Hartmut Neven talks about Google Quantum AI’s breakthrough in quantum error correction

1
00:00:08,000 --> 00:00:10,800
Hello, and welcome to the Physics World Weekly

2
00:00:10,800 --> 00:00:12,740
podcast. I'm Hamish Johnston.

3
00:00:13,294 --> 00:00:15,075
This week, we'll be celebrating

4
00:00:15,615 --> 00:00:17,954
the winners of the Physics World

5
00:00:18,414 --> 00:00:21,234
breakthrough of the year award for 2024.

6
00:00:23,054 --> 00:00:26,274
This episode is supported by the journal Reports

7
00:00:26,414 --> 00:00:28,114
on Progress in Physics,

8
00:00:28,760 --> 00:00:29,500
which offers

9
00:00:30,120 --> 00:00:30,620
unparalleled

10
00:00:31,079 --> 00:00:31,579
visibility

11
00:00:32,039 --> 00:00:33,179
for your groundbreaking

12
00:00:33,799 --> 00:00:34,299
research.

13
00:00:35,320 --> 00:00:38,539
This year's award is all about error correction

14
00:00:38,840 --> 00:00:40,219
in quantum computing,

15
00:00:40,840 --> 00:00:42,859
and we're honoring 2 independent

16
00:00:43,719 --> 00:00:44,219
teams.

17
00:00:44,975 --> 00:00:48,094
I've spoken to the lead researchers of both

18
00:00:48,094 --> 00:00:48,594
groups,

19
00:00:48,975 --> 00:00:50,914
and we're presenting those conversations

20
00:00:51,695 --> 00:00:53,635
in 2 different episodes

21
00:00:54,094 --> 00:00:54,995
of the podcast.

22
00:00:55,454 --> 00:00:57,780
So it's a real bonus this week, not

23
00:00:57,780 --> 00:01:01,539
one, but 2 weekly podcasts for your listening

24
00:01:01,539 --> 00:01:02,039
pleasure.

25
00:01:03,299 --> 00:01:05,479
In this podcast, I'm in conversation

26
00:01:05,859 --> 00:01:07,959
with Google's Hartmut Nevin,

27
00:01:08,500 --> 00:01:10,954
who led a team that has made a

28
00:01:10,954 --> 00:01:12,015
major breakthrough

29
00:01:12,395 --> 00:01:13,295
in implementing

30
00:01:13,674 --> 00:01:15,215
quantum error correction

31
00:01:15,674 --> 00:01:16,575
in a processor

32
00:01:17,034 --> 00:01:17,855
that uses

33
00:01:18,234 --> 00:01:18,734
superconducting

34
00:01:19,515 --> 00:01:20,015
qubits.

35
00:01:20,875 --> 00:01:23,829
In a separate episode, I chat with Mikhail

36
00:01:23,829 --> 00:01:24,329
Lukin

37
00:01:24,630 --> 00:01:25,609
and Dolev

38
00:01:25,909 --> 00:01:26,409
Blufstein

39
00:01:27,030 --> 00:01:28,409
at Harvard University,

40
00:01:29,030 --> 00:01:31,530
who, along with colleagues, have implemented

41
00:01:32,150 --> 00:01:33,609
quantum error correction

42
00:01:33,989 --> 00:01:36,090
on an array of trapped

43
00:01:36,549 --> 00:01:37,849
atomic cubits.

44
00:01:39,024 --> 00:01:42,405
In principle, quantum computers can solve some problems

45
00:01:42,704 --> 00:01:44,084
that cannot be computed

46
00:01:44,465 --> 00:01:45,284
on conventional

47
00:01:45,584 --> 00:01:46,084
processors.

48
00:01:47,024 --> 00:01:47,524
However,

49
00:01:47,905 --> 00:01:50,484
the quantum processors available today

50
00:01:50,864 --> 00:01:52,084
are very susceptible

51
00:01:52,469 --> 00:01:54,489
to disruption by environmental

52
00:01:54,790 --> 00:01:55,290
noise,

53
00:01:55,750 --> 00:01:56,730
and this destroys

54
00:01:57,189 --> 00:02:00,310
the delicate quantum states that are used to

55
00:02:00,310 --> 00:02:02,569
store and process information.

56
00:02:03,510 --> 00:02:05,930
When quantum computing was first proposed,

57
00:02:06,310 --> 00:02:07,130
some physicists

58
00:02:07,674 --> 00:02:09,775
thought that this problem was insurmountable.

59
00:02:10,715 --> 00:02:13,675
But thanks to the development of quantum error

60
00:02:13,675 --> 00:02:14,175
correction,

61
00:02:14,715 --> 00:02:15,215
practical

62
00:02:15,754 --> 00:02:16,895
quantum computers

63
00:02:17,275 --> 00:02:19,294
that can solve useful problems

64
00:02:19,800 --> 00:02:21,419
could soon be a reality.

65
00:02:22,759 --> 00:02:25,180
Quantum error correction works by distributing

66
00:02:25,719 --> 00:02:27,900
1 quantum bit of information,

67
00:02:28,599 --> 00:02:30,219
called a logical cubit,

68
00:02:30,680 --> 00:02:33,580
across several different physical cubits,

69
00:02:34,025 --> 00:02:34,925
such as superconducting

70
00:02:35,385 --> 00:02:35,885
circuits.

71
00:02:36,585 --> 00:02:37,325
In principle,

72
00:02:37,784 --> 00:02:41,145
the robustness of a logical cubit should be

73
00:02:41,145 --> 00:02:41,645
improved

74
00:02:41,944 --> 00:02:42,925
by increasing

75
00:02:43,305 --> 00:02:45,325
the number of physical cubits.

76
00:02:46,025 --> 00:02:47,509
But there's a problem.

77
00:02:48,209 --> 00:02:50,629
Boosting the number of physical cubits

78
00:02:51,009 --> 00:02:51,509
itself

79
00:02:51,969 --> 00:02:52,469
introduces

80
00:02:52,770 --> 00:02:53,270
errors,

81
00:02:53,729 --> 00:02:54,550
and therefore,

82
00:02:54,930 --> 00:02:56,229
creating an optimal

83
00:02:56,610 --> 00:02:58,870
quantum error correction system

84
00:02:59,330 --> 00:03:01,030
is no easy task.

85
00:03:01,754 --> 00:03:04,634
This year, we've awarded one half of the

86
00:03:04,634 --> 00:03:08,074
20 24 breakthrough of the year award to

87
00:03:08,074 --> 00:03:08,574
Hartmut

88
00:03:08,955 --> 00:03:09,455
Niven

89
00:03:09,754 --> 00:03:10,574
and colleagues

90
00:03:10,875 --> 00:03:13,055
at Google Quantum AI

91
00:03:13,770 --> 00:03:14,509
and their collaborators,

92
00:03:15,930 --> 00:03:17,229
and that's for implementing

93
00:03:17,770 --> 00:03:19,229
quantum error correction

94
00:03:19,689 --> 00:03:20,189
below

95
00:03:20,490 --> 00:03:22,669
the surface code threshold

96
00:03:23,129 --> 00:03:24,110
in a superconducting

97
00:03:24,569 --> 00:03:25,069
chip.

98
00:03:26,025 --> 00:03:27,164
For the first time,

99
00:03:27,625 --> 00:03:28,125
exponential

100
00:03:28,664 --> 00:03:29,564
error suppression

101
00:03:30,185 --> 00:03:32,764
in a logical cubit has been achieved

102
00:03:33,305 --> 00:03:35,884
as the number of physical cubits

103
00:03:36,185 --> 00:03:36,685
increases.

104
00:03:37,625 --> 00:03:39,405
And to chat about this achievement,

105
00:03:39,969 --> 00:03:40,469
Hartmut

106
00:03:40,770 --> 00:03:43,110
joins me down the line from California.

107
00:03:44,050 --> 00:03:44,550
Hello.

108
00:03:44,930 --> 00:03:46,229
Welcome to the podcast,

109
00:03:46,610 --> 00:03:47,750
and congratulations

110
00:03:48,370 --> 00:03:49,750
on your team's achievement.

