1
00:00:25,124 --> 00:00:28,866
So there's a lot of ground to cover, right? AI is a really broad

2
00:00:28,946 --> 00:00:32,449
topic. But I want to start with the book, Taming Silicon Valley. And

3
00:00:32,489 --> 00:00:35,851
I specifically want to start with the very end of the book, where in

4
00:00:35,891 --> 00:00:39,133
the acknowledgments, you write how you wrote

5
00:00:39,173 --> 00:00:42,676
this book out of a sense of urgency and disillusionment at lightning speed.

6
00:00:43,076 --> 00:00:46,198
And I'm wondering about, obviously, a lot of stuff has

7
00:00:46,218 --> 00:00:50,121
been going on, but what kind of inspired that sense of urgency?

8
00:00:51,669 --> 00:00:54,711
Well, I think you and I first spoke around the time that I

9
00:00:54,752 --> 00:00:58,094
gave Senate testimony about a year ago. And I think even

10
00:00:58,154 --> 00:01:01,837
then, everybody knew that having a good AI policy

11
00:01:01,877 --> 00:01:05,080
was urgent for the nation. A lot of what we talked about that

12
00:01:05,120 --> 00:01:08,302
day when I was at the Senate was the fact that with social media, the

13
00:01:08,322 --> 00:01:11,945
Senate had not done well. It waited too long, made some bad choices,

14
00:01:12,326 --> 00:01:16,409
and those bad choices became kind of enshrined permanently.

15
00:01:17,143 --> 00:01:20,744
And we don't want to do that with AI. And everybody in the room seemed to understand that.

16
00:01:21,244 --> 00:01:25,126
And yet, I watched the clock and the calendar tick away and

17
00:01:25,206 --> 00:01:28,307
saw that nothing was really happening in the United States. The EU has the

18
00:01:28,387 --> 00:01:31,728
AI Act there. And I think that's to the good. It may not be perfect, but

19
00:01:31,748 --> 00:01:35,049
it's a great start. The state of California is actually trying to

20
00:01:35,089 --> 00:01:38,851
do something now. The federal government, there

21
00:01:38,951 --> 00:01:42,212
is the executive order, which is great, but it's not law. The

22
00:01:42,252 --> 00:01:45,863
way our country works is the president can't make laws. are

23
00:01:45,983 --> 00:01:49,284
actual laws here. Anything a president puts in executive order

24
00:01:49,304 --> 00:01:52,645
can just vanish overnight. And I

25
00:01:52,725 --> 00:01:56,227
increasingly had this sense that nothing was going to happen in the United States

26
00:01:56,267 --> 00:02:00,568
Senate, even though many individual members were in favor. Chuck

27
00:02:00,588 --> 00:02:04,429
Schumer was taking his time with all these listening meetings.

28
00:02:04,869 --> 00:02:08,131
And I sensed that. And what I predicted would happen is

29
00:02:08,171 --> 00:02:11,832
that nothing would happen. And that's, in fact, what has happened. So nothing

30
00:02:11,872 --> 00:02:15,361
has happened. on the policy side, but all

31
00:02:15,381 --> 00:02:19,143
of the things that I worried about at the Senate have kind of materialized. So

32
00:02:19,203 --> 00:02:23,025
I worried that deepfakes would become a huge problem, that misinformation

33
00:02:23,065 --> 00:02:26,267
would become a big problem. And we've seen all those kinds of

34
00:02:26,287 --> 00:02:29,909
things happening. We've seen people manipulate

35
00:02:29,969 --> 00:02:33,651
markets. When I said that, it never happened, but it's already happened

36
00:02:34,151 --> 00:02:37,337
in the US and India. I guess in the U.S.

37
00:02:37,357 --> 00:02:41,199
we're not sure about the motives, but we at least saw a

38
00:02:41,240 --> 00:02:44,562
deepfake move in the market briefly. It's definitely happening

39
00:02:44,582 --> 00:02:47,724
in other stock exchanges around the world. All the kinds of

40
00:02:47,745 --> 00:02:51,267
things I was worried about are happening, and we don't have protections

41
00:02:51,307 --> 00:02:54,450
against them. All the issues of bias. We now know from

42
00:02:54,490 --> 00:02:57,732
an article that just came out that there's covert bias, where if you talk in

43
00:02:58,213 --> 00:03:01,555
African-American English, you get a different response than if you speak

44
00:03:01,575 --> 00:03:04,808
in standard American English. There's so many problems. And we

45
00:03:04,828 --> 00:03:08,172
have no legislation really to address these

46
00:03:08,232 --> 00:03:11,596
problems yet. There are a few bills. There's no kind

47
00:03:11,616 --> 00:03:14,880
of comprehensive approach. And the problems are

48
00:03:14,900 --> 00:03:18,244
just getting worse and worse. We've seen this movie

49
00:03:21,498 --> 00:03:25,623
You mentioned one of those approaches, what's going on in California with SB

50
00:03:27,545 --> 00:03:32,430
1047. Now, that has kind of, to me, been this perfect example

51
00:03:32,690 --> 00:03:36,355
of why regulation is going to continue to be so challenging because

52
00:03:36,395 --> 00:03:39,964
of the reaction to it. the whole lobbying movement against

53
00:03:40,064 --> 00:03:43,746
it, the millions of dollars that are fighting the idea of this. And, you

54
00:03:43,766 --> 00:03:47,188
know, it's this idea that regulation would

55
00:03:50,570 --> 00:03:54,432
And I'm wondering... Yeah, the talking point is a lie. Well, exactly. Like,

56
00:03:55,153 --> 00:03:58,715
sometimes regulation actually fosters

57
00:03:58,755 --> 00:04:01,977
innovation. You know, that's how you get seatbelts, for example, is you have a law

58
00:04:02,017 --> 00:04:05,639
that says you need We need to protect people. We wouldn't have whole industries like

59
00:04:05,960 --> 00:04:09,623
commercial airlines that people were willing to take if we didn't have regulation.

60
00:04:09,703 --> 00:04:13,346
So there's a certain part of Silicon Valley

61
00:04:13,366 --> 00:04:17,029
that wants you to believe that lie, that regulation and innovation

62
00:04:17,089 --> 00:04:21,412
are in diametric opposition. But it's not true. I mean,

63
00:04:21,552 --> 00:04:25,148
we have innovation. In the car industry, this regulation doesn't

64
00:04:25,168 --> 00:04:28,890
stop it. We have regulation in

65
00:04:29,050 --> 00:04:32,332
drugs. That doesn't mean that everybody goes home. It's not like because we

66
00:04:32,372 --> 00:04:35,915
have an FDA, all big pharma just drops out and gives up. It's

67
00:04:36,055 --> 00:04:39,377
ludicrous. And at the same time, we don't want them to put any old

68
00:04:39,457 --> 00:04:44,020
thing out there and have millions

69
00:04:44,060 --> 00:04:47,322
of people take some drug that causes a lot of problems, which we have

70
00:04:47,362 --> 00:04:51,354
seen occasionally. That's why we have a regulatory framework, so we can have innovation,

71
00:04:53,375 --> 00:04:56,676
Right. It's the right kind of innovation, the innovation that

72
00:04:57,376 --> 00:05:00,517
And some of that innovation needs to be towards safety. I mean, a

73
00:05:00,637 --> 00:05:04,138
dirty secret of AI, and maybe it's not even that secret, is nobody knows

74
00:05:04,178 --> 00:05:07,299
how to keep it safe. And nobody actually knows how to rein it in. So

75
00:05:07,319 --> 00:05:10,460
we need innovation on the safety side. And you

76
00:05:10,480 --> 00:05:13,561
have people like Yann LeCun that say, oh, there's no problem. But

77
00:05:13,781 --> 00:05:17,162
even he, when he's pushed to be a little bit more honest, says, well,

78
00:05:17,182 --> 00:05:20,703
there's no problem we can't fix. But we don't have anything encouraging people

79
00:05:22,004 --> 00:05:25,547
You say in the book that the potential upside of AI is enormous,

80
00:05:26,007 --> 00:05:29,209
just as big as the valley says it is. I'm

81
00:05:29,249 --> 00:05:32,531
wondering to you what that upside looks like and what kind

82
00:05:32,551 --> 00:05:35,914
of AI is needed to kind of access that upside. Are

83
00:05:35,934 --> 00:05:39,176
we talking about a general intelligence here for that or is

84
00:05:43,163 --> 00:05:46,985
Not even always machine learning. I mean, look, a calculator is

85
00:05:47,025 --> 00:05:50,807
something that helps innovation in science. Always

86
00:05:50,827 --> 00:05:54,329
has, always will. Just, you know, basic computation, you

87
00:05:54,349 --> 00:05:58,211
know, going back to the 1940s has always helped with science. Not

88
00:05:58,391 --> 00:06:01,493
everything that helps science that comes out of a

89
00:06:01,533 --> 00:06:04,714
computer is going to be AI at all. Not all of it's going to be

90
00:06:04,754 --> 00:06:08,196
machine learning. And, you know, there's every reason

91
00:06:08,236 --> 00:06:11,878
to think that various forms of AI will continue to contribute

92
00:06:11,898 --> 00:06:15,199
to science. Probably the most exciting game in town right now

93
00:06:15,279 --> 00:06:19,041
is AlphaFold or AlphaFold3 from DeepMind, which

94
00:06:19,081 --> 00:06:22,562
has been tremendously helpful. It's changed how people do protein science.

