Quantum Computing's Software Bottleneck — Haiqu's Mykola Myksymenko

WEBVTT
1
00:00:00.240 --> 00:00:04.080
Quantum computing is no longer waiting on physics.
2
00:00:04.320 --> 00:00:07.040
It's waiting on software to stop lying.
3
00:00:08.080 --> 00:00:11.520
Quantum computing has raised billions, produced
4
00:00:11.600 --> 00:00:15.280
headlines, and delivered almost no durable utility.
5
00:00:15.680 --> 00:00:19.360
Hardware improves, claims escalate, Results
6
00:00:19.440 --> 00:00:23.080
remain narrow. My guest today operates in the one layer
7
00:00:23.080 --> 00:00:26.800
where quantum computing either becomes infrastructure
8
00:00:27.190 --> 00:00:29.670
or quietly fails. Mykola
9
00:00:30.070 --> 00:00:33.510
Myksymenko is building the software stack that
10
00:00:33.510 --> 00:00:36.950
decides whether today's quantum machines ever
11
00:00:37.350 --> 00:00:37.750
matter.
12
00:00:44.070 --> 00:00:47.110
Welcome to Startuprad IO,
13
00:00:47.670 --> 00:00:51.270
your podcast and YouTube blog covering the German
14
00:00:51.270 --> 00:00:54.750
startup scene which with news, interviews and
15
00:00:54.990 --> 00:00:56.190
live events.
16
00:00:58.830 --> 00:01:02.470
My guest is Mykola Myksymenko, co founder and
17
00:01:02.470 --> 00:01:05.790
CTO of HiQ. Trained in
18
00:01:05.790 --> 00:01:09.550
theoretical physics at the Max Planck Institute for Complex Systems
19
00:01:09.550 --> 00:01:13.270
at Dresden, University of Gottingen, University of Magdeburg and
20
00:01:13.270 --> 00:01:16.790
the Weizmann Institute. He later led large scale
21
00:01:16.790 --> 00:01:20.340
industrial R and D before founding HiQ.
22
00:01:20.660 --> 00:01:24.260
His company focuses on a problem the industry prefers to
23
00:01:24.260 --> 00:01:27.820
avoid. Quantum hardware exists, but the
24
00:01:27.820 --> 00:01:31.460
execution layer required for real applications does
25
00:01:31.700 --> 00:01:35.340
not. HiQ builds hardware, wear middleware
26
00:01:35.340 --> 00:01:38.820
designed to minimize noise, extend
27
00:01:39.300 --> 00:01:42.580
usable circuit depth, and make current
28
00:01:42.740 --> 00:01:46.100
quantum processors economically relevant.
29
00:01:46.870 --> 00:01:50.310
The company has already worked with IBM, Airbus, BMW,
30
00:01:50.310 --> 00:01:53.670
HSBC and Capgemini, released open source
31
00:01:53.670 --> 00:01:57.030
tooling adopted by the ecosystem, and recently
32
00:01:57.270 --> 00:02:00.790
raised an $11 million seed round while pushing
33
00:02:00.870 --> 00:02:04.710
its first product to market. Welcome Mykola
34
00:02:06.070 --> 00:02:09.790
hi Joe, it's my pleasure to be on this podcast and
35
00:02:09.790 --> 00:02:13.350
it's my pleasure to have made it through the intro without buttering your family
36
00:02:13.430 --> 00:02:17.030
name. Your early
37
00:02:17.270 --> 00:02:21.110
work in theoretical physics trained you to think in
38
00:02:21.110 --> 00:02:24.950
terms of fundamental limits rather than incremental
39
00:02:24.950 --> 00:02:28.790
improvements. How did that mindset shape
40
00:02:28.870 --> 00:02:31.830
the way you look at computing problems today?
41
00:02:32.870 --> 00:02:36.550
I think important moment that we are living
42
00:02:36.550 --> 00:02:39.350
right now is that there are many
43
00:02:39.670 --> 00:02:43.190
exponential technologies which kind of intercept
44
00:02:43.670 --> 00:02:47.330
and accelerate the pace of technology development
45
00:02:47.330 --> 00:02:51.090
that we are facing today. So on one hand there is AI that
46
00:02:51.090 --> 00:02:54.850
everyone is talking about, but on the other hand there is kind of underlying
47
00:02:54.850 --> 00:02:58.050
current which is quantum computing. And
48
00:02:59.090 --> 00:03:02.769
as a scientist you tend to understand like
49
00:03:02.769 --> 00:03:06.330
what are the kind of fundamental processes which might drive some
50
00:03:06.330 --> 00:03:10.170
processes around us. And that's what helps also
51
00:03:10.170 --> 00:03:13.930
to understand how these different currents actually
52
00:03:13.930 --> 00:03:17.250
can accelerate each other. And while quantum computing was
53
00:03:17.570 --> 00:03:21.170
seemingly too far away just like around a decade
54
00:03:21.170 --> 00:03:24.770
ago, today we see it already like
55
00:03:24.930 --> 00:03:28.210
giving some value in various
56
00:03:28.289 --> 00:03:32.010
scientific applications and we see how much it
57
00:03:32.010 --> 00:03:35.650
accelerated over just like last two years. And
58
00:03:36.450 --> 00:03:40.210
while being in this community and in this
59
00:03:42.430 --> 00:03:46.150
scientific endeavor for a while, I
60
00:03:46.150 --> 00:03:49.830
can already kind of predict that within next years we
61
00:03:49.830 --> 00:03:53.550
will see quite a lot of progress in quantum computing. Such
62
00:03:53.550 --> 00:03:56.590
as we may not even predict that in the incremental way.
63
00:04:01.150 --> 00:04:04.510
You then move from academia. We already talked about it
64
00:04:05.070 --> 00:04:08.720
into industry R and D leadership. Where did you
65
00:04:08.720 --> 00:04:12.080
first see the gap between impressive research result
66
00:04:12.320 --> 00:04:15.520
and systems that actually work in production?
67
00:04:17.360 --> 00:04:21.040
Little disclaimer, the systems never work in production
68
00:04:21.040 --> 00:04:24.720
as they normally should. That's true.
69
00:04:24.720 --> 00:04:28.240
But one thing that kind of made me
70
00:04:28.240 --> 00:04:31.920
shift from academia to industry was particularly that I wanted to
71
00:04:32.500 --> 00:04:36.340
work on something which is immediately relevant or immediately
72
00:04:36.580 --> 00:04:40.180
can be applied to some hard problems in real life
73
00:04:40.820 --> 00:04:44.380
and in academia. My background was in
74
00:04:44.380 --> 00:04:48.180
quantum condensed metaphysics. I worked a lot on like predicting
75
00:04:48.420 --> 00:04:51.460
and trying to manipulate
76
00:04:52.820 --> 00:04:55.700
quantum states of matter in order to
77
00:04:56.500 --> 00:05:00.190
discover new materials or new properties of materials. And
78
00:05:00.190 --> 00:05:04.030
while it's very fascinating and very interesting field of research, it
79
00:05:04.030 --> 00:05:07.550
also has quite long horizon in terms of
80
00:05:07.550 --> 00:05:11.270
applicability. So some application of what I was
81
00:05:11.270 --> 00:05:15.070
doing may happen,
82
00:05:15.070 --> 00:05:18.910
for example in real materials or be
83
00:05:18.910 --> 00:05:22.270
discovered in real materials in tens or
84
00:05:22.270 --> 00:05:25.870
twenties years from now. And then I
85
00:05:25.870 --> 00:05:29.460
shifted to industry where the things were on
86
00:05:29.460 --> 00:05:33.100
completely different pace. And here what
87
00:05:33.100 --> 00:05:36.580
made me to make this shift was particularly the progress of
88
00:05:36.580 --> 00:05:39.820
AI. So while working on the
89
00:05:40.620 --> 00:05:44.259
complex, so called the field of complex physics,
90
00:05:44.259 --> 00:05:48.060
of physics of complex systems, I was also interested
91
00:05:48.060 --> 00:05:51.620
in the physics of neural networks. And apparently a lot of
92
00:05:51.620 --> 00:05:55.460
physics theory and methods can be applied into like
93
00:05:55.460 --> 00:05:58.610
theory of learning. That what made me
94
00:06:00.530 --> 00:06:04.130
being able to naturally transit into that
95
00:06:04.130 --> 00:06:07.970
field and made quite actually a good career in industry
96
00:06:08.450 --> 00:06:12.210
where we built actual production grade system based
97
00:06:12.210 --> 00:06:15.970
on the state of the art research which might just appeared in the literature.
