AI and the future of physics

1
00:00:01,280 --> 00:00:03,679
Hello, and welcome to the Physics World Stories

2
00:00:03,679 --> 00:00:06,339
podcast. I'm Andrew Glester. And in this episode,

3
00:00:06,559 --> 00:00:09,300
we're diving into a topic that is revolutionizing

4
00:00:10,000 --> 00:00:11,059
physics itself,

5
00:00:11,679 --> 00:00:12,179
artificial

6
00:00:12,559 --> 00:00:13,059
intelligence.

7
00:00:15,724 --> 00:00:18,364
AI is transforming our world and becoming part

8
00:00:18,364 --> 00:00:20,445
of our day to day, but what about

9
00:00:20,445 --> 00:00:23,324
physics? What happens when we apply machine learning

10
00:00:23,324 --> 00:00:26,045
to the fundamental questions of our universe, and

11
00:00:26,045 --> 00:00:29,670
how is AI accelerating discoveries in particle physics?

12
00:00:30,149 --> 00:00:32,789
Can it help us visualize the unseen through

13
00:00:32,789 --> 00:00:33,289
photography?

14
00:00:33,829 --> 00:00:35,590
Well, and what does the future look like

15
00:00:35,590 --> 00:00:38,090
as AI and physics evolve together?

16
00:00:39,350 --> 00:00:41,909
This episode is inspired by the Institute of

17
00:00:41,909 --> 00:00:45,085
Physics latest white paper on AI in physics,

18
00:00:45,145 --> 00:00:46,284
exploring the challenges,

19
00:00:46,664 --> 00:00:47,164
opportunities,

20
00:00:47,625 --> 00:00:50,505
and ethical questions that come alongside it. We'll

21
00:00:50,505 --> 00:00:53,085
hear from a science photographer at MIT

22
00:00:53,545 --> 00:00:55,405
about AI in her work,

23
00:00:55,719 --> 00:00:57,880
and we'll hear from two people who have

24
00:00:57,880 --> 00:01:00,920
been advising on that white paper, including a

25
00:01:00,920 --> 00:01:01,899
particle physicist

26
00:01:02,520 --> 00:01:04,939
at CERN and the University of Manchester.

27
00:01:05,719 --> 00:01:08,620
But first, Toni Hay formerly of Microsoft,

28
00:01:09,325 --> 00:01:11,905
has spent years at the intersection of computers

29
00:01:12,204 --> 00:01:13,504
and scientific research.

30
00:01:13,965 --> 00:01:16,064
He's one of the consultants on the IOB's

31
00:01:16,125 --> 00:01:18,765
white paper, and he used to be a

32
00:01:18,765 --> 00:01:21,245
particle physicist. I wondered what had made him

33
00:01:21,245 --> 00:01:25,019
move from that into computing. Theoretical particle physics

34
00:01:25,819 --> 00:01:26,560
had essentially

35
00:01:28,939 --> 00:01:31,179
come to a a a a lot of

36
00:01:31,179 --> 00:01:34,619
great crescendo with the gauge theories of quantum

37
00:01:34,619 --> 00:01:37,685
chromodynamics and the standard model. And and you

38
00:01:37,685 --> 00:01:39,685
could read my book, my fifth edition of

39
00:01:39,685 --> 00:01:42,165
my book, gauge series in particle physics, which

40
00:01:42,165 --> 00:01:44,024
is, if you like, an epitaph

41
00:01:45,284 --> 00:01:47,204
of my career in particle physics. So I

42
00:01:47,204 --> 00:01:48,805
was very fortunate to be there at the

43
00:01:48,805 --> 00:01:51,430
critical time. But, yes, I didn't want to

44
00:01:51,430 --> 00:01:52,490
spend my time

45
00:01:53,349 --> 00:01:54,650
doing the same things,

46
00:01:55,430 --> 00:01:58,250
and doing things which are largely irrelevant

47
00:01:58,549 --> 00:02:01,670
to people. Alright? Most people don't care about

48
00:02:01,670 --> 00:02:04,090
particle physics. Most people don't care about,

49
00:02:04,495 --> 00:02:06,594
I would say, they like the pictures from

50
00:02:07,055 --> 00:02:10,354
astrophysics and astronomy, but but, actually, you know,

51
00:02:10,814 --> 00:02:13,694
I'm not sure whether they worry about the

52
00:02:13,694 --> 00:02:15,074
the big bang and the inflationary

53
00:02:15,375 --> 00:02:16,835
period that followed and

54
00:02:17,419 --> 00:02:20,060
the background radiation and stuff like this. So,

55
00:02:20,060 --> 00:02:21,900
I mean, I still find them interesting and

56
00:02:21,900 --> 00:02:23,659
they are fun and and things, but I

57
00:02:23,659 --> 00:02:25,360
do feel that particle physics

58
00:02:26,860 --> 00:02:28,860
evolved into sort of a dead end in

59
00:02:28,860 --> 00:02:29,360
that

60
00:02:30,620 --> 00:02:31,520
it went to

61
00:02:32,125 --> 00:02:34,224
string theory, which is fine,

62
00:02:35,084 --> 00:02:35,584
but

63
00:02:36,925 --> 00:02:38,844
it's sort of metaphysics in that it doesn't

64
00:02:38,844 --> 00:02:40,944
make any predictions that you can check,

65
00:02:41,805 --> 00:02:42,305
and

66
00:02:42,604 --> 00:02:44,844
has excitations at the the Planck level, which

67
00:02:44,844 --> 00:02:46,139
we can never get to.

68
00:02:46,579 --> 00:02:48,379
Oh, and also all the particles I spent

69
00:02:48,379 --> 00:02:50,800
my life caring about, the proton, the pion,

70
00:02:50,860 --> 00:02:51,840
and electron,

71
00:02:53,020 --> 00:02:54,080
they're all zero

72
00:02:54,459 --> 00:02:55,599
mass approximations

73
00:02:56,539 --> 00:02:58,539
because compared to the Planck mass, everything is

74
00:02:58,539 --> 00:03:01,084
zero. So they're there's symmetry breaking effects which

75
00:03:01,084 --> 00:03:02,364
you may get run. Oh, and it's in

76
00:03:02,364 --> 00:03:05,084
the wrong number of dimensions, but maybe maybe

77
00:03:05,084 --> 00:03:06,604
some of them curl up and we get

78
00:03:06,604 --> 00:03:07,504
to four dimensions.

79
00:03:08,444 --> 00:03:09,425
And and really,

80
00:03:09,965 --> 00:03:11,965
it does seem to have evolved into into

81
00:03:11,965 --> 00:03:12,705
sort of

82
00:03:13,210 --> 00:03:16,110
a love of mathematics. And it's very fascinating,

83
00:03:16,169 --> 00:03:18,030
and it's done some wonderful things

84
00:03:18,330 --> 00:03:20,729
in mathematics, but it doesn't have any impact

85
00:03:20,729 --> 00:03:23,610
in particle physics. So I I do I

86
00:03:23,610 --> 00:03:25,069
did feel, yes, increasingly,

87
00:03:25,784 --> 00:03:28,205
particle physics was becoming divorced

88
00:03:28,504 --> 00:03:29,004
from

89
00:03:29,944 --> 00:03:30,685
real life,

90
00:03:31,145 --> 00:03:31,645
and

91
00:03:32,905 --> 00:03:35,064
I got more interested in using computers to

92
00:03:35,064 --> 00:03:36,844
solve them. And then parallel computers,

93
00:03:37,944 --> 00:03:39,405
I was extremely fortunate

94
00:03:41,650 --> 00:03:43,810
to be able to build my team built

95
00:03:43,810 --> 00:03:44,550
and designed

96
00:03:44,930 --> 00:03:45,990
designed and built

97
00:03:46,610 --> 00:03:48,770
a a parallel computer to do physics to

98
00:03:48,770 --> 00:03:51,490
do originally physics, but now I realized there

99
00:03:51,490 --> 00:03:53,010
are other things you can do with it

100
00:03:53,010 --> 00:03:54,949
than physics. And for example,

101
00:03:55,425 --> 00:03:57,685
I in doing e e science,

102
00:03:58,305 --> 00:04:00,465
I really believe things like climate change are

103
00:04:00,465 --> 00:04:03,284
rather more important than particle physics and astronomy.

104
00:04:03,745 --> 00:04:04,245
And,

105
00:04:05,025 --> 00:04:06,645
again, I'm not sure AI

106
00:04:07,185 --> 00:04:09,664
can do a huge amount to accelerate that,

107
00:04:09,664 --> 00:04:10,980
but I think particle

108
00:04:11,280 --> 00:04:12,900
particle physics certainly can't.

109
00:04:13,360 --> 00:04:15,520
So, no, I I welcome the physicists. They're

110
00:04:15,520 --> 00:04:18,100
a wonderful community. They do lots of things,

111
00:04:18,480 --> 00:04:19,120
but but,

112
00:04:20,160 --> 00:04:22,080
I think they're not quite as critical as

113
00:04:22,080 --> 00:04:23,060
they think they are.

114
00:04:24,319 --> 00:04:26,295
Sorry to all our listeners. I mean, that's,

115
00:04:26,295 --> 00:04:28,615
yeah, that's the That's that's that's that's the

116
00:04:28,615 --> 00:04:29,975
way it is. I'm a I I'm a

117
00:04:29,975 --> 00:04:32,314
physicist. I'm giving a talk next week about

118
00:04:32,455 --> 00:04:36,074
Bell's theorem and demonstrating how Bell's theorem demonstrates,

119
00:04:36,535 --> 00:04:39,899
you know, that Einstein's hidden variables were wrong.

120
00:04:40,139 --> 00:04:43,019
And, I still like that, and I'm talking

121
00:04:43,019 --> 00:04:45,660
about difference between John Bell and Einstein and

122
00:04:45,660 --> 00:04:48,699
Bohr. Einstein and Bohr considered correlations at naught

123
00:04:48,699 --> 00:04:51,019
and 90. John Bell, as he used to

124
00:04:51,019 --> 00:04:54,495
delight in saying, because he's Irish, considered correlations

125
00:04:54,634 --> 00:04:57,835
at 37 degrees. Alright? And, then you can

126
00:04:57,835 --> 00:05:01,035
tell the difference. Right? And so, I no.

127
00:05:01,035 --> 00:05:02,495
I I I still think

128
00:05:02,955 --> 00:05:05,055
physics is a wonderful area, and

129
00:05:05,790 --> 00:05:07,490
my hero is still,

130
00:05:07,949 --> 00:05:09,569
Richard Feynman, who

131
00:05:09,949 --> 00:05:12,930
who his wonderful lectures on physics, his wonderful

132
00:05:12,990 --> 00:05:15,389
Cornell lectures he gave, and stuff like that.

133
00:05:15,389 --> 00:05:16,050
And so

134
00:05:16,509 --> 00:05:17,810
and I I worked

135
00:05:18,669 --> 00:05:19,250
with Feynman

136
00:05:20,355 --> 00:05:22,055
to help write up his lectures.

137
00:05:22,435 --> 00:05:24,194
He lectured for the last five years of

138
00:05:24,194 --> 00:05:25,814
his life on computing.

139
00:05:26,275 --> 00:05:28,935
You can find the Feynman lectures on computation

140
00:05:29,154 --> 00:05:31,495
edited by me. You can find them

141
00:05:31,875 --> 00:05:33,175
in a book form, and

142
00:05:33,875 --> 00:05:34,694
they're all about

143
00:05:35,029 --> 00:05:38,649
interesting things and, you know, universality and Turing's

144
00:05:38,709 --> 00:05:39,769
theorem and

145
00:05:40,550 --> 00:05:42,550
and these sort of things. I take it

146
00:05:42,550 --> 00:05:44,870
there's nothing in there about the possibility of

147
00:05:44,870 --> 00:05:45,370
AI?

148
00:05:45,750 --> 00:05:47,689
Yes. Feynman cared about AI.

149
00:05:49,024 --> 00:05:50,865
He called them he didn't like the name

150
00:05:50,865 --> 00:05:53,345
AI. He called them advanced applications, but he

151
00:05:53,345 --> 00:05:54,964
understood, you know, image,

152
00:05:57,504 --> 00:05:59,024
vision, computer vision,

153
00:05:59,345 --> 00:06:01,584
robotics, and things. Yes. He he cared very

154
00:06:01,584 --> 00:06:04,860
much. He had collaborators, one of which on

155
00:06:04,860 --> 00:06:06,699
his original version of the course, he gave

156
00:06:06,699 --> 00:06:09,519
it with two collaborators. One was Carver Mead,

157
00:06:09,660 --> 00:06:11,579
one of the guys who explained why Moore's

158
00:06:11,579 --> 00:06:12,319
Law worked,

159
00:06:12,939 --> 00:06:13,439
and

160
00:06:13,819 --> 00:06:16,539
one was John Hopfield, who won the Nobel

161
00:06:16,539 --> 00:06:19,185
Prize. And he he had a a specific

162
00:06:19,185 --> 00:06:21,584
type of neural network, which is not the

163
00:06:21,584 --> 00:06:24,144
same type as neural network as everybody uses

164
00:06:24,144 --> 00:06:26,225
now. Alright? But but he did win the

165
00:06:26,225 --> 00:06:27,204
Nobel Prize,

166
00:06:30,704 --> 00:06:33,009
and he was originally a physicist.

167
00:06:34,350 --> 00:06:36,430
Hartfield was a physicist. Hinton, on the other

168
00:06:36,430 --> 00:06:37,870
hand, was never a physicist. So,

169
00:06:40,029 --> 00:06:41,569
my computer science friends,

170
00:06:42,349 --> 00:06:43,329
I used to be

171
00:06:43,629 --> 00:06:45,629
have a colleague of famous guy called Jim

172
00:06:45,629 --> 00:06:47,625
Gray, and he was very annoyed about the

173
00:06:47,625 --> 00:06:49,324
physicist claiming engineers

174
00:06:50,425 --> 00:06:53,544
for the guy who invented integrated circuits they

175
00:06:53,544 --> 00:06:55,464
gave a Nobel Prize to. And he was

176
00:06:55,464 --> 00:06:57,704
really, really angry that that he wasn't a

177
00:06:57,704 --> 00:06:59,784
physicist. He was an engineer, and he would

178
00:06:59,784 --> 00:07:01,019
have said the same about

179
00:07:01,979 --> 00:07:02,479
Hinton

180
00:07:03,099 --> 00:07:03,759
and Hopfield,

181
00:07:04,060 --> 00:07:06,220
that they were not physicists. So getting a

182
00:07:06,220 --> 00:07:07,519
Nobel Prize for physics

183
00:07:07,979 --> 00:07:09,899
was a little strange and must have certainly

184
00:07:09,899 --> 00:07:12,479
ruffled a few feathers in the physics community.

185
00:07:13,175 --> 00:07:13,995
And so to

186
00:07:14,454 --> 00:07:16,714
come to the the IOP

187
00:07:19,334 --> 00:07:21,735
document, it's it's a wonderful document, and it's

188
00:07:21,735 --> 00:07:23,194
and it's right to engage,

189
00:07:24,295 --> 00:07:26,694
the attention of physicists because it will become

190
00:07:26,694 --> 00:07:29,189
part of their working life. That's true.

191
00:07:29,490 --> 00:07:31,569
I I just rather doubt when they say,

192
00:07:31,569 --> 00:07:34,050
you know, large numbers of them know all

193
00:07:34,050 --> 00:07:34,790
about it,

194
00:07:35,170 --> 00:07:37,250
that they do know all about it because,

195
00:07:37,250 --> 00:07:39,330
actually, it requires quite a lot of effort

196
00:07:39,330 --> 00:07:39,990
and investment

197
00:07:40,544 --> 00:07:42,164
in computer science technologies,

198
00:07:42,544 --> 00:07:44,004
and I don't think most

199
00:07:44,625 --> 00:07:46,164
particle physicists do that.

200
00:07:46,625 --> 00:07:49,024
It is true, however, that particle physicists have

201
00:07:49,024 --> 00:07:49,524
used,

202
00:07:50,225 --> 00:07:52,464
you know, things like neural networks and other

203
00:07:52,464 --> 00:07:55,500
types of algorithms, which are part of the

204
00:07:55,500 --> 00:07:56,000
general

205
00:07:56,379 --> 00:07:59,580
AI. But what what drives AI now are

206
00:07:59,580 --> 00:08:02,300
purely these deep neural networks and these now

207
00:08:02,300 --> 00:08:02,800
now

208
00:08:03,259 --> 00:08:05,759
the version of them that these transformer networks

209
00:08:05,944 --> 00:08:08,345
that can do these large language models. From

210
00:08:08,345 --> 00:08:11,404
your perspective looking at at AI in society,

211
00:08:12,104 --> 00:08:14,584
not necessarily in physics, but also in physics,

212
00:08:14,584 --> 00:08:15,564
are there sort of

213
00:08:15,944 --> 00:08:16,444
misconceptions

214
00:08:16,904 --> 00:08:18,524
that you can see about it that

215
00:08:19,029 --> 00:08:20,490
that that are out there, or

216
00:08:20,790 --> 00:08:23,209
are we understanding it as it really is?

