Latest / Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™ / Quantum Computing's Software Bottleneck — Haiqu's Mykola Myksymenko

Transcript

Quantum Computing's Software Bottleneck — Haiqu's Mykola Myksymenko · Deep Tech Germany – AI, Robotics & Frontier Innovation by Startuprad.io™

0:00
Quantum computing is no longer waiting on physics.
0:04
It's waiting on software to stop lying.
0:08
Quantum computing has raised billions, produced
0:11
headlines, and delivered almost no durable utility.
0:15
Hardware improves, claims escalate, Results
0:19
remain narrow. My guest today operates in the one layer
0:23
where quantum computing either becomes infrastructure
0:27
or quietly fails. Mykola
0:30
Myksymenko is building the software stack that
0:33
decides whether today's quantum machines ever
0:37
matter.
0:44
Welcome to Startuprad IO,
0:47
your podcast and YouTube blog covering the German
0:51
startup scene which with news, interviews and
0:54
live events.
0:58
My guest is Mykola Myksymenko, co founder and
1:02
CTO of HiQ. Trained in
1:05
theoretical physics at the Max Planck Institute for Complex Systems
1:09
at Dresden, University of Gottingen, University of Magdeburg and
1:13
the Weizmann Institute. He later led large scale
1:16
industrial R and D before founding HiQ.
1:20
His company focuses on a problem the industry prefers to
1:24
avoid. Quantum hardware exists, but the
1:27
execution layer required for real applications does
1:31
not. HiQ builds hardware, wear middleware
1:35
designed to minimize noise, extend
1:39
usable circuit depth, and make current
1:42
quantum processors economically relevant.
1:46
The company has already worked with IBM, Airbus, BMW,
1:50
HSBC and Capgemini, released open source
1:53
tooling adopted by the ecosystem, and recently
1:57
raised an $11 million seed round while pushing
2:00
its first product to market. Welcome Mykola
2:06
hi Joe, it's my pleasure to be on this podcast and
2:09
it's my pleasure to have made it through the intro without buttering your family
2:13
name. Your early
2:17
work in theoretical physics trained you to think in
2:21
terms of fundamental limits rather than incremental
2:24
improvements. How did that mindset shape
2:28
the way you look at computing problems today?
2:32
I think important moment that we are living
2:36
right now is that there are many
2:39
exponential technologies which kind of intercept
2:43
and accelerate the pace of technology development
2:47
that we are facing today. So on one hand there is AI that
2:51
everyone is talking about, but on the other hand there is kind of underlying
2:54
current which is quantum computing. And
2:59
as a scientist you tend to understand like
3:02
what are the kind of fundamental processes which might drive some
3:06
processes around us. And that's what helps also
3:10
to understand how these different currents actually
3:13
can accelerate each other. And while quantum computing was
3:17
seemingly too far away just like around a decade
3:21
ago, today we see it already like
3:24
giving some value in various
3:28
scientific applications and we see how much it
3:32
accelerated over just like last two years. And
3:36
while being in this community and in this
3:42
scientific endeavor for a while, I
3:46
can already kind of predict that within next years we
3:49
will see quite a lot of progress in quantum computing. Such
3:53
as we may not even predict that in the incremental way.
4:01
You then move from academia. We already talked about it
4:05
into industry R and D leadership. Where did you
4:08
first see the gap between impressive research result
4:12
and systems that actually work in production?
4:17
Little disclaimer, the systems never work in production
4:21
as they normally should. That's true.