95
00:06:23,243 --> 00:06:27,184
Everybody in the field uses it or at least knows what it is. That's

96
00:06:27,264 --> 00:06:30,690
not a large language model. more complicated kind

97
00:06:30,710 --> 00:06:34,951
of hybrid system with some careful domain engineering. My

98
00:06:35,191 --> 00:06:38,572
concern mostly right now is with so-called large language models and

99
00:06:38,592 --> 00:06:41,832
chatbots, which have gotten all the attention lately and all the

100
00:06:41,892 --> 00:06:44,993
money or most of the money. Tens of billions of

101
00:06:45,033 --> 00:06:48,734
dollars have gone into them. And they're not actually, I think, a particularly good

102
00:06:48,774 --> 00:06:52,114
way of solving any kind of mission-critical problem. They're fun

103
00:06:52,154 --> 00:06:55,455
to play with, but we're putting way too much money into them

104
00:06:55,932 --> 00:06:59,534
relative to other scientific approaches to AI and other aspects

105
00:06:59,574 --> 00:07:03,156
of science. I mean, people, it's crazy, like, if

106
00:07:03,176 --> 00:07:06,518
an alien dropped in, they would say, these guys were doing pretty

107
00:07:06,558 --> 00:07:09,739
great with science on planet Earth for a while. I mean, they were very slow to

108
00:07:09,779 --> 00:07:12,941
get started, but, you know, eventually they kind of figured out the

109
00:07:12,981 --> 00:07:17,203
scientific method and they're doing great. And then in the, you

110
00:07:17,223 --> 00:07:25,135
know, 2120s, or sorry, 20, excuse me, in the 2020s, They

111
00:07:25,195 --> 00:07:28,878
started doing this really bizarre thing where you had seven companies,

112
00:07:29,319 --> 00:07:32,761
or something like that, all working on exactly the same thing, the

113
00:07:32,821 --> 00:07:36,164
same kind of architecture, with the same kind of data, having the same

114
00:07:36,224 --> 00:07:39,307
kind of problems, all just hoping for a miracle. And

115
00:07:39,327 --> 00:07:43,110
of course that miracle didn't happen. And, you know, thank God in

116
00:07:44,070 --> 00:07:47,213
2026 or whatever, they finally decided to do something different because that was

117
00:07:47,353 --> 00:07:50,976
crazy. And these models were starting to be, you know, a billion dollars

118
00:07:51,036 --> 00:07:54,899
a shot. You know, companies were spending five, 10, 15 billion

119
00:07:54,939 --> 00:07:58,402
dollars on chips and they were all getting the same results. Like you

120
00:07:58,422 --> 00:08:01,605
don't want to run the same experiment seven times. Like, you know,

121
00:08:02,205 --> 00:08:05,768
good scientists like seize the trend. But here, oh, no. Like

122
00:08:06,189 --> 00:08:09,712
the people, the desire for money has has overwhelmed scientific

123
00:08:11,093 --> 00:08:14,295
Right. And so part of that, and you talk about this

124
00:08:14,315 --> 00:08:17,697
a lot in the book as well, which is your point that money and power have

125
00:08:17,737 --> 00:08:20,959
derailed AI from its original mission. And

126
00:08:21,379 --> 00:08:24,841
part of that, too, is this idea of the increasing influence of the unelected tech

127
00:08:24,881 --> 00:08:28,264
leaders on all of our lives where, you know, Sam Altman

128
00:08:28,304 --> 00:08:31,926
can choose to put out whatever model, regardless of whatever harms

129
00:08:31,946 --> 00:08:35,127
it. it may be capable of, and we all

130
00:08:35,167 --> 00:08:38,569
just have to kind of deal with that. I'm wondering about that original mission.

131
00:08:39,069 --> 00:08:42,530
And as you've kind of seen, it seems like AI in

132
00:08:42,570 --> 00:08:45,972
that pure form of trying to help people has maybe lost its

133
00:08:49,293 --> 00:08:52,554
I mean, it hasn't totally lost its way. Like, DeepMind is still, I

134
00:08:52,574 --> 00:08:56,216
think, really trying to do some AI for good. But

135
00:08:56,236 --> 00:08:59,457
it's largely lost its way. And I mean, it's just a

136
00:08:59,497 --> 00:09:02,828
story of money and corruption and A story, I

137
00:09:03,048 --> 00:09:06,569
think, of how AI went from good technical

138
00:09:06,629 --> 00:09:09,750
leaders who really understood the science to people that are better at

139
00:09:09,810 --> 00:09:13,251
marketing and hype and so forth. And they have been rewarded for

140
00:09:13,291 --> 00:09:16,532
that marketing and hype. They've been rewarded by investors. They've

141
00:09:16,552 --> 00:09:19,753
been rewarded by politicians who don't know any better. Now

142
00:09:19,793 --> 00:09:23,254
they're being rewarded by Oprah. It's a

143
00:09:23,294 --> 00:09:26,636
true story. the wrong things are

144
00:09:26,676 --> 00:09:29,997
being rewarded. What's being rewarded right

145
00:09:30,037 --> 00:09:33,118
now is people making crazy claims like that AI is

146
00:09:33,138 --> 00:09:36,600
going to solve physics, which doesn't even make sense. It's not even a coherent claim.

147
00:09:37,140 --> 00:09:40,661
But there's a certain group of people that wants

148
00:09:40,701 --> 00:09:44,383
to feel like we're having this religious moment, we're being in this magical

149
00:09:44,403 --> 00:09:48,485
thing. And so much too much power is being given to

150
00:09:49,765 --> 00:09:53,087
a few people that were in the right place at the right time that can sell

151
00:09:53,167 --> 00:09:57,064
those stories. And the field

152
00:09:57,124 --> 00:10:00,767
of AI, I think, has suffered. It used

153
00:10:00,787 --> 00:10:04,349
to be more of a research endeavor, but now it's a commercial endeavor.

154
00:10:04,549 --> 00:10:07,811
And it also happens that the kind of AI that

155
00:10:07,851 --> 00:10:11,274
people are building right now is very expensive to operate. And that has changed things.

156
00:10:11,334 --> 00:10:14,716
So even you can see within OpenAI's history,

157
00:10:14,756 --> 00:10:17,978
for example, I never loved them, but they were a more

158
00:10:18,018 --> 00:10:21,730
reasonable company early on. And then they realized

159
00:10:21,770 --> 00:10:24,971
that they could make progress, not that it was the only way to make progress, but that they

160
00:10:25,011 --> 00:10:28,572
could make progress by making these very big models and that they were very expensive.

161
00:10:31,952 --> 00:10:35,233
Yeah. And that also, uh, that endeavor kind of pushed

162
00:10:35,833 --> 00:10:39,174
the idea of large language models into everyone's lives in a way of,

163
00:10:40,154 --> 00:10:43,355
you know, dealing with issues of copyright, right. And

164
00:10:46,064 --> 00:10:49,907
The copyright stuff is out of control. I just posted a video

165
00:10:49,927 --> 00:10:53,189
yesterday. Somebody said, wow, look at this amazing new

166
00:10:53,269 --> 00:10:57,371
thing. It's kind of like Dolly, but with videos out

167
00:10:57,391 --> 00:11:00,854
of China. And it showed a lightsaber battle. And

168
00:11:01,154 --> 00:11:04,956
the thing is that these people have obviously not paid for licensing

169
00:11:05,016 --> 00:11:08,458
of Lucasfilm, or I guess maybe Disney, or whoever owns the

170
00:11:08,478 --> 00:11:11,881
properties. But they've trained on it. And so it looks

171
00:11:12,141 --> 00:11:15,436
very much like Star Wars characters. That is Darth Vader.

172
00:11:15,476 --> 00:11:19,739
That is a lightsaber. And it's total copyright violation. Probably they

173
00:11:19,779 --> 00:11:23,061
will eventually get sued. All of this stuff is actually built

174
00:11:24,062 --> 00:11:27,404
on essentially all the data that companies can

175
00:11:27,444 --> 00:11:30,667
scrounge, and that includes copyrighted data. And so you

176
00:11:30,687 --> 00:11:34,229
have these companies that talk about, you know, they want to do AI

177
00:11:34,269 --> 00:11:37,752
for the benefit of humanity, and then they sit here and

178
00:11:37,952 --> 00:11:41,354
steal the property of artists and writers left and right with

179
00:11:41,414 --> 00:11:44,604
no compensation. go around and they say, yeah, we want everybody to

180
00:11:45,085 --> 00:11:48,734
have a universal basic income. But they don't pay licensing rights

181
00:11:50,816 --> 00:11:54,018
Right, and part of that, right, is this idea that, and

182
00:11:54,218 --> 00:11:57,300
you say this in the book, that the view in Silicon Valley is

183
00:11:57,340 --> 00:12:00,702
that anything out there is mine for the taking, right?

184
00:12:01,942 --> 00:12:05,144
And it's kind of... Yeah, and a new line which I don't

185
00:12:05,164 --> 00:12:08,566
have in my book is, let all the lawyers clean it up, which is something Eric

186
00:12:08,586 --> 00:12:11,808
Schmidt said in a private meeting at Stanford. I wish

187
00:12:11,828 --> 00:12:15,430
they could add that in the book because I think it really captures

188
00:12:15,450 --> 00:12:18,892
this whole Silicon Valley ethos right now

189
00:12:19,132 --> 00:12:22,391
of Just move fast, break things, have the lawyers clean it

190
00:12:22,451 --> 00:12:25,994
up. There's no attention to ethics anymore. One

191
00:12:26,014 --> 00:12:29,778
of the things I talk about in the book is how Google used to say, don't be evil. That's

192
00:12:29,818 --> 00:12:33,522
not even an official part of the motto anymore.

193
00:12:33,762 --> 00:12:37,186
And things have really shifted in the last couple of years. I

194
00:12:37,226 --> 00:12:40,629
think the general public doesn't understand how much it's shifted.

195
00:12:40,979 --> 00:12:44,201
So you always had some companies that I would say were marginal on

196
00:12:44,221 --> 00:12:47,482
the ethical side. Meta has never been on top

197
00:12:47,522 --> 00:12:50,624
of my list of ethical companies. But Google, I

198
00:12:50,664 --> 00:12:53,986
think, was more ethical than it has been of

199
00:12:54,026 --> 00:12:57,067
late. Same with Microsoft. They used to talk a

200
00:12:57,108 --> 00:13:00,609
lot about responsible AI. And now everybody's just trying to get market share,

201
00:13:04,175 --> 00:13:07,780
Right. And that necessary shift that we're talking about towards

202
00:13:07,841 --> 00:13:11,005
more ethically minded behavior and the idea of

203
00:13:11,065 --> 00:13:14,671
responsible AI, is regulation the

204
00:13:14,711 --> 00:13:18,076
only way to get that? Because we're talking about such a big shift in culture.