217
00:08:23,589 --> 00:08:24,089
Well,

218
00:08:24,790 --> 00:08:26,970
when I came back in 2015

219
00:08:27,990 --> 00:08:30,410
to the Rutherford Lab, I was amazed

220
00:08:31,145 --> 00:08:33,245
that nobody I could find understood

221
00:08:34,665 --> 00:08:35,165
that,

222
00:08:35,945 --> 00:08:37,565
what you needed to do

223
00:08:38,024 --> 00:08:38,524
AI

224
00:08:39,865 --> 00:08:41,085
post 2012

225
00:08:41,225 --> 00:08:44,009
was large amounts of computing power. And

226
00:08:44,309 --> 00:08:47,589
GPUs like Nvidia's GPUs were one solution. And

227
00:08:47,589 --> 00:08:49,029
that was one of the things that was

228
00:08:49,029 --> 00:08:51,350
found out that actually what made the difference

229
00:08:51,350 --> 00:08:53,829
was the scale of the data and the

230
00:08:53,829 --> 00:08:54,970
scale of the computing.

231
00:08:55,350 --> 00:08:58,065
And everybody here were doing AI includes all

232
00:08:58,065 --> 00:09:00,304
sorts of little algorithms, and they're all mentioned

233
00:09:00,304 --> 00:09:01,044
in the report.

234
00:09:01,424 --> 00:09:04,084
And that they were, if you like, pre

235
00:09:04,304 --> 00:09:05,125
deep learning.

236
00:09:05,504 --> 00:09:07,985
And people were thinking, that's fine. I I

237
00:09:07,985 --> 00:09:09,504
do AI, and I can get money from

238
00:09:09,504 --> 00:09:11,605
the government to do no. The whole purpose

239
00:09:11,789 --> 00:09:13,250
was actually to transform,

240
00:09:14,350 --> 00:09:17,070
and the government putting AI in there money

241
00:09:17,070 --> 00:09:20,289
into AI was because of deep learning. And

242
00:09:20,990 --> 00:09:23,789
now these these transform models and large language

243
00:09:23,789 --> 00:09:26,574
models, and I couldn't find anybody,

244
00:09:27,194 --> 00:09:28,254
who was interested.

245
00:09:28,634 --> 00:09:31,054
And so when I I got a grant,

246
00:09:32,394 --> 00:09:34,794
on AI for science, I called it, I

247
00:09:34,794 --> 00:09:36,574
called it AI for science deliberately

248
00:09:37,209 --> 00:09:39,690
because I knew the politicians didn't know about

249
00:09:39,690 --> 00:09:40,509
machine learning

250
00:09:41,049 --> 00:09:42,889
for for science, which is what it was.

251
00:09:42,889 --> 00:09:43,389
Alright?

252
00:09:44,250 --> 00:09:46,909
Machine learning was a nuance they wouldn't appreciate.

253
00:09:47,610 --> 00:09:48,110
And

254
00:09:49,209 --> 00:09:50,509
I managed to persuade,

255
00:09:51,504 --> 00:09:53,164
it was a collaboration with Turing,

256
00:09:54,345 --> 00:09:56,585
and I managed to persuade them to give

257
00:09:56,585 --> 00:09:57,725
me some

258
00:09:58,105 --> 00:10:00,985
GPU computing power so we could Rutherford will

259
00:10:00,985 --> 00:10:03,884
have offer a GPU AI computing service

260
00:10:04,264 --> 00:10:04,845
to Turing

261
00:10:05,759 --> 00:10:06,259
participants.

262
00:10:07,200 --> 00:10:08,340
And that, I think,

263
00:10:09,600 --> 00:10:10,500
is a challenge

264
00:10:10,960 --> 00:10:14,019
still. And, you know, the question is,

265
00:10:14,639 --> 00:10:17,040
you know, Microsoft and and and others, my

266
00:10:17,040 --> 00:10:19,975
old company, Microsoft is is building data centers

267
00:10:19,975 --> 00:10:22,315
at a at a vast rate, which involves

268
00:10:22,534 --> 00:10:24,315
huge amounts of computing power

269
00:10:24,774 --> 00:10:25,274
and,

270
00:10:26,054 --> 00:10:27,894
cost billions of dollars. And there's no way

271
00:10:27,894 --> 00:10:31,174
that Europe can emulate that because we don't

272
00:10:31,174 --> 00:10:33,450
have any, what are these companies are called,

273
00:10:33,450 --> 00:10:35,309
super scalar companies like,

274
00:10:35,769 --> 00:10:37,610
Amazon, Microsoft, Meta,

275
00:10:38,730 --> 00:10:40,350
Amazon, Microsoft, Meta,

276
00:10:42,330 --> 00:10:45,565
Google. Yes. That's right. So and possibly Apple.

277
00:10:46,264 --> 00:10:47,865
They're the in The US, they're the only

278
00:10:47,865 --> 00:10:50,264
one. China has some sums that could similarly

279
00:10:50,264 --> 00:10:53,464
do that. We don't have companies in Europe

280
00:10:53,464 --> 00:10:56,184
that can put billions a month into building

281
00:10:56,184 --> 00:10:57,644
these centers. And so

282
00:10:58,820 --> 00:11:01,460
Europe's trying to do something, but it it's

283
00:11:01,460 --> 00:11:03,000
it's a complicated business.

284
00:11:03,700 --> 00:11:04,200
And

285
00:11:05,300 --> 00:11:05,960
I think

286
00:11:06,580 --> 00:11:08,820
given the the promise of AI, I think

287
00:11:08,820 --> 00:11:11,139
there will be some national resources that you

288
00:11:11,139 --> 00:11:13,414
can do these things with, but I think

289
00:11:13,414 --> 00:11:14,074
it's complicated.

290
00:11:14,855 --> 00:11:16,534
And, of course, then you have this result

291
00:11:16,534 --> 00:11:18,454
from China. Deep Seek says, oh, you don't

292
00:11:18,454 --> 00:11:21,334
need large amounts of computing power. I'm slightly

293
00:11:21,334 --> 00:11:23,334
skeptical of that, but we we'll wait and

294
00:11:23,334 --> 00:11:25,174
see what happens on there. Yeah. I was

295
00:11:25,174 --> 00:11:27,174
gonna ask about that. You are skeptical of

296
00:11:27,174 --> 00:11:29,339
it, are you? Slightly. Yes. I am. I

297
00:11:29,339 --> 00:11:32,699
suspect that they've actually learned from what The

298
00:11:32,699 --> 00:11:34,799
US companies and there are open solutions,

299
00:11:35,500 --> 00:11:37,039
what they've done. And,

300
00:11:39,339 --> 00:11:41,295
yes, I think that you will still need

301
00:11:41,295 --> 00:11:43,215
large amounts of computing power. And the Chinese

302
00:11:43,215 --> 00:11:44,595
have that. Right? So,

303
00:11:45,855 --> 00:11:47,955
but but I don't know how open

304
00:11:48,975 --> 00:11:50,894
what they've done and what they copied and

305
00:11:50,894 --> 00:11:52,815
what they haven't because, you know, it's difficult

306
00:11:52,815 --> 00:11:53,394
to tell.

307
00:11:54,095 --> 00:11:55,855
As you mentioned earlier, you have an interest

308
00:11:55,855 --> 00:11:57,850
in climate science.

309
00:11:58,149 --> 00:11:58,889
Right. I think

310
00:11:59,269 --> 00:12:01,690
everybody should have. Right? Everybody.

311
00:12:02,389 --> 00:12:05,129
Yeah. No. Absolutely. But there's, you know,

312
00:12:05,590 --> 00:12:08,149
large amounts of computing power is a large

313
00:12:08,149 --> 00:12:09,850
impact on the climate. Right?

314
00:12:11,035 --> 00:12:13,295
Yes. I'm told. I haven't checked this figure

315
00:12:13,435 --> 00:12:15,754
that the the amount is is less than

316
00:12:15,754 --> 00:12:17,375
used for Bitcoin mining.

317
00:12:18,154 --> 00:12:19,215
Right. Okay.

318
00:12:19,595 --> 00:12:21,595
I haven't checked that, but but but it

319
00:12:21,595 --> 00:12:23,215
seems to me a plausible thing,

320
00:12:24,075 --> 00:12:25,295
especially as Bitcoin

321
00:12:26,059 --> 00:12:28,399
and and its variants are now being used,

322
00:12:28,940 --> 00:12:30,860
and they have they use a large amount

323
00:12:30,860 --> 00:12:33,659
of computing power. But but no. That doesn't

324
00:12:33,659 --> 00:12:34,960
worry me so much.

325
00:12:36,299 --> 00:12:38,059
I think they'll come a natural end that

326
00:12:38,059 --> 00:12:39,120
you won't actually

327
00:12:40,700 --> 00:12:43,315
no. Question is how much training data do

328
00:12:43,315 --> 00:12:44,934
you need? Right? And

329
00:12:46,034 --> 00:12:49,254
one of the projects I'm actually interested in

330
00:12:49,875 --> 00:12:52,434
is related to the Institute of Physics thing

331
00:12:52,434 --> 00:12:53,174
is that

332
00:12:54,434 --> 00:12:56,034
what you can do now, you can pick

333
00:12:56,034 --> 00:12:56,759
up a model

334
00:12:57,879 --> 00:12:59,580
that's being trained by ChatGPT

335
00:12:59,960 --> 00:13:02,700
and and and OpenAI and things like that

336
00:13:03,000 --> 00:13:04,299
and the other companies.

337
00:13:05,399 --> 00:13:07,240
And then you can specialize it to your

338
00:13:07,240 --> 00:13:09,044
domain, but you don't know what the thing

339
00:13:09,125 --> 00:13:11,125
has been trained on. You don't know all

340
00:13:11,125 --> 00:13:13,284
the solutions to be trained on Wikipedia. It's

341
00:13:13,284 --> 00:13:15,445
been trained on Reddit. It's been you know?

342
00:13:15,445 --> 00:13:17,284
And what what what is it being trained

343
00:13:17,284 --> 00:13:19,524
on? You don't know. You don't have control.

344
00:13:19,524 --> 00:13:21,605
So one of the things that my friends

345
00:13:21,605 --> 00:13:23,850
in The US and I'm also interested in

346
00:13:24,089 --> 00:13:24,589
is

347
00:13:24,970 --> 00:13:26,589
is is seeing if you can train

348
00:13:27,209 --> 00:13:29,709
a large language model on a corpus

349
00:13:30,089 --> 00:13:30,909
of scientific

350
00:13:31,610 --> 00:13:32,750
data and literature.

351
00:13:33,209 --> 00:13:35,870
And will that give you different solutions

352
00:13:36,250 --> 00:13:38,089
rather than taking what you've got with all

353
00:13:38,089 --> 00:13:39,384
the sort of

354
00:13:40,565 --> 00:13:43,285
extraneous things it's been trained on and then

355
00:13:43,285 --> 00:13:45,684
specializing it. If you train it on on

356
00:13:45,684 --> 00:13:48,825
on scientific data and and and scientific literature,

357
00:13:49,524 --> 00:13:52,100
does that make a difference? And and That's

358
00:13:52,100 --> 00:13:54,500
really interesting. It is really interesting, and I'm

359
00:13:54,500 --> 00:13:57,139
not sure it is it does, actually. But

360
00:13:57,139 --> 00:13:59,379
but it's certainly, they're beginning to look. See,

361
00:13:59,379 --> 00:14:00,039
I work

362
00:14:00,340 --> 00:14:02,340
one of the things that I still would

363
00:14:02,340 --> 00:14:04,600
detain in The US, I'm on the

364
00:14:05,225 --> 00:14:06,845
advanced scientific computing

365
00:14:09,945 --> 00:14:10,445
computing

366
00:14:10,745 --> 00:14:13,865
advanced scientific computing advisory committee for the US

367
00:14:13,865 --> 00:14:16,685
Department of Energy until they've abolished it, alright,

368
00:14:17,305 --> 00:14:19,850
which I don't think they will. But but

369
00:14:20,070 --> 00:14:22,230
the the the Department of Energy has the

370
00:14:22,230 --> 00:14:25,669
the nuclear weapons, which where they fired all

371
00:14:25,669 --> 00:14:27,350
the people who knew about the nuclear weapons

372
00:14:27,350 --> 00:14:30,149
and then had to rehire them because they

373
00:14:30,149 --> 00:14:31,929
suddenly realized well, the previous

374
00:14:32,235 --> 00:14:33,615
previous Trump administration,

375
00:14:33,995 --> 00:14:35,855
the guy went to the Department of Energy

376
00:14:35,914 --> 00:14:37,754
saying he's gonna close it because it was

377
00:14:37,754 --> 00:14:39,754
all about energy and green energy and stuff

378
00:14:39,754 --> 00:14:41,595
like that. But but, actually, it's about where

379
00:14:41,595 --> 00:14:43,995
the bombs are. And eventually, he realized that's

380
00:14:43,995 --> 00:14:45,674
where the bombs are, and you don't really

381
00:14:45,674 --> 00:14:48,389
necessarily want to close that after this step.

382
00:14:49,730 --> 00:14:50,549
So I

383
00:14:50,929 --> 00:14:52,709
but I work with the the nonsecret

384
00:14:53,570 --> 00:14:56,370
part, and there were three supercomputer labs that

385
00:14:56,370 --> 00:14:58,709
I work with. One is Berkeley, One is

386
00:14:59,250 --> 00:15:01,730
Argonne near Chicago, and the other one is

387
00:15:01,730 --> 00:15:03,029
Oak Ridge in Tennessee.

388
00:15:03,554 --> 00:15:05,335
And and there, they have

389
00:15:06,115 --> 00:15:08,514
very the most powerful computers in The US.

390
00:15:08,514 --> 00:15:11,495
They're gigantic things with large numbers. They have

391
00:15:11,794 --> 00:15:14,355
tens of thousands of GPUs on them, so

392
00:15:14,355 --> 00:15:17,495
they can, in fact, do some serious stuff.

393
00:15:17,799 --> 00:15:19,259
They're built for doing supercomputing,

394
00:15:19,879 --> 00:15:21,799
but they can also be used to AI

395
00:15:21,799 --> 00:15:23,980
because they have large numbers of

396
00:15:24,600 --> 00:15:25,899
of of the GPU chips.

397
00:15:26,440 --> 00:15:29,240
They're not necessarily optimally designed for that, but

398
00:15:29,240 --> 00:15:31,615
nonetheless, they're very useful. So that those those

399
00:15:31,615 --> 00:15:33,615
are the community. I think that's a very

400
00:15:33,615 --> 00:15:36,254
valuable community if it isn't destroyed by the

401
00:15:36,254 --> 00:15:37,154
present administration,

402
00:15:37,615 --> 00:15:40,334
but but they're they're they're doing some very

403
00:15:40,334 --> 00:15:42,735
interesting things. If you're not concerned too much

404
00:15:42,735 --> 00:15:44,815
about the climate impact, are there other impacts

405
00:15:44,815 --> 00:15:47,179
of AI that you are more concerned about?

406
00:15:47,960 --> 00:15:49,580
Yes. I'm not concerned about,

407
00:15:50,519 --> 00:15:51,100
you know,

408
00:15:52,679 --> 00:15:55,240
the the the most scary things that taking

409
00:15:55,240 --> 00:15:57,320
over the world and things like this. I

410
00:15:57,320 --> 00:16:00,040
don't actually subscribe to that, that I'm worried

411
00:16:00,040 --> 00:16:00,540
about

412
00:16:01,034 --> 00:16:03,615
the great things also for disinformation,

413
00:16:04,154 --> 00:16:06,095
dis deep fakes, and putting out

414
00:16:06,475 --> 00:16:07,534
all sorts of,

415
00:16:08,475 --> 00:16:11,355
generating all sorts of evil stuff. Yes. I

416
00:16:11,355 --> 00:16:13,674
do. So I my my view on on

417
00:16:13,674 --> 00:16:14,894
that, I hope The UK

418
00:16:15,274 --> 00:16:15,774
has

419
00:16:16,340 --> 00:16:18,360
the search into unethical AI

420
00:16:18,820 --> 00:16:21,080
because, for sure, North Korea,

421
00:16:22,019 --> 00:16:22,519
Iran,

422
00:16:23,060 --> 00:16:26,340
China, Russia are all looking at attacks on

423
00:16:26,340 --> 00:16:26,840
us

424
00:16:27,620 --> 00:16:29,634
with whatever they can do. And so I

425
00:16:29,634 --> 00:16:32,375
don't think they're too concerned about ethical AI.

