Quantum's Real Bottleneck Is Data Encoding, Not Hardware

WEBVTT
1
00:00:00.160 --> 00:00:00.400
Foreign.
2
00:00:09.280 --> 00:00:12.880
Your podcast and YouTube blog covering the German
3
00:00:12.880 --> 00:00:16.480
startup scene with news interviews and
4
00:00:16.560 --> 00:00:17.840
live events.
5
00:00:20.400 --> 00:00:24.160
Hey guys, welcome back to part two of our interview with
6
00:00:24.160 --> 00:00:27.710
Nicola from Haikyuu AI, a quantum
7
00:00:27.710 --> 00:00:31.190
computing startup. He was just power
8
00:00:31.270 --> 00:00:35.110
loading us with information and we were talking about as much
9
00:00:35.110 --> 00:00:38.830
as possible in one episode. So we decided instead of the usual ad break
10
00:00:38.830 --> 00:00:42.510
in the middle, we split this in two episodes. So welcome back here
11
00:00:42.510 --> 00:00:46.150
Mikola. Welcome back and let's
12
00:00:46.150 --> 00:00:49.910
dive right in. Your anomaly detection
13
00:00:49.990 --> 00:00:53.780
work with IBM's Hero processor was
14
00:00:53.860 --> 00:00:57.580
presented as an imperial signal rather than a
15
00:00:57.580 --> 00:01:00.660
claim of quantum advantage. What
16
00:01:00.820 --> 00:01:03.620
exactly did that experiment demonstrate?
17
00:01:04.900 --> 00:01:08.100
Yeah, so that that's actually one of the recent exciting
18
00:01:08.420 --> 00:01:12.260
work in our team. So it's not yet published in the
19
00:01:12.420 --> 00:01:16.020
like proper peer reviewed journal. However
20
00:01:16.260 --> 00:01:19.820
the there is a lot of material that will go in the peer
21
00:01:19.820 --> 00:01:22.380
reviewed publication soon. But
22
00:01:23.420 --> 00:01:26.700
the great result that we got is the following. So
23
00:01:27.740 --> 00:01:31.020
one of the bottlenecks in quantum computers is data
24
00:01:31.020 --> 00:01:34.260
loading. So if you want for example to apply quantum
25
00:01:34.260 --> 00:01:37.580
computers to machine learning applications,
26
00:01:37.740 --> 00:01:41.580
the first problem that you will discover is not
27
00:01:41.580 --> 00:01:45.340
that it's hard to train or like you need to build some
28
00:01:45.340 --> 00:01:49.100
architectures of neural networks or your machine learning algorithm,
29
00:01:49.730 --> 00:01:53.570
but you basically will have troubles with encoding your
30
00:01:53.570 --> 00:01:56.210
data in the quantum computing
31
00:01:57.010 --> 00:02:00.850
memory or the state initial state. And
32
00:02:01.890 --> 00:02:04.930
typically when you do so, you need
33
00:02:05.810 --> 00:02:09.650
so deep or so many operations in your
34
00:02:09.650 --> 00:02:13.210
algorithms that it very quickly
35
00:02:13.210 --> 00:02:17.010
accumulates noise. Because quantum computers are currently noisy, they
36
00:02:17.010 --> 00:02:20.750
have errors and because of those errors your data becomes
37
00:02:20.750 --> 00:02:24.470
corrupted, so it's not operable. So you cannot really run any
38
00:02:24.630 --> 00:02:28.230
machine learning on top of this. So what we did, we
39
00:02:28.230 --> 00:02:31.670
invented a new type of algorithm which
40
00:02:32.070 --> 00:02:35.190
takes advantage of a limited
41
00:02:36.150 --> 00:02:39.830
so called Hilbert space or state space, which
42
00:02:39.830 --> 00:02:43.110
can be generated by
43
00:02:43.830 --> 00:02:47.190
shallow or few operations on the quantum computer
44
00:02:47.970 --> 00:02:51.330
before noise kicks in. So we understand which states
45
00:02:52.050 --> 00:02:55.530
this quantum computer can generate and we will try to take
46
00:02:55.530 --> 00:02:59.170
advantage of that in order to encode data directly in those
47
00:02:59.170 --> 00:03:02.730
states which are not corrupted by noise. So we
48
00:03:02.730 --> 00:03:04.290
created that algorithm and we
49
00:03:06.050 --> 00:03:09.770
decided to make a demo application of this. And first thing that
50
00:03:09.770 --> 00:03:13.570
we did, we took an anomaly detection data set
51
00:03:13.890 --> 00:03:16.850
that was one of the biology anomaly detection where you have
52
00:03:17.490 --> 00:03:21.310
time series of which look
53
00:03:21.310 --> 00:03:25.030
very, very noisy and unstructured. And because of
54
00:03:25.030 --> 00:03:28.350
that it's very hard to treat by classical
55
00:03:29.070 --> 00:03:32.870
machine learning algorithms. So they hardly do not recognize differences in
56
00:03:32.870 --> 00:03:36.430
those time series. And if you apply
57
00:03:36.430 --> 00:03:40.110
now a quantum algorithm
58
00:03:40.110 --> 00:03:43.790
to this data, what quantum algorithms
59
00:03:43.870 --> 00:03:47.660
actually do, in some sense
60
00:03:47.980 --> 00:03:51.740
they lift the dimensionality of this data into much higher
61
00:03:52.300 --> 00:03:56.060
space, high dimensional space. And in that Space you
62
00:03:56.060 --> 00:03:59.500
actually can draw so called decision boundary between different
63
00:03:59.580 --> 00:04:03.260
classes of data. It's very similar to classical
64
00:04:03.260 --> 00:04:06.540
machine learning so called kernel trick where you do the same.
65
00:04:07.260 --> 00:04:11.110
But this is a quantum kernel trick in some sense. And
66
00:04:11.110 --> 00:04:14.630
we notice that if you apply it to the specific complex data,
67
00:04:15.190 --> 00:04:18.310
which is hard for classical
68
00:04:18.390 --> 00:04:20.790
algorithm, it turns out to be
69
00:04:21.750 --> 00:04:25.390
relatively easy for quantum. And we see this first
70
00:04:25.390 --> 00:04:29.190
advantage in the, we don't call it advantage, let's
71
00:04:29.190 --> 00:04:32.790
say improvement in the
72
00:04:33.990 --> 00:04:37.710
ideal simulation. And then we run the same algorithm on
73
00:04:37.710 --> 00:04:41.430
the noisy quantum computer of the 100 qubit scale.
