1
00:00:00,240 --> 00:00:05,080
Speaker 1: Welcome to the Nick, Dick and Pole Show. We are

2
00:00:05,240 --> 00:00:09,240
Speaker 1: this week. So there are people that get really hyped

3
00:00:09,320 --> 00:00:12,120
Speaker 1: up about certain topics, like you can bring up politics

4
00:00:12,119 --> 00:00:15,320
Speaker 1: and they'll get very upset. You can bring up the

5
00:00:15,320 --> 00:00:18,960
Speaker 1: best pizza in New York, the best burrito in San Francisco.

6
00:00:18,640 --> 00:00:21,080
Speaker 2: Cat videos, cat videos.

7
00:00:21,480 --> 00:00:25,160
Speaker 1: But if you say the next two words to Paul Kadrowski,

8
00:00:25,760 --> 00:00:29,880
Speaker 1: you will see the most passionate response from anyone on earth.

9
00:00:29,880 --> 00:00:31,960
Speaker 1: Are you ready? Data centers?

10
00:00:32,280 --> 00:00:34,240
Speaker 3: You know what's funny about it is it actually pisses

11
00:00:34,280 --> 00:00:35,519
Speaker 3: me off that that pisses me off.

12
00:00:35,600 --> 00:00:37,000
Speaker 2: It's one of these matter problems.

13
00:00:37,479 --> 00:00:40,600
Speaker 3: See, like they really pissed me off, but it pisses

14
00:00:40,640 --> 00:00:43,080
Speaker 3: me off that I even care, which just feels ridiculous, right.

15
00:00:43,120 --> 00:00:45,680
Speaker 3: I Mean it's like, I don't know, obsessing about climbing

16
00:00:45,720 --> 00:00:54,640
Speaker 3: gyms or something.

17
00:01:00,480 --> 00:01:02,960
Speaker 1: So for people listening, they're going to think they're literally

18
00:01:03,000 --> 00:01:04,720
Speaker 1: about to turn this off because they're going to be like,

19
00:01:05,000 --> 00:01:07,760
Speaker 1: I do not want to listen to a hour long

20
00:01:07,800 --> 00:01:10,000
Speaker 1: podcast about data centers.

21
00:01:09,800 --> 00:01:12,880
Speaker 2: Right right, especially one that's going to probably run long.

22
00:01:14,319 --> 00:01:19,680
Speaker 1: That's right, But but so on our on our text

23
00:01:19,680 --> 00:01:22,199
Speaker 1: thread for the past, how long has it been dickons.

24
00:01:21,760 --> 00:01:24,720
Speaker 2: He's been sending these data centers quite sometimes six months,

25
00:01:24,840 --> 00:01:25,480
Speaker 2: hundreds of them.

26
00:01:25,640 --> 00:01:30,479
Speaker 1: Yeah, we have been getting conscoundreds of passionate texts from

27
00:01:30,480 --> 00:01:32,760
Speaker 1: Paul about data centers. Do you want to do? You

28
00:01:32,760 --> 00:01:36,560
Speaker 1: want to start this? Like Ki, just kick just lead

29
00:01:36,640 --> 00:01:38,240
Speaker 1: us in, all right, Dick and I are going to

30
00:01:38,280 --> 00:01:40,800
Speaker 1: take a nap and I'll just talk and you just

31
00:01:40,840 --> 00:01:42,479
Speaker 1: what is it the pisses you off so much about

32
00:01:42,560 --> 00:01:43,440
Speaker 1: data centers right now?

33
00:01:43,840 --> 00:01:47,560
Speaker 3: I think it's that people well to at least three things,

34
00:01:47,560 --> 00:01:47,960
Speaker 3: but I'll.

35
00:01:47,800 --> 00:01:49,960
Speaker 2: Give you at least three. At least three. But that's

36
00:01:49,960 --> 00:01:52,000
Speaker 2: at the high level. There's a thousand at a lower level.

37
00:01:52,840 --> 00:01:54,800
Speaker 1: Do we need to set anything up for listeners to say?

38
00:01:54,920 --> 00:01:59,680
Speaker 3: Data centers are these giant air conditioned buildings containing huge

39
00:01:59,760 --> 00:02:03,880
Speaker 3: number members of GPUs and video chips in giant server racks.

