Shai Machnes Machine learning for quantum computing

Show video

hi everyone and welcome to a new  interview in today's episode that   we recorded in the early half of July  2023 I'm talking to Shai Machnes.   Shai is the CEO of qruise and we're going to  partially focus on the work that they're doing   with that company we're also going to focus on  the intersection between Quantum Computing and   machine learning so the intersection between  these two fields is where Shai is focusing his   research and we're going to talk about upcoming  technologies on the machine learning side and how   they extend to quantum computing architectures.  So let's get started with that discussion. Hey Shai, thanks for joining. My pleasure.   Today I figured we could kick this discussion off  by talking about a concept that you've articulated   in the past namely an AI physicist. So you're  working in the intersection I guess you could  

say between machine learning and Quantum Computing  and there's a concept that you've not necessarily   coined but at least stated about AI physicists.  So I would just like to pick your brain a little   bit on what you think about when you when you  say that concept what the ambition is over the   long term of how we can integrate machine  learning with Quantum Computing in general.   Right, so you can do it in one of two ways at  the moment. Well, at the moment there is one way.  

There'll probably be two others going forward.  So right now quantum computers aren't very good   so what you can do is you can use machine learning  to help improve and design quantum computers   which is what our company does uh once we get  better quantum computers one can use quantum   computers to do machine learning and one of  the machine learning things you can do on   future quantum computers is improved quantum  computers right so the so if today to do what   qruise does you need to classically simulate  Quantum devices in the future you could do   the simulation on another quantum device so it  would be a like a bootstrapping process um just   like you know people are now using output from  chat GPT to train other machine learning models um but right now quantum computers aren't very  useful yet they're mostly I would categorize   them as engineering prototypes meaning they're  good enough for physicist Engineers to learn how   to build quantum computers how to use them  Etc um and you know maybe even The Wider   you know potential users and things like  that but you can't run anything useful on   a quantum computer that you won't rather run  on a you know on a classical computer not yet   hmm so for now doing classical machine learning to  help improve quantum computers is what we do okay   so other types of devices like Quantum sensors  and it's a pretty General field before we focus   on that in more detail let's maybe talk about the  other approach first so you you mentioned this   um sort of cyclical behavior and  how you can make these enhancements   is the key reason why you would want to use  quantum computers to then enhance your quantum   computers in the end that you have a much more  natural overlap in terms of the physics associated   with them is that the uh quantum computers will  do best is simulate quantum physics because the   let's see the the memory requirements for  simulating quantum physics is kind of exponential   in the number of qubits or degrees of freedom  let's say of the quantum system uh which is   partly why for example we don't fully understand  high temperature superconductors and we don't have   room temperature superconductors yet because these  are kind of deeply quantum materials and even if   we can write down the equations for you know  one atom and one electron and how they interact   in these materials you have a large number of  electrons interacting with a large number of   atoms and we simply can't solve these equations  not analytically but not even numerically um   quantum computers should be able to do that pretty  directly because it's you know it's a natural fit but as I mentioned not quite yet  right besides the let's say specific   material considerations that quantum  computers might help simulate   are there General so I don't want to get into the  details of qruise yet but this sort of comes into   like almost to the domain of qruise when you  think about optimizing a Quantum system a lot   of the developments that need to happen as I  understand with um especially superconducting   circuits is you want to construct better models  of your Quantum device is is the bottleneck there   right now that you're not able to simulate  Quantum like systems well enough or is it   some other bottleneck I'm thinking like to what  extent having quantum computers in this space   actually helps resolve those kinds of problems  so the problems we're having with Quantum are not   oh they say they are about Quantum but also you  can put them contextualize them in a in a bigger   framework um when you're developing physical  devices that are really at The Cutting Edge um you don't fully understand what's  going on in there so the you know there   is something that's called the forward  forward problem in engineering which is   you know creating a a accurate simulation of  something so that if you you know if you did   your simulation right and you didn't mess up the  construction the device will behave as indicated   by the simulation okay and there are fields of  engineering where this works great right if you   design a house in the appropriate software and  you complete you know compute the stresses and the   loads and everything and and you know you don't  kind of cheat on the materials but you actually   build it it'll be stable just like the simulation  predict and there are other fields where   um let's say electric motors right the  the technology