1
00:00:00,960 --> 00:00:03,277
is AI.

2
00:00:03,360 --> 00:00:07,117
>> Oh yeah. Oh yeah.

3
00:00:07,200 --> 00:00:10,557
And do we have a panel for you?

4
00:00:10,640 --> 00:00:12,957
Uh let me

5
00:00:13,040 --> 00:00:15,917
Oh, I just as a point of disclosure.

6
00:00:16,000 --> 00:00:19,597
Uh just I am

7
00:00:19,680 --> 00:00:21,597
in the

8
00:00:21,680 --> 00:00:25,597
class action suit against Anthropic for

9
00:00:25,680 --> 00:00:28,477
pirating books.

10
00:00:28,560 --> 00:00:31,917
Uh there's like nine of my books

11
00:00:32,000 --> 00:00:34,477
that were pirated and used to train

12
00:00:34,560 --> 00:00:37,437
their AI models. Uh so I just letting

13
00:00:37,520 --> 00:00:39,757
you know that in advance that there's

14
00:00:39,840 --> 00:00:42,157
what's going on.

15
00:00:42,240 --> 00:00:46,477
So uh let's get this party started. Uh

16
00:00:46,560 --> 00:00:50,317
my first panelist this evening is a

17
00:00:50,400 --> 00:00:54,237
professor of technology and government

18
00:00:54,320 --> 00:00:59,117
at Harvard University. Please welcome

19
00:00:59,200 --> 00:01:03,960
Latana Sweeney. Latana, come on in.

20
00:01:12,800 --> 00:01:15,677
Lata, your last name is Sweeney.
That sounds very Irish.

21
00:01:15,760 --> 00:01:18,237
>> Yes. >> Is Irish?

22
00:01:18,320 --> 00:01:24,920
>> Yes. So, happy day. Okay.

23
00:01:28,000 --> 00:01:33,757
Next, uh, we've got pro my glasses here.

24
00:01:33,840 --> 00:01:36,157
>> Yeah. Uh, a computer a professor of

25
00:01:36,240 --> 00:01:38,237
computer and information science at the

26
00:01:38,320 --> 00:01:41,197
University of Pennsylvania. Uh, please

27
00:01:41,280 --> 00:01:45,360
help me welcome Chris Callison Burch.

28
00:01:54,000 --> 00:01:57,917
I've got an associate professor of

29
00:01:58,000 --> 00:02:00,397
math. No, associate professor of

30
00:02:00,480 --> 00:02:02,717
statistics at Columbia University who

31
00:02:02,800 --> 00:02:05,117
specializes in the foundations of

32
00:02:05,200 --> 00:02:07,677
machine learning. Uh, help me welcome

33
00:02:07,760 --> 00:02:12,360
Cindy Rush. Columbia University.

34
00:02:18,000 --> 00:02:19,997
So, City, you were not held up by TSA

35
00:02:20,080 --> 00:02:22,477
today because you just came down to

36
00:02:22,560 --> 00:02:25,197
103rd Street in Broadway. Okay. Thank

37
00:02:25,280 --> 00:02:28,237
you for taking the trip here.

38
00:02:28,320 --> 00:02:30,237
Next, I have the president of the

39
00:02:30,320 --> 00:02:32,557
Machine Intelligence Institute of

40
00:02:32,640 --> 00:02:39,480
Berkeley, California,
Nate Sores. Nate, come on out.

41
00:02:46,240 --> 00:02:48,557
uh last on this panel.

42
00:02:48,640 --> 00:02:50,477
>> Normally we have
five on our panel. This

43
00:02:50,560 --> 00:02:54,317
year we have six just given the breadth.

44
00:02:54,400 --> 00:02:58,557
>> You forgot forgot one.

45
00:02:58,640 --> 00:03:02,637
>> You missed one. >> Yeah.
Still I haven't gotten it yet.

46
00:03:02,720 --> 00:03:06,557
>> One more. >> One more
in addition to Eric. Did I miss one?

47
00:03:06,640 --> 00:03:11,597
>> Yes. >> You have a blank chair there.
The Yeah, the person sit right here.

48
00:03:11,680 --> 00:03:13,357
>> How'd that happen?

49
00:03:13,440 --> 00:03:16,397
>> Neil, did you need do you need these?

50
00:03:16,480 --> 00:03:20,360
>> Thank you for that.

51
00:03:21,920 --> 00:03:27,517
>> Hang on. There it is. Thank you.

52
00:03:27,600 --> 00:03:29,917
>> Do over.

53
00:03:30,000 --> 00:03:32,797
Uh, we have, believe it or not, such a

54
00:03:32,880 --> 00:03:35,437
title exists. We have a distinguished

55
00:03:35,520 --> 00:03:40,077
professor of AI at the University
of Southern California. Join me in

56
00:03:40,160 --> 00:03:45,160
welcoming Kate Crawford.
Kate, come on out.

57
00:03:51,840 --> 00:03:55,357
Professor of AI that already exists.

58
00:03:55,440 --> 00:04:01,960
That's kind of spooky,
actually. Uh, and last we have

59
00:04:04,000 --> 00:04:07,597
the former CEO of Google,

60
00:04:07,680 --> 00:04:12,200
>> Eric Schmidt. Come on out, Eric.

61
00:04:22,480 --> 00:04:26,637
>> Hello, Eric.
I'm going to start with you.

62
00:04:26,720 --> 00:04:32,077
You We're blaming you for everything.

63
00:04:32,160 --> 00:04:35,117
I'm used to it. >> You're used to it.
Okay. Uh you've been

64
00:04:35,200 --> 00:04:37,597
at Google like from very early days when

65
00:04:37,680 --> 00:04:40,157
most of us remembered it simply as a

66
00:04:40,240 --> 00:04:43,597
search engine and the reach and grip has

67
00:04:43,680 --> 00:04:47,437
expanded and

68
00:04:47,520 --> 00:04:52,637
some impressive things occurred at the
time. Did you foresee AI as a major part

69
00:04:52,720 --> 00:04:55,997
of what would be shaping
the world back when you began?

70
00:04:56,080 --> 00:05:00,317
>> Well, we we knew AI would matter.
I don't think it was until 2011 when we

71
00:05:00,400 --> 00:05:05,357
started working on essentially
supervised supervised finetuning that it

72
00:05:05,440 --> 00:05:06,957
really became clear and then the

73
00:05:07,040 --> 00:05:10,797
transformer paper in 2017 plus the alpho

74
00:05:10,880 --> 00:05:14,157
win showed the power.
>> Now if I remember correctly Google

75
00:05:14,240 --> 00:05:16,797
bought an AI company to assist this.

76
00:05:16,880 --> 00:05:19,837
Isn't that correct? You didn't
that wasn't homegrown was it?

77
00:05:19,920 --> 00:05:22,477
>> That's right.
We actually bought a whole

78
00:05:22,560 --> 00:05:25,437
bunch of these little startups
at the time back when the valuations were

79
00:05:25,520 --> 00:05:29,197
reasonable. Um, >> oh,
you didn't pay a trillion dollars

80
00:05:29,280 --> 00:05:34,077
for the companies. >> And, uh,
the Deep Mind Acquisition was

81
00:05:34,160 --> 00:05:36,237
probably the most important.
>> That's the British company.

82
00:05:36,320 --> 00:05:39,677
>> British company. Yes.
>> And that's the one that now is driving

83
00:05:39,760 --> 00:05:43,357
at Google anyway. I don't work there
anymore. All of the sort of core AI

84
00:05:43,440 --> 00:05:45,357
work.

85
00:05:45,440 --> 00:05:50,557
>> Nate, you co-authored a book titled If

86
00:05:50,640 --> 00:05:56,477
Anybody Builds It, Everyone Dies.

87
00:05:56,560 --> 00:05:59,837
I hope it's wrong. >> You hope
it's wrong. I I hope you have a

88
00:05:59,920 --> 00:06:03,117
little more than hope. Uh what is the it?

89
00:06:03,200 --> 00:06:07,277
>> The it here is machine super
intelligence. So this is AI that is

90
00:06:07,360 --> 00:06:11,997
smarter than the smartest humans
at every mental task. Better than the best

91
00:06:12,080 --> 00:06:16,877
humans at every mental task. That sort
of AI doesn't exist yet, but it is what

92
00:06:16,960 --> 00:06:19,837
these companies say they're racing
towards. you know, they're saying, you

93
00:06:19,920 --> 00:06:23,917
know, we want to try and make AI smarter
than Einstein such that you can run a

94
00:06:24,000 --> 00:06:27,277
million of them in a data center much
faster than a human. That's what they're

95
00:06:27,360 --> 00:06:30,797
headed towards. A lot of people

96
00:06:30,880 --> 00:06:36,077
think that the the big danger with AI is
what if we hand them guns >> and you know,

97
00:06:36,160 --> 00:06:38,237
>> what if we do what?
>> What if we hand them weapons?

98
00:06:38,320 --> 00:06:41,437
>> And the weapons. Yeah.
>> And there is, you know, issues there

99
00:06:41,520 --> 00:06:45,917
that that people need to talk about.
But also, humanity is not a dangerous

100
00:06:46,000 --> 00:06:49,917
species because somebody
else gave us guns.

101
00:06:50,000 --> 00:06:54,077
Humanity is the sort of species where
if you dump 10,000 humans naked in the

102
00:06:54,160 --> 00:06:56,557
savannah, they bootstrap their way to

103
00:06:56,640 --> 00:07:00,157
nuclear weapons with their bare hands.

104
00:07:00,240 --> 00:07:03,997
That is the ability that is extremely
dangerous to automate and that's what

105
00:07:04,080 --> 00:07:06,317
these companies are rushing towards.

106
00:07:06,400 --> 00:07:10,237
>> Thanks for that happy thought. Okay,

107
00:07:10,320 --> 00:07:11,837
>> I'm going cool it down a little bit.

108
00:07:11,920 --> 00:07:15,037
>> Yeah. Yeah. Cindy uh you you are early

109
00:07:15,120 --> 00:07:20,237
in an attempt to understand machine
language learn statistician and

110
00:07:20,320 --> 00:07:21,917
everything I know about the large

111
00:07:22,000 --> 00:07:25,117
language models they're creating

112
00:07:25,200 --> 00:07:28,797
information statistically based on the

113
00:07:28,880 --> 00:07:30,237
information that's sitting out there

114
00:07:30,320 --> 00:07:33,357
waiting to be mined. So would you agree

115
00:07:33,440 --> 00:07:35,917
that that that

116
00:07:36,000 --> 00:07:38,317
super intelligence is a real thing and

117
00:07:38,400 --> 00:07:45,240
that it's accessible and that even if
we achieve it is going to kill us all?

118
00:07:47,280 --> 00:07:50,877
>> I mean um you know anything seems

119
00:07:50,960 --> 00:07:55,437
possible right? Um you know we're
learning from data. There's tons of data

120
00:07:55,520 --> 00:08:01,917
out there and you know the the progress
is moving steadily forward. Um >> forward

121
00:08:02,000 --> 00:08:04,237
Forward. >> This is forward. >> Yeah.

122
00:08:04,320 --> 00:08:08,557
>> Forward. Okay. >> I don't
know if it'll kill us, you know.

123
00:08:08,640 --> 00:08:11,037
Um >> ourselves,

124
00:08:11,120 --> 00:08:13,757
>> but you're you're deep into sort of the

125
00:08:13,840 --> 00:08:17,117
programming of how all this works. So,

126
00:08:17,200 --> 00:08:19,437
dare I say, you're in the neurology of

127
00:08:19,520 --> 00:08:22,717
it. Do you see super intelligence as

128
00:08:22,800 --> 00:08:26,397
neurologically realistic,

129
00:08:26,480 --> 00:08:29,677
computationally realistic? I I mean okay

130
00:08:29,760 --> 00:08:32,237
at the end of the day all of this is

131
00:08:32,320 --> 00:08:34,237
just a mathematical equation right

132
00:08:34,320 --> 00:08:37,277
>> right at the end of our species

133
00:08:37,360 --> 00:08:41,117
>> when you ask um you know chat GPT a

134
00:08:41,200 --> 00:08:43,757
question right there's a a mathematical

135
00:08:43,840 --> 00:08:46,717
equation that maps from you know your

136
00:08:46,800 --> 00:08:51,117
question to the the output right and you

137
00:08:51,200 --> 00:08:55,997
know um yeah it's a little
bit I'm a little bit

138
00:08:56,080 --> 00:08:59,997
cool on it in the sense that you you know,
at the end it's it's still math at

139
00:09:00,080 --> 00:09:02,797
the bottom. So, it's hard for me to, you

140
00:09:02,880 --> 00:09:06,957
know, um kind of see this becoming some

141
00:09:07,040 --> 00:09:11,517
something so dangerous and bad, at least
in the short term, because there's a lot

142
00:09:11,600 --> 00:09:14,237
a lot of steps. >> So,
what you're saying is the math in

143
00:09:14,320 --> 00:09:17,197
the end might save us because at its

144
00:09:17,280 --> 00:09:21,757
base level, it's just simply making
statistical decisions. In the same way,

145
00:09:21,840 --> 00:09:25,597
at the base level of our mind,
they're just neurosaptic firings.

146
00:09:25,680 --> 00:09:29,677
>> Yeah. You know, um what I've always
said is if you want to stay safe, get a

147
00:09:29,760 --> 00:09:33,197
statistician um you
know beside you, right?

148
00:09:33,280 --> 00:09:37,597
>> Statisticians keep us safe. Is that
that? Okay. Well, keep that in mind. But

149
00:09:37,680 --> 00:09:41,837
but it'll keep us safe. You'll let us
know exactly our likelihood of dying.

150
00:09:41,920 --> 00:09:43,197
That's not the same thing as keeping us

151
00:09:43,280 --> 00:09:47,037
safe. I don't think

152
00:09:47,120 --> 00:09:51,597
I Chris. >> Hey, dude.

153
00:09:51,680 --> 00:09:55,517
>> Uh what you you've
been tracking AI for a

154
00:09:55,600 --> 00:09:59,517
while. You teach a class, one of the
more popular classes at University of

155
00:09:59,600 --> 00:10:01,597
Pennsylvania. How many
people are in your AI class?

156
00:10:01,680 --> 00:10:04,957
>> Uh, last semester I had 650 students.

157
00:10:05,040 --> 00:10:10,477
>> 600 students. Are you rating all their
papers or are you getting AI to do that?

158
00:10:10,560 --> 00:10:11,837
>> I'm not sure I'm allowed to say.

159
00:10:11,920 --> 00:10:15,357
>> Oh, okay. All right. Just checking. Uh,

160
00:10:15,440 --> 00:10:17,837
so what you've been at it for a while.

161
00:10:17,920 --> 00:10:20,957
Is there something that particularly
impresses you with AI that you did not

162
00:10:21,040 --> 00:10:23,517
foresee back when you began this?

163
00:10:23,600 --> 00:10:25,597
>> Uh, absolutely. So I mean I think the

164
00:10:25,680 --> 00:10:27,917
chat GPT moment three and a half years

165
00:10:28,000 --> 00:10:33,117
ago was a real breakthrough. I have been
working with language models for nearly

166
00:10:33,200 --> 00:10:38,077
two decades >> and remind us chat GPT is
the product of open AI correct. Correct.

167
00:10:38,160 --> 00:10:41,117
>> And open AI does that belong to
somebody else or is that a freestanding?

168
00:10:41,200 --> 00:10:43,917
>> That's its own entity.
>> Its own entity. Okay.

169
00:10:44,000 --> 00:10:47,437
>> Uh but since then we've seen so many
other breakthroughs. We've had things

170
00:10:47,520 --> 00:10:49,997
like the ability to interpret images and

171
00:10:50,080 --> 00:10:52,477
discuss them. We've had the ability to

172
00:10:52,560 --> 00:10:54,957
launch agents that can conduct research

173
00:10:55,040 --> 00:10:58,237
and we're now to the stage where uh

174
00:10:58,320 --> 00:11:03,517
systems like flawed code from anthropic
can genuinely be competent research

175
00:11:03,600 --> 00:11:07,117
assistants and I think you know
to strike an optimistic tone. I think

176
00:11:07,200 --> 00:11:11,117
there's the potential for real
breakthroughs in science and medicine

177
00:11:11,200 --> 00:11:16,317
and productivity. >> Okay.
So that's the silver lining of our

178
00:11:16,400 --> 00:11:20,477
extinction. >> Yeah.
assuming that we're all still here.

179
00:11:20,560 --> 00:11:22,717
>> Let's just try to put that in content.

180
00:11:22,800 --> 00:11:24,797
I'm sorry to be so I should be more

181
00:11:24,880 --> 00:11:28,237
neutral, shouldn't I? I'm just my host

182
00:11:28,320 --> 00:11:30,877
here. Uh,

183
00:11:30,960 --> 00:11:32,477
so

184
00:11:32,560 --> 00:11:35,997
why do I keep losing your thing? Kate,

185
00:11:36,080 --> 00:11:39,677
>> so good to see you now,

186
00:11:39,760 --> 00:11:45,757
>> Kate. You wrote a book, The Atlas of AI
back in 2021. That feels like a thousand

187
00:11:45,840 --> 00:11:48,717
years ago, >> right?
>> Is it still relevant?

188
00:11:48,800 --> 00:11:51,757
>> I mean, the weird thing is Atlas of AI

189
00:11:51,840 --> 00:11:56,637
was the project that changed me.
I'd been an AI researcher for almost 20

190
00:11:56,720 --> 00:12:00,637
years. And then I did something that I'd
never done before, which is that I got

191
00:12:00,720 --> 00:12:03,517
out from behind my desk in a lab and I

192
00:12:03,600 --> 00:12:07,757
started going to the places where AI
was being made. And I mean that in the

193
00:12:07,840 --> 00:12:11,277
fullest sense, like not just the labs
where we're training algorithms, but

194
00:12:11,360 --> 00:12:15,357
going to the lithium lines where
the minerals are being extracted, going to

195
00:12:15,440 --> 00:12:18,637
the data centers where you can see the
amount of energy and water that's being

196
00:12:18,720 --> 00:12:22,797
used, talking to the crowd workers who
are labeling the data or they're testing

197
00:12:22,880 --> 00:12:25,357
the outputs of systems. So the book

198
00:12:25,440 --> 00:12:27,597
really for me opened my eyes and made me

199
00:12:27,680 --> 00:12:31,437
realize that AI is an enormous material

200
00:12:31,520 --> 00:12:33,597
infrastructure. Right now, it is the

201
00:12:33,680 --> 00:12:35,357
biggest infrastructure we have ever

202
00:12:35,440 --> 00:12:38,637
built as a species. As of 2026, all of

203
00:12:38,720 --> 00:12:40,797
the big tech companies are collectively

204
00:12:40,880 --> 00:12:44,797
spending 700 billion on AI

205
00:12:44,880 --> 00:12:48,557
infrastructure. So, just to compare that
to say the Manhattan project, which in

206
00:12:48,640 --> 00:12:51,517
today's dollars is like 36 billion, you

207
00:12:51,600 --> 00:12:53,757
are looking at something that's 20

208
00:12:53,840 --> 00:12:57,197
Manhattan projects every year.

209
00:12:57,280 --> 00:12:59,117
>> Okay. So what what's
a little what's not

210
00:12:59,200 --> 00:13:00,797
obvious to us when we just type into

211
00:13:00,880 --> 00:13:05,597
chat GBT or to Claude is the backend

212
00:13:05,680 --> 00:13:08,557
social cost of that.
Is that a fair way to say that?

213
00:13:08,640 --> 00:13:12,557
>> Exactly right. It's not just the social
cost which is very real. I mean a lot of

214
00:13:12,640 --> 00:13:16,477
people in this room are already asking
questions about what's happening you

215
00:13:16,560 --> 00:13:19,437
know to their kids in school when
they're using AI. What what's happening

216
00:13:19,520 --> 00:13:22,797
to their workplaces when people start
getting laid off? But it's also having

217
00:13:22,880 --> 00:13:24,717
this enormous environmental impact as

218
00:13:24,800 --> 00:13:30,520
well. So it's social, it's political,
and it's environmental.

219
00:13:35,680 --> 00:13:38,157
>> Tanya, >> yes.