98
00:06:15.970 --> 00:06:19.250
And we picked it up and built some systems around that.
99
00:06:21.090 --> 00:06:24.900
That was actually very exciting. And I'm still
100
00:06:24.900 --> 00:06:28.020
kind of using that on one hand, fundamental
101
00:06:28.820 --> 00:06:32.580
understanding of the processes and interest in fundamental science.
102
00:06:32.820 --> 00:06:36.140
But at the same time in industry I got a lot of these
103
00:06:36.140 --> 00:06:39.860
interesting methodologies and approaches which allow
104
00:06:40.100 --> 00:06:43.860
now build very quick
105
00:06:43.860 --> 00:06:46.900
and rapid iterations of research
106
00:06:47.460 --> 00:06:51.300
in the startup environment such that we can actually quickly arrive
107
00:06:51.300 --> 00:06:54.940
into some relevant results which can be
108
00:06:54.940 --> 00:06:57.680
applied already today to some specific problems.
109
00:07:00.400 --> 00:07:04.000
I see you said you had you built quite a career
110
00:07:04.400 --> 00:07:08.120
in the industry. So when did you decide to
111
00:07:08.120 --> 00:07:11.600
find high Q? You factually bet
112
00:07:11.840 --> 00:07:15.480
that software and not the hardware would decide the fate of
113
00:07:15.480 --> 00:07:19.280
quantum computing. What convinced you that this was
114
00:07:19.600 --> 00:07:23.280
the layer where real leverage actually sits?
115
00:07:24.230 --> 00:07:27.510
Yeah, that's exciting story. So
116
00:07:28.390 --> 00:07:31.790
as I shifted to industry, I was always fascinated about
117
00:07:31.790 --> 00:07:35.270
quantum computing. Quantum computing is something that I did my
118
00:07:35.430 --> 00:07:38.990
master's degree. So I started with quantum computing, then
119
00:07:38.990 --> 00:07:42.750
shifted into condensed metaphysics, still worked a lot on non
120
00:07:42.750 --> 00:07:46.550
equilibrium physics and methods of simulating quantum matter
121
00:07:47.270 --> 00:07:51.120
and quantum states and then worked
122
00:07:51.120 --> 00:07:54.880
on AI for a bit. And what I seen in AI was very similar
123
00:07:54.880 --> 00:07:58.720
to what I see today in what happens in quantum computing.
124
00:07:59.200 --> 00:08:02.560
So we have systems which
125
00:08:02.800 --> 00:08:06.480
already are capable of doing interesting results
126
00:08:07.200 --> 00:08:11.000
in the very narrow domain. So, for example, you can apply quantum
127
00:08:11.000 --> 00:08:13.120
computers, modern quantum computers, to
128
00:08:15.030 --> 00:08:18.390
specific simulations of states of matter
129
00:08:18.550 --> 00:08:21.750
or some specific optimization problems which might be
130
00:08:22.230 --> 00:08:25.790
very hard or nearly impossible to solve on the
131
00:08:25.790 --> 00:08:29.110
classical computer. However, there are still classical
132
00:08:29.110 --> 00:08:32.630
counterparts, but in some cases, solving
133
00:08:32.870 --> 00:08:36.710
some of these problems on the quantum computer can be already faster and
134
00:08:36.710 --> 00:08:39.430
sometimes even more convenient. And
135
00:08:40.150 --> 00:08:43.950
that's exactly what I seen in the beginning or
136
00:08:43.950 --> 00:08:47.390
early stages of deep learning, where these systems were already
137
00:08:47.390 --> 00:08:51.230
discovered. And people seen a lot of interesting, fascinating applications in
138
00:08:51.230 --> 00:08:54.790
computer vision in like image
139
00:08:54.790 --> 00:08:58.590
recognition or signal processing.
140
00:08:58.910 --> 00:09:02.510
However, it took around five to 10 years before these
141
00:09:02.510 --> 00:09:06.270
systems became like massively productized. And
142
00:09:07.790 --> 00:09:11.390
one application for that and one background
143
00:09:11.930 --> 00:09:14.250
process which contributed
144
00:09:15.690 --> 00:09:19.370
to that embrace of technology was particularly
145
00:09:19.370 --> 00:09:22.650
building a bunch of the infrastructure
146
00:09:23.370 --> 00:09:26.930
around the research artifacts, for
147
00:09:26.930 --> 00:09:30.490
example, like PyTorch system or TensorFlow systems,
148
00:09:31.050 --> 00:09:34.610
which allowed a lot of people to build these complex
149
00:09:34.610 --> 00:09:38.290
and more and more complex neural
150
00:09:38.290 --> 00:09:41.820
networks easier and train them at scale,
151
00:09:42.380 --> 00:09:45.980
reproduce the results at scale, such that the whole research community
152
00:09:46.060 --> 00:09:49.900
accelerated dramatically. So we kind of lacked
153
00:09:49.900 --> 00:09:53.740
that in quantum computing. So running something on the quantum computer two
154
00:09:53.740 --> 00:09:57.340
years ago was very hard. Like people experimented
155
00:09:57.340 --> 00:10:00.940
literally with like small prototypes of 4 to 10 qubits.
156
00:10:01.420 --> 00:10:05.180
Today we already see community which
157
00:10:05.340 --> 00:10:09.140
runs applications on hundreds of qubits and hundreds of qubits.
158
00:10:09.140 --> 00:10:12.760
This is already a scale which is often hard to
159
00:10:12.760 --> 00:10:16.280
reproduce classically. So you need to be an expert in numerical
160
00:10:16.280 --> 00:10:19.800
techniques in classical high performance computing in order to
161
00:10:19.800 --> 00:10:23.400
reproduce those results. So it went from
162
00:10:23.480 --> 00:10:27.160
extremely toy problems to already more or less
163
00:10:27.240 --> 00:10:30.960
state of the art high performance computing. So I'm very optimistic about where
164
00:10:30.960 --> 00:10:34.040
it all goes. And I think
165
00:10:34.680 --> 00:10:37.880
this intuition of combining different experience
166
00:10:38.200 --> 00:10:41.920
between fields helps me today to shape our research agenda
167
00:10:41.920 --> 00:10:43.780
and prioritizing what to focus on.