111
00:03:50,849 --> 00:03:52,229
Thank you for having me.

112
00:03:53,385 --> 00:03:56,844
So Hartmut, I suppose first things first,

113
00:03:57,784 --> 00:03:59,965
with with my questions. What is

114
00:04:00,344 --> 00:04:01,885
quantum error correction,

115
00:04:02,425 --> 00:04:04,284
and why is it needed?

116
00:04:05,919 --> 00:04:07,459
A quantum error correction

117
00:04:07,759 --> 00:04:08,979
is a necessary

118
00:04:09,439 --> 00:04:09,939
technology

119
00:04:10,879 --> 00:04:13,939
that allows you to scale up to large

120
00:04:14,239 --> 00:04:16,819
quantum computers with many cubits

121
00:04:17,199 --> 00:04:20,144
that can participate in many algorithmic steps.

122
00:04:21,584 --> 00:04:22,884
And and the reason

123
00:04:23,264 --> 00:04:24,805
is because each

124
00:04:25,345 --> 00:04:28,144
individual qubit in a quantum computer, at least

125
00:04:28,144 --> 00:04:31,125
a a quantum computer that exists today,

126
00:04:31,664 --> 00:04:35,125
is is very noisy or subject to failure.

127
00:04:35,729 --> 00:04:37,410
So you you sort of have to club

128
00:04:37,410 --> 00:04:41,269
them together to to get one good cubit,

129
00:04:41,649 --> 00:04:43,349
cubit. Is that how it works?

130
00:04:43,810 --> 00:04:45,569
Yes. The way how I like to think

131
00:04:45,569 --> 00:04:49,169
about it, people often say quantum information is

132
00:04:49,169 --> 00:04:50,069
very fragile.

133
00:04:50,435 --> 00:04:51,895
We need to protect it.

134
00:04:52,835 --> 00:04:54,355
I like to think about it a little

135
00:04:54,355 --> 00:04:55,254
bit differently.

136
00:04:55,714 --> 00:04:57,654
Quantum information is very contagious.

137
00:04:58,514 --> 00:05:00,514
Qubits like to talk to each other, and

138
00:05:00,514 --> 00:05:03,095
they also like to talk to Qubits

139
00:05:03,395 --> 00:05:06,410
outside of our chip, outside of our control,

140
00:05:06,870 --> 00:05:10,009
and that leads to leaking information,

141
00:05:10,789 --> 00:05:13,829
leaking out of the processor, and we need

142
00:05:13,829 --> 00:05:15,209
to prevent it. So

143
00:05:15,669 --> 00:05:18,865
quantum error correction is really a set of

144
00:05:18,865 --> 00:05:20,404
technologies or a technology

145
00:05:20,944 --> 00:05:22,805
that allows us to

146
00:05:24,064 --> 00:05:25,764
control all the information

147
00:05:26,225 --> 00:05:28,084
necessary for a quantum computation.

148
00:05:29,104 --> 00:05:31,680
I see. Okay. And and in the work

149
00:05:31,680 --> 00:05:34,019
that you've done with the Willow processor,

150
00:05:34,560 --> 00:05:36,720
how how have you done that? How have

151
00:05:36,720 --> 00:05:39,519
you dealt with this leakage of information? How

152
00:05:39,519 --> 00:05:40,800
have you made sure that,

153
00:05:41,279 --> 00:05:43,759
the information is where you want it to

154
00:05:43,759 --> 00:05:46,305
be when you do your quantum calculation.

155
00:05:47,324 --> 00:05:47,824
So,

156
00:05:48,845 --> 00:05:49,985
quantum error correction,

157
00:05:50,365 --> 00:05:52,145
like classical error correction,

158
00:05:52,764 --> 00:05:54,064
draws on the

159
00:05:54,444 --> 00:05:57,024
time tested principle in engineering.

160
00:05:57,660 --> 00:05:59,579
That if you want to make something more

161
00:05:59,579 --> 00:06:01,680
stable, you introduce redundancy.

162
00:06:02,539 --> 00:06:05,660
Understand this is, a physics world audience, so

163
00:06:05,660 --> 00:06:08,220
maybe this example is too simple. But,

164
00:06:08,860 --> 00:06:10,879
often say, hey. If you want to

165
00:06:11,564 --> 00:06:14,625
fly, let's say, from Germany here to LA,

166
00:06:15,164 --> 00:06:17,485
if you have an airplane with 1 engine,

167
00:06:17,485 --> 00:06:21,164
that will work. But 2 engines is safer.

168
00:06:21,164 --> 00:06:24,044
And if you have 4 engines, it's yet

169
00:06:24,044 --> 00:06:26,285
better because if one of them fails, you

170
00:06:26,285 --> 00:06:26,785
still

171
00:06:27,250 --> 00:06:28,870
easily make it over.

172
00:06:29,250 --> 00:06:30,629
We use the same principle

173
00:06:31,170 --> 00:06:34,449
in quantum error correction. So we want to

174
00:06:34,449 --> 00:06:38,389
represent 1 logical qubit or the information contained

175
00:06:38,449 --> 00:06:39,990
in 1 logical qubit.

176
00:06:40,290 --> 00:06:43,334
So how we do this is we orchestrate

177
00:06:44,435 --> 00:06:46,995
a set of physical cubits, a little array

178
00:06:46,995 --> 00:06:49,074
of, let's say, 3 by 3 or 5

179
00:06:49,074 --> 00:06:50,855
by 5 or 7 by 7

180
00:06:51,475 --> 00:06:52,455
physical cubits

181
00:06:52,834 --> 00:06:55,334
that make one better protected

182
00:06:55,795 --> 00:06:56,790
logical cubit.

183
00:06:57,589 --> 00:06:59,990
To our delight, what we were able, to

184
00:06:59,990 --> 00:07:00,490
demonstrate

185
00:07:01,189 --> 00:07:04,170
is that as we went to larger arrays

186
00:07:04,230 --> 00:07:05,449
of physical qubits,

187
00:07:05,830 --> 00:07:09,110
the error rate was reduced. So what we

188
00:07:09,110 --> 00:07:11,444
were able to do as we went from

189
00:07:11,444 --> 00:07:13,865
codistance 3 to 5 to 7,

190
00:07:14,245 --> 00:07:16,345
each time as we increase the codistance,

191
00:07:16,884 --> 00:07:19,524
the error rates were reduced by a factor

192
00:07:19,524 --> 00:07:20,185
of 2,

193
00:07:20,564 --> 00:07:22,904
effectively leading to an exponential

194
00:07:23,444 --> 00:07:25,144
reduction in error rate

195
00:07:25,529 --> 00:07:27,149
and therefore creating,

196
00:07:28,330 --> 00:07:32,189
the most convincing prototype of a logical qubit

197
00:07:32,410 --> 00:07:33,230
till to date.

198
00:07:33,930 --> 00:07:36,649
I see. And is that something that surprised

199
00:07:36,649 --> 00:07:37,149
you

200
00:07:37,634 --> 00:07:39,714
when when you set out to do this

201
00:07:39,714 --> 00:07:42,294
research? Were were you expecting to see that

202
00:07:42,354 --> 00:07:43,254
that exponential

203
00:07:43,875 --> 00:07:44,375
effect?

204
00:07:45,235 --> 00:07:48,454
No. Theory had predicted this. It wasn't,

205
00:07:48,834 --> 00:07:50,490
reduced to practice yet.

206
00:07:51,449 --> 00:07:52,649
So we could actually,

207
00:07:53,129 --> 00:07:55,069
predict this, quite well.