205
00:13:19,514 --> 00:13:22,697
Yeah, exactly. Shifting culture. So part of the reason I wrote the

206
00:13:22,717 --> 00:13:26,440
book is I would like the public to hold these companies accountable. If

207
00:13:26,480 --> 00:13:30,023
the public is like, yeah, I understand that this cool

208
00:13:30,083 --> 00:13:33,145
new software I'm using is pillaging artists, but I

209
00:13:33,165 --> 00:13:36,408
don't really care. It's fun to use. Well, I mean, that

210
00:13:36,468 --> 00:13:39,750
sets one expectation. Another expectation we

211
00:13:39,790 --> 00:13:42,953
could say is, I don't want you ripping off artists, because artists are my

212
00:13:42,993 --> 00:13:46,076
friends. And I don't want you ripping off writers, because writers are my friends. And I

213
00:13:46,096 --> 00:13:49,398
don't want you ripping off anybody, because eventually you're going to come from me. And

214
00:13:49,438 --> 00:13:52,739
so I'm going to take a stand right now and say, if

215
00:13:52,779 --> 00:13:56,460
you're not going to ethically source the stuff that you're training on,

216
00:13:56,860 --> 00:14:00,601
I'm not going to be part of this. Come back to me when you've got your licensing squared

217
00:14:00,641 --> 00:14:04,242
away. And you know what? The companies would. It's not that all of these magical products,

218
00:14:04,642 --> 00:14:08,056
and in some ways they are magical, would disappear. be

219
00:14:08,076 --> 00:14:11,599
a little bit more expensive. But look what happened with Apple Music

220
00:14:11,639 --> 00:14:14,822
and Napster and so forth, right? And there was

221
00:14:14,862 --> 00:14:20,367
a brief period in the beginning of Napster around, I guess, around 2000, 2001, where

222
00:14:20,547 --> 00:14:24,350
everybody was getting music for free. And it was really fun for them. And the artists and musicians

223
00:14:24,730 --> 00:14:27,993
were getting totally screwed. And the court said the right thing, which is you need to

224
00:14:28,033 --> 00:14:31,256
license. And you know what? We moved to a model where you

225
00:14:31,276 --> 00:14:35,219
could have what you had in Napster, which was access to everything, almost

226
00:14:35,319 --> 00:14:38,481
everything, but you paid a little bit each month so you pay

227
00:14:38,521 --> 00:14:41,881
ten dollars a month or whatever for apple music and now you can listen

228
00:14:41,921 --> 00:14:45,082
to almost anything and the artists get something they don't get as much as

229
00:14:45,122 --> 00:14:49,043
before and i think that's unfortunate but at least they get something and

230
00:14:50,023 --> 00:14:53,784
we can do that here we can have licensing

231
00:14:53,824 --> 00:14:57,405
there's no reason that that's impossible the companies even

232
00:14:57,465 --> 00:15:00,926
know it you know behind the scenes they are negotiating licensing deals

233
00:15:01,566 --> 00:15:04,773
but their official line is like, you

234
00:15:04,793 --> 00:15:08,075
have to give us this free so we can make the magical AI. We should

235
00:15:08,095 --> 00:15:11,336
not believe that. And the consumers should not believe that

236
00:15:11,416 --> 00:15:14,818
either. The consumers should say, hey, wait a minute, I want to be ethical

237
00:15:14,858 --> 00:15:18,140
here. I want to support artists. I want to support writers. I'm not

238
00:15:18,160 --> 00:15:21,402
going to use this stuff until they sort that out. That would have

239
00:15:23,251 --> 00:15:26,514
Right. And part of the reason that this feels to

240
00:15:26,554 --> 00:15:29,716
me like such a important moment is what you're talking about has

241
00:15:30,017 --> 00:15:33,600
been built up for 20 years. You're talking about the Internet companies

242
00:15:33,640 --> 00:15:36,982
that evolved into the social media companies and now they're the A.I. companies. Right.

243
00:15:37,183 --> 00:15:40,265
And people didn't read the terms of service. They don't want to. A

244
00:15:40,826 --> 00:15:44,429
lot of people don't care how their data is used and how other people's data

245
00:15:44,509 --> 00:15:48,332
is used because data has become the kind of cost of admission to

246
00:15:51,634 --> 00:15:54,876
I think we're like boiling frogs. People have come to accept way too

247
00:15:54,916 --> 00:15:58,037
much. OpenAI just bought a webcam company. They want to be able to

248
00:15:58,678 --> 00:16:02,279
have access to all your files, to all your calendars. And

249
00:16:02,299 --> 00:16:05,661
they just put someone from the NSA on there, or used to be at the NSA, on

250
00:16:05,681 --> 00:16:08,942
their board. I mean, OpenAI is,

251
00:16:09,163 --> 00:16:12,424
in my view, possibly headed towards becoming a surveillance company that

252
00:16:12,444 --> 00:16:15,846
knows everything about you, and then, you

253
00:16:15,886 --> 00:16:19,482
know, monetizes your data. And you get nothing. You

254
00:16:19,502 --> 00:16:23,385
get some free services and that's all you get out of it. And

255
00:16:23,405 --> 00:16:26,747
the world has, I think, been too quick to accept those

256
00:16:26,787 --> 00:16:30,549
kinds of things. And the consequence is that they're

257
00:16:30,569 --> 00:16:34,231
going to, for example, be able to target political advertising to you and manipulate you.

258
00:16:34,972 --> 00:16:38,434
There's new research showing, for example,

259
00:16:38,474 --> 00:16:42,216
that large language models can implant false

260
00:16:42,296 --> 00:16:45,756
memories in people. Like, we have no control, no checks

261
00:16:45,816 --> 00:16:48,997
and balances over what the large language models are

262
00:16:49,017 --> 00:16:52,358
going to teach you, essentially. And they will do it subtly, you

263
00:16:52,398 --> 00:16:56,079
won't notice it, and you can, for example, wind up with false memories. This

264
00:16:56,119 --> 00:16:59,780
is terrifying stuff. Everything that George

265
00:16:59,840 --> 00:17:03,804
Orwell imagined in 1984 is kind of actually happening, and

266
00:17:03,844 --> 00:17:07,006
we're not being loud enough about it. The reason I wrote the book is

267
00:17:07,026 --> 00:17:10,389
to get the public to realize we need to say this is not acceptable. And

268
00:17:10,429 --> 00:17:13,772
yet here we are, it's two months from the election, nobody's even talking about

269
00:17:13,792 --> 00:17:16,875
AI policy. And the next administration is

270
00:17:16,895 --> 00:17:20,077
going to set AI policy in a way that's going to last. You know, just

271
00:17:20,117 --> 00:17:23,359
like once Section 230 was set in

272
00:17:23,419 --> 00:17:26,641
place for social media, the rest of the movie was written at

273
00:17:26,681 --> 00:17:29,783
that moment. So, like, we have to get this right, and we have to get

274
00:17:31,008 --> 00:17:34,230
Speaking about the public's, the importance of the

275
00:17:34,270 --> 00:17:37,731
public in this space and understanding the space, there's

276
00:17:37,751 --> 00:17:41,033
been, and we talked about it, so much hype, so much

277
00:17:41,133 --> 00:17:44,335
marketing, that's not really accurate to the science, right?

278
00:17:44,495 --> 00:17:48,276
What are the most common, I guess, or most frustrating misconceptions about

279
00:17:51,238 --> 00:17:54,639
Maybe the most frustrating is people just think that chatbots are a lot smarter than

280
00:17:54,659 --> 00:17:57,901
they actually are. So they assume, for example, that

281
00:17:58,181 --> 00:18:01,762
chatbots know what they're talking about. And kind of imagine

282
00:18:01,782 --> 00:18:05,004
that they would like fact-check their work or something, but they don't, right?

283
00:18:05,024 --> 00:18:08,446
They hallucinate constantly. In the book I have the example of

284
00:18:08,947 --> 00:18:12,269
one of the systems just making up the fact, making

285
00:18:12,309 --> 00:18:15,391
up the claim, that I have a pet chicken named Henrietta. And I

286
00:18:15,451 --> 00:18:18,653
don't have a pet chicken named Henrietta. It would take two seconds to look

287
00:18:18,713 --> 00:18:22,175
up and do a web search and see if there's any valid source,

288
00:18:22,195 --> 00:18:25,273
because it's such a crazy thing. I mean, Like, I

289
00:18:25,333 --> 00:18:28,555
live in a city. Where am I going to put the pet chicken? Like, it just made no sense

290
00:18:28,575 --> 00:18:31,737
whatsoever. And who names their pet chicken Henrietta? And so

291
00:18:31,757 --> 00:18:34,960
forth. The systems don't actually fact check. But people are

292
00:18:35,000 --> 00:18:38,402
surprised. Like, every day somebody posts on Twitter, I can't believe how dumb this machine

293
00:18:38,442 --> 00:18:41,804
is. Like, somehow we are not communicating enough to the general public

294
00:18:42,084 --> 00:18:45,506
how dumb the machines actually are. And people get taken. Big

295
00:18:45,546 --> 00:18:48,748
businesses all, you know, last year invested massively in

296
00:18:48,788 --> 00:18:52,771
large language models, and then all realized it was disappointing. Over

297
00:18:52,791 --> 00:18:56,088
and over again, people are learning that lesson. It's

298
00:18:56,128 --> 00:18:59,511
been hard for the public to accept. And that's partly because most people have

299
00:18:59,611 --> 00:19:02,934
no training. Like, how do you recognize what

300
00:19:02,975 --> 00:19:06,858
a machine does, or how good it is, or that it's a machine at all? We're

301
00:19:06,878 --> 00:19:10,261
very good at recognizing lions and tigers, because in the

302
00:19:10,301 --> 00:19:13,704
environment of adaptation, we had to run from lions and tigers. We're not

303
00:19:13,784 --> 00:19:16,967
very good at realizing we're being scammed by a bunch of

304
00:19:17,027 --> 00:19:20,491
chatbots that don't know what we're talking about. Most people don't understand how

305
00:19:20,511 --> 00:19:23,652
to resist that. I mean, partly the book is

306
00:19:26,433 --> 00:19:29,733
Right. I think the most interesting thing

307
00:19:29,933 --> 00:19:33,094
to me is kind of watching... The reason chatbots seem to

308
00:19:33,154 --> 00:19:36,415
have grabbed people on and scammed people in the way you're talking about is

309
00:19:36,475 --> 00:19:40,015
they seem to communicate in language. And we're just primed to

310
00:19:40,055 --> 00:19:44,336
assume that if you can talk, you're a you. You're intelligent. And

311
00:19:46,236 --> 00:19:49,357
And so... And then they do little things like they type words out one at

312
00:19:49,397 --> 00:19:53,157
a time. The answer is actually pre-computed probably,

313
00:19:53,177 --> 00:19:56,540
or could be pre-computed, but they do that to make it look more human.

314
00:19:56,720 --> 00:20:00,744
Developers are trying to trick you. The people

315
00:20:00,784 --> 00:20:04,387
who most want to be tricked are most seduced. The more cynical

316
00:20:04,447 --> 00:20:08,351
among us might reach back at it, but

317
00:20:08,871 --> 00:20:12,154
some people want to be seduced by the machines. They want to fall in love with

318
00:20:14,703 --> 00:20:18,724
Yeah, the whole AI girlfriend apps is a whole thing,

319
00:20:18,784 --> 00:20:22,405
but in terms of the limits of the current architecture,

320
00:20:22,645 --> 00:20:26,067
large language models, the backing of generative AI,

321
00:20:26,227 --> 00:20:29,448
you've talked a lot about how deep learning is hitting a wall, which we're

322
00:20:29,828 --> 00:20:33,149
seeing a decent amount of evidence for, and

323
00:20:33,169 --> 00:20:36,710
the idea that... Yeah, probably a better phrase would be reaching a point of diminishing

324
00:20:37,710 --> 00:20:41,251
We've shifted from a regime where there was spectacular progress

325
00:20:41,451 --> 00:20:45,593
over a period of about three years, like 2020 to

326
00:20:46,174 --> 00:20:49,476
2023. We shifted from that to a much more incremental thing. So,

327
00:20:49,836 --> 00:20:53,118
you know, GPT-4 was trained over two years ago.