168
00:10:45.700 --> 00:10:49.180
Out of curiosity, you've been talking about that quantum
169
00:10:49.180 --> 00:10:52.500
computing can now deliver what high performance computing
170
00:10:53.620 --> 00:10:57.300
as state of the art computers can deliver right now. So
171
00:10:57.300 --> 00:11:00.340
I was wondering, if you extrapolate when
172
00:11:00.660 --> 00:11:03.460
quantum computing has its full capability,
173
00:11:04.180 --> 00:11:07.900
what could this mean? Especially everybody is
174
00:11:07.900 --> 00:11:11.560
talking right now about the energy usage
175
00:11:11.560 --> 00:11:15.080
of all those data centers that the big
176
00:11:15.080 --> 00:11:18.640
AI companies or the hyperscalers in general are right
177
00:11:18.640 --> 00:11:22.280
now building, what could that mean? Would it mean like
178
00:11:22.520 --> 00:11:26.280
100% more computing power
179
00:11:26.600 --> 00:11:30.400
per what or where would quantum computing go
180
00:11:30.400 --> 00:11:34.160
there? And then can you give us a very tiny idea if
181
00:11:34.160 --> 00:11:37.580
you have the current LLM models running on a much more
182
00:11:37.580 --> 00:11:41.260
capable quantum computer? Yeah,
183
00:11:41.500 --> 00:11:45.300
that's probably something which is important to dig in. So
184
00:11:45.300 --> 00:11:49.020
many people see quantum computers as just like another GPU
185
00:11:49.020 --> 00:11:52.780
or another large computer which runs
186
00:11:52.780 --> 00:11:56.540
bigger and faster models. This is a little bit
187
00:11:56.620 --> 00:12:00.140
different story. So quantum computers are different kinds of
188
00:12:00.220 --> 00:12:02.790
computations and it's suitable
189
00:12:03.590 --> 00:12:07.310
perfectly to some specific kinds of problems. And it can
190
00:12:07.310 --> 00:12:10.990
be completely useless to different problems. And running classical
191
00:12:10.990 --> 00:12:14.790
computations on the quantum computer, it's not the way
192
00:12:14.790 --> 00:12:18.430
forward. So it's not enough just like to take your favorite
193
00:12:18.430 --> 00:12:22.150
classical computing algorithm and
194
00:12:22.150 --> 00:12:25.270
run it on the quantum computer. You need to completely
195
00:12:26.790 --> 00:12:30.400
rethink the problem and rethink the mathematical approach to your
196
00:12:30.400 --> 00:12:34.240
algorithm to actually use all the
197
00:12:34.240 --> 00:12:36.080
capabilities of
198
00:12:38.160 --> 00:12:41.440
quantum bits there and then tunnelement
199
00:12:41.920 --> 00:12:44.960
the ability to parallelize those computations
200
00:12:45.680 --> 00:12:49.240
using quantum computers. So this is different way of
201
00:12:49.240 --> 00:12:52.880
computations. And that's importantly to
202
00:12:52.960 --> 00:12:56.400
realize that it's probably not immediately
203
00:12:57.040 --> 00:13:00.400
applicable to running large LLMs or
204
00:13:00.480 --> 00:13:03.500
running like conventional
205
00:13:03.900 --> 00:13:06.700
machine learning applications. Actually, if you think about
206
00:13:07.180 --> 00:13:10.460
applying big data problems, one of the
207
00:13:10.620 --> 00:13:14.420
crucial bottleneck in quantum computing is actually loading data
208
00:13:14.420 --> 00:13:18.260
in the quantum computer. So one of the breakthroughs
209
00:13:18.260 --> 00:13:21.180
that we did at Haiku was particularly
210
00:13:21.980 --> 00:13:25.660
unlocking this capability of loading large scale
211
00:13:26.060 --> 00:13:29.450
data sets, which are of the industrial
212
00:13:30.810 --> 00:13:34.610
scale number of features into quantum computers. Before that it was
213
00:13:34.610 --> 00:13:36.730
just impossible. And
214
00:13:39.770 --> 00:13:43.530
there are like small, small improvements here and there which
215
00:13:44.010 --> 00:13:47.730
probably enlarge the number of problems that we can
216
00:13:47.730 --> 00:13:51.330
solve with quantum computers. But before, before
217
00:13:51.330 --> 00:13:54.130
we go into like ubiquitous era,
218
00:13:55.330 --> 00:13:58.770
we will be focused first on the very narrow
219
00:13:58.770 --> 00:14:02.450
problems. And some of them are designing new
220
00:14:02.690 --> 00:14:05.970
molecules and simulating new molecules.
221
00:14:06.850 --> 00:14:10.450
That's a very hard problem for classical computers.
222
00:14:10.690 --> 00:14:14.370
But this is something where quantum computer can
223
00:14:14.370 --> 00:14:17.930
actually help because this is a natural system to tackle with
224
00:14:17.930 --> 00:14:21.330
qubits, it maps almost one to one.
225
00:14:22.530 --> 00:14:25.570
We can think about the problem that I was solving in my
226
00:14:26.530 --> 00:14:30.290
scientific career like a problem of materials. Like if you want
227
00:14:30.290 --> 00:14:32.690
to simulate superconductivity or
228
00:14:33.490 --> 00:14:36.290
magnetism in some specific new
229
00:14:37.490 --> 00:14:40.610
types of metals, then
230
00:14:41.330 --> 00:14:44.610
quantum computers can potentially help with these problems.
231
00:14:45.730 --> 00:14:49.210
Or computational fluid dynamics. That's another
232
00:14:49.210 --> 00:14:51.570
interesting problem which is very
233
00:14:52.450 --> 00:14:56.110
widespread and industry. So companies
234
00:14:56.110 --> 00:14:59.590
in aerospace or automotive run a lot of
235
00:14:59.590 --> 00:15:02.990
computational fluid dynamics simulations and
236
00:15:03.390 --> 00:15:06.910
you very quickly actually kind of hit the glass
237
00:15:06.910 --> 00:15:10.510
ceiling there with a grid
238
00:15:10.670 --> 00:15:14.350
that, that you can simulate on the classical computers. While
239
00:15:14.350 --> 00:15:17.950
on the quantum computers theoretically you can go to much
240
00:15:17.950 --> 00:15:21.790
more denser grid of your simulation and have
241
00:15:21.790 --> 00:15:25.630
much more precise results. However, it's very hard right now to
242
00:15:25.630 --> 00:15:29.070
map this algorithm directly. And we also working on that
243
00:15:29.070 --> 00:15:32.470
helping some companies who are partners to
244
00:15:32.550 --> 00:15:36.350
actually make this transition. So there are A bunch of
245
00:15:36.350 --> 00:15:39.950
problems like this optimization simulation, natural
246
00:15:39.950 --> 00:15:43.390
systems and multiphysics
247
00:15:43.390 --> 00:15:46.870
simulations. These are probably those which will be
248
00:15:47.660 --> 00:15:51.300
on the kind of early, early low hanging
249
00:15:51.300 --> 00:15:52.300
fruits, horizon
250
00:15:55.500 --> 00:15:59.060
and middle to long term. What would it mean to everybody who's now
251
00:15:59.060 --> 00:16:02.860
thinking about investing in data centers who
252
00:16:03.100 --> 00:16:06.140
are already there, announced in vast,
253
00:16:06.780 --> 00:16:10.540
vast numbers? Do you think that'll change
254
00:16:10.940 --> 00:16:13.420
something over time when
255
00:16:14.520 --> 00:16:18.280
quantum computing really becomes productive, especially in
256
00:16:18.280 --> 00:16:21.240
terms of space needed and energy for
257
00:16:21.800 --> 00:16:24.920
computing of some form? Maybe still LLMs
258
00:16:26.200 --> 00:16:29.880
of artificial intelligence. Look,
259
00:16:30.840 --> 00:16:32.360
even from the current perspective,
260
00:16:34.600 --> 00:16:38.280
if you want to simulate some process in nature, like
261
00:16:38.760 --> 00:16:42.530
for example, a molecule, you can do that on
262
00:16:42.530 --> 00:16:46.250
the classical computer to some extent. And I used to
263
00:16:46.250 --> 00:16:50.010
work, for example, on some
264
00:16:50.010 --> 00:16:53.090
of the largest supercomputing centers in Germany.
265
00:16:53.810 --> 00:16:57.090
And those simulations cost
266
00:16:57.970 --> 00:17:01.330
a lot. They can cost easily like $200,000.
267
00:17:02.450 --> 00:17:06.170
And in order to simulate some material or some
268
00:17:06.170 --> 00:17:08.750
specific physical effects while.