4:24
But one thing that kind of made me
4:28
shift from academia to industry was particularly that I wanted to
4:32
work on something which is immediately relevant or immediately
4:36
can be applied to some hard problems in real life
4:40
and in academia. My background was in
4:44
quantum condensed metaphysics. I worked a lot on like predicting
4:48
and trying to manipulate
4:52
quantum states of matter in order to
4:56
discover new materials or new properties of materials. And
5:00
while it's very fascinating and very interesting field of research, it
5:04
also has quite long horizon in terms of
5:07
applicability. So some application of what I was
5:11
doing may happen,
5:15
for example in real materials or be
5:18
discovered in real materials in tens or
5:22
twenties years from now. And then I
5:25
shifted to industry where the things were on
5:29
completely different pace. And here what
5:33
made me to make this shift was particularly the progress of
5:36
AI. So while working on the
5:40
complex, so called the field of complex physics,
5:44
of physics of complex systems, I was also interested
5:48
in the physics of neural networks. And apparently a lot of
5:51
physics theory and methods can be applied into like
5:55
theory of learning. That what made me
6:00
being able to naturally transit into that
6:04
field and made quite actually a good career in industry
6:08
where we built actual production grade system based
6:12
on the state of the art research which might just appeared in the literature.
6:15
And we picked it up and built some systems around that.
6:21
That was actually very exciting. And I'm still
6:24
kind of using that on one hand, fundamental
6:28
understanding of the processes and interest in fundamental science.
6:32
But at the same time in industry I got a lot of these
6:36
interesting methodologies and approaches which allow
6:40
now build very quick
6:43
and rapid iterations of research
6:47
in the startup environment such that we can actually quickly arrive
6:51
into some relevant results which can be
6:54
applied already today to some specific problems.
7:00
I see you said you had you built quite a career
7:04
in the industry. So when did you decide to
7:08
find high Q? You factually bet
7:11
that software and not the hardware would decide the fate of
7:15
quantum computing. What convinced you that this was
7:19
the layer where real leverage actually sits?
7:24
Yeah, that's exciting story. So
7:28
as I shifted to industry, I was always fascinated about
7:31
quantum computing. Quantum computing is something that I did my
7:35
master's degree. So I started with quantum computing, then
7:38
shifted into condensed metaphysics, still worked a lot on non
7:42
equilibrium physics and methods of simulating quantum matter
7:47
and quantum states and then worked
7:51
on AI for a bit. And what I seen in AI was very similar
7:54
to what I see today in what happens in quantum computing.
7:59
So we have systems which
8:02
already are capable of doing interesting results
8:07
in the very narrow domain. So, for example, you can apply quantum
8:11
computers, modern quantum computers, to
8:15
specific simulations of states of matter
8:18
or some specific optimization problems which might be
8:22
very hard or nearly impossible to solve on the
8:25
classical computer. However, there are still classical
8:29
counterparts, but in some cases, solving
8:32
some of these problems on the quantum computer can be already faster and
8:36
sometimes even more convenient. And
8:40
that's exactly what I seen in the beginning or
8:43
early stages of deep learning, where these systems were already
8:47
discovered. And people seen a lot of interesting, fascinating applications in
8:51
computer vision in like image
8:54
recognition or signal processing.
8:58
However, it took around five to 10 years before these
9:02
systems became like massively productized. And
9:07
one application for that and one background
9:11
process which contributed
9:15
to that embrace of technology was particularly
9:19
building a bunch of the infrastructure
9:23
around the research artifacts, for
9:26
example, like PyTorch system or TensorFlow systems,
9:31
which allowed a lot of people to build these complex
9:34
and more and more complex neural
9:38
networks easier and train them at scale,
9:42
reproduce the results at scale, such that the whole research community
9:46
accelerated dramatically. So we kind of lacked
9:49
that in quantum computing. So running something on the quantum computer two
9:53
years ago was very hard. Like people experimented
9:57
literally with like small prototypes of 4 to 10 qubits.
10:01
Today we already see community which
10:05
runs applications on hundreds of qubits and hundreds of qubits.
10:09
This is already a scale which is often hard to
10:12
reproduce classically. So you need to be an expert in numerical
10:16
techniques in classical high performance computing in order to
10:19
reproduce those results. So it went from
10:23
extremely toy problems to already more or less
10:27
state of the art high performance computing. So I'm very optimistic about where
10:30
it all goes. And I think
10:34
this intuition of combining different experience
10:38
between fields helps me today to shape our research agenda
10:41
and prioritizing what to focus on.