328
00:20:53,318 --> 00:20:56,520
Most people don't realize that. And everybody was like talking, oh,

329
00:20:56,620 --> 00:20:59,902
exponential progress, AI gets better and better every day. But

330
00:21:00,282 --> 00:21:03,924
we've only seen minor improvements relative to GPT-4. There's no, you

331
00:21:03,944 --> 00:21:07,066
know, GPT-5 worthy thing. Every time OpenAI comes out with a

332
00:21:07,106 --> 00:21:10,249
new model now, They sit there and like, well, we better not call this

333
00:21:10,289 --> 00:21:15,012
GPT-5 because it's really not that much better than GPT-4. So

334
00:21:15,032 --> 00:21:18,214
what should we call it? And they come out with these answers like GPT-40. Like, what

335
00:21:18,274 --> 00:21:21,897
is that? You know, I think a bunch of these are efforts or

336
00:21:21,917 --> 00:21:25,019
GPT-Turbo, I think was an effort probably to

337
00:21:25,059 --> 00:21:28,301
make GPT-5 and then came out and they're like, yeah, this isn't

338
00:21:28,321 --> 00:21:31,477
really that great. Like, it's a little bit better, but it's not that great. And

339
00:21:31,517 --> 00:21:34,720
so we have hit, I mean, I interrupted your question, but I would say we

340
00:21:34,760 --> 00:21:38,282
haven't absolutely hit a wall overall. There's still little improvement.

341
00:21:38,582 --> 00:21:42,465
We have hit a wall on certain things. So these systems still are

342
00:21:42,525 --> 00:21:45,968
not reliable. They still hallucinate and make up stuff like Henrietta,

343
00:21:46,668 --> 00:21:49,971
the pet chicken. Like that is a wall. Like we have not gotten

344
00:21:50,051 --> 00:21:54,715
past the wall that you don't know what you're getting out of these systems. that

345
00:21:54,775 --> 00:21:58,178
we're just stuck. And then other things like, you know, the graphics get

346
00:21:58,238 --> 00:22:01,541
better. So now we can make movies instead of just still shots.

347
00:22:01,561 --> 00:22:05,424
So there's some progress. But on the core reliability

348
00:22:05,484 --> 00:22:08,827
and reasoning and so forth, it's at best diminishing returns now.

349
00:22:09,082 --> 00:22:12,363
At the same time as we're kind of noticing that, there is so much,

350
00:22:12,943 --> 00:22:16,184
or seems like such a large number of people, right? And

351
00:22:16,364 --> 00:22:19,665
I guess if you're on Twitter like I am, maybe the numbers are

352
00:22:19,705 --> 00:22:22,866
not quite accurate, that think, as you

353
00:22:22,886 --> 00:22:26,287
mentioned, that it's moving exponentially. Every new release

354
00:22:29,538 --> 00:22:32,759
You know, AI influence say that stuff on Twitter and

355
00:22:32,779 --> 00:22:36,340
they got like a million views for the things. Most of them would not know their

356
00:22:36,440 --> 00:22:40,122
ass from an exponential. I mean, an exponential is actually a mathematical thing

357
00:22:40,182 --> 00:22:43,663
where you plot a curve, you fit it and you see like that's

358
00:22:43,703 --> 00:22:46,864
what it would actually mean to be exponential. They don't know

359
00:22:46,884 --> 00:22:50,245
how to do that. Like they're talking about AI and it's like basic, you

360
00:22:50,285 --> 00:22:53,446
know, like grade school math thing or high school math thing that

361
00:22:53,466 --> 00:22:56,747
they just don't understand. Exponential does not mean

362
00:22:57,294 --> 00:23:00,499
that every two months it's a little bit better than last

363
00:23:00,539 --> 00:23:04,044
time. Exponential would be like what we saw before,

364
00:23:04,124 --> 00:23:07,770
where you have these giant qualitative changes every few

365
00:23:09,157 --> 00:23:12,378
Now something that's kind of gone in hand with

366
00:23:12,438 --> 00:23:16,020
all the hype that we're seeing is the investment side

367
00:23:16,040 --> 00:23:19,221
of things and the investment hype and the crazy valuations in

368
00:23:19,241 --> 00:23:23,442
the markets. And as investors have started to realize that

369
00:23:23,482 --> 00:23:27,344
there are diminishing returns, maybe it's not worth all this money, it's

370
00:23:27,404 --> 00:23:31,145
led to the idea of the AI bubble. And you've called OpenAI

371
00:23:31,725 --> 00:23:35,147
possibly the WeWork of AI, that the bubble

372
00:23:38,958 --> 00:23:42,220
I was one of the first people to say there's a bubble. And by that, I

373
00:23:42,260 --> 00:23:45,582
mean a bubble in the valuations. I don't think large language models

374
00:23:45,622 --> 00:23:48,923
will disappear. They're like a tool. We have a lot of different tools for AI.

375
00:23:49,244 --> 00:23:52,866
We need a lot more that we haven't discovered. An analogy would

376
00:23:52,886 --> 00:23:56,708
be like the minute that somebody invented a power screwdriver.

377
00:23:57,008 --> 00:24:00,170
They thought power screwdrivers were going to change the entire world and that

378
00:24:00,190 --> 00:24:03,692
they'd be worth billions. And no, eventually the price would come back down

379
00:24:03,752 --> 00:24:07,114
and, you know, people would pay 30 bucks for them. They'd still be great. People would still love them.

380
00:24:07,174 --> 00:24:10,416
But, you know, they'd be a $30 tool. They wouldn't be like, you

381
00:24:10,436 --> 00:24:13,578
know, constructing your entire home all by itself. And

382
00:24:13,598 --> 00:24:17,080
so that's what's going to happen here. Large language models are not going anywhere.

383
00:24:17,100 --> 00:24:20,262
They have some use. They're not super reliable, but they're good for

384
00:24:20,302 --> 00:24:23,604
brainstorming and autocomplete when you type computer code

385
00:24:23,624 --> 00:24:27,043
and stuff like that. That'll stay. But the idea that open

386
00:24:27,103 --> 00:24:30,724
AI is worth $100 billion is really marginal.

387
00:24:30,824 --> 00:24:33,986
So open AI, right

388
00:24:34,006 --> 00:24:37,307
now it's valued at something like $80 billion. Someone's about to maybe

389
00:24:37,347 --> 00:24:41,209
put in money at a $100 billion valuation. Usually

390
00:24:41,249 --> 00:24:44,711
valuations are relative to profits. They actually have none.

391
00:24:45,891 --> 00:24:50,071
They've not made a profit yet. This is very expensive to operate. These

392
00:24:50,111 --> 00:24:53,433
tools unlike traditional software. You need these expensive chips. You

393
00:24:53,473 --> 00:24:56,714
need very expensive staff You need enormous amounts of

394
00:24:56,774 --> 00:25:00,115
electricity. There's no clear evidence that

395
00:25:00,135 --> 00:25:04,377
they're ever going to make a profit You know if you have to pay 20 billion

396
00:25:04,397 --> 00:25:07,719
dollars a year in order to stay ahead of meta

397
00:25:07,739 --> 00:25:11,080
that's giving a steam stuff away for free Which is also a huge

398
00:25:11,140 --> 00:25:14,642
problem for them and you're making two billion dollars a year. The math just

399
00:25:14,742 --> 00:25:17,830
does not add up And so what I would imagine is at

400
00:25:17,870 --> 00:25:21,451
some point, people are going to stop putting the money in. Right now, OpenAI

401
00:25:21,551 --> 00:25:25,152
is a name brand, and maybe they can get

402
00:25:25,232 --> 00:25:28,833
away with it for a little longer. But they have not transformed the

403
00:25:28,853 --> 00:25:32,014
world in 2024 like they did in 2023. They have

404
00:25:32,114 --> 00:25:35,895
demoed a lot of products that they have not actually delivered. Or

405
00:25:35,935 --> 00:25:40,876
they've delivered products like GPT-4.0. Some

406
00:25:40,916 --> 00:25:44,336
of the things they promised are there and some not. Sal

407
00:25:44,356 --> 00:25:47,717
Khan talking about how you can use it as a tutor. There's no actual tutor

408
00:25:47,778 --> 00:25:50,979
software out there. They talked about Sora, but you can't use it

409
00:25:51,019 --> 00:25:54,940
at home. We don't know how much it would cost. Um, you

410
00:25:54,960 --> 00:25:58,062
know, so they're making demos and now they're going to make an

411
00:25:58,102 --> 00:26:01,303
announcement. Maybe it's going to be called GPT next. They make demos cause

412
00:26:01,323 --> 00:26:05,212
they have learned that making demos drives up the price. But

413
00:26:05,452 --> 00:26:08,835
it doesn't forever, right? So what's gonna, it's a little bit like a Ponzi scheme.

414
00:26:08,895 --> 00:26:11,998
At some point, somebody's gonna say, where is

415
00:26:12,038 --> 00:26:15,141
the money? And in fact, you know, I wrote an article about this a

416
00:26:15,201 --> 00:26:18,684
year ago called, is generative AI

417
00:26:18,744 --> 00:26:22,046
a dud? Or what if generative AI turned out to be a dud or something

418
00:26:22,066 --> 00:26:25,169
like that in my sub stack? So I wrote this article at the time, I

419
00:26:25,649 --> 00:26:29,032
was somewhat alone. There were a few other people maybe saying the same thing. Now, like

420
00:26:29,112 --> 00:26:33,456
every week, somebody's saying that. And the thing about bubbles

421
00:26:34,033 --> 00:26:37,735
is they're all about psychology. So what is a tulip actually worth?

422
00:26:38,335 --> 00:26:41,537
Well, a tulip was mostly worth what people will pay for it.

423
00:26:41,577 --> 00:26:45,098
And so you can, you know, for a few years, people were paying, I don't know, thousands,

424
00:26:45,458 --> 00:26:48,580
tens of thousands of dollars for tulips. And then they all woke up

425
00:26:48,620 --> 00:26:51,761
and said, you know, this is a little crazy. It's just a tulip. It's just a flower. You

426
00:26:51,781 --> 00:26:56,163
can buy another flower for five bucks. Why are you spending, you know,

427
00:26:56,584 --> 00:26:59,885
$10,000 for a tulip? It's not really worth it. And then everybody runs

428
00:26:59,945 --> 00:27:03,307
for the hills. And that's what's gonna happen here, is everybody's gonna

429
00:27:03,327 --> 00:27:06,537
run for the hills. So you could ask, well, who made money and who gets screwed? So the

430
00:27:06,597 --> 00:27:10,318
early employees still make money because they've sold some of the stock, right?