74
00:04:42.190 --> 00:04:45.150
And we seen that this improvement
75
00:04:45.470 --> 00:04:48.670
persists also on the noisy quantum computer.
76
00:04:50.670 --> 00:04:53.390
That's probably the first evidence of
77
00:04:54.110 --> 00:04:57.710
truly quantum data encoding and
78
00:04:57.789 --> 00:05:00.750
truly quantum machine learning algorithm
79
00:05:03.870 --> 00:05:07.550
that applied in tandem in order to improve machine
80
00:05:07.550 --> 00:05:09.550
learning application. And
81
00:05:12.270 --> 00:05:15.730
what's important as well, in this particular blog that we
82
00:05:15.730 --> 00:05:18.010
published, we limited ourselves to
83
00:05:20.490 --> 00:05:24.210
a number of qubits and the depth number of
84
00:05:24.210 --> 00:05:28.050
operations of algorithm which are still simulable on the
85
00:05:28.050 --> 00:05:31.850
classical computer. We did it on purpose in
86
00:05:31.850 --> 00:05:35.410
order to have a benchmark to compare ourselves
87
00:05:35.410 --> 00:05:39.130
with. Otherwise if you will just run it in the
88
00:05:40.800 --> 00:05:43.840
on the quantum computer, in real life, in hardware,
89
00:05:45.360 --> 00:05:48.640
it will be hard to justify that this is not an effect of some noise
90
00:05:48.960 --> 00:05:52.600
or some imperfection of the hardware, rather than actual quantum
91
00:05:52.600 --> 00:05:56.080
effect. So we first proved that here is an ideal simulation,
92
00:05:56.480 --> 00:06:00.320
here we got improvement in the performance, and here is the
93
00:06:00.480 --> 00:06:04.240
hardware result where we still see persistence of that improvement.
94
00:06:05.240 --> 00:06:08.840
And in practice we can still scale
95
00:06:08.840 --> 00:06:12.440
it to the limit, which is beyond possibility
96
00:06:12.440 --> 00:06:16.280
of classical simulation. That's where it's a little bit
97
00:06:16.280 --> 00:06:19.960
harder to control. But we still can actually
98
00:06:20.040 --> 00:06:21.640
load the data at that scale.
99
00:06:25.320 --> 00:06:28.920
You've argued in the past
100
00:06:29.000 --> 00:06:32.340
that the hardest bottleneck in
101
00:06:32.340 --> 00:06:35.580
quantum machine learning is
102
00:06:35.900 --> 00:06:39.460
scalable data encoding rather than the
103
00:06:39.460 --> 00:06:43.140
model architecture. Why is embedding
104
00:06:43.140 --> 00:06:46.300
classical data into quantum computing systems
105
00:06:46.700 --> 00:06:50.220
such a difficult problem? Yeah, that's
106
00:06:50.620 --> 00:06:54.260
another story of a situation when the
107
00:06:54.260 --> 00:06:57.860
machines are here, but we still need
108
00:06:57.860 --> 00:07:01.260
to discover algorithms, like best algorithms to
109
00:07:01.260 --> 00:07:04.910
encode data or manipulate that data on the quantum
110
00:07:04.910 --> 00:07:08.510
computers. And we think we discovered one of the best
111
00:07:08.510 --> 00:07:11.750
algorithms in its class because we literally take advantage of
112
00:07:12.470 --> 00:07:15.750
the part of the state space which is
113
00:07:16.390 --> 00:07:20.230
created on these small scale early machines.
114
00:07:23.030 --> 00:07:26.630
But maybe there could be alternative approaches. So why
115
00:07:26.630 --> 00:07:30.070
it's hard because you are limited with just hundreds of qubits.
116
00:07:30.410 --> 00:07:31.930
And some real world data
117
00:07:34.090 --> 00:07:37.370
can require thousands of features or millions of
118
00:07:37.690 --> 00:07:40.730
features, for example in the image or some
119
00:07:40.890 --> 00:07:44.490
satellite data. And
120
00:07:44.650 --> 00:07:48.410
then you need basically to find the way how to represent
121
00:07:48.490 --> 00:07:52.010
that on hundreds of qubits. So you can have in one
122
00:07:52.010 --> 00:07:55.770
dimensionality you have hundreds of qubits, but then in another dimension you
123
00:07:55.770 --> 00:07:59.500
can kind of re encode your data by adding more and more
124
00:07:59.500 --> 00:08:02.940
operations. And each operation will have like small rotation
125
00:08:03.580 --> 00:08:07.300
or small kind of encoding algorithm, how you encode
126
00:08:07.300 --> 00:08:10.540
more and more features, but that will enlarge the number of
127
00:08:11.340 --> 00:08:14.780
operations that you need to use and depth of your
128
00:08:14.780 --> 00:08:18.620
algorithm. And eventually, eventually your noise kicks
129
00:08:18.620 --> 00:08:22.380
in and noise starts to beat you for every operation
130
00:08:22.380 --> 00:08:25.880
that you actually add into your algorithm. So that's why there is this
131
00:08:25.880 --> 00:08:27.920
huge trade off on one hand. So.
132
00:08:33.120 --> 00:08:36.520
You can still encode more features than number of
133
00:08:36.520 --> 00:08:40.319
qubits, but then you pay for the number of operations. So at some
134
00:08:40.319 --> 00:08:44.000
point you just have very limited room
135
00:08:44.160 --> 00:08:46.640
in which you can do something. And
136
00:08:47.840 --> 00:08:49.040
there are so called
137
00:08:51.130 --> 00:08:54.490
amplitude embedding algorithms
138
00:08:54.650 --> 00:08:58.450
which are more computer more kind of, I would say
139
00:08:58.450 --> 00:09:01.770
than classical quantum computing algorithms which take
140
00:09:01.770 --> 00:09:05.450
advantage of the ideal scenario when you don't have any
141
00:09:05.450 --> 00:09:09.210
errors. But in practice you have quantum computers
142
00:09:09.210 --> 00:09:12.970
which actually have errors and you pay for those operations, so you
143
00:09:12.970 --> 00:09:14.170
cannot really apply those.
144
00:09:19.010 --> 00:09:22.730
You also said embeddings may ultimately
145
00:09:22.730 --> 00:09:25.730
determine whether quantum machine learning
146
00:09:26.210 --> 00:09:29.890
scales at all. What role do they play
147
00:09:30.530 --> 00:09:33.010
in a broader quantum computing stack?