10:45
Out of curiosity, you've been talking about that quantum
10:49
computing can now deliver what high performance computing
10:53
as state of the art computers can deliver right now. So
10:57
I was wondering, if you extrapolate when
11:00
quantum computing has its full capability,
11:04
what could this mean? Especially everybody is
11:07
talking right now about the energy usage
11:11
of all those data centers that the big
11:15
AI companies or the hyperscalers in general are right
11:18
now building, what could that mean? Would it mean like
11:22
100% more computing power
11:26
per what or where would quantum computing go
11:30
there? And then can you give us a very tiny idea if
11:34
you have the current LLM models running on a much more
11:37
capable quantum computer? Yeah,
11:41
that's probably something which is important to dig in. So
11:45
many people see quantum computers as just like another GPU
11:49
or another large computer which runs
11:52
bigger and faster models. This is a little bit
11:56
different story. So quantum computers are different kinds of
12:00
computations and it's suitable
12:03
perfectly to some specific kinds of problems. And it can
12:07
be completely useless to different problems. And running classical
12:10
computations on the quantum computer, it's not the way
12:14
forward. So it's not enough just like to take your favorite
12:18
classical computing algorithm and
12:22
run it on the quantum computer. You need to completely
12:26
rethink the problem and rethink the mathematical approach to your
12:30
algorithm to actually use all the
12:34
capabilities of
12:38
quantum bits there and then tunnelement
12:41
the ability to parallelize those computations
12:45
using quantum computers. So this is different way of
12:49
computations. And that's importantly to
12:52
realize that it's probably not immediately
12:57
applicable to running large LLMs or
13:00
running like conventional
13:03
machine learning applications. Actually, if you think about
13:07
applying big data problems, one of the
13:10
crucial bottleneck in quantum computing is actually loading data
13:14
in the quantum computer. So one of the breakthroughs
13:18
that we did at Haiku was particularly
13:21
unlocking this capability of loading large scale
13:26
data sets, which are of the industrial
13:30
scale number of features into quantum computers. Before that it was
13:34
just impossible. And
13:39
there are like small, small improvements here and there which
13:44
probably enlarge the number of problems that we can
13:47
solve with quantum computers. But before, before
13:51
we go into like ubiquitous era,
13:55
we will be focused first on the very narrow
13:58
problems. And some of them are designing new
14:02
molecules and simulating new molecules.
14:06
That's a very hard problem for classical computers.
14:10
But this is something where quantum computer can
14:14
actually help because this is a natural system to tackle with
14:17
qubits, it maps almost one to one.
14:22
We can think about the problem that I was solving in my
14:26
scientific career like a problem of materials. Like if you want
14:30
to simulate superconductivity or
14:33
magnetism in some specific new
14:37
types of metals, then
14:41
quantum computers can potentially help with these problems.
14:45
Or computational fluid dynamics. That's another
14:49
interesting problem which is very
14:52
widespread and industry. So companies
14:56
in aerospace or automotive run a lot of
14:59
computational fluid dynamics simulations and
15:03
you very quickly actually kind of hit the glass
15:06
ceiling there with a grid
15:10
that, that you can simulate on the classical computers. While
15:14
on the quantum computers theoretically you can go to much
15:17
more denser grid of your simulation and have
15:21
much more precise results. However, it's very hard right now to
15:25
map this algorithm directly. And we also working on that
15:29
helping some companies who are partners to
15:32
actually make this transition. So there are A bunch of
15:36
problems like this optimization simulation, natural
15:39
systems and multiphysics
15:43
simulations. These are probably those which will be
15:47
on the kind of early, early low hanging
15:51
fruits, horizon
15:55
and middle to long term. What would it mean to everybody who's now
15:59
thinking about investing in data centers who
16:03
are already there, announced in vast,
16:06
vast numbers? Do you think that'll change
16:10
something over time when
16:14
quantum computing really becomes productive, especially in
16:18
terms of space needed and energy for
16:21
computing of some form? Maybe still LLMs
16:26
of artificial intelligence. Look,
16:30
even from the current perspective,
16:34
if you want to simulate some process in nature, like
16:38
for example, a molecule, you can do that on
16:42
the classical computer to some extent. And I used to
16:46
work, for example, on some
16:50
of the largest supercomputing centers in Germany.