431
00:27:10,398 --> 00:27:13,759
So early employees of OpenAI do well. The

432
00:27:13,819 --> 00:27:17,440
early investors make money because they take a cut of the money that's invested.

433
00:27:18,160 --> 00:27:21,440
And then the people that invested in them get screwed. So you

434
00:27:21,460 --> 00:27:24,761
have, for example, pension funds. And the people who operate the

435
00:27:24,801 --> 00:27:28,062
pension funds make money because they've invested the pension funds

436
00:27:28,122 --> 00:27:31,823
money. But the people who are actually invested in the pension funds,

437
00:27:32,143 --> 00:27:35,517
the employees of California or whatever, are in the end going to get screwed

438
00:27:35,557 --> 00:27:38,978
because OpenAI is most likely not going to make back

439
00:27:39,018 --> 00:27:42,338
the $100 billion valuation. So like, okay, let's say they issue

440
00:27:42,378 --> 00:27:45,579
stock now at $100 billion. And then next year people are

441
00:27:45,619 --> 00:27:48,760
like, yeah, they're still not really making any money. And

442
00:27:49,000 --> 00:27:52,300
now they need more money. They need it badly. So they take money at

443
00:27:52,320 --> 00:27:55,481
a $50 billion valuation. Well, the people who put in money at

444
00:27:55,801 --> 00:27:58,902
$100 billion valuation just lost half their money. So we're going to

445
00:27:58,942 --> 00:28:02,492
see. And then, you know, what happens with these things is once One

446
00:28:02,532 --> 00:28:05,775
thing goes down, then a lot of them go down. Everybody runs

447
00:28:05,815 --> 00:28:08,977
for the hills. And so I think that's going to happen. Again, the

448
00:28:08,997 --> 00:28:12,259
technology will remain, but people are going to be like,

449
00:28:12,499 --> 00:28:16,202
yeah, maybe NVIDIA is not worth $3 trillion after

450
00:28:16,322 --> 00:28:19,684
all, because NVIDIA has a problem. They make a good product. They

451
00:28:19,704 --> 00:28:23,687
really do. I met Jensen Huang. He's a brilliant CEO.

452
00:28:24,528 --> 00:28:27,878
Everything they do is good. But it's all

453
00:28:27,978 --> 00:28:32,441
kind of premised on the notion that the people who buy their chips,

454
00:28:32,882 --> 00:28:36,044
which are very expensive, tens of thousands of dollars, are

455
00:28:36,064 --> 00:28:39,453
going to make a lot of money with those chips. And if that premise turns

456
00:28:39,553 --> 00:28:42,875
out to not be true, like there's still no killer app

457
00:28:42,895 --> 00:28:46,517
for generative AI. There's just conversation. If that conversation doesn't

458
00:28:46,557 --> 00:28:50,459
become profits, then eventually even NVIDIA winds

459
00:28:50,499 --> 00:28:54,001
up in trouble because eventually people stop placing big

460
00:28:54,142 --> 00:28:57,343
orders for chips because they can't quite figure out what to do with them. I mean, you

461
00:28:57,363 --> 00:29:00,545
know, NVIDIA is not a charity, right? I've

462
00:29:00,585 --> 00:29:03,727
made jokes about Jensen Huang's pension fund being a

463
00:29:03,767 --> 00:29:07,088
charity, but it's not, right? I mean, it's just a joke. You

464
00:29:07,108 --> 00:29:10,551
know, eventually, like open AI, if they don't get

465
00:29:10,872 --> 00:29:14,335
enough money to keep buying all of these chips, then NVIDIA suddenly

466
00:29:14,375 --> 00:29:17,578
loses a big order and then the stock goes down. And so this is

467
00:29:17,618 --> 00:29:20,740
how I think it's going to happen. I don't know if NVIDIA will be first or

468
00:29:20,800 --> 00:29:24,304
last. The psychology of all of this is complicated. But the psychology is

469
00:29:24,444 --> 00:29:27,747
it's not really sustainable at these valuations. It just isn't

470
00:29:28,127 --> 00:29:31,390
because the profits aren't there. Now, some form of AI might be worth

471
00:29:31,450 --> 00:29:34,810
trillions of dollars, which is what everybody's imagining. But generative AI

472
00:29:34,850 --> 00:29:38,031
has technical problems. I don't think it's ever going to be worth that. And so

473
00:29:38,251 --> 00:29:41,592
generative AI, you know, people are eventually going

474
00:29:42,553 --> 00:29:45,654
In terms of those other forms of AI that you were just kind of mentioning there, I

475
00:29:45,694 --> 00:29:49,635
really want to talk about the idea of new paradigms, new architectures, neurosymbolic

476
00:29:49,715 --> 00:29:53,316
AI. I know you've been talking about neurosymbolic AI for a while. What

477
00:29:53,336 --> 00:29:57,297
does that architecture look like and why is it different from

478
00:29:59,398 --> 00:30:03,314
We don't know is the first part of the answer. We

479
00:30:03,354 --> 00:30:06,616
know a general direction to look. There's every reason in the world,

480
00:30:06,656 --> 00:30:09,879
and I will explain that reason. But we haven't really done the work yet, so

481
00:30:09,899 --> 00:30:13,162
we don't really know exactly what it's gonna look like. And in

482
00:30:13,222 --> 00:30:16,465
fact, the thing I'm gonna describe to you, neurosymbolic AI, is

483
00:30:16,645 --> 00:30:20,048
part of an answer, but it's not an answer by itself. I mean, everybody is looking

484
00:30:20,148 --> 00:30:23,571
for AI to be solved magically, and it's just not gonna happen.

485
00:30:23,611 --> 00:30:26,853
Intelligence is, as Chas Firestone and

486
00:30:28,274 --> 00:30:31,955
Brian Shull once said, not one thing but many. There's many different facets

487
00:30:31,995 --> 00:30:35,056
to intelligence. And we've kind of solved one of them, which is

488
00:30:35,116 --> 00:30:38,377
kind of pattern recognition. And there are others we just haven't solved

489
00:30:38,417 --> 00:30:41,578
at all, like reasoning. We don't have a good answer to how to

490
00:30:41,598 --> 00:30:44,879
get a machine to reason. We especially don't have a good answer about how

491
00:30:44,919 --> 00:30:48,240
to get a machine to reason over open-ended problems. So in

492
00:30:48,280 --> 00:30:51,561
very simple things like logic puzzles, machines are great. We've

493
00:30:51,681 --> 00:30:54,942
been able to do that for 75 years pretty well. So

494
00:30:55,102 --> 00:30:58,681
some of the kinds of things that like You know, your 10-year-old reads in school,

495
00:30:59,561 --> 00:31:03,042
you know, A believes in B and doesn't know C, blah,

496
00:31:03,062 --> 00:31:06,263
blah, blah, blah. Those kinds of things we can actually solve pretty well. But

497
00:31:06,683 --> 00:31:11,564
logic in the real world where there's a lot of kind of incomplete information, reasoning

498
00:31:11,944 --> 00:31:15,285
in those cases, we're not that good at. We have

499
00:31:15,505 --> 00:31:18,666
some sense from this classical tradition of AI of how to do

500
00:31:18,746 --> 00:31:22,447
it, but we don't know how to do it at scale when there's a lot of knowledge involved.

501
00:31:22,868 --> 00:31:26,492
Many of your readers or listeners might know Daniel Kahneman's

502
00:31:26,532 --> 00:31:29,956
System 1 and System 2 distinction from his book Thinking Fast

503
00:31:29,996 --> 00:31:33,740
and Slow. System 1 is stuff that's automatic, fast,

504
00:31:33,840 --> 00:31:37,303
reflexive, intuitive, data-driven, and System 2 is

505
00:31:37,364 --> 00:31:41,888
more deliberative, more abstract, more reasoning-like. it's

506
00:31:41,948 --> 00:31:45,493
pretty clear that current AI is good at system one, older

507
00:31:45,533 --> 00:31:49,158
forms of AI are pretty good at system two, not perfect, and

508
00:31:49,178 --> 00:31:52,622
we don't have a good way for those to communicate with each other. That's really

509
00:31:52,642 --> 00:31:55,766
where the advances in AI have to come, I think, is how do you

510
00:31:55,826 --> 00:31:59,171
bridge these traditions? And it's not trivial, because even once you do,

511
00:31:59,211 --> 00:32:02,894
you probably have to put a lot of knowledge in machine interpretable form,

512
00:32:02,954 --> 00:32:06,137
and you have to do something called extracting a cognitive model from a

513
00:32:06,217 --> 00:32:09,339
real-world scene. Nobody knows a general way to do

514
00:32:09,379 --> 00:32:12,682
that. So, you know, some people are like, my timeline for

515
00:32:12,782 --> 00:32:16,245
AGI, meaning when I think general intelligence will come, is

516
00:32:16,265 --> 00:32:20,108
like 2027, or Elon Musk says the end of 2025. And

517
00:32:20,148 --> 00:32:23,332
it's so crazy I offered him a million dollar bet, which he didn't get

518
00:32:23,372 --> 00:32:26,677
back to me on. And it's so crazy that a friend of mine actually upped it to

519
00:32:26,737 --> 00:32:30,161
10 million dollars and still, and there's a Wall Street Journalman after Elon,

520
00:32:30,181 --> 00:32:33,706
and still his people wouldn't respond. It's a completely crazy claim

521
00:32:33,726 --> 00:32:36,830
to say that we will have artificial general intelligence by the end

522
00:32:36,870 --> 00:32:41,233
of 2025. The more you understand the artificial intelligence

523
00:32:41,293 --> 00:32:44,394
at all, the clearer that is. My training was in

524
00:32:44,434 --> 00:32:47,836
how human intelligence develops. And once you look at that, like

525
00:32:47,916 --> 00:32:51,298
what a child learns in the first few years of life, it just becomes

526
00:32:51,478 --> 00:32:55,340
obvious that we're not there yet. We have something that sort of like vaguely approximates

527
00:32:55,380 --> 00:32:58,822
it some of the time and makes complete absurd errors that no human

528
00:32:58,842 --> 00:33:02,183
would ever make a lot of the time. So yes, we need to bring these

529
00:33:02,223 --> 00:33:05,746
two traditions together. That's what I would call neuro-symbolic AI. Part

530
00:33:05,766 --> 00:33:09,170
of the book is advocacy for that, saying, for example,

531
00:33:09,210 --> 00:33:12,594
that people think that whoever has the biggest large language model

532
00:33:12,614 --> 00:33:16,359
is going to win the race to AI. I would say it's whoever sponsors