148
00:09:34.050 --> 00:09:37.370
Yeah, so if you think about any industrial
149
00:09:37.370 --> 00:09:41.050
application, so you always need to
150
00:09:41.050 --> 00:09:44.630
do some kind of state embedding, data
151
00:09:44.630 --> 00:09:47.830
embedding, initial conditions, embedding
152
00:09:48.470 --> 00:09:52.030
and so on. So for example, if you do computational fluid
153
00:09:52.030 --> 00:09:55.510
dynamics, you start with some density
154
00:09:55.510 --> 00:09:59.190
profile of your fluid, so that needs to be
155
00:09:59.190 --> 00:10:02.950
somehow encoded in the initial state of your simulation. So
156
00:10:02.950 --> 00:10:06.710
that's again a data encoding problem. If you
157
00:10:06.710 --> 00:10:10.510
have a machine learning problem, you also need to encode the
158
00:10:10.510 --> 00:10:13.200
data if you have
159
00:10:14.640 --> 00:10:17.840
for example Monte Carlo
160
00:10:18.480 --> 00:10:22.280
problem which people use in finance, for example
161
00:10:22.280 --> 00:10:25.640
for derivatives pricing or some financial risk
162
00:10:25.640 --> 00:10:28.400
estimation. So you start by encoding a
163
00:10:29.440 --> 00:10:33.040
distribution into quantum computer. So again these
164
00:10:33.040 --> 00:10:36.640
distributions can have multidimensional structure and
165
00:10:36.640 --> 00:10:40.440
complex form. And that's again a data encoding
166
00:10:40.440 --> 00:10:44.080
problem, but of a different kind. And if you
167
00:10:44.080 --> 00:10:47.760
don't solve this problem, there is no point of
168
00:10:47.760 --> 00:10:51.320
running any algorithm because like you are killed by noise. And
169
00:10:51.560 --> 00:10:55.400
you can wait for 12 to 20 years for
170
00:10:55.560 --> 00:10:59.160
fault tolerant quantum computers which will be completely without
171
00:10:59.160 --> 00:11:02.840
errors. But that's probably not a good
172
00:11:02.840 --> 00:11:06.620
strategy if you can run already. Something interesting today I'm
173
00:11:06.620 --> 00:11:10.260
just learning quite a lot about quantum computing and I do believe for
174
00:11:10.260 --> 00:11:14.060
me and a lot of non technical people out there, that may
175
00:11:14.220 --> 00:11:17.580
take days, weeks or even months to settle and really
176
00:11:17.580 --> 00:11:21.140
realize what that all means. But let us go a little bit to
177
00:11:21.140 --> 00:11:24.820
HIQ here. You released Rivet, an open
178
00:11:24.820 --> 00:11:28.380
source toolkit for quantum workflow execution.
179
00:11:28.860 --> 00:11:32.700
Why did you decide to open source this layer instead of keeping
180
00:11:32.780 --> 00:11:36.520
everything proprietary? That's a
181
00:11:36.520 --> 00:11:40.120
great question. So we have multiple
182
00:11:40.120 --> 00:11:43.960
tools in our stack and this is
183
00:11:43.960 --> 00:11:47.440
not the only tool that we have. We realize that
184
00:11:47.840 --> 00:11:50.480
there is some technological advantage that we have in
185
00:11:51.600 --> 00:11:55.360
for example, reducing the effects of noise or optimizing
186
00:11:55.600 --> 00:11:58.320
quantum algorithms on more high level optimization level.
187
00:11:59.360 --> 00:12:02.720
And there are other components like
188
00:12:03.470 --> 00:12:07.110
compilation, where we also needed to build some
189
00:12:07.110 --> 00:12:10.670
tools for ourselves and many of those
190
00:12:10.910 --> 00:12:14.750
still don't exist and we still are building a bunch of those for
191
00:12:14.750 --> 00:12:18.510
ourselves and if we found them very
192
00:12:18.510 --> 00:12:22.350
interesting and convenient to use. For example, Rivet
193
00:12:22.590 --> 00:12:25.990
allows you to split your larger algorithm in
194
00:12:25.990 --> 00:12:29.630
chunks and transpile or compile different
195
00:12:29.870 --> 00:12:33.520
chunks independently. And you can
196
00:12:33.520 --> 00:12:36.960
apply different compilation algorithms depending of what different
197
00:12:36.960 --> 00:12:40.640
parts of your algorithm is. And that was very hard to
198
00:12:40.640 --> 00:12:44.440
do before Revit and now it's very easy to do with Revit,
199
00:12:44.920 --> 00:12:48.280
but now where it helps. So our partner
200
00:12:49.400 --> 00:12:53.160
who used this for quantum machine learning application,
201
00:12:53.160 --> 00:12:55.960
discovered that you can use this
202
00:12:56.520 --> 00:12:59.720
partial compilation in order to train quantum
203
00:13:00.040 --> 00:13:03.210
machine learning neural networks. So you can
204
00:13:03.610 --> 00:13:07.450
add layers gradually and compile only part
205
00:13:07.450 --> 00:13:11.170
of your algorithm or neural network instead
206
00:13:11.170 --> 00:13:14.250
of whole every time you update the parameters.
207
00:13:15.130 --> 00:13:18.570
In that case you save up to
208
00:13:18.570 --> 00:13:21.690
100 times on the
209
00:13:21.770 --> 00:13:25.450
classical compute, which you need to spend just like recompiling
210
00:13:25.770 --> 00:13:29.610
your circuit every time you update parameters in neural network. So that's one
211
00:13:29.610 --> 00:13:33.400
small application. But there are plenty of
212
00:13:33.560 --> 00:13:37.280
different other applications that we think people
213
00:13:37.280 --> 00:13:40.120
will benefit from. And for us
214
00:13:41.560 --> 00:13:45.000
it's not a crucial component of
215
00:13:45.080 --> 00:13:48.360
our stack. So we have some components which
216
00:13:48.680 --> 00:13:52.400
actually give us 10 to 100 times advantage over our
217
00:13:52.400 --> 00:13:56.120
competitors. Rivet is of course super
218
00:13:56.120 --> 00:13:59.640
useful, but that's something that we can provide to community
219
00:13:59.800 --> 00:14:03.620
and get them excited. And soon there will be more
220
00:14:03.620 --> 00:14:07.460
open source contributions from us. We are very excited actually
221
00:14:07.460 --> 00:14:10.780
to share some very interesting infrastructural
222
00:14:10.780 --> 00:14:14.460
components. Can you maybe give us a little
223
00:14:14.460 --> 00:14:17.940
tease, a little hint on what that may be here?
