Technology, Connected - Why Quantum Computers Make Too Many Mistakes - Oliver Dial, IBM Quantum

Episode Date: September 13, 2025

Quantum computers make mistakes — a lot of them. One in every thousand calculations can be wrong.In this Thinking on Paper Pocket Edition, Mark and Jeremy speak with Oliver Dial, CTO of IBM Quantum, about how researchers are turning unstable prototypes into practical machines.Oliver explains the difference between error mitigation and fault tolerance, how IBM’s new codes make quantum systems ten times more efficient, and why AI now helps optimize the circuits themselves. He also shares how quantum computing could transform material science, unlocking lighter, stronger, and smarter materials for the next technological age.Please enjoy the show.And remember: Stay curious. Be disruptive. Keep Thinking on Paper.Cheers,Mark & Jeremy--Other ways to connect with us:⁠Listen to every podcast⁠Follow us on ⁠Instagram⁠Follow us on ⁠X⁠Follow Mark on ⁠LinkedIn⁠Follow Jeremy on ⁠LinkedIn⁠Read our ⁠Substack⁠Email: [e-mail]📺 Watch the show on ourdedicatedd YouTube Channel

Quantum computers make a lot of mistakes. Typical error rate for operations is someplace between 1 in 1,000 and 1 in 10,000. So imagine trying to do a computation on a computer that one in 1,000 calculations is just wrong. It's a ridiculously high error rate. How does that error rate compare to classical computers? I don't actually have the air rate for a transistor off the top of my head. It's something on the order 10 to the minus 15. Enough orders of magnitude between them that it's just an unfathom.

the mold number. In practice on your classical processor, you typically don't do anything to correct for the errors other than maybe in the memory and the communication. So to do calculations today, we use something called air mitigation. And what that means is we run the same calculation a lot of times. We run small variants of the same calculation. And we can use that to statistically correct the errors. So even though each individual run had mistakes in it, at the end of the day, we can tell you what the answer would have been without those. That process is pretty expensive. You end up needing to run it exponential number of times in the size of the circuit. And so you have the same overall scaling that you do with a classical computer, but the pre-factors better.

So you end up winning. So this is sort of something that you can do to get into the game. We think it's going to work well enough that we'll show quantum advantage this year or next year. And that would be the point where we're doing something better, faster, cheaper on the quantum computer than we can do on a classical computer. Is that what you mean by quantum advantage? Yep. Quantum utility, something that you can't directly simulate, quantum advantage, something, done better. What is fault tolerance? Air mitigation. We're correcting the air statistically after the

fact. Fault tolerance, you correct them as you run. And that is such a huge difference because it means the errors aren't compounding over the course of the computation. Fixing them as you do, the calculation gives you ability to go to enormously larger circuits. That's that factor of 20,000 that we're talking about in 29 with Starling is the difference between air mitigation, recting after the fact, fault tolerance, correcting as you go. When you're doing this correction of the errors, on the fly as you go. You actually need a substantial classical computer to work out what the errors were and to correct the calculation going forward, and you have to do that real time. So the classical computer has to be able to keep up with the quantum computer while you're doing that.

For both the advantage applications and for these long-term fault-tolerant applications, anything you can do on a classical computer, you should do on a classical computer. We really want to reserve the quantum computer, this precious resource for just the parts of the problem that it's really the only thing that it can solve. One application a lot of people like right now for advantage is something called SQD. It's a technique for simulating a chemistry. And in it, you use the quantum computer to choose what we call a basis, to choose basically how you represent the electrons and the chemical, but then you use the classical computer to actually work out what the energy levels are. And the reason why it works really well is the classical computer, when it does that actual calculation

of the energies, does that calculation accurately. And so you don't need to correct for the kind of faults on the fly like you normally would. And so it's sort of an interesting example of air mitigation. In that, we're typically seeing runtimes on the quantum computers that are measured in hours and runtimes on the classical computers that are also measured in a few hours. But you can trade them off against each other a little bit. If you have a lower noise quantum computer, you can also make the classical part cheaper. You have a noise your quantum computer. You can make up for that by throwing additional classical supercomputing power at it. The way we correct the errors is we use more than one physical cubit to store each logical cubit. And that lets us do parity checks. We can compare the values on