533
00:33:16,379 --> 00:33:19,623
the most innovation to find new things is going to win the race to

534
00:33:20,049 --> 00:33:23,490
Now, in terms of general intelligence, right? And Musk

535
00:33:23,510 --> 00:33:28,212
has his timelines. Musk is a hype man for everything. And,

536
00:33:28,652 --> 00:33:31,793
you know, everyone seems to have their own timeline. And some of

537
00:33:31,833 --> 00:33:35,594
that is tied to the idea of existential risk and

538
00:33:35,714 --> 00:33:39,235
P-Doom, right? And a lot of kind of pseudoscience that

539
00:33:42,897 --> 00:33:47,794
I mean, P-Doom is a real thing in a way, right? audience,

540
00:33:48,095 --> 00:33:51,439
PDoOM means the probability that machines are going to kill us all. My PDoOM is

541
00:33:51,479 --> 00:33:54,662
very low. Some people think it might be like 50% or

542
00:33:54,702 --> 00:33:57,906
even 50% in the next three years. Strangely, some

543
00:33:57,926 --> 00:34:01,190
of those run companies that might cause that if you're to believe their

544
00:34:01,210 --> 00:34:04,513
story. It's a form of hype to say, oh my God, what I'm building

545
00:34:04,573 --> 00:34:07,789
is so amazing and dangerous that might kill

546
00:34:07,829 --> 00:34:11,271
us all. My view is we're not going to annihilate the

547
00:34:11,311 --> 00:34:15,034
human species anytime soon. That that would take considerable work

548
00:34:15,074 --> 00:34:18,176
that is not likely to happen. So I

549
00:34:18,216 --> 00:34:21,377
don't think that scenario is going to happen. But

550
00:34:21,397 --> 00:34:25,720
there's all kinds of other risks associated with AI around misinformation,

551
00:34:25,800 --> 00:34:28,962
disinformation, deepfake porn, bias. I

552
00:34:29,002 --> 00:34:33,024
have a list, as you know, in the book of a dozen different risks There

553
00:34:33,044 --> 00:34:36,666
could be accidental escalation of war. There's many many things that

554
00:34:36,686 --> 00:34:39,947
could happen short of actual extinction Right and the whole

555
00:34:39,987 --> 00:34:43,208
environmental side of things as well. The the massive environmental side

556
00:34:43,228 --> 00:34:46,509
the costs are huge I mean, nobody really knows where this is going that people

557
00:34:46,549 --> 00:34:49,950
are talking about. Well, I have this big model I trained it on the internet didn't

558
00:34:50,010 --> 00:34:53,032
really work So I'll train it on ten times the Internet or a

559
00:34:53,072 --> 00:34:56,233
hundred times the Internet. Well, that's gonna cost a lot of energy I

560
00:34:56,273 --> 00:34:59,514
think one training run might be I'm just making up these numbers might be

561
00:34:59,554 --> 00:35:02,730
like a all the power that the country of Germany uses for

562
00:35:02,750 --> 00:35:06,154
a year at some point. If you keep scaling it up, you're just using insane

563
00:35:06,194 --> 00:35:10,679
amounts of power, insane amounts of water. There's lots of emissions. So

564
00:35:10,839 --> 00:35:13,943
I mean, there's some work, I would say some of the more successful work, to

565
00:35:13,983 --> 00:35:17,446
try to minimize that. But as the models get bigger and bigger, even

566
00:35:20,905 --> 00:35:24,486
Part of what you were saying about the element of human cognition

567
00:35:24,526 --> 00:35:27,686
and human intelligence, where if you study that, we are where we are.

568
00:35:30,127 --> 00:35:33,328
A lot of what I've seen- And we don't just want to replicate it, by the way. Sorry

569
00:35:33,368 --> 00:35:36,668
to interrupt your question, then you can go back. We don't want AI to be just like people.

570
00:35:36,708 --> 00:35:39,949
People have lots of problems. You don't want a calculator to do arithmetic like I

571
00:35:39,989 --> 00:35:43,990
do. I forget to carry the one. I don't pay attention. We're

572
00:35:44,010 --> 00:35:47,791
not here to replicate human intelligence, but we do want to get from

573
00:35:47,871 --> 00:35:51,007
human intelligence the resourcefulness. you know, you can have a

574
00:35:51,107 --> 00:35:54,290
conversation with me in a conversation with someone else on a different topic and

575
00:35:54,350 --> 00:35:57,513
understand them both. Or you can, you know, after we talk, you

576
00:35:57,533 --> 00:36:00,856
can build an Ikea bookshelf or whatever. Like, human beings can do many

577
00:36:00,916 --> 00:36:04,360
different things, even sometimes when the instructions are poor,

578
00:36:04,440 --> 00:36:08,043
you know, you eventually figure it out. It's that resourcefulness of

579
00:36:08,223 --> 00:36:11,386
the adaptiveness of intelligence that we would like to

580
00:36:11,426 --> 00:36:15,110
get into our machines so that we can trust them when we give them new tasks

581
00:36:16,991 --> 00:36:21,314
I don't know if I could do the IKEA bookshelf. It might take me a little while. But

582
00:36:21,614 --> 00:36:24,916
I guess the idea of a general intelligence being possible, right? There's

583
00:36:24,956 --> 00:36:28,418
a lot of debate about whether it will be possible at

584
00:36:28,938 --> 00:36:32,260
I think it's certainly possible. I mean, look, your brain is

585
00:36:32,320 --> 00:36:35,542
just some kind of machine, just like your heart is, you know, a

586
00:36:35,582 --> 00:36:38,784
mechanical pump. Your brain is an information processor. There's

587
00:36:38,804 --> 00:36:42,027
got to be some kind of Um, information processing, we haven't

588
00:36:42,107 --> 00:36:45,730
got the right one, but, you know, I see no principled reason

589
00:36:45,790 --> 00:36:48,972
why we can't build a general intelligence. I just don't think we're doing it

590
00:36:49,032 --> 00:36:52,235
right. I mean, it'd be like, you know, if somebody had

591
00:36:52,275 --> 00:36:55,317
argued in da Vinci's era, you're never going to have a helicopter. That would have been a

592
00:36:55,377 --> 00:36:59,020
wrong argument. It would have turned out you couldn't do it with the materials he

593
00:36:59,060 --> 00:37:02,753
had available. Right. We didn't know enough about internal combustion engines.

594
00:37:02,793 --> 00:37:06,056
We didn't know enough about material science and so

595
00:37:06,117 --> 00:37:09,280
forth. But it wasn't that it was impossible. It was just we we

596
00:37:09,300 --> 00:37:12,564
weren't really ready to do it. That's where we are with AI is like Da Vinci and

597
00:37:12,584 --> 00:37:15,948
helicopters. Like we have the vision now of what this would be like and

598
00:37:15,968 --> 00:37:19,226
why it would be amazing. And we have no reason to think

599
00:37:19,266 --> 00:37:22,689
it's impossible. There was no proof that you couldn't build a helicopter. There's

600
00:37:22,729 --> 00:37:26,172
no proof that you can't build a general intelligence. And I'm

601
00:37:26,332 --> 00:37:29,514
sure that, you know, within 200 years, we'll do it may well

602
00:37:29,534 --> 00:37:32,677
do it in 20 years. I'm also sure that, you

603
00:37:32,697 --> 00:37:36,360
know, we're not going to have it next year that Elon is, you know, either diluted

604
00:37:42,004 --> 00:37:45,187
He did not take it. I mean, that tells you a lot, right? I mean, We

605
00:37:45,227 --> 00:37:48,471
got the bet up to $10 million. It could have been like a symbolic thing. I

606
00:37:48,591 --> 00:37:51,854
mean, obviously, he doesn't need the money. But the fact that he wouldn't even

607
00:37:51,894 --> 00:37:55,738
respond, it doesn't bespeak

608
00:37:59,097 --> 00:38:02,300
Now, with AGI, we're dealing with a hypothetical thing. As you

609
00:38:02,320 --> 00:38:05,583
said, we don't know what it will look like when it will be here. And

610
00:38:05,683 --> 00:38:09,407
there's been a lot of concern on the regulatory front of, you

611
00:38:09,427 --> 00:38:13,431
know, there's so many active harms from AI now.

612
00:38:14,512 --> 00:38:17,955
We shouldn't lose focus on that to regulate for a

613
00:38:18,015 --> 00:38:21,878
potential AGI that we don't know if or when it's coming. How

614
00:38:21,939 --> 00:38:25,702
important is it to kind of balance

615
00:38:25,762 --> 00:38:29,586
that and prepare ourselves in legislation for

616
00:38:31,808 --> 00:38:35,372
What we need to do now is to deal with the current harms

617
00:38:35,412 --> 00:38:39,016
that we know about, most of which are not well handled under American law.

618
00:38:39,136 --> 00:38:42,519
And we need to plant a stake in the ground so that when things

619
00:38:42,959 --> 00:38:46,449
do you know, advanced, we have more intelligent systems

620
00:38:46,489 --> 00:38:50,130
that we have a framework for dealing with them, which includes like communicating with

621
00:38:50,170 --> 00:38:53,711
other countries about risks and sharing information about bad

622
00:38:53,771 --> 00:38:56,891
incidents and so forth. So, like, I think of where we are

623
00:38:56,931 --> 00:39:00,052
right now as a dress rehearsal. It's not the AI that

624
00:39:00,072 --> 00:39:03,613
we're going to have 20 years from now, any more than like the first flip phone

625
00:39:03,653 --> 00:39:06,833
is an iPhone, right? I mean, like the first flip phone was just as

626
00:39:07,273 --> 00:39:10,514
a preview of, you know, what phones could be. And

627
00:39:10,534 --> 00:39:13,875
the AI that we have now is just a preview of where AI is going.