426
00:16:32,434 --> 00:16:34,194
So it's great to see that we're concerned

427
00:16:34,194 --> 00:16:36,274
about it, but so long as some parts

428
00:16:36,274 --> 00:16:37,095
of our establishment

429
00:16:37,794 --> 00:16:39,815
are actually looking at how you counter

430
00:16:40,194 --> 00:16:40,694
unethical

431
00:16:41,074 --> 00:16:43,634
AI. So that's my question, and I and

432
00:16:43,634 --> 00:16:44,340
I'm sure that

433
00:16:45,540 --> 00:16:48,259
some parts of GCHQ or elsewhere are are

434
00:16:48,259 --> 00:16:50,899
doing some stuff like that. I hope Now

435
00:16:50,899 --> 00:16:52,740
regular listeners will know I'm something of a

436
00:16:52,740 --> 00:16:55,059
fan of particle physics, so let's hear from

437
00:16:55,059 --> 00:16:59,139
a particle physicist. Here is Caterina Dolioni. I'm

438
00:16:59,139 --> 00:17:02,164
a professor of particle physics. I work mainly,

439
00:17:02,625 --> 00:17:04,785
at the ATLAS experiment at the Large Hadron

440
00:17:04,785 --> 00:17:05,285
Collider

441
00:17:05,744 --> 00:17:07,904
and where I do data acquisition, real time

442
00:17:07,904 --> 00:17:09,845
analysis, and searches for

443
00:17:10,464 --> 00:17:12,384
dark matter, try to produce it in the

444
00:17:12,384 --> 00:17:12,690
lab.

445
00:17:13,250 --> 00:17:14,529
I'm also very interested in,

446
00:17:15,250 --> 00:17:19,089
software and open science and the environmental impacts

447
00:17:19,089 --> 00:17:20,450
of the research that we do. You say

448
00:17:20,450 --> 00:17:22,210
you're looking for dark matter. Have you found

449
00:17:22,210 --> 00:17:24,049
any? No. Not yet. I mean, we've not

450
00:17:24,450 --> 00:17:26,210
we don't think we've produced it yet. Or

451
00:17:26,210 --> 00:17:28,634
maybe we have produced it, but it's too

452
00:17:28,634 --> 00:17:30,575
rare to be distinguished from the backgrounds

453
00:17:31,275 --> 00:17:33,035
yet. So we don't know. We keep looking.

454
00:17:33,035 --> 00:17:34,795
Okay. Can can you give me a sense

455
00:17:34,795 --> 00:17:36,234
of what that would look like? I I

456
00:17:36,234 --> 00:17:37,674
know it's quite a hard thing to picture,

457
00:17:37,674 --> 00:17:39,355
but what what would it look like in

458
00:17:39,355 --> 00:17:41,460
the data if you saw that? The easiest

459
00:17:41,460 --> 00:17:42,920
answer is you'd see nothing.

460
00:17:44,019 --> 00:17:46,819
There's a because that's the main signature for

461
00:17:46,819 --> 00:17:49,400
what we in in what we're doing of

462
00:17:49,700 --> 00:17:50,920
a dark matter candidate.

463
00:17:51,859 --> 00:17:54,019
And, when I say candidate, it's it's important

464
00:17:54,019 --> 00:17:56,235
to, like, keep that in mind. It's not

465
00:17:56,235 --> 00:17:58,075
guaranteed that that particle that we're going to

466
00:17:58,075 --> 00:17:59,674
see is dark matter. We did many other

467
00:17:59,674 --> 00:18:01,835
experiments to confirm it. But assume that you're

468
00:18:01,835 --> 00:18:04,174
producing dark matter at the Large Hadron Collider

469
00:18:04,394 --> 00:18:07,115
and an experiment detects it. What you're actually

470
00:18:07,115 --> 00:18:08,735
detecting is the missing

471
00:18:09,580 --> 00:18:12,619
transverse energy that is left by particles exiting

472
00:18:12,619 --> 00:18:14,559
your detector without any trace.

473
00:18:15,100 --> 00:18:17,340
And we use more or less conservation of

474
00:18:17,340 --> 00:18:20,480
energy. It's actually conservation of transfer from momentum,

475
00:18:22,220 --> 00:18:23,119
where you have

476
00:18:23,494 --> 00:18:26,235
two visible particles coming in, colliding,

477
00:18:26,775 --> 00:18:29,035
and then a lot of debris coming out,

478
00:18:29,174 --> 00:18:31,414
most of it, you'll be able to detect

479
00:18:31,414 --> 00:18:33,755
with a detector. But the dark matter particles,

480
00:18:34,855 --> 00:18:36,375
you're not going to be able to detect

481
00:18:36,375 --> 00:18:38,380
them. They're dark. They don't interact very much,

482
00:18:38,380 --> 00:18:39,759
so they're just going to escape.

483
00:18:40,059 --> 00:18:41,820
So you're going to see if you sum

484
00:18:41,820 --> 00:18:43,279
the energy that you had at the beginning

485
00:18:43,420 --> 00:18:44,700
and the energy that you had at the

486
00:18:44,700 --> 00:18:46,700
end, you'll see that something is missing. And

487
00:18:46,700 --> 00:18:47,519
this missing

488
00:18:48,220 --> 00:18:51,534
is the signature of potential dark matter. Okay.

489
00:18:51,835 --> 00:18:53,375
It's also the signature of neutrinos.

490
00:18:53,994 --> 00:18:56,075
So those particles exist. Yes. So that's the

491
00:18:56,075 --> 00:18:57,934
main problem. You have a lot of background,

492
00:18:58,394 --> 00:18:59,615
that you need to distinguish.

493
00:19:00,315 --> 00:19:01,914
You need to distinguish what is signal and

494
00:19:01,914 --> 00:19:03,659
what is background. And a lot of the

495
00:19:03,659 --> 00:19:05,339
time, you can only do it with the

496
00:19:05,339 --> 00:19:07,500
analyzing a lot of data and accumulating a

497
00:19:07,500 --> 00:19:08,159
lot of data.

498
00:19:09,099 --> 00:19:10,460
So that that's one of the ways in

499
00:19:10,460 --> 00:19:12,460
which dark matter could appear. I'm not looking

500
00:19:12,460 --> 00:19:15,179
specifically for that. At the moment, I'm, looking

501
00:19:15,179 --> 00:19:15,679
for,

502
00:19:18,355 --> 00:19:20,535
a sort of a sister theory of,

503
00:19:21,475 --> 00:19:24,674
the quantum chromodynamics theory that is called dark

504
00:19:24,674 --> 00:19:25,575
quantum chromodynamics.

505
00:19:26,434 --> 00:19:28,950
So imagine we have a copy of

506
00:19:29,410 --> 00:19:31,890
our plentiful particles and beautiful particles from the

507
00:19:31,890 --> 00:19:33,910
standard model that we all know and love.

508
00:19:34,210 --> 00:19:36,070
There's a copy of that, and it's

509
00:19:36,370 --> 00:19:38,930
dark. We don't see it because it's only

510
00:19:38,930 --> 00:19:39,430
connected

511
00:19:40,384 --> 00:19:43,345
to the Standard Model particles via very weak

512
00:19:43,345 --> 00:19:44,544
interactions or very,

513
00:19:45,345 --> 00:19:47,284
rare particles or very heavy particles.

514
00:19:48,065 --> 00:19:49,984
So we call that a a complete dark

515
00:19:49,984 --> 00:19:51,365
sector somewhere else.

516
00:19:51,825 --> 00:19:53,744
And within this dark sector, that might be

517
00:19:53,744 --> 00:19:54,804
dark matter particles.

518
00:19:55,130 --> 00:19:56,809
It's one of them or a combination of

519
00:19:56,809 --> 00:19:58,970
this dark sector particles makes it makes it

520
00:19:58,970 --> 00:20:00,830
for a a dark matter candidate.

521
00:20:01,289 --> 00:20:03,450
How does AI come into this research that

522
00:20:03,450 --> 00:20:04,809
you're doing here? It comes in quite a

523
00:20:04,809 --> 00:20:06,490
lot because with the amount of data that

524
00:20:06,490 --> 00:20:08,509
we have, we have to be smart

525
00:20:08,970 --> 00:20:11,505
on how we analyze it. And we could

526
00:20:11,505 --> 00:20:12,005
possibly,

527
00:20:12,785 --> 00:20:15,045
do most of the things that we're doing.

528
00:20:15,184 --> 00:20:17,505
Maybe not not all, but most, I would

529
00:20:17,505 --> 00:20:19,285
say, in classical ways.

530
00:20:20,464 --> 00:20:21,664
The thing is that it would take us

531
00:20:21,664 --> 00:20:24,065
much, much longer. It's the same kind of

532
00:20:24,065 --> 00:20:26,429
revolution that we had in particle physics when

533
00:20:26,429 --> 00:20:28,829
people were looking at slides that taking pictures

534
00:20:28,829 --> 00:20:29,329
of,

535
00:20:30,909 --> 00:20:32,990
a collision or taking picture of a certain

536
00:20:32,990 --> 00:20:34,929
process in bubble chambers

537
00:20:35,230 --> 00:20:37,009
and then going from there to computers.

538
00:20:37,710 --> 00:20:40,190
You have something some algorithm, something that really

539
00:20:40,190 --> 00:20:40,690
accelerates

540
00:20:41,674 --> 00:20:43,914
your the speed at which you can gain

541
00:20:43,914 --> 00:20:45,134
insight from the data.

542
00:20:45,434 --> 00:20:47,855
So that's where machine learning is coming from.

543
00:20:48,474 --> 00:20:50,154
This is only one way in which the

544
00:20:50,154 --> 00:20:52,154
field of physics uses machine learning for data

545
00:20:52,154 --> 00:20:53,914
analysis. So in this case, we might not

546
00:20:53,914 --> 00:20:56,649
be the proponents of new algorithms. We're mostly

547
00:20:56,649 --> 00:20:57,149
users,

548
00:20:57,529 --> 00:21:00,109
but it is still having a huge impact.

549
00:21:00,329 --> 00:21:01,849
But there's other people that are trying to

550
00:21:01,849 --> 00:21:04,089
put physics inside into machine learning. So that's

551
00:21:04,089 --> 00:21:06,569
another kind of crosstalk. I don't do that

552
00:21:06,569 --> 00:21:09,049
specifically. I'm more of a someone who is

553
00:21:09,049 --> 00:21:10,509
using machine learning to

554
00:21:11,025 --> 00:21:12,005
get things done.

555
00:21:12,465 --> 00:21:13,765
I mean, the big question

556
00:21:14,225 --> 00:21:15,744
that seems to come up all the time

557
00:21:15,744 --> 00:21:17,125
with this sort of thing is,

558
00:21:17,505 --> 00:21:19,365
does that mean you won't need your PhD

559
00:21:19,424 --> 00:21:20,545
students? I mean, would you

560
00:21:22,225 --> 00:21:23,684
I would not. I mean,

561
00:21:24,000 --> 00:21:25,839
who's going to do any data analysis if

562
00:21:25,839 --> 00:21:27,299
I do teaching all the time?

563
00:21:29,680 --> 00:21:31,700
No. Anyway, jokes aside, it's,

564
00:21:32,320 --> 00:21:34,099
there's a lot of, experience,

565
00:21:34,720 --> 00:21:35,619
I think, that

566
00:21:36,445 --> 00:21:37,725
it's not something that,

567
00:21:38,205 --> 00:21:40,765
even the best large language model is going

568
00:21:40,765 --> 00:21:43,265
to be able to to inject into

569
00:21:43,724 --> 00:21:45,105
this kind of endeavor.

570
00:21:45,965 --> 00:21:48,144
So I'm not worried that anyone's

571
00:21:50,649 --> 00:21:52,730
knowledge based job is going to be in

572
00:21:52,730 --> 00:21:54,569
our field. At least it's gonna be taken

573
00:21:54,569 --> 00:21:55,069
away

574
00:21:55,609 --> 00:21:56,669
by machine learning

575
00:21:57,690 --> 00:21:59,450
simply because we have the tools, but we

576
00:21:59,450 --> 00:22:01,195
need to know how to use it. We

577
00:22:01,195 --> 00:22:03,275
have Copilot for programming, but we still have

578
00:22:03,275 --> 00:22:04,735
to know how to design the software.

579
00:22:05,115 --> 00:22:06,555
So this is the kind of input that

580
00:22:06,555 --> 00:22:08,394
we still need to need to have. Sure.

581
00:22:08,394 --> 00:22:11,134
If you will go for a general artificial

582
00:22:11,195 --> 00:22:13,914
intelligence or something that is bigger and not

583
00:22:13,914 --> 00:22:16,929
necessarily my my field of expertise, then maybe

584
00:22:16,929 --> 00:22:17,909
you can think about

585
00:22:19,490 --> 00:22:22,630
a longer term future and and see what

586
00:22:22,690 --> 00:22:24,690
what that brings. But at the moment, I

587
00:22:24,690 --> 00:22:27,890
think we do need a lot of human

588
00:22:27,890 --> 00:22:29,029
input in the design

589
00:22:29,515 --> 00:22:30,815
and the use of the correct,

590
00:22:31,674 --> 00:22:33,755
algorithm or correct tool for the problem that

591
00:22:33,755 --> 00:22:35,994
you have at hand. You know, five years

592
00:22:35,994 --> 00:22:36,494
ago,

593
00:22:36,875 --> 00:22:39,355
if you told me that AI was gonna

594
00:22:39,355 --> 00:22:41,434
be as big in society as it is

595
00:22:41,434 --> 00:22:43,115
now, I would have said, no. It's not.

596
00:22:43,115 --> 00:22:45,539
Not not in five years. So how quickly

597
00:22:45,539 --> 00:22:47,779
is that future gonna come around where we

598
00:22:47,779 --> 00:22:50,019
do need to think about it replacing people's

599
00:22:50,019 --> 00:22:51,779
jobs and that sort of thing? Have a

600
00:22:51,779 --> 00:22:53,539
crystal ball at the moment. No. Not from

601
00:22:53,539 --> 00:22:56,599
my field. I think anything that we find

602
00:22:56,659 --> 00:22:59,505
boring can possibly be something that is

603
00:23:00,224 --> 00:23:01,045
taken over

604
00:23:01,424 --> 00:23:03,664
by some algorithm, but that's normal. Like, when

605
00:23:03,664 --> 00:23:06,005
you feel like you have everything done,

606
00:23:06,785 --> 00:23:07,984
like, you have everything that is,

607
00:23:09,664 --> 00:23:11,585
well established, and then you you get a

608
00:23:11,585 --> 00:23:13,765
robot or, you know, something

609
00:23:14,519 --> 00:23:16,119
machine like doing it for you. It has

610
00:23:16,119 --> 00:23:17,980
happened before and that hasn't really

611
00:23:18,680 --> 00:23:21,160
made the so it has made changes, but

612
00:23:21,160 --> 00:23:22,539
it hasn't made changes

613
00:23:23,000 --> 00:23:23,900
for the worse

614
00:23:24,359 --> 00:23:27,075
if handled correctly from the the workers' perspective.

615
00:23:27,075 --> 00:23:29,154
So you can't say, oh, well, bye. I'm

616
00:23:29,154 --> 00:23:31,075
not meeting you anymore. That that's not a

617
00:23:31,075 --> 00:23:32,914
nice thing to say to someone who was

618
00:23:32,914 --> 00:23:34,294
doing the same job as,

619
00:23:34,595 --> 00:23:36,515
you know, a robot or or something. You

620
00:23:36,515 --> 00:23:37,815
need to find parts

621
00:23:38,115 --> 00:23:40,115
in which these people can be appreciated, can

622
00:23:40,115 --> 00:23:42,269
be recognized. So that, I think, my concern

623
00:23:42,269 --> 00:23:44,529
maybe is more about that than about

624
00:23:44,990 --> 00:23:45,650
the replacement

625
00:23:45,950 --> 00:23:47,170
of its

626
00:23:49,549 --> 00:23:50,750
there might be a way to do it,

627
00:23:50,750 --> 00:23:51,789
but we have to do it in a

628
00:23:51,789 --> 00:23:53,650
in a way that's sustainable for people,

629
00:23:54,190 --> 00:23:55,330
not just maximizing,

630
00:23:56,515 --> 00:23:59,154
scientific profit, whatever you want to call it.

631
00:23:59,154 --> 00:24:00,375
Yeah. Yeah. Yeah. Yeah.