is quite well understood   so I can simulate an electric motor in  detail and what the simulation indicates   would be the performance very likely will  actually be the performance in practice   um but there are Technologies  where this is not the case   um for example Quantum Technologies but but not  only for example um let's say um silicon photonics   so there are chips that have kind of photonic  Pathways that are fabricated kind of generally   in the same kind of lithographic  approaches as you would make chips but Humanity in 2023 cannot control the optical  length of these Pathways to sufficient accuracy   because for the system to behave the way we  wanted to we need to control the kind of the   lengths of these things to within a small  fraction of a wavelength and we don't know   how to do that yet so you fabricate the device  but the the phase difference between the two   kind of Pathways is kind of random why because we  don't have sufficiently good manufacturing control   um and and therefore you need like once  the device comes out of manufacturing you   need per device to actually measure  what's going on and make Corrections um and in Quantum it's partially that um  we have let's say in superconducting qubits   um kind of at the bottom of the qubit is  a very small super sensitive uh component   that's based on quantum tunneling known as the  Josephine Junction Quantum tunneling has kind   of exponential sensitivity to kind of changes in  the in distance and potential and other things   so you know a few atoms move to the left to the  right and now this thing behaves differently   we don't know how to make things with Atomic  accuracy right Humanity hasn't figured that out   yet so we we make qubits and they don't come out  exactly the way we planned but on top of that you   have other issues so you control these qubits with  electromagnetic fields which are from these kind   of boxes of very high-end electronics that can  manipulate fields at the nanosecond time scale   but like everything nothing's perfect so on these  controls you know you get I don't know 12 bits of   accuracy in the voltage right but the 13th qubit  is randomish but it's actually even kind of worse   because the way it's random is problematic and  um you you need to start accounting for it and   measuring exactly how much noise you have and  how much crosstalk you have between all the   wires and so the the the distance between what  you design and what you are able to fabricate   there's some Gap there and that's what limiting  the performance of quantum computers so we need   tools to kind of in some sense reverse engineer  the devices we we ourselves are building because   we plan them to some accuracy we fabricate them to  to the best knowledge of humanity you know today   and yet that gap between what  we actually get and what we plan   is kind of the pain point for performance of  the device and then having tools to kind of   understand in great detail what's going on and  you know given you know you could have for a   certain Quantum operation you could easily have  between five and ten different sources of noise   simultaneously and to be able to understand  which of these is actually the pain point   which of these is actually the limiting  factor which tells you what you need to   focus on when you're developing the next iteration  of hardware this whole thing is really really hard but it's it's not limited to Quantum  and in Quantum Computing the equations   that describe the Dynamics of the  system are the Schrodinger equation but in other systems it could be you know electromagnetic waves or I don't know the propagation of   light in matter which is an electromagnetic equation  but in a different frequency range or other things   you could have phonons which are essentially slight vibrations that also affect things   um and sometimes it's really silly things like the 50 hertz of the electrical signal from the wall   somehow weakly propagating through all the devices and all the electronics and and it's it's tiny it's like a 0.1 percent or whatever but when you're trying to get to four and five   nines accuracy we're trying to get a  quantum operation to be 99.999% accurate   right you get like the tenth of a  thousandth of the noise from somewhere and and that's what the field is wrestling with and everybody's wrestling with it in different   you know different Hardwares different attempts to get there one thing I want to linger on for a   second that that struck me was you talked about josephson junctions and the way those are being   fabricated you said that depending on where the atoms actually end up in the fabrication process   that will have quite significant effects on the actual um well model or the actual product that   you get in the end is there also a Time Dimension to that I'm thinking the architecture as such or   the material as such ones that has been fabricated is that does that have a tendency to shift over   time let's say yeah so it's not just so it's not just that you need to characterize all the   noises you have right but it's also really on the material Level you have changes it's not   just on the material Level so I'll tell you  something we experienced in the last 48 hours   okay uh in in the last 48 hours we're working on this uh superconducting quantum computer   and the lab had a glitch in the cooling mechanism  so if the chip is usually 300th of a degree above   absolute zero it went all the way up to one  degree above absolute zero right that's still   a lot colder than outer space right it's  really really cold but it wasn't enough   that first nothing worked and then once it kind  of it took it takes hours to cool it down back   um atoms shifted because of this higher  temperature and where they settled wasn't   exactly the same place they were a previous