16:53
And those simulations cost
16:57
a lot. They can cost easily like $200,000.
17:02
And in order to simulate some material or some
17:06
specific physical effects while.
17:09
And you can run those simulations sometimes for weeks in order
17:13
to get some reasonable result. And we already
17:16
witness some interesting types of these
17:20
simulations where we can simulate the same system,
17:24
still simulatable classically, but we can simulate it already today on the
17:27
quantum computer, and we can get results maybe in
17:31
10 minutes. So even from that
17:35
perspective, like running large
17:37
supercomputing cluster for weeks, or
17:42
running a quantum computer for a few minutes, I can already
17:46
hint that you actually probably spent much less energy
17:50
on that. When we will be able
17:54
to run machine learning simulations at scale, I would say
17:57
that that probably will transit also to machine learning
18:01
simulations. I need to be honest that
18:05
machine learning is still hard to solve
18:08
directly on the quantum computer. But there are very specific
18:11
problems, for example, anomaly detection, which
18:15
potentially quantum computers can actually solve more
18:20
better than classical computers. And we already
18:24
see few of those examples here and there.
18:29
But ubiquitous machine learning, where we,
18:32
for example, apply it to LLMs, it's
18:36
probably some years away from now.
18:41
Some years. Okay, I see. Let's get back to
18:45
the original questions I envisioned for this interview. For
18:49
founders watching quantum from the outside,
18:52
noise, not qubit count, is the real reason
18:56
most pilots fail quietly. What actually
19:00
happens inside these systems
19:03
that makes noise such a dominant constraint?
19:09
Yeah, so you should think about quantum computing as
19:14
basically early stages of classical computers. Like
19:18
try to remember what happened in classical computing
19:22
in between 40s and 60s. So these
19:26
were like, these were
19:29
large machines. Like sometimes they were like of the size of a room
19:33
and they were very hard to operate. They did
19:37
not have error correction at the time. So
19:41
basically they were operating with valves. And those
19:44
valves can very easily overheat or become malfunctioning.
19:49
So Literally the first bugs were in those
19:52
computers, just because some things did not function well. So you
19:56
should be a high level hardware operator
20:00
and understand the low level algorithmic theory in
20:03
order to run something on these machines.
20:07
And that was for a while before error correction was
20:11
introduced to these machines. And today we are in
20:14
very similar era in the quantum computers. So these
20:18
machines already exist, they have hundreds of qubits,
20:21
hundreds of qubits are hard to simulate. And we already have
20:25
evidences from Google, from IBM and other companies
20:29
that you can simulate on these machines some very
20:33
complicated states of matter
20:36
or quantum states, which are
20:40
impossible or will take extremely long time
20:43
to simulate classically. So
20:47
we have those evidences. Now the question is
20:51
how we can use these machines for something practical. And that's where
20:55
the bottleneck is. Because in order to run something on the
20:59
quantum computers, you need to be
21:03
very much an expert in how to fight noise and
21:06
imperfections in these machines. They are not perfect, and
21:10
in order to apply error correction to them, you need thousands of qubits,
21:13
and we are not yet in that regime, and better fidelity of those
21:17
qubits. So we are in the low qubit regimes and noisy machines.
21:22
But apparently
21:26
even in this regime you can extract something useful.
21:30
And in order to do that, a lot of expertise is needed.
21:33
What we try to introduce is
21:37
to some extent democratization of ability
21:41
for a wider community of scientists,
21:45
of algorithm researchers, of material scientists,
21:49
quantum chemists use quantum computers for
21:53
their own simulations and try to discover something
21:56
useful. And in order to do that, we abstracted away
22:00
all those low level manipulations with
22:05
noise, with improvement of performance,
22:08
with reducing the impact of noise on the algorithm
22:12
performance, such that scientists can purely focus on science
22:16
and on the mapping their problem to the quantum algorithm,
22:20
and they can completely forget about what happens
22:23
under that layer. And we think that that is
22:27
crucial for today, because today we need more and more scientists
22:31
to work with quantum computers and discover those applications. The more people
22:34
will run on quantum computers, the sooner we actually will
22:38
discover userful applications, as we did with classical
22:42
computing, as we did with deep
22:45
learning applications and other use cases. You
22:49
often say the quantum software stack
22:52
simply doesn't exist. Yet when companies start
22:56
experimenting with quantum today, where do they usually
22:59
underestimate the integration complexity?