269
00:17:09.630 --> 00:17:13.390
And you can run those simulations sometimes for weeks in order
270
00:17:13.390 --> 00:17:16.430
to get some reasonable result. And we already
271
00:17:16.750 --> 00:17:20.350
witness some interesting types of these
272
00:17:20.350 --> 00:17:22.670
simulations where we can simulate the same system,
273
00:17:24.430 --> 00:17:27.990
still simulatable classically, but we can simulate it already today on the
274
00:17:27.990 --> 00:17:31.710
quantum computer, and we can get results maybe in
275
00:17:31.710 --> 00:17:35.160
10 minutes. So even from that
276
00:17:35.160 --> 00:17:37.800
perspective, like running large
277
00:17:37.880 --> 00:17:41.720
supercomputing cluster for weeks, or
278
00:17:42.280 --> 00:17:46.120
running a quantum computer for a few minutes, I can already
279
00:17:46.520 --> 00:17:49.800
hint that you actually probably spent much less energy
280
00:17:50.360 --> 00:17:54.080
on that. When we will be able
281
00:17:54.080 --> 00:17:57.760
to run machine learning simulations at scale, I would say
282
00:17:57.760 --> 00:18:01.560
that that probably will transit also to machine learning
283
00:18:01.560 --> 00:18:05.080
simulations. I need to be honest that
284
00:18:05.080 --> 00:18:07.960
machine learning is still hard to solve
285
00:18:08.520 --> 00:18:11.960
directly on the quantum computer. But there are very specific
286
00:18:11.960 --> 00:18:15.159
problems, for example, anomaly detection, which
287
00:18:15.400 --> 00:18:18.440
potentially quantum computers can actually solve more
288
00:18:20.360 --> 00:18:23.560
better than classical computers. And we already
289
00:18:24.520 --> 00:18:27.000
see few of those examples here and there.
290
00:18:29.070 --> 00:18:32.870
But ubiquitous machine learning, where we,
291
00:18:32.870 --> 00:18:36.510
for example, apply it to LLMs, it's
292
00:18:36.510 --> 00:18:39.790
probably some years away from now.
293
00:18:41.390 --> 00:18:45.230
Some years. Okay, I see. Let's get back to
294
00:18:45.470 --> 00:18:49.030
the original questions I envisioned for this interview. For
295
00:18:49.030 --> 00:18:52.350
founders watching quantum from the outside,
296
00:18:52.670 --> 00:18:56.470
noise, not qubit count, is the real reason
297
00:18:56.470 --> 00:18:59.910
most pilots fail quietly. What actually
298
00:19:00.070 --> 00:19:03.190
happens inside these systems
299
00:19:03.750 --> 00:19:07.350
that makes noise such a dominant constraint?
300
00:19:09.670 --> 00:19:13.510
Yeah, so you should think about quantum computing as
301
00:19:14.550 --> 00:19:17.670
basically early stages of classical computers. Like
302
00:19:18.390 --> 00:19:21.350
try to remember what happened in classical computing
303
00:19:22.350 --> 00:19:26.030
in between 40s and 60s. So these
304
00:19:26.030 --> 00:19:29.630
were like, these were
305
00:19:29.630 --> 00:19:33.230
large machines. Like sometimes they were like of the size of a room
306
00:19:33.710 --> 00:19:37.510
and they were very hard to operate. They did
307
00:19:37.510 --> 00:19:40.590
not have error correction at the time. So
308
00:19:41.070 --> 00:19:44.710
basically they were operating with valves. And those
309
00:19:44.710 --> 00:19:48.350
valves can very easily overheat or become malfunctioning.
310
00:19:49.150 --> 00:19:52.750
So Literally the first bugs were in those
311
00:19:52.750 --> 00:19:56.390
computers, just because some things did not function well. So you
312
00:19:56.390 --> 00:19:59.630
should be a high level hardware operator
313
00:20:00.030 --> 00:20:03.870
and understand the low level algorithmic theory in
314
00:20:03.870 --> 00:20:06.990
order to run something on these machines.
315
00:20:07.470 --> 00:20:11.190
And that was for a while before error correction was
316
00:20:11.190 --> 00:20:14.790
introduced to these machines. And today we are in
317
00:20:14.790 --> 00:20:18.350
very similar era in the quantum computers. So these
318
00:20:18.350 --> 00:20:21.510
machines already exist, they have hundreds of qubits,
319
00:20:21.830 --> 00:20:25.030
hundreds of qubits are hard to simulate. And we already have
320
00:20:25.510 --> 00:20:29.270
evidences from Google, from IBM and other companies
321
00:20:29.430 --> 00:20:32.550
that you can simulate on these machines some very
322
00:20:33.190 --> 00:20:36.710
complicated states of matter
323
00:20:36.950 --> 00:20:40.150
or quantum states, which are
324
00:20:40.230 --> 00:20:43.750
impossible or will take extremely long time
325
00:20:43.750 --> 00:20:46.720
to simulate classically. So
326
00:20:47.840 --> 00:20:51.280
we have those evidences. Now the question is
327
00:20:51.600 --> 00:20:55.440
how we can use these machines for something practical. And that's where
328
00:20:55.440 --> 00:20:59.160
the bottleneck is. Because in order to run something on the
329
00:20:59.160 --> 00:21:02.480
quantum computers, you need to be
330
00:21:03.280 --> 00:21:06.480
very much an expert in how to fight noise and
331
00:21:06.480 --> 00:21:09.840
imperfections in these machines. They are not perfect, and
332
00:21:10.160 --> 00:21:13.960
in order to apply error correction to them, you need thousands of qubits,
333
00:21:13.960 --> 00:21:17.670
and we are not yet in that regime, and better fidelity of those
334
00:21:17.670 --> 00:21:21.190
qubits. So we are in the low qubit regimes and noisy machines.
335
00:21:22.070 --> 00:21:24.710
But apparently
336
00:21:26.310 --> 00:21:29.430
even in this regime you can extract something useful.
337
00:21:30.070 --> 00:21:33.830
And in order to do that, a lot of expertise is needed.
338
00:21:33.910 --> 00:21:36.950
What we try to introduce is
339
00:21:37.990 --> 00:21:41.270
to some extent democratization of ability
340
00:21:41.670 --> 00:21:45.190
for a wider community of scientists,
341
00:21:45.750 --> 00:21:49.330
of algorithm researchers, of material scientists,
342
00:21:49.410 --> 00:21:52.850
quantum chemists use quantum computers for
343
00:21:53.010 --> 00:21:56.850
their own simulations and try to discover something
344
00:21:56.930 --> 00:22:00.210
useful. And in order to do that, we abstracted away
345
00:22:00.690 --> 00:22:04.290
all those low level manipulations with
346
00:22:05.090 --> 00:22:08.050
noise, with improvement of performance,
347
00:22:08.610 --> 00:22:12.370
with reducing the impact of noise on the algorithm
348
00:22:12.370 --> 00:22:16.220
performance, such that scientists can purely focus on science
349
00:22:16.460 --> 00:22:20.060
and on the mapping their problem to the quantum algorithm,
350
00:22:20.060 --> 00:22:23.500
and they can completely forget about what happens
351
00:22:23.820 --> 00:22:27.500
under that layer. And we think that that is
352
00:22:27.500 --> 00:22:31.300
crucial for today, because today we need more and more scientists
353
00:22:31.300 --> 00:22:34.900
to work with quantum computers and discover those applications. The more people
354
00:22:34.900 --> 00:22:38.700
will run on quantum computers, the sooner we actually will
355
00:22:38.700 --> 00:22:42.370
discover userful applications, as we did with classical
356
00:22:42.370 --> 00:22:45.890
computing, as we did with deep
357
00:22:45.890 --> 00:22:49.490
learning applications and other use cases. You
358
00:22:49.490 --> 00:22:52.930
often say the quantum software stack
359
00:22:52.930 --> 00:22:56.050
simply doesn't exist. Yet when companies start
360
00:22:56.050 --> 00:22:59.730
experimenting with quantum today, where do they usually
361
00:22:59.810 --> 00:23:03.170
underestimate the integration complexity?