208
00:07:55,610 --> 00:07:57,610
What is important if you want to achieve

209
00:07:57,610 --> 00:07:58,430
such a result

210
00:07:58,810 --> 00:07:59,310
is

211
00:08:00,410 --> 00:08:03,129
your it's a system engineering challenge. So it's

212
00:08:03,129 --> 00:08:05,629
not good enough if just your

213
00:08:05,954 --> 00:08:08,274
single qubit gates are very good or just

214
00:08:08,274 --> 00:08:10,995
your 2 qubit gates are very good. Your

215
00:08:10,995 --> 00:08:11,814
state preparation,

216
00:08:12,194 --> 00:08:15,254
your readout, all components have to be,

217
00:08:15,954 --> 00:08:16,694
very well,

218
00:08:17,235 --> 00:08:20,060
engineered and have to be what is called

219
00:08:20,120 --> 00:08:22,939
below threshold. See, this only works,

220
00:08:25,159 --> 00:08:27,339
or if you want to have more cubits

221
00:08:28,439 --> 00:08:29,659
but less error.

222
00:08:30,039 --> 00:08:33,274
This only works if your cubits have achieved

223
00:08:33,575 --> 00:08:36,455
a certain basic quality, and that is known

224
00:08:36,455 --> 00:08:37,274
as the field

225
00:08:37,735 --> 00:08:41,034
as being below threshold. So if all components

226
00:08:41,975 --> 00:08:44,315
of the system are reasonably good,

227
00:08:44,789 --> 00:08:47,529
then you can orchestrate them into something

228
00:08:47,990 --> 00:08:51,029
really very good. That is essentially how quantum

229
00:08:51,029 --> 00:08:52,169
error correction works.

230
00:08:52,709 --> 00:08:54,950
I see. And and and so the system

231
00:08:54,950 --> 00:08:57,370
that you have available at the moment,

232
00:08:58,894 --> 00:09:00,495
would you describe it as a sort of

233
00:09:00,495 --> 00:09:01,554
a proof of principle

234
00:09:02,175 --> 00:09:05,215
system? Or are you able to to actually

235
00:09:05,215 --> 00:09:06,434
use it to solve

236
00:09:06,975 --> 00:09:07,475
practical,

237
00:09:08,654 --> 00:09:11,774
computing problems, and and even problems that can't

238
00:09:11,774 --> 00:09:12,514
be solved

239
00:09:13,059 --> 00:09:16,120
easily by a a conventional classical computer.

240
00:09:17,139 --> 00:09:17,639
So

241
00:09:18,659 --> 00:09:19,159
the

242
00:09:19,940 --> 00:09:21,959
quantum error correction demonstration

243
00:09:22,339 --> 00:09:25,319
made just a single good logical qubit.

244
00:09:25,725 --> 00:09:28,125
Of course, a single logical qubit is not

245
00:09:28,125 --> 00:09:31,644
good enough to run any interesting algorithm. You

246
00:09:31,644 --> 00:09:33,504
will need, many of them.

247
00:09:33,884 --> 00:09:36,284
And also, you need it yet better. We

248
00:09:36,284 --> 00:09:39,325
achieved roughly a 1 in a 1000 error

249
00:09:39,325 --> 00:09:39,659
rate,

250
00:09:40,139 --> 00:09:42,720
which means that then you can

251
00:09:43,179 --> 00:09:45,519
run about a 1,000 operations

252
00:09:46,139 --> 00:09:47,279
in your algorithm.

253
00:09:47,819 --> 00:09:49,740
Because the way you can think about error

254
00:09:49,740 --> 00:09:51,259
rates, if it's, let's say, 1 in a

255
00:09:51,259 --> 00:09:53,339
1000 or 1 in a 1000000, what it

256
00:09:53,339 --> 00:09:56,085
means, you have your cubits, you apply your,

257
00:09:56,485 --> 00:09:56,985
gates,

258
00:09:57,365 --> 00:09:59,605
and then you have the 1 in a

259
00:09:59,605 --> 00:10:01,065
1000 or 1 in a 1000000

260
00:10:01,445 --> 00:10:02,904
chance that you

261
00:10:03,524 --> 00:10:05,865
crash your machine. You get a blue screen

262
00:10:06,004 --> 00:10:08,345
and you have to restart your computation.

263
00:10:08,940 --> 00:10:11,120
So 1 in a 1000 error rate means

264
00:10:11,420 --> 00:10:14,620
we can really run algorithms with about a

265
00:10:14,620 --> 00:10:17,120
1,000 gates. So this gives you a certain

266
00:10:17,580 --> 00:10:18,720
limit of complexity.

267
00:10:20,139 --> 00:10:21,759
Many of the famous algorithms,

268
00:10:22,460 --> 00:10:22,960
for,

269
00:10:23,855 --> 00:10:24,754
quantum simulations

270
00:10:25,214 --> 00:10:26,914
or factoring large numbers,

271
00:10:27,294 --> 00:10:29,214
they often need way more,

272
00:10:29,695 --> 00:10:32,495
gates than a 1,000 and need 1,000,000 or

273
00:10:32,495 --> 00:10:33,315
even 1,000,000,000.

274
00:10:33,934 --> 00:10:36,014
So, therefore, we have to,

275
00:10:36,495 --> 00:10:36,995
make

276
00:10:37,934 --> 00:10:41,399
yet lower error rate logical cubits. And we

277
00:10:41,399 --> 00:10:44,040
can, of course, do this, rather easily in

278
00:10:44,040 --> 00:10:46,060
a way because once you have an exponential

279
00:10:46,200 --> 00:10:46,700
reduction,

280
00:10:47,080 --> 00:10:48,940
you just ask for this algorithm

281
00:10:49,480 --> 00:10:51,580
what error rate do I need. And then

282
00:10:51,720 --> 00:10:54,695
you go to the appropriate code distance, which,

283
00:10:55,154 --> 00:10:57,634
let's say if you want something really low,

284
00:10:57,634 --> 00:11:00,434
like 1 in 10,000,000,000 error rate, that's something

285
00:11:00,434 --> 00:11:01,894
we want to eventually achieve.

286
00:11:03,075 --> 00:11:05,235
Then you go to a code distance of

287
00:11:05,235 --> 00:11:07,794
21 or 23, or if you improve the

288
00:11:07,794 --> 00:11:08,294
overall

289
00:11:09,059 --> 00:11:11,220
hardware, maybe we can do it with 19

290
00:11:11,220 --> 00:11:14,279
or 17. Meaning, then you have a 19

291
00:11:14,340 --> 00:11:15,160
by 19

292
00:11:15,860 --> 00:11:18,980
array of physical cubits that make one very

293
00:11:18,980 --> 00:11:20,279
good, very low

294
00:11:20,660 --> 00:11:22,544
logical error rate, cubit.

295
00:11:23,024 --> 00:11:24,704
And then, of course and you don't need

296
00:11:24,704 --> 00:11:27,204
just one logical cubit, but you will need,

297
00:11:28,544 --> 00:11:31,284
let's say, a1000 of them. And that is

298
00:11:31,424 --> 00:11:32,325
since the endpoint

299
00:11:33,184 --> 00:11:35,924
of our development roadmap. Our roadmap

300
00:11:36,384 --> 00:11:37,024
calls for

301
00:11:37,870 --> 00:11:40,449
it consists of 6 salient milestones.

302
00:11:41,069 --> 00:11:43,389
And the last milestone, at least for now

303
00:11:43,389 --> 00:11:44,209
on this roadmap,

304
00:11:45,149 --> 00:11:47,409
calls for building a 1,000,000

305
00:11:47,870 --> 00:11:48,909
physical qubit,

306
00:11:50,269 --> 00:11:51,250
quantum processor,

307
00:11:51,825 --> 00:11:53,845
which would, with today's methods,

308
00:11:55,504 --> 00:11:59,345
translate into about a 1,000 highly protected logical

309
00:11:59,345 --> 00:11:59,845
qubits.

310
00:12:00,304 --> 00:12:02,804
Now you have a system that can confidently

311
00:12:03,665 --> 00:12:05,159
generate commercial value.

312
00:12:06,039 --> 00:12:07,720
I see. And and can you talk a

313
00:12:07,720 --> 00:12:10,539
bit about the the the physical hardware?

314
00:12:11,559 --> 00:12:13,639
The am I right in thinking that you're

315
00:12:13,639 --> 00:12:14,459
you're using,

316
00:12:15,240 --> 00:12:15,740
superconducting

317
00:12:16,120 --> 00:12:17,579
qubits. Is that right?