23:03
Yeah. So when we interact with
23:07
industries, what we frequently see that
23:11
many companies start setting up the quantum programs.
23:16
They understand that this is a technology of the future, and
23:19
sooner or later this will dramatically transform
23:23
of what they do in many of the heavy high
23:27
performance computing simulations.
23:31
And in practice, this is not
23:34
only about building Software. It's also about deep,
23:39
low level expertise in happening something to run
23:43
on the real hardware. And that's where we see real gap.
23:47
So I know few
23:50
companies who actually have some expertise in
23:54
running simulations on real hardware at
23:59
the state of the art scale.
24:02
Most of the simulations which people do are of the
24:06
toy scale. But you can run more. And
24:10
in order to do so you need deep expertise
24:16
in, for example, tailoring noise such that you
24:20
can use various
24:24
symmetric properties in order to simplify the noise
24:28
that is in the hardware. Then you can do a
24:32
bunch of tricks in order to suppress its effect on the
24:35
algorithmic performance. You can also work on the
24:38
algorithm itself in order to reduce the number of operations
24:42
that operate in that algorithm and drive
24:46
you to the resulting state. So
24:50
all of these tricks are complex in nature.
24:54
And frankly speaking, you need to be expert in just doing
24:57
this for some very short period of time
25:01
while these computers are in this state of
25:06
performance. And I don't
25:10
think anyone in industry should actually
25:14
spend significant amount of time
25:18
focusing on this low level optimization. So this should be done
25:22
automatically. And that can be done by a
25:26
middleware stack. And in quantum computing it's almost
25:29
non existent. So if you think about again, like parallel
25:33
to classical computers, we have seen
25:36
software driving the adoption
25:40
revolution several times already. For example, when we had the
25:43
first personal computers, this kind of
25:47
middleware revolution was happening because of adoption of
25:51
BASIC language. So that was introduced
25:55
by today's Microsoft, such that
25:59
more and more quantum enthusiasts were able to actually program
26:03
computers in a much easier way than they did before.
26:08
Then the same kind of revolution happens recently
26:11
with introducing of GPUs
26:15
and CUDA language or
26:19
framework to program those GPUs.
26:23
So in this case, again like a middleware stack,
26:26
in this case, CUDA helped a lot of scientists
26:30
and experts in AI, for example, to use
26:34
GPUs for their computational needs.
26:37
And currently we are in this very same kind
26:41
of stage with quantum computers. So right now I would say
26:45
there is no CUDA or no BASIC for
26:49
quantum computers. It's still very hard to program them. You literally work
26:52
on the quantum assembly level. And it takes
26:56
a lot of effort to run something and we want to abstract that
27:00
away. Giving people instruments and tools such
27:03
that they can think about the problem, can think about like what would
27:07
be the actual use of
27:11
quantum computers in their particular problems, or how to decompose or
27:15
change the algorithm itself, rather than thinking how
27:18
to optimize it to run on the hardware so that that we can take and
27:22
abstract away from the user,
27:26
we're talking middleware here that sits
27:30
exactly in a layer where responsibility shifts from
27:34
the hardware isn't ready to the execution model is wrong.
27:38
What does that layer actually do
27:42
inside the quantum workflow?
27:45
Maybe you can dumb it down for non physics majors.
27:50
Right, so imagine that we have hardware and apparently a
27:54
lot of hardware providers, they don't focus that much on the software
27:58
layer just because the hardware is so hard to build.
28:02
There are so many problems that you need to be laser focused in order to
28:05
make your qubits better, more stable,
28:10
scale them to a larger quantity and so on. So there are a lot
28:14
of challenges there in hardware stack and then there
28:18
are people who worked on application sites. So we still need to
28:21
discover a lot of quantum algorithms in a lot of different
28:26
algorithmic areas and then understand which
28:30
actual real life problems map to those quantum algorithms.
