it shut? But it’s an unusual bomb, because it doesn’t just have a timer or something. It is connected to some kind of quantum measurement. It could be, for example, whether or not a particular atom decays. Atoms that are radioactive can decay from one to another, and that is a quantum event that may or may not happen, and if in fact if unobserved, we would say it has and hasn’t happened. That was the original example.
Simon Benjamin [00:27:31] – But actually, it could be anything. We could put in there one qubit of a quantum computer if we wanted to make it in more contemporary language, and we could say, I’ll put the qubit in a superposition of zero and one, and then after exactly one minute, inside the box, a little measuring device is going to go and look at that qubit. The qubit will either be zero or one. But now, I shut the box and seal it off, and I just wait five minutes. So after one minute, the measuring device looks at the atom, and if it saw state one, it activated the bomb. And if it saw state zero, it did not activate the bomb. And then at that point, it’s disarmed, if it was off. Now, what quantum theory would tell you is that at that moment that the measuring device looked at the atom, what happened was the superposition that the atom was in of zero plus one, it’s in a sense spread to the measuring device. Now you’ve got the measuring device saw state zero in a superposition with the measuring device saw state one. But the measuring device is connected to the bomb. In the very next instant, you have a superposition
Simon Benjamin [00:28:37] – of the measuring device saw state zero and the bomb didn’t go off and the cat’s just bored in the box. And the measuring device saw state one, activated the bomb, it blew up, and the cat is… Get too graphic about that. But, so, it’s a scaling up thought experiment, and it’s saying, look, now, after one minute, and well leave the box shut for five minutes, so for the following four minutes, inside the box, it’s not that one or the other has happened and we just don’t know. That would be the sort of classical explanation. It’s that both those things have happened. They are in superposition with each other, and that is an actual state of the universe. It’s distinct from being one or the other,
Simon Benjamin [00:29:18] – or it’s one or the other and we don’t know. Literally both those things are the case inside the box. Then of course when you open the box, that’s a measurement act from the rest of the universe onto the little sub-universe, and now, it will be one thing or the other, or, depending, by the way, on how you interpret the grand scheme of things, you could say the whole universe goes into a superposition of, the cat is dead and the cat isn’t. That was designed to highlight how weird quantum theory is. There’s nothing, we would never attempt that experiment. I suppose if you were going to do it, you would do it with something other than a live cat. The point is to make the microscopic world where weirdness, maybe we can kind of mentally brush it aside and say, yeah, whatever, small things are weird. But there’s nothing keeping it small, except the technological challenge of doing stuff like what I just described. In a way, a quantum computer itself is a useful and humane version of a Schrödinger’s cat experiment, because we imagine having a huge number of components
Simon Benjamin [00:30:22] – that are all in not just two states but in multiple possible states simultaneously. We expect to be able to use that for useful things, and we mustn’t open the box, as in measure the qubits, prematurely.
Craig Cannon [00:30:33] – How does that relate to using an atom as a qubit?
Simon Benjamin [00:30:37] – Right, exactly. We want the atom to be like the cat. We want it to be able to be in a superposition of dead and alive, zero and one, and maybe we want to have a bunch of them. But we need to keep them, our atoms, totally isolated from the rest of the world, just like if we, in our Schrödinger cat picture, if the box was a bit leaky, if you were like, I’m just going to peek, I can’t, I’m just going to keep one eye on it, it spoils the experiment. If you’re measuring it, one thing or the other will happen. We will not have the superposition. We need to take our atoms and isolate them from the rest of the world, and here’s how we do it. When I say we, you know, I’m being very, I don’t do it. They do it. First off, you want to have a vacuum, because you don’t want your atoms talking to the rest of the stuff that the world is made of, right. You have something called a vacuum chamber, which is literally a box which is very tightly sealed and has some special sort of ports in it that allow you to, for example, pump the atmosphere out of it.
Simon Benjamin [00:31:36] – That’s your starting point, is a vacuum chamber. That already is a little bit exotic compared to conventional computers, because they don’t need to have a vacuum. But the good news is that vacuums aren’t particularly hard. Of all the various exotic things that we need when we are thinking about building quantum computers, sometimes, it’s super low temperatures, there’s always challenges, a vacuum is actually reasonably mundane. In fact, in the early days of computing, we had vacuum tubes. We need better vacuums than that, but we can have better vacuums. First you have your box, literally a box, and you make sure that it’s a vacuum in there, except for, of course, the components of your technology, right? Ion trap, what does it mean? An ion is just an atom that has had one electron flicked off of it. And this means, so now you’re thinking of your atom. You’ve got the core of your atom, and then you’ve got the electrons whizzing around it. The whole thing has, is electrically neutral in that there are as many as electrons as their are protons, and so, it doesn’t do anything special with electric fields. It has no, has no net electric charge. But, so what you do is you deliberately flick off one of the electrons. Now, your atom, which might be, for example, for the guys here, a calcium atom, has one too few electrons.
Simon Benjamin [00:32:55] – Now, what that means is that now, it has a net positive charge, which means you can push and pull it around with electric fields. You can make it come over here or push it away. You can actually manipulate the atom, hold it, push it, pull it, without touching it with anything.
Craig Cannon [00:33:11] – In other words, it’s floating in the vacuum.
Simon Benjamin [00:33:13] – Floating in the vacuum.
Craig Cannon [00:33:14] – You can push it and pull it, and you can.
Simon Benjamin [00:33:14] – You can push it and pull it, because we don’t want it just bouncing around inside our vacuum chamber. We want it to be in the mid, heart of our technology. An ion is just an atom that has had one or, for us, just one, electron flicked off, so that now it has a charge, so that now we have a kind of handle on it, right? What do you do next? You have a chip, which looks to the naked eye, until you really study it, a bit like any conventional microchip, and the chip has metal elements on the top of it, often gold, and those are designed to just create electric fields. You charge them up, and now they have an electric field around them. They push and pull and push and pull your ion to keep it in one place, not touching the chip, but essentially floating above the chip. Here, you’ve got your vacuum chamber. Inside it, you’ve got your ion trap chip, which is just a bunch of little metal elements. Dirt cheap ones, dirt cheap. Actually very expensive, because what they do is they, they design a particular trap layout, and then they have to get it manufactured.