224
00:14:17.940 --> 00:14:21.580
Exclusive for our startup rate IO audience. Yeah.
225
00:14:21.580 --> 00:14:25.180
So one of the
226
00:14:25.180 --> 00:14:28.900
advantages that we have is ability to
227
00:14:29.380 --> 00:14:33.140
build very complex middleware stacks.
228
00:14:33.930 --> 00:14:37.690
So when we started operating at the space,
229
00:14:38.170 --> 00:14:41.130
we realized it's not that easy to combine different
230
00:14:42.970 --> 00:14:46.650
tools on the middleware layer. So a lot of these,
231
00:14:48.170 --> 00:14:51.930
either ours which are available
232
00:14:51.930 --> 00:14:55.570
from our competitors or open source, they have
233
00:14:55.570 --> 00:14:58.410
their own interfaces, they have their own
234
00:15:00.430 --> 00:15:04.190
like infrastructure in which they embed. So it's very hard
235
00:15:04.190 --> 00:15:07.990
to combine like for example, compilation layer from one
236
00:15:07.990 --> 00:15:11.550
provider, error mitigation layer from another
237
00:15:11.710 --> 00:15:15.470
provider error Correction layer from the search provider and so on.
238
00:15:15.790 --> 00:15:19.550
So that's where we realize that we can
239
00:15:19.550 --> 00:15:22.990
build an infrastructure which allows
240
00:15:23.390 --> 00:15:26.200
us to, to
241
00:15:26.840 --> 00:15:30.600
basically create a kind of a
242
00:15:30.600 --> 00:15:34.200
tissue into which we place those components
243
00:15:34.280 --> 00:15:38.120
and which connects them very naturally, such that
244
00:15:38.120 --> 00:15:41.920
we can build probably some of the most
245
00:15:41.920 --> 00:15:45.640
complex middleware stacks
246
00:15:45.800 --> 00:15:49.640
on the market, where each component is either state
247
00:15:49.640 --> 00:15:53.150
of the art or better than state of the art. But when they
248
00:15:53.150 --> 00:15:56.510
combine, when we combine all of them together, we can reach
249
00:15:56.990 --> 00:16:00.350
performance beyond any other results in the industry.
250
00:16:00.510 --> 00:16:04.150
So that's something that
251
00:16:04.150 --> 00:16:06.910
we think would be very beneficial
252
00:16:08.910 --> 00:16:12.590
for the broader community. And we are working on
253
00:16:13.870 --> 00:16:16.670
open sourcing part of this infrastructure such that
254
00:16:17.700 --> 00:16:21.540
you can combine our stack with
255
00:16:22.260 --> 00:16:25.940
stacks of other companies of open source
256
00:16:26.580 --> 00:16:30.380
tools. And as a researcher, you will be able
257
00:16:30.380 --> 00:16:33.940
to study how, for example, I can
258
00:16:33.940 --> 00:16:37.300
build some specific
259
00:16:37.860 --> 00:16:41.700
noise mitigation with some specific error noise tailoring
260
00:16:42.020 --> 00:16:45.860
technique and study how they combine together.
261
00:16:45.940 --> 00:16:49.740
So it's typically not that easy to do now. You don't need
262
00:16:49.740 --> 00:16:53.420
to spend like months or half a year of
263
00:16:53.420 --> 00:16:57.220
doing that manually. You will have infrastructure which will allow you to do that.
264
00:16:58.180 --> 00:17:01.820
Talking about infrastructure here, my understanding is some of
265
00:17:01.820 --> 00:17:04.980
the future releases that are coming from HIQ
266
00:17:05.460 --> 00:17:09.260
will cost money. And for founders evaluating HiQ, the
267
00:17:09.260 --> 00:17:12.500
Q question becomes whether they are buying
268
00:17:12.580 --> 00:17:15.840
infrastructure leverage or, or temporary optimization.
269
00:17:16.560 --> 00:17:20.360
How would you describe that distinction to a potential
270
00:17:20.360 --> 00:17:23.760
customer? That's a good question. So on one hand
271
00:17:24.240 --> 00:17:27.760
I have my moments. Yeah,
272
00:17:29.520 --> 00:17:31.280
on one hand we have
273
00:17:33.520 --> 00:17:36.880
some of the best in class components and I would say that
274
00:17:36.960 --> 00:17:40.640
you're correct that these are components which
275
00:17:40.640 --> 00:17:43.800
allow you to get, for example, better
276
00:17:44.200 --> 00:17:47.800
error mitigation or better optimization of your algorithm,
277
00:17:48.040 --> 00:17:51.760
which you can apply immediately today. But as computers will
278
00:17:51.760 --> 00:17:54.840
get better, probably noise will become less of a problem.
279
00:17:55.640 --> 00:17:58.840
So maybe you will not need to care that much about noise.
280
00:17:59.560 --> 00:18:02.920
So sooner or later these noise mitigation
281
00:18:03.400 --> 00:18:06.520
tools will probably become less relevant.
282
00:18:07.240 --> 00:18:10.920
However, the infrastructure will still be there. So there will be new components
283
00:18:10.920 --> 00:18:13.240
which need to be integrated. There will be new
284
00:18:15.080 --> 00:18:18.280
challenges of integrating error correction,
285
00:18:20.600 --> 00:18:22.760
encoding of logical qubits,
286
00:18:23.960 --> 00:18:27.640
decoding mechanism, and so on and so forth,
287
00:18:28.040 --> 00:18:31.760
which will need a special software layer
288
00:18:31.760 --> 00:18:34.520
to integrate with each other. So
289
00:18:35.490 --> 00:18:39.170
what's exciting about what we are building today, we
290
00:18:39.170 --> 00:18:42.810
can integrate different middleware components which
291
00:18:42.810 --> 00:18:46.090
allow us to run at scale. At present quantum
292
00:18:46.090 --> 00:18:49.810
computers. Tomorrow the same infrastructure would allow us to run
293
00:18:51.170 --> 00:18:54.450
error correction, logical qubits
294
00:18:54.450 --> 00:18:58.250
encoding, and the day after tomorrow it would allow us
295
00:18:58.250 --> 00:19:01.410
to run and orchestrate different quantum algorithms.
296
00:19:02.050 --> 00:19:05.650
So it's the same infrastructure. It scales as we mature with the,
297
00:19:06.650 --> 00:19:08.810
with the quantum hardware Performance.