so we  had to so things kind of work but the accuracy was   very bad so we we had to re-measure all the  qubit frequencies and the all the parameters   of the chip because everything drifted because  you heat it up you you but not all the   way up to room temperature like from you know 0.03  above absolute zero to one degree of absolute zero  and that was still enough to mess things up  so these things are kind of painfully sensitive   um there are other Technologies by  the way that are far less sensitive   so okay when you say would you say  superconducting architectures are the   most sensitive to these kinds of of things  or when you say other architectures   that are less sensitive what are you referring to  there so to to kind of take the The Other Extreme   um NV-centers meaning you take nanodiamonds  or artificially manufactured diamonds and   you shoot nitrogen atoms in them uh sometimes  the nitrogen atom will kick out a carbon atom   and that creates like a Quantum system that  you can use as a qubit it's a bit a bit more   elaborate but generally speaking uh this is valid  now this system is quantum at room temperature   right yeah so you don't need to cool it  at all and and for example trapped ions   uh you know some most variants don't need  cooling either but they need excellent vacuum   however it's not that hard to build a hundred  qubit superconducting quantum computer because if   you know how to build 10 going from 10 to 100 is  with superconducting it's mostly more of the same   it's an engineering challenge that's so much  scientific but with trapped ions and NV- centers and others there are these inherent limits that  make it really hard to scale beyond a certain size sorry for interrupting what would you say some  different advantages some scale better some   are more sensitive some are this some of that  so it's not clear that's why people are trying   multiple directions because for every technology  people are working on You Can Count advantages and   disadvantages and it's not clear you know where  we'll make the Breakthrough mm-hmm can we can we   stay on uh trapped Ion on computers for just a  second and particularly focus on the scaling up   of Trapped iron quantum computers compared to  superconducting ones is your view that uh that   would be sort of easier to do I mean this is all  maybe in sort of parenthesis because it's hard to   predict what roadblocks would potentially be ahead  but your general view of these two architectures   comparing them would you say there are more like  significant roadblocks for let's say scaling up   trapped ion quantum computers as opposed to  superconducting ones because the reason why   I'm asking this is I I actually talked to someone  before who was under the impression that scaling   up trapped ion quantum computers could actually  be um quite replicable as well because you have   similar like because you circumvent some of the  fabrication issues that you have with for example   uh the josephson junctions if you take that like  the atoms that you use in a trapped iron quantum   computer are identical so that would potentially  be an argument for scaling it up right so um the the atoms are identical right so so but  the environment in which they are held the   magnetic fields the potentials used  to control them are not exactly identical   um and we need everything to be identical right  the other is that superconducting qubits are   fabricated with let's say lithography type  techniques right and we know how to make chips   with a billion transistors today so from that  perspective the the let's say the chip part it's a little bit easier to scale  however if you think of Trapped ions um so people used to do linear traps which was kind   of you have range magnetic fields so  the old ions come like sit in a row that Rainer black used to kind of spearhead  that he got to 14 qubits with a lot of effort   and then he basically stopped because  you couldn't scale bigger than that since then people have come up with  Solutions segmented perhaps or ideas but the so I'll say this if you have a 10 qubit   quantum like superconducting quantum chip going  to 100 is more of the same you're going to have   a lot of problems with wires and it gets a little  big and there's crosstalk it's not trivial but   you don't need kind of the fundamental change of  architecture to get to 100 trapped ions you need   to start playing around with the architecture  because you can't do it in one simple track   so people do uh their ideas where the the  ions themselves move and you move them around   but then you're you're looking at uh something  that you know maybe a little bit like a pinball   machine that has a hundred balls and you  you need to be very careful with all of   them because if you lose one the quantum State  goes kind of you you open up the box and the   cat either dies or meows but you know it's  not a quantum cat anymore mm-hmm so so in in   some sense you can think of trapped irons as  this is kind of really nanomechanical system   so that comes with its own difficulties that said  the these systems can be very very well isolated   so let's say as far as I know the records for  their best single and two qubit gates are all   in trapped ion systems but the biggest systems  are uh superconducting and not trapped ions now then again you have neutral atoms reedberg  systems where they can do larger number of atoms   but then you you have some difficulty  with controlling individual atoms   you can do kind of global operations but  it's more difficult to kind of do individual   Gates let's call it and as a result these  systems are very well suited for simulation   but for let's say what I would call  digital Quantum circuits it's a bit harder   so again there are advantages and disadvantages  to every approach um Intel is working