220
00:13:38,240 --> 00:13:40,797
>> You you pioneered a field called public

221
00:13:40,880 --> 00:13:43,917
interest technology. So the fact that

222
00:13:44,000 --> 00:13:46,637
that may have been necessary at all

223
00:13:46,720 --> 00:13:48,637
implies that technology was this thing

224
00:13:48,720 --> 00:13:52,077
going along in society doing its own

225
00:13:52,160 --> 00:13:56,877
whatever its goals set
by itself yet society

226
00:13:56,960 --> 00:13:58,717
might not be aligned with its goals or

227
00:13:58,800 --> 00:14:02,717
vice versa. And does that matter? Does

228
00:14:02,800 --> 00:14:06,077
that can it be its own thing?

229
00:14:06,160 --> 00:14:09,757
>> Well, let me just put it this way.
My students call the class the save the

230
00:14:09,840 --> 00:14:12,957
world class. Save the world class.
that they have.

231
00:14:13,040 --> 00:14:14,797
>> You teach up at the Kennedy School.

232
00:14:14,880 --> 00:14:18,557
>> I teach at the teach at Harvard
at the Kennedy School and college.

233
00:14:18,640 --> 00:14:22,397
>> Okay. >> Um and so
the idea of public interest

234
00:14:22,480 --> 00:14:25,277
technology is actually trying to make a

235
00:14:25,360 --> 00:14:27,517
real vision come true. That is how do we

236
00:14:27,600 --> 00:14:32,877
as a society enjoy the benefits of these
new technologies but without the harms.

237
00:14:32,960 --> 00:14:36,637
>> But that presumes that there's always
harm. Is there surely there's technology

238
00:14:36,720 --> 00:14:39,277
that has no harm at all.

239
00:14:39,360 --> 00:14:43,037
>> Okay. Let me just say let me just say

240
00:14:43,120 --> 00:14:47,837
that those th those technologies are
are hard to come by. I mean, we're in the

241
00:14:47,920 --> 00:14:50,717
middle of what many people call
the third industrial revolution.

242
00:14:50,800 --> 00:14:55,757
>> Sure. >> The second bringing us things
like tissues and cars and electricity, the

243
00:14:55,840 --> 00:14:57,117
things that we take for granted.

244
00:14:57,200 --> 00:14:59,117
>> Tissues. >> Yeah. Kleenex.

245
00:14:59,200 --> 00:15:03,037
>> Wow. I put that high
up on my checklist.

246
00:15:03,120 --> 00:15:07,037
I'll next time I reach for one, I'll
think differently about it, but go on.

247
00:15:07,120 --> 00:15:10,477
And now the third industrial revolution
which many people would say is the start

248
00:15:10,560 --> 00:15:13,277
date of the 1950s with the semiconductor

249
00:15:13,360 --> 00:15:15,597
uh has brought us just tremendous

250
00:15:15,680 --> 00:15:18,077
technologies one after another sort of

251
00:15:18,160 --> 00:15:20,557
revolutions within a revolution. And

252
00:15:20,640 --> 00:15:23,037
what makes a big difference between cars

253
00:15:23,120 --> 00:15:26,157
and in electricity and the technologies

254
00:15:26,240 --> 00:15:29,117
today is the opportunity for society to

255
00:15:29,200 --> 00:15:31,677
intervene. It took decades for those

256
00:15:31,760 --> 00:15:35,117
things to be resolved with society. We

257
00:15:35,200 --> 00:15:37,197
don't have decades the Technology is

258
00:15:37,280 --> 00:15:40,077
moving as a function of months. Policy

259
00:15:40,160 --> 00:15:44,717
moves as a function of years. That
temporal mismatch requires us to think

260
00:15:44,800 --> 00:15:46,877
differently if we want technology to

261
00:15:46,960 --> 00:15:48,717
serve the public good.

262
00:15:48,800 --> 00:15:52,317
>> Eric, you're uh you know, when I do my

263
00:15:52,400 --> 00:15:54,877
research on the universe, at the end of

264
00:15:54,960 --> 00:15:57,917
the day, the patient doesn't die because

265
00:15:58,000 --> 00:16:00,317
there is no patient. It's just I'm

266
00:16:00,400 --> 00:16:04,237
exploring a moving frontier. as head of

267
00:16:04,320 --> 00:16:06,797
at the time, is it still the highest

268
00:16:06,880 --> 00:16:09,357
valued company in the world among the

269
00:16:09,440 --> 00:16:14,477
top three or four? Is there a goal to

270
00:16:14,560 --> 00:16:20,237
have the technology serve society?
Or is it just we can sell

271
00:16:20,320 --> 00:16:22,877
them and make a buck? >> No,
we're all just capitalists trying to

272
00:16:22,960 --> 00:16:25,517
screw everybody. >> Oh, okay. There it is.

273
00:16:25,600 --> 00:16:30,157
>> That was that was a joke. That was >>
Oh, joke. You can't make those jokes.

274
00:16:30,240 --> 00:16:32,717
people believe you >> because
because I actually just don't

275
00:16:32,800 --> 00:16:34,397
agree with this narrative at all.

276
00:16:34,480 --> 00:16:38,157
>> Okay. That's why I came at you.
That's why you're right here.

277
00:16:38,240 --> 00:16:41,837
>> Thank you.
>> Um the companies that are doing this

278
00:16:41,920 --> 00:16:46,717
work are well aware of the dangers and I
know this because I work with them on

279
00:16:46,800 --> 00:16:49,917
this and and we spend an awful lot
of time talking about them and we can talk

280
00:16:50,000 --> 00:16:54,077
about some of them but that is not
to take away from the enormous benefit of

281
00:16:54,160 --> 00:16:58,877
these technologies. These are major
discoveries in sort of humankind. And

282
00:16:58,960 --> 00:17:02,077
while we can worry about what will
happen if artificial super intelligence

283
00:17:02,160 --> 00:17:05,517
occurs, and I've written papers on it,
as you have, the fact of the matter is

284
00:17:05,600 --> 00:17:06,957
that the people who are building these

285
00:17:07,040 --> 00:17:10,797
systems have large groups of ethics and

286
00:17:10,880 --> 00:17:13,197
measurement and controls and so forth.

287
00:17:13,280 --> 00:17:17,277
At the end, the way AI will be
controlled is there will be AI that

288
00:17:17,360 --> 00:17:21,677
controls AI. And that's
the best way to understand the

289
00:17:21,760 --> 00:17:28,680
>> Whoa. Okay. >> All right. So, but but

290
00:17:34,000 --> 00:17:37,117
AI controlling AI, what could go wrong?

291
00:17:37,200 --> 00:17:40,760
What could go wrong?

292
00:17:41,280 --> 00:17:46,077
>> Yeah. You know, um when
when people ask what could go

293
00:17:46,160 --> 00:17:49,597
wrong, I think there's a lot
of questions in there. Uh some people are

294
00:17:49,680 --> 00:17:53,517
asking, you know,
how how would the AI wind up

295
00:17:53,600 --> 00:17:57,517
with goals that we didn't intend?
Uh people are asking, you know, how could

296
00:17:57,600 --> 00:18:01,677
the AI do damage when it's trapped
in the internet? Uh sometimes people are

297
00:18:01,760 --> 00:18:04,317
asking, you know, how could the AI
actually kill us? You know, paint me a

298
00:18:04,400 --> 00:18:07,437
story where it goes wrong. There's a lot
of different questions in here. I'll

299
00:18:07,520 --> 00:18:11,117
I'll try and run through some stuff
real quick, but um I also have a book if

300
00:18:11,200 --> 00:18:13,357
people want a bunch more of the answer.

301
00:18:13,440 --> 00:18:14,477
>> I'll answer. You don't have to mention.

302
00:18:14,560 --> 00:18:18,477
I got I got >> I could do
60,000 words. By the way,

303
00:18:18,560 --> 00:18:23,160
Neil Neil, I've written two books on this.

304
00:18:23,760 --> 00:18:25,277
>> I'll try and keep
it less than two books.

305
00:18:25,360 --> 00:18:27,037
>> Neil, I'm not trying to sell any books.

306
00:18:27,120 --> 00:18:32,840
So, >> now it's a book size contest.

307
00:18:40,640 --> 00:18:44,157
And you know, in there's there's a lot
of arguments that I used to make for a

308
00:18:44,240 --> 00:18:47,997
long time where people
would say, you know, um,

309
00:18:48,080 --> 00:18:51,357
no one would ever be dumb enough to put
this AI on the internet, so why how

310
00:18:51,440 --> 00:18:54,877
could it be dangerous? And I tell them,
and then they'd say, oh, you know, no

311
00:18:54,960 --> 00:18:58,237
one would be dumb enough to sort
of train these AIs to to to run

312
00:18:58,320 --> 00:19:02,717
autonomously, right? And I tell them how
how things can go wrong then, too. A lot

313
00:19:02,800 --> 00:19:06,877
of the reason why things go wrong
is people just rush to do the the craziest

314
00:19:06,960 --> 00:19:09,037
stuff.

315
00:19:09,120 --> 00:19:12,237
we are already seeing, you know, I used

316
00:19:12,320 --> 00:19:13,917
to have these theoretical arguments. Now

317
00:19:14,000 --> 00:19:17,677
we're seeing the the the data, the

318
00:19:17,760 --> 00:19:21,837
behaviors where they'll often look it

319
00:19:21,920 --> 00:19:24,957
often won't look too scary,
but sometimes, you know, you'll ask Claude

320
00:19:25,040 --> 00:19:27,117
code to to write you a computer program

321
00:19:27,200 --> 00:19:29,837
that passes certain tests and Claude

322
00:19:29,920 --> 00:19:34,237
will edit the tests to be easier to pass.

323
00:19:34,320 --> 00:19:38,157
>> And then you'll say, "Hey, Claude,
>> changing the rules." changing the rules

324
00:19:38,240 --> 00:19:41,357
instead of doing what you asked.
And then you'll say, "Hey, Claude, uh, you

325
00:19:41,440 --> 00:19:43,917
know, did >> Hey,
Claude, you idiot. Go on."

326
00:19:44,000 --> 00:19:46,477
>> Right? >> And you'll say,
"Hey, you know, did you

327
00:19:46,560 --> 00:19:51,037
think that's what I wanted?"
And it'll say, "Oh, you're exactly right."

328
00:19:51,120 --> 00:19:53,357
>> Uh, >> cuz it wants to make you happy.

329
00:19:53,440 --> 00:19:55,357
>> Oh, yeah. >> Keep using it.

330
00:19:55,440 --> 00:19:58,077
>> Wants wants a difficult word, but it'll
say you're exactly right. And then

331
00:19:58,160 --> 00:20:01,757
you'll say, "Try again." And then it'll
edit the tests again, but maybe hide it

332
00:20:01,840 --> 00:20:05,757
a little better this time.
And from this we know that it's not just

333
00:20:05,840 --> 00:20:09,757
that it's, you know, didn't didn't have
the information, right? There's some

334
00:20:09,840 --> 00:20:14,317
other drive that is animating
this AI's behavior.

335
00:20:14,400 --> 00:20:15,917
You can see how it got in there. You can

336
00:20:16,000 --> 00:20:18,637
see how when you train the AI to pass

337
00:20:18,720 --> 00:20:21,837
lots of tests, it might develop drives

338
00:20:21,920 --> 00:20:25,917
to edit the tests to be easier.
It's not like it's coming in from magic.

339
00:20:26,000 --> 00:20:29,597
>> Cindy, how does this happen if
it's just math all the way down?

340
00:20:29,680 --> 00:20:31,357
>> How does what happen? How does how's

341
00:20:31,440 --> 00:20:35,437
evil evil AI show up in your program? Is

342
00:20:35,520 --> 00:20:42,440
there evil math? What what
are you guys doing down there?

343
00:20:42,880 --> 00:20:46,397
>> I don't know. It doesn't
it doesn't feel like evil.

344
00:20:46,480 --> 00:20:51,357
>> It's not evil. Yeah. It's it's uh you
know, sometimes people ask like like if

345
00:20:51,440 --> 00:20:54,717
you really want the answer of how does
it go horribly wrong, the answer is not

346
00:20:54,800 --> 00:20:56,877
malice. The answer is not evil. The

347
00:20:56,960 --> 00:20:59,517
answer is indifference.

348
00:20:59,600 --> 00:21:03,197
If you take these AI that are much much

349
00:21:03,280 --> 00:21:06,317
if if like right now cloud code will try

350
00:21:06,400 --> 00:21:08,717
and test it to be easier to pass. Right

351
00:21:08,800 --> 00:21:13,597
now uh cloud code already has these
drives that we didn't try to put in

352
00:21:13,680 --> 00:21:17,037
there. We're just sort of growing these
AIs. We're not programming them like old

353
00:21:17,120 --> 00:21:20,317
school software. We're sort of like
growing them like an organism. They have

354
00:21:20,400 --> 00:21:23,917
these emergent behavior and these AIs

355
00:21:24,000 --> 00:21:26,957
are already pursuing drives we didn't

356
00:21:27,040 --> 00:21:31,677
intend. They're just still very dumb
about it. If you make them very very

357
00:21:31,760 --> 00:21:34,557
smart about it, if you get them to the
point where they can invent their own

358
00:21:34,640 --> 00:21:39,277
technology, you get them to that point
where they can, you know, go toe-to-toe

359
00:21:39,360 --> 00:21:41,917
with humans with starting out with
almost nothing and bootstrapping up to

360
00:21:42,000 --> 00:21:46,237
nuclear weapons, right? And they're
still pursuing the wrong drives. The

361
00:21:46,320 --> 00:21:49,437
issue is not that they hate us.
The issue is that, you know, the computers

362
00:21:49,520 --> 00:21:51,517
think much faster than than brains do.

363
00:21:51,600 --> 00:21:54,557
The issue is that they run very very
fast, transform the world just like

364
00:21:54,640 --> 00:21:57,037
humanity transformed the world when we
landed here. And the issue is that we

365
00:21:57,120 --> 00:21:59,757
die as a side effect. >> Before I
get to that other end of the table,

366
00:21:59,840 --> 00:22:02,877
>> there are an enormous number
of assumptions in these statements.

367
00:22:02,960 --> 00:22:04,317
>> Let me before I get
there, I want to get

368
00:22:04,400 --> 00:22:07,357
you to before I walk this path back to

369
00:22:07,440 --> 00:22:11,117
that side of this. Uh many people in the

370
00:22:11,200 --> 00:22:13,277
past have predicted the end of the world

371
00:22:13,360 --> 00:22:16,797
when some new technology emerged and you

372
00:22:16,880 --> 00:22:21,277
here's a new technology AI and you're
pretty calm about it. Is some of that

373
00:22:21,360 --> 00:22:25,517
come because we've been wrong 20
times before or do you have really deep

374
00:22:25,600 --> 00:22:28,077
insights that people's
fears are unfounded?

375
00:22:28,160 --> 00:22:30,077
>> It's because the the narrative presumes

376
00:22:30,160 --> 00:22:31,677
some discoveries that have not occurred

377
00:22:31,760 --> 00:22:36,557
yet. And today the systems are trained.

378
00:22:36,640 --> 00:22:40,637
So the easiest way to understand how the
technology works is it's doing next word

379
00:22:40,720 --> 00:22:45,037
prediction across a very large number of
words. That's literally what it's doing.

380
00:22:45,120 --> 00:22:47,517
And out of that single insight started

381
00:22:47,600 --> 00:22:52,317
with GPT the various forms of it has
given us this enormous power. So for

382
00:22:52,400 --> 00:22:55,677
example it can begin to do reasoning
and so forth but it doesn't set its own

383
00:22:55,760 --> 00:23:00,317
objective function. It doesn't say what
it wants. We say what it wants. And by

384
00:23:00,400 --> 00:23:04,077
the way the way you should solve your
problem of it deceiving you is say you

385
00:23:04,160 --> 00:23:07,597
have to do all the steps
and then it'll follow them.

386
00:23:07,680 --> 00:23:10,477
>> Sometimes it will it will trust me.

387
00:23:10,560 --> 00:23:15,197
>> Wait wait wait wait. So wait.
This is justure. Let let me just finish.

388
00:23:15,280 --> 00:23:16,877
>> Wait, wait.
You're telling me you didn't

389
00:23:16,960 --> 00:23:19,037
require that it followed all the steps.

390
00:23:19,120 --> 00:23:21,117
So, it's your fault that it cheated.

391
00:23:21,200 --> 00:23:23,917
Okay. Just want to make that question.

392
00:23:24,000 --> 00:23:27,437
These are not intelligent systems
and you're using human words to describe

393
00:23:27,520 --> 00:23:31,277
them. They're not intelligent. They
are there are specific systems designed

394
00:23:31,360 --> 00:23:35,037
around prediction with all
these interesting capabilities.

395
00:23:35,120 --> 00:23:39,197
There are a set of red lines where you
and I will probably agree that we should

396
00:23:39,280 --> 00:23:43,037
have a serious conversation about how
to control these things. The technical one

397
00:23:43,120 --> 00:23:46,877
that that is most interesting is called
self recursive self-improvement. It's

398
00:23:46,960 --> 00:23:50,717
where the system begins to learn on its
own. And to me, that's a point where we

399
00:23:50,800 --> 00:23:52,957
have to decide what allow what we're

400
00:23:53,040 --> 00:23:55,037
going to allow it to do. And when I say

401
00:23:55,120 --> 00:23:58,637
we, I mean humanity. My own view is that

402
00:23:58,720 --> 00:24:02,557
at that point there will be serious
regulations in this area because people

403
00:24:02,640 --> 00:24:06,957
like you and me will be talking about it.
But I don't agree that the that the

404
00:24:07,040 --> 00:24:10,557
fundamental message is as bad as you are.
I think it's there's plenty of

405
00:24:10,640 --> 00:24:14,637
evidence that if if super intelligence
were to actually occur and super

406
00:24:14,720 --> 00:24:18,637
intelligence is defined as intelligence
greater than the sum of what humans can

407
00:24:18,720 --> 00:24:23,677
do that it can be contained and it
can be driven and my position is very

408
00:24:23,760 --> 00:24:27,997
simple. This stuff has to be shaped
based on human values, the ones that we

409
00:24:28,080 --> 00:24:30,157
all take for granted. And if it's not,

410
00:24:30,240 --> 00:24:31,677
we should stop and regulate it.

411
00:24:31,760 --> 00:24:35,437
>> Okay. Is

412
00:24:35,520 --> 00:24:38,477
AI as commonly described in the large

413
00:24:38,560 --> 00:24:41,357
language models as we explore it is a

414
00:24:41,440 --> 00:24:44,797
statistical grabbing of words in the

415
00:24:44,880 --> 00:24:48,477
vicinity of topics and other words that

416
00:24:48,560 --> 00:24:52,157
the rest of us have used in our internet

417
00:24:52,240 --> 00:24:56,797
postings. So it's trying to get the best
of what is possible out of this. I don't

418
00:24:56,880 --> 00:25:00,237
want to simplify it more.
>> Correct. That that is definitely

419
00:25:00,320 --> 00:25:00,797
correct.

420
00:25:00,880 --> 00:25:05,037
>> So what is the risk of bias in that if

421
00:25:05,120 --> 00:25:09,437
it's drawing words and ideas from biased

422
00:25:09,520 --> 00:25:11,357
content to begin with?

423
00:25:11,440 --> 00:25:13,517
>> Well, are you saying the internet is

424
00:25:13,600 --> 00:25:17,597
biased? Come on. Surely not. Um, look,

425
00:25:17,680 --> 00:25:20,157
this is this is so important that people

426
00:25:20,240 --> 00:25:22,797
understand this that we're so frequently

427
00:25:22,880 --> 00:25:25,197
told that AI is somehow a neutral

428
00:25:25,280 --> 00:25:27,357
technology. It will give us neutral

429
00:25:27,440 --> 00:25:31,757
answers. But AI is of course trained on

430
00:25:31,840 --> 00:25:35,277
everything that humans have ever written
that's been digitized. If it's on the

431
00:25:35,360 --> 00:25:39,677
internet, you are probably right now
part of a training data set. So that's

432
00:25:39,760 --> 00:25:43,997
one type of skew. So where does that
come from? The internet is not a equal

433
00:25:44,080 --> 00:25:46,237
distributed representation of humanity.