298
00:19:14.490 --> 00:19:18.290
In our interview, I'm not sure if it was this part or part
299
00:19:18.290 --> 00:19:21.650
one. You've suggested that the first three
300
00:19:21.650 --> 00:19:25.370
customers for quantum computing middleware are teams already
301
00:19:25.450 --> 00:19:29.050
struggling with failed pilots. What
302
00:19:29.130 --> 00:19:32.810
pattern do you see in companies that come to you
303
00:19:33.380 --> 00:19:37.060
after those early experiments? Yeah,
304
00:19:38.740 --> 00:19:40.020
that's actually quite funny.
305
00:19:42.500 --> 00:19:46.180
We've seen that several times. The companies start their
306
00:19:46.820 --> 00:19:48.980
quantum program and
307
00:19:50.820 --> 00:19:52.740
they never run anything on the hardware.
308
00:19:56.660 --> 00:20:00.260
That means they have very expensive, high maintenance
309
00:20:00.420 --> 00:20:03.970
quantum hardware in their laboratories.
310
00:20:04.370 --> 00:20:08.170
That's good when they have like typically they don't. I think there
311
00:20:08.170 --> 00:20:11.810
are few, few institutions in the world who actually have
312
00:20:12.690 --> 00:20:15.090
quantum hardware in their own environment.
313
00:20:16.770 --> 00:20:20.610
Let's say majority actually has one or another
314
00:20:20.610 --> 00:20:24.330
partnership with quantum hardware provider. However,
315
00:20:24.330 --> 00:20:28.120
it's quite interesting, curious to see that. And many
316
00:20:28.120 --> 00:20:31.840
of them run very few experiments on real hardware. And
317
00:20:31.840 --> 00:20:35.560
the reason for that they struggle when they
318
00:20:35.560 --> 00:20:39.400
try to run something on the hardware. They just get poor
319
00:20:39.400 --> 00:20:43.080
results and kind of disappointed by this
320
00:20:43.080 --> 00:20:46.520
process and just focus on the algorithmic performance or
321
00:20:46.520 --> 00:20:48.840
algorithm design. However that
322
00:20:50.200 --> 00:20:54.000
on one hand it allows you to discover new algorithm, but at some
323
00:20:54.000 --> 00:20:57.840
point you still need to run those algorithms on the quantum hardware in order to
324
00:20:57.840 --> 00:21:01.380
build intuition what runs what not. For example, when we worked with
325
00:21:02.900 --> 00:21:06.540
one large financial institution first we
326
00:21:06.540 --> 00:21:09.380
discovered ability to
327
00:21:09.860 --> 00:21:13.700
encode heavy tailed financial
328
00:21:13.780 --> 00:21:16.500
distributions into quantum hardware
329
00:21:17.300 --> 00:21:21.100
with few operations. Then this
330
00:21:21.100 --> 00:21:24.940
allowed us to run now algorithms on top of
331
00:21:24.940 --> 00:21:28.740
this data that we load. But it was still
332
00:21:28.740 --> 00:21:32.460
hard to do so before us people just
333
00:21:32.620 --> 00:21:36.060
did the same exercise on like three to five to six
334
00:21:36.060 --> 00:21:39.740
qubits maybe. And we were able to scale this
335
00:21:39.740 --> 00:21:43.380
to dozens of qubits. And the
336
00:21:43.380 --> 00:21:47.020
reason we were able to do so because when we started to experiment,
337
00:21:47.020 --> 00:21:50.380
we faced some other challenges on the algorithmic level.
338
00:21:50.620 --> 00:21:54.390
And that's actually inspired us to think differently. And we
339
00:21:54.390 --> 00:21:57.910
discovered some algorithmic tricks which allowed us to reduce the
340
00:21:57.910 --> 00:22:01.230
depths of the algorithm itself such that we could
341
00:22:01.470 --> 00:22:05.310
load the data and then run the algorithm at the largest possible scale
342
00:22:05.710 --> 00:22:09.470
on the real hardware. And before us, for example, the same
343
00:22:09.630 --> 00:22:13.390
company actually worked with most of the
344
00:22:13.390 --> 00:22:16.830
major quantum hardware providers.
345
00:22:17.310 --> 00:22:20.950
And while hardware was available, they were not able
346
00:22:20.950 --> 00:22:23.510
to particularly execute this specific
347
00:22:26.310 --> 00:22:30.030
and my assumption is not any
348
00:22:30.030 --> 00:22:33.670
E commerce shoppers currently investing in quantum computing. It
349
00:22:33.670 --> 00:22:37.469
will be those that require like simulations of the highest
350
00:22:37.469 --> 00:22:39.670
complexity drug discovery
351
00:22:41.830 --> 00:22:44.550
molecules. Stuff like this comes to mind.
352
00:22:46.230 --> 00:22:49.990
Yeah, so these are very non trivial problems right now.
353
00:22:49.990 --> 00:22:53.790
And obviously not every company needs to invest in
354
00:22:53.790 --> 00:22:57.190
quantum computing today. So mostly these are companies
355
00:22:57.270 --> 00:23:01.030
who literally have large or
356
00:23:01.350 --> 00:23:04.870
heavy high performance computing loads.
357
00:23:05.350 --> 00:23:08.790
So for example automotive, aerospace, those companies
358
00:23:08.790 --> 00:23:12.510
obviously have a lot of such applications. On one hand you have a lot
359
00:23:12.510 --> 00:23:16.230
of those computational fluid dynamics workflows. This can be hundreds
360
00:23:16.230 --> 00:23:19.070
of thousands
361
00:23:20.750 --> 00:23:24.510
dollars or could be millions of dollars per year spent.
362
00:23:26.430 --> 00:23:30.190
I had roommates who were studying engineering and he could really scare
363
00:23:30.190 --> 00:23:33.150
them with fluid dynamics. Yeah,
364
00:23:34.190 --> 00:23:37.910
it's very non trivial problem. It's actually beaten to death in some sense in quantum
365
00:23:37.910 --> 00:23:41.590
classical computing, such that it's so well optimized that it's very
366
00:23:41.590 --> 00:23:45.390
hard to beat immediately those results. But as soon as quantum
367
00:23:45.390 --> 00:23:49.190
computers will get good enough. So that
368
00:23:49.190 --> 00:23:52.950
will happen very quickly and the performance
369
00:23:53.270 --> 00:23:56.710
of the quantum computational fluid dynamics algorithm
370
00:23:57.590 --> 00:24:01.150
will be quite impressive. But today we are on the toy
371
00:24:01.150 --> 00:24:04.830
examples scale and the same happens with optimization
372
00:24:04.830 --> 00:24:07.910
problems. That's another kind of potentially low hanging fruit.