Simon Benjamin [00:34:15] – But in terms of its components, it’s just some bits of metal on a surface. It’s much, much, much more basic than, say, a silicon chip. The ion trap chip itself is nothing special, and the atoms are just atoms, but you’ve now trapped the atoms in the form of ions, because this gives us the ability to have a grip on them, floating in the middle of a high-quality vacuum, not touching anything, and inside the box. Now, that’s a beautifully isolated system. What you can do is you can have several atoms. You essentially have a gap, and there’s another one, and another one, and another one, and then, the first thing that you find, or something exciting that you find, is that a quantum superposition will now last a very long time compared to almost any other way of doing things. For the superconducting qubits that many researchers are excited, and we do do work on that here in Oxford as well, the decay time, the amount of time that can go by before the wonderful zero-one superposition just degrades is the tiniest fraction of a second, of the order down to a microsecond, something like that. For us, the guys over the road have done 50 seconds without, let’s say, any particularly advanced or special tricks, and it’s also, 10 minutes is also achievable by another, by another lab, using some extra tricks. But 50 seconds, which is a proper amount of time, right? It’s taken me longer than that to explain it, but that’s how long, without any intervention, without any special
Simon Benjamin [00:35:47] – tricks and techniques to keep things alive, if you just put, just, I mean, I keep being conscious of how I’m being dismissive of some of the most…
Craig Cannon [00:35:56] – Is it super cooled, room temperature?
Simon Benjamin [00:35:57] – No. This is at room temperature. You take your ion. You use a laser actually to put it into a zero and one superposition, and then you come back, and you see if it’s still in the superposition you put it in later. One way you can do that is you can reverse. You can use the canceling laser pulse, the opposite effect of the one you just did. You take your zero, your zero state and make it zero plus one, and then you come back, and you do exactly the reverse process. And if nothing has changed, that should take you back to state zero. But if the thing is degraded, then who knows. Again, you’ll be in some random state. You can come back 50 seconds later and have a good chance that it’s the, you know, that indeed, you get back to the initial state, which means nothing went wrong for 50 seconds. If 50 seconds isn’t enough for you, there are some techniques which a Chinese group has pioneered which you can push that out to 10 minutes. But the thing is that 50 seconds is enough,
Simon Benjamin [00:36:55] – because in that time, you could have done a huge number of little processing operations, low-level processing operations inside the machine. You could have got tons of work done in those 50 seconds, because you can do gate operations on the order of many hundreds of microseconds. You could get vast amounts of done, stuff done, or, in fact, a fraction of that now. So you could get a vast amount done in the time, what’s called the decoherence time, which is the lifetime of your qubit. In the Schrödinger cat picture, the cat would actually not stay alive and dead for, permanently, because in any real experiment, the box would always be a little bit leaky, leaky for information. After some amount of time, which might be, you know, a minute or a year, the cat would actually indeed be in one state or the other, and so that, we call that the decoherence time. It’s the case for all quantum approaches as well, and one of the big measures for how well you’re doing is, what’s your decoherence time. Our decoherence time is excellent, because our system is incredibly well isolated from the rest of the world.
Craig Cannon [00:38:05] – Are we swapping in new qubits every time the… after 50 seconds, throw in another one, and then scale from there?
Simon Benjamin [00:38:13] – That’s right. 50 seconds is a long time, but it’s not forever. So your two options are this. One, do a calculation that, from the start of the calculation to the end is a fair bit shorter than 50 seconds. That means everything will just behave as it should, and well before 50 seconds, you’re already measuring the system, getting the answer out. Now, for the first quantum algorithms that we made, the first serious ones that might do something useful, that perhaps we’ll see in the coming year, that’s probably going to be the approach. Just go for it. I can get stuff done. However, it’s not really the answer for a long-running calculation that might take hours, days, because that’s… You can’t suppress decoherence for that long. The same techniques we were talking about before that will generally allow you to correct errors, including errors that you made because you didn’t control things perfectly, will also hoover up these occasional, and in fact, if you are doing, let’s think about it. If you’re doing maybe hundreds of thousands of operations per second, and those ones have a one chance in 1,000 of going wrong, then you’re making a huge number of errors per second. The fact that, after 50 seconds, the thing would go wrong anyway, is just a drop in the ocean. If you’re already on top of your game in terms of controlling that rate of errors, which we know that we can be, then that 50 seconds is no problem. It’s actually so much of a weaker effect.
Craig Cannon [00:39:43] – It’s trivial.
Simon Benjamin [00:39:44] – We don’t even bother putting it into our calculations, usually, because, you’re dealing with thousands of times, like, it’s like a fire hose compared to a dripping faucet. We’re dealing with the fire hose. We don’t bother putting in the faucet. And to be honest with you, the only reason in my calculations I would put that in for an ion trap calculation is if the referee told me to. The referee of my paper, who maybe didn’t 100% understand it, says, this is an exciting paper, but I can’t believe that they’ve left out the effect of natural decoherence.
Craig Cannon [00:40:13] – Seriously.
Simon Benjamin [00:40:14] – I’m like, all right, so then I politely reply and I say, we thank the referee for pointing out our omission, and we have now put that in. What that meant was, you know, that the effective rate of errors has gone up by the tiniest amount.
Craig Cannon [00:40:25] – You guys are so polite.