28:34
So there are just a handful of those today. So I would say there
28:38
are two communities. One is low level on the hardware, another community
28:41
works more on algorithmic side and they
28:45
do not overlap that much. So that where
28:49
we sit. So it's not only Weeb, there are
28:52
other companies working in this space, but the focus is to
28:56
provide the middle layer between low level
29:01
hardware which operates with pulses and low level
29:05
assembly languages, quantum assembly languages
29:09
and higher level abstractions for algorithm
29:12
definition. And that's apparently not
29:16
that easy. And in the future we need to understand that the whole
29:20
stack is being changing as well at the same
29:23
time, because as computers get better,
29:27
we introduce new and new
29:30
paradigms. For example, in few years from now we
29:34
will have full tolerant quantum
29:37
computers, so they will operate with error
29:41
correction loops in the middle in order to
29:45
first encode more qubits in a single
29:48
logical qubit and then correct errors if those appear.
29:53
So that's already another kind of
29:56
algorithmic subroutine which need
30:00
to be somehow integrated in this low level stack.
30:05
And so in order to do so with many of these components,
30:08
we need some kind of orchestration
30:12
layer which will orchestrate when and which tools we should
30:15
use. There should be
30:19
ability to run on one hand real time
30:24
error mitigation and error suppression, on the other hand
30:28
error correction cycles and finally some
30:32
finalized post processing of still logical errors
30:36
which might still happen in those computers. So there
30:40
are many, many complexity points here
30:45
and this becomes more and more complex actually field, but
30:49
the stack itself kind of grows in the real time right now
30:53
and today it's not that complex. But
30:57
I imagine that in 10 years from now that
31:01
will be very similar to what we see in classical computers today, which
31:04
is very complex Actually middleware stack, which we even don't know,
31:08
like typical programmers don't even know what's happening to their
31:12
algorithm after they compile. I had to smile
31:16
because there was a setback for the IBM stock price a few days
31:19
ago that because
31:24
Claude the AI model could start working
31:27
COBOL code. And so I was wondering at one point
31:31
in 10 years if there's somebody who's trying to get COBOL running
31:35
on quantum computing. So every.
31:38
But my understanding is basically every system, every software out
31:42
there that is running on the current
31:46
silicon based wafer technology
31:50
will need to be redone for the quantum computing age.
31:56
Not really. So we don't need to push every
31:59
software to the quantum computers. So not all of the problems
32:03
need quantum computers to be solved.
32:06
So these problems that are
32:11
mostly required quantum computers are those which are
32:16
complex in nature, which require enormous
32:19
amount of memory in order to run some transformations
32:23
in data structures or enormous amount of compute.
32:29
And these are specific classes of problem
32:33
and other problems like for example, you don't need
32:37
quantum computers on your phone or like a quantum browser or
32:40
whatever. So this is completely different
32:44
space. And in principle, in the future we will see
32:48
basically these two technologies cooperate together. And we already
32:51
have those hybrid quantum classical algorithms where you
32:55
have some subroutines running on the classical computer,
32:58
while those more complex algorithms
33:02
can be decomposed and run separately on
33:06
the, on the quantum computer where you get the result and then integrated back into
33:10
your classical workflow. So these, these technologies will
33:13
cooperate. So I would say that that's the
33:17
future that we will see. For example, like on the phone, every day you
33:21
call your Uber. Imagine when you do that, you solve the
33:25
complex optimization problem such that your driver
33:28
needs to navigate the shortest path to your location.
33:32
In the future, if quantum computers will be ubiquitous, that
33:36
tiny optimization problem will probably be delegated
33:40
to a quantum computer to solve maybe fraction of a second faster.
33:46
Okay, but
33:49
we took a little detour. My frequent listeners already know that.