632
00:24:01,234 --> 00:24:02,994
So but I'm sorry. I will get on

633
00:24:02,994 --> 00:24:04,515
to the more positive things in a minute,

634
00:24:04,515 --> 00:24:06,615
obviously. But but so you're

635
00:24:06,994 --> 00:24:08,615
doing teaching as well. Does

636
00:24:08,914 --> 00:24:11,430
does students' use of AI concern you?

637
00:24:11,830 --> 00:24:12,330
Yes.

638
00:24:12,789 --> 00:24:14,789
In not in a way that, I mean,

639
00:24:14,789 --> 00:24:16,869
I encourage my students to use I teach

640
00:24:16,869 --> 00:24:19,190
a programming course. So it's something that, if

641
00:24:19,190 --> 00:24:20,950
you don't use AI, you use Stack Overflow.

642
00:24:20,950 --> 00:24:22,789
If you don't use Stack Overflow, use Google.

643
00:24:22,789 --> 00:24:24,505
If you don't use Google, you ask your

644
00:24:24,505 --> 00:24:26,605
friend. So there's always been a case of

645
00:24:26,984 --> 00:24:29,144
I need to find this information, this punctual

646
00:24:29,144 --> 00:24:30,984
information of something that is happening to me

647
00:24:30,984 --> 00:24:32,585
and I don't know how to solve. Where

648
00:24:32,585 --> 00:24:33,404
do I go?

649
00:24:33,784 --> 00:24:35,865
And this is it's fine to go to,

650
00:24:36,664 --> 00:24:38,184
I think large language models,

651
00:24:39,539 --> 00:24:39,620
or

652
00:24:40,820 --> 00:24:42,820
you know, it it's it's replacing it's it's

653
00:24:42,820 --> 00:24:45,080
just evolving something that was happening before.

654
00:24:45,700 --> 00:24:48,340
What I'm not entirely sure I understand at

655
00:24:48,340 --> 00:24:49,559
the moment is how,

656
00:24:50,180 --> 00:24:52,039
the perspective has shifted between,

657
00:24:52,580 --> 00:24:54,440
I want to learn something for myself

658
00:24:54,924 --> 00:24:56,365
because I want to have that experience. I

659
00:24:56,365 --> 00:24:57,904
want to be able to apply that experience,

660
00:24:58,125 --> 00:25:00,125
and I want to reach a goal with

661
00:25:00,125 --> 00:25:00,705
a minimum

662
00:25:01,404 --> 00:25:03,345
amount of effort possible

663
00:25:03,724 --> 00:25:06,125
because that is the trick, I think. That

664
00:25:06,205 --> 00:25:07,289
this is the tricky part.

665
00:25:08,730 --> 00:25:10,029
When we're tired and,

666
00:25:11,369 --> 00:25:12,890
we have a lot of things to do,

667
00:25:12,890 --> 00:25:15,609
we want to just find the easiest way

668
00:25:15,609 --> 00:25:17,769
to get something done. But if you make

669
00:25:17,769 --> 00:25:20,350
your entire course or your entire programming course,

670
00:25:21,450 --> 00:25:22,890
work like that, what have you learned at

671
00:25:22,890 --> 00:25:23,234
the end?

672
00:25:23,795 --> 00:25:25,394
So one thing that we're trying to to

673
00:25:25,394 --> 00:25:27,174
encourage students about is,

674
00:25:27,795 --> 00:25:30,275
share your prompts. Ask the right questions. If

675
00:25:30,275 --> 00:25:31,654
you use this kind of tools,

676
00:25:32,035 --> 00:25:33,634
you have to use them responsibly. You can't

677
00:25:33,634 --> 00:25:35,759
just ask do the assignment for me because

678
00:25:35,920 --> 00:25:36,660
we found,

679
00:25:37,039 --> 00:25:39,140
hallucinations in our in our assignments.

680
00:25:39,599 --> 00:25:41,359
There is that that that is the immediate,

681
00:25:41,599 --> 00:25:44,400
the immediate problem, but also very complicated solutions

682
00:25:44,400 --> 00:25:46,720
for problems that if you just had thought

683
00:25:46,720 --> 00:25:48,494
five minutes, you would have found yourself.

684
00:25:49,454 --> 00:25:51,015
So I'm not saying that this is going

685
00:25:51,015 --> 00:25:53,055
to be overcome. There's going to be better

686
00:25:53,055 --> 00:25:56,095
versions of, Charge GPTs or Copilots that will

687
00:25:56,095 --> 00:25:56,914
do this better.

688
00:25:58,654 --> 00:25:59,555
But it somehow,

689
00:26:00,734 --> 00:26:01,634
doesn't stimulate

690
00:26:02,250 --> 00:26:04,250
learning for oneself. And I think this is

691
00:26:04,250 --> 00:26:05,789
an important part of,

692
00:26:07,930 --> 00:26:09,309
at least a physics degree.

693
00:26:10,009 --> 00:26:12,250
Because we're not I mean, so maybe we

694
00:26:12,250 --> 00:26:13,450
do it we do it also for the

695
00:26:13,450 --> 00:26:15,134
money. We're gonna do it exclusively for the

696
00:26:15,134 --> 00:26:16,975
money. We do it also for our pleasure,

697
00:26:16,975 --> 00:26:17,634
for our,

698
00:26:18,654 --> 00:26:20,414
will to understand the world. And if we're

699
00:26:20,414 --> 00:26:22,174
just asking someone else to understand the world

700
00:26:22,174 --> 00:26:24,195
for us, then where does it leave us?

701
00:26:25,134 --> 00:26:27,535
But I think there's also good things about

702
00:26:27,535 --> 00:26:28,035
the

703
00:26:28,654 --> 00:26:29,795
the the use of

704
00:26:30,380 --> 00:26:31,920
artificial intelligence for

705
00:26:32,620 --> 00:26:35,180
for teaching, for solving teaching problems, for solving

706
00:26:35,180 --> 00:26:38,320
learning problems. It democratizes knowledge quite a lot.

707
00:26:38,620 --> 00:26:40,460
A lot of people have similar access to

708
00:26:40,460 --> 00:26:42,224
knowledge that they might not have had before.

709
00:26:43,825 --> 00:26:45,105
And then there's all the,

710
00:26:45,585 --> 00:26:47,904
the I'm not gonna fall into that hole

711
00:26:47,904 --> 00:26:49,505
at the moment because I would talk about

712
00:26:49,505 --> 00:26:50,325
it for hours.

713
00:26:51,105 --> 00:26:53,424
Environmental sustainability of these kind of tools. So

714
00:26:53,424 --> 00:26:55,539
what are we doing? We're just continuing to

715
00:26:55,539 --> 00:26:57,779
use resources without thinking because it's good for

716
00:26:57,779 --> 00:26:58,920
us because it eases

717
00:26:59,380 --> 00:27:01,880
our understanding. It accelerates our our knowledge.

718
00:27:02,420 --> 00:27:04,660
How do we use it responsibly? My wondering

719
00:27:04,660 --> 00:27:06,099
is do we get to a point in

720
00:27:06,099 --> 00:27:07,160
physics where

721
00:27:07,845 --> 00:27:10,884
it doesn't actually matter if AI is doing

722
00:27:10,884 --> 00:27:11,785
all the discoveries?

723
00:27:12,325 --> 00:27:13,304
Because and

724
00:27:13,765 --> 00:27:15,765
there's a black box though, isn't there, of

725
00:27:15,924 --> 00:27:17,704
we don't know how it did it.

726
00:27:18,325 --> 00:27:20,404
Well, I I think there's a science fiction

727
00:27:20,404 --> 00:27:22,660
story about that. Don't remember the title or

728
00:27:22,660 --> 00:27:24,100
the author at the moment, but I I

729
00:27:24,100 --> 00:27:26,440
brought it up before. And that is where,

730
00:27:26,980 --> 00:27:28,900
in the far future, AI is doing all

731
00:27:28,900 --> 00:27:30,660
discoveries, and there are a few months just

732
00:27:30,660 --> 00:27:32,119
to reverse engineer those.

733
00:27:33,059 --> 00:27:34,275
Figure out how they've done

734
00:27:34,755 --> 00:27:36,195
it. Yeah. So imagine it could be AI,

735
00:27:36,195 --> 00:27:37,795
could be an alien that comes and brings

736
00:27:37,795 --> 00:27:39,955
you up an amazing technology, and then you're

737
00:27:39,955 --> 00:27:41,714
like, okay. How can I use it? Why

738
00:27:41,955 --> 00:27:43,894
how is it there? How is it working?

739
00:27:44,434 --> 00:27:45,894
You know, it's it's kind of reverse

740
00:27:46,275 --> 00:27:48,820
the scientific process because you're not making a

741
00:27:48,820 --> 00:27:51,059
discovery yourself. You're just figuring out how it's

742
00:27:51,059 --> 00:27:53,460
done. But it wouldn't be too different from

743
00:27:53,460 --> 00:27:55,460
reverse engineering a piece of code for someone

744
00:27:55,460 --> 00:27:56,600
who has not left documentation.

745
00:27:57,860 --> 00:27:59,460
You still have to figure things out. So

746
00:27:59,779 --> 00:28:01,860
Can you see that as a possible future

747
00:28:01,860 --> 00:28:03,595
with AI? I mean, I can see a

748
00:28:03,595 --> 00:28:05,194
lot of possible futures for AI, and I'm

749
00:28:05,194 --> 00:28:07,034
worried about different things in the world apart

750
00:28:07,275 --> 00:28:09,694
like, that are not that thing before.

751
00:28:10,875 --> 00:28:11,194
But,

752
00:28:11,994 --> 00:28:12,815
I think it

753
00:28:13,434 --> 00:28:15,515
it it could happen, but I'm still thinking

754
00:28:15,515 --> 00:28:15,839
that

755
00:28:16,720 --> 00:28:18,480
the human brain is the best possible kind

756
00:28:18,480 --> 00:28:20,740
of AI that we got. And the collaborations

757
00:28:20,799 --> 00:28:23,119
of humans is something that is that brings

758
00:28:23,119 --> 00:28:24,019
in serendipitous,

759
00:28:25,359 --> 00:28:27,440
discoveries as well. Not to say that maybe

760
00:28:27,440 --> 00:28:28,720
at some point, we're not going to be

761
00:28:28,720 --> 00:28:30,159
able to simulate that, but I think the

762
00:28:30,159 --> 00:28:31,220
power in the

763
00:28:31,785 --> 00:28:33,725
individuals and collaborations

764
00:28:34,345 --> 00:28:36,025
of of humans is not going to be

765
00:28:36,025 --> 00:28:37,244
something that is easily

766
00:28:37,785 --> 00:28:38,285
matched

767
00:28:39,144 --> 00:28:40,684
by AI. So we'll still

768
00:28:41,545 --> 00:28:43,085
so if the future is

769
00:28:43,945 --> 00:28:45,650
that it's not too near

770
00:28:47,230 --> 00:28:48,830
because we we're not going to be able

771
00:28:48,830 --> 00:28:50,849
to be I I have faith in

772
00:28:51,309 --> 00:28:53,630
in our in in the human aspect of

773
00:28:53,630 --> 00:28:55,309
of science and research, doing a lot of

774
00:28:55,309 --> 00:28:57,390
collaborative science. I think that there's things there

775
00:28:57,390 --> 00:29:00,284
that cannot be replaced by by AI. I

776
00:29:00,284 --> 00:29:02,284
think most people listening will have had some

777
00:29:02,284 --> 00:29:04,845
experience now of using something like chat GPT

778
00:29:04,845 --> 00:29:05,904
or the other,

779
00:29:07,164 --> 00:29:10,365
generative AI word based things, maybe even images.

780
00:29:10,365 --> 00:29:12,365
But how do you actually use AI in

781
00:29:12,365 --> 00:29:13,919
the work that you're doing? How what what

782
00:29:13,919 --> 00:29:15,759
do you actually do with it? So we're

783
00:29:15,759 --> 00:29:16,500
not using,

784
00:29:17,200 --> 00:29:18,259
generative AI

785
00:29:18,879 --> 00:29:21,039
too much. We're figuring out how to do

786
00:29:21,039 --> 00:29:22,659
it, and, there are some

787
00:29:23,119 --> 00:29:26,079
uses. There's, this thing is called ATLAS GPT.

788
00:29:26,079 --> 00:29:28,315
So it's an it's an experiment in within

789
00:29:28,315 --> 00:29:29,054
our experiment

790
00:29:29,595 --> 00:29:32,315
that is trolling the, knowledge base of the

791
00:29:32,315 --> 00:29:32,815
experiment.

792
00:29:33,595 --> 00:29:35,434
And, you ask a question, you get an

793
00:29:35,434 --> 00:29:35,934
answer.

794
00:29:36,954 --> 00:29:39,275
Now an interesting thing, about the training of

795
00:29:39,275 --> 00:29:40,875
that, and I think it's it's part of

796
00:29:40,875 --> 00:29:43,480
the the reason why there's still a lot

797
00:29:43,480 --> 00:29:46,220
of human input needed in this this machinery,

798
00:29:46,759 --> 00:29:49,019
is that if the documentation that it's,

799
00:29:49,480 --> 00:29:50,539
crawling is,

800
00:29:51,320 --> 00:29:53,000
that it's been trained on is obsolete, it's

801
00:29:53,000 --> 00:29:54,194
gonna give you obsolete answers.

802
00:29:55,315 --> 00:29:57,554
And so how do you make sure that

803
00:29:57,554 --> 00:29:58,855
the training data is

804
00:29:59,234 --> 00:29:59,734
proper?

805
00:30:00,355 --> 00:30:02,194
A human has to go there and clean

806
00:30:02,194 --> 00:30:03,815
up the pages of the documentation.

807
00:30:04,595 --> 00:30:07,075
So you still are feeding some human information

808
00:30:07,075 --> 00:30:09,130
into this machine. It's just digesting it and

809
00:30:09,130 --> 00:30:10,910
giving it to you in a different form.

810
00:30:11,289 --> 00:30:12,829
So this is just a simple

811
00:30:13,130 --> 00:30:15,690
example of why I think we still need

812
00:30:15,690 --> 00:30:16,990
human in in the loop.

813
00:30:17,769 --> 00:30:19,849
But so this is not the main use

814
00:30:19,849 --> 00:30:21,450
that we have in our field for for

815
00:30:21,450 --> 00:30:21,884
AI.

816
00:30:22,765 --> 00:30:25,724
We mostly use algorithms that analyze large amount

817
00:30:25,884 --> 00:30:27,664
large number of features that give us,

818
00:30:28,765 --> 00:30:31,164
some insight in the data that we would

819
00:30:31,164 --> 00:30:32,625
not have had otherwise.

820
00:30:33,724 --> 00:30:34,545
So this is

821
00:30:35,329 --> 00:30:37,650
we have different stages in, in our experiments.

822
00:30:37,650 --> 00:30:39,190
One stage is to reconstruct

823
00:30:40,130 --> 00:30:42,150
from the signals that the detector gives

824
00:30:43,170 --> 00:30:43,829
you. Some

825
00:30:44,130 --> 00:30:46,549
there there is a particle with this energy.

826
00:30:46,769 --> 00:30:48,390
So that process we call it reconstruction.

827
00:30:49,250 --> 00:30:51,134
And here you can have a lot of,

828
00:30:51,695 --> 00:30:52,595
neural networks

829
00:30:53,615 --> 00:30:54,755
of different sorts

830
00:30:55,134 --> 00:30:57,055
to do this kind of reconstructions. Imagine you

831
00:30:57,055 --> 00:30:58,894
have a lot of points in space in

832
00:30:58,894 --> 00:31:00,734
your detector because it's like a big digital

833
00:31:00,734 --> 00:31:02,815
camera, and you want to reconstruct tracks of

834
00:31:02,815 --> 00:31:05,660
particles. This is a gigantic combinatorics problem.

835
00:31:06,519 --> 00:31:08,759
Graph neural networks can do it well and

836
00:31:08,759 --> 00:31:11,079
give you an answer of where what tracks

837
00:31:11,079 --> 00:31:13,500
you got with fewer fakes than other algorithms,

838
00:31:13,559 --> 00:31:14,220
for example.

839
00:31:15,845 --> 00:31:18,404
Or you can, use it for clustering problems.

840
00:31:18,404 --> 00:31:20,244
That's one of the other classic uses of

841
00:31:20,244 --> 00:31:21,384
AI. So if you have

842
00:31:21,765 --> 00:31:24,404
a a detector that gives you, spray of

843
00:31:24,404 --> 00:31:27,365
particles and the spray of particles represented by

844
00:31:27,365 --> 00:31:28,825
a bunch of energy deposits,

845
00:31:29,420 --> 00:31:31,819
then you can use, AI to cluster this

846
00:31:31,819 --> 00:31:34,059
energy deposits in a more efficient way, for

847
00:31:34,059 --> 00:31:35,759
example. So that's one thing that,

848
00:31:36,299 --> 00:31:37,119
they can do.

