Unlocking Universal Fault Tolerance with Cat Qubits and Delving into Advanced Quantum Concepts

[00:00:00] Welcome to the Quantum State, a podcast exploring the latest research and innovation in quantum computing.

[00:00:06] Join us as we dive into ground-breaking breakthroughs, trends, and use shaping the quantum landscape.

[00:00:31] Hi everyone and welcome back to the Quantum State.

[00:00:34] Today we have Theo Peronen, which I'm going to ask you to pronounce your name as well who is the co-founder of Alice and Bob.

[00:00:42] So please introduce yourself correctly now that I've butchered that and tell us a little bit about yourself and how you got into the quantum computing space.

[00:00:51] Thank you Anastasia, so indeed I'm Theo Perona, I'm the CEO and co-founder of Alice and Bob.

[00:00:58] I'm a physicist by training and passion to entrepreneur later on.

[00:01:04] I did my studies at Ecole Polytechnique and specialized in quantum physics at Ecole Norma de Chibitier, here in France and Paris.

[00:01:14] And did my PhD actually on super connecting qubits for quantum information.

[00:01:21] And we started with Rafael Liscan, my co-founder, Alice and Bob right before defending our PhDs back in 2020 so four years ago.

[00:01:31] Great, so tell us also what does Alice and Bob do?

[00:01:35] Yeah, so Alice and Bob were building a universal, faulty quantum computer out of a new kind of super connecting circuits called the cat qubits.

[00:01:47] And the idea of the cat qubits is really to design a qubit from the ground up for four tolerance, for the ideal quantum computing.

[00:01:58] And the thing that is incredible with cat qubits and I'm sure we're going to talk about that bit during this conversation is that they're able to correct autonomously half of the errors that limit the performances of a quantum computer.

[00:02:15] And by doing so, they provide a shortcut to four tolerance that we're building upon.

[00:02:21] So at Alice and Bob were a hardware company, we do everything from the math, the design of the chip all the way up in the integration of the system and the firmware of the machine.

[00:02:33] For our listeners, your team is quite interested in shores algorithm in particular and that's been one of the high impact things that you've been thinking about.

[00:02:43] What do you see as the implications of that? Is it going to be useful in the future for people who are shifting away from the cryptography that it potentially compromises anyway?

[00:02:55] That's a very good question. Actually, the reason why we go for shore algorithm as a benchmark is because shore algorithm has several features.

[00:03:07] The first one is a historical one and the second would be that is it's a very cleanly defined use case.

[00:03:17] And the last one is that it's actually one of the toughest use case out there. So to say a bit more so historically shore algorithm, I mean game as a kind of a breakthrough.

[00:03:29] The moment in time where people started to realize that quantum computers were not limited in terms of potential only to simulating quantum physics.

[00:03:41] It was a new class of machine able to tackle problems that were, I mean, that had no other options to be solved.

[00:03:52] And so this is when quite a few physicists mathematician started to investigate more deeply quantum computing.

[00:04:02] Now, as a benchmark, the good thing is that when you say, well, I'm going to benchmark an architecture to see how many qubits what are the hypotheses to be able to run shore algorithm to break RSA 2048.

[00:04:18] Well, there's nothing undefined there. You have a very cleanly defined problem and so you can compare apples to apples when you benchmark to architecture.

[00:04:31] And then the good thing for us I mean from a mindset of building a useful machine, the fact that we know breaking RSA 2048 is actually one of the most stringent use case out there.

[00:04:47] So one thing I like to say is that if you can run shore, you can run anything.

[00:04:52] So that way you can think of this at the end goal at least for now.

[00:04:59] I mean computing in general is a never ending story obviously but it's a good target to have in mind.

[00:05:08] Is you wonder verifies well?

[00:05:10] Yeah, exactly.

[00:05:12] Yeah, it's a good benchmark.

[00:05:15] Now from a useful I mean is it going to be used?

[00:05:20] I actually hope not.

[00:05:24] But that's a personal take here.

[00:05:26] I believe for most of internet security and things like that will have plenty of time to pivot away.

[00:05:33] Now it's a known thing that many intelligence services at the moment or not very nice players are starting to fill up hard drives with encrypted data that they hope to decipher later on but apart from that.