362
00:23:03.410 --> 00:23:07.170
Yeah. So when we interact with
363
00:23:07.170 --> 00:23:09.970
industries, what we frequently see that
364
00:23:11.130 --> 00:23:14.730
many companies start setting up the quantum programs.
365
00:23:16.010 --> 00:23:19.610
They understand that this is a technology of the future, and
366
00:23:19.770 --> 00:23:22.650
sooner or later this will dramatically transform
367
00:23:23.370 --> 00:23:27.210
of what they do in many of the heavy high
368
00:23:27.210 --> 00:23:30.410
performance computing simulations.
369
00:23:31.210 --> 00:23:34.810
And in practice, this is not
370
00:23:34.810 --> 00:23:38.650
only about building Software. It's also about deep,
371
00:23:39.710 --> 00:23:42.990
low level expertise in happening something to run
372
00:23:43.630 --> 00:23:46.910
on the real hardware. And that's where we see real gap.
373
00:23:47.230 --> 00:23:50.190
So I know few
374
00:23:50.990 --> 00:23:54.829
companies who actually have some expertise in
375
00:23:54.829 --> 00:23:58.430
running simulations on real hardware at
376
00:23:59.150 --> 00:24:02.190
the state of the art scale.
377
00:24:02.830 --> 00:24:06.470
Most of the simulations which people do are of the
378
00:24:06.470 --> 00:24:10.100
toy scale. But you can run more. And
379
00:24:10.900 --> 00:24:14.180
in order to do so you need deep expertise
380
00:24:16.820 --> 00:24:20.620
in, for example, tailoring noise such that you
381
00:24:20.620 --> 00:24:24.140
can use various
382
00:24:24.140 --> 00:24:27.780
symmetric properties in order to simplify the noise
383
00:24:28.180 --> 00:24:32.020
that is in the hardware. Then you can do a
384
00:24:32.020 --> 00:24:35.320
bunch of tricks in order to suppress its effect on the
385
00:24:35.320 --> 00:24:38.720
algorithmic performance. You can also work on the
386
00:24:38.720 --> 00:24:42.320
algorithm itself in order to reduce the number of operations
387
00:24:42.400 --> 00:24:46.240
that operate in that algorithm and drive
388
00:24:46.240 --> 00:24:49.680
you to the resulting state. So
389
00:24:50.000 --> 00:24:53.840
all of these tricks are complex in nature.
390
00:24:54.240 --> 00:24:57.880
And frankly speaking, you need to be expert in just doing
391
00:24:57.880 --> 00:25:01.360
this for some very short period of time
392
00:25:01.760 --> 00:25:05.410
while these computers are in this state of
393
00:25:06.930 --> 00:25:10.610
performance. And I don't
394
00:25:10.610 --> 00:25:12.850
think anyone in industry should actually
395
00:25:14.210 --> 00:25:17.170
spend significant amount of time
396
00:25:18.850 --> 00:25:22.290
focusing on this low level optimization. So this should be done
397
00:25:22.290 --> 00:25:26.010
automatically. And that can be done by a
398
00:25:26.010 --> 00:25:29.730
middleware stack. And in quantum computing it's almost
399
00:25:29.730 --> 00:25:33.290
non existent. So if you think about again, like parallel
400
00:25:33.290 --> 00:25:36.450
to classical computers, we have seen
401
00:25:36.930 --> 00:25:40.050
software driving the adoption
402
00:25:40.130 --> 00:25:43.930
revolution several times already. For example, when we had the
403
00:25:43.930 --> 00:25:47.770
first personal computers, this kind of
404
00:25:47.770 --> 00:25:50.930
middleware revolution was happening because of adoption of
405
00:25:51.250 --> 00:25:55.050
BASIC language. So that was introduced
406
00:25:55.050 --> 00:25:58.460
by today's Microsoft, such that
407
00:25:59.180 --> 00:26:02.780
more and more quantum enthusiasts were able to actually program
408
00:26:03.900 --> 00:26:06.940
computers in a much easier way than they did before.
409
00:26:08.140 --> 00:26:11.820
Then the same kind of revolution happens recently
410
00:26:11.900 --> 00:26:15.420
with introducing of GPUs
411
00:26:15.420 --> 00:26:19.020
and CUDA language or
412
00:26:19.500 --> 00:26:22.700
framework to program those GPUs.
413
00:26:23.100 --> 00:26:26.610
So in this case, again like a middleware stack,
414
00:26:26.850 --> 00:26:30.130
in this case, CUDA helped a lot of scientists
415
00:26:30.530 --> 00:26:34.330
and experts in AI, for example, to use
416
00:26:34.330 --> 00:26:36.770
GPUs for their computational needs.
417
00:26:37.810 --> 00:26:41.410
And currently we are in this very same kind
418
00:26:41.410 --> 00:26:45.210
of stage with quantum computers. So right now I would say
419
00:26:45.210 --> 00:26:48.690
there is no CUDA or no BASIC for
420
00:26:49.170 --> 00:26:52.890
quantum computers. It's still very hard to program them. You literally work
421
00:26:52.890 --> 00:26:56.650
on the quantum assembly level. And it takes
422
00:26:56.650 --> 00:27:00.210
a lot of effort to run something and we want to abstract that
423
00:27:00.290 --> 00:27:03.970
away. Giving people instruments and tools such
424
00:27:03.970 --> 00:27:07.570
that they can think about the problem, can think about like what would
425
00:27:07.570 --> 00:27:10.850
be the actual use of
426
00:27:11.410 --> 00:27:15.010
quantum computers in their particular problems, or how to decompose or
427
00:27:15.010 --> 00:27:18.810
change the algorithm itself, rather than thinking how
428
00:27:18.810 --> 00:27:22.090
to optimize it to run on the hardware so that that we can take and
429
00:27:22.090 --> 00:27:23.410
abstract away from the user,
430
00:27:26.600 --> 00:27:30.440
we're talking middleware here that sits
431
00:27:30.520 --> 00:27:34.280
exactly in a layer where responsibility shifts from
432
00:27:34.600 --> 00:27:38.440
the hardware isn't ready to the execution model is wrong.
433
00:27:38.680 --> 00:27:42.000
What does that layer actually do
434
00:27:42.000 --> 00:27:43.880
inside the quantum workflow?
435
00:27:45.720 --> 00:27:48.600
Maybe you can dumb it down for non physics majors.
436
00:27:50.690 --> 00:27:54.530
Right, so imagine that we have hardware and apparently a
437
00:27:54.530 --> 00:27:58.290
lot of hardware providers, they don't focus that much on the software
438
00:27:58.290 --> 00:28:01.810
layer just because the hardware is so hard to build.
439
00:28:02.210 --> 00:28:05.809
There are so many problems that you need to be laser focused in order to
440
00:28:05.890 --> 00:28:08.210
make your qubits better, more stable,
441
00:28:10.770 --> 00:28:14.570
scale them to a larger quantity and so on. So there are a lot
442
00:28:14.570 --> 00:28:18.380
of challenges there in hardware stack and then there
443
00:28:18.380 --> 00:28:21.900
are people who worked on application sites. So we still need to
444
00:28:21.900 --> 00:28:24.900
discover a lot of quantum algorithms in a lot of different
445
00:28:26.740 --> 00:28:30.420
algorithmic areas and then understand which
446
00:28:30.500 --> 00:28:34.220
actual real life problems map to those quantum algorithms.
