Deep connections: why two AI pioneers won the Nobel Prize for Physics

1
00:00:07,839 --> 00:00:10,719
Hello, and welcome to the Physics World Weekly

2
00:00:10,719 --> 00:00:12,580
podcast. I'm Hamish Johnston.

3
00:00:13,154 --> 00:00:15,074
This week, we're going to be chatting about

4
00:00:15,074 --> 00:00:15,894
the 2024

5
00:00:17,074 --> 00:00:18,695
Nobel Prize for Physics,

6
00:00:19,154 --> 00:00:20,454
which rather surprisingly

7
00:00:21,074 --> 00:00:24,134
was given for work done in machine learning

8
00:00:24,355 --> 00:00:25,974
and artificial intelligence.

9
00:00:27,309 --> 00:00:30,210
This episode is brought to you by Smaract,

10
00:00:30,910 --> 00:00:33,090
elevating high precision positioning,

11
00:00:33,630 --> 00:00:35,570
metrology, and automation

12
00:00:36,030 --> 00:00:37,809
to empower your breakthroughs.

13
00:00:38,670 --> 00:00:41,729
Shape the future with Smaract Technologies.

14
00:00:43,454 --> 00:00:44,195
On Tuesday,

15
00:00:44,495 --> 00:00:47,615
the Nobel Prize for Physics was awarded to

16
00:00:47,615 --> 00:00:48,594
John Hopfield

17
00:00:49,054 --> 00:00:50,435
of Princeton University

18
00:00:50,975 --> 00:00:54,594
and Geoffrey Hinton of the University of Toronto

19
00:00:55,090 --> 00:00:56,469
for their foundational

20
00:00:57,090 --> 00:00:58,710
discoveries and inventions

21
00:00:59,170 --> 00:01:03,510
that enable machine learning and artificial neural networks.

22
00:01:04,609 --> 00:01:07,489
Needless to say, we didn't see that one

23
00:01:07,489 --> 00:01:09,670
coming here at physics world.

24
00:01:10,215 --> 00:01:13,915
We normally don't consider machine learning and AI

25
00:01:14,295 --> 00:01:16,234
to be disciplines of physics.

26
00:01:16,614 --> 00:01:17,754
And while Hotfield

27
00:01:18,055 --> 00:01:18,875
is a physicist,

28
00:01:19,495 --> 00:01:20,875
Hinton is not.

29
00:01:21,575 --> 00:01:23,834
But we really shouldn't have been surprised.

30
00:01:24,879 --> 00:01:27,219
AI is beginning to change our world,

31
00:01:27,680 --> 00:01:29,140
and some of the mathematics

32
00:01:29,519 --> 00:01:30,420
that it uses

33
00:01:30,799 --> 00:01:33,219
is deeply rooted in physics.

34
00:01:40,854 --> 00:01:43,274
To talk about the prize and the connections

35
00:01:43,494 --> 00:01:45,435
between AI and physics,

36
00:01:46,055 --> 00:01:48,555
I'm joined down the line from California

37
00:01:49,174 --> 00:01:50,075
by Anil

38
00:01:50,375 --> 00:01:50,875
Ananthaswami,

39
00:01:51,814 --> 00:01:55,034
who has written the book, Why Machines Learn,

40
00:01:55,390 --> 00:01:56,689
the Elegant Maths

41
00:01:56,990 --> 00:01:57,969
Behind Modern

42
00:01:58,270 --> 00:01:58,770
AI.

43
00:01:59,790 --> 00:02:02,049
Hi, Anil. Welcome to the podcast.

44
00:02:02,590 --> 00:02:04,670
Hi, Hamish. Thank you very much for having

45
00:02:04,670 --> 00:02:06,210
me. It's my pleasure.

46
00:02:06,750 --> 00:02:09,389
So, Anil, here at, at Physics World, we

47
00:02:09,389 --> 00:02:10,935
were really surprised

48
00:02:11,634 --> 00:02:13,955
by this prize. We were completely caught off

49
00:02:13,955 --> 00:02:17,254
guard if truth be told. Were you surprised?

50
00:02:18,914 --> 00:02:22,534
I was surprised. I woke up, yesterday morning,

51
00:02:22,834 --> 00:02:25,175
and my Twitter feed was a buzz.

52
00:02:26,010 --> 00:02:27,469
Yeah. I I was surprised,

53
00:02:28,090 --> 00:02:30,569
but not necessarily in a bad way. I

54
00:02:30,569 --> 00:02:32,270
think I was pleasantly surprised.

55
00:02:33,210 --> 00:02:35,210
I suppose here at Physics World, we don't

56
00:02:35,210 --> 00:02:36,110
really consider,

57
00:02:37,449 --> 00:02:38,990
AI and machine learning

58
00:02:39,370 --> 00:02:40,030
to be,

59
00:02:41,394 --> 00:02:42,455
to be physics.

60
00:02:42,914 --> 00:02:43,235
But,

61
00:02:44,275 --> 00:02:45,174
of course, Hotfield

62
00:02:45,715 --> 00:02:46,614
is a physicist.

63
00:02:46,995 --> 00:02:48,215
And there's lots of

64
00:02:48,594 --> 00:02:49,094
mathematics

65
00:02:49,555 --> 00:02:51,254
and, and and I suppose,

66
00:02:51,555 --> 00:02:52,614
physics related

67
00:02:53,074 --> 00:02:53,574
mathematics

68
00:02:53,955 --> 00:02:54,455
in,

69
00:02:55,155 --> 00:02:57,400
AI and machine learning. So

70
00:02:57,780 --> 00:03:00,020
maybe it's not that surprising. And and I

71
00:03:00,020 --> 00:03:02,340
would guess that you're, you know I think

72
00:03:02,340 --> 00:03:05,060
I was pleasantly surprised because I thought that,

73
00:03:05,460 --> 00:03:08,260
yes, this is a very important issue in

74
00:03:08,260 --> 00:03:09,719
society at the moment.

75
00:03:10,185 --> 00:03:12,985
And, you know, it's great that the, that

76
00:03:12,985 --> 00:03:15,805
the Nobel Prize Committee is is is focusing

77
00:03:16,425 --> 00:03:17,645
the public's attention

78
00:03:18,105 --> 00:03:18,765
on it.

79
00:03:19,145 --> 00:03:20,105
Yeah. I mean, I think,

80
00:03:20,905 --> 00:03:22,745
you know, when you look at all the

81
00:03:22,745 --> 00:03:23,245
reactions

82
00:03:24,530 --> 00:03:25,430
across the board,

83
00:03:26,050 --> 00:03:28,610
there are certainly people who are upset about

84
00:03:28,610 --> 00:03:29,110
this,

85
00:03:30,370 --> 00:03:33,430
thinking that the standards for awarding the nobles

86
00:03:33,489 --> 00:03:34,870
are a bit inconsistent.