40
00:02:05,000 --> 00:02:07,080
Speaker 3: They've been around forever. They've been around for as long

41
00:02:07,120 --> 00:02:09,720
Speaker 3: as we've had web services, back early days of Amazon,

42
00:02:09,720 --> 00:02:11,760
Speaker 3: early days of Google, early days of Yahoo, we've had

43
00:02:11,800 --> 00:02:13,920
Speaker 3: data centers. So that's nothing new, right. The idea that

44
00:02:13,960 --> 00:02:17,040
Speaker 3: data centers are somehow new is of course just wrong.

45
00:02:17,120 --> 00:02:19,880
Speaker 3: The difference is that all of a sudden spending on

46
00:02:19,960 --> 00:02:21,920
Speaker 3: data centers went from being kind of like this, like

47
00:02:22,000 --> 00:02:23,919
Speaker 3: growing four percent a year to growing like twenty two

48
00:02:23,960 --> 00:02:25,360
Speaker 3: percent a year just in the last two and.

49
00:02:25,320 --> 00:02:25,799
Speaker 2: A half years.

50
00:02:25,840 --> 00:02:26,800
Speaker 1: Right, but that's because of AI.

51
00:02:26,960 --> 00:02:34,360
Speaker 3: Right, Really, don't piss me off. So yeah, it has

52
00:02:34,400 --> 00:02:35,880
Speaker 3: to do with this AI stuff. But the point is

53
00:02:35,880 --> 00:02:40,160
Speaker 3: it's all of a sudden taken off and everyone's kind

54
00:02:40,160 --> 00:02:43,160
Speaker 3: of playing this game of trying to say, you have

55
00:02:43,160 --> 00:02:45,600
Speaker 3: one hundred megawat data center, I can build a giga

56
00:02:45,600 --> 00:02:47,720
Speaker 3: wat data center. You can build a giga wat data center.

57
00:02:47,720 --> 00:02:49,959
Speaker 3: I can build a ten giga wat data center. And

58
00:02:50,440 --> 00:02:54,840
Speaker 3: everyone's piling into this business of proposing to construct giant

59
00:02:54,919 --> 00:02:57,440
Speaker 3: data centers all over the US, all over North America,

60
00:02:57,440 --> 00:02:59,680
Speaker 3: all over the world, and so on, and all in

61
00:03:00,080 --> 00:03:04,200
Speaker 3: large part because of this notional idea that it's the future.

62
00:03:04,240 --> 00:03:07,840
Speaker 3: These are the new engines of capitalism, that if I

63
00:03:07,880 --> 00:03:09,760
Speaker 3: don't have a data center, it's not just that I

64
00:03:09,760 --> 00:03:12,399
Speaker 3: can't fuel AI. I might as well just shut down

65
00:03:12,400 --> 00:03:16,760
Speaker 3: the economy because from a regional economic development standpoint, that's

66
00:03:16,800 --> 00:03:18,519
Speaker 3: what matters. And you'll hear this all the time from

67
00:03:18,520 --> 00:03:21,359
Speaker 3: officials in Georgia and Virginia and everywhere else. Why are

68
00:03:21,360 --> 00:03:23,680
Speaker 3: you so obsessed about these things. It's because it's an

69
00:03:23,680 --> 00:03:27,480
Speaker 3: annuity income for regional tax base, and it's because it's

70
00:03:27,480 --> 00:03:27,960
Speaker 3: the future.

71
00:03:28,080 --> 00:03:30,560
Speaker 2: They're keeping told that, they're keeping told it's the future.

72
00:03:30,760 --> 00:03:35,760
Speaker 2: Let's talk about all the first of all, ignoring ignoring

73
00:03:35,840 --> 00:03:40,720
Speaker 2: the climate change impact, which is like, yes, it's the

74
00:03:40,720 --> 00:03:43,480
Speaker 2: same as other that missus Lincoln always the play ignoring

75
00:03:43,760 --> 00:03:51,040
Speaker 2: but ignoring the climate change impact. There are multiple, multiple,

76
00:03:51,080 --> 00:03:55,880
Speaker 2: multiple multiplicious oh endless. I mean the first one or

77
00:03:55,960 --> 00:03:58,640
Speaker 2: one of one of the most obvious that's really only

78
00:03:58,680 --> 00:04:01,080
Speaker 2: starting to be I mean this morning we were looking

79
00:04:01,120 --> 00:04:03,800
Speaker 2: at the news and seeing that it's starting to be

80
00:04:03,800 --> 00:04:06,400
Speaker 2: traded this way, and the debt from some of these companies,

81
00:04:06,760 --> 00:04:11,600
Speaker 2: the temporal disconnect between the life of the asset GPUs yeah,

82
00:04:11,640 --> 00:04:15,520
Speaker 2: and the term on the debt of upwards forty years

83
00:04:15,520 --> 00:04:17,480
Speaker 2: an Oracles case. Yeah, you know, you've got a two

84
00:04:17,520 --> 00:04:20,680
Speaker 2: and a half year asset being you know, finance, it's

85
00:04:20,680 --> 00:04:22,600
Speaker 2: supposed to be rolled over every year.