[00:05:52] Yeah, I think it's more of a math problem as a benchmark.

[00:05:56] So this is how I view it.

[00:05:58] Yeah another thing is right it's a pretty deep circuit.

[00:06:01] So when we talk about financial and pharmaceutical applications VQE these are what we call kind of the shallow circuits in that they don't have a lot of gates in our own shores is actually a pretty deep one as well which kind of to your point is if you can run shores maybe can run anything at that point.

[00:06:18] That means we have solved a lot of those issues in the hardware.

[00:06:22] Yeah, actually I think it's a more general statement.

[00:06:27] The fact that when you look at all the algorithm with a promise of a super polynomial or an exponential speed up they tend to be pretty deep in terms of circuit size.

[00:06:42] And the mindset we're in at Alice and Bob is really to say well if you want to perform classical computers because those are already so fast you're going to need a super polynomial speed up I mean when you compare to modern GPUs for example they're just able to run so many operation per second it's it's staggering.

[00:07:06] So you need the that kind of speed up this you need to be able to run deep circuits and this is why we're trying to build a full turn on computer so that we can get our rights that are in the range then to the minus eight and and beyond.

[00:07:25] No, that's the next one point actually because sometimes when we talk about growers algorithm and search you know a lot of people actually just compare that to the worst case classical search and don't think about the GPUs and other more efficient algorithms that we do have a classical systems and then that speed up doesn't become that impressive.

[00:07:44] Yeah, I mean with just a quadratic speed up you need your problem to be extremely big to actually start to kick in in terms of speed up so yeah it's not that realistic maybe longer term.

[00:08:02] At some point when the quantum computers would be more mature they be able to take on a wider class of algorithm and actually grow up and the phase amplification underlying it is a is a pretty general method I mean it can be applied to a very wide variety of use case but because it just took water out of speed up it's not sufficient to justify building a machine as of today.

[00:08:30] Whereas use cases with super put in the speed up like chemistry or for sure even HL even though some conditions apply there and those are the one that drives us.

[00:08:47] So when people talk about shores algorithm they're usually talking about the factoring algorithm but there's his other algorithm we published in the same year for the street log.

[00:08:59] Which has relevance to electric breaking elliptic curve cryptography so I wonder if you could just sort of give an explanation for a audience how elliptic curve encryption works and why his algorithm is able to crack that.

[00:09:15] That's kind of a curve ball because this one is deaf so yeah so indeed actually so maybe to say it's simply so shore algorithm was designed to factor large number into into crimes or I think that's how I learned it.

[00:09:40] And this is very useful when you want to break RSA the the Stolars cryptography but when you look at the underlying sub algorithm what is under the hood of trial algorithm especially the Fourier transform.

[00:09:56] Well it also applies to some other types of cryptography called elliptic curves which I mean I'm not the expert here won't be very good at explaining it simply but it relates to a pretty complex math problem where you're trying to find roots of an equation.

[00:10:20] And this is the same kind of sub rootings from a quantum point of view the thing that is fun it that electric curve cryptography is much harder for classical computer to break than RSA.

[00:10:36] And this for example in applications like Bitcoin where it's used they tend to use keys that are much smaller because when you try to think on the of classical attacks you don't need such long keys but from a quantum computer point of view those two are completely equivalent.

[00:10:59] So actually elliptic curves with today's keys size will be broken a bit earlier than the RSA.

[00:11:09] So I mean you've spoken about how shores algorithm is a good benchmarking algorithm as a generic representative of what a quantum technology is capable of doing what do you see as the fundamental roadblocks generally speaking and in particular what are you doing to address it with the approaches that you're using.

[00:11:28] Yeah the challenge is the the size and the accuracy required to implement that.

[00:11:36] And so something that is known by the physicist community for quite some time now it the tremendous amount of resources required to correct for quantum errors meaning that if you start with slightly faulty quantum bits and you want to get some very good quantum bits that we're using.

[00:11:57] We called logical qubits that are composite system made out of several physical one bits well with the standard approach to you need typically a ratio of 1000 physical on bit per logical one.