the physical cubits and see if one of them is split and fix it. The ratio of the number of physical cubits you need to the number of the logical cubits you get, people call the rate of the code, and it's super important. Because the bigger that ratio it is, the harder it is to build the quantum computer that you want. If you look at the best air correcting codes people were talking about a couple of years ago for a useful logical cubit, you need it on the order of a thousand physical cubits per logical cubit. To put that in perspective, this big machine behind me right now, that guy has about 450 cubits in it. So you needed something about twice as big as that in terms of number of cubits to store even one logical cubit. And one logical cubit is useless.

I can simulate that on my iPhone. You really need about a hundred of them before you're even doing anything interesting. So part of this announcement was an architecture for this computer built with a code that's about 10 times more efficient. And so that means for this code to store, in this case, a dozen logical cubits that come in 12 packs, we need about 400 physical cubits. So again, referring to the system behind me, that's about a 12 pack of cubits.

If I had about 10 of those systems, I would have something that's more powerful than the most powerful classical computer in the world. Now, to put that in perspective, this is the lobby of our building. I'm a physicist. Condensed matter physics study like weird quantum states of matter.

And I got into quantum computing to build that Blue J system. It was like the dream. You know, this large-scale fault-tolerant quantum computer, the first quantum computer that can really run any quantum algorithm that you want to throw at it. And so it's always been kind of a game of, well, we're working really hard on the qubits. We're working really hard on small air correcting codes, all these little technology proofs for these air-mitigated systems that they're computationally useful, but they're tricky to use well, in all honesty. And to have the technology have advanced so far that at this point in my career, I can say, you know what? I think that in eight years we'll have built that thing. That's just amazing that it's watching this entire field come to a pinnacle in the next decade that has me just delighted.

Pickle tink. Does AI play a role in the IBM roadmap? And if so, how? So we talk about bits, neurons, and cubits. classical computer, AI, and quantum is being what's going to power the future of computing. And at the end of the day, IBM and IBM research in particular is all about the future of computing, figuring out how it's going to influence all of our lives together for the better. I think there's going to be some really interesting stuff in quantum plus AI, actually.

We talked a little bit about scaling and quantum advantage and be able to do something faster, better with a quantum computer than a classical. There are actually some very specific AI problems for which you can prove that a quantum computer would have an advantage. And that's sort of surprising to be. I mean, AI is very heuristic. It's very much, we do all this stuff and it kind of works, but it's really hard to prove why. And so beyond that, as we get quantum computers big enough to explore AI problems, I think we're going to see heuristics where quantum and AI play together very nicely as well. We use AI a lot in quantum. We talked a little bit about compiling quantum circuits. And it's a lot like a classical compiler for

software that we have an optimizing compiler. It finds places where you have redundant things inside of your quantum algorithm simplifies them so that it runs faster. Right now, because we have such high air rates, if your circuit runs faster, it runs better. That's super important. And actually, we have a lot of AI-powered what we call transpiler passes that go through and improve quantum circuits. So the reinforcement goes in both directions. To the form factor in the footprint of quantum data centers is looking very different

than traditional data center footprints. Well, and it's looking very different from how anyone would have predicted it would have looked two years ago, that this advancement in the air correcting, codes and the idea of building these modular fault tower quantum computers is really bringing this from the kind of large-scale engineering you would think of like a space program or something like that into something that looks much more like a traditional IT sort of system. You're at dinner party. What do you say to people though?

Why quantum? Why does it matter to me? Why does it matter to humanity? I really like to pick on just one of those applications that I mentioned material science. How do we simulate semiconductors, superconductors, strange novel materials? And the reason why I like to pick that one is, what are the ages of man named? We have the Stone Age, we have the Iron Age, we have the Bronze Age. That is how important materials are to us because it dictates what we can build.

And so if quantum computers can bring forward material science so we can develop new light or stronger materials, superconductors that work at higher temperatures, any of those things, just one, one run that invents one thing, has the potential to change everything for everyone. And