Simon Benjamin [00:40:26] – All the diagrams look exactly the same. Ion traps are a very beautiful system in that they’ve taken nature’s natural quantum unit and successfully isolated it, or a bunch of them, from the rest of the universe, and we, the fact that they successfully isolate them is seen by the fact that they will last in one of these delicate quantum states for a long time. Plus, we’ve also now learned to control them and get them to talk to each other to this very high level. They are actually the gold standard of qubits. But, they aren’t the only approach you’ll hear about; in fact, if you look online, you will hear more about the superconducting qubits, which the approach, the main approach, that Google and IBM and even Intel are taking. That’s a different approach, and it has different strengths and weaknesses.
Craig Cannon [00:41:22] – Let’s assume, let’s go with yours for now, because the question I’m curious about is, the way you tell it to me, like, okay, this is working, this is good. Fair enough. Like, let’s assume that.
Simon Benjamin [00:41:32] – No, no, no, that’s how I think of it.
Craig Cannon [00:41:34] – Scaling it is the challenge, right?
Simon Benjamin [00:41:36] – Scaling it is the challenge. Actually, it’s been a few years now since these very nice numbers were obtained. 2014 for this 99.9%. Why in 2015 have we not already built the quantum computer? Because now, theorists are saying, look, I’ve got some blueprints and designs and protocols.
Craig Cannon [00:41:53] – We’re good.
Simon Benjamin [00:41:55] – Make it all work, exactly, and the lab guys are like, “Well, good news, we’ve actually cracked that number,” and it’s like, boom, let’s go, right? Well, it is the problem of scaling. It’s one thing to have put a small number of qubits in a box in the lab and worked on it for a month, and then on a really successful Tuesday afternoon, you get beautiful data out to prove it all works, and then you get a paper. That’s now science sort of progresses. It’s a gap between that and figuring out, A, how would you make that work not on a lucky Tuesday afternoon but every single time in a robust way, so that’s a kind of, an engineering problem, really. It’s taking away the things, the uncertainties that were basically not making the experiment work on Monday, but on Tuesday, it did. That’s just a development process. But moreover, it’s saying, how can I go from having two or three of these guys behaving themselves to having thousands of these guys? That’s the scaling problem. Everyone is focusing all their attention on the scaling problem. Now, the first interesting number is actually about 50 qubits. Why? Here I have to say a phrase that is not my favorite phrase. Quantum supremacy. Sounds a bit racist or something, but it isn’t. Quantum supremacy also sounds like the end of the road
Simon Benjamin [00:43:17] – for everything that isn’t quantum.
Craig Cannon [00:43:19] – Well, it sounds like AGI.
Simon Benjamin [00:43:22] – It’s like, “Oh, now we’re, it’s all over. We’ve got quantum supremacy.”
Craig Cannon [00:43:24] – We’re done. You don’t have a job.
Simon Benjamin [00:43:26] – Quantum supremacy is, in that sense, a rather hyped-up phrase, but it refers to something that is very exciting and that is, we hope, about to happen. Here’s my question for you. How big does my quantum computer have to be before it could potentially be useful. Well, how could we even answer that? One way is this. We could say, my ordinary computer, especially if it’s a supercomputer, can pretend to be a quantum computer. We know the laws of physics that are governing the quantum system. We know the equations. We can put them, digitize them and put them into software on a conventional computer, and then we can say to the conventional computer, “Okay, you are now simulating or emulating, if you like, a quantum computer, and if I set my quantum computer going into this algorithm, what would I see?” Now, for one qubit, that’s easy. Actually, one cubit, we could do it on the whiteboard. I mean, we would just think about what happens with one or two. For, let’s say 15 qubits, I could write an app, it’s not in my pocket, I could write an app for my iPhone that would simulate a quantum computer that has perhaps 15 qubits pretty easily. For 30, 29 qubits, I would need a very expensive laptop, or a nice laptop, and I would be able to use that machine to pretend to be a quantum computer with 29 or 30 qubits. If I have access to some of the world’s largest supercomputers, I could push it to perhaps 45, that’s been done, or a little higher, in terms of the number of qubits. Now, what I mean by this is,
Simon Benjamin [00:45:04] – I don’t want to make any approximations. I don’t want to make it a bit like a quantum computer. I want the conventional computer to exactly replicate what a real quantum computer would do, full, the full Monty. If that’s what I want, I, with enormous expense, like renting time on a top-10 supercomputer, I could push into the high 40s in the number of qubits. But the thing is, every time I add one more qubit, I double, slightly more than double, actually, the difficulty of the task. In fact, I double the amount of memory I would need. 45 qubits required 0.5, this wasn’t my work, this was work elsewhere in the community, required 0.5, half a petabyte of RAM.
Craig Cannon [00:45:48] – RAM, okay, of traditional RAM.
Simon Benjamin [00:45:50] – Yeah, like in your supercomputer. Which would be distributed over a bunch of boxes.
Craig Cannon [00:45:53] – Exactly.
Simon Benjamin [00:45:54] – Now, if you wanted 46, you’d need a petabyte. If you wanted 47, you’d need two petabytes. 47 is not that much more interesting than 45, but you would build a machine four times as big, right?
Craig Cannon [00:46:06] – Yes.
Simon Benjamin [00:46:07] – It’s this exponential increase, which is exactly what we expect, because a quantum computer is supposed to be exponentially more powerful than a conventional computer, for certain tasks. It’s no surprise that when we try to get our conventional computer to pretend to be a quantum computer, we can get so far, but we’re on this incredibly punishing curve. Quantum supremacy is a word that’s used, sometimes people say quantum advantage or, there is actually a phrase which I would love to advocate, but I can’t with a straight face, which would be quantum inimitability. Because no one can say it. I even have to think.
Craig Cannon [00:46:45] – Spelling it’s difficult, yeah.
Simon Benjamin [00:46:46] – It just sounds too technical and boring. But that’s what’s really happening, because once, let’s say 50 qubits.
Craig Cannon [00:46:53] – Sure.