[00:12:13] And when you look at shore algorithm and you put all the numbers well the study that made by Google back in 2019 I think showed that it would take about 20 million physical quantum bits to be able to run shore algorithm and break RSA to 40.

[00:12:34] And this is kind of the main challenge here because I mean today IBM is leading the path with about 1000 physical so we're missing four orders of magnitude here to get there now.

[00:12:51] This is not the definite answer to the problem and Alison Bob is pioneering a new path simplification of this.

[00:13:00] The idea is to say that when you look at those 20 million required actually only 0.03% of the qubits are here to compute meaning the remaining 99.97% are overhead.

[00:13:16] What we do at Alison Bob is that we make those qubits compute and so what we've done in the past is that we design another type of qubit called the cat qubit that corrects half of its errors natively fewer hardwired feedback loop.

[00:13:35] And by doing so we gain a square root ratio on the overhead so we're basically left to 30 to 1 instead of 1000 to 1.

[00:13:45] And so what we showed is that if you take all the algorithm together you end up requiring only 350,000 instead of 20 million so you gain a 60 fold improvement and this is already very impactful because the cost and the complexity of the machine is somewhat direct.

[00:14:04] It's somewhat directly proportional to the number of qubits so when you gain a 60 fold improvement while you slash the price of the machine by 98%.

[00:14:14] So it's pretty significant but still I mean it's still 350,000 qubits so we're not there yet and the last result where we just published a few days ago should have to keep improving upon that.

[00:14:30] So we gain another almost 4 fold when now slightly below 100,000 qubits required.

[00:14:40] Yeah so this is a great segue into the paper and like I mentioned we have been discussing this paper a lot on this podcast it's one of our favorite topics to talk about because it really kind of changed the game this year in in accelerating I think for while we thought that want to winter was kind of here but everyone was just going to be able to do that.

[00:14:59] But everyone was just quiet really working on great things so the paper is called computing 256 bit elliptic curve logarithm in 9 hours with 126,133 cat qubits.

[00:15:12] So we mentioned these a little before but can you go more in depth what are cat qubits?

[00:15:17] Yeah so cat qubits a new kind of super connecting qubits in the implementation and then they're kind of a new paradigm because when you think of the standard super connecting qubit called the transmon it's basically the most simple qubit you could think of.

[00:15:41] So we have a very good communication with the capacitor and the Josef Shangrong and Zatze.

[00:15:46] It's a highly harmonic system and the only thing you can do to improve its performance is work on material science I mean increase the quality of your fabrication.

[00:15:57] And what we change at the mindset when designing cat qubit that we realized over the past decade that correcting for hours means basically finding a way to extract entropy from the system continuously without extracting information without creating the coherence on your system.

[00:16:20] Cat qubits is a system it's actually a slightly in an almost harmonic oscillator so an LC resonator, a non dinner if you may.

[00:16:31] Where we engineer a coupling with its environment so that we can continuously inject energy and extract entropy from the system without leaking information.

[00:16:44] And so by doing so basically we're inventing the static RAM, the SRAM of quantum computer in the center just like the SRAM it's slightly more complex than the dynamic RAM which is just a capacitor but it's always powered always stable and it's perfectly meant for compute.

[00:17:07] So cat qubits I mentioned a non dinner in which through this mechanism we stabilize two possible states of this antenna, two possible state of the same amplitude but opposite phases.

[00:17:23] Just like if you would take a pendulum and drive it vertically at twice its frequency, you will see that you would stabilize two possible positions of this pendulum actually the maps is exactly equivalent.

[00:17:39] And what we showed experimentally and that was actually the beginning of Alice and Bob in 2020 in nature physics.

[00:17:47] And now we improve the panda pretty significantly is that the bit flip time, the time between two errors of bit flip in this qubit increases exponentially with the amplitude of this oscillation with the actual number of microwave photons we put in the room.

[00:18:08] And it does so only at the linear cost on the other type of errors called phase.

[00:18:15] So basically to wrap up cat qubits it's a driven qubit that is able to increase exponentially its bit flip time at the linear cost on phase.

[00:18:26] Without any pre factor so it's perfectly meant for fault tolerance but it's terribly bad at niscat noise intermediate scale quantum computing.