849
00:31:37,819 --> 00:31:39,679
Tagging, we call it identifying

850
00:31:40,140 --> 00:31:42,619
particle a from particle b. That's another really

851
00:31:42,619 --> 00:31:42,940
big,

852
00:31:44,194 --> 00:31:46,194
problem that we use for we use AI

853
00:31:46,194 --> 00:31:49,414
for. So is this particle a quark,

854
00:31:49,954 --> 00:31:51,894
derived from a quark of type b?

855
00:31:52,835 --> 00:31:54,835
Then this is the beauty core. It has

856
00:31:54,835 --> 00:31:55,335
the

857
00:31:56,299 --> 00:31:57,980
this party will have the tendency to fly

858
00:31:57,980 --> 00:32:00,059
a little away from the interaction vertex. So

859
00:32:00,059 --> 00:32:00,720
you find

860
00:32:01,019 --> 00:32:02,480
a neural network that can

861
00:32:02,940 --> 00:32:05,579
distinguish what is a probe particle from a

862
00:32:05,579 --> 00:32:06,960
particle that's long lived.

863
00:32:07,500 --> 00:32:09,419
So this is something that we use quite

864
00:32:09,419 --> 00:32:10,335
a lot for this

865
00:32:11,454 --> 00:32:13,774
well, maybe basic task, but it's not quite

866
00:32:13,774 --> 00:32:15,954
data analysis. Data analysis, I have

867
00:32:16,414 --> 00:32:19,375
10 b's and five electrons. What is the

868
00:32:19,375 --> 00:32:22,414
process that gave it this output to me?

869
00:32:22,414 --> 00:32:24,194
I mean, very, very simple. Right?

870
00:32:24,980 --> 00:32:26,740
We we have many more particles in our

871
00:32:26,740 --> 00:32:27,240
events.

872
00:32:27,619 --> 00:32:29,059
And then you can use all of this

873
00:32:29,059 --> 00:32:31,299
for data analysis because you can you have,

874
00:32:31,539 --> 00:32:33,460
the the problem there is to distinguish signal

875
00:32:33,460 --> 00:32:34,119
for background,

876
00:32:34,420 --> 00:32:36,740
and you have a variety of supervised and

877
00:32:36,740 --> 00:32:37,240
unsupervised

878
00:32:37,619 --> 00:32:37,964
methods

879
00:32:38,845 --> 00:32:39,904
to to make this,

880
00:32:40,765 --> 00:32:41,265
distinction.

881
00:32:41,884 --> 00:32:44,204
One thing that I've been dabbling with, but,

882
00:32:45,005 --> 00:32:45,904
is unsupervised

883
00:32:46,204 --> 00:32:46,704
learning,

884
00:32:47,484 --> 00:32:49,644
what, people call the anomaly detection or the

885
00:32:49,644 --> 00:32:50,910
outlier detection methods.

886
00:32:51,470 --> 00:32:53,470
It's the the kind of algorithm that ping

887
00:32:53,470 --> 00:32:54,849
your credit card when you're

888
00:32:55,150 --> 00:32:56,670
abroad. So it's the same thing that we

889
00:32:56,670 --> 00:32:58,590
use. And the advantage is that we don't

890
00:32:58,590 --> 00:33:01,710
know what to expect from the the new

891
00:33:01,710 --> 00:33:02,210
physics.

892
00:33:02,670 --> 00:33:04,750
Ideally, we just we know very well. We

893
00:33:04,750 --> 00:33:06,345
know standard model very well, and we have

894
00:33:06,505 --> 00:33:08,424
plenty of standard model data. Most of our

895
00:33:08,424 --> 00:33:10,744
collision and most of our collisions that we

896
00:33:10,744 --> 00:33:12,605
analyze are standard model processes.

897
00:33:13,144 --> 00:33:15,304
So can we get an algorithm to learn

898
00:33:15,304 --> 00:33:16,525
how this looks like

899
00:33:16,984 --> 00:33:18,825
and then tell us if there's any difference

900
00:33:18,825 --> 00:33:19,484
or deviations

901
00:33:20,730 --> 00:33:21,390
from the

902
00:33:21,690 --> 00:33:24,490
the data that the, that we have. Now

903
00:33:24,490 --> 00:33:26,570
it's conceptually, this is very easy, and it's

904
00:33:26,570 --> 00:33:28,509
beautiful. It works super well.

905
00:33:28,890 --> 00:33:29,630
In practice,

906
00:33:30,090 --> 00:33:31,230
there's a lot

907
00:33:31,849 --> 00:33:33,609
of pitfalls and the things that one needs

908
00:33:33,609 --> 00:33:35,884
to understand before calling an anomaly

909
00:33:36,184 --> 00:33:37,244
seen by an algorithm,

910
00:33:37,865 --> 00:33:38,605
new physics.

911
00:33:39,065 --> 00:33:39,805
For example,

912
00:33:40,184 --> 00:33:41,964
detector noise that you weren't expecting

913
00:33:42,984 --> 00:33:43,484
or,

914
00:33:45,705 --> 00:33:48,184
you know, clustering of a specific process that

915
00:33:48,184 --> 00:33:50,279
was too rare to be seen in your

916
00:33:50,359 --> 00:33:52,279
previous data analysis, but now you found that

917
00:33:52,279 --> 00:33:53,660
out and if not your physics.

918
00:33:54,599 --> 00:33:55,900
So trying to understand

919
00:33:56,200 --> 00:33:56,940
what the

920
00:33:57,720 --> 00:34:00,519
known unknowns are and what the unknown unknowns

921
00:34:00,519 --> 00:34:03,325
are, that the that's the key that's the

922
00:34:03,325 --> 00:34:05,085
key problem there, not necessarily what I gotta

923
00:34:05,085 --> 00:34:07,184
be using for your anomaly detection.

924
00:34:08,605 --> 00:34:10,444
So I I I'm very interested in that,

925
00:34:10,684 --> 00:34:12,444
that side of things because it also has,

926
00:34:15,005 --> 00:34:16,065
an impact on

927
00:34:16,760 --> 00:34:19,579
more basic data position things like data compression.

928
00:34:20,839 --> 00:34:22,679
So if you don't, if you want to

929
00:34:22,679 --> 00:34:25,339
compress using machine compress data using machine learning,

930
00:34:26,039 --> 00:34:27,320
are you going to just,

931
00:34:27,719 --> 00:34:30,565
wipe out any anomalies because the compression will

932
00:34:30,565 --> 00:34:32,425
reconstruct will compress them badly?

933
00:34:33,045 --> 00:34:35,204
It's only the anomalies, and then you'll just

934
00:34:35,204 --> 00:34:36,965
put them back into the bulk, and then

935
00:34:36,965 --> 00:34:38,105
you've lost your signal.

936
00:34:38,885 --> 00:34:41,204
And so this is all interesting interplay that,

937
00:34:41,719 --> 00:34:43,480
you have to think about it when you're

938
00:34:43,480 --> 00:34:46,440
using AI and physics. Yeah. Yeah. But so

939
00:34:46,440 --> 00:34:49,000
you do need your students to be able

940
00:34:49,000 --> 00:34:50,219
to understand things

941
00:34:50,839 --> 00:34:53,319
Yeah. Without the AI, the use of AI.

942
00:34:53,319 --> 00:34:54,839
They need to understand how to use it

943
00:34:55,000 --> 00:34:56,679
Yeah. But also how to do it if

944
00:34:56,679 --> 00:34:58,804
it wasn't there. I think there's also another

945
00:34:58,804 --> 00:35:00,025
aspect that the communities,

946
00:35:00,404 --> 00:35:02,424
has realized that you need to have,

947
00:35:04,005 --> 00:35:04,505
reproducible

948
00:35:05,764 --> 00:35:07,844
science. This is generally a pillar of what

949
00:35:07,844 --> 00:35:08,585
we're doing.

950
00:35:09,045 --> 00:35:10,644
You can also go into the open science.

951
00:35:10,644 --> 00:35:12,570
So you want someone else, not you, maybe

952
00:35:12,570 --> 00:35:15,230
the general public even, to reproduce, you know,

953
00:35:16,090 --> 00:35:17,869
reasonable manner what you're doing.

954
00:35:18,650 --> 00:35:20,030
AI doesn't make this easy,

955
00:35:21,210 --> 00:35:23,550
mostly because of the complexity and the

956
00:35:23,849 --> 00:35:25,769
the complexity of the networks, but also the

957
00:35:25,769 --> 00:35:28,804
computational resources needed for running large algorithms.

958
00:35:29,344 --> 00:35:30,565
So how do you make,

959
00:35:33,505 --> 00:35:34,724
AI based analysis

960
00:35:35,344 --> 00:35:36,565
accessible to others,

961
00:35:36,864 --> 00:35:39,525
understandable by others, reproducible by others?

962
00:35:39,984 --> 00:35:40,484
And

963
00:35:41,230 --> 00:35:43,150
sometimes the easy answer is, well, you don't.

964
00:35:43,150 --> 00:35:44,829
You give them a version of the analysis

965
00:35:44,829 --> 00:35:46,050
that sees the same thing

966
00:35:46,670 --> 00:35:48,369
that is not using AI.

967
00:35:48,750 --> 00:35:50,510
So you still need someone to do that

968
00:35:50,510 --> 00:35:52,910
part. And, usually, when you want to convince

969
00:35:52,910 --> 00:35:55,070
someone that you've made the discovery, you will

970
00:35:55,070 --> 00:35:57,284
be asked, I think, over the field, can

971
00:35:57,284 --> 00:35:58,264
you please now

972
00:35:59,045 --> 00:36:00,344
do something that is,

973
00:36:02,405 --> 00:36:04,664
that at least indicates in the right direction

974
00:36:05,125 --> 00:36:06,105
that you have

975
00:36:06,405 --> 00:36:06,905
something

976
00:36:07,284 --> 00:36:09,160
solid with your AI algorithm.

977
00:36:09,539 --> 00:36:11,460
I think what we're missing is also the,

978
00:36:11,700 --> 00:36:14,180
some more crosstalk with people that are working

979
00:36:14,180 --> 00:36:14,579
on,

980
00:36:14,980 --> 00:36:16,420
explainable AI and,

981
00:36:18,180 --> 00:36:20,420
because we don't we don't want to treat

982
00:36:20,420 --> 00:36:22,574
our data analysis as a black box.

983
00:36:23,454 --> 00:36:25,554
We can't. That's not scientific method.

984
00:36:26,734 --> 00:36:27,135
And,

985
00:36:27,534 --> 00:36:29,474
this kind of, theoretical advances,

986
00:36:29,855 --> 00:36:32,014
when we try to to work with people

987
00:36:32,014 --> 00:36:33,795
that work on this, this topic,

988
00:36:34,255 --> 00:36:36,255
are still very, very far from our,

989
00:36:37,359 --> 00:36:37,859
understanding.

990
00:36:39,119 --> 00:36:40,019
People are doing

991
00:36:40,400 --> 00:36:42,719
marvelous work, and, it all makes sense if

992
00:36:42,719 --> 00:36:45,359
you think about it in from afar and

993
00:36:45,359 --> 00:36:46,719
when they explain to you. But how do

994
00:36:46,719 --> 00:36:48,559
we get that into the physics field? How

995
00:36:48,559 --> 00:36:50,054
do we how do we make sure that

996
00:36:50,614 --> 00:36:53,355
we're using the state of the art? Mhmm.

997
00:36:53,414 --> 00:36:56,295
Caterina was also involved in the IOP's white

998
00:36:56,295 --> 00:36:58,295
paper. It came out of a workshop that

999
00:36:58,295 --> 00:36:59,994
wants to take the temperature off

1000
00:37:00,855 --> 00:37:03,929
how AI is impacting physics and where are

1001
00:37:03,929 --> 00:37:05,690
we going with that. Because I think the

1002
00:37:05,690 --> 00:37:07,469
IOP is a very,

1003
00:37:09,210 --> 00:37:12,269
good, very strong stakeholder, but also can influence

1004
00:37:12,409 --> 00:37:14,170
where we're going also because of the impact

1005
00:37:14,170 --> 00:37:15,869
on teaching, undergraduate teaching.

1006
00:37:16,570 --> 00:37:18,554
So there was a workshop in October, I

1007
00:37:18,554 --> 00:37:19,454
believe, that,

1008
00:37:20,315 --> 00:37:20,815
gathered

1009
00:37:21,114 --> 00:37:22,335
information from,

1010
00:37:23,034 --> 00:37:23,775
the participants

1011
00:37:24,155 --> 00:37:26,494
in a variety of of forms and

1012
00:37:26,954 --> 00:37:29,295
the then summary this white paper summarizes

1013
00:37:29,594 --> 00:37:31,289
the the findings of this workshop.

1014
00:37:31,849 --> 00:37:33,929
And here, it's the AI is really broad,

1015
00:37:33,929 --> 00:37:34,590
so it's

1016
00:37:34,969 --> 00:37:35,789
machine learning,

1017
00:37:36,090 --> 00:37:38,410
like the traditional machine learning, but then there's

1018
00:37:38,410 --> 00:37:39,309
also other

1019
00:37:39,849 --> 00:37:42,010
all other kinds of things. So it's, the

1020
00:37:42,010 --> 00:37:44,204
participants were also from different fields.

1021
00:37:45,565 --> 00:37:47,905
But in general, it's, it came out that,

1022
00:37:48,364 --> 00:37:50,065
of course, AI is useful.

1023
00:37:50,525 --> 00:37:51,025
It's

1024
00:37:53,244 --> 00:37:56,065
and physics can play and is already playing

1025
00:37:56,489 --> 00:37:58,650
a a special role in AI and can

1026
00:37:58,650 --> 00:37:59,469
also be

1027
00:38:00,809 --> 00:38:02,510
highlight that can also be highlighted

1028
00:38:03,690 --> 00:38:04,670
highlighted further.

1029
00:38:05,849 --> 00:38:07,530
And and at the moment, we can say

1030
00:38:07,530 --> 00:38:09,309
that I think physics needs AI

1031
00:38:09,974 --> 00:38:11,434
because of this, data,

1032
00:38:11,974 --> 00:38:14,155
reconstruction, data processing, data analysis.

1033
00:38:15,175 --> 00:38:16,535
Where we have a lot of datasets, a

1034
00:38:16,535 --> 00:38:18,454
lot of features, then AI is really making

1035
00:38:18,454 --> 00:38:20,215
a difference. Sorry, Scott. Just to go to

1036
00:38:20,215 --> 00:38:22,695
the environmental thing, is it the case that

1037
00:38:22,695 --> 00:38:24,869
it was it will speed things up to

1038
00:38:24,869 --> 00:38:26,089
such a degree that

1039
00:38:26,469 --> 00:38:28,730
that will mean it's using less resources

1040
00:38:29,269 --> 00:38:31,989
because we're not Hopefully. Okay. I think this

1041
00:38:31,989 --> 00:38:33,609
was a a recent, UN,

1042
00:38:34,150 --> 00:38:34,949
report that,

1043
00:38:35,509 --> 00:38:37,190
where it can use AI to improve the

1044
00:38:37,190 --> 00:38:37,769
the environment.

1045
00:38:38,454 --> 00:38:40,054
Right? So it's it's kind of coming full

1046
00:38:40,054 --> 00:38:40,554
circle.

1047
00:38:41,494 --> 00:38:44,135
But I think we don't know enough or

1048
00:38:44,135 --> 00:38:46,054
at least it's not transparent enough on how

1049
00:38:46,054 --> 00:38:47,815
we run our resource and how how much

1050
00:38:47,815 --> 00:38:48,474
our resources

1051
00:38:48,855 --> 00:38:49,355
cost.

1052
00:38:50,295 --> 00:38:50,795
So,

1053
00:38:51,734 --> 00:38:52,875
there's a number of

1054
00:38:53,210 --> 00:38:54,210
groups that are,

1055
00:38:54,650 --> 00:38:56,329
trying to do that. We also have a

1056
00:38:56,329 --> 00:38:57,469
UKIF. It's the

1057
00:38:57,849 --> 00:39:00,650
European Coalition for AI and Fundamental Sciences, and

1058
00:39:00,650 --> 00:39:03,630
we have a group on on, environmental sustainability.

1059
00:39:04,009 --> 00:39:06,224
It's not yet taken off, but it hopefully

1060
00:39:06,224 --> 00:39:06,724
will.

1061
00:39:07,505 --> 00:39:09,985
That are just informing people. That that's the

1062
00:39:09,985 --> 00:39:11,125
main the main point.