33:55
But we're still talking middleware here. And by
33:59
operating in this layer, haiku forces
34:03
customers to confront whether they are buying, learning,
34:06
signaling or real performance. What
34:09
kinds of motivations do you actually see
34:13
when companies start exploring quantum
34:17
computing? Right. The
34:21
challenge in any of the exponential technology
34:25
is that initially it seems
34:29
like nothing happens. The progress is so slow that it
34:33
seems like that the real value will be years ahead and you don't need to
34:36
invest to it today. So you can just sit and observe
34:40
what happens. On the other hand,
34:44
the nature of exponential technologies says
34:48
that at some point you have a Very
34:51
rapid growth of technological adoption and
34:56
capabilities of this technology. And I would say
34:59
that's what everyone sees in AI today,
35:03
but that what we will see to start happening in quantum very
35:07
soon. And the reason for that is that
35:12
it's not. First of all we have a pace of technology which is
35:15
accelerating. And as being in this part of the industry I literally
35:19
see that three years ago it was almost impossible to run
35:23
anything. So there were just few people, handful of people who were
35:26
running something on the hardware. Today there are many
35:30
people in the world who run on the IBM's largest
35:34
machines for free via cloud.
35:38
And today you can immediately access
35:42
largest scale quantum computers in the cloud.
35:46
And that's still
35:50
not enough to immediately give you the business value.
35:53
But that's something, if you see this
35:57
perspective, that's already something quite impressive.
36:01
Another thing that we see from our perspective
36:05
just few years ago,
36:08
running something on these largest machines could cost you
36:12
easily tens to hundred thousand
36:16
dollars if you want to get truly
36:19
interesting new results. With the
36:23
help of technologies like ours, we can
36:26
reduce this cost to tens of dollars or up
36:30
to hundreds of dollars. Such that we dramatically
36:33
democratize the level of
36:37
accessibility of these technologies. Such that now any student can
36:40
afford running so called utility scale
36:44
experiments on the quantum
36:47
computers available to him or her today.
36:52
And that again accelerates everything. Because now there will be more
36:56
experiments, now there will be more discoveries, now there will be more
36:59
algorithms and the whole cycle accelerates. And we will
37:03
see the adoption getting faster and faster and faster.
37:07
At the same time there is a hardware progress which we see in parallel
37:11
which is again moving pretty fast.
37:15
Companies actually hit their milestones in
37:19
different qubit modalities. In superconducting
37:23
qubits, we see growing number of
37:26
superconducting qubits grids. We see error
37:30
correction codes already applied on the superconducting
37:33
computers. We see trapped ion computers actually
37:37
also growing in number of qubits and in the speed
37:41
of running operations on these computers.
37:44
Neutral site atom computers also move forward very fast, like
37:48
soon. There are a few companies who promise photonic
37:52
quantum computers. So there are a lot of different
37:56
kind of underlying currents which accelerate the
38:00
whole pace of this technological progress. So
38:04
I'm very positive and like very optimistic about the future of quantum
38:07
computers. Me too. But would have also
38:11
understood is that many current quantum
38:15
initiatives may be premature.
38:19
From your perspective, when should companies pause
38:23
experimentation rather than accelerate it?
38:28
Right. So I would say I probably
38:31
had to say it in the previous question, but
38:35
one of important aspect, like why it's important to do something today.
38:40
Suppose tomorrow there are quantum computers which operate at
38:43
scale. We can run any
38:46
algorithm we wish. The challenge would be
38:51
to discover those algorithms. There are just handful. And
38:55
not every problem is so easily mappable to quantum
38:59
algorithm. So it will take another few years to just
39:02
discover algorithms for their particular business problems.
39:07
Then it takes typically in enterprises, it takes few years in
39:11
order to adopt those algorithms into
39:14
enterprise software ecosystem, integrate them, test
39:18
them, make sure that they produce the right
39:21
values in some specific toic conditions.
39:26
So like stress test them in different scenarios. So
39:29
all that takes typically time. For example, in finance, if you want to
39:33
implement a new algorithm which would
39:37
help, for example in some decision
39:41
making, the roadmap to
39:45
implement this algorithm can take sometimes few years.