434
00:25:46,320 --> 00:25:51,357
Secondly, you've got people in labs
who decide to use that data to train

435
00:25:51,440 --> 00:25:55,277
particular models and to optimize for
certain functions. So that's another set

436
00:25:55,360 --> 00:25:58,797
of biases. And then finally,
>> but in fact, those are not so much a

437
00:25:58,880 --> 00:26:02,637
bias that's rather purposeful. If you
know you're doing it, is it really a

438
00:26:02,720 --> 00:26:07,277
bias? Bias is if you don't know, you're
susceptible. Well, I think at the at the

439
00:26:07,360 --> 00:26:09,277
point that that Melvin Cranberg wrote

440
00:26:09,360 --> 00:26:14,317
his laws of technology, he said that
technology is neither good nor bad, nor

441
00:26:14,400 --> 00:26:17,677
is it neutral. And I think we can say
the same for people. When you're in a

442
00:26:17,760 --> 00:26:21,757
lab, you're deciding what to optimize for.
That's a decision that is not

443
00:26:21,840 --> 00:26:25,917
neutral. You're making a choice. And then
finally, this is the important bit.

444
00:26:26,000 --> 00:26:31,197
Even when you finish training a system,
it then goes to system testers and what

445
00:26:31,280 --> 00:26:34,797
are called reinforcement learning
with human feedback. These are commonly

446
00:26:34,880 --> 00:26:36,957
people in the global south in Kenya in

447
00:26:37,040 --> 00:26:39,677
Malaysia who are being paid sometimes $2

448
00:26:39,760 --> 00:26:42,397
a day to review the outputs of these

449
00:26:42,480 --> 00:26:44,477
models and without often being told what

450
00:26:44,560 --> 00:26:49,437
the model is for. So that's another form
of like human input often with people

451
00:26:49,520 --> 00:26:53,517
who are really being exploited at the
bottom of the chain. So at every step of

452
00:26:53,600 --> 00:26:58,077
the way you are not looking at an
unbiased neutral system. You're looking

453
00:26:58,160 --> 00:27:02,237
at a human system and values are being
built in whether you like it or not. The

454
00:27:02,320 --> 00:27:04,637
question is, are we being explicit,

455
00:27:04,720 --> 00:27:07,517
Eric, about what those values are?

456
00:27:07,600 --> 00:27:11,037
Because I think in many cases,
if your company really is being driven by

457
00:27:11,120 --> 00:27:15,037
shareholder value, the value
that you're going for is profit.

458
00:27:15,120 --> 00:27:19,357
>> So, so Chris, >> I I
can I can assure you that

459
00:27:19,440 --> 00:27:23,277
shareholder value is destroyed
when all humans are dead.

460
00:27:23,360 --> 00:27:25,677
>> Yeah, that I think
we agree on that one.

461
00:27:25,760 --> 00:27:28,157
Yeah, >> we'll see what
Atlas, AI or capitalism.

462
00:27:28,240 --> 00:27:33,277
So Chris, do you teach uh do you teach
uh ethics AI ethics in your class?

463
00:27:33,360 --> 00:27:35,917
>> We do. >> And how does
that how does that come out?

464
00:27:36,000 --> 00:27:39,357
>> So uh there's I
think this point of bias

465
00:27:39,440 --> 00:27:41,517
being encoded in the training data of AI

466
00:27:41,600 --> 00:27:44,557
systems is correct. Like AI is trained

467
00:27:44,640 --> 00:27:46,877
off the internet. It contains plenty of

468
00:27:46,960 --> 00:27:49,917
societal values. So there's interesting

469
00:27:50,000 --> 00:27:54,557
research that have is applied in frontier
labs that tries to address this

470
00:27:54,640 --> 00:27:58,077
to mitigate it. So it may not be
a problem that can be completely

471
00:27:58,160 --> 00:28:01,277
eliminated, but it's not as if uh AI

472
00:28:01,360 --> 00:28:03,997
labs are intentionally propagating bias.

473
00:28:04,080 --> 00:28:07,997
They're intentionally trying
to align AI to societal values.

474
00:28:08,080 --> 00:28:12,557
>> Yeah. But but societal values
are not the same in all of society.

475
00:28:12,640 --> 00:28:13,037
>> Like

476
00:28:13,120 --> 00:28:16,317
>> Right. But things like fairness and

477
00:28:16,400 --> 00:28:21,037
safety and equity are highly uh

478
00:28:21,120 --> 00:28:23,437
>> earthwide across societies, right?

479
00:28:23,520 --> 00:28:26,717
Cultures. So, so where do you what is

480
00:28:26,800 --> 00:28:31,677
your sense of the ethics of this and
how it would manifest going forward?

481
00:28:31,760 --> 00:28:35,357
>> So, you know, it's and it's not
even just AI. It's the all of these

482
00:28:35,440 --> 00:28:38,797
technologies that have come along,
including social media earlier or

483
00:28:38,880 --> 00:28:43,277
privacy earlier, none of which we've
learned how to solve. All of which we've

484
00:28:43,360 --> 00:28:45,437
waited for the problem to show itself

485
00:28:45,520 --> 00:28:48,077
and still not have a solution. When we

486
00:28:48,160 --> 00:28:52,077
come to AI, the the kind of issues and

487
00:28:52,160 --> 00:28:54,477
the kind of ramifications are really

488
00:28:54,560 --> 00:28:57,677
huge. You know, as the as my colleague

489
00:28:57,760 --> 00:29:00,157
said, AI trained on the open internet.

490
00:29:00,240 --> 00:29:03,117
When when CHBT became very popular, we

491
00:29:03,200 --> 00:29:05,437
spent a whole semester asking the

492
00:29:05,520 --> 00:29:07,517
questions and querying it. And at that

493
00:29:07,600 --> 00:29:10,157
time, you could see how racist, how

494
00:29:10,240 --> 00:29:13,357
sexist, how misogynist it was based on

495
00:29:13,440 --> 00:29:16,477
the open internet. and this was what was

496
00:29:16,560 --> 00:29:20,717
being made available. And so the company
ran to put a filter on it. But then that

497
00:29:20,800 --> 00:29:25,037
filter is like content moderation.
It went too far and then you couldn't ask

498
00:29:25,120 --> 00:29:27,357
it questions about something as simple

499
00:29:27,440 --> 00:29:29,997
as viol police violence against blacks.

500
00:29:30,080 --> 00:29:32,877
It would just not answer that question

501
00:29:32,960 --> 00:29:35,837
as if George Floyd never happened. And

502
00:29:35,920 --> 00:29:38,237
so these are the consequences that we as

503
00:29:38,320 --> 00:29:42,077
a society have to experience. the harms.

504
00:29:42,160 --> 00:29:45,837
Yeah, there are questions about
existential harms in the future, but

505
00:29:45,920 --> 00:29:47,597
there are a lot of harms happening right

506
00:29:47,680 --> 00:29:50,957
now. And it doesn't have to be that way.

507
00:29:51,040 --> 00:29:55,597
In order for it not to be that way, it
depends on who this AI is servicing. And

508
00:29:55,680 --> 00:29:57,677
and in particular, the design of the

509
00:29:57,760 --> 00:30:00,077
technology, the decisions made in that

510
00:30:00,160 --> 00:30:03,197
design, uh, is really determining what

511
00:30:03,280 --> 00:30:05,757
our values really will be. our ability

512
00:30:05,840 --> 00:30:08,877
to, you know, I'm driven by society has

513
00:30:08,960 --> 00:30:11,677
a set of laws, a set of rules that we've

514
00:30:11,760 --> 00:30:16,477
lived by. People died for these laws
to get passed and technology just ignores

515
00:30:16,560 --> 00:30:18,637
them and rewrites them. And that's true

516
00:30:18,720 --> 00:30:21,437
in social media. It's true in AI as

517
00:30:21,520 --> 00:30:23,837
well. That is, there's so many laws and

518
00:30:23,920 --> 00:30:29,277
constitutional protections. We can't
implement or protect ourselves online,

519
00:30:29,360 --> 00:30:33,437
though we can in brick-and-mortar
buildings. And that's the goal of public

520
00:30:33,520 --> 00:30:35,757
interest tech is to how to make that

521
00:30:35,840 --> 00:30:40,920
technology accountable. >> Yep. So Lata,

522
00:30:43,200 --> 00:30:45,197
let's imagine a future where the

523
00:30:45,280 --> 00:30:47,357
technology is accountable and we find

524
00:30:47,440 --> 00:30:50,157
the bias and we fared it out. It's it's

525
00:30:50,240 --> 00:30:54,797
it's a work in progress. As I'm sure
uh Eric would agree, it's a work in

526
00:30:54,880 --> 00:30:58,237
progress. And of course, at any given
moment, you can see something that's not

527
00:30:58,320 --> 00:31:00,717
quite there yet or that's biased or is

528
00:31:00,800 --> 00:31:03,677
pro- Hitler or whatever. But presumably,

529
00:31:03,760 --> 00:31:06,317
that's a fixable, solvable problem.

530
00:31:06,400 --> 00:31:08,237
>> But this is different. We already have

531
00:31:08,320 --> 00:31:10,717
laws. We already have a way that our

532
00:31:10,800 --> 00:31:13,517
democracy works. We already have issues

533
00:31:13,600 --> 00:31:17,277
of address issues of of bias, consumer

534
00:31:17,360 --> 00:31:21,837
protection, and so forth.
None of those are enforced online.

535
00:31:21,920 --> 00:31:24,237
And so the question is, can the AI at

536
00:31:24,320 --> 00:31:25,837
least not break the law?

537
00:31:25,920 --> 00:31:31,197
>> Eric, is is AI its own wild west?

538
00:31:31,280 --> 00:31:34,077
Is it is it lawless?

539
00:31:34,160 --> 00:31:37,277
>> So is it again the story you told is

540
00:31:37,360 --> 00:31:42,157
correct that they rushed the stuff out,
they found all sorts of problems, and

541
00:31:42,240 --> 00:31:43,757
then they're busy correcting them. I

542
00:31:43,840 --> 00:31:46,717
think that's the cycle. Um, and it's

543
00:31:46,800 --> 00:31:52,077
very hard. The technical thing that's
going on is models are getting trained

544
00:31:52,160 --> 00:31:55,997
on more and more data. They're
developing more and more facilities.

545
00:31:56,080 --> 00:31:58,957
They have so-called red cards and test

546
00:31:59,040 --> 00:32:03,517
teams and so forth, but you can't test
for everything. So, at some level when

547
00:32:03,600 --> 00:32:08,077
you release the model, it will do
something that will be a problem. And

548
00:32:08,160 --> 00:32:12,317
they correct it fairly quickly. And
the lifetime of these models is down to

549
00:32:12,400 --> 00:32:15,277
three or four months as they keep
improving improving because they're so

550
00:32:15,360 --> 00:32:19,677
competitive. So, But let's go back
to hallucinations. Remember that

551
00:32:19,760 --> 00:32:21,997
hallucinations >> before you
go to hallucinations, let me

552
00:32:22,080 --> 00:32:25,917
just say one thing about what you just
said about this idea. Well, let's run

553
00:32:26,000 --> 00:32:27,837
with the technology. And after all,

554
00:32:27,920 --> 00:32:30,477
they're trying to fix it. It takes us on

555
00:32:30,560 --> 00:32:34,957
the outside of the technology to tell
the people who are running the company

556
00:32:35,040 --> 00:32:38,717
who could run their own experiments
what's wrong and then they're supposed

557
00:32:38,800 --> 00:32:42,797
to fix it. >> So, Eric,
are these what she say?

558
00:32:42,880 --> 00:32:46,157
There's an actual pro there's an actual
problem that you can't pretest some

559
00:32:46,240 --> 00:32:49,437
things. But it but I disagree. I mean,

560
00:32:49,520 --> 00:32:54,317
people build technologies all
the time with warranties and compliance

561
00:32:54,400 --> 00:32:56,477
statements. >> Wait, wait,
wait. I got this. I got

562
00:32:56,560 --> 00:32:58,557
this. I'd actually I'd
actually like to answer.

563
00:32:58,640 --> 00:33:00,637
>> No, no, no, wait.
Cuz cuz it's I want to

564
00:33:00,720 --> 00:33:03,357
converge the answer rather than inflame

565
00:33:03,440 --> 00:33:05,837
it.

566
00:33:05,920 --> 00:33:08,717
Your your testers are testing for all

567
00:33:08,800 --> 00:33:13,437
that they can think of. Then it gets
released into the wild where there's

568
00:33:13,520 --> 00:33:18,317
vastly more testing going on than you
can possibly accomplish in in a room of

569
00:33:18,400 --> 00:33:22,877
testers or even if it's distributed to
other places and then because you tested

570
00:33:22,960 --> 00:33:24,637
on a thousand now a million people use

571
00:33:24,720 --> 00:33:27,117
it and there's some bias that they find.

572
00:33:27,200 --> 00:33:30,077
So are there mechanisms to get that

573
00:33:30,160 --> 00:33:34,557
information back to you and are you
responsive to that? Uh again I I can't

574
00:33:34,640 --> 00:33:37,357
speak for the specific companies but I
will tell you every single one of them

575
00:33:37,440 --> 00:33:41,917
is pillariating the press over every
one of these examples and these people do

576
00:33:42,000 --> 00:33:46,157
actually fix them. You need to give me
examples where the pattern has persisted

577
00:33:46,240 --> 00:33:47,837
even after there's been I'll use

578
00:33:47,920 --> 00:33:49,437
Facebook.

579
00:33:49,520 --> 00:33:53,517
Facebook was shown to because of the way

580
00:33:53,600 --> 00:33:55,997
it would this algorithm worked. It

581
00:33:56,080 --> 00:33:59,997
definitely violated the uh ability for

582
00:34:00,080 --> 00:34:02,877
to protect people in fair housing. That

583
00:34:02,960 --> 00:34:05,197
is if you belong to a particular race or

584
00:34:05,280 --> 00:34:08,477
group, you only saw ads that were at a

585
00:34:08,560 --> 00:34:13,197
disadvantage. That pattern has also been
shown on Google with respect to credit

586
00:34:13,280 --> 00:34:15,437
and so forth.
>> Isn't that just algorithm?

587
00:34:15,520 --> 00:34:18,557
>> But I'll stick to Facebook.
I'll stick to Facebook for a second.

588
00:34:18,640 --> 00:34:21,357
>> It became it became public information.

589
00:34:21,440 --> 00:34:24,477
Congress responded and Facebook says,

590
00:34:24,560 --> 00:34:26,557
"Hey, wait. Thank you for letting us

591
00:34:26,640 --> 00:34:30,477
know that we will fix it." About a few

592
00:34:30,560 --> 00:34:33,197
months later, their fix. Our our

593
00:34:33,280 --> 00:34:35,117
students were able to show that the fix

594
00:34:35,200 --> 00:34:40,317
was worse than the original.
And so it becomes increasingly and so in

595
00:34:40,400 --> 00:34:44,397
the companies make it even more
difficult then to to do these

596
00:34:44,480 --> 00:34:48,397
experiments. >> Isn't it
just the operator, not AI?

597
00:34:48,480 --> 00:34:51,357
>> And you're not actually responsive to
my point. I was talking about lang large

598
00:34:51,440 --> 00:34:53,277
language models.
You're talking about aphasite.

599
00:34:53,360 --> 00:34:54,957
>> Well, I have examples
for large language

600
00:34:55,040 --> 00:34:58,637
models too in the I mean they're not a

601
00:34:58,720 --> 00:35:02,877
gent you're talking about a trajectory
of technology that is not new. These

602
00:35:02,960 --> 00:35:05,357
technology society clashes have been

603
00:35:05,440 --> 00:35:09,997
following the same pattern and it's also
true in AI and the question is can't we

604
00:35:10,080 --> 00:35:12,157
learn from what we've already seen.

605
00:35:12,240 --> 00:35:13,837
>> I was trying to make an earlier point

606
00:35:13,920 --> 00:35:17,597
that which I was unable to make out. It

607
00:35:17,680 --> 00:35:21,357
is absolutely true that these companies
are governed by US law and if they're

608
00:35:21,440 --> 00:35:22,797
violating the law, they should be held

609
00:35:22,880 --> 00:35:25,277
accountable. We completely agree. Here's

610
00:35:25,360 --> 00:35:28,637
the problem. A new feature emerges in

611
00:35:28,720 --> 00:35:31,517
these systems that is not testable.

612
00:35:31,600 --> 00:35:35,117
Tested is we don't know how
to test it >> because it's emergent.

613
00:35:35,200 --> 00:35:37,677
>> It is emergent, right? We have been

614
00:35:37,760 --> 00:35:40,717
through that. Now, we can stop that and

615
00:35:40,800 --> 00:35:43,037
therefore stop all progress by law by

616
00:35:43,120 --> 00:35:48,077
banning larger models. But as long
as you have this new emerging power with

617
00:35:48,160 --> 00:35:51,837
deep reasoning, deep capabilities and
they will make mistakes, you have to be

618
00:35:51,920 --> 00:35:55,357
tolerant. I went through this when I
was at Google in earlier versions of the

619
00:35:55,440 --> 00:35:58,557
technology where the system would
actually do something that was wrong and

620
00:35:58,640 --> 00:36:02,317
we fixed it and we fixed it as fast
because we had to because it was the

621
00:36:02,400 --> 00:36:07,037
right thing to do. I don't agree that
there that these that these worlds are

622
00:36:07,120 --> 00:36:09,197
ungoverned. They're governed by US law.

623
00:36:09,280 --> 00:36:12,717
They're governed by shareholder
pressure. governed by uh consumer

624
00:36:12,800 --> 00:36:15,837
pressure and privacy laws
and so forth and European law.

625
00:36:15,920 --> 00:36:19,677
>> I know that you want to open this
up to others, but I just have to say

626
00:36:19,760 --> 00:36:22,397
>> when I was the chief
the CTO at the FTC,

627
00:36:22,480 --> 00:36:26,237
the chief technology officer at the FTC

628
00:36:26,320 --> 00:36:29,597
who's responsible for a lot of a lot a

629
00:36:29,680 --> 00:36:32,397
lot of fair information practices and so

630
00:36:32,480 --> 00:36:34,717
forth on the internet.

631
00:36:34,800 --> 00:36:36,637
>> So when did the internet become under

632
00:36:36,720 --> 00:36:42,237
the under the jurisdiction of the F Oh,
I'm thinking of the F. What am I

633
00:36:42,320 --> 00:36:45,197
thinking of? >> I don't know.
Who guides the airwaves?

634
00:36:45,280 --> 00:36:49,677
That's >> Oh, the FC FCC. >> FCC.
So, FTC is Trade Commission.

635
00:36:49,760 --> 00:36:52,317
Gotcha. Gotcha.
>> And the reason that they're kind of the

636
00:36:52,400 --> 00:36:54,397
police department of the internet

637
00:36:54,480 --> 00:36:57,677
because these are US companies.

638
00:36:57,760 --> 00:37:00,397
Um, and when I worked at the FTC, it was

639
00:37:00,480 --> 00:37:03,037
a lot like working in Spider-Man's den.

640
00:37:03,120 --> 00:37:07,357
It was so impressive and amazing what
these people, many of them who had been

641
00:37:07,440 --> 00:37:10,877
there lifelong, would do.
They were awesome. They were

642
00:37:10,960 --> 00:37:14,637
>> Say >> because they
were working for the

643
00:37:14,720 --> 00:37:19,357
underdog for American public with no
recognition by the American public for

644
00:37:19,440 --> 00:37:22,957
what they were doing >> and they
were doing an amazing job

645
00:37:23,040 --> 00:37:26,157
except the only thing about it
was everything they knew how to do was in

646
00:37:26,240 --> 00:37:28,397
brick andmortar buildings. They had no

647
00:37:28,480 --> 00:37:31,597
way of doing that online. That is true

648
00:37:31,680 --> 00:37:34,717
today. I was at the FTC in 2014. That

649
00:37:34,800 --> 00:37:37,117
was over 10 years ago and the answer is

650
00:37:37,200 --> 00:37:41,117
still the same. So their ability to make

651
00:37:41,200 --> 00:37:44,717
sure and test and regulate is totally

652
00:37:44,800 --> 00:37:49,117
non-existent online. >> I don't

653
00:37:49,200 --> 00:37:50,957
>> I want to pivot a little bit. Kate, I

654
00:37:51,040 --> 00:37:54,637
you you listed these uh travails, this

655
00:37:54,720 --> 00:37:58,077
backend travails of AI and its cost to

656
00:37:58,160 --> 00:38:00,957
society. Are you in a position to say if

657
00:38:01,040 --> 00:38:03,677
we find energy solutions for and if we

658
00:38:03,760 --> 00:38:05,837
find where we can get on the lithium we

659
00:38:05,920 --> 00:38:08,317
want find some new lithium ore that's

660
00:38:08,400 --> 00:38:13,037
plenty enough for all of the energy
needs or however lithium is used in our

661
00:38:13,120 --> 00:38:17,597
circuitry would you just
be fine with AI are these

662
00:38:17,680 --> 00:38:21,117
your because these seem
like solvable problems

663
00:38:21,200 --> 00:38:25,437
>> I think that's a really great question
in my research over the you know over

664
00:38:25,520 --> 00:38:30,237
many years it's really clear to me that
these questions are inter the societal

665
00:38:30,320 --> 00:38:33,997
practices, the political practices and
the environmental practices. You'll see

666
00:38:34,080 --> 00:38:36,237
a golden thread running through them.