87
00:03:36,289 --> 00:03:38,689
Others are wondering whether the Nobel Committee has

88
00:03:38,689 --> 00:03:40,425
given into the AI hype

89
00:03:41,784 --> 00:03:43,085
and things like that.

90
00:03:43,625 --> 00:03:46,205
And, of course, questions about whether machine learning

91
00:03:46,745 --> 00:03:47,564
really is,

92
00:03:48,425 --> 00:03:49,485
physics belongs,

93
00:03:49,784 --> 00:03:51,705
you know, should we be talking about machine

94
00:03:51,705 --> 00:03:53,784
learning and physics in the same breath and

95
00:03:53,784 --> 00:03:54,444
so on?

96
00:03:56,050 --> 00:03:58,710
I have also seen completely different

97
00:03:59,090 --> 00:03:59,830
and more,

98
00:04:00,610 --> 00:04:04,050
inclusive and positive reactions about this Nobel. There

99
00:04:04,050 --> 00:04:06,550
are people who have very strongly argued that,

100
00:04:07,810 --> 00:04:08,310
physics

101
00:04:08,764 --> 00:04:10,284
plays a very important role,

102
00:04:10,604 --> 00:04:11,745
in machine learning,

103
00:04:13,004 --> 00:04:15,185
as does neuroscience, as does mathematics,

104
00:04:16,044 --> 00:04:16,605
you know,

105
00:04:17,084 --> 00:04:17,824
and computational,

106
00:04:19,564 --> 00:04:22,819
or other computer science. So I think, you

107
00:04:22,819 --> 00:04:24,819
know, you can you can easily imagine the

108
00:04:24,819 --> 00:04:27,639
furor if this prize had been awarded

109
00:04:28,259 --> 00:04:29,319
in medicine

110
00:04:29,699 --> 00:04:32,279
and or neuroscience or something like that.

111
00:04:34,419 --> 00:04:36,645
So my take is that, you know, physics

112
00:04:36,645 --> 00:04:37,145
does,

113
00:04:38,004 --> 00:04:38,504
inform,

114
00:04:38,964 --> 00:04:40,425
machine learning in very,

115
00:04:41,285 --> 00:04:43,444
strong ways. And and the reverse is also

116
00:04:43,444 --> 00:04:45,545
true that machine learning is now playing

117
00:04:45,925 --> 00:04:48,004
an important role in how physics is being

118
00:04:48,004 --> 00:04:48,485
done,

119
00:04:48,884 --> 00:04:50,279
and we can talk about that.

120
00:04:51,000 --> 00:04:52,839
That's right. Yeah. We'll we'll chat about that

121
00:04:52,839 --> 00:04:55,480
a bit later. But, first, I I wanted

122
00:04:55,480 --> 00:04:57,720
to ask you about the winners. Now John

123
00:04:57,720 --> 00:05:01,240
Hopfield is a, a bona fide physicist. He

124
00:05:01,240 --> 00:05:04,514
started out his career in condensed matter physics.

125
00:05:04,514 --> 00:05:06,194
Can can you tell us a bit about

126
00:05:06,194 --> 00:05:06,694
him

127
00:05:07,074 --> 00:05:09,254
and, and and his contributions

128
00:05:09,634 --> 00:05:11,814
to machine learning and AI?

129
00:05:12,514 --> 00:05:14,675
Yeah. Like you said, he started off in

130
00:05:14,675 --> 00:05:17,154
condensed matter physics. And, you know, at some

131
00:05:17,154 --> 00:05:19,475
point in his career in the seventies or

132
00:05:19,475 --> 00:05:19,975
so,

133
00:05:20,410 --> 00:05:22,810
he kind of felt like he had used

134
00:05:22,810 --> 00:05:25,529
up all his particular talents as he called

135
00:05:25,529 --> 00:05:26,750
them, in

136
00:05:27,529 --> 00:05:29,529
in solid state physics. And he was looking

137
00:05:29,529 --> 00:05:30,029
for

138
00:05:30,490 --> 00:05:32,970
new avenues of research, and he actually ended

139
00:05:32,970 --> 00:05:34,615
up moving from

140
00:05:34,995 --> 00:05:37,735
solid state physics and condensed matter physics to,

141
00:05:38,834 --> 00:05:41,795
studying the dynamics of biochemical reactions. And he

142
00:05:41,795 --> 00:05:45,154
made some seminal contributions there. And and the

143
00:05:45,154 --> 00:05:46,995
irony was that he was at Princeton at

144
00:05:46,995 --> 00:05:47,574
the time,

145
00:05:47,930 --> 00:05:50,490
and what he was doing wasn't considered physics

146
00:05:50,490 --> 00:05:52,350
and he ended up moving to Caltech.

147
00:05:54,009 --> 00:05:56,350
And, it was at Caltech that he then

148
00:05:56,490 --> 00:05:58,410
started thinking further about,

149
00:05:59,209 --> 00:06:01,149
how to then take his understanding

150
00:06:01,529 --> 00:06:03,935
of the dynamics of, these,

151
00:06:04,394 --> 00:06:05,294
you know, biochemical,

152
00:06:06,394 --> 00:06:07,454
reaction networks,

153
00:06:09,034 --> 00:06:11,514
to other fields. He was, in particular, very

154
00:06:11,514 --> 00:06:12,495
interested in,

155
00:06:13,274 --> 00:06:15,115
seeing if he could make a contribution to

156
00:06:15,115 --> 00:06:17,214
neuroscience, to computational neuroscience.

157
00:06:17,970 --> 00:06:18,449
And,

158
00:06:19,329 --> 00:06:21,329
and he kept looking for a problem that

159
00:06:21,329 --> 00:06:23,329
he wanted to solve. And eventually, he found

160
00:06:23,329 --> 00:06:25,029
something in neuroscience where

161
00:06:25,490 --> 00:06:26,389
his ideas,

162
00:06:27,649 --> 00:06:29,490
that he had been developing so far could

163
00:06:29,490 --> 00:06:30,149
be applied.

164
00:06:31,084 --> 00:06:32,144
And this was,

165
00:06:32,604 --> 00:06:36,464
the problem of, associative memories, and associative memories

166
00:06:36,524 --> 00:06:38,285
are you know, we all have them.