on the on   the quantum dots in Silicon idea which is  even closer to the kind of lithographic   um you know chip industry approach to things and  of course there are also difficulties there so   and I think that the largest they got is somewhere  between 10 and 20. um okay so it do the reason you   have people working on so many directions and  and very smart people is that there's no clear   winner at this problem that's right so you know  today every chip you can find with really tiny uh   exceptions are all kind of mosfet transistors you  know one technology took over the world right yeah   we're not there in Quantum we haven't found  our transistor the thing you can replicate   like a billion times and and you  have multiple Labs commercial academic   exploring you know multiple candidates and trying  to come up with new candidates all the time   because all the options we have are not that great  all really really hard to kind of push forward   now you know there are so many smart people working  on this I wouldn't be surprised if somebody   figures something out and and suddenly we have a  leap um on the other hand you know there are also   a lot of smart people working on Fusion and there  are a lot of approaches to do fusion and fusion is   you know five years away for the last 50 years so  it's extremely hard to predict it's extremely hard one thought I had and we could use this maybe  to to tailor the discussion a bit to machine   learning as well um when you mentioned before  the incident today like with the superconducting   quantum computer and the refrigerator that wasn't  quite at the right temperature are you seeing any   sort of tools that go beyond just characterizing  the architectures in a better way but actually   help with these kinds of everyday considerations  for running these systems I'm thinking that AI I   mean obviously the sort of the the natural thing  your mind tends towards when you hear that term   AI combined with with quantum physics is you  can use it to characterize your systems better   get sort of better performance but in terms of  helping out like the all the surrounding tasks   that need to be done are you seeing right now  in in the academical space that there are lots   of tools popping up for those kinds of things  or is that still lagging behind a little bit um it very much depends at which level you're looking at so  there's a lot of activity in machine   learning for science in general there's a lot  of activity in machine learning for anything   whatever you want to put there's a lot of activity  there right now um but it hasn't coalesced into   you know kind of large like large very powerful  mature software packages or something like that the if you want let's say   simulations for Quantum systems on the simpler  side you have packages like QuTiP an open   source package it's really really good right  but it kind of only goes so far which is fine um there aren't really you know for error  mitigation which is this idea that you can   you have a noisy quantum computer and if you take  not just run the original calculation that you   wanted but you modified in in really smart ways  and you run multiple calculations you can kind of   uh figure out from the various results what would  the result would have been if there was no noise   so there's a very powerful  package called medic to do that   um do there are other parts but there's no kind  of a comprehensive solution okay understood one   other development I mean that's been really  coming along heavily is all the GPT models   oh yeah have you seen any applications of that in  the the overall space that you're in I'm thinking   that that to me doesn't come very naturally like  how you would actually apply a technology like   that to a space like Quantum Computing perhaps  I'm wrong there yeah not yet right I mean Chat GPT today is like um let's say a very  Junior programmer yeah so you can tell it   to read the Qiskit so Qiskit for example a  very uh very nice package from IBM open source   to do things having to do with compilation and  searches it's a good collection very useful thing   you can tell Chat GPT okay um read the  documentation for Qiskit and then Implement   for me I don't know the shore algorithm or seven  qubits and it'll figure out how to do that but   it's it's very much a junior and when it comes  to doing things that are let's say research level   it's not and I don't think it'll get there  anytime soon because the way Chat GPT thinks of math   is is Too Human by the way if you ask Chat GPT  to sum up 15 numbers you know and you don't   have the same number of digits or something  it'll almost always give you the wrong match   yeah now it's kind of ridiculous this is a  computer to do the inference that's   the the conversation you're having and it's trying  to add up it's you know utilizing many many many   billions of assembly instructions you know or the  equivalent on a GPU yeah where it could do this   with like five assembly instructions if it felt  like a computer and not like a human um LLMs large   language models were designed to process human  language or process fuzzy reality sort of things   math is not that like there is no fuzziness one  and one is always right there's this joke that   you know you go to an accountant you ask it how  much is one plus one he says well how much do   you need it to be yeah the the with with um with  math and physics when you manipulate algebraic   equations there is no room for fuzziness if you  if you ask Chat GPT to solve a math problem like   a high school or you know undergrad math book  if you just ask for the solution the chance   of getting an error is much greater than if  you ask it to detail all the steps along the way   that's a very human thinking when we train  in universities when we train you know people   