86
00:04:23,040 --> 00:04:25,960
Speaker 1: Can you explain that to people that go for it? Wait,

87
00:04:26,200 --> 00:04:29,000
Speaker 1: hold on, you can explain it. You ran a company

88
00:04:29,480 --> 00:04:31,279
Speaker 1: they had no debt, though I like to defer to

89
00:04:31,320 --> 00:04:34,120
Speaker 1: Paul and then come in for the joke. Can you

90
00:04:34,160 --> 00:04:36,719
Speaker 1: explain what he just said in English.

91
00:04:36,839 --> 00:04:39,400
Speaker 3: So data centers are expensive, yeah, but money has to

92
00:04:39,400 --> 00:04:40,159
Speaker 3: come from somewhere.

93
00:04:40,320 --> 00:04:40,560
Speaker 1: Yeah.

94
00:04:40,680 --> 00:04:43,680
Speaker 3: Companies don't want to fund them entirely from profits because

95
00:04:43,680 --> 00:04:45,920
Speaker 3: it's too much money, so they raise debt to do it,

96
00:04:46,360 --> 00:04:48,599
Speaker 3: and then they raise more debt later on because you

97
00:04:48,640 --> 00:04:51,400
Speaker 3: have to keep refinancing these things. But the debt has

98
00:04:51,440 --> 00:04:54,760
Speaker 3: a term, often thirty years, very long term debt. But

99
00:04:54,920 --> 00:04:57,640
Speaker 3: you pay the debt, pay off the debt, and make

100
00:04:57,880 --> 00:05:02,520
Speaker 3: interest payments based on the use renting the GPUs. GPUs

101
00:05:02,560 --> 00:05:05,159
Speaker 3: don't last very long for at least two reasons. One

102
00:05:05,240 --> 00:05:08,479
Speaker 3: is that technology changes pretty quickly, so you know, every

103
00:05:08,480 --> 00:05:10,400
Speaker 3: four or five years, you just turn them over because

104
00:05:10,400 --> 00:05:13,440
Speaker 3: there's some newer chip from Nvidio or whatever else and then.

105
00:05:13,520 --> 00:05:15,279
Speaker 3: But the other reason is that if you use chips

106
00:05:15,320 --> 00:05:17,680
Speaker 3: for training, it's kind of like using a car for

107
00:05:17,760 --> 00:05:20,880
Speaker 3: street racing. You're running a chip flat out seven days

108
00:05:20,880 --> 00:05:23,320
Speaker 3: a week, twenty four hours a day, so that the

109
00:05:23,839 --> 00:05:27,120
Speaker 3: thermal degradation the chip breaks down fast. So it's not

110
00:05:27,320 --> 00:05:29,840
Speaker 3: just that the chips have to be changed every few years.

111
00:05:29,880 --> 00:05:32,279
Speaker 3: But if you use a chip for training, it's like

112
00:05:32,320 --> 00:05:34,279
Speaker 3: heat stress. It's just sitting there in this really hot

113
00:05:34,400 --> 00:05:35,720
Speaker 3: environment running flat out.

114
00:05:35,800 --> 00:05:40,800
Speaker 2: So there are two things that companies do to gussy

115
00:05:40,960 --> 00:05:46,279
Speaker 2: up the looks of that's thank you, I went old

116
00:05:46,320 --> 00:05:47,640
Speaker 2: West there. Yeah, that's good.

117
00:05:48,440 --> 00:05:49,240
Speaker 1: One of the two things.

118
00:05:49,279 --> 00:05:52,280
Speaker 2: One they take the debt and they move it off

119
00:05:52,320 --> 00:05:55,279
Speaker 2: the balance sheet and do it through some SPV. You

120
00:05:55,400 --> 00:05:59,040
Speaker 2: one is reminded. One might say, isn't that what Enron did?