167
00:06:39,245 --> 00:06:41,425
Imagine you have experienced something,

168
00:06:43,800 --> 00:06:44,939
some some episodic

169
00:06:45,879 --> 00:06:46,699
memory where

170
00:06:47,319 --> 00:06:50,300
this the thing that you experienced had strong

171
00:06:50,519 --> 00:06:52,620
smells or, you know, there was

172
00:06:53,160 --> 00:06:55,000
a song playing in the background and that

173
00:06:55,000 --> 00:06:58,055
thing becomes embedded in your memory. And then

174
00:06:58,115 --> 00:06:59,175
many months later,

175
00:06:59,475 --> 00:07:01,915
a fragment of that smell or a a

176
00:07:01,915 --> 00:07:04,194
a fragment of that song that you heard,

177
00:07:05,235 --> 00:07:06,055
comes into,

178
00:07:06,834 --> 00:07:09,314
your experience and that entire memory is recalled

179
00:07:09,314 --> 00:07:12,214
and and and this is associative memory. And

180
00:07:13,370 --> 00:07:14,910
hopefully realized that

181
00:07:15,370 --> 00:07:17,790
the computational problem of trying to solve,

182
00:07:18,730 --> 00:07:20,990
how to store and recall memories,

183
00:07:21,769 --> 00:07:23,689
could be something he could apply his talents

184
00:07:23,689 --> 00:07:25,389
to, and that was the

185
00:07:26,115 --> 00:07:26,995
beginning of,

186
00:07:27,634 --> 00:07:30,375
his work on Hopfield Networks. He essentially,

187
00:07:31,875 --> 00:07:32,375
designed,

188
00:07:32,915 --> 00:07:35,574
what are today called the current neural networks.

189
00:07:36,355 --> 00:07:37,654
And he showed how,

190
00:07:38,275 --> 00:07:41,254
you know, the principles that from condensed metaphysics,

191
00:07:42,379 --> 00:07:44,300
this are this is the physics of spin

192
00:07:44,300 --> 00:07:44,800
glasses,

193
00:07:45,819 --> 00:07:47,519
could be applied to

194
00:07:47,979 --> 00:07:48,479
train,

195
00:07:48,939 --> 00:07:49,759
such associative

196
00:07:50,860 --> 00:07:53,819
memory networks and recall such memories given a

197
00:07:53,819 --> 00:07:57,120
fragment of the information or given corrupted information.

198
00:07:57,904 --> 00:07:59,605
Yeah. It is. I mean, it is incredible.

199
00:07:59,665 --> 00:07:59,985
The,

200
00:08:00,545 --> 00:08:02,404
yeah, I suppose the connection between

201
00:08:02,785 --> 00:08:05,764
spin glasses, you know, a huge problem in

202
00:08:05,824 --> 00:08:07,904
in condensed matter physics, and,

203
00:08:08,704 --> 00:08:10,564
machine learning and AI.

204
00:08:11,050 --> 00:08:12,430
Despite the fact that Hopfield,

205
00:08:12,889 --> 00:08:15,370
is a physicist, I have to admit, that

206
00:08:15,370 --> 00:08:18,509
I hadn't really heard of him until, yesterday,

207
00:08:19,530 --> 00:08:20,670
or I should say,

208
00:08:21,449 --> 00:08:22,830
Tuesday when the,

209
00:08:23,290 --> 00:08:26,115
when the, when the announcement was made. But

210
00:08:26,115 --> 00:08:27,254
I had heard of Hinton,

211
00:08:27,875 --> 00:08:29,175
who's not a physicist.

212
00:08:29,875 --> 00:08:31,955
And I think that's particularly because he I

213
00:08:31,955 --> 00:08:32,855
think he's become

214
00:08:33,235 --> 00:08:34,375
a bit of a celebrity

215
00:08:34,914 --> 00:08:36,695
even before winning the prize.

216
00:08:37,460 --> 00:08:38,600
He's made some

217
00:08:39,059 --> 00:08:39,559
warnings

218
00:08:39,860 --> 00:08:40,360
about

219
00:08:41,139 --> 00:08:43,639
artificial intelligence and how they could affect

220
00:08:44,740 --> 00:08:47,660
society. So so who is Geoffrey Hinton? And,

221
00:08:48,100 --> 00:08:49,559
and what what has he done,

222
00:08:50,259 --> 00:08:52,815
in the world of AI and machine learning?

223
00:08:54,075 --> 00:08:56,095
Oh, Geoffrey Hinton's fame,

224
00:08:56,634 --> 00:08:57,774
regarding AI,

225
00:08:58,794 --> 00:08:59,294
goes,

226
00:08:59,834 --> 00:09:02,315
beyond just warning us about the dangers of

227
00:09:02,315 --> 00:09:04,649
AI. He he truly is someone who has,

228
00:09:05,290 --> 00:09:07,070
been instrumental in,

229
00:09:07,769 --> 00:09:08,430
you know,

230
00:09:09,290 --> 00:09:10,670
neural network research.

231
00:09:11,129 --> 00:09:13,389
So for instance, if you go back to

232
00:09:13,769 --> 00:09:16,269
the late 19 fifties and early 19 sixties

233
00:09:16,330 --> 00:09:17,230
when the first,

234
00:09:18,410 --> 00:09:19,870
neural networks were

235
00:09:20,445 --> 00:09:22,305
designed and built. These were,

236
00:09:22,764 --> 00:09:25,725
so called perceptrons or single layer neural networks

237
00:09:25,725 --> 00:09:28,225
that were designed by Frank Rosenblatt,

238
00:09:28,605 --> 00:09:30,705
who was a, Cornell University,

239
00:09:32,524 --> 00:09:33,024
psychologist.

240
00:09:34,559 --> 00:09:36,659
Those networks had big limitations.

241
00:09:37,199 --> 00:09:39,299
And, in the 19 sixties,

242
00:09:40,480 --> 00:09:44,179
MIT researchers, Marvin Minsky and Seymour Papert, they

243
00:09:44,399 --> 00:09:46,019
pretty much called poor

244
00:09:46,559 --> 00:09:47,445
rather they

245
00:09:48,004 --> 00:09:50,725
kind of pour cold water on, neural network

246
00:09:50,725 --> 00:09:53,684
research by pointing out that these single layer

247
00:09:53,684 --> 00:09:55,785
neural networks couldn't really solve

248
00:09:57,125 --> 00:09:57,945
very important

249
00:09:58,325 --> 00:10:00,644
or even very simple problems of a certain

250
00:10:00,644 --> 00:10:01,144
kind.

251
00:10:01,879 --> 00:10:04,279
All of that research died, and there were

252
00:10:04,279 --> 00:10:07,000
very few people who persisted and who believed

253
00:10:07,000 --> 00:10:08,700
that neural networks would eventually,

254
00:10:09,240 --> 00:10:11,720
solve the kinds of problems that, they are

255
00:10:11,720 --> 00:10:14,460
actually solving now. And one of those researchers

256
00:10:14,519 --> 00:10:15,580
was Geoff Hinton.

257
00:10:16,325 --> 00:10:18,105
He was doing his PhD,

258
00:10:18,644 --> 00:10:19,384
in Edinburgh,

259
00:10:20,644 --> 00:10:22,825
and, you know, despite his advisors,

260
00:10:23,924 --> 00:10:26,184
lack of interest in neural networks, he persisted.