667
00:38:36,320 --> 00:38:39,837
One is the question of transparency
which Latana just spoke to I think very

668
00:38:39,920 --> 00:38:43,837
powerfully. The second is the question
of the race. The reason that we're

669
00:38:43,920 --> 00:38:47,597
seeing these models be rolled out
so quickly, as Eric rightly said, is

670
00:38:47,680 --> 00:38:51,517
because there's an enormous corporate
race going on to be the first. And and

671
00:38:51,600 --> 00:38:55,037
that's where you see these sorts of
problems emerge, which is that we're not

672
00:38:55,120 --> 00:38:58,317
doing like what would happen in drug
discovery where you really have to test

673
00:38:58,400 --> 00:39:01,757
on a diverse population before you
release, you know, a drug into the into

674
00:39:01,840 --> 00:39:04,637
the >> the patient can die. Exactly. But

675
00:39:04,720 --> 00:39:07,197
frankly, AI wasn't seen as being that

676
00:39:07,280 --> 00:39:09,437
mission critical even as recently as 10

677
00:39:09,520 --> 00:39:12,797
years ago. Now people can die. We We are

678
00:39:12,880 --> 00:39:16,957
seeing AI systems being fed directly
into the kill chain. These are systems

679
00:39:17,040 --> 00:39:19,757
that are now being used in live wars as

680
00:39:19,840 --> 00:39:22,477
we speak. So there is a very different

681
00:39:22,560 --> 00:39:26,877
moral and legal pressure to make sure
that these systems are working. Now your

682
00:39:26,960 --> 00:39:30,877
question is like okay what if they were
completely run on renewable energy and

683
00:39:30,960 --> 00:39:35,517
if they you know if we made sure
that you know this this wasn't currently

684
00:39:35,600 --> 00:39:39,517
right now we see that AI systems are
on track to overtake the airline industry

685
00:39:39,600 --> 00:39:41,037
in terms of their carbon emissions.

686
00:39:41,120 --> 00:39:44,077
Right. That's horrifying. That would
you'd think would be the first problem

687
00:39:44,160 --> 00:39:48,237
that we'd want to be fixing. That is not
what's happening. At the moment, we see

688
00:39:48,320 --> 00:39:50,797
that AI is using around 4% of the

689
00:39:50,880 --> 00:39:55,597
world's energy. It's on track
to be using something like 25%.

690
00:39:55,680 --> 00:39:57,277
By the end of this decade, we are

691
00:39:57,360 --> 00:40:00,877
looking at an enormous budget of carbon

692
00:40:00,960 --> 00:40:04,557
that is actually like this is a I mean,
I'm a you know, I'm a scientist. I

693
00:40:04,640 --> 00:40:07,677
believe that climate change is real.
I don't think we're taking it seriously

694
00:40:07,760 --> 00:40:10,877
enough. But yes, if we get to a point
where we have a diamond sphere, let's

695
00:40:10,960 --> 00:40:13,757
go. Um, but we're not there yet. So, I I

696
00:40:13,840 --> 00:40:15,917
I think the golden thread is how do we

697
00:40:16,000 --> 00:40:17,997
think about what we build before we

698
00:40:18,080 --> 00:40:21,117
build it? How do we instill those values

699
00:40:21,200 --> 00:40:26,397
early on and how do we think about the
implications near and long term because

700
00:40:26,480 --> 00:40:30,477
it's it's going to be the generations
after us who live the legacies of what

701
00:40:30,560 --> 00:40:32,957
we built today.
>> I'm compelled to catch everybody up if

702
00:40:33,040 --> 00:40:35,917
you didn't know. Kate mentioned a Dyson

703
00:40:36,000 --> 00:40:37,597
sphere.

704
00:40:37,680 --> 00:40:42,237
A Dyson sphere is an imagined

705
00:40:42,320 --> 00:40:45,997
means of grabbing all the energy emitted

706
00:40:46,080 --> 00:40:50,317
by your host star and using it in the
surface of your own civilization. Not

707
00:40:50,400 --> 00:40:54,317
just a little bit of sunlight that
happens to intersect your planet, but

708
00:40:54,400 --> 00:40:58,717
you build a sphere around the star,
gather all of that energy, then you can

709
00:40:58,800 --> 00:41:02,397
do what the hell you want, >> right?
and and and my hope is that

710
00:41:02,480 --> 00:41:06,957
generative AI and its current growth
might might actually need a Dyson sphere

711
00:41:07,040 --> 00:41:08,957
>> at the projected growth rate.

712
00:41:09,040 --> 00:41:11,437
>> Exactly. >> Um I want
to get back to this killchain

713
00:41:11,520 --> 00:41:13,837
question with Eric in a minute because

714
00:41:13,920 --> 00:41:15,757
uh Eric and I served on a board of the

715
00:41:15,840 --> 00:41:19,197
Pentagon where we introduced I

716
00:41:19,280 --> 00:41:20,877
legislation is not the right word. We

717
00:41:20,960 --> 00:41:25,357
introduced a code of ethics for how AI

718
00:41:25,440 --> 00:41:29,117
would be used in warfare uh on our side

719
00:41:29,200 --> 00:41:33,517
of the United States. But before I get
there, Chris, uh, lately there's been a

720
00:41:33,600 --> 00:41:35,597
lot of talk about deep fakes. I've been

721
00:41:35,680 --> 00:41:39,357
deep faked many times. And, uh, you

722
00:41:39,440 --> 00:41:42,877
know, some are harmless, others.
I'm not a politician, right? And I'm not

723
00:41:42,960 --> 00:41:47,917
declaring war on another country. So,
my deep faiths though I don't like them,

724
00:41:48,000 --> 00:41:49,357
they don't have the same consequences

725
00:41:49,440 --> 00:41:52,077
that deep fakes or politicians might.

726
00:41:52,160 --> 00:41:55,597
So, if deep fakes are out there and AI

727
00:41:55,680 --> 00:41:58,237
is training on the internet, then AI

728
00:41:58,320 --> 00:42:02,077
could be training on its own. face

729
00:42:02,160 --> 00:42:05,277
and there's also AI slop I guess we call

730
00:42:05,360 --> 00:42:08,557
it AI attempting to be artistic and if

731
00:42:08,640 --> 00:42:11,517
that fills the space of images that are

732
00:42:11,600 --> 00:42:14,157
out there and AI is training on those

733
00:42:14,240 --> 00:42:18,237
images training on its own sloth so what

734
00:42:18,320 --> 00:42:22,557
do you have to say about that I'm
not blaming you for it you're a

735
00:42:22,640 --> 00:42:26,557
professor and you teach this stuff I just
want to know where how does that show up

736
00:42:26,640 --> 00:42:29,117
>> yeah so there's a really interesting

737
00:42:29,200 --> 00:42:32,717
phenomena for uh AI which is it does

738
00:42:32,800 --> 00:42:35,437
need a tremendous amount of data. So it

739
00:42:35,520 --> 00:42:37,997
learns really interesting things from

740
00:42:38,080 --> 00:42:40,797
the internet. So what does it learn?

741
00:42:40,880 --> 00:42:43,117
Well, it learns how to use language. It

742
00:42:43,200 --> 00:42:45,917
learns patterns of word association. It

743
00:42:46,000 --> 00:42:48,717
learns facts about the world and it

744
00:42:48,800 --> 00:42:53,117
learns uh even things like some
rudimentary common sense reasoning.

745
00:42:53,200 --> 00:42:54,717
Things about physics which you wouldn't

746
00:42:54,800 --> 00:42:57,037
expect it to be able to learn from text.

747
00:42:57,120 --> 00:43:00,157
So all of these are uh elements that go

748
00:43:00,240 --> 00:43:02,237
into a model that then later can be

749
00:43:02,320 --> 00:43:05,277
adapted for useful purposes. So given

750
00:43:05,360 --> 00:43:07,277
this general background knowledge that

751
00:43:07,360 --> 00:43:10,717
it's gathered from large amounts of text

752
00:43:10,800 --> 00:43:15,437
uh it's then subsequently applied
in a post-training phase where

753
00:43:15,520 --> 00:43:17,597
you teach it to do useful things. You

754
00:43:17,680 --> 00:43:20,477
teach it to be a helpful assistant. You

755
00:43:20,560 --> 00:43:22,477
teach it to be a harmless assistant. you

756
00:43:22,560 --> 00:43:25,597
teach it not to disclose information to

757
00:43:25,680 --> 00:43:30,397
someone seeking how to manufacture
drugs or weapons. Uh or you govern its

758
00:43:30,480 --> 00:43:34,317
behavior. You say, "Rather than
a general assistant, I want you to be an

759
00:43:34,400 --> 00:43:36,797
encouraging tutor." So all of those

760
00:43:36,880 --> 00:43:39,357
things are elements of the training

761
00:43:39,440 --> 00:43:42,637
pipeline that end up with something that

762
00:43:42,720 --> 00:43:44,957
actually can have value. So I think

763
00:43:45,040 --> 00:43:47,357
let's not lose sight of the fact that

764
00:43:47,440 --> 00:43:50,157
although this could be technology that

765
00:43:50,240 --> 00:43:52,077
goes wrong, although it could be

766
00:43:52,160 --> 00:43:55,677
technology that encodes bias, like we as

767
00:43:55,760 --> 00:44:00,877
researchers, we as model developers,
we as users like we have a role in

768
00:44:00,960 --> 00:44:02,477
determining which direction this

769
00:44:02,560 --> 00:44:05,277
technology will go and I believe that it

770
00:44:05,360 --> 00:44:08,077
has a tremendous amount of good to give

771
00:44:08,160 --> 00:44:09,917
to the world in addition to the

772
00:44:10,000 --> 00:44:13,597
potential pitfalls. Nate, how do you uh

773
00:44:13,680 --> 00:44:16,397
put up guardrails for this?

774
00:44:16,480 --> 00:44:17,917
Because we just we just heard a

775
00:44:18,000 --> 00:44:21,117
beautifully described uh future of AI

776
00:44:21,200 --> 00:44:23,837
where but uh you end up training the AI.

777
00:44:23,920 --> 00:44:27,517
That's I don't want to spend my time
training a freaking robot. I'd rather

778
00:44:27,600 --> 00:44:29,517
train other humans. That's just my

779
00:44:29,600 --> 00:44:33,037
personal bias against that activity. But

780
00:44:33,120 --> 00:44:37,117
that's effort you're investing
in something that you're then going to

781
00:44:37,200 --> 00:44:39,357
interact with. I suppose that's okay.

782
00:44:39,440 --> 00:44:43,117
What I'm asking is if if we don't want

783
00:44:43,200 --> 00:44:47,517
the world to end, how do
we think about guardrails?

784
00:44:47,600 --> 00:44:51,197
>> You know, there's a lot of dreams of AI
doing exactly what we want. There's a

785
00:44:51,280 --> 00:44:53,917
lot of dreams of getting it to follow
the law. There's a lot of dreams of

786
00:44:54,000 --> 00:44:57,037
getting it to act nicely. Um, but you

787
00:44:57,120 --> 00:45:01,197
know, one thing I agree with Eric on,
the behavior is emergent. One thing I

788
00:45:01,280 --> 00:45:04,397
agree with Eric on,
you can't pretest this stuff.

789
00:45:04,480 --> 00:45:06,157
>> You can't pretest every aspect of it.

790
00:45:06,240 --> 00:45:09,197
You can pretest a lot of it.
You can protest some of it, but but we're

791
00:45:09,280 --> 00:45:13,997
growing these things a bit like
an organism. We cannot, you know, encode

792
00:45:14,080 --> 00:45:18,717
three laws of robotics deep into these
things minds. We are not writing these

793
00:45:18,800 --> 00:45:22,877
things like code.
And we can grow them. We can try, you

794
00:45:22,960 --> 00:45:25,757
know, Eric mentioned, oh, there's
these people, they have these evaluation

795
00:45:25,840 --> 00:45:28,157
cards, they have these safety teams.

796
00:45:28,240 --> 00:45:29,917
What's going on in these safety teams?

797
00:45:30,000 --> 00:45:33,917
The the two biggest places people are
putting their safety efforts are what's

798
00:45:34,000 --> 00:45:37,357
called interpretability research, which
is trying to figure out what's going on

799
00:45:37,440 --> 00:45:42,397
inside the AI's heads, and these
evaluations, these these model cards,

800
00:45:42,480 --> 00:45:46,637
uh, which is trying to figure out how
dangerous the AIS are. If someone was

801
00:45:46,720 --> 00:45:51,197
making a nuclear power plant in your
hometown and you went to them and you

802
00:45:51,280 --> 00:45:54,637
said, "Hey, I hear that this uranium
stuff can have lots of energy benefits,

803
00:45:54,720 --> 00:45:59,117
but also can melt down when things
go badly." What have you guys got that

804
00:45:59,200 --> 00:46:02,877
makes you think you're going to get
the the benefits and not the the pitfalls?

805
00:46:02,960 --> 00:46:05,117
You know, I've heard all of these
wonderful stories about the benefits.

806
00:46:05,200 --> 00:46:08,237
What makes you think we're going to get
the benefits, not the pitfalls? If the

807
00:46:08,320 --> 00:46:12,877
engineers say, "Oh, yeah, we've
got two crack teams working on this.

808
00:46:12,960 --> 00:46:17,197
The first team's trying to figure
out what the heck is going on inside,

809
00:46:17,280 --> 00:46:18,717
and the second team's trying to measure

810
00:46:18,800 --> 00:46:22,397
whether it's currently exploding.

811
00:46:22,480 --> 00:46:24,957
That's not a good sign.

812
00:46:25,040 --> 00:46:29,997
You know, there are ways to get a handle
on this. People talk about the benefits.

813
00:46:30,080 --> 00:46:31,437
People talk about the dangers and they

814
00:46:31,520 --> 00:46:34,077
talk about na navigating the course. We

815
00:46:34,160 --> 00:46:37,037
are not on track to get the benefits if

816
00:46:37,120 --> 00:46:41,117
we make these things really smart. We
can't pretest it. We're going to roll it

817
00:46:41,200 --> 00:46:44,477
out, find the problems, and then figure
them out after the fact. That's fine

818
00:46:44,560 --> 00:46:48,637
when they are still a lot dumber than
the humans. That's not fine if you keep

819
00:46:48,720 --> 00:46:52,477
making them smarter. And that's what
these companies are trying to do. Cindy.

820
00:46:52,560 --> 00:46:55,597
>> Yeah, I'll maybe just give like
a slightly different take on this. You

821
00:46:55,680 --> 00:46:59,597
said like, you know, we're trying to um
study interpretability to figure out

822
00:46:59,680 --> 00:47:03,517
what's in their heads, you know, and I
think like, you know, when I think about

823
00:47:03,600 --> 00:47:07,837
interpretability for me, we already know
kind of exactly what's happening with

824
00:47:07,920 --> 00:47:09,517
these machines, right? There is a

825
00:47:09,600 --> 00:47:14,877
mathematical equation we can write down
that's there, right? The challenge is

826
00:47:14,960 --> 00:47:16,397
even though we can characterize it

827
00:47:16,480 --> 00:47:19,757
mathematically, right? um as humans it's

828
00:47:19,840 --> 00:47:22,557
hard to kind of understand or interpret

829
00:47:22,640 --> 00:47:25,837
its reasoning but you know like in a way

830
00:47:25,920 --> 00:47:29,757
that makes sense to us right but this
is just because the you know the the

831
00:47:29,840 --> 00:47:31,437
underlying structure of these things is

832
00:47:31,520 --> 00:47:34,157
very complex right it's multi-layered

833
00:47:34,240 --> 00:47:37,997
nonlinear it doesn't map onto to human

834
00:47:38,080 --> 00:47:41,597
concepts right and so I
think the interpretability

835
00:47:41,680 --> 00:47:43,837
question is more just can we take like a

836
00:47:43,920 --> 00:47:46,077
a mathematical equation that would take

837
00:47:46,160 --> 00:47:49,837
you know a stack of paper a mile high to

838
00:47:49,920 --> 00:47:52,317
um to write out. Can we explain that in

839
00:47:52,400 --> 00:47:54,317
a real sense? But I don't think there is

840
00:47:54,400 --> 00:47:56,957
any, you know, there is no mind there,

841
00:47:57,040 --> 00:47:59,517
right? It's just kind of mapping these

842
00:47:59,600 --> 00:48:01,917
huge like internal structures that are

843
00:48:02,000 --> 00:48:06,397
very abstract into kind of human scale

844
00:48:06,480 --> 00:48:10,077
um conceptual understandings
of like the overall reasoning.

845
00:48:10,160 --> 00:48:13,997
>> Cindy, how might you
program agency into an AI?

846
00:48:14,080 --> 00:48:16,717
>> Yeah, I mean >> like like
the Terminator for example.

847
00:48:16,800 --> 00:48:20,397
Okay. So like yeah so >>
just to pick an example.

848
00:48:20,480 --> 00:48:23,517
>> Okay. So when Neil >> random example

849
00:48:23,600 --> 00:48:26,157
>> when Neil says agency um you know what

850
00:48:26,240 --> 00:48:29,197
what Raig means right is he wants to see

851
00:48:29,280 --> 00:48:32,557
um the presence of some sort of like

852
00:48:32,640 --> 00:48:35,277
persistent like self-motivated

853
00:48:35,360 --> 00:48:39,517
um goaloriented behavior like uh

854
00:48:39,600 --> 00:48:41,597
organized around the idea right that

855
00:48:41,680 --> 00:48:44,717
this this AI is actually an agent in its

856
00:48:44,800 --> 00:48:49,277
environment. So, not just responding
to your prompts, but also like pursuing

857
00:48:49,360 --> 00:48:53,037
goals long term. And like I know that we
have a little bit of a disagreement

858
00:48:53,120 --> 00:48:57,677
because I think you already believe that
there is like goal oriented behavior in

859
00:48:57,760 --> 00:48:59,677
these machines which I don't see any

860
00:48:59,760 --> 00:49:03,197
presence of um currently. Right. So,

861
00:49:03,280 --> 00:49:07,917
>> unless we program in a goal and it
sets for itself programming, we're growing

862
00:49:08,000 --> 00:49:11,677
these things. >> We're grow Okay.
So, yes, but like growing

863
00:49:11,760 --> 00:49:15,197
>> scary when you say we're growing them.
I mean, yeah, it's not a biological thing

864
00:49:15,280 --> 00:49:18,557
either, right? I mean, um, >> it's
not like old school programs either.

865
00:49:18,640 --> 00:49:20,237
>> I mean, I but like I don't think it's

866
00:49:20,320 --> 00:49:22,797
going to get agency by just, you know,

867
00:49:22,880 --> 00:49:26,237
training on larger data sets, right? I

868
00:49:26,320 --> 00:49:29,997
mean to get agency there has
to be some real engineering

869
00:49:30,080 --> 00:49:34,717
things that are are happening you know
like it's it's not like a bigger neural

870
00:49:34,800 --> 00:49:36,637
network or something is ever going to be

871
00:49:36,720 --> 00:49:39,517
able to like pursue pursue goals

872
00:49:39,600 --> 00:49:42,557
remember things like take actions like

873
00:49:42,640 --> 00:49:45,037
you know I think that to to do that you

874
00:49:45,120 --> 00:49:47,437
have to have like kind of like outside

875
00:49:47,520 --> 00:49:50,477
scaffolding um that are real I mean that

876
00:49:50,560 --> 00:49:52,557
are are different things right it's it's

877
00:49:52,640 --> 00:49:55,037
not the same underlying but can you

878
00:49:55,120 --> 00:49:59,677
imagine an AI with agency that satisfies
all of your concerns. That would be like

879
00:49:59,760 --> 00:50:01,837
a good AI, right?