318
00:12:18,535 --> 00:12:20,934
Yes. That's that's correct. There's a different ways

319
00:12:20,934 --> 00:12:22,154
how you can represent

320
00:12:23,254 --> 00:12:26,075
a quantum bit, a cubit. Essentially, any

321
00:12:26,535 --> 00:12:29,195
quantum mechanical two system state

322
00:12:30,375 --> 00:12:31,169
can be used.

323
00:12:32,049 --> 00:12:33,029
We use superconducting

324
00:12:33,490 --> 00:12:33,990
cubits.

325
00:12:34,529 --> 00:12:35,029
So,

326
00:12:35,809 --> 00:12:37,970
anybody here on the audience who has let's

327
00:12:37,970 --> 00:12:39,970
say when they were teenager built a little

328
00:12:39,970 --> 00:12:42,690
radio out of an electronics kit, they may

329
00:12:42,690 --> 00:12:44,549
be familiar with LC

330
00:12:45,009 --> 00:12:45,509
circuits

331
00:12:46,414 --> 00:12:49,134
or electrical oscillator where you can sync the

332
00:12:49,134 --> 00:12:49,794
the electrons

333
00:12:50,174 --> 00:12:51,955
slash back and forth between

334
00:12:52,575 --> 00:12:54,355
a capacitor and an inductor.

335
00:12:55,214 --> 00:12:57,875
Our cubits are basically this LC,

336
00:12:58,254 --> 00:12:58,754
circuits,

337
00:12:59,929 --> 00:13:01,309
but they are superconducting,

338
00:13:03,129 --> 00:13:06,009
and live at very low temperature below the

339
00:13:06,009 --> 00:13:07,950
transition temperature of the superconductor.

340
00:13:08,730 --> 00:13:09,470
So they

341
00:13:09,850 --> 00:13:11,070
essentially implement

342
00:13:12,065 --> 00:13:13,205
a quantum mechanical

343
00:13:13,664 --> 00:13:14,725
harmonic oscillator.

344
00:13:15,504 --> 00:13:16,004
And

345
00:13:16,865 --> 00:13:19,365
people may remember from a 1st semester

346
00:13:19,904 --> 00:13:20,804
quantum physics,

347
00:13:21,504 --> 00:13:22,725
class that

348
00:13:23,024 --> 00:13:24,965
in a quantum mechanical oscillator,

349
00:13:25,345 --> 00:13:26,004
you have

350
00:13:26,600 --> 00:13:27,100
discretized

351
00:13:27,879 --> 00:13:29,179
energy levels.

352
00:13:29,959 --> 00:13:33,000
And we use the lowest energy level and

353
00:13:33,000 --> 00:13:36,139
the first excited energy level as our logical

354
00:13:36,279 --> 00:13:37,899
zero and logical one.

355
00:13:38,804 --> 00:13:41,365
I see. And and how how do you

356
00:13:41,365 --> 00:13:42,825
connect up these,

357
00:13:43,764 --> 00:13:44,665
these superconducting

358
00:13:45,045 --> 00:13:48,024
oscillators? How exactly do you get communication

359
00:13:48,404 --> 00:13:49,144
and coordination

360
00:13:49,764 --> 00:13:51,625
so that you can do your error correction

361
00:13:51,684 --> 00:13:52,425
and, ultimately,

362
00:13:53,350 --> 00:13:53,850
computations.

363
00:13:54,629 --> 00:13:56,710
So so there are different ways. Once you

364
00:13:56,710 --> 00:13:57,350
have your,

365
00:14:00,230 --> 00:14:01,370
I'll see oscillators,

366
00:14:01,830 --> 00:14:03,290
your, your cubits.

367
00:14:03,910 --> 00:14:05,509
Maybe one piece I didn't,

368
00:14:06,230 --> 00:14:06,629
say,

369
00:14:07,590 --> 00:14:08,945
that's still important.

370
00:14:09,404 --> 00:14:10,705
I told you we essentially,

371
00:14:12,605 --> 00:14:13,105
implement

372
00:14:14,045 --> 00:14:16,545
cubits as quantum mechanical

373
00:14:17,165 --> 00:14:17,665
oscillators.

374
00:14:19,404 --> 00:14:20,845
And I told you that we use the

375
00:14:20,845 --> 00:14:22,445
lowest energy level and see,

376
00:14:23,179 --> 00:14:25,500
the next highest energy level as our 0

377
00:14:25,500 --> 00:14:28,319
and 1. We do one more piece because

378
00:14:28,379 --> 00:14:29,679
in a harmonic,

379
00:14:30,940 --> 00:14:33,980
quantum mechanical oscillator, then there will also be

380
00:14:33,980 --> 00:14:36,299
a level 2, an energy level 3, and

381
00:14:36,299 --> 00:14:38,794
so on. And they all have exactly,

382
00:14:39,814 --> 00:14:41,194
the same energy

383
00:14:41,495 --> 00:14:41,975
distance,

384
00:14:42,454 --> 00:14:42,954
hbar,

385
00:14:43,495 --> 00:14:44,634
times omega.

386
00:14:45,095 --> 00:14:47,495
So that is not so good because, if

387
00:14:47,495 --> 00:14:49,174
I have my qubit, let's say, in the

388
00:14:49,174 --> 00:14:51,095
first excited state and I send it a

389
00:14:51,095 --> 00:14:54,559
pile with with just this energy difference, and

390
00:14:54,559 --> 00:14:56,320
I can exactly make sure it goes back

391
00:14:56,320 --> 00:14:58,720
to my 0. It might also go up

392
00:14:58,720 --> 00:14:59,040
to,

393
00:14:59,840 --> 00:15:02,480
level 2, and we don't want this because

394
00:15:02,480 --> 00:15:05,414
this is outside of the code space. So

395
00:15:05,414 --> 00:15:07,495
we put one more ingredient in, and that

396
00:15:07,495 --> 00:15:10,394
is the Josephson junction. And the Josephson junction

397
00:15:10,695 --> 00:15:12,794
is a nonlinear circuit element,

398
00:15:13,095 --> 00:15:15,274
and it makes our oscillator

399
00:15:15,654 --> 00:15:17,035
slightly un harmonic.

400
00:15:17,389 --> 00:15:20,190
What this means is now the energy levels

401
00:15:20,190 --> 00:15:21,169
are not equidistant

402
00:15:21,470 --> 00:15:21,970
anymore,

403
00:15:22,350 --> 00:15:23,169
but the

404
00:15:24,190 --> 00:15:27,230
first separation between 0 and 1 is a

405
00:15:27,230 --> 00:15:30,929
little bit larger energy difference than between 12,

406
00:15:31,470 --> 00:15:34,254
and that is still larger than 23. So

407
00:15:34,254 --> 00:15:36,595
you get this letter of shrinking,

408
00:15:38,414 --> 00:15:39,954
distance energy levels.

409
00:15:40,254 --> 00:15:42,834
And that's quite useful because now we can

410
00:15:43,054 --> 00:15:43,794
send in

411
00:15:44,254 --> 00:15:46,514
control palaces that will only

412
00:15:46,815 --> 00:15:47,634
cause transitions

413
00:15:48,240 --> 00:15:49,379
between the lowest,

414
00:15:49,759 --> 00:15:52,240
two energy levels and the others out of

415
00:15:52,240 --> 00:15:55,600
code space are not involved. So that really

416
00:15:55,600 --> 00:15:56,100
completes,

417
00:15:56,799 --> 00:15:57,539
the superconducting

418
00:15:57,919 --> 00:15:58,419
qubit.