261
00:10:26,325 --> 00:10:28,264
He kept doing it. And

262
00:10:29,360 --> 00:10:32,480
it was sometime in the mid 19 eighties

263
00:10:32,480 --> 00:10:32,980
that,

264
00:10:33,840 --> 00:10:37,379
Geoffrey Hinton basically with David Jummelhart and,

265
00:10:38,639 --> 00:10:42,000
Williams came up with the paper that showed

266
00:10:42,000 --> 00:10:42,500
how

267
00:10:43,105 --> 00:10:45,264
deep neural networks could be trained using the

268
00:10:45,264 --> 00:10:48,085
back propagation algorithms. They showed how

269
00:10:48,465 --> 00:10:51,264
these networks could learn very important features that

270
00:10:51,264 --> 00:10:51,764
exists,

271
00:10:52,304 --> 00:10:53,045
in data.

272
00:10:53,665 --> 00:10:55,665
And, that was the beginning of the deep

273
00:10:55,665 --> 00:10:56,565
learning revolution.

274
00:10:57,169 --> 00:10:58,629
This particular Nobel,

275
00:10:59,409 --> 00:11:01,649
that Jeff Hinton has been given is for

276
00:11:01,649 --> 00:11:05,269
work that actually predates the back propagation algorithm.

277
00:11:05,569 --> 00:11:08,069
It was for something called Boltzmann's machine.

278
00:11:09,345 --> 00:11:09,745
And,

279
00:11:10,144 --> 00:11:11,845
Boltzmann machines are,

280
00:11:12,625 --> 00:11:15,345
something that follow on from Hopfield's work. So

281
00:11:15,345 --> 00:11:17,825
Hopfield designed Hopfield networks that we were just

282
00:11:17,825 --> 00:11:18,644
talking about.

283
00:11:19,424 --> 00:11:22,304
And Boltzmann machines are further advanced on that,

284
00:11:22,304 --> 00:11:22,625
and,

285
00:11:23,264 --> 00:11:25,820
they use ideas from statistical physics.

286
00:11:26,920 --> 00:11:27,399
So,

287
00:11:27,879 --> 00:11:28,379
honestly,

288
00:11:28,840 --> 00:11:31,639
you know, Jeff Hinton's contributions to AI are

289
00:11:31,639 --> 00:11:32,920
just immense because,

290
00:11:33,559 --> 00:11:34,460
he literally,

291
00:11:35,240 --> 00:11:37,320
was one of the key figures that led

292
00:11:37,320 --> 00:11:39,240
to the deep learning revolution that we are

293
00:11:39,240 --> 00:11:39,980
seeing today.

294
00:11:40,825 --> 00:11:41,884
I see. And

295
00:11:42,504 --> 00:11:44,764
am I right in saying that he resigned

296
00:11:45,144 --> 00:11:45,644
recently

297
00:11:46,184 --> 00:11:47,085
from a position

298
00:11:47,544 --> 00:11:48,764
was it at Google,

299
00:11:50,264 --> 00:11:52,684
because he wanted to speak out

300
00:11:53,159 --> 00:11:55,399
about AI. Can can you talk a bit

301
00:11:55,399 --> 00:11:58,299
about that? What what were some of his

302
00:11:58,440 --> 00:11:59,339
key concerns?

303
00:12:01,399 --> 00:12:03,240
Yes. You're right. I think he did resign

304
00:12:03,240 --> 00:12:04,220
from Google.

305
00:12:04,679 --> 00:12:07,000
He said that he resigned specifically so that

306
00:12:07,000 --> 00:12:10,995
he could talk about the potential dangers of,

307
00:12:11,455 --> 00:12:11,955
AI,

308
00:12:13,134 --> 00:12:13,535
and,

309
00:12:14,415 --> 00:12:16,975
and he's been, he's been talking about that

310
00:12:16,975 --> 00:12:19,695
openly now. He is basically making people aware

311
00:12:19,695 --> 00:12:21,075
that we should be,

312
00:12:22,580 --> 00:12:25,299
you know, paying attention to what's coming down

313
00:12:25,299 --> 00:12:27,139
the line, that if we are not able

314
00:12:27,139 --> 00:12:27,639
to

315
00:12:28,340 --> 00:12:30,200
regulate and control this technology,

316
00:12:30,899 --> 00:12:32,919
this might pose a danger to,

317
00:12:33,620 --> 00:12:35,855
you know, all of us. So he's been

318
00:12:35,995 --> 00:12:38,235
pretty vocal about it, and it's been quite

319
00:12:38,235 --> 00:12:38,735
surprising

320
00:12:39,115 --> 00:12:39,615
how,

321
00:12:41,034 --> 00:12:43,914
you know, people like him and Joshua Benjio

322
00:12:43,914 --> 00:12:46,394
and others have changed their mind over the

323
00:12:46,394 --> 00:12:47,534
last few years.

324
00:12:47,860 --> 00:12:49,779
Because if you go back just 2 years

325
00:12:49,779 --> 00:12:51,879
or so, we weren't hearing such,

326
00:12:52,740 --> 00:12:54,259
fears voiced by,

327
00:12:55,220 --> 00:12:57,299
the main players some of the main players

328
00:12:57,299 --> 00:12:59,639
in AI. Not all of them, of course.

329
00:13:00,580 --> 00:13:03,815
So yeah. I mean, that's something he's been

330
00:13:03,815 --> 00:13:05,894
very concerned about, and he's been very vocal

331
00:13:05,894 --> 00:13:06,855
about in the last,

332
00:13:07,335 --> 00:13:09,754
in a year or so. We've spoken about

333
00:13:09,975 --> 00:13:10,955
spin glasses.

334
00:13:12,055 --> 00:13:13,355
We've spoken about,

335
00:13:14,535 --> 00:13:15,035
Boltzmann

336
00:13:15,415 --> 00:13:18,820
machines. I mean, this sounds like some real

337
00:13:18,820 --> 00:13:21,379
physics here. Can you is it possible for

338
00:13:21,379 --> 00:13:22,679
you to give us a flavor

339
00:13:23,379 --> 00:13:25,960
of of how the concepts of physics

340
00:13:26,340 --> 00:13:28,040
have come into machine learning

341
00:13:28,419 --> 00:13:30,919
and AI and and how they're used?

342
00:13:31,725 --> 00:13:33,424
So John Hopfield basically,

343
00:13:35,085 --> 00:13:35,985
designed his,

344
00:13:36,845 --> 00:13:37,825
Hopfield network,

345
00:13:39,004 --> 00:13:41,184
and based it on the icing model,

346
00:13:42,365 --> 00:13:43,664
of magnetic materials.

347
00:13:44,410 --> 00:13:47,070
And the idea there was that his artificial

348
00:13:47,210 --> 00:13:50,110
neural network had neurons that were bidirectionally

349
00:13:50,730 --> 00:13:51,230
connected.