[00:18:36] So cat qubits I mean this the name is inherited from of course the famous shooting at a paracadal concept and there the ideas you've got big physical systems in super position with one another.

[00:18:51] And this is I take it how the name comes about in this case to we've got a super position of two quantum states but large as far as qubits normally go.

[00:19:02] So when you when you learn quantum physics the first thing they teach was when you have a super position of large things it doesn't work right deco here's.

[00:19:12] But here it seems that we're getting the opposite effect you've got a super position of two relatively large or not quite macroscopic but not as small as usual either and yet it's giving us all this error reduction how does that come about.

[00:19:27] Yeah it's it's a very good thing question I mean we could talk from there for hours because yeah it's an interesting challenge.

[00:19:39] So indeed in our typical cat qubits we each qubit is comprising of about 10 photon let's say in terms of orders of magnitude so quite a few I mean excitation compared to your transplant that is on your box either containing zero or one pattern but in some sense this is this is pretty general meaning when you want to correct for errors.

[00:20:08] And this is very general you need redundancy actually from a maths point of view you need a larger Hilbert space that you that you burn most of it to stabilize a subspace of a sub manifold of dimension two.

[00:20:26] So so in when you think of the standard approach to quantum error fraction while the surface code is somewhat of the same you use plenty of of two level systems that you use in harmony together to create this logical qubits is two level system but at the end of the day it's also a large macroscopic system with plenty of excitation in it.

[00:20:52] But the difference in cat qubits is that it's extremely hard way efficient meaning you get all that redundancy directly from a single component on your chip because this component is a box that can contain many photons.

[00:21:09] So it's yeah it's kind of the surprise of quantum error correction is that indeed you can get larger system that are more stable than a smaller system.

[00:21:21] I mean there are always trade-offs whenever we talk about any sort of error corrections the trade-off so if you've got a superposition of whatever it is a large number of photons then you lose one of them boom flip the phase like do logical bit flip right so there's a trade-off in some other degree of freedom inevitably.

[00:21:38] Yeah exactly that's the that's exactly and that's a very good point so that what I mentioned by saying there is a cost but the cost is only linear in the number of photons see the for for a cat qubit to suffer a phase trip indeed it only takes the loss of the gain of one photon and one can very easily guess that this is proportional to the number of photons involved in the cat qubit.

[00:22:08] So as you increase the size of the number of photon you're going to increase linearly without any pre factor your your phase trip break but since you're gaining exponentially on the bit flip and this can be understood as kind of two potential whale that gets further and further apart for about a year that grows with the size of the system.

[00:22:33] And if you've done a bit of quantum mechanics you know that the probability to tunnel through this bar here reduces exponentially with the size of the bar here so so this is how you get this rate of now for experts are there the approach we do Alice and Bob by correcting bit flips first and then phase trip through redundancy is kind of a bacon sure like approach so it's a it's a type of a

[00:23:03] code that corrects one errors and then the other and this cannot get you to arbitrarily low error rate there is an optimum a maximum limit to that but what is good for us is that this optimum is fine or I mean it's in 10 to the minus 8 or 10 to the

[00:23:22] minus 9 error rate with reasonable hardware hypothesis so indeed that there is a cost to pay but we believe it's a it's a very reasonable approach at least for the first 10 to 20 years of full tournaments.

[00:23:40] So I know one of the really notable advantages of cat cubits is that you can use them to exploit the fact that noise usually has a bias one type of noise is more common than another and whereas with cubits you can't keep your code immune to the bias and do processing with it with the cat cubits you can so it preserves the noise bias.

[00:24:10] Because you're working with this higher dimensional thing using many photons.

[00:24:16] But I was wondering like I know is that some of the earlier work at least from a few years ago seem to show that the state preparation wasn't so great and also measurements also weren't so great.

[00:24:30] With some of these codes can you comment on on some of your strategies to improve those numbers.

[00:24:37] So so maybe to comment on the first on the noise bias and then come to this panel states preparation and measurement.

[00:24:47] The noise bias indeed I mean you have plenty of cubits that exhibit noise bias meaning I estimate between a bit flips and face people to one and two for five.