coming out of high school and we take them  through undergrad and Graduate Studies in PhD   we train them that in certain things you  should not think like a human you should   try and think like a computer because the  way you know if you have a set of equations   the decision what step to take is perhaps based on  intuition and experience but once you've decided   I'll move the variables from you know the left  side to the right side there's only one correct   way to do that there is no kind of there's no  wiggle room and trying to get and even in high   school you know Elementary School we teach kids to  do arithmetic but when they get to high school we   say no don't do arithmetic by hand takes too much  time and you're more likely to make mistakes this   is a calculator use the calculator and right  now the approach the main approach is to get   is you know there's this shiny new Hammer  these LLMs and they do wonderful things   but not everything is a nail and qruise is going  in a direction of okay if not everything is a   nail what's next and and we have some ideas but  yeah not quite yet ready to to speak about it but   but maybe we can segue two qruise then and speak  generally about the kinds of problems you're   solving with that framework yeah would you just  like to give a quick high level pitch sort of   of what qruise does and then we'll get into some  more details right so so qruise is developing   um essentially machine learning physicists right  if I said it earlier that you know right now   language models aren't very good at uh at math  and physics We believe We Know Why and and the   long-term goal is to to help people in development  Labs to develop new sort of devices   uh by having you know like an infinite number  of Junior physicists working with them 24 / 7.  

now we're not going to replace you know  the the more Senior People anytime soon   uh and just having you know a lot of Junior  people in software doesn't solve all the problems   like you know there's this kind of you know a  thousand monkeys on a thousand keyboards don't   uh generate Shakespeare and you can't take a  thousand PhD students and tell them build me   a quantum computer they need somebody with more  experience to guide them so the software won't   be able to replace the more senior judgment and  strategy but the day-to-day a lot of it can be   done by by software by Machine learning and what  we're focusing right now is um is what's known as   this inverse problem of okay I've built something it  doesn't quite work the way I've planned now I need   to understand why and we we originally built the  software very much for Quantum technology but it   turns out it's a little bit like you know Amazon  started building a bookstore once they had you   know the the web infrastructure and fulfillment  and Logistics all that stuff worked out they   realized that hey I can actually use it almost  directly for selling absolutely everything else we we kind of started out  trying to provide algorithms   to to help quantum computers work better  and understand why they're not perfect   and it turns out that if you replace the  simulation with from schrodinger equation   to the Maxwell equations and suddenly you can  solve very similar problems in many other fields   fluid mechanics or you know whatever it is  because the the the the methodology and ideas   that sit on top of the simulation don't  really care what's inside the simulation they're quite General so the way we train   physicists to kind of think  about how do you analyze data to figure out what's going on this  methodology is quite general not very   dependent on which sub-sub field of physics you're  at right and and and the equivalent algorithms   are similarly not very dependent on  the field of course you need to to   have let's say a very reliable simulation in  whatever field you do but I'll give you an example   let's have a I have a model of this of a system  and I I optimized it to be as close as possible   to the data but one of the parameters in the model  of the system still has a very big uncertainty   why because my data doesn't include sufficient  information to narrow down the uncertainty um you can ask the computer to design a set of  experiments that once executed will give you   information that once analyzed can narrow the uncertainty of this parameter which is extremely   useful and important but all these algorithms  don't really care what sort of physics there is   all they need is that they need their simulation  and they need the simulation to have certain   mathematical properties but once you have that you  can do what's known as Bayesian experiment design   in almost any field of physics so if I understand that correctly you're  saying essentially that you would develop an   algorithm that identifies not only if a system is  underfitted let's say but also what the root cause   like where you would most likely find the root  cause of that underfitting where this model would   need additional parameterization in order to fit  the data in a good way is that roughly the right   way of thinking about it so there are two issues here okay is my model is too simplistic it   doesn't include the right physical phenomena it's  missing some parameters uh and I need to figure   out what to add to the model this at the moment is  beyond the capabilities to do automatically okay   this is where you need kind of physical intuition  Etc and experience and that's really hard   but let's say you have a model that  actually you know doesn't do a very bad job   but you still want to know the value of certain  parameters to high accuracy but you haven't   collected the right data you want the computer  to figure out what's the right data to collect   so that you you can narrow down these parameters  so this we can do automatically the the