Simon Benjamin [00:46:54] – You could, with an enormous effort, build a conventional computer that can simulate 50.
Craig Cannon [00:46:59] – Right.
Simon Benjamin [00:46:59] – But that would need a large portion, that would be like a…
Craig Cannon [00:47:04] – A massive.
Simon Benjamin [00:47:04] – Bigger than the current supercomputers, right, and why would you, because then you still wouldn’t be able to do 51. That point at which is just becomes ridiculous to bother trying to match the power of the quantum computer, that turning point is referred to as quantum supremacy, or you might say quantum inimitability, because you literally can’t imitate the quantum computer with any sensible, sane amount of classical computing power. What that means is that there’s no point. If you think you’ve got a clever idea for what to use a quantum computer for that’s going to change the world, if your idea involves much less than 50 qubits, you’re wrong, because you may have a very nice idea, but what we would do is we would just make that into a program that runs on a conventional computer, and you could just have it. You don’t need to have a quantum computer. Once you’re above 50 qubits, let’s says 64, because that’s a nice, binary number, and it’s much more than 50. There’s no way you’ll be simulating completely a 64-qubit quantum computer on any kind of classical hardware. If you can make that machine, it has the potential to do things that we can’t do by any other means. Now, something you should, I should immediately mention at this point is, we don’t actually know anything, the theorists have not worked out anything, that a 64-qubit quantum computer can do that’s super useful.
Craig Cannon [00:48:33] – Including breaking cryptography.
Simon Benjamin [00:48:35] – Right, breaking codes and stuff like that, which is useful if you’re the NSA or whatever.
Craig Cannon [00:48:39] – Sure.
Simon Benjamin [00:48:40] – Breaking codes is in the category of things that needs at least thousands of qubits, but because it’s a big, tough, long-running task, it also needs this whole error-correction thing to be going on, and that boosts the size of it. Because once you say, “Oh, wait, I need error correction,” then you have to do this thing we talked about where each logical qubit is actually a bunch of physical qubits, and for a really long algorithm, it might need to be a lot of them. You’ll leap from thousands into millions of qubits to do the code-breaking stuff. Actually, we’ve got this enormous gap between the point at which quantum computers could be useful for something, which is 50, and the point at which they’re definitely useful for a whole bunch of stuff that we’ve worked out on paper, which is more like a million qubits. 50 to a million, that’s a pretty big gap. We need there to be stuff in that gap, because otherwise, the thing I like to say is it’s a bit like showing an iPhone to a guy from 1965. You’re lie, good news. The stuff you’re working on, these big clunky computers, are going to give us this, and everyone’s going to have one, right, in the first world, anyway. We’re all going to have these amazing machines. The guy might be super inspired by that and might jump out of bed in the morning thinking, okay, I’m part of this epic quest, right, but when that guy goes to their boss and tries to get funding for 1966s, right, it’s no good to say, in 2018, there’s going to be these awesome things.
Simon Benjamin [00:50:05] – That’s our problem for quantum computing as well. What we’ve worked out on paper is stuff that needs pretty big computers, and we don’t know how long it would take to get there. Maybe hopefully not decades, but it could take definitely more than a decade to get to the point where we have millions of well-behaved qubits. On the other hand, what we believe is about to happen in the coming year is that people will start to bring out qubits, sorry, quantum devices that are at or just a little bit over the quantum supremacy threshold. This number, a lot of people are racing to deliver 50 qubits. It’s this kind of magic number. Those machines, however, won’t, as far as we know, be immediately useful for stuff that people really want to do. We’ll be able to test our their quantum-ness and show them off with a whole bunch of ideas, but it won’t break codes. It won’t, what are some of the other things people are excited about? Breaking codes is often mentioned, and it isn’t a very positive application, by the way, breaking codes, but it is very interesting, because the difficulty of breaking codes is so well-established, and therefore, if a new machine comes along and can do it, that machine must have something special about it that the old ones didn’t have. That’s why it’s so exciting as a sort of showcase.
Craig Cannon [00:51:18] – But as a counterpoint to that, the code could also be created by a quantum computer, correct, and defended.
Simon Benjamin [00:51:22] – Right, that’s true. That’s true. So quantum giveth and taketh away, or the other way around, taketh away and giveth, because a quantum computer would crack the codes that we rely on today.
Craig Cannon [00:51:32] – On a classical computer.
Simon Benjamin [00:51:33] – Yes. We can also design communication systems that are protected by quantum physics, where basically, an eavesdropper who tries to see what the communication is would cause a necessary disturbance, like the uncertainty principle. It’s necessary, if you measure a quantum system, to disturb it, and that would be detected. Quantum can also offer you, it’s almost like a salesman. We’re like, bad news. All your current technology for security is going to become iffy. Good news, have a look at this brochure. The things we are excited about that quantum computers will one day do, code breaking, not the most exciting, but a very important proof of the power of the machines, but what people are really more excited about, quantum-enabled discovery is a phrase which basically means, in science, and also in industrial R&D like drug discovery and so on, there’s an enormous amount of trial and error, because it’s not possible to use software to predict what’s going to happen. We can’t predict what would make a great superconductor, and in that way, just go, or say, all right, we’ll go make that. That will be an amazing superconductor. Trial and error, trial and error. We’re not very good yet at predicting how to synthesize complex molecules, so chemists who have an enormous amount of intuition and years of experience just go and try it, and they occasionally get a breakthrough. Trial and error, trial and error. If we could have a far more powerful machine for predicting the behavior of these things, we could take the trial and error out. Imagine you’re a chemist.