373
00:24:09.030 --> 00:24:12.630
We are working right now in the lower qubit regime where we
374
00:24:12.630 --> 00:24:16.350
have optimization problems of so called few
375
00:24:16.350 --> 00:24:19.870
degrees of freedoms. And typically those are very easy for
376
00:24:19.870 --> 00:24:23.310
classical computers. But again as soon as you reach
377
00:24:23.470 --> 00:24:27.269
few thousand degrees of freedom, that already becomes
378
00:24:27.269 --> 00:24:30.350
very very hard. And that will require
379
00:24:31.070 --> 00:24:34.710
several hundred to thousands of qubits. And again
380
00:24:34.710 --> 00:24:38.030
that's not too far away if you look on the roadmaps of what
381
00:24:38.350 --> 00:24:41.970
hardware provider offer us. So yeah, I'm
382
00:24:42.130 --> 00:24:45.850
quite optimistic where we're going. And again chemistry problems
383
00:24:45.850 --> 00:24:49.490
and material design problems are probably the lowest hanging fruits
384
00:24:49.490 --> 00:24:53.250
here. You're
385
00:24:53.250 --> 00:24:56.970
very optimistic. But let's talk a little bit in the future and if
386
00:24:56.970 --> 00:25:00.570
hardware progress slows for the next couple of years, your
387
00:25:00.570 --> 00:25:04.330
thesis either becomes even more important or much
388
00:25:04.330 --> 00:25:07.490
harder to prove. How do you think about that scenario?
389
00:25:08.460 --> 00:25:11.180
Hopefully hardware progress will not slow. So we have
390
00:25:12.140 --> 00:25:15.660
several competing technologies on the
391
00:25:15.660 --> 00:25:19.420
roadmap. So there are many different qubit modalities.
392
00:25:19.500 --> 00:25:23.340
I would say it's again very similar to what happened in classical
393
00:25:23.340 --> 00:25:24.620
computing where we had
394
00:25:26.780 --> 00:25:29.820
multiple transistor technologies up until
395
00:25:30.140 --> 00:25:33.940
60s, mid-60s, 70s before we settled
396
00:25:33.940 --> 00:25:37.420
on the specific architecture that we
397
00:25:37.580 --> 00:25:41.370
using up until today. And the same story
398
00:25:41.370 --> 00:25:44.890
kind of repeats in quantum computing. So we have multiple qubit
399
00:25:44.890 --> 00:25:48.610
technologies. We still fight for scaling those
400
00:25:48.610 --> 00:25:52.170
number of qubits, make them more stable, but there is
401
00:25:52.250 --> 00:25:55.530
consistent progress and it's actually quite impressive.
402
00:25:57.130 --> 00:26:00.970
I remember when I was in university actually building
403
00:26:01.690 --> 00:26:05.410
some theoretical arguments of what will be the
404
00:26:05.410 --> 00:26:08.880
fastest time of flipping qubits or whatever. And
405
00:26:11.040 --> 00:26:14.560
this looked like very much a theoretical
406
00:26:14.560 --> 00:26:16.560
exercise which will never be
407
00:26:18.240 --> 00:26:21.920
even considered in real world experiment
408
00:26:21.920 --> 00:26:25.560
on some hardware. And then in just
409
00:26:25.560 --> 00:26:28.480
few years I would say we started seeing the first
410
00:26:29.360 --> 00:26:33.040
quantum computers operating with
411
00:26:33.040 --> 00:26:36.800
a number of qubits where you can already run some algorithms. So
412
00:26:36.800 --> 00:26:40.600
it's you, once you consider this Perspective, everything becomes
413
00:26:40.600 --> 00:26:44.320
like clear that we are literally sitting on the exponent and this exponent
414
00:26:44.320 --> 00:26:46.400
will actually soon lift off.
415
00:26:50.640 --> 00:26:54.240
That's a very optimistic, forward looking
416
00:26:54.240 --> 00:26:58.040
statement. And also looking forward, I've been looking a little bit
417
00:26:58.040 --> 00:27:01.440
at your roadmap and this hints at something closer to
418
00:27:01.440 --> 00:27:04.480
quantum operating system than a single tool.
419
00:27:05.870 --> 00:27:09.710
What would a complete quantum software stack
420
00:27:09.950 --> 00:27:12.510
actually look like? Exactly. So
421
00:27:14.030 --> 00:27:17.630
that's not fully settled now, but obviously
422
00:27:19.230 --> 00:27:22.870
you can already draw a lot of parallels with
423
00:27:22.870 --> 00:27:26.630
classical computing and we don't need to reinvent
424
00:27:26.630 --> 00:27:30.310
the wheel here literally. So we have different type of
425
00:27:30.310 --> 00:27:33.990
hardware technology, but the software
426
00:27:33.990 --> 00:27:37.630
stack is very similar. So we need some kind of
427
00:27:37.630 --> 00:27:41.390
quantum assembly. So that's already done. We need some kind
428
00:27:41.390 --> 00:27:45.190
of quantum intermediate representation for lower
429
00:27:45.190 --> 00:27:48.150
level transpilation and
430
00:27:48.470 --> 00:27:52.270
compilation of our algorithm. That's something that a number of
431
00:27:52.270 --> 00:27:56.030
companies currently working with and we also involved in some
432
00:27:56.030 --> 00:27:59.760
of these initiatives. Then as you move upper and
433
00:27:59.760 --> 00:28:03.560
upper and upper to the stack, at some point you get to the algorithms. Like
434
00:28:03.560 --> 00:28:06.560
someone wants to run, for example,
435
00:28:07.120 --> 00:28:10.800
a quantum Monte Carlo algorithm. Another person can
436
00:28:10.880 --> 00:28:14.480
have developed an algorithm for loading data
437
00:28:15.040 --> 00:28:18.720
and then you want maybe to combine those algorithms somehow.
438
00:28:18.800 --> 00:28:22.360
So for that you need something similar to
439
00:28:22.360 --> 00:28:25.840
operating system, where operating system, it's a
440
00:28:26.160 --> 00:28:29.960
system which orchestrates data and compute flows between
441
00:28:29.960 --> 00:28:33.760
different applications. And today you
442
00:28:33.760 --> 00:28:37.520
already see some early stages of that on the level of middleware.