419
00:15:59,120 --> 00:16:01,679
Now once I have a superconducting qubit, how

420
00:16:01,679 --> 00:16:03,894
can I couple them? There there are various

421
00:16:03,894 --> 00:16:06,375
ways how you can, couple them. There are

422
00:16:06,375 --> 00:16:06,875
capacitive

423
00:16:07,254 --> 00:16:09,995
couplings, inductive couplings. You can use

424
00:16:10,535 --> 00:16:11,035
little,

425
00:16:11,495 --> 00:16:11,995
qubits,

426
00:16:12,535 --> 00:16:15,175
in between that act as a coupler. You

427
00:16:15,254 --> 00:16:17,735
they just have to get into interaction. So

428
00:16:17,735 --> 00:16:20,600
you you can maybe think of the mechanical

429
00:16:20,820 --> 00:16:22,279
analog, you know, pendulums,

430
00:16:22,580 --> 00:16:25,240
and, you put a little spring between them,

431
00:16:25,379 --> 00:16:27,059
and then they feel each other. That is

432
00:16:27,059 --> 00:16:27,559
an

433
00:16:27,940 --> 00:16:29,240
interaction between,

434
00:16:30,019 --> 00:16:32,919
2 oscillators. So there are multiple design choices

435
00:16:32,980 --> 00:16:34,815
you have. And what's the exact,

436
00:16:35,434 --> 00:16:37,834
best choice for coupling is that is actually

437
00:16:37,834 --> 00:16:39,855
still a bit of matter of research.

438
00:16:40,714 --> 00:16:42,414
I see. And and

439
00:16:42,955 --> 00:16:44,095
you mentioned that,

440
00:16:44,634 --> 00:16:46,714
I suppose, your ultimate goal is to get

441
00:16:46,714 --> 00:16:48,634
a a quantum processor with

442
00:16:49,250 --> 00:16:51,110
that offers about a 1,000,

443
00:16:52,370 --> 00:16:52,870
logical

444
00:16:53,329 --> 00:16:55,970
cubits. And so so that would require many,

445
00:16:55,970 --> 00:16:58,070
many more actual physical

446
00:16:58,850 --> 00:16:59,350
superconducting

447
00:17:00,289 --> 00:17:02,209
cubits. Is is that the sort of thing

448
00:17:02,209 --> 00:17:03,750
that that can be miniaturized

449
00:17:04,210 --> 00:17:04,710
onto

450
00:17:05,365 --> 00:17:07,525
onto a chip? Or or would that have

451
00:17:07,525 --> 00:17:09,684
to take up a, let's say, an entire

452
00:17:09,684 --> 00:17:10,184
lab,

453
00:17:11,284 --> 00:17:12,265
at a university?

454
00:17:13,284 --> 00:17:13,765
So,

455
00:17:14,085 --> 00:17:15,065
roughly speaking,

456
00:17:17,044 --> 00:17:19,365
our our cubits are not super small. Actually,

457
00:17:19,365 --> 00:17:20,789
if I were to give you a chip,

458
00:17:21,830 --> 00:17:24,009
which has an array of cubits and you

459
00:17:24,230 --> 00:17:26,150
squint, you can actually see,

460
00:17:26,710 --> 00:17:29,269
the individual cubits. They are a little bit

461
00:17:29,269 --> 00:17:30,730
smaller than a square

462
00:17:31,109 --> 00:17:31,609
millimeter.

463
00:17:32,390 --> 00:17:33,529
So, essentially,

464
00:17:34,714 --> 00:17:37,835
out of these components, we would make larger

465
00:17:37,835 --> 00:17:39,214
and larger arrays.

466
00:17:40,315 --> 00:17:41,375
But currently,

467
00:17:41,914 --> 00:17:43,775
a little bit as opposed to

468
00:17:44,714 --> 00:17:45,214
classical,

469
00:17:45,674 --> 00:17:46,174
CMOS,

470
00:17:46,714 --> 00:17:47,214
technologies,

471
00:17:47,910 --> 00:17:50,869
we are not aiming or it's not a

472
00:17:50,869 --> 00:17:52,789
priority for us to make the,

473
00:17:53,269 --> 00:17:54,890
cubits smaller and smaller

474
00:17:55,190 --> 00:17:55,690
because,

475
00:17:56,150 --> 00:17:58,330
we actually like that they have the size

476
00:17:58,390 --> 00:18:02,105
because then it's easier to control them with,

477
00:18:02,805 --> 00:18:05,365
microwave pulses, so it's easier to read them

478
00:18:05,365 --> 00:18:07,924
out. So we would just have to make,

479
00:18:09,365 --> 00:18:12,985
sufficiently large chips. Eventually, you have to, join

480
00:18:13,205 --> 00:18:14,039
multiple chips

481
00:18:14,519 --> 00:18:17,019
to make a chip that has the surface

482
00:18:17,080 --> 00:18:18,220
area to,

483
00:18:18,600 --> 00:18:20,759
cover, let's say, a 1,000 or even a

484
00:18:20,759 --> 00:18:21,240
1000000,

485
00:18:21,640 --> 00:18:22,700
physical cubits.

486
00:18:23,880 --> 00:18:25,660
I see. And and so

487
00:18:25,960 --> 00:18:27,820
where are you in in the development,

488
00:18:28,805 --> 00:18:31,045
of that at the moment? Have you have

489
00:18:31,045 --> 00:18:33,224
you managed to to create,

490
00:18:34,244 --> 00:18:37,845
a system of integrated physical cubits that's large

491
00:18:37,845 --> 00:18:38,345
enough

492
00:18:38,644 --> 00:18:40,184
to give you a 1,000

493
00:18:40,724 --> 00:18:43,240
logical cubits? Or is that something that's down

494
00:18:43,240 --> 00:18:44,539
the road a bit further?

495
00:18:45,480 --> 00:18:49,079
So I mentioned the road map that, the

496
00:18:49,079 --> 00:18:51,259
Google Quantum AI team published.

497
00:18:52,119 --> 00:18:55,019
And this road map consists of 6 milestones.

498
00:18:55,545 --> 00:18:56,924
So the first milestone

499
00:18:57,465 --> 00:19:00,525
was we achieved it, in 2019.

500
00:19:01,384 --> 00:19:03,545
It was showing for the very first time

501
00:19:03,545 --> 00:19:04,285
that a quantum

502
00:19:05,144 --> 00:19:05,644
processor

503
00:19:06,184 --> 00:19:06,684
could

504
00:19:07,305 --> 00:19:10,490
compute a task in minutes. That's a zen

505
00:19:10,490 --> 00:19:14,009
fastest supercomputer would have needed 10000 years to

506
00:19:14,009 --> 00:19:14,509
do.

507
00:19:15,450 --> 00:19:18,509
The second milestone we also achieved already was

508
00:19:19,369 --> 00:19:20,829
similar to the current experiment,

509
00:19:21,369 --> 00:19:23,369
was to show that as we go from

510
00:19:23,369 --> 00:19:26,044
code distance 3 to 5, the error rate

511
00:19:26,184 --> 00:19:27,004
comes down.

512
00:19:27,384 --> 00:19:28,845
But that milestone was

513
00:19:30,105 --> 00:19:32,585
defined as just being the break even point.

514
00:19:32,585 --> 00:19:35,244
So, yes, the error rate did come down,

515
00:19:35,304 --> 00:19:36,825
but just by a hair,

516
00:19:37,224 --> 00:19:40,200
4%. So that was not that impressive yet.

517
00:19:40,440 --> 00:19:41,639
So the the current,

518
00:19:42,200 --> 00:19:44,039
result, which we refer to as a yard

519
00:19:44,039 --> 00:19:46,859
stone, is actually between 2 big milestone

520
00:19:47,319 --> 00:19:47,819
that

521
00:19:48,359 --> 00:19:48,859
improved,

522
00:19:49,480 --> 00:19:53,079
on milestone 2 by it's not 4% anymore.

523
00:19:53,079 --> 00:19:54,894
It's by a factor of 2 that the

524
00:19:54,894 --> 00:19:56,815
error rate came down and it came down

525
00:19:56,815 --> 00:19:59,695
twice from, again, code is 3 to 5

526
00:19:59,695 --> 00:20:00,355
to 7.