Simon Benjamin [00:53:06] – You want to synthesize this big, fancy molecule, because you believe that it will help with Alzheimer’s or something like that. It’s exactly the right molecular shape. Instead of trial and error, you just set that task to your quantum computer. In the morning, the quantum computer, you come back, the quantum computer says, well, I had a think about it. If you do this and then this and then this and then this with this chemical and this chemical, you will synthesize that thing. That would supercharge the rate of progress that we have in various areas. In molecular synthesis, in materials, design and discovery, and could give us, perhaps it’s hype-y for me to say it, but almost a golden age of rapidly discovering new materials, new chemicals, new drugs. That would be very exciting.
Craig Cannon [00:53:46] – How many qubits are we talking around here?
Simon Benjamin [00:53:48] – Right, exactly. The ideas that the theorists have works out thoroughly already would again seem to need millions of qubits. But, there’s hope that some of these tasks may actually work with far fewer qubits, and in particular, some of these tasks may not need the error correction thing. The error correction thing is great as a principle, but it requires you to make the quantum computer much bigger, because you have the logical qubits, which turn into a bunch of physical qubits. If you don’t need that, if you can just use the direct atoms, ions of your computer as the logical qubits, then you can maybe get stuff done with 100 qubits or a couple of hundred qubits, probably a bit more than 50, but that could be an application that sits near to the, the kind of end of the chasm, you know, 50 to a million, that’s this big gap. We’re at the 50 end. Is there a stepping stone, a very interesting stepping stone, which would be the ability to simulate chemistry and material systems that’s actually quite near our end of the chasm? That would be super exciting, right?
Craig Cannon [00:54:55] – Does this relate back to your network design before because you planted the seed there a while ago.
Simon Benjamin [00:55:00] – Yeah, I did plant the seed, so now I kind of want to harvest it. If we’re now saying, look, I understand this fault tolerance, this threshold, all that’s exciting stuff, but that stuff lives at the other side of the chasm, that’s what you do with your million-qubit quantum computer.
Craig Cannon [00:55:13] – That’s a tough sell.
Simon Benjamin [00:55:14] – To make it be able to run forever. But now let’s focus on what we might get done in the next two or three years. What are these first machines, these embryonic or adolescent quantum computers going to be able to do? Now, we come back to connectivity. If I have to swap my qubits 10 times or 100 times before it gets next to the guy it now needs to talk to, that’s very debilitating, because now I will pick up loads of error on the way over there, and my whole algorithm will be very burdened with error. We aren’t using these clever error-correction techniques. We can’t afford it in terms of the number of qubits. We’ve only got 200. By the way, 200 qubits, I mean, 50 is where we want to be in the coming year. Hopefully that will become 200 fairly quickly.
Craig Cannon [00:55:57] – The baseline, where are we at right now?
Simon Benjamin [00:55:59] – Where we’re at right now is that various labs around the world can give you 20 qubits. Ish. 10 to 20. What’s been announced and trailered, sometimes announced, sometimes kind of unofficially announced by some of the big companies is that they’re working on a 50-qubit machine. Because that’s the obvious goal, right, to get to 50. That’s typically in a grid that’s like a seven-by-seven, that would give you 49, a grid of qubits. But those plans, those designs are nearest-neighbor ones, which means if you want the guy in one corner to talk to the other, it’ll be swap, swap, swap, swap, swap, swap, swap, swap, swap. You’ll get a lot of error in there. We’re in the regime of like, tens of qubits, trying to push to 50. Now, the network approach we’re taking is a very different philosophy.
Simon Benjamin [00:56:46] – The number of qubits we have for a long time has just been a handful of them, like five, and we’re not trying to scale that up. And why not? Because five is far too few to do anything useful. What we’re trying to do is have five inside a box that behave really well and make a beautiful little, useless, but beautiful little quantum computer. Then to understand how to link that box to another box with an optical link. The idea is that if you can crack that, if you can have a small quantum computer and make it link to another small quantum computer that’s sitting right next to it, and if you’ve mastered that, then nothing stops you, except your checkbook, I guess, from making loads of them and linking them up. That’s our approach to scalability. It’s not to have the qubits directly talking to each other, and then you say, well, I’ve done 20. Let’s see if I can do 50. That was super hard. Let me shoot for 64. It’s getting more and more complicated. I’m having to control more and more parameters. It’s a nightmare, which is the way it’s been up to now, anyway. Instead, we’re saying, look, let’s just get really good at making a small module, a fixed size, and learn how to plug, to connect two modules so they can talk to each other. It’s a quantum computer that’s divided into little pieces, and each piece is independent, and it’s pretty much plug-and-play in the sense that you just, if you had 50 of these modules, then that would be a very powerful machine. If you manufacture another 10, you just plug them in, and now it’s an even more powerful machine, right?
Craig Cannon [00:58:13] – This optical link doesn’t have the same degree of disturbance?
Simon Benjamin [00:58:16] – That’s a very good question. How does the optical link work? The optical link does indeed have a lower quality to it than the internal operations, which will be super good. At the 99.9 that we’ve already achieved. The reason people don’t do this, because it sounds like a great solution. Like, why doesn’t everyone do that?
Simon Benjamin [00:58:37] – Everyone ought to make little quantum computers and link them up. It’s the answer, right? It is the answer, but there is a problem, which is that the links, there’s nothing in the physics that means the links can’t be beautiful and very clean and very high quality, but in terms of the experiments that have been achieved so far, the links are not great. You might have, inside your box, one part in a thousand is your error rate, but for the links, one part in 10 might be your error rate. Then you think, well, that’s no good, because if I’m joining up these beautiful little components with crappy connections, the whole machine is actually going to be crappy, not beautiful. There’s an answer to that, which is very much like what you would do in conventional technology, which is that you use the links several times, and by using it several times, you can effectively get one better quality communication out of it. Your phone, I keep reaching for my phone. Your phone, if it’s on the edge of your, if it’s seeing one bar, right, of your wireless, that’s at the point where it’s very iffy whether it can connect or not. But when you load a web page, it loads more slowly, but it still loads perfectly. It’s not like the images are full of static.