[00:25:00] And the thing that is pretty unique with cat cubits is the fact that we have a gate set that allows us to do bias preserving gates meaning that you can understand this pretty simply.

[00:25:19] If each of my cat cubits exhibit let's say no bit flips but only face trips and I only later on correct for face trip.

[00:25:29] I should not at any point from my algorithm be able to convert a face trip into a bit flip because I'm never going to catch it up later on.

[00:25:39] So I need each and every of my gates at the physical level to preserve this bias.

[00:25:47] Which for example prevent us from implementing at the physical level gates like the Haramar gate because by definition this would convert a bit flip into a face trip and vice versa.

[00:26:00] What we am I mean at the beginning of a lesson about what started the company the fact that we already had an architecture with bias preserving gates that let us build logical cubit and then at the logical cubit level have a universal set of gate.

[00:26:16] So we have a set of gate that we can do with the two fully plus some measurement and preparation and the signal gate.

[00:26:27] And with that you can reconstruct and compile the Haramar gate but since you are at the logical level where you corrected bit flip and face trip and you're no longer somewhat noise bias, then you're good to go.

[00:26:39] Yeah pretty pretty exciting actually we are in our last papers this summer we show at least on single cubit gate. We haven't published yet on our two cubit gates that those gates are indeed bias preserving meaning the probability of having a bit flips during a single cubit gate is down to something like 10 to the minus six also or maybe or 10 to the minus seven I don't remember exactly.

[00:27:03] It's very slow it's very small where the probability of having a face trip is of a few percent now regarding the state preparation and measurement well it's it really depends because again it's biased so if you want to prepare zero or one or if you want to measure along the z axis meaning whether you're in zero or one this you can because you have such a great bit flip.

[00:27:33] So you can get as many times as you want I mean you can you can be extremely accurate there now we don't have a Haramar gate and we need to measure and prepare along the x axis as well.

[00:27:48] Yeah this is where we're working pretty pretty strongly because what we've showed it that it's this is directly proportional to the actually the square root of the strength of our autonomous stabilization so the more non linear system is the more able to autonomously correct its errors through this engineered dissipation the better our gates and we're actually pretty close to the fresh.

[00:28:18] So we're actually going to be able to do a little bit more of the natural of quantum error correction and we believe our latest chip helium one with 16 get will be under the threshold for our project in good I mean listeners can go and read the block post by Sebastian as well chief of experiment that puts everything in perspective here about the challenge basically we're not that far from fresh.

[00:28:46] So how much closer do you think this approach takes us to getting to something large and scalable like implementation of cracking elliptic curve compared to all the other mainstream architectures that we're familiar with.

[00:29:00] So the way I like to to view it currently is to say I mean cracking elliptic curves is a bit further down the road the question for me is when do we start out performing classical computers at the wide variety of tasks that is not just buzzing sampling or simulating a quantum system with a quantum system.

[00:29:24] I mean it's kind of a totological problem. It's a bit of a treat we say it simulates itself.

[00:29:34] Yeah exactly so that's not very impactful for me the right target the moment where quantum computers start to be in orbit if you want to build the analogy with them with the space race is when you get about I mean between 50 and 100

[00:29:53] logical qubits with their rate that are low enough meaning lower than 10 to the minus six at least and this is when you can start running algorithm for the physics for example for simulating spin chains or thing like that when you can when you have a machine that becomes relevant for researchers and scientists out there.

[00:30:17] And so what we showed with our latest LDPC plus cat architecture so low density by the check that I mentioned before where we gain another four fault improvement is that we could build 100 logical qubits with their rate of 10 to the minus eight with as few as 1500 cats and 1500 cats is something IBM is close to be able to manufacture.

[00:30:43] So our internal roadmap for example aims at at delivering such a machine in 2028 and from there while it becomes commercially viable and you can scale from there and get to elliptic curves I don't know a few years later I guess but with our novel architecture I guess elliptic curves would take something like 40,000 cats to run and sure slightly below 100,000.

[00:31:13] For RSA so can you go a little bit more in depth you you kind of alluded to this new paper that was just released a few days ago maybe a week ago so you can can talk a little bit more about the key findings of that and what helps you kind of move to this next stage of development.