1063
00:39:12,065 --> 00:39:13,985
Computing is not free in general. We think

1064
00:39:13,985 --> 00:39:15,585
that it's free because we are not paying

1065
00:39:15,585 --> 00:39:17,765
the bill. We're not paying the electricity bill,

1066
00:39:17,949 --> 00:39:18,690
but the electricity bill

1067
00:39:18,989 --> 00:39:19,969
translates into

1068
00:39:20,309 --> 00:39:20,809
a,

1069
00:39:21,469 --> 00:39:21,969
environmental

1070
00:39:22,910 --> 00:39:24,910
bill, servers, and the there's a life cycle

1071
00:39:24,910 --> 00:39:27,170
assessment of all of that stuff. So it's

1072
00:39:27,390 --> 00:39:30,190
trying to make the wisest choices to obtain

1073
00:39:30,190 --> 00:39:31,710
the results that we want. It doesn't mean

1074
00:39:31,710 --> 00:39:34,284
turning everything off because it consumes less or

1075
00:39:34,284 --> 00:39:37,244
not using the rate latest and greatest AI

1076
00:39:37,244 --> 00:39:39,965
algorithm because it consumes too much. Just trying

1077
00:39:39,965 --> 00:39:40,465
to,

1078
00:39:41,565 --> 00:39:42,625
map the problem

1079
00:39:43,324 --> 00:39:45,485
to the algorithm in a way that also

1080
00:39:45,485 --> 00:39:47,965
includes environmental sustainability as one of the axis,

1081
00:39:47,965 --> 00:39:49,349
not just speed or

1082
00:39:49,889 --> 00:39:51,510
that 1% more or,

1083
00:39:51,969 --> 00:39:52,949
you know, efficiency.

1084
00:39:54,050 --> 00:39:56,389
Felice Frankel is a science photographer

1085
00:39:56,849 --> 00:39:58,949
and research scientist at the Massachusetts

1086
00:39:59,489 --> 00:40:00,630
Institute of Technology

1087
00:40:01,055 --> 00:40:04,414
or MIT. In a recent daily briefing on

1088
00:40:04,414 --> 00:40:05,795
nature.com,

1089
00:40:05,855 --> 00:40:09,394
Felice wrote about what science photos do that

1090
00:40:09,454 --> 00:40:11,635
AI generated images can't.

1091
00:40:12,255 --> 00:40:14,575
Most of the work I do is in

1092
00:40:14,575 --> 00:40:17,280
science photography. That is I make pictures

1093
00:40:17,740 --> 00:40:18,720
of the science,

1094
00:40:19,500 --> 00:40:20,559
or I create

1095
00:40:21,260 --> 00:40:21,760
images

1096
00:40:22,780 --> 00:40:23,280
that,

1097
00:40:23,900 --> 00:40:26,960
that are metaphoric in as far as describing

1098
00:40:27,099 --> 00:40:28,800
what the research is about.

1099
00:40:29,414 --> 00:40:32,135
And, of course, everybody is thinking about AI,

1100
00:40:32,135 --> 00:40:34,875
and I'm I'm now looking at what

1101
00:40:35,255 --> 00:40:36,635
AI can do photographically.

1102
00:40:37,655 --> 00:40:40,715
Not not in the sciences, but I'm seeing

1103
00:40:41,015 --> 00:40:42,855
in the, you know, in the world around

1104
00:40:42,855 --> 00:40:43,559
us this

1105
00:40:44,119 --> 00:40:44,619
extraordinary

1106
00:40:45,960 --> 00:40:47,420
tool that can create

1107
00:40:47,880 --> 00:40:49,019
an image from

1108
00:40:49,400 --> 00:40:49,900
nothing,

1109
00:40:50,599 --> 00:40:51,099
basically.

1110
00:40:52,519 --> 00:40:53,340
And so

1111
00:40:54,039 --> 00:40:57,160
I started thinking, oh, boy. You know? Maybe

1112
00:40:57,160 --> 00:40:58,680
I'll be out of a job at one

1113
00:40:58,680 --> 00:40:59,704
point because

1114
00:41:00,325 --> 00:41:01,305
AI someday

1115
00:41:01,844 --> 00:41:02,664
will be able

1116
00:41:03,204 --> 00:41:03,704
to

1117
00:41:04,644 --> 00:41:05,125
depict,

1118
00:41:06,325 --> 00:41:06,825
research

1119
00:41:07,285 --> 00:41:08,025
art completely

1120
00:41:08,325 --> 00:41:08,825
artificially.

1121
00:41:10,005 --> 00:41:11,144
That is to say,

1122
00:41:11,605 --> 00:41:12,030
not

1123
00:41:13,949 --> 00:41:16,530
creating an image from the thing,

1124
00:41:17,550 --> 00:41:20,289
but will develop an image from pixels

1125
00:41:20,590 --> 00:41:23,650
out of out of the large language

1126
00:41:24,030 --> 00:41:25,010
model. So

1127
00:41:25,485 --> 00:41:25,985
it's

1128
00:41:26,445 --> 00:41:28,684
it was it started about a year ago

1129
00:41:28,684 --> 00:41:30,465
when I when I when I was thinking,

1130
00:41:30,765 --> 00:41:31,825
am I in trouble?

1131
00:41:32,365 --> 00:41:34,704
And, you know, I'm the fact is

1132
00:41:35,405 --> 00:41:37,965
I have to be realistic. I'm next month,

1133
00:41:37,965 --> 00:41:40,625
I'm turning 80, I will have you know.

1134
00:41:40,929 --> 00:41:43,349
I can't believe that I'm that old. Right?

1135
00:41:43,890 --> 00:41:45,890
And so I think I'm thinking about my

1136
00:41:45,890 --> 00:41:46,390
future,

1137
00:41:47,250 --> 00:41:49,650
but it but it's it was it's an

1138
00:41:49,650 --> 00:41:50,150
interesting

1139
00:41:50,449 --> 00:41:51,829
thing to think about,

1140
00:41:52,530 --> 00:41:53,590
if in fact,

1141
00:41:53,894 --> 00:41:56,295
I will be taken over even if I

1142
00:41:56,295 --> 00:41:57,034
were young

1143
00:41:57,494 --> 00:41:59,894
by AI. And and so the article was

1144
00:41:59,894 --> 00:42:00,875
is investigating

1145
00:42:01,335 --> 00:42:01,994
that idea.

1146
00:42:02,695 --> 00:42:03,675
And where,

1147
00:42:03,974 --> 00:42:05,355
I did some experiments,

1148
00:42:05,735 --> 00:42:08,534
I I have a a photograph that I

1149
00:42:08,534 --> 00:42:09,515
made of

1150
00:42:10,849 --> 00:42:12,289
science that, in fact,

1151
00:42:12,769 --> 00:42:14,230
Munji Gewendi's science

1152
00:42:14,690 --> 00:42:16,849
got a Nobel Prize last year for this

1153
00:42:16,849 --> 00:42:17,349
work.

1154
00:42:18,050 --> 00:42:20,630
It it's an image that you're looking at

1155
00:42:20,849 --> 00:42:21,349
nanocrystals

1156
00:42:21,890 --> 00:42:24,230
in in vials, different colors,

1157
00:42:24,735 --> 00:42:27,235
fluorescent different wavelengths. Let's leave it at that.

1158
00:42:27,615 --> 00:42:30,735
And so I asked AI to with my

1159
00:42:30,735 --> 00:42:31,235
prompt

1160
00:42:32,175 --> 00:42:35,135
to create x, y, and z. And very

1161
00:42:35,295 --> 00:42:37,954
I used various models, and they were terrible.

1162
00:42:39,179 --> 00:42:41,920
They were cartoon like. They were silly,

1163
00:42:42,699 --> 00:42:43,199
but

1164
00:42:43,659 --> 00:42:45,359
I could see it happening.

1165
00:42:45,980 --> 00:42:47,039
At some point,

1166
00:42:47,900 --> 00:42:50,460
AI will be able to create an image

1167
00:42:50,460 --> 00:42:51,280
that looks

1168
00:42:51,820 --> 00:42:52,559
and represents

1169
00:42:53,195 --> 00:42:54,574
like the real science.

1170
00:42:57,594 --> 00:43:00,655
The key is how are we going to

1171
00:43:01,434 --> 00:43:02,574
create a system

1172
00:43:03,514 --> 00:43:06,094
to judge whether an image for

1173
00:43:06,394 --> 00:43:06,894
submission

1174
00:43:07,880 --> 00:43:09,500
is AI or not.

1175
00:43:09,880 --> 00:43:11,820
And so at the very end

1176
00:43:12,440 --> 00:43:15,340
of the article, I list some ideas of

1177
00:43:15,719 --> 00:43:17,340
when you submit an image,

1178
00:43:17,880 --> 00:43:20,760
you you ask the researcher to say, is

1179
00:43:20,760 --> 00:43:22,059
this an AI image?

1180
00:43:22,364 --> 00:43:25,405
If so, what model you used? What prompt

1181
00:43:25,405 --> 00:43:27,244
did you use? You know, a number of

1182
00:43:27,244 --> 00:43:27,744
questions.

1183
00:43:28,364 --> 00:43:29,664
But in the end,

1184
00:43:30,764 --> 00:43:33,985
the key is that we should not permit

1185
00:43:34,764 --> 00:43:36,945
any AI image to

1186
00:43:37,309 --> 00:43:38,130
be presented

1187
00:43:38,590 --> 00:43:40,690
as a record of the science.

1188
00:43:41,389 --> 00:43:42,289
Yes. AI

1189
00:43:42,670 --> 00:43:45,630
will be very good at creating images that

1190
00:43:45,630 --> 00:43:46,449
are explanatory.

1191
00:43:47,789 --> 00:43:48,289
Conceptually,

1192
00:43:48,829 --> 00:43:49,809
even structurally,

1193
00:43:50,429 --> 00:43:52,449
it sort of looks like this,

1194
00:43:53,255 --> 00:43:54,315
but never

1195
00:43:54,695 --> 00:43:56,715
accept it as the record

1196
00:43:57,015 --> 00:43:59,195
of the science. And, unfortunately,

1197
00:43:59,655 --> 00:44:00,954
we have to trust

1198
00:44:01,494 --> 00:44:02,235
the submission.

1199
00:44:03,255 --> 00:44:04,235
If AI

1200
00:44:04,934 --> 00:44:05,434
keeps

1201
00:44:06,215 --> 00:44:06,715
improving,

1202
00:44:07,880 --> 00:44:10,119
you can never see an occasion where it

1203
00:44:10,119 --> 00:44:12,360
gets to the point where we can accept

1204
00:44:12,360 --> 00:44:14,599
it for that. Yeah. Because it it's the

1205
00:44:14,599 --> 00:44:15,099
intention.

1206
00:44:15,480 --> 00:44:17,480
That is the that's the key to the

1207
00:44:17,480 --> 00:44:17,980
submission.

1208
00:44:18,680 --> 00:44:19,660
If your intention

1209
00:44:20,519 --> 00:44:22,860
is to explain the science,

1210
00:44:24,574 --> 00:44:27,554
fine. If your intention is to say this

1211
00:44:27,775 --> 00:44:29,635
is a record of the science,

1212
00:44:30,815 --> 00:44:33,954
that's verboten for me and for anybody, really.

1213
00:44:34,335 --> 00:44:36,574
Now I don't know the answer of how

1214
00:44:36,574 --> 00:44:37,074
to

1215
00:44:37,534 --> 00:44:38,514
teach honesty.

1216
00:44:39,800 --> 00:44:41,820
I mean, we we've seen manipulated

1217
00:44:42,199 --> 00:44:45,320
images. We there's a whole list of images

1218
00:44:45,320 --> 00:44:47,880
that are we're looking at that have been

1219
00:44:47,880 --> 00:44:48,380
manipulated

1220
00:44:49,880 --> 00:44:50,860
to, in fact,

1221
00:44:52,275 --> 00:44:54,775
say what the researcher wanted to say

1222
00:44:55,714 --> 00:44:58,614
because the science wasn't there. So that's that's

1223
00:44:58,994 --> 00:45:02,275
the manipulation of images is something that we've

1224
00:45:02,275 --> 00:45:04,594
been around that's been around for years. Even,

1225
00:45:04,594 --> 00:45:05,335
for example,

1226
00:45:05,710 --> 00:45:07,090
as I say in the article,

1227
00:45:07,390 --> 00:45:08,369
these glorious

1228
00:45:08,750 --> 00:45:11,329
images that we see of of the universe

1229
00:45:11,949 --> 00:45:13,969
from the James Webb and the Hubble,

1230
00:45:14,510 --> 00:45:16,369
those are all highly manipulated

1231
00:45:16,670 --> 00:45:19,784
images, those colors. People think the universe looks

1232
00:45:19,784 --> 00:45:22,344
like that. It doesn't. But we but we'd

1233
00:45:22,344 --> 00:45:24,605
say that these have been falsely colored.

1234
00:45:25,144 --> 00:45:27,484
At least, we should be saying that. And

1235
00:45:28,025 --> 00:45:28,525
but,

1236
00:45:29,864 --> 00:45:31,565
so changing an image

1237
00:45:32,829 --> 00:45:35,010
is something that we've been doing for

1238
00:45:35,389 --> 00:45:38,530
for a while, but it's starting from scratch

1239
00:45:39,070 --> 00:45:40,769
to create an image that

1240
00:45:41,230 --> 00:45:42,849
literally never existed

1241
00:45:44,349 --> 00:45:45,089
to represent

1242
00:45:45,550 --> 00:45:46,289
the thing

1243
00:45:46,905 --> 00:45:49,545
is, in fact, the problem. Because you have

1244
00:45:49,545 --> 00:45:50,045
this

1245
00:45:50,344 --> 00:45:52,664
photography interest. Right? You have the science interest

1246
00:45:52,664 --> 00:45:54,045
and the photography interest.

1247
00:45:54,505 --> 00:45:57,724
Does looking at an AI photograph, a photograph

1248
00:45:57,785 --> 00:45:59,864
that's been generated, not an image, but a

1249
00:45:59,864 --> 00:46:01,960
photograph that's been generated by AI, AI. You

1250
00:46:01,960 --> 00:46:04,219
know, something to look like a photograph.

1251
00:46:04,519 --> 00:46:07,019
Does that offend your sort of artistic

1252
00:46:07,559 --> 00:46:08,059
sensibilities

1253
00:46:08,360 --> 00:46:09,019
as well?

1254
00:46:09,320 --> 00:46:11,239
Let me first say I'm not an artist,

1255
00:46:11,239 --> 00:46:12,699
and that's kind of important.

1256
00:46:13,880 --> 00:46:14,380
Artistically,

1257
00:46:14,760 --> 00:46:17,194
I'm very I'm blown away by what is

1258
00:46:17,394 --> 00:46:19,315
what people are able to do with AI.

1259
00:46:19,315 --> 00:46:20,454
It's just remarkable.

1260
00:46:21,394 --> 00:46:22,775
But as a scientist,

1261
00:46:24,034 --> 00:46:25,335
I I'm worried

1262
00:46:26,355 --> 00:46:26,855
because

1263
00:46:27,795 --> 00:46:29,094
it's about truth

1264
00:46:29,714 --> 00:46:30,454
as as

1265
00:46:31,460 --> 00:46:32,980
as what I'm trying to do when I

1266
00:46:32,980 --> 00:46:34,440
make an image. Remember,

1267
00:46:34,980 --> 00:46:37,319
when I make a photograph an image,

1268
00:46:37,940 --> 00:46:39,400
it is a representation

1269
00:46:39,940 --> 00:46:40,839
of the work.

1270
00:46:41,139 --> 00:46:42,920
It is not the work.

1271
00:46:43,244 --> 00:46:45,184
Is is it a it's a representation.

1272
00:46:46,045 --> 00:46:47,184
So there's always

1273
00:46:48,045 --> 00:46:50,384
in my picture some sort of manipulation.

1274
00:46:50,924 --> 00:46:53,324
The very nature of making an image is

1275
00:46:53,324 --> 00:46:53,904
a manipulation

1276
00:46:54,204 --> 00:46:54,944
of reality.

1277
00:46:55,404 --> 00:46:57,880
You know, I'm not, you're not showing everything.

1278
00:46:57,940 --> 00:46:59,400
I'm I'm framing it,

1279
00:46:59,700 --> 00:47:01,480
and so that's my initial

1280
00:47:02,099 --> 00:47:02,599
manipulation.

1281
00:47:03,460 --> 00:47:04,840
But I I

1282
00:47:05,140 --> 00:47:06,760
I'm not doing anything

1283
00:47:07,700 --> 00:47:08,840
to the data.

1284
00:47:09,140 --> 00:47:11,079
That's that's the bottom line.