Simon Benjamin [00:59:44] – And that’s because your phone can keep requesting. It uses a code to protect the information. It can also request. It can say, I didn’t get that. Send it to me again. We know now how to use the same kind of idea in the quantum setting to say, I’ve got a bit of a crappy link between what we might call the Alice module and the Bob. We love Alice and Bob in the quantum field, A and B, right.
Craig Cannon [01:00:04] – Fair. So you’re sending packets, essentially, it’s not the whole thing.
Simon Benjamin [01:00:09] – Like that. I’m being a little bit loose. The actual thing we do is something called entanglement, which we haven’t discussed, but that’s the other big, most people have heard. Ooh, quantum involves something called entanglement, right, which Einstein called spooky and blah, blah, blah. The test of whether your quantum computer can really be broken into little modules and wired up, one way to ask that question precisely is to say, can you achieve good quality entanglements between two modules? That would be the proof that they are actually part of the same quantum state, part of the same quantum machine. If you cannot get entanglement between the Alice module and the Bob module, then your machine will forever really just be a bunch of separate units that aren’t teaming up to create one quantum state, one quantum calculation. You could say then, all right, I’ve got a link between my two boxes. Let’s think about that link. What quality of entanglement can it create between the Alice ion trap and the Bob one? I might say, well, not very good, only 90%. Then you’re like, well, that’s not very good. That’s not good enough. You’re not going to. What we actually do is we create a couple of qubits
Simon Benjamin [01:01:15] – that are entangled, 90%, not very good, and we store them. Remember, we can store stuff practically forever. Well, 50 seconds. Then we create another, what we would call, another technical phrase, a bell pair, which means two guys that are entangled with each other. And again, they’ll be in 90% quality. They’ll be, you know, not good, not as high as we want. But now we’ve got two what I would call crappy bell pairs, and what we know how to do is to sacrifice one in order to boost the quality of the other. This is too use an analogy, but it’s a little bit like if you, someone’s trying to talk to you, and it’s a little bit staticky. If you hear it once, you’re not sure, but if you hear it twice, even though it’s staticky both times, you’re able to infer more reliably what that, what that information was. That’s a very loose analogy. But basically, we take two poor-quality communication channel uses, and we make it into one good one.
Simon Benjamin [01:02:15] – That is the solution to that problem. As long as your boxes have very good memory, they do have a very good memory, as long as they have very good quality internal operations, which they do have, then you can take a poor-quality link and boost it by using it a few times into effectively a very good quality link, as good, actually, as the internal operations. We actually plan to build a machine which will be made out of modules, which, each one is a small quantum computer, and by the way, so small that on its own, it’s not good for anything. Even as few as five qubits. Then we have modules, and we link them up, and when two of these modules need to talk to each other to become part of a single, unified quantum state, we’ll use our link, and our link will need to be used several times and then purified or distilled is the phrase we use in order to make it as if it was a very beautiful, very perfect link. And now, these two modules will be connected in a high-fidelity, high-quality way.
Craig Cannon [01:03:12] – Does this perform at the same degree that your 10-qubit quantum computer? Well, you’re stitching them all together. Is it performing at the same speed as one of these 50-qubit quantum computers?
Simon Benjamin [01:03:22] – Yeah, that’s another great question. Initially, no. Because the links are, individually, they’re fast, in that they can be about a megahertz, right, or something like that, but the problem is, it’s a technical problem, but a very interesting one that we’re having at the moment, is that the links are optical. What happens is a particle of light called a photon comes out of your ion and, flies out of your ion. It needs to be caught and put into an optical fiber, and that link needs to be going on at the other end as well. The problem is, at the moment, most of the light gets lost. It doesn’t, it comes out of the ion, and it just hits the wall of the chamber. We know how to solve that problem, by essentially building what we would call a cavity, which is a way of catching and controlling light, into the system, and we’re working on it. But for the first generation of the technology that we envisage, we won’t have that. It’s a little bit, two steps down the road. Instead, we’ll have to just deal with the fact that we, that these links are kind of slow in that you try to use the link and nothing happens.
Simon Benjamin [01:04:27] – You try to use the link and nothing happens. You try and you succeed, and then you have to do that a few times. That slows you down a lot. The actual rate at which the clock speed, if you like, of our modular quantum computer will be limited by the rate of connection, in version one. But, for version one, we don’t really care how long it takes to run, because we’ll be doing small problems that can fit in 50 or 100 or 200 qubits. They’ll be fast anyway, and tit doesn’t really matter if they take, it’s not, this is oversimplifying, but it doesn’t necessarily matter if the calculation gets done in 10 microseconds or in 10 milliseconds or indeed in 10 seconds, because it’s a small calculation anyway that we’re trying to do. What we envisage is a first-generation of this technology, which, so you’ve drilled right down to it, actually. It has interesting properties versus these 50, these seven-by-seven grids. It’s slower, but it’s higher fidelity, and it’s higher connectivity. We believe that it will actually be, you would use them for different things, but if I’m going to choose one, I’m going to say ours is better.
Simon Benjamin [01:05:34] – As things mature and you have more and more modules and you start to take on more and more challenging problems and it starts to become frustrating, that you’re limited by the link speed, then we have this solution that will basically stop us wasting all the light, catch all the light, and then we come right up comparably to the speed of other approaches, whilst still having the very high connectivity and the free scalability and so on. What we actually picture is if you’ve seen those photos of, say, a Google server farm, these very big facilities where you’ve got… You look in different directions, you’ve got these pillars running off into the distance. You can picture something like that for a single quantum computer, if you want, where you have a bunch of modules, and they link, and you have the fibers coming up and going, you know, in sort of cable guides and coming back down again. You build your quantum computer in a big room. But why not? Rooms are relatively cheap, compared to quantum computers, so it doesn’t really matter if it’s big or small. It means that the end user won’t directly have a quantum computer in their phone for any time, but to be honest, there’s no approach to quantum computing that’s at the same sort of level of maturity as superconducting qubits or ion traps that could possibly fit a quantum computer in your phone. But that doesn’t matter, because the way you use your phone to do the more challenging tasks, even speech recognition, is typically in the cloud, right, we would say in the cloud.