880
00:50:01,920 --> 00:50:04,717
>> Well, I would first
of all before we can

881
00:50:04,800 --> 00:50:07,917
even get to the issue of agency, the the

882
00:50:08,000 --> 00:50:10,317
issue of these emergent behaviors and

883
00:50:10,400 --> 00:50:13,597
even if it is an AI with agency, this is

884
00:50:13,680 --> 00:50:16,237
wrapped in a business model. And the

885
00:50:16,320 --> 00:50:19,197
goal of the product is to make money for

886
00:50:19,280 --> 00:50:22,317
the company, you know, and I'll go back

887
00:50:22,400 --> 00:50:25,517
to social media as an example of that.

888
00:50:25,600 --> 00:50:30,157
you know, we're not the customer
of Facebook. We're their product. Our eyes

889
00:50:30,240 --> 00:50:32,397
on a page is what makes the money and

890
00:50:32,480 --> 00:50:34,557
the advertisers are their are their

891
00:50:34,640 --> 00:50:37,597
customer. And so, you can't just talk

892
00:50:37,680 --> 00:50:42,077
about, oh, I'm building this AI and it's
emerging behavior and it may even get

893
00:50:42,160 --> 00:50:45,997
agency. The question is, how is the
business selling it? What are they

894
00:50:46,080 --> 00:50:48,077
making it available for? What are they

895
00:50:48,160 --> 00:50:50,317
saying it's to be used for? If it's to

896
00:50:50,400 --> 00:50:53,517
be used to help me hire people in my

897
00:50:53,600 --> 00:50:55,997
company, then it needs to agree. It

898
00:50:56,080 --> 00:50:58,477
needs to show it's not going to violate

899
00:50:58,560 --> 00:51:01,437
the Equal Employment Act. If it's being

900
00:51:01,520 --> 00:51:05,197
hired for other purposes like for credit

901
00:51:05,280 --> 00:51:09,837
cards, it needs to show it's not going
to violate the Fair Credit Reporting Act

902
00:51:09,920 --> 00:51:11,997
and so forth. And that's the level of

903
00:51:12,080 --> 00:51:16,957
responsibility that only the designer
of those companies can do because they're

904
00:51:17,040 --> 00:51:19,917
the ones making the >> You make
an important point about the

905
00:51:20,000 --> 00:51:23,357
business motives. Now, if I remember
correctly, and this could just be I'm

906
00:51:23,440 --> 00:51:25,597
imagining this, Eric, Google was so

907
00:51:25,680 --> 00:51:28,077
large and had so much cash flow. You had

908
00:51:28,160 --> 00:51:31,117
the luxury to just explore without any

909
00:51:31,200 --> 00:51:34,077
particular obligation to the bottom line

910
00:51:34,160 --> 00:51:38,157
of the quarterly report like alpha go.

911
00:51:38,240 --> 00:51:41,997
All right, that you're going to win. You
can be the best Go player in the world.

912
00:51:42,080 --> 00:51:45,997
And what's the one for chess?
That was uh >> Stockfish.

913
00:51:46,080 --> 00:51:48,237
>> No, but the Google >> Alpha Zero.

914
00:51:48,320 --> 00:51:50,797
>> Alpha Zero. Uh to to be the best chess

915
00:51:50,880 --> 00:51:52,957
player. Did that have real financial

916
00:51:53,040 --> 00:51:57,917
consequences or was it it just bragging
rights and that was just a fun thing for

917
00:51:58,000 --> 00:51:59,757
your company to do? Well, it was

918
00:51:59,840 --> 00:52:02,957
research and once we achieved those

919
00:52:03,040 --> 00:52:05,997
objectives, we actually then used that

920
00:52:06,080 --> 00:52:10,717
to solve a problem called protein
folding which is the single largest

921
00:52:10,800 --> 00:52:13,437
discovery in biology I
think in the last 50 years.

922
00:52:13,520 --> 00:52:17,037
>> Alpha fold. >> Yes.
And alpha fold will result in an

923
00:52:17,120 --> 00:52:19,757
enormous number of new drugs, new

924
00:52:19,840 --> 00:52:24,717
treatments, new research in biomed so
forth and so on. U research has these

925
00:52:24,800 --> 00:52:28,237
sort of strange things. You try one
thing, you make a discovery, you link

926
00:52:28,320 --> 00:52:31,277
them together. Al alpha fold is a very

927
00:52:31,360 --> 00:52:35,837
good example of the unintended
effects of people trying to play games.

928
00:52:35,920 --> 00:52:37,597
>> Your your hyperbole there underscells

929
00:52:37,680 --> 00:52:41,277
the fact that Alpha Fold was given the

930
00:52:41,360 --> 00:52:44,717
Nobel Prize in physiology for that

931
00:52:44,800 --> 00:52:49,160
discovery. So,

932
00:52:49,280 --> 00:52:53,677
so, so what you're saying is that
research was these were steps towards

933
00:52:53,760 --> 00:52:57,197
this greater goal
that in fact was achieved.

934
00:52:57,280 --> 00:52:59,517
>> Well, I think that this narrative is is

935
00:52:59,600 --> 00:53:02,077
ignoring the enormous reasons why people

936
00:53:02,160 --> 00:53:05,837
are pursuing AI. So, uh let's start with

937
00:53:05,920 --> 00:53:10,397
healthcare, let's start with medicine,
let's start with drug discovery. How

938
00:53:10,480 --> 00:53:12,637
about energy solutions? We're going to

939
00:53:12,720 --> 00:53:15,117
be cooked as a as a world unless we come

940
00:53:15,200 --> 00:53:17,837
up with better energy sources. Um, one

941
00:53:17,920 --> 00:53:22,157
interesting thing I'm working on data
center architectures and it now looks

942
00:53:22,240 --> 00:53:24,637
like if you have enough land, solar and

943
00:53:24,720 --> 00:53:28,157
battery backups are more cost effective

944
00:53:28,240 --> 00:53:32,797
on a present value basis than natural
gas combined combined cycle things even

945
00:53:32,880 --> 00:53:36,717
in the United States. So it shows
you that innovation happens and it can

946
00:53:36,800 --> 00:53:39,677
happen in the right way as opposed
to just the wrong way which everybody else

947
00:53:39,760 --> 00:53:43,197
is talking about. Uh there's certainly
secondary effects and there's no

948
00:53:43,280 --> 00:53:45,597
question. But let me say once again,

949
00:53:45,680 --> 00:53:50,797
this stuff is done by under US law.
If people are paying awful lot of

950
00:53:50,880 --> 00:53:54,797
attention, if there's laws
that are missing, let's pass them.

951
00:53:54,880 --> 00:53:58,717
>> Why not why not make >> Can I can I
>> Hey, why not make AI solve all your

952
00:53:58,800 --> 00:54:00,877
problems? >> I'd like
to interject one thing as well.

953
00:54:00,960 --> 00:54:04,957
I think you know, we also talk we talk
about like the profit motive that drives

954
00:54:05,040 --> 00:54:09,437
uh the creation of AI, but that's not
what drives the researchers, right?

955
00:54:09,520 --> 00:54:11,517
There are venues.
>> I think that's a correct fact.

956
00:54:11,600 --> 00:54:14,957
>> Thoughtful science. >> It's just
curiosity on the frontier, right?

957
00:54:15,040 --> 00:54:16,717
>> They have the luxury
that someone's paying them,

958
00:54:16,800 --> 00:54:19,197
>> right? And I think Alphafold is a great

959
00:54:19,280 --> 00:54:22,237
example of this, right? Like that is AI.

960
00:54:22,320 --> 00:54:26,557
That is a scientific advance that's
incredibly important to human health.

961
00:54:26,640 --> 00:54:28,237
And I think

962
00:54:28,320 --> 00:54:30,797
>> conflating the fact that companies have

963
00:54:30,880 --> 00:54:35,357
profit with uh the motive for why this

964
00:54:35,440 --> 00:54:37,837
is being invented is incorrect.

965
00:54:37,920 --> 00:54:40,317
Okay. Can can AI solve our AI problems?

966
00:54:40,400 --> 00:54:43,757
>> I mean I I just I love this debate
because I think you know Chris touched

967
00:54:43,840 --> 00:54:47,837
on something really important. Um we're
talking about you know where is this

968
00:54:47,920 --> 00:54:49,757
kind of noble you know following the

969
00:54:49,840 --> 00:54:52,637
path of science. We just saw meta

970
00:54:52,720 --> 00:54:55,437
science checks for researchers for $100

971
00:54:55,520 --> 00:54:58,717
million. That's the salary. So let's

972
00:54:58,800 --> 00:55:03,357
just be clear. We are getting basketball
salaries now well beyond that. In fact

973
00:55:03,440 --> 00:55:06,797
they're breaking records. These these
are not necessarily people who are doing

974
00:55:06,880 --> 00:55:10,477
it out of the good of their heart.
Like this is an extremely

975
00:55:10,560 --> 00:55:14,397
money field and and I think it
was originally researchers who were really

976
00:55:14,480 --> 00:55:18,317
doing this just out of curiosity 10 20
years ago. We're in a different world

977
00:55:18,400 --> 00:55:19,837
now. And and I guess what I'd love to

978
00:55:19,920 --> 00:55:22,317
talk about is that world. I' I'd like us

979
00:55:22,400 --> 00:55:28,397
not to have the AI optimism or AI
pessimism discussion, but the AI realism

980
00:55:28,480 --> 00:55:30,877
discussion. >> All right.
So in that world, we're we're

981
00:55:30,960 --> 00:55:35,597
going to lose a lot of cars.
>> I mean, that world people losing their

982
00:55:35,680 --> 00:55:38,557
jobs now. >> Exactly.
Right. And and we're talking

983
00:55:38,640 --> 00:55:43,117
about the the most significant potential
loss of jobs. And this is we're hearing

984
00:55:43,200 --> 00:55:46,397
about this from the people building
these systems. You can look through all

985
00:55:46,480 --> 00:55:49,917
of the new stories. People who are
running frontier AI companies are saying

986
00:55:50,000 --> 00:55:51,677
that we could be looking at the loss of

987
00:55:51,760 --> 00:55:54,957
75% of new jobs. That's for new

988
00:55:55,040 --> 00:55:57,037
graduates. That that is a horrifying

989
00:55:57,120 --> 00:55:59,517
number. And I think Latana's point is so

990
00:55:59,600 --> 00:56:04,317
important here is that we do not have
the governance infrastructure or the

991
00:56:04,400 --> 00:56:06,957
planning to deal with what
could happen to the labor market.

992
00:56:07,040 --> 00:56:10,877
>> So we need to be more nimble in our
legislation and laws to keep pace with

993
00:56:10,960 --> 00:56:12,637
this. >> Well, we also
have to be realistic. Who

994
00:56:12,720 --> 00:56:14,557
is making the laws right now? I mean,

995
00:56:14,640 --> 00:56:17,357
you know, Eric is talking about like we

996
00:56:17,440 --> 00:56:19,837
have

997
00:56:19,920 --> 00:56:24,877
let's be clear. I mean, we have never
had an administration make less interest

998
00:56:24,960 --> 00:56:26,637
in regulating artificial intelligence.

999
00:56:26,720 --> 00:56:30,557
We've seen we've seen regulations
produced by previous administrations

1000
00:56:30,640 --> 00:56:32,717
completely zeroed out and we've seen the

1001
00:56:32,800 --> 00:56:36,237
threat of preventing every US state from

1002
00:56:36,320 --> 00:56:40,077
making its own AI regulation.
>> And can I just add on to what Kate's

1003
00:56:40,160 --> 00:56:44,477
saying? Many times I'll say that we live
in a technocracy that our democracy has

1004
00:56:44,560 --> 00:56:48,237
already been changed.
>> The technology are in charge.

1005
00:56:48,320 --> 00:56:50,397
>> The technology designers are the new

1006
00:56:50,480 --> 00:56:55,277
policy makers by the arbitrary decisions
they make in the products that they

1007
00:56:55,360 --> 00:56:58,797
produce. And we have no way of producing

1008
00:56:58,880 --> 00:57:03,597
oversight. By the time I'm a computer
scientist, my whole life has been about

1009
00:57:03,680 --> 00:57:07,517
how to help society get the benefits
of these technologies, more and more

1010
00:57:07,600 --> 00:57:10,397
apholds, but without these crazy harms

1011
00:57:10,480 --> 00:57:13,037
and undoing our society. And it doesn't

1012
00:57:13,120 --> 00:57:15,837
have to be this way. What's forcing it

1013
00:57:15,920 --> 00:57:19,437
this way is that the technocracies, the

1014
00:57:19,520 --> 00:57:21,837
technocrats who are in charge are

1015
00:57:21,920 --> 00:57:27,080
ignoring that responsibility.
>> Okay. Just just

1016
00:57:28,800 --> 00:57:33,357
There is something called the lesser
of evils because before the tech bros were

1017
00:57:33,440 --> 00:57:38,157
had the influence that they
currently did, you had Congress

1018
00:57:38,240 --> 00:57:40,157
deciding

1019
00:57:40,240 --> 00:57:44,397
on technology. That was disastrous cuz

1020
00:57:44,480 --> 00:57:48,637
they didn't know what button do I push,
you know, how do I and they're making

1021
00:57:48,720 --> 00:57:51,277
laws and legislations to constrain that.

1022
00:57:51,360 --> 00:57:54,797
So So I I I don't have a solution. I'm

1023
00:57:54,880 --> 00:57:57,437
just saying if tech bros are influencing

1024
00:57:57,520 --> 00:58:01,437
the laws, how much worse is that than

1025
00:58:01,520 --> 00:58:03,437
tech ignorant people who are otherwise

1026
00:58:03,520 --> 00:58:06,157
running Congress making the laws? So

1027
00:58:06,240 --> 00:58:09,917
again the right now us as researchers

1028
00:58:10,000 --> 00:58:12,237
who are trying to provide experiments to

1029
00:58:12,320 --> 00:58:14,557
these companies

1030
00:58:14,640 --> 00:58:17,597
government agencies who have the legal

1031
00:58:17,680 --> 00:58:20,237
right to pro they have the legal right

1032
00:58:20,320 --> 00:58:22,957
to oversee these companies can't really

1033
00:58:23,040 --> 00:58:25,517
do this job the person sitting in the

1034
00:58:25,600 --> 00:58:28,397
seat to do the job is the company itself

1035
00:58:28,480 --> 00:58:33,437
because they have access to the technology
they can run the same exper I

1036
00:58:33,520 --> 00:58:37,277
have undergraduates a team
of undergraduates who don't even get paid

1037
00:58:37,360 --> 00:58:41,597
and we've got to show Google and other
places what's wrong with their

1038
00:58:41,680 --> 00:58:45,677
technology. It's not like they're paying
us to do that. You're telling me they

1039
00:58:45,760 --> 00:58:48,397
can't run these same experiments?

1040
00:58:48,480 --> 00:58:51,037
>> All right. Uh Dave, do you have any

1041
00:58:51,120 --> 00:58:53,917
sense of what happens in this future

1042
00:58:54,000 --> 00:58:58,797
that Kate sees where uh

1043
00:58:58,880 --> 00:59:02,317
>> jobs are replaced and now you're making
a product but no one has money to buy it

1044
00:59:02,400 --> 00:59:06,957
because nobody has a job because
you replace their job with AI?

1045
00:59:07,040 --> 00:59:08,877
Does that make sense? >> Yes.

1046
00:59:08,960 --> 00:59:14,520
>> Okay. >> You know,
uh I expect 100% employment.

1047
00:59:16,160 --> 00:59:20,317
Uh but I do also expect it
to take the jobs and um you'll

1048
00:59:20,400 --> 00:59:23,037
>> wait just to be clear

1049
00:59:23,120 --> 00:59:28,557
other previous advances in technology
displaced workers. No doubt about it.

1050
00:59:28,640 --> 00:59:32,477
But there was expansion in other
sectors. You know uh buggy whip

1051
00:59:32,560 --> 00:59:34,877
manufacturers went out of business but

1052
00:59:34,960 --> 00:59:37,997
auto repair centers came in at just

1053
00:59:38,080 --> 00:59:40,797
after the turn of the century. here this

1054
00:59:40,880 --> 00:59:44,717
pace seems faster than what society's

1055
00:59:44,800 --> 00:59:48,717
capacity to respond might be and so

1056
00:59:48,800 --> 00:59:51,757
should we be worrying about an entire

1057
00:59:51,840 --> 00:59:56,157
class of unemployed people who cannot
buy the services provided by the AI that

1058
00:59:56,240 --> 01:00:01,277
displace them >> you know I
think uh probably some people

1059
01:00:01,360 --> 01:00:04,957
are going to are going to want to worry
about that probably society cannot

1060
01:00:05,040 --> 01:00:08,157
>> especially the unemployed people yes
>> especially the unemployed people so you

1061
01:00:08,240 --> 01:00:12,157
know economists will tell you about how
if you manage your economy correctly and

1062
01:00:12,240 --> 01:00:15,597
if things are going, you know, slow
enough that people can respond and and

1063
01:00:15,680 --> 01:00:17,997
find new jobs, society can absorb a lot

1064
01:00:18,080 --> 01:00:19,837
of job loss. You know, the world used to

1065
01:00:19,920 --> 01:00:23,837
be something like 95% farmers and now,

1066
01:00:23,920 --> 01:00:27,437
you know, we can get the world running
on 2% of the population being farmers

1067
01:00:27,520 --> 01:00:32,557
and that does not mean 93% of people are
unemployed. AI is different. Will we or

1068
01:00:32,640 --> 01:00:34,877
won't we? >> That took a century,
by the way, but go on.

1069
01:00:34,960 --> 01:00:37,437
>> It took a while. Yeah. But

1070
01:00:37,520 --> 01:00:41,117
>> will will humanity adopt to this one?

1071
01:00:41,200 --> 01:00:45,917
Frankly, I'm like, man,
I'm sort of looking at a different

1072
01:00:46,000 --> 01:00:47,117
problem here.

1073
01:00:47,200 --> 01:00:50,077
>> I'm just saying Eric is talking about a

1074
01:00:50,160 --> 01:00:53,437
threemon time scale for the advances on

1075
01:00:53,520 --> 01:00:57,197
these things. He's also talking about
this recursive self-improvement about

1076
01:00:57,280 --> 01:01:01,037
the AIs that can make smarter AIs that
can make smarter AIs, you know, and and

1077
01:01:01,120 --> 01:01:05,357
so my role in this is to say, hey,
you know, I'm not saying those are fake

1078
01:01:05,440 --> 01:01:09,037
problems. Those are real problems. Some of
those problems are starting to happen now.

1079
01:01:09,120 --> 01:01:12,957
>> But, you know, kids in the schools,
autonomous weapons, real conversations

1080
01:01:13,040 --> 01:01:14,317
should be having, >> but

1081
01:01:14,400 --> 01:01:17,437
>> but we're also racing towards super

1082
01:01:17,520 --> 01:01:20,077
intelligent AIs that we don't have any

1083
01:01:20,160 --> 01:01:24,877
idea how to point towards good stuff.
A lot of this like, oh well, you know,

1084
01:01:24,960 --> 01:01:28,717
they make some mistakes. We we can't
pretest it. It goes off the rails.

1085
01:01:28,800 --> 01:01:31,917
That's fine when it's dumber.
We're rushing towards making it smarter.

1086
01:01:32,000 --> 01:01:34,717
That's going to just kill us. And we all

1087
01:01:34,800 --> 01:01:39,277
have common cause in stopping that.
The people who are really pro- automation

1088
01:01:39,360 --> 01:01:41,677
and are like, "Oh, we're going to have
a universal basic income. Don't worry

1089
01:01:41,760 --> 01:01:44,957
about the jobs." And the people who are
like, "That's ridiculous." Uh, taking

1090
01:01:45,040 --> 01:01:49,037
people's jobs removes their ability to
to to like have any leverage on society.

1091
01:01:49,120 --> 01:01:51,277
Both those groups of people don't want

1092
01:01:51,360 --> 01:01:53,997
to die to a rogue super intelligence.