1285
00:47:11,815 --> 00:47:14,315
If I start manipulating the data,

1286
00:47:15,414 --> 00:47:18,535
then then I I I'm making a terrible

1287
00:47:18,535 --> 00:47:21,094
mistake, and that's I try very hard. And

1288
00:47:21,094 --> 00:47:22,554
when if I do anything,

1289
00:47:23,335 --> 00:47:26,235
like remove a dust particle, for example,

1290
00:47:26,670 --> 00:47:29,889
I always indicate that I've done so. Always.

1291
00:47:30,429 --> 00:47:32,190
In sort of captions and that sort of

1292
00:47:32,190 --> 00:47:35,389
thing. Yes. Exact yeah. Absolute for example, I

1293
00:47:35,389 --> 00:47:36,510
have a book out,

1294
00:47:37,469 --> 00:47:40,315
a series of books called the visual elements,

1295
00:47:42,054 --> 00:47:45,034
communicating science and engineering, and the first element

1296
00:47:45,094 --> 00:47:46,154
is photography.

1297
00:47:47,255 --> 00:47:49,414
And I taught it's very it's a handbook,

1298
00:47:49,414 --> 00:47:51,539
and it's I've I'm told it's very good.

1299
00:47:51,780 --> 00:47:54,679
And so I say at the very beginning,

1300
00:47:55,059 --> 00:47:56,519
all of these images

1301
00:47:57,460 --> 00:47:58,440
have been digitally

1302
00:47:58,739 --> 00:47:59,239
enhanced

1303
00:48:00,900 --> 00:48:01,400
because

1304
00:48:01,860 --> 00:48:03,800
I I wanted you to pay attention

1305
00:48:04,574 --> 00:48:08,114
to the process, not necessarily the the the

1306
00:48:08,414 --> 00:48:08,914
distraction.

1307
00:48:10,255 --> 00:48:12,655
And so at the very beginning, I and

1308
00:48:12,655 --> 00:48:14,355
and when I say enhanced,

1309
00:48:14,655 --> 00:48:17,710
I'm talking about removing dust particles and which

1310
00:48:17,710 --> 00:48:18,210
really

1311
00:48:18,590 --> 00:48:20,190
but but you have to know that I've

1312
00:48:20,190 --> 00:48:20,930
done that.

1313
00:48:21,390 --> 00:48:23,710
So I always indicate if I if I

1314
00:48:23,710 --> 00:48:26,150
do anything like that. Okay. That's good. But

1315
00:48:26,269 --> 00:48:28,750
so it because you've you've touched on this

1316
00:48:28,750 --> 00:48:29,385
that there's

1317
00:48:29,945 --> 00:48:32,744
a concern about honesty, right, and and how

1318
00:48:32,744 --> 00:48:34,445
we ensure honesty.

1319
00:48:34,824 --> 00:48:36,905
At times, when you look around the world,

1320
00:48:36,905 --> 00:48:38,844
it feels like it's a runaway train

1321
00:48:39,224 --> 00:48:40,045
of dishonesty.

1322
00:48:40,744 --> 00:48:41,244
And

1323
00:48:41,784 --> 00:48:43,569
how we hold on to it, as you

1324
00:48:43,569 --> 00:48:45,489
say, it's difficult. But could you give us

1325
00:48:45,489 --> 00:48:48,049
some suggestions of how we Yeah. As as

1326
00:48:48,049 --> 00:48:49,730
the people who are who do care about

1327
00:48:49,730 --> 00:48:51,809
it, what do we do? Yeah. I I

1328
00:48:51,809 --> 00:48:52,309
think

1329
00:48:52,849 --> 00:48:53,989
now this might be

1330
00:48:54,609 --> 00:48:55,909
naive on my part.

1331
00:48:56,375 --> 00:48:59,034
But I think the more people understand

1332
00:48:59,335 --> 00:49:00,155
our process,

1333
00:49:01,335 --> 00:49:04,155
like, for example, how I make a particular

1334
00:49:04,375 --> 00:49:04,875
image,

1335
00:49:06,214 --> 00:49:07,994
the more we will engage

1336
00:49:08,454 --> 00:49:09,835
people to understand

1337
00:49:11,010 --> 00:49:13,429
what can be or what cannot be,

1338
00:49:13,809 --> 00:49:14,309
if

1339
00:49:14,690 --> 00:49:15,510
if if they

1340
00:49:15,890 --> 00:49:18,869
if they see if they understand, for example,

1341
00:49:19,570 --> 00:49:20,070
that

1342
00:49:20,530 --> 00:49:21,510
how how NASA

1343
00:49:22,784 --> 00:49:25,184
or the the James Webb people color, if

1344
00:49:25,184 --> 00:49:26,804
they actually see that,

1345
00:49:27,505 --> 00:49:29,844
then it's a new part of your thinking.

1346
00:49:30,224 --> 00:49:32,005
So for example, I'll give you a personal

1347
00:49:32,944 --> 00:49:35,284
experience. I used to be a a choral

1348
00:49:35,344 --> 00:49:37,824
singer. My voice is gone at this point.

1349
00:49:37,824 --> 00:49:39,320
That's very disturbing.

1350
00:49:40,019 --> 00:49:42,119
But I used to be a very serious

1351
00:49:42,180 --> 00:49:42,680
auditioned

1352
00:49:43,219 --> 00:49:44,119
choral singer.

1353
00:49:44,579 --> 00:49:45,079
Because

1354
00:49:45,380 --> 00:49:46,599
I know the music,

1355
00:49:47,860 --> 00:49:50,760
because I've sung this particular song piece,

1356
00:49:51,539 --> 00:49:53,880
when I hear another chorus sing it,

1357
00:49:54,474 --> 00:49:55,214
I'm engaged

1358
00:49:55,514 --> 00:49:58,094
more with it, and, actually, I can

1359
00:49:58,394 --> 00:50:00,795
sort of tell when something is not quite

1360
00:50:00,795 --> 00:50:04,414
right. It's because I know more about it.

1361
00:50:04,714 --> 00:50:06,494
Same with cooking, for example.

1362
00:50:06,929 --> 00:50:08,769
I'm I'm a very good cook, and I

1363
00:50:08,769 --> 00:50:09,909
could read a recipe

1364
00:50:10,769 --> 00:50:13,250
and know that, uh-uh, I'm not gonna do

1365
00:50:13,250 --> 00:50:16,769
this part because I'm experienced with it. I

1366
00:50:16,769 --> 00:50:20,150
maintain that engaging people in our process,

1367
00:50:21,234 --> 00:50:24,215
especially the, yeah, the next generation of researchers,

1368
00:50:24,514 --> 00:50:27,554
even if they're not making pictures, but engaging

1369
00:50:27,554 --> 00:50:28,695
them in the process,

1370
00:50:30,034 --> 00:50:30,775
I believe,

1371
00:50:31,394 --> 00:50:33,574
might push all of us into

1372
00:50:33,954 --> 00:50:34,454
understanding

1373
00:50:34,835 --> 00:50:36,135
what is not

1374
00:50:37,130 --> 00:50:38,489
right. What what do you think? Do you

1375
00:50:38,489 --> 00:50:40,650
think that's correct? I mean, no. I don't.

1376
00:50:40,650 --> 00:50:43,210
I I mean, it I'm a lecturer in

1377
00:50:43,210 --> 00:50:44,110
science communication.

1378
00:50:44,410 --> 00:50:47,309
Right? Oh. Yeah. It's my thing. So

1379
00:50:48,090 --> 00:50:50,170
you're preaching to the converted, but I I'm

1380
00:50:50,329 --> 00:50:52,125
you know, there's people listening who might not

1381
00:50:52,605 --> 00:50:55,344
agree. I'm always teaching about communicating the scientific

1382
00:50:55,405 --> 00:50:57,744
process and helping people to understand that process

1383
00:50:58,045 --> 00:51:00,304
is a really big part of science communication.

1384
00:51:00,445 --> 00:51:01,425
It's not about

1385
00:51:01,804 --> 00:51:04,364
scientific facts. You know, science at school quite

1386
00:51:04,364 --> 00:51:06,684
often is about learning facts, and there's a

1387
00:51:06,684 --> 00:51:07,744
bit about the process

1388
00:51:08,159 --> 00:51:11,119
because but, you know, the scientific process of

1389
00:51:11,119 --> 00:51:13,460
vaccines, if people understood how that was

1390
00:51:14,159 --> 00:51:15,139
those new vaccines

1391
00:51:15,920 --> 00:51:18,079
in inverted commas that came along during the

1392
00:51:18,079 --> 00:51:20,319
the pandemic, if people know the process knew

1393
00:51:20,319 --> 00:51:23,284
the process that had gone into producing those,

1394
00:51:23,344 --> 00:51:25,585
there wouldn't be as much fear, I think.

1395
00:51:25,585 --> 00:51:27,984
As much. I'm not saying it's it's it's

1396
00:51:27,984 --> 00:51:30,545
an all or nothing thing, but they're they're

1397
00:51:30,545 --> 00:51:33,045
engaged in thinking. This is the

1398
00:51:33,664 --> 00:51:36,484
ongoing issue that I did talk to colleagues

1399
00:51:36,625 --> 00:51:37,125
about.

1400
00:51:37,440 --> 00:51:39,940
People are not interested in thinking.

1401
00:51:40,800 --> 00:51:41,699
It's hard.

1402
00:51:42,559 --> 00:51:45,219
You know? And and they want quick answers.

1403
00:51:46,000 --> 00:51:48,000
I don't know how to engage people in

1404
00:51:48,000 --> 00:51:48,500
thinking.

1405
00:51:50,394 --> 00:51:52,155
There should be a way to make it

1406
00:51:52,155 --> 00:51:52,655
rewarding

1407
00:51:53,035 --> 00:51:56,015
to think. I mean, as scientists, we are,

1408
00:51:56,075 --> 00:51:58,394
and, you know, we're rewarded in the thinking

1409
00:51:58,394 --> 00:51:58,894
process.

1410
00:51:59,355 --> 00:52:00,815
But for the most part,

1411
00:52:01,755 --> 00:52:04,015
most people just want to be told

1412
00:52:04,849 --> 00:52:07,670
what to to choose either a or b

1413
00:52:08,130 --> 00:52:09,349
and not why.

1414
00:52:10,289 --> 00:52:12,550
My hope is if we start,

1415
00:52:13,090 --> 00:52:13,829
for example,

1416
00:52:14,130 --> 00:52:16,369
on a on a simple level, if we

1417
00:52:16,369 --> 00:52:17,910
start creating visuals

1418
00:52:18,530 --> 00:52:19,750
that are engaging

1419
00:52:20,585 --> 00:52:21,804
and gives permission

1420
00:52:22,105 --> 00:52:22,844
to people

1421
00:52:23,224 --> 00:52:24,684
who to ask questions.

1422
00:52:25,784 --> 00:52:28,045
You people are not frightened of images.

1423
00:52:28,904 --> 00:52:31,484
It's a means of engaging them to ask

1424
00:52:31,625 --> 00:52:32,125
questions.

1425
00:52:32,820 --> 00:52:34,519
And once you get them,

1426
00:52:35,219 --> 00:52:36,519
frankly, it's a seduction

1427
00:52:37,619 --> 00:52:40,579
to ask a question. My thinking is that

1428
00:52:40,579 --> 00:52:41,880
there's a next step

1429
00:52:42,500 --> 00:52:44,820
so that, for example, I'm coming out with

1430
00:52:44,820 --> 00:52:46,599
a young adult's book in

1431
00:52:47,025 --> 00:52:48,085
in the fall.

1432
00:52:48,944 --> 00:52:50,005
It's for teenagers.

1433
00:52:50,304 --> 00:52:52,164
It's called Phenomenal Moments.

1434
00:52:52,944 --> 00:52:55,984
And the idea is that everything around us

1435
00:52:55,984 --> 00:52:56,724
is science,

1436
00:52:57,184 --> 00:53:00,304
period. That's it. Everything we look at, everything

1437
00:53:00,304 --> 00:53:02,005
we touch is about science.

1438
00:53:02,699 --> 00:53:05,280
So the whole book is about everyday

1439
00:53:05,579 --> 00:53:06,079
phenomena,

1440
00:53:07,179 --> 00:53:09,260
and the pictures, I'd like to think, are

1441
00:53:09,260 --> 00:53:11,179
beautiful, but you can't you don't know what

1442
00:53:11,179 --> 00:53:13,199
they are. It's sort of a guessing game.

1443
00:53:13,820 --> 00:53:14,320
And

1444
00:53:15,385 --> 00:53:16,845
the my hope is that

1445
00:53:17,305 --> 00:53:20,045
they the kids will look at the picture.

1446
00:53:20,585 --> 00:53:22,505
They'll see the caption about what it is

1447
00:53:22,505 --> 00:53:23,485
that it is.

1448
00:53:23,864 --> 00:53:26,505
And then when they start walking through the

1449
00:53:26,505 --> 00:53:29,819
park one day, they're gonna see something like

1450
00:53:29,819 --> 00:53:31,920
what they just saw in the book.

1451
00:53:32,299 --> 00:53:35,359
The picture I'm making is engaging them

1452
00:53:36,539 --> 00:53:37,039
to

1453
00:53:37,420 --> 00:53:37,920
remember,

1454
00:53:38,299 --> 00:53:40,319
perhaps, when they see it again,

1455
00:53:41,579 --> 00:53:43,454
they're gonna know what that what is

1456
00:53:43,934 --> 00:53:44,994
because the picture

1457
00:53:45,454 --> 00:53:46,355
is a means

1458
00:53:46,894 --> 00:53:48,755
of getting them interested.

1459
00:53:49,454 --> 00:53:52,015
It's very simple. I mean, I'm not doing

1460
00:53:52,015 --> 00:53:53,875
anything brilliant here, but

1461
00:53:54,255 --> 00:53:56,035
but I think that with pictures,

1462
00:53:56,500 --> 00:53:58,980
we can we can get more people to

1463
00:53:58,980 --> 00:54:01,000
start looking and thinking about.

1464
00:54:01,300 --> 00:54:03,960
Absolutely. And but in that sort of use,

1465
00:54:04,260 --> 00:54:06,739
would you see AI being a useful tool

1466
00:54:06,739 --> 00:54:10,280
in in generating Yeah. Well yeah. Oh, boy.

1467
00:54:12,074 --> 00:54:12,574
Yes.

1468
00:54:12,954 --> 00:54:14,734
Yes. I think it can be

1469
00:54:15,034 --> 00:54:18,394
as long as we it is indicated that

1470
00:54:18,394 --> 00:54:19,775
this image was

1471
00:54:20,394 --> 00:54:21,614
done with AI.

1472
00:54:22,554 --> 00:54:23,855
That's the primary.

1473
00:54:25,099 --> 00:54:26,480
And whether or not

1474
00:54:26,780 --> 00:54:29,579
we could get the AI image maker to

1475
00:54:29,579 --> 00:54:31,739
do that is a whole other thing. I

1476
00:54:31,739 --> 00:54:33,099
don't I don't know. I don't know how

1477
00:54:33,099 --> 00:54:34,460
to do that. Yeah. I think it's a

1478
00:54:34,460 --> 00:54:36,320
brilliant example, the NASA images.

1479
00:54:37,114 --> 00:54:39,755
You know, how if we can understand how

1480
00:54:39,755 --> 00:54:42,554
that works, then there's there's an there's another

1481
00:54:42,554 --> 00:54:44,315
level of interest in that for me. I'm

1482
00:54:44,315 --> 00:54:46,315
sort of okay. So how have they colored

1483
00:54:46,315 --> 00:54:48,394
those images? And there's there's a there's a

1484
00:54:48,394 --> 00:54:50,409
level of intrigue about AI, which I think

1485
00:54:50,409 --> 00:54:53,050
disappears in a few years when everything's it's

1486
00:54:53,050 --> 00:54:55,050
just gonna be a thing. But, there's a

1487
00:54:55,050 --> 00:54:56,809
there's a sort of, oh, that one's created

1488
00:54:56,809 --> 00:54:58,269
by AI. That's interesting.

1489
00:54:58,969 --> 00:55:00,969
Or this one's created using that piece of

1490
00:55:00,969 --> 00:55:02,809
software. That's interesting. I don't I didn't know

1491
00:55:02,809 --> 00:55:04,570
about that. And it adds a it adds

1492
00:55:04,570 --> 00:55:05,869
a level to it. But I

1493
00:55:06,224 --> 00:55:08,545
I I can totally see that. What I

1494
00:55:08,545 --> 00:55:10,545
struggle with, and this is slightly other to

1495
00:55:10,545 --> 00:55:13,025
this, you know, I can't imagine logging into

1496
00:55:13,025 --> 00:55:13,525
Netflix

1497
00:55:14,065 --> 00:55:14,805
and going,

1498
00:55:15,664 --> 00:55:17,664
which one of these were created by AI?