Simon Benjamin [01:06:59] – As long as somewhere, there’s a big, fat, quantum computer behaving itself, achieving, you know, having a low error rate, having the necessary connectivity, being big enough to tackle the tasks, then in principal, that can be doing jobs not just for the scientists and researchers, but if you’ve been able to think of an application, so machine learning might be one of them, that actually benefits the enduser, the enduser can also use the quantum computer. We picture this kind of thing: a modular machine where you have as many as modules as you need to get the task done, and by the way, at the moment, a module, the vacuum chamber in the lab, is this kind of size.
Craig Cannon [01:07:35] – Oh, really?
Simon Benjamin [01:07:36] – Quite chunky. Not too bad.
Craig Cannon [01:07:39] – I saw that photo floating around. Someone here won an award, right, for their floating…
Simon Benjamin [01:07:41] – Oh, yeah, they are actually quite pretty, because they’re sort of shiny metal, and they’ve got these cool things in the side. They look a bit like steampunk technology or something. They’re about this size. The reason they’re that size is because in the lab, you want to be able to get your hand in there. You don’t, it’s human-scale technology. You know, you want to get screwdrivers in there and so on. But there’s nothing inherent to the technology that would make them be that. Actually, they would work a little bit better if they were miniaturized. You could shrink that down to a far smaller little vacuum chamber, and that’s your little quantum computer. And then you want 50 or 100 or 200 of those, but that’s not so bad. Just lay them out. And so, that’s our picture. That’s a thing we’d love to build, actually. There is no fundamental impediment to that, actually. If someone came to us and gave us a billion dollars, we could have a crack at building that tomorrow. But no one has given us a billion dollars, so instead, we’re taking a slightly more sort of traditional, academic, conservative route of thrasing out the physics of two of these guys, and then having a think about how to sort of come up with a next generation which is a bit better, and then building four, and then building 16 and 32 and 64. But, there will certainly be a point, and I think at much earlier than 32. Once you’ve cracked the physics of even two of these guys, you’re good to go, in the sense that, you know,
Simon Benjamin [01:08:57] – you can just partner with a company that would manufacture these things, and you can just have your quantum computer.
Craig Cannon [01:09:01] – Stitched together and go.
Simon Benjamin [01:09:02] – In fact, this, this network approach is, I think, the only one, that if you wanted to Manhattan Project it or moonshot it, one of these big projects that have happened in the past where you have a goal that’s extremely challenging and you just decide you’re going to make it happen, if you want to do that with a quantum computer, if you said, “Look, I’m impatient to wait 10 years, 15 years for the big quantum computers to come, I want it to come now,” the network approach is the only one that could potentially deliver that, because we just, when we think about approaches where the qubits are right next to each other and just make that structure bigger and bigger and bigger, people are having heroic efforts now to deliver seven-by-seven grid, but from talking to those researchers, it doesn’t seem to me that they will then immediately know how to do a 10-by-10 grid. That will bring a whole new raft of problems and levels of control challenge and so on. At some point, presumably, you do know how to just keep on scaling with a relatively modest amount of new investment. But with the modular approach, as soon as you’ve got two guys that link really well, you can have a million if you can afford the cost of one times a million, right? So I mean, there is no difference. The system doesn’t change.
Craig Cannon [01:10:16] – On the timeline side, are we talking within a couple years that you guys are going to make this happen?
Simon Benjamin [01:10:23] – Well, it depends on how much interest there is in actually trying to build the machine that’s potentially useful. Academics, left to their own devices, will tend to go hard for the interesting scientific results, but they don’t really, we don’t really know how to tackle the problem of building a big facility. That’s not a problem for a university professor.
Craig Cannon [01:10:44] – Nor the funding. You don’t have the same incentive as a company.
Simon Benjamin [01:10:47] – No, exactly. A company would be saying to its shareholders or investors, look, here’s our year one, here’s our year two, here’s our year five milestones, and what we’ll be able to achieve at each stage. The academics work in a different way. But what I will say to you is that certainly, sorting out thoroughly and demonstrating this idea of two modules, which fully link together to form a single unit in practice, so to the programmer, it’s a single quantum computer. The engineer knows that it’s been broken into two pieces with an optical link. Getting that demonstrated is an immediate goal for us in the next year, getting it thoroughly demonstrated. Whether we then go, hooray, let’s build a thousand of them, right, I would love to, by the way.
Simon Benjamin [01:11:28] – I would love to do that. There was nothing more exciting than saying, we’ve cracked it. The more naturally academic route would be to say, okay, we’ve got that; now let’s get that cavity thing in. Now let’s do this, now let’s do this and just keep making the physics better and more exciting and more well-controlled. But I do believe that once you’ve got two modules that talk to each other really well and have all these lovely properties, you are, you have a green light to go large on that, if you want to. But it is, this is probably boring, but it’s, it’s an interesting problem all around the quantum computing field, not just for people who want to build network machines, that you think, how do you actually get from the lab to some kind of device? That’s a problem that people have met in many different areas over the years, but it’s particularly acute for quantum computing, because it is such a difficult to understand technology. It does require such a lot of detailed know-how. It’s not easy to do, let’s say, tech transfer out of the laboratory into some kind of commercial setting. But, it’s happening. There are spin-out companies coming from academics all over the place now as they start to wrestle with this question of how can I have an investor-facing aspect of my research where I can, indeed, start to think about these exciting questions, so let’s build it.
Craig Cannon [01:12:48] – Yeah. Do you think now is the time, just kind of wrapping up, that people are going to start rolling out cloud compute or are we not quite there yet?