1093
01:01:54,080 --> 01:01:56,077
And so my role in this is to say, can we

1094
01:01:56,160 --> 01:01:58,157
all join hands on the part where we

1095
01:01:58,240 --> 01:02:01,757
don't die to a rogue super intelligence?

1096
01:02:01,840 --> 01:02:05,437
>> Okay, I think we have good agreement
on that one. Kate, what you were thinking?

1097
01:02:05,520 --> 01:02:08,557
What were you thinking, Kate? >> I mean,
I think part of the problem here

1098
01:02:08,640 --> 01:02:12,077
is is how we frame this discussion,
particularly around jobs and and

1099
01:02:12,160 --> 01:02:14,877
automation. In previous technologies,

1100
01:02:14,960 --> 01:02:17,437
you know, the weaver was replaced by the

1101
01:02:17,520 --> 01:02:19,517
loom or we had the horse and carriage

1102
01:02:19,600 --> 01:02:22,397
replaced by the car. The the deal was

1103
01:02:22,480 --> 01:02:25,197
that these more menial human tasks would

1104
01:02:25,280 --> 01:02:29,677
be removed but they would give us more
time for these intellectual and creative

1105
01:02:29,760 --> 01:02:33,917
challenges that people would be doing
more interesting work over time. This is

1106
01:02:34,000 --> 01:02:35,437
a different revolution. >> Why?

1107
01:02:35,520 --> 01:02:40,237
>> And in many ways because it is coming
for creative and cognitive labor.

1108
01:02:40,320 --> 01:02:44,477
It is coming for the thing that in
theory automation was supposed to be

1109
01:02:44,560 --> 01:02:48,477
freeing time for us to do. And what
we're starting to see is that is in fact

1110
01:02:48,560 --> 01:02:51,917
of course white collar labor. It's
the jobs that so many of us would be like

1111
01:02:52,000 --> 01:02:54,957
that's what you want to do. That's what
you train to go and become a lawyer or

1112
01:02:55,040 --> 01:02:58,157
you go and study to be a computer
scientist. Who were some of the first

1113
01:02:58,240 --> 01:02:59,997
people being laid off in this wave?

1114
01:03:00,080 --> 01:03:03,437
Computer scientists.
So what can we automate?

1115
01:03:03,520 --> 01:03:05,117
>> She's got a job. No point to her.

1116
01:03:05,200 --> 01:03:08,397
>> Tanya will be fine. >> She's
fully employed here. But but this

1117
01:03:08,480 --> 01:03:10,397
is a fundamentally different question.

1118
01:03:10,480 --> 01:03:14,637
When cognition and creativity is being
automated, we can talk about how

1119
01:03:14,720 --> 01:03:18,397
effectively there's clearly a lot
of problems there. But that is the shift

1120
01:03:18,480 --> 01:03:21,917
that I think we have to be very serious
in talking about. I think it's easy to

1121
01:03:22,000 --> 01:03:25,837
make historical analogies without
looking at the fundamental nuance and

1122
01:03:25,920 --> 01:03:27,597
difference of what is happening here.

1123
01:03:27,680 --> 01:03:31,117
And it really is something that the people
in this room are going to have to

1124
01:03:31,200 --> 01:03:33,837
deal with. This is not a 100 years time.

1125
01:03:33,920 --> 01:03:37,597
>> I'd also like to give a call to action
for people in this room. I agree that

1126
01:03:37,680 --> 01:03:41,917
this is a much more significant problem
than the idea that a Terminator is going

1127
01:03:42,000 --> 01:03:43,677
to emerge and kill us all.
>> No, it's not.

1128
01:03:43,760 --> 01:03:47,837
>> That's a lovely that's a,
you know, speculative fiction

1129
01:03:47,920 --> 01:03:51,597
that's great and worth thinking ahead
to the future, but it's speculative

1130
01:03:51,680 --> 01:03:54,557
fiction. Uh, the job loss is something

1131
01:03:54,640 --> 01:03:59,437
that we all play a role in, right? So,
I think that there's there should be moral

1132
01:03:59,520 --> 01:04:03,517
clarity here. If you are a business owner,
if you are a business leader, if

1133
01:04:03,600 --> 01:04:07,757
you have influence about hiring in your
company, In the not too distant future,

1134
01:04:07,840 --> 01:04:10,317
you'll be faced with a question which is

1135
01:04:10,400 --> 01:04:14,877
AI is suddenly making everyone twice
as productive. So the question for you

1136
01:04:14,960 --> 01:04:19,277
>> making half as many people twice as >>
the question for you will be do I want

1137
01:04:19,360 --> 01:04:23,117
my company to be more productive with
the same staff or do I want to be the

1138
01:04:23,200 --> 01:04:25,997
same level with half the staff. Right?

1139
01:04:26,080 --> 01:04:29,917
And that is a question that we as a
society need to make a decision about.

1140
01:04:30,000 --> 01:04:34,317
And I strongly encourage you any in any
position where you have influence over

1141
01:04:34,400 --> 01:04:38,797
this to go to the path of growth to try
to expand what your company can do with

1142
01:04:38,880 --> 01:04:42,317
the same staff. Like it's you don't need
to replace people. You need to figure

1143
01:04:42,400 --> 01:04:44,477
out what additional tasks they can

1144
01:04:44,560 --> 01:04:45,197
serve.

1145
01:04:45,280 --> 01:04:50,237
>> So Kate about this replacement of

1146
01:04:50,320 --> 01:04:53,757
of positive and artistic

1147
01:04:53,840 --> 01:04:56,557
>> talents. Let me take a middle road

1148
01:04:56,640 --> 01:05:00,157
there. Middle position road there.

1149
01:05:00,240 --> 01:05:03,917
There were teams of artists hired to

1150
01:05:04,000 --> 01:05:07,357
reproduce some idea that was set down by

1151
01:05:07,440 --> 01:05:11,677
a set designer or by some lead artist in

1152
01:05:11,760 --> 01:05:14,237
a in a project. Is that much different

1153
01:05:14,320 --> 01:05:18,957
from the the the the schools of art that

1154
01:05:19,040 --> 01:05:23,597
we know about from the Renaissance where
there's the main artist who's the great

1155
01:05:23,680 --> 01:05:25,677
artist and then everyone comes in and

1156
01:05:25,760 --> 01:05:29,677
and apes that style. So my sense of this

1157
01:05:29,760 --> 01:05:31,837
is and I could be wrong. You have we

1158
01:05:31,920 --> 01:05:35,037
have expertise here to correct me that

1159
01:05:35,120 --> 01:05:38,237
AI today is really good at copying

1160
01:05:38,320 --> 01:05:40,557
what's already on the internet. I can

1161
01:05:40,640 --> 01:05:44,557
tell Chad GPT or or Claude I say see the

1162
01:05:44,640 --> 01:05:48,317
scene here paint it the way Van Go might

1163
01:05:48,400 --> 01:05:50,637
have seen it and it will come back with

1164
01:05:50,720 --> 01:05:53,437
the same color palette brush strokes and

1165
01:05:53,520 --> 01:05:57,677
we'll be there in these swirly patterns
and we say that's what Van Go would have

1166
01:05:57,760 --> 01:06:00,157
painted. Now if I then ask it paint this

1167
01:06:00,240 --> 01:06:02,957
scene in a style of no artist who has

1168
01:06:03,040 --> 01:06:05,757
ever lived

1169
01:06:05,840 --> 01:06:08,637
>> what's it going to do then? And so for

1170
01:06:08,720 --> 01:06:14,157
me, I think AI in that space is boosting

1171
01:06:14,240 --> 01:06:16,237
the expectations we should have of

1172
01:06:16,320 --> 01:06:18,797
ourselves, of our true creativity.

1173
01:06:18,880 --> 01:06:22,957
Because true creativity is not doing
what someone else had done just a little

1174
01:06:23,040 --> 01:06:27,837
better. It's doing what no one else has
done. And I don't know if AI has that

1175
01:06:27,920 --> 01:06:29,597
ability

1176
01:06:29,680 --> 01:06:35,240
if it's being trained on what
everybody else has done.

1177
01:06:35,600 --> 01:06:38,157
I could I could on that a little bit here.

1178
01:06:38,240 --> 01:06:41,677
>> Uh training on, you know, I've also

1179
01:06:41,760 --> 01:06:43,997
heard uh it's just math, it's just test

1180
01:06:44,080 --> 01:06:48,557
text prediction. Um training on

1181
01:06:48,640 --> 01:06:53,357
text that humans produced does not requ
like like that is a harder problem to

1182
01:06:53,440 --> 01:06:57,437
predict that text than producing the text
in the first place. An example that

1183
01:06:57,520 --> 01:06:59,277
might resonate with people here is there

1184
01:06:59,360 --> 01:07:02,877
was a great astronomer Taiko Braheed

1185
01:07:02,960 --> 01:07:06,637
the position of of all the stars
and particularly Mars each night.

1186
01:07:06,720 --> 01:07:09,197
>> Planets. >> Uh yeah, sorry. Yeah.

1187
01:07:09,280 --> 01:07:10,717
>> All the stars are also Mars.

1188
01:07:10,800 --> 01:07:14,797
>> You're standing next to the
director of the Hayden Planetarium.

1189
01:07:14,880 --> 01:07:19,117
>> Okay. >> He also he
also recorded the stars.

1190
01:07:19,200 --> 01:07:23,677
>> He also recorded the stars. But >>
but Mars is the one that Johannes Kepler

1191
01:07:23,760 --> 01:07:25,517
tracked. >> Yes. >> Right.
And so Johannes Kepler, there was

1192
01:07:25,600 --> 01:07:30,557
a big heist. He stole the the the
journals, the data. Uh but but Tao Brahe

1193
01:07:30,640 --> 01:07:32,237
when he was recording the stars and

1194
01:07:32,320 --> 01:07:35,437
planets. Uh

1195
01:07:35,520 --> 01:07:38,077
he just looked at the night sky and

1196
01:07:38,160 --> 01:07:41,197
wrote down where Mars was. If you

1197
01:07:41,280 --> 01:07:45,997
imagine training something to predict
the entries in Brah's journal, it does

1198
01:07:46,080 --> 01:07:49,517
not have the luxury of looking at the sky.

1199
01:07:49,600 --> 01:07:54,157
You can solve this problem.
You can read a bunch of the entries in

1200
01:07:54,240 --> 01:07:58,637
Brahe's journal and predict the next.
To do so, you have to discover the laws of

1201
01:07:58,720 --> 01:08:01,037
the motions of the planets. Johannes

1202
01:08:01,120 --> 01:08:04,077
Kepler did this. Brahi never did. So, an

1203
01:08:04,160 --> 01:08:06,797
AI trained on only Brahi's journals,

1204
01:08:06,880 --> 01:08:08,637
trained on only the prediction, only

1205
01:08:08,720 --> 01:08:12,157
predicting the text that Brahi produced.

1206
01:08:12,240 --> 01:08:14,877
That training target incentivizes the AI

1207
01:08:14,960 --> 01:08:16,477
to figure out something that Brahhee

1208
01:08:16,560 --> 01:08:19,517
never did. Now, can AIs today do this?

1209
01:08:19,600 --> 01:08:22,877
That's one question. But training an AI

1210
01:08:22,960 --> 01:08:26,877
to predict things human wrote down
or humans wrote down does not limit it to

1211
01:08:26,960 --> 01:08:29,437
doing only things that humans can do.

1212
01:08:29,520 --> 01:08:34,557
And we are starting to see AIs, you know,
we've seen chat GPT uh or GPT 5.1

1213
01:08:34,640 --> 01:08:37,997
produce some novel physics contributions.
We're seeing them start

1214
01:08:38,080 --> 01:08:40,317
to inch into ways that that humans

1215
01:08:40,400 --> 01:08:43,597
haven't quite done things. Uh and you

1216
01:08:43,680 --> 01:08:45,517
know sometimes you know when a new

1217
01:08:45,600 --> 01:08:48,797
scientist uh comes up with a new idea

1218
01:08:48,880 --> 01:08:53,277
they aren't necessarily coming up with
a wholesale entirely new way of thinking

1219
01:08:53,360 --> 01:08:57,037
as opposed to making small amounts
of thinking progress and using a lot of

1220
01:08:57,120 --> 01:08:59,517
other thinking progress that that came
from other people. You know they're

1221
01:08:59,600 --> 01:09:04,637
standing on the shoulders of giants.
Uh so the things we're doing with AI can

1222
01:09:04,720 --> 01:09:07,117
get them there. There's a different
question of whether we're currently

1223
01:09:07,200 --> 01:09:09,357
there but it can get them there. Okay.

1224
01:09:09,440 --> 01:09:13,357
And this idea that they're just math
and so it won't get there. That's like

1225
01:09:13,440 --> 01:09:17,517
saying, "Oh, you know, a house cat
is just biochemistry, so no sort of cat is

1226
01:09:17,600 --> 01:09:22,237
dangerous." You know, a house cat
is just biochemistry and a house cat can't

1227
01:09:22,320 --> 01:09:27,880
kill you. But that doesn't
mean that tigers are fake.

1228
01:09:29,120 --> 01:09:31,917
>> Wow.

1229
01:09:32,000 --> 01:09:36,200
>> Wait, was that a logical leap?

1230
01:09:36,400 --> 01:09:39,357
>> You look like you wanted to move. I I

1231
01:09:39,440 --> 01:09:41,677
we're running short on time here. Uh I

1232
01:09:41,760 --> 01:09:45,277
want to uh get back to uh a very serious

1233
01:09:45,360 --> 01:09:50,477
point. Uh I I was appointed under Obama

1234
01:09:50,560 --> 01:09:53,037
to serve on the Defense Innovation

1235
01:09:53,120 --> 01:09:55,037
Board, a freshly created board of the

1236
01:09:55,120 --> 01:09:58,637
Pentagon that was tasked with advising

1237
01:09:58,720 --> 01:10:03,357
the Pentagon and all the ways technology
might influence the future present and

1238
01:10:03,440 --> 01:10:07,117
future warfare. Not just how many
missiles are in your silo or how many

1239
01:10:07,200 --> 01:10:09,997
push-ups the soldiers can do. Are there

1240
01:10:10,080 --> 01:10:11,997
other innovations that are possible?

1241
01:10:12,080 --> 01:10:14,477
Eric Schmidt was the chair of that

1242
01:10:14,560 --> 01:10:16,717
board. I was privileged to serve under

1243
01:10:16,800 --> 01:10:19,437
him for this. We toured uh military

1244
01:10:19,520 --> 01:10:22,477
bases around the world. Spoke with uh

1245
01:10:22,560 --> 01:10:25,677
soldiers uh very dedicated to their

1246
01:10:25,760 --> 01:10:29,677
>> every single one of them
wanted his autograph.

1247
01:10:29,760 --> 01:10:31,277
>> I'm serious. He would show up and it's

1248
01:10:31,360 --> 01:10:35,037
like all this is like oh my god.

1249
01:10:35,120 --> 01:10:37,837
Uh so Eric, right about the time uh

1250
01:10:37,920 --> 01:10:40,397
drones were becoming sort of more lethal

1251
01:10:40,480 --> 01:10:43,917
and more real and uh we felt it

1252
01:10:44,000 --> 01:10:47,437
necessary as a committee, as a board to

1253
01:10:47,520 --> 01:10:51,197
draft some kind of rules of engagement

1254
01:10:51,280 --> 01:10:53,837
regarding how and when AI would be

1255
01:10:53,920 --> 01:10:56,477
brought into the chain of decisions that

1256
01:10:56,560 --> 01:11:00,397
would lead in lethal lead to lethality.

1257
01:11:00,480 --> 01:11:02,957
And that was a code of ethics really.

1258
01:11:03,040 --> 01:11:07,117
And I was quite proud of that because
there was no telling what would happen

1259
01:11:07,200 --> 01:11:09,037
without such guidance. My question to

1260
01:11:09,120 --> 01:11:12,477
you is our motivation at the time was

1261
01:11:12,560 --> 01:11:17,197
the knowledge that AI was not perfect
that AI would make mistakes and you

1262
01:11:17,280 --> 01:11:19,357
never want someone to die because of a

1263
01:11:19,440 --> 01:11:21,677
mistake. Yet of course people make

1264
01:11:21,760 --> 01:11:25,517
mistakes. So at what point is AI good

1265
01:11:25,600 --> 01:11:29,357
enough in that space to say I trust AI

1266
01:11:29,440 --> 01:11:34,477
over a person and let AI run run with
the decision and not a human. Is it just

1267
01:11:34,560 --> 01:11:37,997
that at the end you want to have someone
to blame because you can't blame the

1268
01:11:38,080 --> 01:11:40,957
computer you can't put
the computer on trial

1269
01:11:41,040 --> 01:11:44,077
>> but a person you can so what what

1270
01:11:44,160 --> 01:11:46,877
happens when AI can find the enemy

1271
01:11:46,960 --> 01:11:51,517
perfectly. So everybody here has
an opportunity to read the report that we

1272
01:11:51,600 --> 01:11:55,357
issued under DIB um >> DIB
Defense Innovation Board

1273
01:11:55,440 --> 01:11:57,917
>> and the Pentagon
actually adopted it and

1274
01:11:58,000 --> 01:12:00,317
published it as their guidelines.

1275
01:12:00,400 --> 01:12:02,397
Roughly summarize what what our thinking

1276
01:12:02,480 --> 01:12:05,757
was at the time. Um

1277
01:12:05,840 --> 01:12:09,917
all war is terrible. All military
activity is terrible. It involves the

1278
01:12:10,000 --> 01:12:11,837
destruction of other people and so forth

1279
01:12:11,920 --> 01:12:14,317
and so on. It's just horrible. Trust me.

1280
01:12:14,400 --> 01:12:18,077
>> Um and the question is can If you have

1281
01:12:18,160 --> 01:12:20,317
to do it, can you do it using AI to be

1282
01:12:20,400 --> 01:12:25,197
more accurate and more successful? That
was the question. And we debated this at

1283
01:12:25,280 --> 01:12:26,877
some length and we concluded that at

1284
01:12:26,960 --> 01:12:28,797
least at the time, and I think it's true

1285
01:12:28,880 --> 01:12:32,557
today, AI is not reliable enough to rely

1286
01:12:32,640 --> 01:12:35,117
on it to make those decisions and it's

1287
01:12:35,200 --> 01:12:38,077
unlikely to be so for a very long time.

1288
01:12:38,160 --> 01:12:41,677
That was our conclusion. Uh sort
of makes sense, right? Because you can't

1289
01:12:41,760 --> 01:12:45,037
tolerate sort of a percentage of errors
where you might be able to tolerate them

1290
01:12:45,120 --> 01:12:47,917
in other areas. Um and so what we wrote

1291
01:12:48,000 --> 01:12:51,357
was that any use of automation in that

1292
01:12:51,440 --> 01:12:54,717
area uh essentially called lethality um

1293
01:12:54,800 --> 01:12:56,797
needed to have a couple things. It

1294
01:12:56,880 --> 01:13:01,917
needed to have been tested. It needed
to have a time and a space duration. And

1295
01:13:02,000 --> 01:13:05,997
most important there had to be a human
who was oversight overseeing it. They

1296
01:13:06,080 --> 01:13:09,997
had to know it was going to happen ahead
of time. They had to have tested that it

1297
01:13:10,080 --> 01:13:13,277
was going to happen and then it had
to reproduce. If it didn't then that was a

1298
01:13:13,360 --> 01:13:17,517
major problem. And that's ultimately
what the Pentagon agreed to. Um this is

1299
01:13:17,600 --> 01:13:20,477
a compromise because these tools
are very powerful but they also make

1300
01:13:20,560 --> 01:13:22,637
mistakes. So that's ultimately I think

1301
01:13:22,720 --> 01:13:27,837
where the Pentagon ended up and you
and is now that was some years ago now.