447
00:28:34.220 --> 00:28:38.060
So there are just a handful of those today. So I would say there
448
00:28:38.060 --> 00:28:41.540
are two communities. One is low level on the hardware, another community
449
00:28:41.700 --> 00:28:45.480
works more on algorithmic side and they
450
00:28:45.480 --> 00:28:49.160
do not overlap that much. So that where
451
00:28:49.160 --> 00:28:52.960
we sit. So it's not only Weeb, there are
452
00:28:52.960 --> 00:28:56.760
other companies working in this space, but the focus is to
453
00:28:56.760 --> 00:29:00.120
provide the middle layer between low level
454
00:29:01.800 --> 00:29:05.200
hardware which operates with pulses and low level
455
00:29:05.200 --> 00:29:08.120
assembly languages, quantum assembly languages
456
00:29:09.000 --> 00:29:12.520
and higher level abstractions for algorithm
457
00:29:12.520 --> 00:29:16.220
definition. And that's apparently not
458
00:29:16.220 --> 00:29:20.020
that easy. And in the future we need to understand that the whole
459
00:29:20.020 --> 00:29:23.620
stack is being changing as well at the same
460
00:29:23.620 --> 00:29:26.740
time, because as computers get better,
461
00:29:27.380 --> 00:29:29.460
we introduce new and new
462
00:29:30.580 --> 00:29:34.300
paradigms. For example, in few years from now we
463
00:29:34.300 --> 00:29:37.820
will have full tolerant quantum
464
00:29:37.820 --> 00:29:41.650
computers, so they will operate with error
465
00:29:41.650 --> 00:29:45.090
correction loops in the middle in order to
466
00:29:45.330 --> 00:29:48.690
first encode more qubits in a single
467
00:29:48.770 --> 00:29:52.210
logical qubit and then correct errors if those appear.
468
00:29:53.170 --> 00:29:56.930
So that's already another kind of
469
00:29:56.930 --> 00:30:00.610
algorithmic subroutine which need
470
00:30:00.610 --> 00:30:04.050
to be somehow integrated in this low level stack.
471
00:30:05.020 --> 00:30:08.860
And so in order to do so with many of these components,
472
00:30:08.860 --> 00:30:12.220
we need some kind of orchestration
473
00:30:12.300 --> 00:30:15.900
layer which will orchestrate when and which tools we should
474
00:30:15.900 --> 00:30:17.980
use. There should be
475
00:30:19.820 --> 00:30:23.340
ability to run on one hand real time
476
00:30:24.220 --> 00:30:27.580
error mitigation and error suppression, on the other hand
477
00:30:28.540 --> 00:30:31.820
error correction cycles and finally some
478
00:30:32.730 --> 00:30:36.530
finalized post processing of still logical errors
479
00:30:36.530 --> 00:30:40.290
which might still happen in those computers. So there
480
00:30:40.290 --> 00:30:43.610
are many, many complexity points here
481
00:30:45.050 --> 00:30:48.330
and this becomes more and more complex actually field, but
482
00:30:49.690 --> 00:30:53.370
the stack itself kind of grows in the real time right now
483
00:30:53.370 --> 00:30:56.890
and today it's not that complex. But
484
00:30:57.210 --> 00:31:01.000
I imagine that in 10 years from now that
485
00:31:01.000 --> 00:31:04.760
will be very similar to what we see in classical computers today, which
486
00:31:04.760 --> 00:31:08.520
is very complex Actually middleware stack, which we even don't know,
487
00:31:08.760 --> 00:31:12.520
like typical programmers don't even know what's happening to their
488
00:31:12.520 --> 00:31:16.240
algorithm after they compile. I had to smile
489
00:31:16.240 --> 00:31:19.800
because there was a setback for the IBM stock price a few days
490
00:31:19.800 --> 00:31:22.280
ago that because
491
00:31:24.120 --> 00:31:27.800
Claude the AI model could start working
492
00:31:27.800 --> 00:31:31.480
COBOL code. And so I was wondering at one point
493
00:31:31.480 --> 00:31:35.240
in 10 years if there's somebody who's trying to get COBOL running
494
00:31:35.240 --> 00:31:38.600
on quantum computing. So every.
495
00:31:38.840 --> 00:31:42.520
But my understanding is basically every system, every software out
496
00:31:42.520 --> 00:31:45.320
there that is running on the current
497
00:31:46.200 --> 00:31:49.960
silicon based wafer technology
498
00:31:50.520 --> 00:31:54.280
will need to be redone for the quantum computing age.
499
00:31:56.160 --> 00:31:59.720
Not really. So we don't need to push every
500
00:31:59.720 --> 00:32:03.400
software to the quantum computers. So not all of the problems
501
00:32:03.400 --> 00:32:05.280
need quantum computers to be solved.
502
00:32:06.960 --> 00:32:09.040
So these problems that are
503
00:32:11.440 --> 00:32:15.200
mostly required quantum computers are those which are
504
00:32:16.240 --> 00:32:19.960
complex in nature, which require enormous
505
00:32:19.960 --> 00:32:23.680
amount of memory in order to run some transformations
506
00:32:23.680 --> 00:32:27.380
in data structures or enormous amount of compute.
507
00:32:29.620 --> 00:32:32.420
And these are specific classes of problem
508
00:32:33.540 --> 00:32:37.340
and other problems like for example, you don't need
509
00:32:37.340 --> 00:32:40.860
quantum computers on your phone or like a quantum browser or
510
00:32:40.860 --> 00:32:43.780
whatever. So this is completely different
511
00:32:44.660 --> 00:32:47.300
space. And in principle, in the future we will see
512
00:32:48.020 --> 00:32:51.620
basically these two technologies cooperate together. And we already
513
00:32:51.620 --> 00:32:55.300
have those hybrid quantum classical algorithms where you
514
00:32:55.380 --> 00:32:58.640
have some subroutines running on the classical computer,
515
00:32:58.880 --> 00:33:02.000
while those more complex algorithms
516
00:33:02.880 --> 00:33:06.680
can be decomposed and run separately on
517
00:33:06.680 --> 00:33:10.240
the, on the quantum computer where you get the result and then integrated back into
518
00:33:10.240 --> 00:33:13.680
your classical workflow. So these, these technologies will
519
00:33:13.760 --> 00:33:17.400
cooperate. So I would say that that's the
520
00:33:17.400 --> 00:33:21.120
future that we will see. For example, like on the phone, every day you
521
00:33:21.520 --> 00:33:24.890
call your Uber. Imagine when you do that, you solve the
522
00:33:25.210 --> 00:33:28.930
complex optimization problem such that your driver
523
00:33:28.930 --> 00:33:31.690
needs to navigate the shortest path to your location.
524
00:33:32.570 --> 00:33:36.330
In the future, if quantum computers will be ubiquitous, that
525
00:33:36.810 --> 00:33:40.570
tiny optimization problem will probably be delegated
526
00:33:40.570 --> 00:33:44.330
to a quantum computer to solve maybe fraction of a second faster.
527
00:33:46.170 --> 00:33:47.850
Okay, but
528
00:33:49.920 --> 00:33:53.600
we took a little detour. My frequent listeners already know that.