Simon Benjamin [01:12:58] – I think there are already companies which aim to do that.
Craig Cannon [01:13:01] – I think so too.
Simon Benjamin [01:13:02] – In that sense, we are there. I’m excited by that, and yet also a little bit worried by it, because of course when a startup company especially, but even a titan like IBM, when they need to explain their actions to their investors and excite their investors and shareholders, they’re going to pitch things in a very optimistic way that an academic would normally be more cagey about. There’s that tension that I already see in the field between the story that you’ll hear from, and the worry is that it will create too much expectation. There will be too many stories appearing in The Economist and whatever about how quantum is going to revolutionize everything. Quantum supremacy, as exciting as the phenomenon that it refers to is, is a little bit, I’m afraid, feeding into that hype, because if you’re an investor and you’ve read a little bit about this, you’re like, “Right, I’m going to invest in this company that will achieve a 50-qubit machine. It will get quantum supremacy, and then, profit.”
Simon Benjamin [01:14:05] – And there is no, it’s one of those step one, step two, step four is profit, or whatever that joke goes like, right? We don’t yet know useful things to do with a small quantum computer. We’re furiously working on that, but we haven’t talked about machine learning. That’s another sort of interesting possibility that’s being explored, to put it simply. We talked about the enabling more rapid discoveries, which is very exciting, but we can’t yet prove on paper that a small machine will be able to do those things. We’re furiously investigating it. The risk is that people get too excited, they think things are going to happen in the next one to two years, in terms of not just the exciting machines coming out, but in terms of actual, useful breakthroughs. This drug was developed on the Quantum 3000, that kind of stuff, within the next couple of years, which might happen, but could very easily not happen, and, you know, the hype machine starts to feed itself. In artificial intelligence, we were talking about this, there’ve been a couple of AI winters which the guys, now, when I heard the phrase AI winter, I assumed it was like, from when Skynet launches all the nukes, right, and then we’re, like a sci-fi thing. But AI winter is just the field started to attract loads of attention, there were loads of investors, and then it didn’t quite deliver fast enough, and then suddenly it became toxic to say that you were working in AI, and that was like nonsense and silly stuff for academics. Not once, but twice.
Simon Benjamin [01:15:28] – And for a couple of years, the field struggles to get funding and then starts to build up again.
Craig Cannon [01:15:32] – I know people working in machine learning now that are very concerned. They’re like, “We want to be able to continue working on this.”
Simon Benjamin [01:15:38] – Exactly. Machine learning is now working in the sense that it’s working in Facebook, it’s giving us self-driving cars, and yet somehow, expectations still manage just to run ahead of it, right? People just can’t stay calm. There’s even a possibility that it will happen again.
Craig Cannon [01:15:52] – For sure.
Simon Benjamin [01:15:53] – Be over-invested. It might be that that’s just the way things have to happen, but it would be kind of nice, as an academic, a slightly more sort of conservative, I want to see more linear progression. I’d rather that the quantum computing field did not overinflate and then collapse and then we all have to really struggle to get funding for a while, and then it can build up again. It would be nice if we could actually moderate things.
Craig Cannon [01:16:14] – Yeah, hopefully.
Simon Benjamin [01:16:15] – But who knows? I’m more excited than I’ve ever been to see these, because we are now getting, on the verge of getting machines that will behave in ways that cannot be predicted, cannot be simulated. We’ll be in the regime where we’re genuinely discovering how the machine behaves by having one, and that’s what we’ve been dreaming of all these years, and there is also the potential to just really commit and try and build a big machine, which I would love to do, but who knows.
Craig Cannon [01:16:45] – Last question, then. If someone wants to get involved and start working in quantum computing somehow, where do you think they can provide the most value?
Simon Benjamin [01:16:53] – Tthe answer to that question is a richer answer than it was even a couple of years ago. A couple of years ago, the real answer would be, to be honest, you need to be in university as a young person, you need to choose the right course options, and you just need to go into academia. Now, there’s a richer set of ways you can get involved. Still that’s the primary one, to be honest. But there are companies that are recruiting people with all kinds of expertise now, including, for example, software engineers who don’t know anything about quantum, but they do know about writing high-performance software. I work with some such people, right? Because one of the things we want to do is push the limits of how well our convention computers can pretend to be quantum computers. It’s important also to check that the emerging quantum computers are really doing what they say they’re doing. For example, if you are a programmer who would love to get involved with this, there are now opportunities where you can provide that part of the puzzle, or if you’re a systems engineer who doesn’t, again, have a huge background in quantum physics but wants to be part of this process, the good news is, people are now trying to build these complex machines that are really too much for the laboratory and are really complex machines that need to be developed in a separate project. We are looking for people who know a lot
Simon Benjamin [01:18:06] – about systems integration, all these kind of problems, and they don’t need to understand the deep properties of the quantum aspects of it. Getting a vacuum chamber that has a good optical interface is a set of problems which enable a quantum computer, but you don’t need to understand quantum superposition, and the field is trying to diversify to include more such people, not just professors whose merit, measure of success is can they get a paper into Science or Nature.
Craig Cannon [01:18:34] – Publish.
Simon Benjamin [01:18:35] – But people who are outstanding engineers and programmers who don’t care about papers into Nature but do care about meeting their goals and satisfying their contracts, essentially. We are diversified. It is leaving academia, or at least spreading out. And so I would say, anyone who has a lot of technical expertise in these kind of areas or anyone who’s a very good programmer, the door’s now open to participate in the quantum revolution.
Craig Cannon [01:19:00] – That’s great. Thank you so much for your time.
Simon Benjamin [01:19:01] – Okay, thank you. It’s been fun.
Craig Cannon [01:19:04] – All right, thanks for listening. As always, you can find the transcript and video at blog.ycombinator.com, and if you have a second, it would be awesome to give us a rating and review wherever you find your podcasts. See you next time.