39:49
That's why it's important to start developing these algorithms today. Imagine
39:53
that your competitor already did it today. And then
39:57
in the year from now, quantum computers suddenly
40:00
start to showcase value. So you still will need
40:04
to to take few years in order to build, test,
40:08
integrate the workflow in your organization. So you will be
40:12
already losing. And since it's exponential technology, you will be losing
40:15
a lot. So it's very similar to what we see today
40:19
in AI. The same story can repeat itself
40:24
in quantum computing. So stay in the game.
40:28
It may cost you, but if you're not part of the race, it
40:32
may cost you more. Okay, this is the
40:35
point where founders decide whether quantum computing is a
40:39
distraction or an execution problem they will
40:42
eventually have to face. For Bayer
40:46
evaluating Haikyuu, the real value is not
40:50
speed ups in isolation, but making
40:53
quantum experiments behave less like bespoke research
40:57
projects. What changes when your software
41:01
sits in the execution pipeline? Yeah, so
41:04
when, when we integrate our software, the very first effort that you
41:08
see, suddenly you can run much more experiments for the same cost.
41:13
So typical enterprise programs are not that
41:16
expensive on the scale. So these are typically few people
41:21
and few hundred thousand dollars, up to a million could be
41:24
in compute costs. But that limits you
41:28
in the number of experiments that you can run
41:33
and various use cases that you can check and
41:37
empirically experiment with. So with
41:40
our software we can not speed up, but
41:44
reduce the cost of those experiments by
41:47
partially also speeding them up in
41:52
the factor of ten or even hundreds.
41:56
And that basically leads you to a situation when
42:00
for the same budget you can run much more, and
42:04
that can much faster bring you to
42:08
better intuition, better understanding of the technology,
42:13
and better and better
42:16
algorithmic pipeline for your specific business
42:20
use cases. So it's not immediately that you will solve
42:24
all the problems, but you will get much more
42:27
intuition and operational knowledge. Of
42:31
how to run something on the quantum computers. And hopefully
42:36
with few customers with whom we are working today, we also
42:39
will bring them to the edge of quantum
42:44
usability and userfulness when they can start up operating
42:47
these systems along with classical high performance
42:51
computing systems.
42:55
You framed the buying decision around reducing the
42:59
total time of learning rather than just the costs per
43:03
run. How should companies think about the
43:06
economics of experimenting with quantum computing
43:10
today? It's quite expensive.
43:16
We don't think about the total cost. But
43:20
what other economics? Future payouts, future
43:23
risks, current risk, future payouts. Right. So it
43:27
seemingly running something on quantum computers seems like a lot, right?
43:31
So like one hour of execution on the quantum computer can
43:35
cost up to $50,000 or even more, depending
43:39
on the technology and the company
43:43
who is providing that machine to you. But in
43:47
practice, the time
43:51
spent on the machine and time spent for example in high
43:55
performance computing clusters, these are different types.
43:59
And for example, some problems you
44:03
can still solve on the quantum computer, like let's say in one
44:07
hour or two hours. So it seemingly costs a lot,
44:11
but these problems might be completely impossible to solve
44:14
on the classical computer. And for example, if you're a
44:18
chemistry company and you need to create new
44:23
molecular structures
44:27
to have specific properties, you need to run those simulations
44:31
daily and a lot of them and
44:35
eventually you are literally limited by the size of molecules that
44:38
you can simulate. And if quantum computers will
44:42
be of a particularly good quality, you will be able to
44:46
simulate them in minutes or maybe hours
44:50
instead of weeks or months on the classical
44:53
computer. Or sometimes you cannot do that at all on the
44:57
classical computer. So it's eventually we will just bring
45:01
get to the economy where you either can run something or you
45:05
can't. So that's, it's not even like
45:09
in terms of economics, it's just like there are some applications
45:13
which are just impossible to will be impossible to run classically.
45:17
You're doing very, very great. But I'm afraid we are way
45:21
over time here. So I think we are now running around
45:25
30, 35 to 40 minutes in maybe
45:29
45 minutes in total. But usually that's the length of
45:33
one of our interviews. So I'm afraid we
45:37
do a little cutting here and go into part number two.
45:45
That's all folks. Find more news streams,
45:49
events and
45:50
interviews@www.startuprad
45:54
IO. Remember, sharing is caring.
46:03
Sat.