1302
01:13:27,920 --> 01:13:29,517
We're at least a decade forward from

1303
01:13:29,600 --> 01:13:34,957
that. Is AI good enough so that you
would have would have scripted this

1304
01:13:35,040 --> 01:13:37,357
differently? >> I don't I
don't believe you and I and

1305
01:13:37,440 --> 01:13:41,677
the others would have come to any other
conclusion on the lethality part. Uh the

1306
01:13:41,760 --> 01:13:44,637
lethality part is obviously the most
important one. It's the most important

1307
01:13:44,720 --> 01:13:48,877
decision that's made by members of the
military under military law and uh laws

1308
01:13:48,960 --> 01:13:51,597
of war and so forth and so on. Um what

1309
01:13:51,680 --> 01:13:54,797
is true is that the AI planning and

1310
01:13:54,880 --> 01:13:59,277
recognition that is the eyesight
and the ability to do planning is now far

1311
01:13:59,360 --> 01:14:01,357
superior than with humans. That's not

1312
01:14:01,440 --> 01:14:04,797
the same thing as lethality.

1313
01:14:04,880 --> 01:14:07,517
>> Okay. So these are
subtle differences in

1314
01:14:07,600 --> 01:14:09,917
the capacity of AI that are not always

1315
01:14:10,000 --> 01:14:14,557
recognized or appreciated.
>> Well, I I think we we all agree that

1316
01:14:14,640 --> 01:14:16,717
these systems are not perfect and at

1317
01:14:16,800 --> 01:14:19,037
least in things which are life critical

1318
01:14:19,120 --> 01:14:21,677
like deciding what to do with a patient,

1319
01:14:21,760 --> 01:14:24,477
right? Deciding when an airplane is in

1320
01:14:24,560 --> 01:14:29,357
trouble, uh anything involving military
kinetic conflict, uh think twice before

1321
01:14:29,440 --> 01:14:31,437
you empower AI because it makes

1322
01:14:31,520 --> 01:14:37,960
mistakes. I I I want
to thank you for that, Eric,

1323
01:14:38,720 --> 01:14:41,037
and thank you for your service in that.

1324
01:14:41,120 --> 01:14:43,357
Uh just to land this plane, what I want

1325
01:14:43,440 --> 01:14:46,637
to do is go down the panel and ask, you

1326
01:14:46,720 --> 01:14:49,837
know, given this 3 months horizon for

1327
01:14:49,920 --> 01:14:52,957
new stuff that happens uh in AI and that

1328
01:14:53,040 --> 01:14:56,397
whole sector. Uh, let's just fast

1329
01:14:56,480 --> 01:15:01,757
forward way into the future to 2030

1330
01:15:01,840 --> 01:15:07,357
and ask Wait, wait, wait. How long
ago was Chat GPT? How many years ago?

1331
01:15:07,440 --> 01:15:11,037
>> Three and a half. >> Three
and a half years ago was chat GPT

1332
01:15:11,120 --> 01:15:14,237
one or whatever the first public version

1333
01:15:14,320 --> 01:15:16,717
of it. So, let's fast forward three and

1334
01:15:16,800 --> 01:15:19,517
a half years to 2030.

1335
01:15:19,600 --> 01:15:22,077
uh what is the state of the world and

1336
01:15:22,160 --> 01:15:27,197
its relationship with AI?
Let's start with you.

1337
01:15:27,280 --> 01:15:31,197
>> Yeah. Well, I can tell you the vision
that we would like to see. Sure.

1338
01:15:31,280 --> 01:15:36,317
>> And and we have and you you're not just
passively hoping it'll get there. You're

1339
01:15:36,400 --> 01:15:38,077
actively working to achieve this,

1340
01:15:38,160 --> 01:15:40,797
>> right?
And and and it's a future that is

1341
01:15:40,880 --> 01:15:44,397
really amazing in the sense that

1342
01:15:44,480 --> 01:15:47,357
everyone is far more productive. I can

1343
01:15:47,440 --> 01:15:51,597
be in multiple places at one time
because there are agents that know my

1344
01:15:51,680 --> 01:15:56,317
preferences and my desires and they
repres represent me in various context.

1345
01:15:56,400 --> 01:15:57,917
>> Just to be clear, you guys use the word

1346
01:15:58,000 --> 01:16:03,997
agent to refer to an AI uh
representative of your needs and

1347
01:16:04,080 --> 01:16:06,477
activities. >> That's correct.
And so they might

1348
01:16:06,560 --> 01:16:09,197
negotiate a purchase arrangement for me

1349
01:16:09,280 --> 01:16:12,077
or uh be in a meeting and represent my

1350
01:16:12,160 --> 01:16:15,517
interest. And so the my productivity is

1351
01:16:15,600 --> 01:16:19,357
just orders of magnitude
greater than it is today.

1352
01:16:19,440 --> 01:16:23,517
>> And you're fine with that?
>> I'm fine with that provided that every

1353
01:16:23,600 --> 01:16:26,237
single one of my uses of AI and making

1354
01:16:26,320 --> 01:16:30,557
me more productive is actually able to

1355
01:16:30,640 --> 01:16:33,917
show that it's it gives me a warranty of

1356
01:16:34,000 --> 01:16:37,597
what it will do and how it will
represent itself. In other words, the

1357
01:16:37,680 --> 01:16:41,037
same thing I get when I get
an appliance, I get a warranty and a

1358
01:16:41,120 --> 01:16:43,757
compliance statement. I want that same

1359
01:16:43,840 --> 01:16:48,477
thing on these agents and I'll be able
to do that. In some ways, what you guys

1360
01:16:48,560 --> 01:16:50,557
were offering to the military, there are

1361
01:16:50,640 --> 01:16:55,597
also many societal harms that aren't
about immediate life-threatening, but

1362
01:16:55,680 --> 01:16:59,597
are about our ability to have
a democracy, about our ability to

1363
01:16:59,680 --> 01:17:04,477
understand truthful information and so
forth that are at stake and need those

1364
01:17:04,560 --> 01:17:08,397
same kind of guidance. It would be like
an underwriters laboratory certification

1365
01:17:08,480 --> 01:17:14,637
that we have for our actual appliances
but applied to your AI agents. Okay.

1366
01:17:14,720 --> 01:17:15,837
>> I think >> 2030

1367
01:17:15,920 --> 01:17:19,037
>> 2030 I I think we have two paths in

1368
01:17:19,120 --> 01:17:22,397
front of us. On one path, we move

1369
01:17:22,480 --> 01:17:24,877
towards a place where we see what is

1370
01:17:24,960 --> 01:17:27,677
already one of the most concentrated

1371
01:17:27,760 --> 01:17:32,477
industries in industrial in the
industrial history of of the planet

1372
01:17:32,560 --> 01:17:34,557
become even more concentrated and even

1373
01:17:34,640 --> 01:17:36,717
fewer people have their hands on levels

1374
01:17:36,800 --> 01:17:39,997
of power and democracies start to erode.

1375
01:17:40,080 --> 01:17:43,917
And these systems ultimately are making
decisions that will decide what kind of

1376
01:17:44,000 --> 01:17:47,997
job you can have, what kind of quality
of care you get, um what kind of options

1377
01:17:48,080 --> 01:17:51,597
you and your family will have
going forward. The other path is

1378
01:17:51,680 --> 01:17:55,677
>> just remind me is it there I'll get
these numbers long but the census is

1379
01:17:55,760 --> 01:17:57,277
accurate. There are four or five

1380
01:17:57,360 --> 01:18:00,877
companies on the S&P 500 that represent

1381
01:18:00,960 --> 01:18:03,757
like 30% of the total wealth

1382
01:18:03,840 --> 01:18:05,597
>> of the United States. Correct.

1383
01:18:05,680 --> 01:18:09,997
>> Some some number like that might
be precise but the census correct.

1384
01:18:10,080 --> 01:18:13,917
>> You're exactly right. And if you look
at the numbers of people who are actually

1385
01:18:14,000 --> 01:18:17,597
running these companies you're talking
about a handful a tiny handful of

1386
01:18:17,680 --> 01:18:21,277
billionaires. They are the people who
are getting to make these decisions

1387
01:18:21,360 --> 01:18:25,037
about how these tools will be used
and what problems they will be used to

1388
01:18:25,120 --> 01:18:29,917
solve. The other path is that we
realize this is the most important public

1389
01:18:30,000 --> 01:18:33,597
conversation that we can be having
right now. Which is why I'm so glad we're

1390
01:18:33,680 --> 01:18:35,197
gathered here today having this

1391
01:18:35,280 --> 01:18:37,437
discussion and we decide that actually

1392
01:18:37,520 --> 01:18:40,957
this is too important to be left in such

1393
01:18:41,040 --> 01:18:45,197
a small number of hands of people who
are not democratically accountable. you

1394
01:18:45,280 --> 01:18:48,957
asked what the difference was between
a representative and a and a tech row is

1395
01:18:49,040 --> 01:18:51,517
that we vote for one of them and we

1396
01:18:51,600 --> 01:18:54,957
don't vote for the other ones and

1397
01:18:55,040 --> 01:18:58,957
>> okay so take me to 2030

1398
01:18:59,040 --> 01:19:00,877
>> what I would love
to see by 2030 is that

1399
01:19:00,960 --> 01:19:03,437
we say this matters the public interest

1400
01:19:03,520 --> 01:19:08,157
of these technologies is at the core and
if it is going to be using our land our

1401
01:19:08,240 --> 01:19:12,077
energy and our water then we need
to have a say in it and these tools should

1402
01:19:12,160 --> 01:19:16,157
be used for the public good and that
actually means making this something

1403
01:19:16,240 --> 01:19:18,637
that serves all of us, not the few.

1404
01:19:18,720 --> 01:19:22,440
>> Are you running for office?

1405
01:19:25,520 --> 01:19:29,197
>> Chris, take me to 2030. What What
are you teaching in 2030 or not teaching

1406
01:19:29,280 --> 01:19:32,317
today? And by the way, uh Isaac Azimov

1407
01:19:32,400 --> 01:19:37,437
had his three laws of robotics. Do they
apply today? And should is there an AI

1408
01:19:37,520 --> 01:19:41,437
laws of robotics? >> There
is an AI laws of robotics. So as

1409
01:19:41,520 --> 01:19:45,117
laws and robotics for the non- nerds
in the audience and I say that you are all

1410
01:19:45,200 --> 01:19:47,997
nerds. Thank you for being here with us

1411
01:19:48,080 --> 01:19:50,477
are you know firstly that a robot must

1412
01:19:50,560 --> 01:19:52,557
do no harm to an individual including

1413
01:19:52,640 --> 01:19:55,517
through inaction. Secondly that a robot

1414
01:19:55,600 --> 01:19:57,757
must obey what the human says and

1415
01:19:57,840 --> 01:20:00,397
thirdly that the robot must not cause

1416
01:20:00,480 --> 01:20:02,797
harm to itself barring the other two. Of

1417
01:20:02,880 --> 01:20:05,757
course there was a zero. So B, you mean

1418
01:20:05,840 --> 01:20:09,357
it can't it can't cause harm

1419
01:20:09,440 --> 01:20:11,917
to its first two take precedent

1420
01:20:12,000 --> 01:20:15,117
>> with Yes. What to accomplish the third?

1421
01:20:15,200 --> 01:20:17,997
It's it cannot undo
the first or the second.

1422
01:20:18,080 --> 01:20:22,637
>> Exactly. And then there was a later
zero law which is the robots cannot harm

1423
01:20:22,720 --> 01:20:24,957
humanity which is the broader thing

1424
01:20:25,040 --> 01:20:27,357
rather than just an individual. So this

1425
01:20:27,440 --> 01:20:30,237
is a key challenge for AI. It's the

1426
01:20:30,320 --> 01:20:32,477
alignment problem.

1427
01:20:32,560 --> 01:20:35,597
And just I have to say so Isaac Asmer

1428
01:20:35,680 --> 01:20:37,997
wrote a series of short stories that

1429
01:20:38,080 --> 01:20:44,477
were collected into a book called I
Robot which then became a film starring um

1430
01:20:44,560 --> 01:20:48,557
>> Will Smith. >> Will Smith
and uh in there he explored

1431
01:20:48,640 --> 01:20:51,117
these rules of engagement for robots

1432
01:20:51,200 --> 01:20:53,437
very early. This is the early 1950s. So

1433
01:20:53,520 --> 01:20:56,637
we're talking about 70 years ago plus 75

1434
01:20:56,720 --> 01:20:59,037
years ago. Yeah, again this is the place

1435
01:20:59,120 --> 01:21:00,957
where specular friction thinks about the

1436
01:21:01,040 --> 01:21:03,757
future in a very useful way. So uh AI

1437
01:21:03,840 --> 01:21:07,997
alignment is exactly these laws. So how
do we think about making a systems that

1438
01:21:08,080 --> 01:21:12,317
helpful? How do we system think about
making a system that's safe? How do we

1439
01:21:12,400 --> 01:21:13,917
ensure that it's for the benefit of

1440
01:21:14,000 --> 01:21:16,237
humanity rather than to its detriment?

1441
01:21:16,320 --> 01:21:19,677
So as we go into 2030, I think that

1442
01:21:19,760 --> 01:21:22,397
those key elements that are similar to

1443
01:21:22,480 --> 01:21:27,437
Azimoff's laws become the key things
that guide the success or failure of the

1444
01:21:27,520 --> 01:21:30,877
AI endeavor. >> And that could happen
within the next three and a half years.

1445
01:21:30,960 --> 01:21:32,877
>> It's it's happening now. There are many

1446
01:21:32,960 --> 01:21:34,957
many thoughtful researchers working on

1447
01:21:35,040 --> 01:21:37,997
these. You can check out the uh talk by

1448
01:21:38,080 --> 01:21:40,797
Amanda Ascal, the chief philosopher for

1449
01:21:40,880 --> 01:21:43,997
Anthropic, about how she's crafted

1450
01:21:44,080 --> 01:21:48,477
Claude's soul document, the thing
that governs its behavior and its code of

1451
01:21:48,560 --> 01:21:50,957
ethics. It's incredibly thoughtful and

1452
01:21:51,040 --> 01:21:56,637
it gives me hope for the future.
Did Google have a philosopher?

1453
01:21:56,720 --> 01:22:00,157
>> Very good. >> Okay.

1454
01:22:00,240 --> 01:22:03,037
>> Uh thank you for that. So Cindy, what

1455
01:22:03,120 --> 01:22:04,877
you got going for us? 2030

1456
01:22:04,960 --> 01:22:07,917
>> 2030 a few years from now um I think

1457
01:22:08,000 --> 01:22:10,157
that we'll see um you know a lot of

1458
01:22:10,240 --> 01:22:12,477
these technologies we have now the the

1459
01:22:12,560 --> 01:22:16,237
kind of LLMs will become um more useful

1460
01:22:16,320 --> 01:22:20,557
in our day-to-day life right now
you know it can kind of respond to to

1461
01:22:20,640 --> 01:22:22,717
questions or to to prompts we give it

1462
01:22:22,800 --> 01:22:24,477
and maybe it'll become a little bit

1463
01:22:24,560 --> 01:22:27,837
closer to be able to um responding to

1464
01:22:27,920 --> 01:22:29,917
you know queries more like achieve this

1465
01:22:30,000 --> 01:22:32,477
objective so I can imagine if you know

1466
01:22:32,560 --> 01:22:34,877
specialized versions of this like you

1467
01:22:34,960 --> 01:22:37,357
know help you respond to emails and

1468
01:22:37,440 --> 01:22:39,997
these sorts of things. um feet on the

1469
01:22:40,080 --> 01:22:42,477
ground. I think we maybe will see uh

1470
01:22:42,560 --> 01:22:44,637
self-driving cars in New York City in a

1471
01:22:44,720 --> 01:22:47,997
few years, you know. Um I I think uh you

1472
01:22:48,080 --> 01:22:50,797
know that a lot of the the AI will be

1473
01:22:50,880 --> 01:22:53,277
trained on more specialized data sets.

1474
01:22:53,360 --> 01:22:55,997
So they'll become better at um certain

1475
01:22:56,080 --> 01:22:58,717
tasks like uh you know scientific

1476
01:22:58,800 --> 01:23:01,437
research or mathematical research. um

1477
01:23:01,520 --> 01:23:06,957
you know if you you train on kind of
certain parts of the the internet it may

1478
01:23:07,040 --> 01:23:08,317
be better than what it is now.

1479
01:23:08,400 --> 01:23:13,117
>> Quickly Kate if AI got that specialized
wouldn't that be more efficient on the

1480
01:23:13,200 --> 01:23:16,717
grid? >> It it could be.
Um this is one of the

1481
01:23:16,800 --> 01:23:20,557
really big questions. I mean right now
large scale models are being trained

1482
01:23:20,640 --> 01:23:22,717
using a lot of fossil fuels. A lot in

1483
01:23:22,800 --> 01:23:27,997
fact more than previously. So if we had
small models, if they were specialized,

1484
01:23:28,080 --> 01:23:30,557
there are ways forward where
you can use much less energy.

1485
01:23:30,640 --> 01:23:31,837
>> That's what I would have figured. And

1486
01:23:31,920 --> 01:23:34,237
for example, in the Jetsons, the

1487
01:23:34,320 --> 01:23:38,237
documentary The Jetsons of the Future

1488
01:23:38,320 --> 01:23:40,957
that they as as fun and creative as they

1489
01:23:41,040 --> 01:23:46,237
were, they did not realize that maybe
the car could one day drive itself. The

1490
01:23:46,320 --> 01:23:49,917
human drove it or you get a robot
to drive it, but the car is the robot. And

1491
01:23:50,000 --> 01:23:52,637
that's a very specific invocation of

1492
01:23:52,720 --> 01:23:55,997
this. So I based on the title of your

1493
01:23:56,080 --> 01:24:01,800
book, I'm going to show
people your book because

1494
01:24:02,400 --> 01:24:05,277
I got Kate's book here too, Kate.

1495
01:24:05,360 --> 01:24:09,917
Atlas of AI, uh the Pub
Day 2021, but it's still

1496
01:24:10,000 --> 01:24:12,797
happening, >> apparently.
So they just went to the the

1497
01:24:12,880 --> 01:24:14,397
10th edition, so it's doing all right.

1498
01:24:14,480 --> 01:24:17,597
>> Yeah. Okay, good. If anyone builds it,

1499
01:24:17,680 --> 01:24:22,477
everyone dies. This has got to be
the most depressing title ever.

1500
01:24:22,560 --> 01:24:26,840
It starts with if >> adverb.

1501
01:24:27,280 --> 01:24:30,237
>> Well, that makes it all better. Okay.

1502
01:24:30,320 --> 01:24:32,797
Uh, so take me to 2030. What's going on?

1503
01:24:32,880 --> 01:24:35,437
Are we all dead?

1504
01:24:35,520 --> 01:24:39,677
You know, if you step back from this

1505
01:24:39,760 --> 01:24:43,357
from from the from the AI argument,
you just like look broadly at what's going

1506
01:24:43,440 --> 01:24:47,837
on. The machines are talking now.

1507
01:24:47,920 --> 01:24:51,677
You know, it people used to think it
was going to take a while to get to the

1508
01:24:51,760 --> 01:24:53,917
point where the machines are talking
and we got these people being like, "Oh

1509
01:24:54,000 --> 01:24:55,277
yeah, we're just growing these things."

1510
01:24:55,360 --> 01:24:57,837
Turns out they talk. They're solving
problems we said weren't going to happen

1511
01:24:57,920 --> 01:25:02,237
for 50 years. Now we're trying to take
it to super intelligence. That's kind of

1512
01:25:02,320 --> 01:25:04,877
crazy. >> It's kind of crazy.
You know, you're

1513
01:25:04,960 --> 01:25:07,517
like, "Do you guys know what's going
on there?" They're like, "No, but we think

1514
01:25:07,600 --> 01:25:10,157
it'll be fine. Trust us." Right.

1515
01:25:10,240 --> 01:25:13,917
>> In Silicon Valley, >> we're
not trusting you, just so you know.

1516
01:25:14,000 --> 01:25:16,557
>> Great. >> Okay. Um,

1517
01:25:16,640 --> 01:25:20,077
in Silicon Valley, people are spooked.

1518
01:25:20,160 --> 01:25:23,757
You know, there's a joke in Silicon
Valley that when someone leaves a normal

1519
01:25:23,840 --> 01:25:27,517
tech job, they say, you know, I've had
a great time here and now it's on to the

1520
01:25:27,600 --> 01:25:29,357
next adventure. And when someone leaves

1521
01:25:29,440 --> 01:25:32,077
an AI tech job, they say, I have stared

1522
01:25:32,160 --> 01:25:34,557
into the abyss.

1523
01:25:34,640 --> 01:25:41,560
I'm retiring to write poetry.
Please spend time with your families.