529
00:33:55.440 --> 00:33:59.280
But we're still talking middleware here. And by
530
00:33:59.360 --> 00:34:02.560
operating in this layer, haiku forces
531
00:34:03.200 --> 00:34:06.640
customers to confront whether they are buying, learning,
532
00:34:06.640 --> 00:34:09.760
signaling or real performance. What
533
00:34:09.840 --> 00:34:13.120
kinds of motivations do you actually see
534
00:34:13.520 --> 00:34:17.230
when companies start exploring quantum
535
00:34:17.230 --> 00:34:20.910
computing? Right. The
536
00:34:21.310 --> 00:34:24.510
challenge in any of the exponential technology
537
00:34:25.390 --> 00:34:29.230
is that initially it seems
538
00:34:29.230 --> 00:34:33.030
like nothing happens. The progress is so slow that it
539
00:34:33.030 --> 00:34:36.710
seems like that the real value will be years ahead and you don't need to
540
00:34:36.710 --> 00:34:40.350
invest to it today. So you can just sit and observe
541
00:34:40.819 --> 00:34:43.059
what happens. On the other hand,
542
00:34:44.659 --> 00:34:48.299
the nature of exponential technologies says
543
00:34:48.299 --> 00:34:51.739
that at some point you have a Very
544
00:34:51.739 --> 00:34:55.299
rapid growth of technological adoption and
545
00:34:56.019 --> 00:34:58.979
capabilities of this technology. And I would say
546
00:34:59.379 --> 00:35:02.979
that's what everyone sees in AI today,
547
00:35:03.699 --> 00:35:07.419
but that what we will see to start happening in quantum very
548
00:35:07.419 --> 00:35:10.370
soon. And the reason for that is that
549
00:35:12.130 --> 00:35:15.650
it's not. First of all we have a pace of technology which is
550
00:35:15.650 --> 00:35:19.490
accelerating. And as being in this part of the industry I literally
551
00:35:19.490 --> 00:35:23.210
see that three years ago it was almost impossible to run
552
00:35:23.210 --> 00:35:26.890
anything. So there were just few people, handful of people who were
553
00:35:26.890 --> 00:35:30.570
running something on the hardware. Today there are many
554
00:35:30.570 --> 00:35:33.970
people in the world who run on the IBM's largest
555
00:35:34.450 --> 00:35:37.980
machines for free via cloud.
556
00:35:38.780 --> 00:35:42.060
And today you can immediately access
557
00:35:42.220 --> 00:35:45.580
largest scale quantum computers in the cloud.
558
00:35:46.620 --> 00:35:50.060
And that's still
559
00:35:50.380 --> 00:35:53.980
not enough to immediately give you the business value.
560
00:35:53.980 --> 00:35:57.580
But that's something, if you see this
561
00:35:57.580 --> 00:36:00.380
perspective, that's already something quite impressive.
562
00:36:01.440 --> 00:36:04.000
Another thing that we see from our perspective
563
00:36:05.040 --> 00:36:06.160
just few years ago,
564
00:36:08.720 --> 00:36:12.480
running something on these largest machines could cost you
565
00:36:12.640 --> 00:36:16.200
easily tens to hundred thousand
566
00:36:16.200 --> 00:36:18.800
dollars if you want to get truly
567
00:36:19.360 --> 00:36:23.080
interesting new results. With the
568
00:36:23.080 --> 00:36:26.480
help of technologies like ours, we can
569
00:36:26.480 --> 00:36:30.280
reduce this cost to tens of dollars or up
570
00:36:30.280 --> 00:36:33.640
to hundreds of dollars. Such that we dramatically
571
00:36:33.640 --> 00:36:37.160
democratize the level of
572
00:36:37.160 --> 00:36:40.880
accessibility of these technologies. Such that now any student can
573
00:36:40.880 --> 00:36:44.320
afford running so called utility scale
574
00:36:44.320 --> 00:36:47.720
experiments on the quantum
575
00:36:47.720 --> 00:36:50.000
computers available to him or her today.
576
00:36:52.240 --> 00:36:56.000
And that again accelerates everything. Because now there will be more
577
00:36:56.000 --> 00:36:59.720
experiments, now there will be more discoveries, now there will be more
578
00:36:59.720 --> 00:37:03.520
algorithms and the whole cycle accelerates. And we will
579
00:37:03.520 --> 00:37:07.040
see the adoption getting faster and faster and faster.
580
00:37:07.920 --> 00:37:11.440
At the same time there is a hardware progress which we see in parallel
581
00:37:11.760 --> 00:37:14.800
which is again moving pretty fast.
582
00:37:15.520 --> 00:37:19.120
Companies actually hit their milestones in
583
00:37:19.840 --> 00:37:23.120
different qubit modalities. In superconducting
584
00:37:23.120 --> 00:37:26.260
qubits, we see growing number of
585
00:37:26.900 --> 00:37:30.340
superconducting qubits grids. We see error
586
00:37:30.340 --> 00:37:33.860
correction codes already applied on the superconducting
587
00:37:33.860 --> 00:37:37.220
computers. We see trapped ion computers actually
588
00:37:37.700 --> 00:37:41.300
also growing in number of qubits and in the speed
589
00:37:41.300 --> 00:37:44.500
of running operations on these computers.
590
00:37:44.740 --> 00:37:48.580
Neutral site atom computers also move forward very fast, like
591
00:37:48.580 --> 00:37:52.420
soon. There are a few companies who promise photonic
592
00:37:52.420 --> 00:37:55.960
quantum computers. So there are a lot of different
593
00:37:56.600 --> 00:38:00.400
kind of underlying currents which accelerate the
594
00:38:00.400 --> 00:38:04.120
whole pace of this technological progress. So
595
00:38:04.120 --> 00:38:07.600
I'm very positive and like very optimistic about the future of quantum
596
00:38:07.600 --> 00:38:11.320
computers. Me too. But would have also
597
00:38:11.480 --> 00:38:14.760
understood is that many current quantum
598
00:38:15.480 --> 00:38:18.920
initiatives may be premature.
599
00:38:19.480 --> 00:38:23.160
From your perspective, when should companies pause
600
00:38:23.160 --> 00:38:26.480
experimentation rather than accelerate it?
601
00:38:28.000 --> 00:38:31.720
Right. So I would say I probably
602
00:38:31.720 --> 00:38:35.360
had to say it in the previous question, but
603
00:38:35.760 --> 00:38:39.200
one of important aspect, like why it's important to do something today.
604
00:38:40.240 --> 00:38:43.800
Suppose tomorrow there are quantum computers which operate at
605
00:38:43.800 --> 00:38:46.920
scale. We can run any
606
00:38:46.920 --> 00:38:50.730
algorithm we wish. The challenge would be
607
00:38:51.610 --> 00:38:55.210
to discover those algorithms. There are just handful. And
608
00:38:55.210 --> 00:38:59.010
not every problem is so easily mappable to quantum
609
00:38:59.010 --> 00:39:02.650
algorithm. So it will take another few years to just
610
00:39:02.650 --> 00:39:06.330
discover algorithms for their particular business problems.
611
00:39:07.530 --> 00:39:11.330
Then it takes typically in enterprises, it takes few years in
612
00:39:11.330 --> 00:39:14.670
order to adopt those algorithms into
613
00:39:14.750 --> 00:39:18.550
enterprise software ecosystem, integrate them, test
614
00:39:18.550 --> 00:39:21.950
them, make sure that they produce the right
615
00:39:21.950 --> 00:39:24.910
values in some specific toic conditions.
616
00:39:26.030 --> 00:39:29.670
So like stress test them in different scenarios. So
617
00:39:29.670 --> 00:39:33.310
all that takes typically time. For example, in finance, if you want to
618
00:39:33.390 --> 00:39:36.830
implement a new algorithm which would
619
00:39:37.950 --> 00:39:41.480
help, for example in some decision
620
00:39:41.560 --> 00:39:45.200
making, the roadmap to
621
00:39:45.200 --> 00:39:48.200
implement this algorithm can take sometimes few years.
622
00:39:49.400 --> 00:39:53.240
That's why it's important to start developing these algorithms today. Imagine
623
00:39:53.240 --> 00:39:56.760
that your competitor already did it today. And then
624
00:39:57.080 --> 00:40:00.360
in the year from now, quantum computers suddenly
625
00:40:00.840 --> 00:40:04.520
start to showcase value. So you still will need
626
00:40:04.520 --> 00:40:08.350
to to take few years in order to build, test,
627
00:40:08.670 --> 00:40:12.310
integrate the workflow in your organization. So you will be
628
00:40:12.310 --> 00:40:15.990
already losing. And since it's exponential technology, you will be losing
629
00:40:15.990 --> 00:40:19.750
a lot. So it's very similar to what we see today
630
00:40:19.750 --> 00:40:23.470
in AI. The same story can repeat itself
631
00:40:24.430 --> 00:40:27.870
in quantum computing. So stay in the game.