1499
00:55:17,664 --> 00:55:19,184
That's the one I'm gonna sit down and

1500
00:55:19,184 --> 00:55:20,704
watch tonight. You know, I want like, you're

1501
00:55:20,704 --> 00:55:21,765
talking about music,

1502
00:55:22,359 --> 00:55:22,859
individual

1503
00:55:23,159 --> 00:55:24,859
differences between choirs.

1504
00:55:25,400 --> 00:55:28,139
And then there's, you know, the the nuances

1505
00:55:28,280 --> 00:55:30,380
of actors. There's the nuances of

1506
00:55:30,679 --> 00:55:32,139
photographers. There's the nuances

1507
00:55:32,599 --> 00:55:35,019
of filmmakers. There's the nuances of artists,

1508
00:55:35,400 --> 00:55:36,699
individual artists. And

1509
00:55:37,000 --> 00:55:39,385
and I I I worry that

1510
00:55:39,764 --> 00:55:40,505
if we

1511
00:55:40,885 --> 00:55:44,264
allow AI in image creation at any level,

1512
00:55:44,324 --> 00:55:45,784
then we we end up.

1513
00:55:46,244 --> 00:55:46,744
Yeah.

1514
00:55:48,324 --> 00:55:48,824
Yeah.

1515
00:55:49,925 --> 00:55:51,304
You're right. I mean,

1516
00:55:52,070 --> 00:55:54,650
the question is, will there be a time

1517
00:55:55,670 --> 00:55:58,329
when we will see that AI really

1518
00:55:58,710 --> 00:55:59,530
is missing

1519
00:56:00,949 --> 00:56:01,929
that creativity

1520
00:56:02,389 --> 00:56:04,730
that only a human can bring?

1521
00:56:05,905 --> 00:56:08,385
It let me quickly go back to the

1522
00:56:08,385 --> 00:56:10,085
image that I was talking about,

1523
00:56:10,464 --> 00:56:11,764
the AI image

1524
00:56:12,065 --> 00:56:12,565
that,

1525
00:56:14,464 --> 00:56:15,605
Dali created

1526
00:56:16,065 --> 00:56:17,284
of these vials.

1527
00:56:19,099 --> 00:56:21,019
As I said, it was very cartoon like.

1528
00:56:21,019 --> 00:56:22,880
There were all kinds of mistakes. But,

1529
00:56:23,660 --> 00:56:24,160
interestingly,

1530
00:56:25,420 --> 00:56:28,400
the the model created little dots,

1531
00:56:29,820 --> 00:56:30,719
in the vials.

1532
00:56:31,194 --> 00:56:33,674
And the model put a couple of the

1533
00:56:33,674 --> 00:56:34,174
dots

1534
00:56:34,875 --> 00:56:36,494
on the surface of the table.

1535
00:56:36,795 --> 00:56:38,894
That was an aesthetic decision

1536
00:56:40,315 --> 00:56:41,214
that AI

1537
00:56:41,914 --> 00:56:43,110
decided to do.

1538
00:56:44,070 --> 00:56:45,050
It was stupid.

1539
00:56:45,750 --> 00:56:46,730
It was silly.

1540
00:56:47,750 --> 00:56:48,250
Maybe

1541
00:56:49,430 --> 00:56:51,610
the machine will never be able

1542
00:56:52,710 --> 00:56:54,170
to be as creative

1543
00:56:54,470 --> 00:56:57,085
as we. I I actually don't know. I

1544
00:56:57,085 --> 00:56:59,344
mean, I don't know enough about that world.

1545
00:56:59,885 --> 00:57:01,965
And maybe, you know, what I should do

1546
00:57:01,965 --> 00:57:03,885
is talk to people who do know. Well,

1547
00:57:03,885 --> 00:57:04,545
I did.

1548
00:57:04,925 --> 00:57:06,844
I did before I wrote the article, and

1549
00:57:06,844 --> 00:57:08,605
they all agree we have to have some

1550
00:57:08,605 --> 00:57:10,065
kind of guardrails.

1551
00:57:10,364 --> 00:57:13,109
That's that's a that's a done deal. But

1552
00:57:13,109 --> 00:57:15,109
the question that you're asking, which is a

1553
00:57:15,109 --> 00:57:17,529
very important question, is will AI

1554
00:57:18,630 --> 00:57:20,489
be as creative as a human,

1555
00:57:21,589 --> 00:57:23,929
and can we discern that difference?

1556
00:57:25,605 --> 00:57:28,105
I think at at least at this point,

1557
00:57:28,405 --> 00:57:30,485
I don't think it come it will come

1558
00:57:30,724 --> 00:57:31,385
I think

1559
00:57:31,765 --> 00:57:35,224
there's always something that the human can do

1560
00:57:36,324 --> 00:57:39,519
that AI now can't do. But will it

1561
00:57:39,519 --> 00:57:40,659
happen in the future?

1562
00:57:42,639 --> 00:57:43,699
Probably, yes.

1563
00:57:45,519 --> 00:57:47,539
So there goes the answer to that.

1564
00:57:49,279 --> 00:57:50,960
We spoke to Tony Hay earlier in the

1565
00:57:50,960 --> 00:57:53,359
podcast, and I wonder who he thought should

1566
00:57:53,359 --> 00:57:55,114
read the IOP's

1567
00:57:55,494 --> 00:57:56,315
white paper

1568
00:57:56,775 --> 00:57:59,175
on AI and physics. Oh, I think it's

1569
00:57:59,175 --> 00:58:00,695
got a a a lot of things that

1570
00:58:00,695 --> 00:58:02,315
I absolutely agree with.

1571
00:58:04,054 --> 00:58:06,135
It is very physics oriented. It says, you

1572
00:58:06,135 --> 00:58:07,974
know, physics is the only field with large,

1573
00:58:07,974 --> 00:58:10,829
well curated datasets and theories. Well, there are

1574
00:58:10,829 --> 00:58:13,950
things like chemistry, possibly biology, material science, a

1575
00:58:13,950 --> 00:58:15,950
few other things who might object to that

1576
00:58:15,950 --> 00:58:17,809
statement. But but but,

1577
00:58:19,869 --> 00:58:21,869
no. It's it's it's it's a good thing

1578
00:58:21,869 --> 00:58:22,369
to

1579
00:58:22,864 --> 00:58:25,444
to to galvanize the community that,

1580
00:58:27,184 --> 00:58:28,864
AI is going to be in their future,

1581
00:58:28,864 --> 00:58:30,324
and I I do see that

1582
00:58:30,625 --> 00:58:32,885
that that you could make an AI assistant,

1583
00:58:33,105 --> 00:58:34,804
which was really very effective

1584
00:58:35,909 --> 00:58:37,989
in advising you in your day to day

1585
00:58:37,989 --> 00:58:39,289
job and in your work,

1586
00:58:39,909 --> 00:58:41,210
as a as a physicist.

1587
00:58:41,510 --> 00:58:43,269
So I I I think that with the

1588
00:58:43,269 --> 00:58:46,570
large language models, you will find that there

1589
00:58:48,469 --> 00:58:49,609
are AI assistants

1590
00:58:50,414 --> 00:58:53,215
for physics, and that will be part of

1591
00:58:53,215 --> 00:58:54,914
many people's lives. Alright?

1592
00:58:55,775 --> 00:58:59,055
And I think people understanding the strengths and

1593
00:58:59,055 --> 00:59:00,275
weaknesses of it

1594
00:59:02,190 --> 00:59:03,630
and the fact that you need to do

1595
00:59:03,630 --> 00:59:05,809
skills. And so the the things it recommends

1596
00:59:05,869 --> 00:59:07,949
are very sensible things, and it's good that

1597
00:59:07,949 --> 00:59:10,449
the the physics community is aware

1598
00:59:10,750 --> 00:59:11,570
of the potential.

1599
00:59:12,269 --> 00:59:13,889
But I did, for example,

1600
00:59:14,355 --> 00:59:17,554
very much approve of the chemistry Nobel Prize

1601
00:59:17,554 --> 00:59:18,215
this year,

1602
00:59:18,675 --> 00:59:19,175
which

1603
00:59:20,994 --> 00:59:22,614
was awarded for for

1604
00:59:22,994 --> 00:59:23,974
protein folding

1605
00:59:26,034 --> 00:59:27,574
and gave it to DeepMind

1606
00:59:28,034 --> 00:59:29,175
two people from DeepMind,

1607
00:59:29,660 --> 00:59:31,820
and also a colleague of mine from University

1608
00:59:31,820 --> 00:59:33,660
of Washington where I used to have a

1609
00:59:33,660 --> 00:59:34,559
joint position,

1610
00:59:34,860 --> 00:59:37,200
David Baker, who's been doing it for years.

1611
00:59:37,420 --> 00:59:39,200
And I think that was a good thing,

1612
00:59:39,260 --> 00:59:39,760
and

1613
00:59:40,219 --> 00:59:42,160
and will actually have ramifications,

1614
00:59:43,605 --> 00:59:47,125
huge ramifications in all sorts of omics type

1615
00:59:47,125 --> 00:59:47,625
stuff.

1616
00:59:48,565 --> 00:59:50,885
So so I see that there's really exciting

1617
00:59:50,885 --> 00:59:53,684
applications in that area, drugs and cures the

1618
00:59:53,684 --> 00:59:56,164
diseases and things like that. Physics is less

1619
00:59:56,164 --> 00:59:57,224
clear. I mean,

1620
00:59:59,180 --> 01:00:01,660
material science is probably the major hope that

1621
01:00:01,660 --> 01:00:03,840
you actually will find something really exciting,

1622
01:00:04,619 --> 01:00:05,840
and that would be

1623
01:00:06,539 --> 01:00:07,680
a good thing to do.

1624
01:00:07,980 --> 01:00:08,880
If you did

1625
01:00:09,980 --> 01:00:10,480
just

1626
01:00:10,860 --> 01:00:11,360
program,

1627
01:00:11,820 --> 01:00:12,320
inform

1628
01:00:12,619 --> 01:00:14,240
the larger language model

1629
01:00:14,835 --> 01:00:18,114
with physics information. Right? All it knew was

1630
01:00:18,114 --> 01:00:19,175
the physics information.

1631
01:00:19,474 --> 01:00:21,715
Would that make a really good physicist, or

1632
01:00:21,715 --> 01:00:23,474
does it do you need other things as

1633
01:00:23,474 --> 01:00:25,715
well? That those no. That that's that's the

1634
01:00:25,715 --> 01:00:28,275
sort of question that's interesting me. And and

1635
01:00:28,275 --> 01:00:28,949
you see,

1636
01:00:30,070 --> 01:00:32,630
we we use deep learning on to help

1637
01:00:32,630 --> 01:00:34,550
analyze some of the data at the lab,

1638
01:00:34,550 --> 01:00:35,849
which is usually

1639
01:00:36,150 --> 01:00:38,550
in in images. And if for images, it's

1640
01:00:38,550 --> 01:00:40,949
it's ideal, except that you have to train

1641
01:00:40,949 --> 01:00:43,375
it on ground truth. And once you've trained

1642
01:00:43,375 --> 01:00:45,054
it on ground truth where you know the

1643
01:00:45,054 --> 01:00:45,554
answer,

1644
01:00:46,255 --> 01:00:48,335
it can then go and see data it

1645
01:00:48,335 --> 01:00:50,755
hasn't seen before and make decisions.

1646
01:00:53,134 --> 01:00:56,275
If you take something like quantum computing, right,

1647
01:00:56,760 --> 01:00:58,699
which is something I care about. And

1648
01:00:59,320 --> 01:01:01,640
at this point, I advertise my lectures with

1649
01:01:01,640 --> 01:01:02,140
Feynman,

1650
01:01:02,599 --> 01:01:05,800
the Feynman lectures on computation. The question is

1651
01:01:05,800 --> 01:01:06,699
we do simulations

1652
01:01:07,400 --> 01:01:08,380
at all different

1653
01:01:08,864 --> 01:01:11,105
scales, but we don't do a fully quantum

1654
01:01:11,105 --> 01:01:12,085
mechanical simulation

1655
01:01:12,625 --> 01:01:13,125
with

1656
01:01:14,864 --> 01:01:16,945
with everything in in terms of the Hilbert

1657
01:01:16,945 --> 01:01:18,644
space just grows exponentially,

1658
01:01:19,425 --> 01:01:22,385
and and that computers don't have enough memory

1659
01:01:22,385 --> 01:01:24,150
to handle more than the, you know, small

1660
01:01:24,150 --> 01:01:26,469
number of electrons and things like that, a

1661
01:01:26,469 --> 01:01:28,710
handful. But quantum computing can do it very

1662
01:01:28,710 --> 01:01:30,949
easily, and it grows linearly, and it's it's

1663
01:01:30,949 --> 01:01:32,789
it's much easier to do a big system

1664
01:01:32,789 --> 01:01:35,050
there. Question is, what would be different

1665
01:01:35,670 --> 01:01:37,530
by doing it that? Because we've done

1666
01:01:38,255 --> 01:01:39,074
a whole range

1667
01:01:39,695 --> 01:01:42,735
of modeling at various levels, which are, yes,

1668
01:01:42,735 --> 01:01:43,235
approximations,

1669
01:01:43,934 --> 01:01:46,255
but they're are we going to find really

1670
01:01:46,255 --> 01:01:48,494
some new things by doing that? And that's

1671
01:01:48,494 --> 01:01:50,690
an interesting question. But but

1672
01:01:52,349 --> 01:01:54,429
and people talk about, oh, this wonderful stuff

1673
01:01:54,429 --> 01:01:54,929
on

1674
01:01:55,309 --> 01:01:57,889
quantum machine learning. Now I don't know

1675
01:01:58,670 --> 01:01:59,650
of any machine,

1676
01:02:01,389 --> 01:02:01,889
quantum

1677
01:02:02,525 --> 01:02:03,025
computer

1678
01:02:03,325 --> 01:02:06,684
that can really do serious calculations yet. I

1679
01:02:06,684 --> 01:02:07,985
think it's getting close,

1680
01:02:08,684 --> 01:02:10,224
closer than I thought it would.

1681
01:02:10,925 --> 01:02:11,425
And

1682
01:02:12,605 --> 01:02:13,505
I don't know

1683
01:02:13,965 --> 01:02:16,684
whether quantum machine learning or quantum AI, if

1684
01:02:16,684 --> 01:02:17,190
you like,

1685
01:02:17,670 --> 01:02:19,510
is a real real thing. But that's that's

1686
01:02:19,510 --> 01:02:20,889
something that that, you

1687
01:02:21,269 --> 01:02:23,190
know, smart people should look at and young

1688
01:02:23,190 --> 01:02:25,609
kids could could find really that an interesting

1689
01:02:25,670 --> 01:02:26,489
thing to do.

1690
01:02:26,949 --> 01:02:27,449
And

1691
01:02:27,829 --> 01:02:29,050
and then you could also

1692
01:02:30,389 --> 01:02:32,394
that's starting with the physics model.

1693
01:02:32,775 --> 01:02:34,954
And then the question is,

1694
01:02:35,335 --> 01:02:37,335
does it make any difference whether you've trained

1695
01:02:37,335 --> 01:02:37,994
it on,

1696
01:02:38,295 --> 01:02:41,114
you know, Wikipedia and Reddit or scientific literature?

1697
01:02:41,335 --> 01:02:43,275
Again, I won't make. Yeah.

1698
01:02:43,610 --> 01:02:46,090
Yeah. Okay. Well, maybe we'll find out in

1699
01:02:46,090 --> 01:02:48,650
a future episode of the podcast at some

1700
01:02:48,650 --> 01:02:50,170
point. But No. That would that would be

1701
01:02:50,170 --> 01:02:52,329
nice. Yes. Yeah. There's no it's an exciting

1702
01:02:52,329 --> 01:02:52,829
time.

1703
01:02:53,449 --> 01:02:56,030
I'd like to thank Tony Hay, Felice Frankel,

1704
01:02:56,255 --> 01:02:58,574
and Caterina Dolioni for talking to me for

1705
01:02:58,574 --> 01:03:01,235
this episode of the Physics World Stories podcast.

1706
01:03:01,535 --> 01:03:03,775
You can find links to their work and,

1707
01:03:03,775 --> 01:03:06,094
of course, the IOP's white paper on AI

1708
01:03:06,094 --> 01:03:08,914
and physics on physicsworld.com.

1709
01:03:09,329 --> 01:03:11,250
We'll be back soon with something else from

1710
01:03:11,250 --> 01:03:13,650
this wonderful world of physics, and thank you

1711
01:03:13,650 --> 01:03:14,950
very much for listening.