527
00:20:00,734 --> 00:20:02,815
So milestone 3, which is sort of the

528
00:20:02,815 --> 00:20:06,174
midpoint of our road map, will be a

529
00:20:06,174 --> 00:20:06,674
single

530
00:20:06,975 --> 00:20:08,674
very good logical qubit

531
00:20:09,009 --> 00:20:11,650
with a 1 in a 1000000 logical error

532
00:20:11,650 --> 00:20:12,150
rate.

533
00:20:12,690 --> 00:20:13,970
And then from there, we,

534
00:20:15,169 --> 00:20:17,750
now maybe I mentioned the remaining milestones.

535
00:20:18,210 --> 00:20:20,470
Milestone 4 is then having several,

536
00:20:20,849 --> 00:20:23,669
logical cubits of about that quality

537
00:20:24,244 --> 00:20:26,825
and to have a gate set, a universal

538
00:20:27,205 --> 00:20:28,184
set of gates,

539
00:20:28,884 --> 00:20:29,705
gate operations

540
00:20:30,404 --> 00:20:30,904
between

541
00:20:31,205 --> 00:20:32,744
those logical cubits.

542
00:20:33,205 --> 00:20:35,384
As we then scale up through milestones,

543
00:20:35,924 --> 00:20:36,424
56,

544
00:20:36,725 --> 00:20:38,025
we get a 100,000

545
00:20:38,404 --> 00:20:39,465
or even a1000000

546
00:20:39,960 --> 00:20:40,779
physical cubits,

547
00:20:41,559 --> 00:20:42,839
allowing us to make,

548
00:20:43,720 --> 00:20:46,279
more and more logical cubits, let's say, with

549
00:20:46,279 --> 00:20:48,380
current technologies or current estimates

550
00:20:48,839 --> 00:20:50,380
would be a 100 or 1000.

551
00:20:50,759 --> 00:20:52,839
Most likely by the time we reach those

552
00:20:52,839 --> 00:20:53,339
milestones,

553
00:20:54,295 --> 00:20:57,894
quantum error correction technologies have improved further. Our

554
00:20:57,894 --> 00:20:59,595
hardware has improved further,

555
00:20:59,894 --> 00:21:02,934
and we get a few more logical cubits

556
00:21:02,934 --> 00:21:05,494
out of a given set of or a

557
00:21:05,494 --> 00:21:07,755
given number of physical cubits.

558
00:21:08,360 --> 00:21:10,039
I see. Okay. And do and do you

559
00:21:10,039 --> 00:21:11,980
have any any sort of feeling for,

560
00:21:12,759 --> 00:21:13,820
the the timescale,

561
00:21:14,840 --> 00:21:16,840
you know, when when you will get to

562
00:21:16,840 --> 00:21:17,500
a situation

563
00:21:17,799 --> 00:21:18,299
where

564
00:21:18,680 --> 00:21:20,700
you can you you can start

565
00:21:21,134 --> 00:21:24,335
connecting up these very good logical cubits. Is

566
00:21:24,335 --> 00:21:25,934
that something that is that like a 5

567
00:21:25,934 --> 00:21:27,955
year thing or a 10 year

568
00:21:28,815 --> 00:21:32,174
plan? More about 5 years. We expect that

569
00:21:32,174 --> 00:21:35,695
we will have feature complete quantum computer with

570
00:21:35,695 --> 00:21:36,210
at least

571
00:21:37,009 --> 00:21:40,069
a 100 logical cubits, hopefully more like 1,000

572
00:21:40,450 --> 00:21:41,750
by the end of this decade.

573
00:21:42,450 --> 00:21:45,569
So, I've been asked this by reporters before,

574
00:21:45,569 --> 00:21:48,464
oh, is quantum computing like nuclear fusion, which

575
00:21:48,625 --> 00:21:49,444
famously is

576
00:21:49,744 --> 00:21:50,644
people quit,

577
00:21:51,744 --> 00:21:52,964
20 years out.

578
00:21:53,265 --> 00:21:55,744
Quantum computing is not like this. We publish

579
00:21:55,744 --> 00:21:57,744
the road map and we pretty much knock

580
00:21:57,744 --> 00:21:59,204
out the milestones as

581
00:22:00,304 --> 00:22:01,940
clockwork. So we are making

582
00:22:02,319 --> 00:22:04,740
good progress. Of course, it's a very ambitious

583
00:22:04,960 --> 00:22:07,380
road map and could we be delayed by

584
00:22:07,759 --> 00:22:09,839
a year or 2? Absolutely, that could happen,

585
00:22:09,839 --> 00:22:11,839
but so far, so good. We have pretty

586
00:22:11,839 --> 00:22:15,434
much stayed to true to the predicted timeline.

587
00:22:16,214 --> 00:22:17,974
I see. And and when you get to

588
00:22:17,974 --> 00:22:19,355
a 1,000 cubits,

589
00:22:20,694 --> 00:22:22,934
logical cubits, I mean, I would have thought

590
00:22:22,934 --> 00:22:25,275
that there were practical things

591
00:22:25,734 --> 00:22:27,034
that you could do

592
00:22:27,414 --> 00:22:28,794
with a 1,000

593
00:22:29,230 --> 00:22:29,730
cubits?

594
00:22:30,990 --> 00:22:33,649
You you know, may maybe not solving universal

595
00:22:33,869 --> 00:22:36,609
problems, but are there are there specific

596
00:22:37,069 --> 00:22:40,269
problems, maybe problems in science that you could

597
00:22:40,269 --> 00:22:43,569
tackle with a 1,000 logical cubit machine?

598
00:22:43,904 --> 00:22:46,785
So with a 1,000 logical qubit machine, we

599
00:22:46,785 --> 00:22:48,545
certainly can do many,

600
00:22:48,945 --> 00:22:50,085
useful things.

601
00:22:50,865 --> 00:22:52,644
So we can start to then,

602
00:22:53,105 --> 00:22:53,605
simulate

603
00:22:54,785 --> 00:22:56,884
processes relevant to drug development.

604
00:22:57,345 --> 00:22:57,845
We

605
00:22:58,650 --> 00:23:01,070
develop algorithms that would help,

606
00:23:02,730 --> 00:23:04,990
with the design of nuclear fusion reactors.

607
00:23:05,690 --> 00:23:06,750
There would be applications

608
00:23:07,210 --> 00:23:07,710
in

609
00:23:08,170 --> 00:23:08,670
making

610
00:23:09,049 --> 00:23:10,990
batteries for electric cars

611
00:23:12,575 --> 00:23:14,815
better in the sense of you can charge

612
00:23:14,815 --> 00:23:17,875
them quicker or they are lighter,

613
00:23:18,734 --> 00:23:20,815
less dangerous to burn. I mean, wouldn't it

614
00:23:20,815 --> 00:23:22,674
be awesome if you could have airplanes

615
00:23:23,055 --> 00:23:26,174
as that operate on ion air batteries? In

616
00:23:26,174 --> 00:23:28,890
principle, that's possible. Those designs

617
00:23:30,069 --> 00:23:32,569
have a higher energy density like kerosene.

618
00:23:33,349 --> 00:23:35,369
But today, they're still too brittle.

619
00:23:35,910 --> 00:23:38,869
You can't reliably put those into airplanes yet.

620
00:23:39,429 --> 00:23:40,650
But with,

621
00:23:42,150 --> 00:23:43,210
a quantum computer,

622
00:23:43,555 --> 00:23:46,035
you can hasten the development of such a

623
00:23:46,035 --> 00:23:46,934
device like,

624
00:23:47,315 --> 00:23:48,775
batteries for airplanes

625
00:23:49,394 --> 00:23:52,214
because you don't today, if there's an

626
00:23:52,835 --> 00:23:53,335
electrochemistry

627
00:23:54,035 --> 00:23:56,759
engineer and she has an idea, oh, I

628
00:23:56,759 --> 00:23:59,019
think this is a better cathode material.

629
00:23:59,320 --> 00:24:01,400
Let's put this in. The only way to

630
00:24:01,400 --> 00:24:03,640
test it today is to build this battery,

631
00:24:03,640 --> 00:24:05,480
take it to her to lab, and measure

632
00:24:05,480 --> 00:24:08,059
it. That's a very slow process.