628
00:39:13,915 --> 00:39:17,177
And we don't fully glimpse it yet. But

629
00:39:17,217 --> 00:39:20,618
we already see that there's pressing need to know

630
00:39:20,678 --> 00:39:24,440
what to do with it. And we've already seen how the

631
00:39:24,520 --> 00:39:27,761
companies have gone from publicly saying, oh, yeah, we need

632
00:39:27,801 --> 00:39:31,883
regulation. We want this to be safe for everybody, saying

633
00:39:31,923 --> 00:39:35,104
those things in public, like when Altman said exactly those

634
00:39:35,124 --> 00:39:40,648
things next to me at the US Senate around the time we met, like

635
00:39:40,908 --> 00:39:44,490
privately and sometimes even publicly opposing actual legislation coming

636
00:39:44,531 --> 00:39:47,752
up with excuses like OpenAI said about the California bill what

637
00:39:47,792 --> 00:39:51,415
we need this to be federal. Well that's true we do need federal

638
00:39:51,455 --> 00:39:54,837
law here but like we got to start somewhere and the state law

639
00:39:54,877 --> 00:39:58,078
is here it's not really going to make the federal law Harder the federal law

640
00:39:58,098 --> 00:40:01,339
can supersede the state law if we need to do that like this is

641
00:40:01,379 --> 00:40:04,620
a bogus argument It's an argument of the company that doesn't actually want

642
00:40:04,660 --> 00:40:07,881
to be regulated despite whatever they might tell the public and like

643
00:40:08,462 --> 00:40:12,123
we should be suspicious and we should start to view these companies like cigarette companies

644
00:40:12,163 --> 00:40:15,364
that you know would downplay risks and try

645
00:40:15,404 --> 00:40:18,525
to play the public with expensive lobbying campaigns and

646
00:40:18,985 --> 00:40:22,188
you know, secretive research that, you know, they would put out

647
00:40:22,228 --> 00:40:25,591
under other names and all that. Like, we should think these

648
00:40:25,651 --> 00:40:28,934
guys have a lot of money at stake and they are playing us.

649
00:40:29,315 --> 00:40:32,658
I mean, a whole section of the book is really about how they tend to do that,

650
00:40:34,105 --> 00:40:37,467
Towards the end of the book, you answer a question that

651
00:40:37,667 --> 00:40:41,429
I have asked myself many times, and you say, we

652
00:40:41,469 --> 00:40:44,590
should not stop AI. We should insist that it be

653
00:40:44,610 --> 00:40:47,932
made safe, better, and more trustworthy. At the same time,

654
00:40:47,972 --> 00:40:51,214
there's been a lot of other, I guess, discourse on

655
00:40:51,234 --> 00:40:55,016
that idea from a bunch of other people, right? There's all these categories of

656
00:40:55,116 --> 00:40:58,778
folks within AI, and the kind of Doomer argument, right?

657
00:40:58,838 --> 00:41:02,380
And the pause AI, and then those kinds of folks saying, shut it down now, oh

658
00:41:02,420 --> 00:41:05,822
my god. And I'm wondering, like, that side of the argument, and

659
00:41:06,102 --> 00:41:09,284
part of that argument is not all existential AGI. A

660
00:41:09,324 --> 00:41:12,706
lot of it is focused on the current harms that we're seeing and there's no redress for

661
00:41:12,746 --> 00:41:16,708
them. Is there anything of value in the idea of slowing

662
00:41:18,109 --> 00:41:22,292
I think I would start by making a distinction between research and deployment. So

663
00:41:23,413 --> 00:41:26,875
there is no research that I know about right now that

664
00:41:26,915 --> 00:41:30,217
seems, like, so deadly that I would say, like, just don't do it. I

665
00:41:30,257 --> 00:41:33,360
could imagine such a thing, but I don't... I don't see it

666
00:41:33,400 --> 00:41:36,502
from what I read about. On the

667
00:41:36,542 --> 00:41:40,145
other hand, are these things problematic? Should they be deployed now?

668
00:41:40,265 --> 00:41:43,587
I can see a real argument saying, no, until you get your house in

669
00:41:43,647 --> 00:41:46,790
order, don't deploy it. So this is what we do with

670
00:41:46,830 --> 00:41:50,173
medicines, right? We say, make a safety case and

671
00:41:50,373 --> 00:41:53,575
make a case that the risks outweigh, sorry, that the

672
00:41:53,635 --> 00:41:57,018
benefits outweigh the risks. And if you can't do that, we don't say,

673
00:41:57,238 --> 00:42:00,601
never make your drug, but we say, come back, tell

674
00:42:00,621 --> 00:42:03,739
me how you're gonna mitigate the risks. Um, you

675
00:42:03,779 --> 00:42:07,161
know, maybe it's a simple thing. You can just tell people, don't take it on an empty stomach

676
00:42:07,201 --> 00:42:10,483
and we're good to go. And now, you know, we have a solution here. Um,

677
00:42:11,063 --> 00:42:14,285
don't trust, you know, let's have a big public campaign and say, don't

678
00:42:14,345 --> 00:42:17,827
trust these chatbots. I mean, that's what you need for, you know, one particular thing.

679
00:42:18,047 --> 00:42:21,409
I think there are other problems that are not so easily solved. Like these systems are

680
00:42:21,549 --> 00:42:25,191
clearly discriminating and they're clearly being used in job decisions

681
00:42:25,252 --> 00:42:28,473
and that's not cool. And you could make an argument that the cost of

682
00:42:28,513 --> 00:42:31,836
society for discriminatory behavior and hiring is

683
00:42:31,916 --> 00:42:35,233
pretty high. and you could say that's a reason

684
00:42:35,293 --> 00:42:38,715
why we should slow down on deployment or pause on deployment and

685
00:42:38,735 --> 00:42:41,856
say look you're perfectly well in the ship software of

686
00:42:41,896 --> 00:42:45,098
this sort but we need a solution here figure this out you know

687
00:42:45,118 --> 00:42:48,280
it takes you six months great if it takes you two years then so be it

688
00:42:48,300 --> 00:42:51,661
like come back to us when you have a handle

689
00:42:51,721 --> 00:42:54,863
on this what about misinformation can you make a watermark that

690
00:42:54,963 --> 00:42:58,305
actually works so we'll know what's generated by a machine can

691
00:42:58,345 --> 00:43:01,546
you guarantee your stuff will be labeled you know what can you do for

692
00:43:01,626 --> 00:43:04,768
us here so that Society doesn't pick up all

693
00:43:04,808 --> 00:43:08,069
the costs. Another thing I keep thinking about is all

694
00:43:08,109 --> 00:43:11,370
these companies that used to pump toxic chemicals into

695
00:43:11,410 --> 00:43:14,651
the water and society had to pick up the cost. There's this phrase, I

696
00:43:14,671 --> 00:43:17,952
don't know who made it up, I wish I had, which is to privatize the

697
00:43:18,212 --> 00:43:21,433
benefits and socialize the cost. We've seen that happen a

698
00:43:21,553 --> 00:43:24,874
lot of times in history and I don't want to see that with AI because I'm

699
00:43:24,994 --> 00:43:28,095
part of the field. I built an AI company, I researched it

700
00:43:28,115 --> 00:43:31,386
a lot. I don't want to see AI privatizing the

701
00:43:31,406 --> 00:43:34,649
benefits and socializing the cost. And that's exactly what I'm

702
00:43:34,669 --> 00:43:38,052
seeing. And that's why I wrote the book, is because I don't think that's fair.

703
00:43:41,556 --> 00:43:45,820
In terms of the benefits, too, which part of this equation, as

704
00:43:45,840 --> 00:43:49,844
we've talked about, is the benefits in terms of great

705
00:43:49,884 --> 00:43:53,488
stock valuations and tons of money that's pouring into this kind of small chunk

706
00:43:53,528 --> 00:43:57,271
of people running things. But on the other side, you

707
00:43:57,311 --> 00:44:00,474
know, looking through the chatbot curtain to the

708
00:44:00,554 --> 00:44:03,837
applications of AI and often just

709
00:44:03,877 --> 00:44:07,621
machine learning that are being used, what

710
00:44:07,681 --> 00:44:11,564
stuff is going on there that has you excited

711
00:44:11,705 --> 00:44:15,488
or hopeful, I guess, for the positive impact that

712
00:44:17,650 --> 00:44:20,893
My hope is more around things like DeepMind's AlphaFold than it

713
00:44:20,993 --> 00:44:24,556
is around chatbots per se. Chatbots, to

714
00:44:24,596 --> 00:44:28,479
me, have some value, but not great value. They're

715
00:44:28,499 --> 00:44:31,861
good for brainstorming. You have a human in the loop who can check it. They're

716
00:44:31,881 --> 00:44:35,784
good for computers, basically

717
00:44:35,804 --> 00:44:38,987
for a form of autocorrect that helps coders write

718
00:44:39,007 --> 00:44:43,250
faster. Even there, there's some risks, like quality

719
00:44:43,270 --> 00:44:46,793
of security may go down, some studies have shown. I

720
00:44:46,833 --> 00:44:50,636
worry about their use in medicine. AlphaFold

721
00:44:50,696 --> 00:44:54,699
is a tool built for a problem, which is take this protein

722
00:44:54,759 --> 00:44:58,482
sequence, the sequence of amino acids,

723
00:44:59,302 --> 00:45:02,865
and tell me what its three-dimensional structure is. That itself turns

724
00:45:02,905 --> 00:45:06,347
out to be an incredibly valuable problem because so much of biology is

725
00:45:06,407 --> 00:45:09,690
really about three-dimensional jigsaw puzzles. So having that

726
00:45:09,770 --> 00:45:13,174
tool, which is not a chatbot, is super helpful. And

727
00:45:13,534 --> 00:45:16,957
we should be putting more money into those kinds of things. And I'm glad to see that

728
00:45:16,997 --> 00:45:20,701
Google slash DeepMind has put money into those things. Chatbots themselves,

729
00:45:20,941 --> 00:45:24,625
I'm just not that excited. And I see all of these negative

730
00:45:24,665 --> 00:45:28,008
consequences. Or another study I saw this morning was about

731
00:45:28,309 --> 00:45:31,691
tutoring. It was a math problem

732
00:45:31,711 --> 00:45:35,533
thing with Turkish students, a thousand subjects, and

733
00:45:35,553 --> 00:45:39,296
it was very helpful when kids were working on the practice problems,

734
00:45:39,316 --> 00:45:42,598
but by the time the actual exam came, whatever benefit was there

735
00:45:42,738 --> 00:45:46,660
was lost. Probably, I would hypothesize, though I'm not sure, because

736
00:45:47,280 --> 00:45:50,562
chatbots sort of like, you know, give you some memorized problems

737
00:45:50,582 --> 00:45:53,964
you remember in the moment, but they're not really teaching you the conceptual stuff very well,

738
00:45:54,004 --> 00:45:57,746
and they're also giving you an illusion of understanding things better

739
00:45:57,786 --> 00:46:01,527
than you actually do, which is not particularly healthy. But whatever reason,

740
00:46:01,547 --> 00:46:05,049
like the data there on the educational benefit was actually very,

741
00:46:05,149 --> 00:46:08,790
very weak. And so, you know, over and over we're seeing people

742
00:46:08,830 --> 00:46:12,131
get excited about applications and they don't pan out as well as people thought.