350
00:13:52,730 --> 00:13:53,230
And

351
00:13:53,529 --> 00:13:56,809
they had dynamics the very in much the

352
00:13:56,809 --> 00:13:57,850
same manner that,

353
00:13:58,490 --> 00:13:59,629
magnetic materials,

354
00:14:00,595 --> 00:14:04,215
when their spins are disturbed, will settle back

355
00:14:04,675 --> 00:14:06,055
into a stable state

356
00:14:06,675 --> 00:14:07,735
and become ferromagnetic

357
00:14:08,195 --> 00:14:11,575
because that particular state represents an energy minimum.

358
00:14:11,634 --> 00:14:13,815
He figured out how to

359
00:14:15,179 --> 00:14:16,079
depict the

360
00:14:16,539 --> 00:14:18,879
energy, quote, unquote energy of

361
00:14:19,899 --> 00:14:21,279
a a recurrent neural network

362
00:14:21,740 --> 00:14:22,240
and

363
00:14:22,620 --> 00:14:24,319
showed how you could store

364
00:14:24,860 --> 00:14:25,360
information

365
00:14:27,054 --> 00:14:28,915
or a memory into these networks

366
00:14:29,295 --> 00:14:32,035
in a way that the stored memory represented

367
00:14:32,095 --> 00:14:34,274
an energy minimum of that network.

368
00:14:34,735 --> 00:14:36,815
And then, if you were to preserve that

369
00:14:36,815 --> 00:14:39,955
network, so which essentially meant you were corrupting

370
00:14:40,254 --> 00:14:40,835
the memory,

371
00:14:41,440 --> 00:14:44,000
the network could dynamically find its way back

372
00:14:44,000 --> 00:14:46,399
to that energy minimum. And that then when

373
00:14:46,399 --> 00:14:47,940
it reached that energy minimum,

374
00:14:48,480 --> 00:14:50,320
you could just read off the outputs of

375
00:14:50,320 --> 00:14:50,980
the neurons

376
00:14:51,360 --> 00:14:53,059
and recover that memory.

377
00:14:53,440 --> 00:14:53,940
So,

378
00:14:55,264 --> 00:14:57,904
it was very, very strongly based on his

379
00:14:57,904 --> 00:14:59,105
understanding of,

380
00:14:59,824 --> 00:15:01,365
you you know, the icing model.

381
00:15:01,985 --> 00:15:04,144
And then so, obviously, physics played a very

382
00:15:04,144 --> 00:15:07,125
large part in the development of Hopfield networks.

383
00:15:07,460 --> 00:15:07,960
And,

384
00:15:08,340 --> 00:15:10,740
Anil, Hinton, you know, as you said, took

385
00:15:10,740 --> 00:15:12,279
took this idea further

386
00:15:12,740 --> 00:15:13,240
using,

387
00:15:14,019 --> 00:15:16,679
using concepts from physics to create this,

388
00:15:17,299 --> 00:15:20,040
Boltzmann machine. And and those are based on,

389
00:15:20,179 --> 00:15:21,160
I'm I'm guessing,

390
00:15:21,644 --> 00:15:23,504
ideas from statistical physics.

391
00:15:24,524 --> 00:15:26,545
Yes. That's absolutely correct. So,

392
00:15:27,004 --> 00:15:29,024
Hopfield networks are deterministic.

393
00:15:29,884 --> 00:15:30,205
And,

394
00:15:31,404 --> 00:15:34,764
what Hinton did was Hinton, along with Terry

395
00:15:34,764 --> 00:15:36,225
Sechnovsky and others,

396
00:15:37,129 --> 00:15:38,570
What he did was,

397
00:15:38,970 --> 00:15:39,470
to

398
00:15:39,850 --> 00:15:41,309
make his networks,

399
00:15:42,009 --> 00:15:43,309
stochastic, and

400
00:15:43,690 --> 00:15:44,750
he did take

401
00:15:45,129 --> 00:15:46,509
strong cues from,

402
00:15:47,450 --> 00:15:47,950
thermodynamics

403
00:15:48,490 --> 00:15:49,950
and Boltzmann distributions.

404
00:15:50,834 --> 00:15:54,454
So, Boltzmann machines are also recurrent neural networks,

405
00:15:55,875 --> 00:15:58,274
which means that the artificial neurons, which are

406
00:15:58,274 --> 00:16:00,694
computational units, are bidirectionally connected.

407
00:16:01,554 --> 00:16:02,054
Recurrent,

408
00:16:02,834 --> 00:16:05,495
the the kind of recurrent neural networks that

409
00:16:06,370 --> 00:16:09,649
Hinton uses, in Boltzmann machines have so called

410
00:16:09,649 --> 00:16:12,450
visible neurons, which are basically the neurons that

411
00:16:12,450 --> 00:16:14,210
you can access, but they also have hidden

412
00:16:14,210 --> 00:16:14,710
neurons.

413
00:16:15,889 --> 00:16:18,290
And but the entire thing is stochastic. And

414
00:16:18,290 --> 00:16:20,070
the whole idea is that

415
00:16:21,014 --> 00:16:23,514
you ought to be able to model probability

416
00:16:23,735 --> 00:16:24,235
distributions

417
00:16:24,615 --> 00:16:25,434
over data.

418
00:16:26,215 --> 00:16:27,434
So you're learning

419
00:16:28,934 --> 00:16:31,254
features or patterns that exist in data in

420
00:16:31,254 --> 00:16:33,675
order to be able to model probability distributions

421
00:16:33,815 --> 00:16:34,555
over data.

422
00:16:35,360 --> 00:16:38,639
And these are, the dynamics of these networks

423
00:16:38,639 --> 00:16:39,379
are designed,

424
00:16:39,919 --> 00:16:42,179
such that the network finds its way,

425
00:16:43,120 --> 00:16:46,159
to some thermodynamic equilibrium. Again, the the use

426
00:16:46,159 --> 00:16:46,980
of the word

427
00:16:47,304 --> 00:16:47,804
thermodynamic

428
00:16:48,105 --> 00:16:50,605
equilibrium or energy state, these are just proxies.

429
00:16:50,665 --> 00:16:53,725
They're obviously, these are pieces of software, so

430
00:16:54,105 --> 00:16:55,245
they don't have,

431
00:16:55,625 --> 00:16:57,625
you know, energy in the same way that

432
00:16:57,625 --> 00:17:00,985
physical systems do. But conceptually, they're operating in

433
00:17:00,985 --> 00:17:02,009
in the same way.

434
00:17:02,649 --> 00:17:05,230
And so Boltzmann machines are essentially

435
00:17:06,009 --> 00:17:09,529
modeling probability distributions over data, and that's a

436
00:17:09,529 --> 00:17:12,029
very important part of generative AI today.