633
00:24:09,079 --> 00:24:11,019
With a quantum computer, you can

634
00:24:11,894 --> 00:24:14,375
simulate this in silico, so to speak, where

635
00:24:14,375 --> 00:24:14,875
you

636
00:24:15,255 --> 00:24:15,755
now,

637
00:24:17,174 --> 00:24:17,674
simulate,

638
00:24:18,055 --> 00:24:20,615
the properties of this battery and see how

639
00:24:20,615 --> 00:24:23,115
fast it would charge or how quickly,

640
00:24:23,974 --> 00:24:24,474
electrons

641
00:24:24,855 --> 00:24:25,355
diffuse.

642
00:24:25,669 --> 00:24:27,609
And then only the very best designs,

643
00:24:28,710 --> 00:24:31,049
that checked out, you take into the lab,

644
00:24:31,990 --> 00:24:34,890
measure in practice how well this would work.

645
00:24:35,509 --> 00:24:37,450
I see. And and what about,

646
00:24:38,230 --> 00:24:38,890
I suppose,

647
00:24:40,095 --> 00:24:42,355
applications that maybe people would associate

648
00:24:43,134 --> 00:24:44,275
with Google?

649
00:24:45,134 --> 00:24:46,755
Things like, information

650
00:24:47,295 --> 00:24:47,795
processing,

651
00:24:48,974 --> 00:24:49,714
the optimization

652
00:24:50,255 --> 00:24:51,394
of of searches,

653
00:24:51,934 --> 00:24:52,335
and,

654
00:24:52,894 --> 00:24:53,954
I suppose AI

655
00:24:54,409 --> 00:24:56,990
when it comes to dealing with large quantities

656
00:24:57,130 --> 00:24:57,789
of information.

657
00:24:58,169 --> 00:24:59,390
Could a a 1,000,

658
00:25:00,809 --> 00:25:03,529
logical cubit machine be useful for that? Or

659
00:25:03,529 --> 00:25:04,669
would you really need

660
00:25:05,049 --> 00:25:06,190
lots more cubits,

661
00:25:06,569 --> 00:25:08,809
logical cubits to to do those sort of

662
00:25:08,809 --> 00:25:09,309
applications?

663
00:25:10,465 --> 00:25:13,424
Yeah. So what we talked about so far

664
00:25:13,424 --> 00:25:13,985
is this,

665
00:25:14,545 --> 00:25:17,525
application area is referred to as quantum simulation.

666
00:25:18,144 --> 00:25:20,465
And we often refer to this as Richard

667
00:25:20,465 --> 00:25:20,965
Feynman's,

668
00:25:21,904 --> 00:25:23,125
killer app.

669
00:25:23,744 --> 00:25:24,244
Because

670
00:25:24,589 --> 00:25:25,490
it was Feynman

671
00:25:25,869 --> 00:25:26,769
who famously,

672
00:25:28,109 --> 00:25:29,650
realized for the first time

673
00:25:30,029 --> 00:25:33,070
that it's actually his quote. Nature is not

674
00:25:33,070 --> 00:25:35,230
classical, and damn it. And if you wanna

675
00:25:35,230 --> 00:25:38,130
simulate nature, you better make it quantum mechanical.

676
00:25:38,625 --> 00:25:42,005
So that's questions like the dynamics of,

677
00:25:43,265 --> 00:25:44,325
chemical reactions

678
00:25:44,865 --> 00:25:45,365
or

679
00:25:45,825 --> 00:25:49,125
properties of magnetic materials or making very

680
00:25:49,424 --> 00:25:49,924
low

681
00:25:50,384 --> 00:25:50,884
resistance,

682
00:25:51,664 --> 00:25:52,964
materials for

683
00:25:53,345 --> 00:25:53,845
electronics.

684
00:25:54,430 --> 00:25:55,490
And all these

685
00:25:56,910 --> 00:25:57,410
problems

686
00:25:59,070 --> 00:25:59,570
entail

687
00:26:00,910 --> 00:26:01,410
quantum

688
00:26:02,509 --> 00:26:04,450
phenomena or these are systems

689
00:26:04,910 --> 00:26:08,269
where quantum effects are important and simulating those

690
00:26:08,269 --> 00:26:09,615
well is sort of the

691
00:26:10,095 --> 00:26:13,535
baseline killer app for quantum computers. But you're

692
00:26:13,535 --> 00:26:15,454
in quite right. It's not limited to this

693
00:26:15,454 --> 00:26:16,115
at all.

694
00:26:16,414 --> 00:26:18,755
Today, we know about 60,

695
00:26:19,214 --> 00:26:19,714
algorithms

696
00:26:20,174 --> 00:26:22,755
that have a scaling advantage, which means,

697
00:26:23,539 --> 00:26:27,079
as, the problems get larger, the quantum computer

698
00:26:27,140 --> 00:26:29,240
can do it more efficiently. Meaning,

699
00:26:29,619 --> 00:26:31,859
it can do the quantum algorithms, can do

700
00:26:31,859 --> 00:26:35,640
it with fewer, sometimes way fewer steps than

701
00:26:35,965 --> 00:26:37,105
a classical computer.

702
00:26:37,884 --> 00:26:40,065
And way fewer steps, I mean, can be

703
00:26:40,605 --> 00:26:41,105
exponential

704
00:26:41,725 --> 00:26:43,345
reduction or a quadratic,

705
00:26:43,965 --> 00:26:46,225
reduction in the number of steps

706
00:26:46,684 --> 00:26:47,424
you need.

707
00:26:49,299 --> 00:26:50,919
And, for example, with optimization,

708
00:26:51,379 --> 00:26:55,159
which is another killer app because optimization problems

709
00:26:55,619 --> 00:26:58,899
are so pervasive. They are, key in machine

710
00:26:58,899 --> 00:26:59,399
learning.

711
00:26:59,700 --> 00:27:00,919
They're key in engineering.

712
00:27:01,220 --> 00:27:04,904
They're important in finance. There's hardly any

713
00:27:05,445 --> 00:27:06,904
area that doesn't,

714
00:27:07,285 --> 00:27:09,224
require the solution of optimization

715
00:27:09,525 --> 00:27:10,025
problems.

716
00:27:10,325 --> 00:27:12,984
And we have known since the nineties that,

717
00:27:13,845 --> 00:27:17,309
for any optimization problem, you at least get

718
00:27:17,309 --> 00:27:17,970
a quadratic,

719
00:27:18,910 --> 00:27:19,970
speed up in

720
00:27:21,390 --> 00:27:21,890
scaling.

721
00:27:23,309 --> 00:27:26,210
But quadratic is not as good as exponential

722
00:27:26,430 --> 00:27:27,170
and therefore,

723
00:27:28,109 --> 00:27:28,634
it would

724
00:27:29,355 --> 00:27:29,835
put,

725
00:27:31,195 --> 00:27:32,654
quantum enhanced optimization

726
00:27:33,355 --> 00:27:34,575
rather far out.

727
00:27:34,954 --> 00:27:37,595
But to our delight, our team has developed

728
00:27:37,595 --> 00:27:39,855
a new algorithm that's called the DQI,

729
00:27:40,474 --> 00:27:43,214
algorithms, stands for decoded quantum interference.

730
00:27:44,160 --> 00:27:45,380
And this algorithm,

731
00:27:47,039 --> 00:27:47,539
seems

732
00:27:48,240 --> 00:27:50,180
to give us an exponential

733
00:27:51,759 --> 00:27:52,580
speed up

734
00:27:53,039 --> 00:27:54,180
in optimization

735
00:27:55,039 --> 00:27:57,140
for certain classes of optimization

736
00:27:57,440 --> 00:27:57,940
problems.