part of trying to figure  out what to add to the model   we don't have a very smart way of doing it so we  can just try stuff you know like you know throw   stuff against the wall see what sticks right  and if it's a computer it can throw things   really fast so maybe you can make some progress  this way but that's kind of a Brute Force   approach to things and we have ideas of a more  refined approach but we're not there yet   okay when we talk about this being generalizable  as well I don't want to dig in too much into   the details of what fields that would be  generalizable towards but just in terms of   um your where you're coming from where you're  saying it do you mean that it's generalizable   across let's say architectures I would assume  these kinds of approaches are but even more   fundamentally than that perhaps yeah yeah so okay  so so you know we came from Quantum Computing and   Quantum sensor and our academic background  is control of these types of systems um so we had to deal with very different  systems right from room temperature NV-centers to milli Kelvin superconducting  systems and also time scales that   vary from the nanosecond to the micro  sector to the millisecond um this kind of forced us to create  approaches that are actually very flexible   because although we were dealing with always  things that fall under the kind of headline   of quantum Computing or Quantum technology but  if you if you kind of look down into the detail   they're actually very different from each other so  kind of that led us to to seek approaches that are   that are General and are not very specific to the  physics and and once we developed them for Quantum   we realized that hey we made something that's  actually even more flexible than we realized   but the the flexibility stems from the fact that  over the last for me 15 years for the some of   the professors it's more than 25. they've been  dealing with a wide range of physical systems   all of them still under the title of you know  Quantum technology but they're very different   so we had no choice but to to develop flexible  approaches that aren't too specific to the   architecture or to the type of system right  when you're trying to optimize a Quantum gate   you you you're looking for ways to do it  that'll work no matter if it's NV-center   or trapped ion or superconducting and you  end up with this kind of flexible approach   that's actually even more flexible than you  thought and you can take it into other places   okay one consideration when it comes to this  offering and another initiative that you've   been involved in so I know you've been involved in  OpenSuperQ I think that's now actually evolved   to OpenSuperQPlus these days I don't  know if you're still part of that project um okay so are you able to use these Technologies  now that you've developed as part of qruise   so the story of Cruz is kind of  intimately tied to OpenSuperQ and OpenSuperQPlus   okay a lot more of the original OpenSuperQ so we actually started as   an academic open source project that  was developed as part of OpenSuperQ right we needed to provide control algorithms  and you know these sort of things for for open   OpenSuperQ and that that's how we created  the not the methodology the methodology   predated that but the software and then  when we created the qruise the company   uh we essentially forked the open source uh  project and continued developing it as proprietary   because it was MIT license and we could do that  okay and just then to help people with   putting putting the um inventions into context  now in the uh OpenSuperQ project what are you   actually doing that requires the involvement of  the qruise products if you just want to paint a   picture of like what what the target of that  project is and how that's being facilitated   okay so so the idea was that okay you have this  quantum computer every qubit is slightly different   and a-priori you don't even know the exact  parameters of the qubit and yet you want to make   to to create the same Gates the same operations  on all the qubits and and they have to   be very very accurate right but so  you're trying to do as essentially um yeah having kind of think of it as you know you get very a set of  very different cars with different steering   different tires different I don't know what  engines and they need to kind of Traverse this   really complex path and they need to do it  exactly exactly the same to the millimeter   but they're all slightly different I mean they're  all cars but they're all slightly different and we   we had to build algorithms that will you know be able to  do that be able to understand quickly kind of the   particular parameters of each qubit each car and  then figure out how do we need to tweak the the   driving which superconducting qubits are driven  by electromagnetic fields in the gigahertz range   which is why are known as microwaves right then  the macro if you have like at home is like 2.4   gigahertz or something whereabouts which is why  when Wi-Fi was 2.4 gigahertz it failed when the   microwave was on because they're basically using  the same frequencies so superconducting qubits are   driven by by microwaves uh typically not as low  as 2.4 or like five six seven eight gigahertz