437
00:17:12,809 --> 00:17:15,130
I see. And and so some of the

438
00:17:15,130 --> 00:17:15,630
chatter,

439
00:17:16,089 --> 00:17:17,789
you know, that I I came across

440
00:17:18,234 --> 00:17:18,974
on Tuesday

441
00:17:19,595 --> 00:17:20,494
about this,

442
00:17:21,115 --> 00:17:23,774
the this particular Nobel Prize is that

443
00:17:24,714 --> 00:17:27,434
not only is there lots of physics that's

444
00:17:27,434 --> 00:17:28,734
gone into the development

445
00:17:29,115 --> 00:17:30,255
of machine learning

446
00:17:30,794 --> 00:17:32,654
and, artificial intelligence,

447
00:17:34,140 --> 00:17:35,759
Now physicists are

448
00:17:36,059 --> 00:17:36,559
using

449
00:17:36,860 --> 00:17:38,400
those tools in,

450
00:17:38,779 --> 00:17:40,940
you know, I suppose, a a wide number

451
00:17:40,940 --> 00:17:42,160
of fields from,

452
00:17:42,700 --> 00:17:43,200
developing

453
00:17:43,900 --> 00:17:44,640
new materials

454
00:17:45,100 --> 00:17:47,039
to, analyzing data

455
00:17:47,420 --> 00:17:47,900
from,

456
00:17:48,625 --> 00:17:50,085
particle physics collisions.

457
00:17:50,464 --> 00:17:52,244
Can can you give us a little flavor

458
00:17:52,304 --> 00:17:54,565
of how physicists are using

459
00:17:54,865 --> 00:17:57,904
machine learning and AI in their day to

460
00:17:57,904 --> 00:17:58,644
day work?

461
00:17:59,585 --> 00:18:02,484
Yeah. I mean, we saw today the

462
00:18:03,029 --> 00:18:05,049
chemistry novel going to the

463
00:18:05,350 --> 00:18:07,910
people who figured out the protein folding problem

464
00:18:07,910 --> 00:18:10,970
using machine learning alpha fold, and that's one,

465
00:18:11,910 --> 00:18:13,450
you know, again, that'll be a

466
00:18:13,910 --> 00:18:14,890
point of contention.

467
00:18:16,005 --> 00:18:18,424
But, you know, protein folding has also been

468
00:18:19,525 --> 00:18:21,945
partly a physics problem, and that has been,

469
00:18:22,325 --> 00:18:24,184
you know, solved using machine learning.

470
00:18:25,285 --> 00:18:27,065
There have been numerous attempts

471
00:18:28,259 --> 00:18:31,160
in using machine learning to discover, for instance,

472
00:18:31,299 --> 00:18:34,660
new quantum optics experiments. So given all the

473
00:18:34,660 --> 00:18:35,160
tools

474
00:18:35,860 --> 00:18:36,360
that

475
00:18:36,820 --> 00:18:37,320
experimental,

476
00:18:38,100 --> 00:18:41,140
quantum optics physicist uses on their optical bench,

477
00:18:41,140 --> 00:18:44,115
you can use machine learning algorithms to design

478
00:18:44,894 --> 00:18:46,194
new circuits that

479
00:18:46,494 --> 00:18:49,694
humans may not have discovered because these algorithms

480
00:18:49,694 --> 00:18:52,174
are able to search a very, very large

481
00:18:52,174 --> 00:18:53,154
space of solutions,

482
00:18:53,615 --> 00:18:56,115
and they're optimized to find the best possible

483
00:18:56,174 --> 00:18:56,674
circuits.

484
00:18:58,179 --> 00:18:58,919
For instance,

485
00:18:59,380 --> 00:19:01,480
you know, it's already been shown that such,

486
00:19:02,099 --> 00:19:03,799
machine learning models can,

487
00:19:04,339 --> 00:19:06,839
find better ways to do entanglement swapping,

488
00:19:07,220 --> 00:19:09,380
which may not have been discovered where if

489
00:19:09,380 --> 00:19:11,319
it weren't for these machine learning models.

490
00:19:12,375 --> 00:19:14,615
There have been efforts on there are ongoing

491
00:19:14,615 --> 00:19:18,295
efforts, for instance, to discover symmetries that exist,

492
00:19:18,615 --> 00:19:21,095
in in data for, you know, data that

493
00:19:21,095 --> 00:19:23,115
might be coming out of the LHC.

494
00:19:23,654 --> 00:19:26,375
Of course, once, the machine learning models find

495
00:19:26,375 --> 00:19:26,875
symmetries,

496
00:19:27,720 --> 00:19:29,980
making sense of what these symmetries

497
00:19:30,440 --> 00:19:31,799
actually mean is a much,

498
00:19:32,279 --> 00:19:35,240
larger physics question, and, I don't think AI

499
00:19:35,240 --> 00:19:38,119
at this point, is capable of answering those

500
00:19:38,119 --> 00:19:38,619
questions.

501
00:19:39,894 --> 00:19:42,394
There are also efforts to speed up observation

502
00:19:42,535 --> 00:19:45,494
cosmology. You can you can train machine learning

503
00:19:45,494 --> 00:19:45,994
models

504
00:19:46,694 --> 00:19:47,194
using,

505
00:19:47,974 --> 00:19:48,954
pairs of,

506
00:19:50,134 --> 00:19:52,214
you know, data where on one side, on

507
00:19:52,214 --> 00:19:53,674
the input side are

508
00:19:55,299 --> 00:19:56,200
low resolution

509
00:19:56,740 --> 00:19:57,240
hydrodynamical

510
00:19:57,700 --> 00:20:00,119
simulations, and on the output side are

511
00:20:00,740 --> 00:20:03,779
high resolution versions of the same simulation. So

512
00:20:03,779 --> 00:20:05,640
you you create enough of those,

513
00:20:06,259 --> 00:20:08,119
training data. You train your

514
00:20:08,500 --> 00:20:10,359
machine learning models to

515
00:20:10,924 --> 00:20:12,704
correlate these low resolution,

516
00:20:13,724 --> 00:20:16,065
simulations with the high resolution simulations.

517
00:20:16,444 --> 00:20:18,765
And then once you've trained them, then anytime

518
00:20:18,765 --> 00:20:21,164
you need a new, high risk simulation, all

519
00:20:21,164 --> 00:20:22,859
you have to do is well, I'm saying

520
00:20:22,859 --> 00:20:24,220
all you have to do, but, you know,

521
00:20:24,220 --> 00:20:26,000
with caveats, but you basically,

522
00:20:27,420 --> 00:20:28,619
generate low res,

523
00:20:29,660 --> 00:20:32,059
simulations. And then in fractions of a second,

524
00:20:32,059 --> 00:20:34,000
the machine learning model will give you,

525
00:20:34,539 --> 00:20:36,559
the high res version of that simulation.