737
00:27:58,315 --> 00:28:01,674
We don't quite understand yet which classes those

738
00:28:01,674 --> 00:28:04,394
are, but if this were to pan out,

739
00:28:04,394 --> 00:28:05,934
this would be super exciting

740
00:28:06,315 --> 00:28:08,255
because you can think of optimization

741
00:28:08,634 --> 00:28:11,035
as puzzle solving. And let's say if you

742
00:28:11,035 --> 00:28:12,000
wanted to

743
00:28:12,640 --> 00:28:13,700
build an AI,

744
00:28:15,039 --> 00:28:17,279
then, of course, an AI that is better

745
00:28:17,279 --> 00:28:19,839
in puzzle solving, that is the one you

746
00:28:19,839 --> 00:28:21,779
will wanna have. So, therefore,

747
00:28:22,319 --> 00:28:25,119
my prediction is that in the future, if

748
00:28:25,119 --> 00:28:27,460
you have a quantum AI playing

749
00:28:28,214 --> 00:28:30,934
chess or go against an AI, the quantum

750
00:28:30,934 --> 00:28:31,914
AI will win.

751
00:28:32,694 --> 00:28:34,075
I see. And and

752
00:28:34,775 --> 00:28:35,994
because of that exponential,

753
00:28:37,494 --> 00:28:39,494
effect, do do does that mean that you

754
00:28:39,494 --> 00:28:40,315
could conceivably

755
00:28:40,694 --> 00:28:41,755
implement those

756
00:28:42,529 --> 00:28:44,710
algorithms on a a 1,000

757
00:28:45,089 --> 00:28:47,569
cubit machine? You you might not have to

758
00:28:47,569 --> 00:28:51,589
wait until you've got 10,000 logical cubits or

759
00:28:51,809 --> 00:28:54,849
a 1000000 logical cubits. That the it's so

760
00:28:54,849 --> 00:28:57,089
efficient that you could implement it on a

761
00:28:57,089 --> 00:28:57,589
small

762
00:28:58,125 --> 00:28:59,105
quantum computer.

763
00:29:00,125 --> 00:29:01,105
We should definitely

764
00:29:01,404 --> 00:29:05,244
see first compelling examples of optimization problems in

765
00:29:05,244 --> 00:29:06,045
the range of,

766
00:29:06,525 --> 00:29:07,505
1,000 variables.

767
00:29:08,045 --> 00:29:09,644
Of course, more is better if you could

768
00:29:09,644 --> 00:29:10,799
have problems with,

769
00:29:11,440 --> 00:29:12,580
10,000 or,

770
00:29:13,200 --> 00:29:16,000
100,000 variables. That would be better. But but

771
00:29:16,000 --> 00:29:18,799
certain optimization problems are very hard, and you

772
00:29:18,799 --> 00:29:21,840
definitely can find the optimal solutions for a

773
00:29:21,840 --> 00:29:23,220
1,000 variable problem.

774
00:29:23,625 --> 00:29:25,625
And if we could show that the quantum

775
00:29:25,625 --> 00:29:26,525
computer can

776
00:29:26,904 --> 00:29:28,365
find better solutions,

777
00:29:29,865 --> 00:29:30,924
don't wanna mislead,

778
00:29:31,384 --> 00:29:32,265
the audience here.

779
00:29:33,065 --> 00:29:33,724
It is

780
00:29:34,744 --> 00:29:38,440
not known today, and many suspect it is

781
00:29:38,440 --> 00:29:39,580
actually not correct,

782
00:29:40,039 --> 00:29:41,180
that quantum computers

783
00:29:42,759 --> 00:29:44,779
can solve very hard optimization

784
00:29:45,640 --> 00:29:46,140
problems

785
00:29:46,840 --> 00:29:47,340
perfectly

786
00:29:47,720 --> 00:29:48,220
either.

787
00:29:49,000 --> 00:29:51,285
But what they often can do is they

788
00:29:51,285 --> 00:29:54,325
give you a much better approximate solutions than

789
00:29:54,325 --> 00:29:57,125
what is classically attainable. So they solve the

790
00:29:57,125 --> 00:29:57,625
puzzles

791
00:29:57,924 --> 00:30:00,485
not necessarily to full optimality, but they can

792
00:30:00,485 --> 00:30:01,545
solve the puzzles

793
00:30:02,005 --> 00:30:03,065
better than classical

794
00:30:03,365 --> 00:30:03,865
computers.

795
00:30:04,750 --> 00:30:06,829
I see. And I don't think, Hartmut, I

796
00:30:06,829 --> 00:30:07,970
I haven't asked you,

797
00:30:08,670 --> 00:30:09,890
specifically about

798
00:30:10,269 --> 00:30:12,589
the Willow chip. So how many,

799
00:30:13,710 --> 00:30:14,769
physical cubits,

800
00:30:15,390 --> 00:30:16,450
does it integrate?

801
00:30:18,304 --> 00:30:20,644
So the Willow chip has a 105,

802
00:30:21,265 --> 00:30:22,164
physical cubits,

803
00:30:22,785 --> 00:30:23,845
and they have,

804
00:30:24,464 --> 00:30:25,525
very high quality.

805
00:30:25,984 --> 00:30:26,724
The physical

806
00:30:27,825 --> 00:30:30,224
error rates, let's say, for our 2 cubit

807
00:30:30,224 --> 00:30:30,630
gate

808
00:30:31,109 --> 00:30:33,369
gates is just 1 in a 1000.

809
00:30:33,829 --> 00:30:34,549
So this,

810
00:30:34,950 --> 00:30:39,130
allows to run the most complex quantum algorithms

811
00:30:39,430 --> 00:30:41,289
today on the Willow chip.

812
00:30:41,670 --> 00:30:43,750
Well, that's great, Hartmut. Thanks so much for

813
00:30:43,750 --> 00:30:44,809
coming on the podcast.

814
00:30:45,875 --> 00:30:48,195
Oh, most welcome. Pleasure to be on your

815
00:30:48,195 --> 00:30:48,695
show.

816
00:30:55,955 --> 00:30:58,375
You're listening to 1 of 2 podcasts

817
00:30:58,829 --> 00:31:01,329
with our breakthrough of the year winners.

818
00:31:01,950 --> 00:31:04,450
The other features Harvard University's

819
00:31:04,990 --> 00:31:06,049
Mikhail Lukin

820
00:31:06,349 --> 00:31:07,970
and Dolev Blufstein,

821
00:31:08,750 --> 00:31:10,750
and you can find it on the Physics

822
00:31:10,750 --> 00:31:11,650
World website,

823
00:31:12,029 --> 00:31:14,945
or at your favorite podcast provider.

824
00:31:15,724 --> 00:31:18,285
You can also read more about our top

825
00:31:18,285 --> 00:31:20,144
10 breakthroughs of 2024

826
00:31:21,005 --> 00:31:22,865
on the Physics World website.

827
00:31:23,565 --> 00:31:26,705
This served as the shortlist for our breakthrough

828
00:31:26,845 --> 00:31:29,470
of the year award, and it covers a

829
00:31:29,470 --> 00:31:30,849
range of fantastic

830
00:31:31,230 --> 00:31:32,529
research in Physics.

831
00:31:32,990 --> 00:31:34,529
So do check it out.

832
00:31:35,150 --> 00:31:37,309
I'm afraid that's all the time we have

833
00:31:37,309 --> 00:31:38,690
for this week's podcast.

834
00:31:39,150 --> 00:31:42,529
Thanks to Hartmut Nevin for joining me today,

835
00:31:42,784 --> 00:31:45,345
and a special thanks to our producer Fred

836
00:31:45,345 --> 00:31:45,845
Ailes.

837
00:31:46,784 --> 00:31:49,505
Physics World's coverage of the breakthrough of the

838
00:31:49,505 --> 00:31:53,424
year is supported by Reports on Progress in

839
00:31:53,424 --> 00:31:53,924
Physics,

840
00:31:54,625 --> 00:31:55,365
which offers

841
00:31:55,825 --> 00:31:56,325
unparalleled

842
00:31:57,024 --> 00:31:57,524
visibility

843
00:31:57,839 --> 00:31:58,819
for your groundbreaking

844
00:31:59,440 --> 00:31:59,940
research.

845
00:32:00,480 --> 00:32:02,900
You can find the journal at iopscience.org.