121
00:05:59,240 --> 00:06:02,920
Speaker 2: And you would be correct, that is what. Yeah, let's

122
00:06:02,960 --> 00:06:05,880
Speaker 2: not put this right here, let's move this over here

123
00:06:05,920 --> 00:06:08,920
Speaker 2: and to a third party and have And the second

124
00:06:08,920 --> 00:06:11,840
Speaker 2: thing they do is, of course they don't mark down

125
00:06:11,880 --> 00:06:13,920
Speaker 2: the ad They don't you know, they don't mark down

126
00:06:13,960 --> 00:06:15,480
Speaker 2: the life of the acid over two and a half years.

127
00:06:15,520 --> 00:06:17,520
Speaker 2: They're like, well, really, these things are going to last

128
00:06:17,560 --> 00:06:18,799
Speaker 2: like six to eight years.

129
00:06:18,640 --> 00:06:20,479
Speaker 3: Which is which which was true in the old world

130
00:06:20,520 --> 00:06:22,120
Speaker 3: of data centers. So I was running like an S

131
00:06:22,200 --> 00:06:25,680
Speaker 3: three cloud for Amazon for storage. You could realistically say

132
00:06:25,720 --> 00:06:27,760
Speaker 3: most of the hardware was six or seven years. But

133
00:06:27,800 --> 00:06:31,120
Speaker 3: it was completely different because the more people who use

134
00:06:31,360 --> 00:06:33,720
Speaker 3: Amazon for storage, the more money I make is Amazon,

135
00:06:33,720 --> 00:06:35,440
Speaker 3: I don't have to keep adding chips for can I

136
00:06:35,680 --> 00:06:36,520
Speaker 3: can I still talk here?

137
00:06:36,600 --> 00:06:37,880
Speaker 2: No question?

138
00:06:38,040 --> 00:06:40,280
Speaker 1: That's really important for this right when you build a

139
00:06:40,360 --> 00:06:43,479
Speaker 1: data center. How much of the money is sixty percent,

140
00:06:44,240 --> 00:06:46,520
Speaker 1: which is the sixty percent for the chokes, So the

141
00:06:46,680 --> 00:06:49,240
Speaker 1: chips are sixty percent, So the building's only forty percent.

142
00:06:49,279 --> 00:06:52,279
Speaker 1: The air conditioning, the food trucks, all the other nonsense.

143
00:06:52,440 --> 00:06:54,719
Speaker 3: There's a few or few food trucks in you might think.

144
00:06:54,560 --> 00:06:55,920
Speaker 1: Because there's not many people working there.

145
00:06:56,160 --> 00:06:57,840
Speaker 3: Well, no, it's really tiny. Yeah that's one of them.

146
00:06:57,920 --> 00:06:59,600
Speaker 3: But that's actually an important point. These things are not

147
00:06:59,720 --> 00:07:01,880
Speaker 3: huge that was going to be.

148
00:07:02,040 --> 00:07:05,359
Speaker 2: That's they go into these communities and we're going to

149
00:07:05,400 --> 00:07:08,440
Speaker 2: create you know, at least five hundred, even even a

150
00:07:08,480 --> 00:07:12,280
Speaker 2: low number like five hundred, they the community doesn't end

151
00:07:12,360 --> 00:07:12,640
Speaker 2: up going.

152
00:07:12,640 --> 00:07:15,240
Speaker 1: Well, this happened. My dad lives in Racine, Wisconsin, and

153
00:07:15,280 --> 00:07:18,560
Speaker 1: they were going to build a massive data center there.

154
00:07:20,920 --> 00:07:22,840
Speaker 1: I forget who was the guys who make the chips

155
00:07:22,880 --> 00:07:28,880
Speaker 1: for the iPhone. No, no, I forget the name of

156
00:07:28,880 --> 00:07:31,520
Speaker 1: the company. Anyway, So they had all these bids and

157
00:07:31,520 --> 00:07:33,120
Speaker 1: they offered all these deals, and they said they were

158
00:07:33,120 --> 00:07:34,920
Speaker 1: going to build all these houses, and it turned out

159
00:07:34,920 --> 00:07:38,840
Speaker 1: like they needed no housing. Uh and uh yeah.

160
00:07:38,680 --> 00:07:40,680
Speaker 3: But it's even it's even worse than that in some

161
00:07:40,720 --> 00:07:42,520
Speaker 3: ways because of the if you take the total cost

162
00:07:42,560 --> 00:07:44,840
Speaker 3: of the data center and the tax abatements that people give.