526
00:20:37,174 --> 00:20:38,534
If you had tried to do it in

527
00:20:38,534 --> 00:20:39,275
the normal,

528
00:20:39,815 --> 00:20:41,815
sort of computational way, that might have taken

529
00:20:41,815 --> 00:20:44,375
something like a 1000000 CPU hours. So there

530
00:20:44,375 --> 00:20:46,394
are many, many things that are happening,

531
00:20:46,855 --> 00:20:50,075
where machine learning is being is now influencing

532
00:20:50,214 --> 00:20:52,799
the way physics is done. Maybe, even in

533
00:20:52,799 --> 00:20:55,440
condensed matter physics, the the search or in

534
00:20:55,440 --> 00:20:57,299
material science, the search for,

535
00:20:58,159 --> 00:20:58,659
new

536
00:20:58,960 --> 00:20:59,460
materials.

537
00:21:00,240 --> 00:21:02,099
You know, it's very, very hard to,

538
00:21:03,039 --> 00:21:06,259
model create complex models of interatomic interactions.

539
00:21:06,634 --> 00:21:08,894
But if you can train machine learning models

540
00:21:09,035 --> 00:21:11,674
to learn the patterns and learn the interactions,

541
00:21:11,674 --> 00:21:14,015
and then they can be used to predict

542
00:21:14,394 --> 00:21:16,015
new materials with new properties,

543
00:21:16,875 --> 00:21:18,634
all of that stuff is happening. So this

544
00:21:18,634 --> 00:21:20,795
is a two way street. I mean, physics

545
00:21:20,795 --> 00:21:21,295
has

546
00:21:22,039 --> 00:21:22,940
very definitely

547
00:21:23,400 --> 00:21:23,900
influenced,

548
00:21:24,440 --> 00:21:26,119
machine learning in AI and,

549
00:21:26,759 --> 00:21:28,759
you know, and machine learning in AI, I

550
00:21:28,759 --> 00:21:30,460
think, is paying it back now.

551
00:21:31,480 --> 00:21:34,519
We should probably mention one more profound way

552
00:21:34,519 --> 00:21:35,980
in which physics has influenced

553
00:21:36,279 --> 00:21:37,065
AI, which

554
00:21:37,544 --> 00:21:39,404
most listeners would have,

555
00:21:40,504 --> 00:21:42,764
played around with these AIs and these are,

556
00:21:43,065 --> 00:21:44,744
you know, about a year and a half

557
00:21:44,744 --> 00:21:45,244
ago,

558
00:21:46,264 --> 00:21:47,724
we were abuzz with

559
00:21:48,345 --> 00:21:50,904
the image generation models that came out of,

560
00:21:51,464 --> 00:21:52,365
these companies.

561
00:21:53,009 --> 00:21:55,590
For instance, DALL E or Stable Diffusion,

562
00:21:56,130 --> 00:21:57,269
well, those use

563
00:21:57,650 --> 00:22:00,150
very strongly used principles from non equilibrium,

564
00:22:00,769 --> 00:22:01,269
thermodynamics,

565
00:22:01,570 --> 00:22:03,830
and and it would not be possible without

566
00:22:04,049 --> 00:22:05,670
a heavy influence from physics.

567
00:22:06,045 --> 00:22:07,565
I mean, one thing that that I really

568
00:22:07,565 --> 00:22:09,404
find amazing, Anil, is that,

569
00:22:10,285 --> 00:22:13,025
this year, we've got one Nobel Prize

570
00:22:13,325 --> 00:22:14,305
for the development

571
00:22:15,085 --> 00:22:16,225
of a new technology.

572
00:22:17,005 --> 00:22:18,625
And then the next day,

573
00:22:19,005 --> 00:22:21,345
another Nobel Prize goes to

574
00:22:21,670 --> 00:22:22,970
somebody who's used

575
00:22:23,349 --> 00:22:24,170
that technology

576
00:22:24,789 --> 00:22:26,970
to solve a a really, really

577
00:22:27,430 --> 00:22:27,930
important

578
00:22:28,470 --> 00:22:28,970
scientific

579
00:22:29,269 --> 00:22:30,650
problem. I mean, this

580
00:22:31,190 --> 00:22:33,289
this field seems to be moving

581
00:22:33,924 --> 00:22:36,825
so quickly. Is is that something that attracted

582
00:22:36,884 --> 00:22:37,384
you

583
00:22:37,684 --> 00:22:40,964
to, machine learning and AI when you decided

584
00:22:40,964 --> 00:22:42,105
to write your book?

585
00:22:43,525 --> 00:22:46,164
To be honest, I started writing the book

586
00:22:46,164 --> 00:22:47,960
before this hype,

587
00:22:48,500 --> 00:22:49,480
became apparent,

588
00:22:49,779 --> 00:22:51,059
and I don't want to call it hype

589
00:22:51,059 --> 00:22:53,140
in a negative way. Let's you know, there's

590
00:22:53,140 --> 00:22:54,819
so much that has happened in the last

591
00:22:54,819 --> 00:22:57,720
2 years. I started writing this in 2020

592
00:22:57,779 --> 00:22:58,919
when all of this was

593
00:22:59,299 --> 00:23:01,859
completely not on my horizon. I was just

594
00:23:01,859 --> 00:23:02,359
purely

595
00:23:03,295 --> 00:23:04,355
fascinated by,

596
00:23:05,134 --> 00:23:07,455
the mathematics of machine learning. And I really

597
00:23:07,455 --> 00:23:08,255
felt like,

598
00:23:08,734 --> 00:23:10,575
you know, that there's a story to be

599
00:23:10,575 --> 00:23:13,555
told about all of the mathematics that underpins

600
00:23:13,775 --> 00:23:14,755
modern AI.

601
00:23:15,309 --> 00:23:17,390
And as it happened about halfway through the

602
00:23:17,390 --> 00:23:19,170
writing of the book, chat GPT

603
00:23:19,470 --> 00:23:21,950
happened and DALL E and stable diffusion happened

604
00:23:21,950 --> 00:23:25,090
and, you know, things just changed. And the

605
00:23:25,309 --> 00:23:27,390
alpha 4, alpha go, all of these things

606
00:23:27,390 --> 00:23:28,289
have been happening.

607
00:23:28,954 --> 00:23:30,714
And, you know, the last 4 years have

608
00:23:30,714 --> 00:23:31,934
been just insanely,

609
00:23:33,115 --> 00:23:33,855
fast moving.

610
00:23:34,795 --> 00:23:36,634
And I I my suspicion is that we

611
00:23:36,634 --> 00:23:38,654
are still at the very beginnings of this.

612
00:23:40,315 --> 00:23:42,315
I see. And and was it difficult when

613
00:23:42,315 --> 00:23:43,529
you're writing the book?

614
00:23:44,090 --> 00:23:47,150
You know, I'm guessing you you probably learned

615
00:23:47,210 --> 00:23:47,950
new things,

616
00:23:48,410 --> 00:23:50,349
new and exciting things about AI

617
00:23:50,890 --> 00:23:52,590
almost on a daily basis.