something of that nature so we needed to figure out  how to drive the qubits very precisely and to   tailor it for each qubit in each device and if  you heat it up the chip and cool it down again   you needed to do a do-over because you  know the engines are now different and the   tires are all different it's still the same  car but it's not the exactly exactly the same   car and you need to kind of fine tune everything  again and that was that was the role of our team   okay how would you say the resources for a  project a lot like that compared to some of   the other players in the space like one of  the considerations I have is I would assume   that like for example the development of machine  learning models it's quite resource intensive and   the players that have large budgets that they're  able to allocate to these kinds of projects are   able to get a long way especially when it comes  to scaling up the system so like the largest   superconducting quantum computers come from  like Google IBM these these kinds of players   how um well budgeted would you say the product  was from the perspective of actually uh driving   this kind of innovation and how far would  you say you've come so so far in that project so I I think that's the wrong question you  can correct it if you want in Europe   were trying to compete with the big commercial  companies with university-led product projects   and Building 100 cubic point of  computer is a huge engineering feat   and one to which universities  are not well suited in my mind   if I were kind of in charge of project funding  instead of giving you know 50 to 100 million to   a project to build a big quantum computer I will  give you know 10 million or 5 million to a bunch   of universities and tell them to look for better  qubits to build the best 10 qubit chip possible   and then patent it you know and then create a commercial   company to scale it up yeah the  budget you've got in universities isn't as large but there are other issues um European approaches  to distribute to have everything distributed   multiple countries multiple universities Etc  these things benefit from a concentration and in IBM everybody working on  quantum or almost everybody is in   Upstate New York at least for working on the hardware um it's really hard if people are at different  universities and then they have different   motivations and different agendas and everybody  is in multiple projects so they have multiple   priorities and you know they I would  change the way Quantum funding is done in Europe   a lot more small projects to make breakthroughs   in qubit design circuit design gate design that sort  of thing which is needed in the field   and the leave of scaling up to commercial  companies because if you created you know   a 10 qubit chip that has uh you know far  better fidelities than anybody on the market   and you you know and you patented it you'll have  a line of VCS kind of around the block waiting to   give you money and on the other hand and this is  the sort of thing that universities do well you   know go very very deep understand everything you  know to the finest detail come up with new ideas   do essentially technical proof of concept on the  other hand there are things that universities   don't do quite as well which is I think which  is kind of large large projects that are   you know have a very significant engineering  component now sometimes you have no choice right   because if you're trying to build LHC then you  know there's no commercial side to this equation   so you have to do it in an academic environment  but here universities in Europe are competing   against commercial companies in the US right and  and the skill sets are different and the strengths   are different and we kind of we can choose our  fights and I think we're choosing the wrong fight understood so I think that leads us very well  to to the last question that I was going to ask   you which is specifically about what parts  of research are gaining sufficient attention   in your view and which ones would require  more attention and more perhaps funding   from on behalf of the of the University so you  mentioned that your opinion is that a university   should focus for example on qubit design but  do you have any more uh concrete yeah yeah   I have but I want to maybe clarify this is not the University's decision it's the funding  agency's decision because if the funding agencies   are putting out you know 200 million euros  that we're going to fund three collaborations   to build quantum computers we're going to  give each 70 million or something right but   this is what the universities will do and if the  funding agencies said okay we're gonna fund now   25 programs to build better qubits and it's  small it's like five million euro project   uh and and you know you you could have  it everywhere it doesn't matter but   every project is localized so that people  can sit and work together I think overall   you'll get better results but it's not the  University's choice it's the funding agency's choice   that's one another thing is that you ask what's  not covered well and I think machine learning um a person today can finish a PhD in you  know quantum physics you know work on   building quantum computers and never take a single  course in machine learning I think in 2023 that's insane I think machine learning is clearly becoming  a very basic tool right you don't need to to   develop new approaches in machine learning but  just like you can't really do let's say Quantum   technology without knowing how to program  you you can't do Quantum technology without   properly understanding the tools that machine  learning make available the approaches that machine   learning make available you don't need  to design new neural nets but you do need to know   how to kind of utilize existing architectures  in an effective way and you know most okay I   don't know if most but probably  most people finishing PhD in Quantum technology   you know don't go through machine learning  training and I think that's a big mistake good so then more resources to to machine learning   or more prioritization perhaps  towards machine learning at least   we're overtime now and I think we should cut  it here I really enjoyed having the discussion   Shai so thanks a lot for that yeah I hope you  have a great rest of your evening uh you too

2023-09-03

Show video