743
00:46:12,631 --> 00:46:15,772
I'm just not sure chatbots are, to use that phrase from

744
00:46:15,812 --> 00:46:19,354
Star Wars, the droids we're looking for. Like there's,

745
00:46:21,520 --> 00:46:24,764
undoubtedly huge benefit from AI. Two other things I love

746
00:46:24,824 --> 00:46:28,970
in AI are GPS navigation directions. I

747
00:46:28,990 --> 00:46:32,174
travel all the time. That's fantastic. I'd be really disappointed to

748
00:46:32,214 --> 00:46:35,719
give that up. I actually love maps, but it's a pain

749
00:46:35,739 --> 00:46:39,163
to work with a map when you're in the city and it's not on a grid and

750
00:46:39,544 --> 00:46:42,631
whatever. And so, you know, I love GPS directions and

751
00:46:42,651 --> 00:46:46,033
I love Google search or I use DuckDuckGo for privacy reasons

752
00:46:46,053 --> 00:46:49,296
or whatever, but, you know, using

753
00:46:49,336 --> 00:46:52,478
machine learning to help you do web search is fantastic. So

754
00:46:52,658 --> 00:46:56,281
like, it's not that I don't like AI, but chatbots, you

755
00:46:56,321 --> 00:46:59,583
know, if you ask me their positive value, I'm just not sure. Like

756
00:46:59,963 --> 00:47:03,246
if chatbots disappeared from the planet Earth, at least until we could figure out

757
00:47:03,666 --> 00:47:07,369
how to make them work as well as they pretend to work, you

758
00:47:07,409 --> 00:47:10,670
know, that'd be okay with me. On Twitter the other day, I said,

759
00:47:10,890 --> 00:47:14,333
I won't remember the exact word, but I basically said, chatbots answer

760
00:47:14,393 --> 00:47:17,676
everything with the apparent authority of an encyclopedia. They

761
00:47:17,756 --> 00:47:21,058
sound like they know what they're talking about, but the reliability of

762
00:47:21,118 --> 00:47:24,281
a magic eight ball, like, I don't need that in

763
00:47:24,321 --> 00:47:28,244
my life. And, you know, I'm surprised as many people do. I'd

764
00:47:30,182 --> 00:47:33,923
I also read the study of the Turkish students. Very fascinating. I

765
00:47:34,304 --> 00:47:37,385
would hope for a lot more studies about that kind of thing. And we're seeing a

766
00:47:37,425 --> 00:47:40,586
lot of integrations of

767
00:47:40,666 --> 00:47:43,967
these current systems, which, as you said, are not reliable. They do hallucinate. We're

768
00:47:43,987 --> 00:47:47,949
seeing it pushed into education in a lot of concerning ways. And

769
00:47:47,969 --> 00:47:51,534
you kind of talked about what's lost, right, in students learning through

770
00:47:51,834 --> 00:47:55,319
whatever this unreliable tutor. Now, in

771
00:47:55,339 --> 00:47:58,643
the goal to get better AI, to get AI that can reason, that is

772
00:47:58,723 --> 00:48:02,378
reliable, that doesn't hallucinate. I've

773
00:48:02,438 --> 00:48:05,780
been wondering if that would be such a good thing, because

774
00:48:05,860 --> 00:48:09,101
right now we kind of have the backstop of we know, or if you know, you know

775
00:48:09,181 --> 00:48:12,703
these systems are not reliable, don't trust them, right? The

776
00:48:12,743 --> 00:48:16,344
idea of this complete and further integration of my GBT

777
00:48:18,885 --> 00:48:22,107
You know, I always get painted as hating AI, but in principle I

778
00:48:22,147 --> 00:48:25,328
love things like AI tutors. Like, if

779
00:48:25,368 --> 00:48:28,890
they could work, you know, they didn't hallucinate anything,

780
00:48:29,567 --> 00:48:33,071
they could reliably build a model of what the student

781
00:48:33,171 --> 00:48:36,614
understands about the problem, you could make them in

782
00:48:36,655 --> 00:48:39,938
principle cheaply enough, then you could help all kinds of

783
00:48:39,958 --> 00:48:43,081
people who can't afford individual tutors right now. I think that'd be

784
00:48:43,101 --> 00:48:46,325
great. Like I see no moral reason not

785
00:48:46,365 --> 00:48:49,588
to do that. I could see arguments that like if you had that technology would

786
00:48:49,608 --> 00:48:52,865
it be dangerous in some other way and we obviously have to look at those things

787
00:48:52,905 --> 00:48:56,126
and understand the whole, but like in principle, I love these

788
00:48:56,226 --> 00:48:59,607
use cases. I love the use case of, you know, home

789
00:48:59,647 --> 00:49:02,948
doctor because every doctor is overburdened and many people in

790
00:49:02,968 --> 00:49:06,189
the world don't have enough access to doctors. I love the idea of

791
00:49:06,329 --> 00:49:09,690
a, you know, dermatology app that you can really trust

792
00:49:09,750 --> 00:49:13,351
because not everybody has access. So, you know, I really want

793
00:49:13,391 --> 00:49:16,572
to see this stuff work, but I just think that

794
00:49:16,692 --> 00:49:20,033
what we have right now doesn't work that well. We're overselling it, that

795
00:49:20,473 --> 00:49:23,794
the overselling, it has multiple problems, including the fact that

796
00:49:23,814 --> 00:49:27,135
people get bad answers. The fact that there may be, I think there already is starting

797
00:49:27,155 --> 00:49:30,557
to be public pushback. You know, I think in some ways that Silicon

798
00:49:30,577 --> 00:49:34,358
Valley is its own worst enemy right now by constantly overselling things.

799
00:49:34,398 --> 00:49:37,899
We're getting to like a boy who cried wolf thing. Like if somebody comes

800
00:49:37,959 --> 00:49:41,263
along now with a perfect home doctor in an app, A

801
00:49:41,283 --> 00:49:44,565
lot of people are going to be like, yeah, I don't know. I tried this thing

802
00:49:44,585 --> 00:49:47,786
a couple of years ago, and it told me to take this medicine, and

803
00:49:47,806 --> 00:49:51,008
I got really sick, and forget it. Or with

804
00:49:51,068 --> 00:49:54,369
driverless cars, we're pushing them out too fast. If we suddenly put

805
00:49:54,969 --> 00:49:58,491
millions of them on the road, they would have accidents every day, and people would die, and

806
00:49:58,531 --> 00:50:01,973
people might cancel the whole thing. And it's better that we have at

807
00:50:01,993 --> 00:50:06,843
least somewhat of a phased rollout. overselling

808
00:50:06,903 --> 00:50:10,525
this stuff has a consequence, and it just doesn't work

809
00:50:10,585 --> 00:50:13,726
that well yet. We may see in software, where people have

810
00:50:13,746 --> 00:50:17,487
been using a lot for coding, that the code can't be maintained very

811
00:50:17,527 --> 00:50:20,748
well. So in software, one challenge is to write code that

812
00:50:20,808 --> 00:50:24,250
works, and another is to come back to it a year, five years, 10 years,

813
00:50:24,550 --> 00:50:27,971
20 years later, and have it still work, and be able to change

814
00:50:27,991 --> 00:50:31,232
it, because circumstances change. The most famous example of this is probably the

815
00:50:31,272 --> 00:50:34,786
Y2K problem, where people Program dates

816
00:50:34,846 --> 00:50:38,709
with two digits instead of four, and it cost millions of dollars

817
00:50:38,749 --> 00:50:42,272
of work had to be done. The public doesn't know too much about it anymore because

818
00:50:42,292 --> 00:50:47,335
they actually managed to solve that problem. It was a big mess at the time. Every

819
00:50:47,376 --> 00:50:50,638
piece of software breaks eventually because something changes. You're

820
00:50:50,658 --> 00:50:54,401
calling some other piece of software. And you need, the

821
00:50:54,421 --> 00:50:57,643
thing about programming is like in the moment that you write it, you know how it works. But then

822
00:50:57,884 --> 00:51:01,346
if you come back later, you want your code to be clear, well documented, et

823
00:51:01,687 --> 00:51:05,219
cetera, et cetera. There's a big risk We

824
00:51:05,259 --> 00:51:08,419
call it technical risk or technical debt that the code that people are

825
00:51:08,439 --> 00:51:11,920
going to write with these machines They don't really understand it and it's

826
00:51:11,960 --> 00:51:16,481
kind of like kludge together. I mean hodgepodge together and

827
00:51:17,362 --> 00:51:21,283
You know, we have trouble fixing it. So we may see some downstream costs as

828
00:51:23,228 --> 00:51:26,429
Right, we're talking about all these problems. And there's so many.

829
00:51:26,449 --> 00:51:31,411
And I think for a lot of people, it's really overwhelming and

830
00:51:31,471 --> 00:51:35,092
frightening in ways that range from, I'm

831
00:51:35,112 --> 00:51:38,653
going to lose my job to, oh my God, what's going to happen? I've

832
00:51:38,673 --> 00:51:42,634
seen the Terminator movie, right? Whether or not that's legit. What

833
00:51:42,654 --> 00:51:46,036
should people be doing here? How do people process this?

834
00:51:48,516 --> 00:51:51,867
I would like to see people get loud about

835
00:51:51,947 --> 00:51:55,708
this, talk to their congresspeople, write op-eds,

836
00:51:56,368 --> 00:51:59,509
maybe consider boycotts. So, you know, one that I suggest at

837
00:51:59,549 --> 00:52:02,969
the end of the book, and I'm going to set something up on this at

838
00:52:03,069 --> 00:52:06,450
teamingsiliconvalley.org, is maybe

839
00:52:06,490 --> 00:52:10,171
we should boycott any art that isn't properly licensed.

840
00:52:10,751 --> 00:52:14,272
I mean, any, sorry, any generative AI that isn't, that's

841
00:52:14,392 --> 00:52:17,781
using art that isn't properly licensed. Right? We

842
00:52:17,801 --> 00:52:21,643
should just say, look, as consumers, we want to have all

843
00:52:21,663 --> 00:52:25,404
of this stuff be ethically sourced. And we'll just wait. We'll

844
00:52:25,744 --> 00:52:29,346
sit and wait until you do that. And so I think

845
00:52:29,466 --> 00:52:33,107
collective action here to say, we want AI

846
00:52:33,147 --> 00:52:36,349
to be done in an ethical way, in a way that

847
00:52:36,409 --> 00:52:39,690
is good for everybody and not just a few people. We're going to stand by

848
00:52:39,730 --> 00:52:43,185
that. Like, if we could actually get enough people To

849
00:52:43,245 --> 00:52:46,969
agree on that, we could change Silicon

850
00:52:46,989 --> 00:52:50,953
Valley. Silicon Valley used to, I think, care

851
00:52:51,013 --> 00:52:54,277
more about the consumers. And nowadays has an attitude of like,

852
00:52:54,297 --> 00:52:57,640
yeah, we're just gonna make money off those people. And we

853
00:52:59,282 --> 00:53:02,785
Yeah, absolutely. Well, Gary, I really appreciate your time. This