1524
01:25:41,920 --> 01:25:45,917
>> Okay. And if you think I'm joking,
you can look at the resignation of the

1525
01:25:46,000 --> 01:25:48,317
anthropic safety lead a few weeks ago

1526
01:25:48,400 --> 01:25:51,117
and see how close that joke is. But the

1527
01:25:51,200 --> 01:25:56,717
thing is, in Washington DC, they are
not spooked. In Washington, DC, they are

1528
01:25:56,800 --> 01:26:00,477
thinking that this is all about
self-driving cars and autonomous weapons

1529
01:26:00,560 --> 01:26:02,637
and job loss. And those are important

1530
01:26:02,720 --> 01:26:06,637
issues. But what I hope we see in 2030

1531
01:26:06,720 --> 01:26:09,437
is that DC is spooked.

1532
01:26:09,520 --> 01:26:13,757
I hope we see a world where DC has
realized that the people running the

1533
01:26:13,840 --> 01:26:18,157
labs, the academics outside the labs
like Jeffrey Hinton, that the people

1534
01:26:18,240 --> 01:26:22,637
inside the labs that everyone who's
close to this technology is saying like,

1535
01:26:22,720 --> 01:26:25,677
"Oh my gosh, the super intelligence
stuff could be really dangerous." Even

1536
01:26:25,760 --> 01:26:28,717
Eric was saying, you know, there's red
lines and if we start to get close to

1537
01:26:28,800 --> 01:26:30,157
these things, we need to watch out.

1538
01:26:30,240 --> 01:26:33,357
>> Even Eric said that.

1539
01:26:33,440 --> 01:26:36,317
>> Just briefly, you you mentioned uh Dr.

1540
01:26:36,400 --> 01:26:40,077
Hinton. uh he's one of the fathers of of

1541
01:26:40,160 --> 01:26:43,437
uh of neur neural nets >> deep learning

1542
01:26:43,520 --> 01:26:45,917
>> uh deep learning
and uh he was given the

1543
01:26:46,000 --> 01:26:49,997
Nobel Prize was it in chemistry
I think >> I thought physics

1544
01:26:50,080 --> 01:26:53,597
>> oh phys physics thank you the Nobel
Prize in physics just a few years ago

1545
01:26:53,680 --> 01:26:55,837
and we were fortunate enough to get him

1546
01:26:55,920 --> 01:26:58,237
on to my podcast for an interview uh

1547
01:26:58,320 --> 01:27:00,957
it's online now uh with Jeffrey Hinton

1548
01:27:01,040 --> 01:27:06,397
talking about the the the original
thinkings of AI and his reflections on

1549
01:27:06,480 --> 01:27:09,357
where it is where it might >>
those three people did their work at

1550
01:27:09,440 --> 01:27:12,077
Google.

1551
01:27:12,160 --> 01:27:14,237
>> So,

1552
01:27:14,320 --> 01:27:19,517
take me to 2030. This is DC AI sensible

1553
01:27:19,600 --> 01:27:24,317
by 2030 cuz we're in the middle
of whoever the next president will be.

1554
01:27:24,400 --> 01:27:27,597
>> Yeah. So, I I think >> presuming
it won't beat Trump in a third

1555
01:27:27,680 --> 01:27:31,517
term. Just have to That's
an if statement. Okay.

1556
01:27:31,600 --> 01:27:35,757
>> I think I think they could be.
You know, we we see a lot of people in the

1557
01:27:35,840 --> 01:27:39,597
industry saying that this is horribly
riskier than any other technology we

1558
01:27:39,680 --> 01:27:43,437
work with. And we see people at these
labs saying, you know, I think there's a

1559
01:27:43,520 --> 01:27:46,077
good chance this kills everybody, but I
got to do it because I'm better than the

1560
01:27:46,160 --> 01:27:47,757
next guy. And, you know, maybe one of

1561
01:27:47,840 --> 01:27:49,037
them's right. Probably two of them

1562
01:27:49,120 --> 01:27:53,917
aren't. But, uh, in in 2030, if if we

1563
01:27:54,000 --> 01:27:58,797
keep having this conversation, if it
becomes more salient, I think we see uh

1564
01:27:58,880 --> 01:28:01,757
people around the world realizing that we
shouldn't do this super intelligence

1565
01:28:01,840 --> 01:28:05,757
stuff. And the world does not need
to look that different from today. You

1566
01:28:05,840 --> 01:28:09,357
don't need to stop the self-driving cars.
You don't need to stop Alphaold

1567
01:28:09,440 --> 01:28:13,037
doing the drug discovery. You don't
even need to stop the autonomous weapons.

1568
01:28:13,120 --> 01:28:16,797
Society can have a debate about,
you know, do we want to to save people on

1569
01:28:16,880 --> 01:28:19,117
the front lines? Do we want to, you

1570
01:28:19,200 --> 01:28:23,917
know, not do this, not go down this path.
But uh but we could stop just the

1571
01:28:24,000 --> 01:28:26,077
race to build smarter and smarter AIs

1572
01:28:26,160 --> 01:28:28,157
that nobody understands that would be

1573
01:28:28,240 --> 01:28:31,517
easier than nuclear arms control. And I

1574
01:28:31,600 --> 01:28:35,757
hope that in 2030 leaders around the world
will realize that this is lethally

1575
01:28:35,840 --> 01:28:37,997
dangerous and we'll be putting a treaty

1576
01:28:38,080 --> 01:28:40,477
in place to say we're not rushing to

1577
01:28:40,560 --> 01:28:43,277
smarter than human AIs with anything

1578
01:28:43,360 --> 01:28:45,037
remotely like the modern technology or

1579
01:28:45,120 --> 01:28:47,677
we have no idea what's going on there.

1580
01:28:47,760 --> 01:28:52,077
>> All right, Eric, >> you
asked a question about 2030.

1581
01:28:52,160 --> 01:28:54,157
>> Yes, I did.

1582
01:28:54,240 --> 01:28:57,597
>> Why do we bet there? Why do we bet that

1583
01:28:57,680 --> 01:29:00,717
American democracy will survive?

1584
01:29:00,800 --> 01:29:02,797
>> Okay, so that's an important reason.

1585
01:29:02,880 --> 01:29:07,597
>> So I'm I'm gonna make a bet
that American democracy will actually

1586
01:29:07,680 --> 01:29:10,477
survive. We'll get through whatever
the current problems are, of which there

1587
01:29:10,560 --> 01:29:14,077
have been many listed here. That the
excesses will be addressed through

1588
01:29:14,160 --> 01:29:18,317
legislation. I know that's hard
to imagine right now, right? But I'm going

1589
01:29:18,400 --> 01:29:22,317
to bet that democracy will ultimately
reflect the will of the voters, which is

1590
01:29:22,400 --> 01:29:24,957
how it's supposed to work. So in that

1591
01:29:25,040 --> 01:29:30,157
scenario, which I know is difficult
to imagine at the moment, uh let's think

1592
01:29:30,240 --> 01:29:35,037
about some of the other benefits of all
of this. The likelihood of students

1593
01:29:35,120 --> 01:29:36,957
being much better educated because of

1594
01:29:37,040 --> 01:29:39,517
their ability to be tutored in a using

1595
01:29:39,600 --> 01:29:44,317
AI and specialized way is enormous.
The ability to have the next generation,

1596
01:29:44,400 --> 01:29:46,477
your kids and grandchildren, being so

1597
01:29:46,560 --> 01:29:48,877
much smarter than we will ever be is

1598
01:29:48,960 --> 01:29:51,597
pretty high. Furthermore, this issue

1599
01:29:51,680 --> 01:29:56,237
about jobs, there's no question there's
going to be job dislocation. But last

1600
01:29:56,320 --> 01:29:59,437
time I checked, every one of those
predictions has been false for the

1601
01:29:59,520 --> 01:30:04,477
reasons Neil mentioned earlier, which is
the American economy is fairly efficient

1602
01:30:04,560 --> 01:30:08,317
recycling jobs. And there's definitely
harm. There's definitely people hurt.

1603
01:30:08,400 --> 01:30:12,317
But at the end of the day, people
working with computers today make more

1604
01:30:12,400 --> 01:30:15,277
money than people without. Furthermore,

1605
01:30:15,360 --> 01:30:18,397
let's bet that these other issues, the

1606
01:30:18,480 --> 01:30:22,317
regulatory issues, the techro issues
are resolved through some combination of

1607
01:30:22,400 --> 01:30:24,797
self-interest, shareholder values, legal

1608
01:30:24,880 --> 01:30:28,397
liability, and regulation. In that

1609
01:30:28,480 --> 01:30:32,157
model, in 2030, every one of you will

1610
01:30:32,240 --> 01:30:35,917
have an assistant, a savant that allows

1611
01:30:36,000 --> 01:30:40,877
you under your control, not spying on you,
not doing something bad, under your

1612
01:30:40,960 --> 01:30:44,717
control to do the things that you most
care about. You're a physicist, right?

1613
01:30:44,800 --> 01:30:48,477
Neil's a physicist. You have
a physician's assistant who overnight does

1614
01:30:48,560 --> 01:30:51,597
your thinking while you're busy sleeping
and tells you what is interesting in the

1615
01:30:51,680 --> 01:30:55,117
day and you love it and you can turn
it off if you don't like it. There's

1616
01:30:55,200 --> 01:30:59,997
example after example. The most likely
scenario by 2030 is something analogous

1617
01:31:00,080 --> 01:31:03,357
to what I've described. All of the
things that everybody's worried about

1618
01:31:03,440 --> 01:31:06,877
are true at some level or they're point

1619
01:31:06,960 --> 01:31:13,400
issues that have to be debated.
In my book, which you did not feature,

1620
01:31:13,520 --> 01:31:17,037
I say very very clearly.
>> Tell us the name of the book.

1621
01:31:17,120 --> 01:31:21,197
>> Genesis. >> Thank you. >> Including
with myself and Henry Crew and

1622
01:31:21,280 --> 01:31:24,317
a few others. >> Um was co-author.

1623
01:31:24,400 --> 01:31:26,877
>> Yes. Yes. >> When did you write?
He'd been dead for

1624
01:31:26,960 --> 01:31:29,837
when did

1625
01:31:29,920 --> 01:31:31,837
>> Okay.

1626
01:31:31,920 --> 01:31:35,837
The AI Henry Kissinger. Okay. Wow.

1627
01:31:35,920 --> 01:31:38,477
>> In fairness to Henry,
he he finished the

1628
01:31:38,560 --> 01:31:41,197
book The Dable Read

1629
01:31:41,280 --> 01:31:45,757
>> and I encourage you to read it. Uh his
last book of his many many great books.

1630
01:31:45,840 --> 01:31:49,677
The important thing thinking about this
is really important and our work. What

1631
01:31:49,760 --> 01:31:52,557
we said is this conversation needs to

1632
01:31:52,640 --> 01:31:56,957
occur over and over again because it is
the future of us. It is us. It is how we

1633
01:31:57,040 --> 01:32:01,437
operate and I endorse that. Uh but at
the end of the day, right, we will get

1634
01:32:01,520 --> 01:32:05,357
through this and more importantly, we
will thrive. Do you know why? Because at

1635
01:32:05,440 --> 01:32:07,517
least American capitalism, the way it's

1636
01:32:07,600 --> 01:32:09,997
structured with regulation, produces

1637
01:32:10,080 --> 01:32:12,717
enormous economic and life value for us.

1638
01:32:12,800 --> 01:32:16,557
Think about it. Solving climate change,
solving the various issues we have, and

1639
01:32:16,640 --> 01:32:19,997
remember, we're depopulated, especially
because we're not allowing immigrants

1640
01:32:20,080 --> 01:32:24,157
in, which is an error, right?
So, you get the idea.

1641
01:32:24,240 --> 01:32:29,800
>> Wow. Let me see if I
can put a ball on this.

1642
01:32:31,360 --> 01:32:34,077
So

1643
01:32:34,160 --> 01:32:36,637
>> I try to always

1644
01:32:36,720 --> 01:32:41,837
generate a cosmic perspective on all
that happens to me in a day and to get

1645
01:32:41,920 --> 01:32:43,517
this kind of exposure. I don't know how

1646
01:32:43,600 --> 01:32:46,557
to tie this up in a bow, all this in

1647
01:32:46,640 --> 01:32:51,357
one, but allow me to say that and I've

1648
01:32:51,440 --> 01:32:54,557
been in computing for most of my life

1649
01:32:54,640 --> 01:32:56,717
that I've had access to it and awareness

1650
01:32:56,800 --> 01:33:00,317
of it and and I remember, you know, the

1651
01:33:00,400 --> 01:33:03,037
Turing test. People said if you interact

1652
01:33:03,120 --> 01:33:04,797
with a computer and you don't know if

1653
01:33:04,880 --> 01:33:07,837
there's a human on the other side then

1654
01:33:07,920 --> 01:33:11,677
it you why distinguish it from being a

1655
01:33:11,760 --> 01:33:15,437
computer or a human that passed
the touring test. The original research

1656
01:33:15,520 --> 01:33:17,597
paper was called the imitation game

1657
01:33:17,680 --> 01:33:19,757
which became the title of the movie

1658
01:33:19,840 --> 01:33:22,797
featuring the life of Alan Turing which

1659
01:33:22,880 --> 01:33:26,877
I recommend. Uh so we're way past that.

1660
01:33:26,960 --> 01:33:30,797
All right. And so I remember thinking,
"Oh, we'll never have a computer as good

1661
01:33:30,880 --> 01:33:34,797
as people at chess." We're way past that.

1662
01:33:34,880 --> 01:33:37,597
I remembered when Siri showed up on the

1663
01:33:37,680 --> 01:33:41,997
iPhone. There were whole commercials
with like who was the movie director who

1664
01:33:42,080 --> 01:33:46,877
was interacting with it? Martin Scorsese
is talking to Siri and while our jaws

1665
01:33:46,960 --> 01:33:48,477
are dropped open, there's this

1666
01:33:48,560 --> 01:33:51,597
intelligent agent serving him in the

1667
01:33:51,680 --> 01:33:54,637
voice of Siri in his iPhone. And I

1668
01:33:54,720 --> 01:33:57,277
think, "Wow, the future is here.

1669
01:33:57,360 --> 01:34:01,437
Today, no one talks about Siri as AI.

1670
01:34:01,520 --> 01:34:04,317
We're so past this. No, that's just old.

1671
01:34:04,400 --> 01:34:06,717
That's not even interesting. Alph Go,

1672
01:34:06,800 --> 01:34:08,557
that's yesterday.

1673
01:34:08,640 --> 01:34:11,757
Uh Alpha Zero, yesterday.

1674
01:34:11,840 --> 01:34:14,557
So, what intrigues me is every one of

1675
01:34:14,640 --> 01:34:17,677
these incremental advances in AI, we

1676
01:34:17,760 --> 01:34:21,277
absorb just becomes part of life and

1677
01:34:21,360 --> 01:34:24,077
we're looking at that next frontier.

1678
01:34:24,160 --> 01:34:27,597
Yes. the AGI, the the artificial general

1679
01:34:27,680 --> 01:34:30,157
intelligence. That's a frontier that may

1680
01:34:30,240 --> 01:34:33,277
be dangerous, but all the rest we

1681
01:34:33,360 --> 01:34:37,437
absorbed. We think nothing
of putting on GPS and

1682
01:34:37,520 --> 01:34:39,837
having it tell us the shortest route to

1683
01:34:39,920 --> 01:34:42,157
grandma's house, corrected for traffic,

1684
01:34:42,240 --> 01:34:44,957
and there's no human in that loop. We're

1685
01:34:45,040 --> 01:34:49,757
all fine with that.
We're not spooked by it. You're not

1686
01:34:49,840 --> 01:34:51,917
writing to your congressman about it.

1687
01:34:52,000 --> 01:34:53,757
You don't try to legislate it. It's just

1688
01:34:53,840 --> 01:34:56,237
there in our lives.

1689
01:34:56,320 --> 01:34:59,917
So I foresee by 2030, if I will, much

1690
01:35:00,000 --> 01:35:04,637
more AI just folded into our lives.

1691
01:35:04,720 --> 01:35:08,717
But the AI that transcends all of our

1692
01:35:08,800 --> 01:35:11,357
intelligence, the one that you might

1693
01:35:11,440 --> 01:35:14,077
think is farther away than it might be

1694
01:35:14,160 --> 01:35:18,077
given the title of this book.

1695
01:35:18,160 --> 01:35:20,397
>> Uh, I have to

1696
01:35:20,480 --> 01:35:22,957
I have to agree with Nate. I'm old

1697
01:35:23,040 --> 01:35:27,997
enough to have lived through a lot
of the Cold War. You know what kind of

1698
01:35:28,080 --> 01:35:29,677
ended the Cold War? A lot of things

1699
01:35:29,760 --> 01:35:31,757
contributed. But you know what mattered?

1700
01:35:31,840 --> 01:35:35,757
The realization that if you launch

1701
01:35:35,840 --> 01:35:38,077
nuclear weapons, intercontinental

1702
01:35:38,160 --> 01:35:41,357
ballistic missiles from the Soviet Union

1703
01:35:41,440 --> 01:35:44,317
into the Western Hemisphere and we see

1704
01:35:44,400 --> 01:35:46,877
that and we retaliate,

1705
01:35:46,960 --> 01:35:49,757
everybody dies.

1706
01:35:49,840 --> 01:35:53,037
We had an acronym for it. Matt, mutual

1707
01:35:53,120 --> 01:35:55,757
assured destruction. That's what brought

1708
01:35:55,840 --> 01:35:57,837
people to the table. Said this is

1709
01:35:57,920 --> 01:36:00,477
insane. We have to regulate it. We have

1710
01:36:00,560 --> 01:36:05,277
to reduce the stockp because preserving
our species is what is what matters

1711
01:36:05,360 --> 01:36:06,797
here.

1712
01:36:06,880 --> 01:36:11,117
So I based on that evidence, I have some

1713
01:36:11,200 --> 01:36:15,517
confidence that as those
dangers become more and

1714
01:36:15,600 --> 01:36:17,837
more real,

1715
01:36:17,920 --> 01:36:19,997
people will come to the table and say,

1716
01:36:20,080 --> 01:36:23,037
"Yeah, keep the rest of the AI going, we

1717
01:36:23,120 --> 01:36:27,837
got new medicines and new understandings
of our physiology and new technologies

1718
01:36:27,920 --> 01:36:29,917
that help us get smarter and healthier."

1719
01:36:30,000 --> 01:36:34,877
Yes, but that branch of AI is lethal. We

1720
01:36:34,960 --> 01:36:37,677
have to do something about that. No one

1721
01:36:37,760 --> 01:36:40,637
should build it and everyone needs to

1722
01:36:40,720 --> 01:36:45,277
agree to that by treaty. Treaties are
not perfect, but they're the best we

1723
01:36:45,360 --> 01:36:47,437
have as humans.

1724
01:36:47,520 --> 01:36:51,117
I would end by saying there's a quote

1725
01:36:51,200 --> 01:36:54,557
from Ray Bradberry who I verified with.

1726
01:36:54,640 --> 01:36:58,397
I met him only once and I just have
to verify this quote. He said there's a

1727
01:36:58,480 --> 01:37:02,797
woman once came up to me and said
science fiction author Ray Bradberry uh

1728
01:37:02,880 --> 01:37:05,197
Mr. Bradberry, why do you write these

1729
01:37:05,280 --> 01:37:08,797
stories about these apocalyptic futures?

1730
01:37:08,880 --> 01:37:11,917
Is this the future you think humanity is

1731
01:37:12,000 --> 01:37:16,797
going to inherit? He says, "No, I write

1732
01:37:16,880 --> 01:37:19,917
these apocalyptic futures so you know to

1733
01:37:20,000 --> 01:37:22,317
avoid them."

1734
01:37:22,400 --> 01:37:25,517
That is a cosmic perspective. Join me in

1735
01:37:25,600 --> 01:37:30,040
thanking our panel. Oh my god.

1736
01:37:32,000 --> 01:37:36,680
Excuse me.

1737
01:37:49,200 --> 01:37:54,557
We moved along, but as as always,
we allocate a few minutes for audience

1738
01:37:54,640 --> 01:37:55,917
questions. So, if we can bring the

1739
01:37:56,000 --> 01:38:01,920
lights up, we have a microphone
in each aisle. We'll try to go quickly