163
00:07:44,880 --> 00:07:47,320
Speaker 3: They give them guaranteed power, guaranteed water, they cut the

164
00:07:47,360 --> 00:07:50,200
Speaker 3: real estate tax. In many cases, you're paying over fifty

165
00:07:50,240 --> 00:07:51,600
Speaker 3: million per job created.

166
00:07:52,000 --> 00:07:55,160
Speaker 2: So just give me the money. Yeah, I don't want that.

167
00:07:55,240 --> 00:07:56,880
Speaker 2: I got a great way to save you twenty five

168
00:07:56,880 --> 00:07:59,720
Speaker 2: million dollars. Give me twenty five million dollars and don't

169
00:07:59,720 --> 00:07:59,920
Speaker 2: do that.

170
00:08:00,360 --> 00:08:02,440
Speaker 3: Don't do the data center. Cut at the middleman.

171
00:08:03,120 --> 00:08:05,760
Speaker 1: Okay, So let me just ask the obvious question here.

172
00:08:05,440 --> 00:08:06,720
Speaker 2: Wait, there's more.

173
00:08:07,040 --> 00:08:10,000
Speaker 1: Why Why should anyone who still listening to this podcast

174
00:08:10,040 --> 00:08:10,960
Speaker 1: is maybe two or three people.

175
00:08:11,000 --> 00:08:13,600
Speaker 3: Oh, you know what, here's what I actually I'm gonna,

176
00:08:13,640 --> 00:08:15,520
Speaker 3: I'm gonna I'm going to flame out for a second.

177
00:08:15,680 --> 00:08:17,840
Speaker 3: I actually think it's ridiculous that people aren't paying attention

178
00:08:18,040 --> 00:08:20,440
Speaker 3: because it's probably the single most important thing that could

179
00:08:20,480 --> 00:08:23,760
Speaker 3: cause the next major financial crisis in the world. And

180
00:08:23,800 --> 00:08:26,840
Speaker 3: the reason is because all major prior financial crises have

181
00:08:26,880 --> 00:08:28,880
Speaker 3: been tied to one of four things. Right, It's either

182
00:08:28,960 --> 00:08:32,480
Speaker 3: an amazing technology story that gets people over excited, really

183
00:08:32,520 --> 00:08:37,120
Speaker 3: loose credit, real estate, or some kind of government backstop

184
00:08:37,160 --> 00:08:39,200
Speaker 3: for something that causes people to do too much of

185
00:08:39,240 --> 00:08:40,960
Speaker 3: a thing. And it could have been Fanny and Freddie

186
00:08:40,960 --> 00:08:42,559
Speaker 3: back in the financial crisis. It could have been the

187
00:08:42,600 --> 00:08:45,640
Speaker 3: south Sea bubble back in the eighteenth century. This nonsense

188
00:08:45,679 --> 00:08:50,320
Speaker 3: has all four it's the first buy. It's got a

189
00:08:50,360 --> 00:08:53,800
Speaker 3: real estate component. Data centers are essentially built on the

190
00:08:53,840 --> 00:08:56,360
Speaker 3: idea of scarce land land that has water, power and

191
00:08:56,400 --> 00:08:58,520
Speaker 3: interconnect I can't just build one anywhere I want to,

192
00:08:58,800 --> 00:09:01,559
Speaker 3: so there's a scarcity compone that's causing people to speculate,

193
00:09:01,640 --> 00:09:04,560
Speaker 3: like a kind of mini Chinatown about real estate. There's

194
00:09:04,559 --> 00:09:07,240
Speaker 3: a there's a credit component. We're seeing that already with

195
00:09:07,320 --> 00:09:09,400
Speaker 3: all these private credit firms all coming and saying, no,

196
00:09:09,840 --> 00:09:11,640
Speaker 3: pick me, I can create the STVA for you. This

197
00:09:11,640 --> 00:09:14,160
Speaker 3: will be exciting, right, So there's a credit component. There's

198
00:09:14,160 --> 00:09:16,840
Speaker 3: a government component, because the governments have now all decided

199
00:09:17,440 --> 00:09:19,679
Speaker 3: we have to win because China can't win, so we're

200
00:09:19,679 --> 00:09:21,920
Speaker 3: gonna That's why we got into this big discussion I

201
00:09:21,960 --> 00:09:23,440
Speaker 3: think this week about whether or not.