[00:31:31] So the.

[00:31:33] Our previous architecture was extremely naive in some sense I mean you quite natively for bit flips and then you use the most naive error correcting code for correct classical one because you only have one type of errors that is the repetition code just basically a chain of cats where you do a majority vote to correct for errors now we knew.

[00:31:56] Quite some time that there are more advanced quantum or I mean classical error correcting code and that can be applied here.

[00:32:06] So we worked on the type of codes that are called load and city by to check that are used I mean that have been invented in the 60s and that are used in plenty of today's technology like 5G or why if I believe.

[00:32:25] Sorry and and the finding that that we're I mean something we realized is that combining cats with LDPC codes works tremendously well because people have been looking at LDPC codes for quantum computing for quite some time there's a beautiful paper by IBM released last summer on that.

[00:32:50] And the thing is in the paper that the promises are great but the engineering challenges pretty daunting because it takes a special kind of connectivity you need to be able to connect to bits that are pretty far apart from one another so this complexify the design of your chip quite significantly plus it's not very clear how you can run gates on such code.

[00:33:17] So it's I think their paper is titled on a memory and not quantum processor here what we showed is that because cat qubits only live one type of error you can use not a quantum LDPC code but a classical LDPC code.

[00:33:35] And this lifts quite a few no go for your M of math and more precisely we show in this paper that we can achieve not only a 4 fold gain on overhead but also be able to run gates and also do so while only requiring local.

[00:33:55] I mean nearest neighbor or next nearest neighbor at most connectivity which is something that is currently pretty mundane into pep connecting circuits so the finding of this theory paper that we have an architecture.

[00:34:12] And feasible with today's technology that opens the path to get to to those machine of 100 logical to be 10 more pretty soon.

[00:34:26] A little detail question on this because I find it's very interesting using the LDPC codes so many of them have a structure that has a symmetry under transition.

[00:34:42] So the way the code is structured and so that like sort of a natural for a system with periodic boundaries.

[00:34:50] So how can you get around that issue with your architecture do you have a way to work around for that?

[00:34:57] This might be slightly beyond my ability and I'm going to need the team here something I can come on town it the fact that actually the type of because LDPC code is actually a family of codes.

[00:35:11] It's not one given but the one that work best on our architecture now actually called a subfamily of cellular automata.

[00:35:24] So basically small packets of our operating code that you propagate on it so this kind of echoes this as this translational invariance you mentioned and indeed there are boundaries condition but here you're going to need to invite my chief of theory Jeremy to get a more detailed answer.

[00:35:50] It's quite a fascinating result. It's nice to see that some of the really well studied advances in classical computing actually can be made to work in a quantum regime by using some very clever ideas like you've managed to use this larger space of photons to encode quantum information.

[00:36:18] And yeah that's quite nice.

[00:36:21] Yeah now the challenge is on the experiment side because having such a nice architecture is a driver for the team.

[00:36:31] Now it's great promises that we will be able to claim victory once the chip is actually running that arrow cracking code.

[00:36:43] So currently we're working on a prototype of logical tribit still using a repetition code with a 16 cat called helium one that will be communicating more about this year and then we'll move on to a bit more than 40 where we'll be able to have a very long live logical tribit and also start demonstrating the LDPC plus cat architecture.

[00:37:11] And later on move to large scale machine that can actually deliver impact on real use cases.

[00:37:19] I just curious also have you thought about using your architecture for material design.

[00:37:31] So maybe either synthesizing new materials or are running simulations that are just beyond what classical computers can do but also aren't so complicated that you would need hundreds of thousands of qubits.

[00:37:49] Yeah so the my understanding of those algorithm actually the team is currently working on the detailed resource estimation there to provide numbers but my understanding for material science and chemistry in general is that indeed you need fewer qubits and it start to kk in at a few hundreds of logical qubits.

[00:38:10] But the challenge at the moment where we are that you need even better fidelity.

[00:38:18] To deliver impact you need them yeah, fidelity that are pretty good in 10 to the minus 11 or more the difference with an algorithm like sure is that here you can actually average in some sense.