618
00:23:53,289 --> 00:23:54,509
Were were you constantly

619
00:23:54,809 --> 00:23:55,309
rewriting

620
00:23:55,964 --> 00:23:58,144
your book or adding new chapters?

621
00:23:58,845 --> 00:24:00,365
I mean, are are you gonna be be

622
00:24:00,365 --> 00:24:02,605
adding a a a chapter now about,

623
00:24:04,125 --> 00:24:07,025
the structure of, proteins, for example?

624
00:24:08,140 --> 00:24:08,380
Well,

625
00:24:09,099 --> 00:24:11,500
I kinda got lucky in the sense that

626
00:24:11,500 --> 00:24:15,440
my focus was on the basic mathematical principles

627
00:24:15,660 --> 00:24:16,480
that underlie,

628
00:24:17,099 --> 00:24:19,599
machine learning. And to be honest,

629
00:24:20,140 --> 00:24:20,960
that that mathematics,

630
00:24:22,714 --> 00:24:24,555
for me at least, in my book, more

631
00:24:24,555 --> 00:24:28,174
or less stops in in the 19 nineties.

632
00:24:28,714 --> 00:24:29,214
So,

633
00:24:29,674 --> 00:24:32,075
it's kind of a historical account of the

634
00:24:32,075 --> 00:24:34,394
math you need to understand what's happening today,

635
00:24:34,394 --> 00:24:35,214
and and

636
00:24:35,515 --> 00:24:37,980
and so that stuff doesn't change.

637
00:24:38,519 --> 00:24:41,339
For instance, we were talking earlier of how

638
00:24:41,880 --> 00:24:44,200
Hinton and Rumelhart and others figured out the

639
00:24:44,200 --> 00:24:46,859
back propagation algorithm. So that's pretty much

640
00:24:47,799 --> 00:24:49,880
the ending of my book. I come to

641
00:24:49,880 --> 00:24:52,705
that, you know, the ability to train large

642
00:24:52,845 --> 00:24:53,984
deep neural networks

643
00:24:54,365 --> 00:24:55,585
is based on this

644
00:24:56,045 --> 00:24:58,384
one single back propagation algorithm.

645
00:24:58,684 --> 00:25:00,605
And it doesn't matter how large the network

646
00:25:00,605 --> 00:25:02,785
is, you know, you could you could train

647
00:25:03,085 --> 00:25:05,805
a network with 1 hidden layer or a

648
00:25:05,805 --> 00:25:08,019
100 hidden layers. The algorithm is same. So

649
00:25:08,019 --> 00:25:10,180
if you understand that, you understand what's happening

650
00:25:10,180 --> 00:25:10,680
today.

651
00:25:12,500 --> 00:25:14,660
I might have to add a chapter or

652
00:25:14,660 --> 00:25:17,539
2 at some point, but, during the course

653
00:25:17,539 --> 00:25:19,400
of the writing of it, I wasn't concerned

654
00:25:19,460 --> 00:25:22,440
with the the rapid changes because

655
00:25:23,005 --> 00:25:24,924
my focus was on being able to tell

656
00:25:24,924 --> 00:25:27,184
the reader that, okay, here are some really

657
00:25:27,804 --> 00:25:28,304
simple,

658
00:25:28,924 --> 00:25:29,424
interesting,

659
00:25:30,044 --> 00:25:33,005
and honestly, in my mind, subjectively speaking, very

660
00:25:33,005 --> 00:25:36,464
elegant mathematics that underpins machine learning.

661
00:25:37,390 --> 00:25:40,349
I see. And so, you you mentioned extra

662
00:25:40,349 --> 00:25:41,950
chapters. I mean, it sounds to me like

663
00:25:41,950 --> 00:25:43,309
you could write,

664
00:25:44,029 --> 00:25:45,329
another book about

665
00:25:45,950 --> 00:25:47,089
exciting applications

666
00:25:47,390 --> 00:25:47,710
of,

667
00:25:48,509 --> 00:25:51,089
artificial intelligence. Do do you have any plans

668
00:25:51,335 --> 00:25:51,734
to do that, or,

669
00:25:52,455 --> 00:25:54,714
do do do you have other interests

670
00:25:55,095 --> 00:25:55,755
these days?

671
00:25:57,174 --> 00:25:58,075
I possibly,

672
00:25:58,775 --> 00:26:00,634
if I were to write something,

673
00:26:01,494 --> 00:26:03,734
it probably won't be about applications. I think

674
00:26:03,734 --> 00:26:05,835
applications, honestly, for me personally,

675
00:26:07,015 --> 00:26:09,710
I I don't find them, that exciting in

676
00:26:09,710 --> 00:26:10,609
terms of storytelling.

677
00:26:11,789 --> 00:26:13,329
So I I I gravitate

678
00:26:13,710 --> 00:26:14,929
much more to,

679
00:26:15,390 --> 00:26:16,049
the basic,

680
00:26:16,509 --> 00:26:18,450
science side of things or the fundamentals.

681
00:26:18,990 --> 00:26:21,804
So I'm still looking for another idea. Some

682
00:26:21,964 --> 00:26:23,724
there are some thoughts in my head, but

683
00:26:23,724 --> 00:26:25,744
nothing has really clarified yet.

684
00:26:26,845 --> 00:26:30,204
I see. Okay. Well, thanks, Anil. Thanks so

685
00:26:30,204 --> 00:26:32,464
much for coming on to the podcast

686
00:26:32,765 --> 00:26:36,119
and, and sharing your thoughts about this year's

687
00:26:36,419 --> 00:26:38,039
Nobel Prize for Physics.

688
00:26:38,419 --> 00:26:39,559
And you can read

689
00:26:39,859 --> 00:26:42,099
much more about the prize on the Physics

690
00:26:42,099 --> 00:26:43,000
World website.

691
00:26:43,859 --> 00:26:46,599
And on the website, you can also find

692
00:26:46,659 --> 00:26:48,839
a review of Anil's book.

693
00:26:49,220 --> 00:26:50,839
Thanks for being on the podcast.

694
00:26:52,234 --> 00:26:54,075
Hamish, thanks you very thank you very much.

695
00:26:54,075 --> 00:26:55,214
This has been a pleasure.

696
00:27:02,075 --> 00:27:04,654
This episode was supported by SmartAct,

697
00:27:05,500 --> 00:27:08,319
empowering breakthroughs in science and technology

698
00:27:08,859 --> 00:27:10,640
with high precision positioning,

699
00:27:11,099 --> 00:27:11,599
metrology,

700
00:27:12,140 --> 00:27:13,839
and automation solutions.

701
00:27:14,460 --> 00:27:14,960
Visit

702
00:27:15,339 --> 00:27:15,839
smartact.com

703
00:27:17,259 --> 00:27:18,240
to learn more.