632
00:40:28.510 --> 00:40:32.080
It may cost you, but if you're not part of the race, it
633
00:40:32.240 --> 00:40:35.840
may cost you more. Okay, this is the
634
00:40:35.840 --> 00:40:39.640
point where founders decide whether quantum computing is a
635
00:40:39.640 --> 00:40:42.880
distraction or an execution problem they will
636
00:40:42.960 --> 00:40:46.720
eventually have to face. For Bayer
637
00:40:46.800 --> 00:40:50.160
evaluating Haikyuu, the real value is not
638
00:40:50.320 --> 00:40:53.360
speed ups in isolation, but making
639
00:40:53.600 --> 00:40:57.280
quantum experiments behave less like bespoke research
640
00:40:57.600 --> 00:41:01.280
projects. What changes when your software
641
00:41:01.360 --> 00:41:04.560
sits in the execution pipeline? Yeah, so
642
00:41:04.880 --> 00:41:08.600
when, when we integrate our software, the very first effort that you
643
00:41:08.600 --> 00:41:12.080
see, suddenly you can run much more experiments for the same cost.
644
00:41:13.200 --> 00:41:16.880
So typical enterprise programs are not that
645
00:41:16.880 --> 00:41:20.480
expensive on the scale. So these are typically few people
646
00:41:21.120 --> 00:41:24.900
and few hundred thousand dollars, up to a million could be
647
00:41:24.900 --> 00:41:28.700
in compute costs. But that limits you
648
00:41:28.700 --> 00:41:31.340
in the number of experiments that you can run
649
00:41:33.500 --> 00:41:36.940
and various use cases that you can check and
650
00:41:37.020 --> 00:41:40.700
empirically experiment with. So with
651
00:41:40.780 --> 00:41:44.460
our software we can not speed up, but
652
00:41:44.460 --> 00:41:47.580
reduce the cost of those experiments by
653
00:41:47.900 --> 00:41:50.530
partially also speeding them up in
654
00:41:52.770 --> 00:41:55.730
the factor of ten or even hundreds.
655
00:41:56.530 --> 00:42:00.170
And that basically leads you to a situation when
656
00:42:00.170 --> 00:42:03.330
for the same budget you can run much more, and
657
00:42:04.370 --> 00:42:07.490
that can much faster bring you to
658
00:42:08.130 --> 00:42:11.090
better intuition, better understanding of the technology,
659
00:42:13.010 --> 00:42:16.570
and better and better
660
00:42:16.570 --> 00:42:20.250
algorithmic pipeline for your specific business
661
00:42:20.250 --> 00:42:24.010
use cases. So it's not immediately that you will solve
662
00:42:24.010 --> 00:42:27.770
all the problems, but you will get much more
663
00:42:27.770 --> 00:42:31.610
intuition and operational knowledge. Of
664
00:42:31.610 --> 00:42:35.450
how to run something on the quantum computers. And hopefully
665
00:42:36.010 --> 00:42:39.610
with few customers with whom we are working today, we also
666
00:42:39.690 --> 00:42:42.650
will bring them to the edge of quantum
667
00:42:44.090 --> 00:42:47.910
usability and userfulness when they can start up operating
668
00:42:47.910 --> 00:42:51.590
these systems along with classical high performance
669
00:42:51.670 --> 00:42:52.790
computing systems.
670
00:42:55.990 --> 00:42:59.750
You framed the buying decision around reducing the
671
00:42:59.750 --> 00:43:03.310
total time of learning rather than just the costs per
672
00:43:03.310 --> 00:43:06.950
run. How should companies think about the
673
00:43:06.950 --> 00:43:10.670
economics of experimenting with quantum computing
674
00:43:10.670 --> 00:43:12.150
today? It's quite expensive.
675
00:43:16.630 --> 00:43:19.270
We don't think about the total cost. But
676
00:43:20.400 --> 00:43:23.760
what other economics? Future payouts, future
677
00:43:23.760 --> 00:43:27.440
risks, current risk, future payouts. Right. So it
678
00:43:27.600 --> 00:43:31.280
seemingly running something on quantum computers seems like a lot, right?
679
00:43:31.840 --> 00:43:35.559
So like one hour of execution on the quantum computer can
680
00:43:35.559 --> 00:43:38.880
cost up to $50,000 or even more, depending
681
00:43:39.520 --> 00:43:41.600
on the technology and the company
682
00:43:43.520 --> 00:43:47.140
who is providing that machine to you. But in
683
00:43:47.140 --> 00:43:50.540
practice, the time
684
00:43:51.660 --> 00:43:55.420
spent on the machine and time spent for example in high
685
00:43:55.420 --> 00:43:58.700
performance computing clusters, these are different types.
686
00:43:59.660 --> 00:44:03.420
And for example, some problems you
687
00:44:03.420 --> 00:44:07.180
can still solve on the quantum computer, like let's say in one
688
00:44:07.180 --> 00:44:10.380
hour or two hours. So it seemingly costs a lot,
689
00:44:11.020 --> 00:44:14.530
but these problems might be completely impossible to solve
690
00:44:14.850 --> 00:44:18.410
on the classical computer. And for example, if you're a
691
00:44:18.410 --> 00:44:21.970
chemistry company and you need to create new
692
00:44:23.330 --> 00:44:24.690
molecular structures
693
00:44:27.649 --> 00:44:31.250
to have specific properties, you need to run those simulations
694
00:44:31.250 --> 00:44:34.130
daily and a lot of them and
695
00:44:35.010 --> 00:44:38.810
eventually you are literally limited by the size of molecules that
696
00:44:38.810 --> 00:44:42.620
you can simulate. And if quantum computers will
697
00:44:42.620 --> 00:44:46.380
be of a particularly good quality, you will be able to
698
00:44:46.380 --> 00:44:50.020
simulate them in minutes or maybe hours
699
00:44:50.180 --> 00:44:53.940
instead of weeks or months on the classical
700
00:44:53.940 --> 00:44:57.460
computer. Or sometimes you cannot do that at all on the
701
00:44:57.460 --> 00:45:00.740
classical computer. So it's eventually we will just bring
702
00:45:01.300 --> 00:45:05.020
get to the economy where you either can run something or you
703
00:45:05.020 --> 00:45:08.200
can't. So that's, it's not even like
704
00:45:09.800 --> 00:45:13.640
in terms of economics, it's just like there are some applications
705
00:45:13.640 --> 00:45:17.080
which are just impossible to will be impossible to run classically.
706
00:45:17.560 --> 00:45:20.920
You're doing very, very great. But I'm afraid we are way
707
00:45:21.159 --> 00:45:24.920
over time here. So I think we are now running around
708
00:45:25.640 --> 00:45:29.440
30, 35 to 40 minutes in maybe
709
00:45:29.440 --> 00:45:33.160
45 minutes in total. But usually that's the length of
710
00:45:33.240 --> 00:45:36.760
one of our interviews. So I'm afraid we
711
00:45:37.300 --> 00:45:40.900
do a little cutting here and go into part number two.
712
00:45:45.700 --> 00:45:49.300
That's all folks. Find more news streams,
713
00:45:49.460 --> 00:45:50.620
events and
714
00:45:50.620 --> 00:45:54.100
interviews@www.startuprad
715
00:45:54.340 --> 00:45:57.540
IO. Remember, sharing is caring.
716
00:46:03.310 --> 00:46:10.840
Sat.