[00:38:37] And so you can have a trade off between time and space in some sense.

[00:38:45] So this obviously is not a favorable skating but it can let you grab the last few orders of magnitude that you're missing.

[00:38:54] What do you see is the single most important issue that quantum computer scientists need to be working to overcome for the future of the field.

[00:39:04] I guess my understanding is you're focusing on their correction to a large extent in a pretty novel way.

[00:39:10] What do you see is the really big issues?

[00:39:14] Yeah, so they are something that Google and us start to encounter because when you get to longer lifetime actually what you have is a more sensitive probe.

[00:39:28] And so you start probing events that are more rare.

[00:39:34] And more precisely one of the challenge that we might encounter still a bit soon to say but at least this is part of what gives me a bad night.

[00:39:46] Are the source of correlated errors and here's kind of story time so I don't know if you remember the catastrophe of Fukushima.

[00:39:56] At Fukushima they had four water from for the primary secret.

[00:40:03] And so this answered redundancy and they said well the probability of one failure is low but the probability of all four pumps failing is extremely extremely low so we're good to go.

[00:40:15] But the problem with that all the four pumps were in the same room and when the tsunami hit the whole room was flooded.

[00:40:24] So you had a correlated error here and this led to the catastrophe.

[00:40:30] And this is true for every code out there they are meant to correct for uncorrelated errors.

[00:40:39] So once you have a source let's say imagine a mefer write that hits my chip I'm going to have a correlated error on all my qubits and I'm not going to be able to correct that.

[00:40:54] In our case it's not matter of right it's likely to be a high energy physics like a mune or thing like that that hit the chip every 10 seconds or 100 seconds and create phonons all over the chip which leads to correlated errors.

[00:41:12] So this is something we're working on we have quite a few ideas on how to correct that but yeah as you as you start getting to longer life time longer, longer coherence time.

[00:41:22] You start encountering novel physics, rarer events and those will need dedicated engineering tricks to circumvent the rate.

[00:41:34] So you've alluded a little bit to your roadmap and timelines but I want to ask you the million dollar question we always ask is how far away are we from seeing quantum computers actually solving a real world problem that a classical computer can't.

[00:41:49] And also on your end what you think that will be when we can actually break encryption let's say you know in a reasonable amount of time before the heat death of the universe.

[00:42:00] So today I'm pretty confident that quantum computer will become useful in a regular base on a regular basis before the end of the decade.

[00:42:14] Meaning as soon as you start having those hundreds of logical qubits with sufficiently lower rates and I mean we're well positioned but we're not the only one on that race.

[00:42:25] I mean you have some terrific results by pure R for example and they have their own engineering challenges but I mean you have improvement across the board of platforms for quantum computing.

[00:42:41] So it seems unlikely that will stumble completely and won't be able to run that so for the double negative here but you see my point we target 2028 for such a machine but if we're a bit late I'm pretty sure it's still going to be before the end of the decade.

[00:43:00] Now to break encryption it's a it's a bigger problem but if you assume sort of a more slow of the number of qubits.

[00:43:10] That should start falling in I don't know 2033 2035 something like that.

[00:43:19] So what are you envision as the next big breakthrough in quantum computing or you think should be the next big breakthrough?

[00:43:29] Yeah so two types of answers there I think we're still missing the sputnik moment of quantum computing but this is not the breakthrough I want to answer but first let me say a word about this sputnik moment.

[00:43:45] Sputnik was a was a useless milestone in some sense historically I mean it's a machine that just went out there and beeped in space for three months also but it's signal that mankind had made the machine that could escape gravity at least for some time.

[00:44:08] And that triggered the beginning of the space era in the 60s.

[00:44:14] Now for quantum computing we're still missing the sputnik moment in the sense escaping the coherent having a man made machine that is completely decoupled from the rest of the universe that has no causal link with our class you all will.

[00:44:34] So proper logical qubit and that would be a strong trigger now this is going to happen in the next couple of years I'm pretty sure.

[00:44:45] But this is not really a breakthrough in terms of something we haven't for seen what I'm hoping for and I think there's plenty of room for that is a completely new class of quantum algorithm.