443
00:28:37.600 --> 00:28:41.360
That's where we operate right now. But as we will
444
00:28:41.360 --> 00:28:45.120
move up the stack sooner or later, we will not work
445
00:28:45.120 --> 00:28:48.920
with isolated quantum programs. But similarly to how
446
00:28:48.920 --> 00:28:52.770
we work with classical computers, we can run multiple programs
447
00:28:52.770 --> 00:28:55.850
in parallel. Those programs communicate between each other.
448
00:28:56.410 --> 00:29:00.090
So eventually we will get to that future in quantum computing
449
00:29:00.090 --> 00:29:03.610
and you will need infrastructure for that. So
450
00:29:03.770 --> 00:29:07.290
the things that we today are building for this lower level
451
00:29:07.290 --> 00:29:10.970
stack, they naturally
452
00:29:11.450 --> 00:29:14.850
project into that future, basically
453
00:29:14.850 --> 00:29:18.610
offering these capabilities of allowing different programs to communicate,
454
00:29:18.610 --> 00:29:20.090
which is with each other.
455
00:29:25.930 --> 00:29:29.490
At this stage, the only remaining question is whether you're
456
00:29:29.490 --> 00:29:33.170
still structurally positioned to act on what
457
00:29:33.170 --> 00:29:35.370
you've just heard. Guys out there,
458
00:29:37.130 --> 00:29:40.490
let's get in the. In the very last few questions,
459
00:29:40.810 --> 00:29:44.650
Mikola here. Deep tech founders often have to build
460
00:29:44.650 --> 00:29:48.210
conviction long before the validation appears.
461
00:29:48.370 --> 00:29:52.050
How do you personally manage that psychological
462
00:29:52.050 --> 00:29:55.570
tension? Yeah,
463
00:29:55.570 --> 00:29:59.410
that's. On one hand, that's hard. On the other
464
00:29:59.410 --> 00:30:02.690
hand, I would say most of the deep tech founders, they have this
465
00:30:03.730 --> 00:30:07.410
very strong intuition. These are typically scientists or like people
466
00:30:07.410 --> 00:30:11.250
who are deeply in some technological area for all their
467
00:30:11.250 --> 00:30:14.800
life. So it's not based just like on the
468
00:30:15.200 --> 00:30:18.880
blind vision. It's all based actually.
469
00:30:19.280 --> 00:30:22.400
It's deeply rooted in Science. So our team is
470
00:30:22.800 --> 00:30:25.520
majority, probably 90% of our team
471
00:30:26.640 --> 00:30:30.320
scientists. So they have PhDs or some high degrees
472
00:30:31.280 --> 00:30:34.560
from good universities. And we have
473
00:30:34.880 --> 00:30:38.470
deep understanding where everything is heading, where
474
00:30:38.470 --> 00:30:41.990
hardware providers are moving, what is the
475
00:30:41.990 --> 00:30:45.430
roadmap, what is marketing roadmap and what is realistic
476
00:30:45.430 --> 00:30:49.070
roadmap. We also can
477
00:30:49.070 --> 00:30:52.270
always look back in what happened in classical computing.
478
00:30:52.750 --> 00:30:56.390
I'm a huge fan actually of scientific and engineering
479
00:30:56.390 --> 00:30:59.950
history. I like to draw a lot of parallels between what
480
00:30:59.950 --> 00:31:03.790
happened in the past and what lessons we can actually bring to the future.
481
00:31:04.710 --> 00:31:08.310
And also, even while working in the industry
482
00:31:08.390 --> 00:31:11.190
for just six years before
483
00:31:12.390 --> 00:31:15.510
founding Haiku, I witnessed deep learning
484
00:31:15.750 --> 00:31:19.470
revolution and I witnessed what happened to Blockchain,
485
00:31:19.470 --> 00:31:23.190
for example. And I've noticed like
486
00:31:23.190 --> 00:31:26.150
some similarities in two of these technologies
487
00:31:26.710 --> 00:31:30.270
and how. And also what's
488
00:31:30.270 --> 00:31:34.020
interesting, augmented reality and virtual reality, that's another kind
489
00:31:34.020 --> 00:31:37.500
of exponential technology. So while working in these
490
00:31:38.540 --> 00:31:42.100
areas, I witnessed what worked, what didn't work, for
491
00:31:42.100 --> 00:31:45.260
example, what made AI
492
00:31:46.460 --> 00:31:50.220
widely adopted and why, for example, augmented reality
493
00:31:50.220 --> 00:31:54.060
is still not there. Like why not everyone wearing augmented reality
494
00:31:54.060 --> 00:31:57.780
glasses and we still have mobile phones around. Everyone predicted that there will
495
00:31:57.780 --> 00:32:01.470
be no mobile phones already. So
496
00:32:01.550 --> 00:32:05.110
yeah, so a lot of these, you can draw parallels and think about
497
00:32:05.110 --> 00:32:08.910
your particular space and where it's heading. And that's,
498
00:32:08.910 --> 00:32:12.750
I would say, what kind of gives me a lot of
499
00:32:13.470 --> 00:32:17.150
confidence about what we are doing and understanding that what
500
00:32:17.150 --> 00:32:20.030
we are doing is right. We
501
00:32:20.750 --> 00:32:23.390
so far did not need to pivot too much.
502
00:32:24.830 --> 00:32:28.110
We had actually quite well defined trajectory
503
00:32:28.610 --> 00:32:32.130
and we achieved very well defined milestones on the way
504
00:32:32.930 --> 00:32:36.530
to our goal. So I think that's what makes us
505
00:32:37.410 --> 00:32:40.770
very unique and at the same time confident in the future.
506
00:32:42.290 --> 00:32:45.890
I would partly agree with you. I've learned especially
507
00:32:46.050 --> 00:32:49.810
in economic history. I love the saying, history does not repeat
508
00:32:49.810 --> 00:32:53.290
itself, but it rhymes. You've
509
00:32:53.290 --> 00:32:57.040
spent years operating IT areas where being directly
510
00:32:57.040 --> 00:33:00.280
right can take a long time to prove.
511
00:33:00.680 --> 00:33:04.160
What mental models helped you navigate that
512
00:33:04.160 --> 00:33:06.360
uncertainty? I would say
513
00:33:09.400 --> 00:33:13.000
I operate as a researcher, I always treat myself as scientist.