[00:44:59] And this is at the moment we're still relying on the same handful of quantum algorithm I think that there was just one new that appeared in 2022 but with no clear application yet.

[00:45:13] But when you think about it there has been so few people working on designing novel quantum algorithm compared to classical algorithm.

[00:45:23] And it's a ton of room for brilliant ideas in quantum computing here.

[00:45:30] And I'm yeah, I think we could be very surprised to see novel algorithm come forward and again if you do the historical parallel and you think you know this famous quote by the CEO of IBM in the 40s that there is maybe a world market for five computers.

[00:45:51] It's kind of the same here because if you vote about computing in the 40s who are saying well I want to break an in mom and I want to decipher codes and I want to compute trajectories for ballistic basically.

[00:46:07] And you didn't have that many more ideas on what to do with a computer history prove them so wrong.

[00:46:15] And I think it's going to repeat itself with quantum computing looking back at your own history what what spent the biggest unexpected thing you've come across.

[00:46:25] I mean the trajectory of a listen more than myself is a is the story of so many good stars aligning I mean it's a it takes quite a bit of luck for this to happen.

[00:46:43] Not not saying that we have succeeded at anything but at least to to create this opportunity.

[00:46:49] So what surprised me the most is the level of support and engagement and how many of the physicists that were surrounding Rafael and myself while we were PhD student that jumped in the adventure with us.

[00:47:06] Not only the researchers but also the institution the academic labs even at the political level all being so supportive that was pretty surprising.

[00:47:17] I mean I expected the tougher time hiring the team initially but we ended up with a pretty nice line of a physicist and we managed to get however everyone on board.

[00:47:31] So our listeners are academics in the quantum space but we also have a lot of people who are students that are really interested in getting into quantum computing or people from different domains that are thinking that quantum is going to be really important to their careers.

[00:47:45] So what advice would you give to someone starting out in the quantum journey?

[00:47:52] I think it's about reading a lot and reading a lot of very different things not only scientific papers but also get a sense of the dynamic of the whole ecosystem.

[00:48:06] And this helped me quite a lot at the beginning of my career when I was still a student just to actually I got in one on computing by reading IEEE spectrum at the time and seeing the first you know pop science article about quantum computing and seeing the momentum grow.

[00:48:28] I was the Nobel Prize of Sergei Roche and David Noemain on in 2012 and kind of understanding the momentum and the dynamic.

[00:48:39] What I was a student there in 2012, 2013 actually my peers were all learning about machine learning and they are now starting companies like mistral in France.

[00:48:52] I think now is the right time to get into quantum to be in five to seven years.

[00:48:58] The one starting the quantum mistral company of the quantum of NEI in some sense.

[00:49:05] And just to finish up can you share maybe your most memorable moments while working with Alice and Bob?

[00:49:16] I mean it's going so fast the the plenty of great times but I think what moved me the most was this September when we had an internal conference for a little bit by an off site.

[00:49:38] And this is when I realized that Alice and Bob was no longer just my project with Raphael where now 80 people so many things happening seeing all those projects internally that some of them I was not completely aware of but surprised by plenty of results.

[00:49:58] Yeah, I mean it was extremely moving to see this group, this company becoming something somewhat independent of myself and yeah that was moving to see the collective and the team play there as well.

[00:50:19] Great.

[00:50:21] So thank you Theo for coming on the podcast where can we find more information about your papers publications, your personal social media if you have that.

[00:50:31] Yeah, you can follow us on LinkedIn, Twitter or X should I say and the website is Alice dash Bob dot com and we publish everything we do on archive.

[00:50:45] So yeah, no secreties there going happen.

[00:50:48] Get your bits are very new and buzzing system in general there is plenty of room for research and don't hesitate to reach out happy to collaborate.

[00:50:59] Great and for listeners thank you for listening to this episode if you are on YouTube make sure to comment down below if you have any questions.

[00:51:07] Any guests you like to see any follow ups we can bring people back make sure to like and subscribe on Apple or wherever you get your podcasts and thank you so much Peter and Gavin for joining as well.

[00:51:19] Thank you, thanks.

[00:51:21] Thanks so much for having me.