514
00:33:16.120 --> 00:33:19.000
And fortunately my scientific area
515
00:33:20.120 --> 00:33:23.680
is such that you find
516
00:33:23.840 --> 00:33:26.400
a lot of beauty and
517
00:33:28.080 --> 00:33:30.320
excitement in what you are doing
518
00:33:32.320 --> 00:33:36.080
from those interesting theories and
519
00:33:36.080 --> 00:33:39.840
interesting results that you start getting
520
00:33:40.720 --> 00:33:43.600
even on the daily basis. And
521
00:33:44.480 --> 00:33:48.080
at the same time, the work in theoretical physics, I think it's unique
522
00:33:48.080 --> 00:33:51.690
because like as many theoretical physicists know, that
523
00:33:52.250 --> 00:33:56.090
results of your work is typically
524
00:33:56.330 --> 00:34:00.170
seen in many, many, many years from now.
525
00:34:01.450 --> 00:34:04.250
So I think that helps a lot in deep tech
526
00:34:05.050 --> 00:34:08.650
or in the area which we are today. On one hand,
527
00:34:09.610 --> 00:34:13.330
you understand that probably results will not be immediate and that's
528
00:34:13.330 --> 00:34:16.490
hard. Maybe it's worth just going
529
00:34:17.050 --> 00:34:19.920
back to AI space, which is booming today.
530
00:34:21.360 --> 00:34:24.480
But on the other hand,
531
00:34:26.400 --> 00:34:30.240
we understand the roadmap. We understand that when we will hit that
532
00:34:30.720 --> 00:34:33.520
milestone and when quantum computing will actually
533
00:34:35.120 --> 00:34:38.680
start to be much
534
00:34:38.680 --> 00:34:42.280
broaderly adopted, that will
535
00:34:42.280 --> 00:34:45.760
be entirely different game. And once we will
536
00:34:46.960 --> 00:34:49.120
be one of the first companies
537
00:34:51.120 --> 00:34:54.680
in that age at that moment when that
538
00:34:54.680 --> 00:34:57.280
happens, that can bring us to
539
00:35:00.480 --> 00:35:04.240
some absolutely unbelievable, I would say,
540
00:35:04.240 --> 00:35:06.880
results. So yeah, it's very interesting
541
00:35:09.200 --> 00:35:12.810
to kind of understand where it might bring us
542
00:35:13.450 --> 00:35:17.050
and. And when you see that roadmap quite clearly,
543
00:35:17.050 --> 00:35:18.330
that's actually quite exciting.
544
00:35:21.610 --> 00:35:25.210
I usually close my interviews with two questions. I'll just
545
00:35:25.210 --> 00:35:29.010
combine them into one because you've been talking that you've been
546
00:35:29.010 --> 00:35:32.850
saying that you are a company. If quantum
547
00:35:32.850 --> 00:35:36.130
computing really takes off, you are on the verge of
548
00:35:36.130 --> 00:35:39.870
profiting from that. Or are you open to talk
549
00:35:39.870 --> 00:35:43.710
to new investors who would like to join you on this journey and as well
550
00:35:43.710 --> 00:35:44.830
talented employees?
551
00:35:47.630 --> 00:35:51.390
Sure. We are currently on very early stage, so we just closed the
552
00:35:51.950 --> 00:35:55.550
seed round and obviously we will be venture
553
00:35:55.550 --> 00:35:59.310
backed for a while. So we continue talking to investors. We might
554
00:35:59.470 --> 00:36:03.230
get some extensions of rounds. We will have next
555
00:36:03.230 --> 00:36:06.940
rounds. So this is a continuous process and when you are
556
00:36:06.940 --> 00:36:10.580
a startup you need always be raising.
557
00:36:11.460 --> 00:36:15.220
So that's the mode we operate. And my co founder
558
00:36:15.220 --> 00:36:18.780
actually always speaks to investors, also speak to
559
00:36:18.780 --> 00:36:22.500
investors, but probably not as frequently. But
560
00:36:22.740 --> 00:36:26.020
yeah, that's a mode that we are working in and we
561
00:36:26.020 --> 00:36:29.460
certainly are always looking for exciting
562
00:36:30.750 --> 00:36:34.590
and energetic talent and for that we
563
00:36:34.670 --> 00:36:38.430
actually operate very much as like a research
564
00:36:38.430 --> 00:36:41.870
group. So we have many internship
565
00:36:41.870 --> 00:36:45.390
programs, we collaborate with many universities. We try actually
566
00:36:45.390 --> 00:36:48.910
to give young people opportunity to join us
567
00:36:48.910 --> 00:36:52.750
earlier and grow with us because like the stuff that we are
568
00:36:52.750 --> 00:36:56.590
doing is basically at the bleeding edge
569
00:36:57.030 --> 00:36:59.430
of some of these areas in middleware
570
00:37:01.030 --> 00:37:04.550
which are very hard to find anywhere in academia today.
571
00:37:07.430 --> 00:37:11.030
So that's how we recruit new
572
00:37:11.590 --> 00:37:15.270
talented people. But also we are from time to time
573
00:37:15.510 --> 00:37:17.190
offering full time positions
574
00:37:19.350 --> 00:37:23.070
and these are regularly announced. So I
575
00:37:23.070 --> 00:37:24.230
would say both.
576
00:37:27.460 --> 00:37:31.300
Only thing left for me to say. Mikola, thank you very much.
577
00:37:31.460 --> 00:37:34.740
Best of luck and thank you very much for joining us from Western
578
00:37:34.740 --> 00:37:38.460
Ukraine for recording this interview. Thank you,
579
00:37:38.460 --> 00:37:40.660
Joe. It was a pleasure talking to you.
580
00:37:41.860 --> 00:37:45.380
Quantum computing will not be saved by optimism
581
00:37:45.380 --> 00:37:48.620
or funding cycles. It will survive only if
582
00:37:48.620 --> 00:37:51.700
execution becomes honest, repeatable and useful.
583
00:37:52.410 --> 00:37:56.250
Mykola Maximenko is working at that
584
00:37:56.410 --> 00:37:59.850
boundary. Haikyuu is available at Haikyuu
585
00:38:00.090 --> 00:38:03.730
AI. This conversation continues wherever
586
00:38:03.730 --> 00:38:07.130
serious decisions about frontier technologies are
587
00:38:07.450 --> 00:38:08.650
made. Thank you.
588
00:38:13.530 --> 00:38:17.050
That's all folks. Find more news streams,
589
00:38:17.210 --> 00:38:18.380
events and
590
00:38:18.380 --> 00:38:22.740
interviews@www.startuprat.IO.
591
00:38:23.140 --> 00:38:25.140
remember, sharing is car.