202
00:09:23,559 --> 00:09:25,040
Speaker 2: Yeah, that was the word backstop.

203
00:09:27,240 --> 00:09:29,280
Speaker 3: I thought I didn't mean that, but that's implicit and

204
00:09:29,360 --> 00:09:31,240
Speaker 3: a lot of what these companies are doing. They approach

205
00:09:31,280 --> 00:09:33,679
Speaker 3: other countries and say you need to build data centers

206
00:09:33,720 --> 00:09:35,920
Speaker 3: because don't you want to have a Norwegian AI. Don't

207
00:09:35,920 --> 00:09:38,440
Speaker 3: you want to have a Swiss AI, Mongolian AI. But

208
00:09:38,559 --> 00:09:40,959
Speaker 3: that's all this cynical idea of trying to wrap AI

209
00:09:41,040 --> 00:09:43,160
Speaker 3: in a flag. And so you need you need each

210
00:09:43,160 --> 00:09:45,480
Speaker 3: of these pieces, but you also need the technology story.

211
00:09:45,520 --> 00:09:46,880
Speaker 2: You need a really great story.

212
00:09:46,920 --> 00:09:48,680
Speaker 3: It's not enough to just say, like, you know, we'll

213
00:09:48,679 --> 00:09:51,880
Speaker 3: build small free you know, Tamagatchi's or something. We need

214
00:09:51,880 --> 00:09:54,200
Speaker 3: some kind of great technology story. And so this has

215
00:09:54,240 --> 00:09:56,760
Speaker 3: one of the best technology stories ever because it actually

216
00:09:56,760 --> 00:09:59,240
Speaker 3: happens to be true. AI is incredibly important, It actually

217
00:09:59,280 --> 00:10:02,079
Speaker 3: works to change people's lives. So you put all those

218
00:10:02,120 --> 00:10:05,920
Speaker 3: four pieces together and there's no way this doesn't become

219
00:10:06,000 --> 00:10:08,120
Speaker 3: one of the largest financial bubbles in history. So the

220
00:10:08,160 --> 00:10:11,240
Speaker 3: idea that people say, oh, data center is boring, screw you.

221
00:10:11,679 --> 00:10:13,280
Speaker 3: I mean, this is like the sort of stuff that

222
00:10:13,320 --> 00:10:15,839
Speaker 3: takes down economy. This is has the potential to be.

223
00:10:15,880 --> 00:10:17,000
Speaker 2: Larger than all the other ones.

224
00:10:17,080 --> 00:10:19,880
Speaker 1: We five billion or something this year, what's the numbers.

225
00:10:20,200 --> 00:10:21,839
Speaker 2: It's going to be closer to three hundred.

226
00:10:21,559 --> 00:10:24,920
Speaker 3: Billion, Jesus, and then and then probably two trillion over

227
00:10:24,960 --> 00:10:25,440
Speaker 3: the next too.

228
00:10:25,800 --> 00:10:30,440
Speaker 2: So all all of this now going on while are.

229
00:10:31,240 --> 00:10:35,120
Speaker 2: You know, you need power to plug in to make

230
00:10:35,160 --> 00:10:39,640
Speaker 2: the GPS go, which we're horrible at. Our grid is

231
00:10:39,640 --> 00:10:44,440
Speaker 2: already trash. And you know, compared to the grid in China,

232
00:10:44,800 --> 00:10:48,439
Speaker 2: which they have ample power, they have ample power.

233
00:10:48,480 --> 00:10:51,520
Speaker 3: And it's also this that there's this perverse incentive to

234
00:10:51,640 --> 00:10:54,040
Speaker 3: agree to add a data center to the grid because

235
00:10:54,120 --> 00:10:57,280
Speaker 3: regional politicians want it for all the reasons we've talked about.

236
00:10:57,640 --> 00:11:00,040
Speaker 3: The grid want, the utilities want it because it's a

237
00:11:00,160 --> 00:11:02,600
Speaker 3: very predictable source. I know how much load there's going

238
00:11:02,640 --> 00:11:04,599
Speaker 3: to be. I can add it. It's not like speculating it.

239
00:11:04,720 --> 00:11:06,560
Speaker 3: I know what it's going to do. But the problem

240
00:11:06,600 --> 00:11:09,720
Speaker 3: is adding it adds well as much as gigawatts of

241
00:11:09,720 --> 00:11:12,480
Speaker 3: an additional power needs to the grid, and where's that

242
00:11:12,520 --> 00:11:13,160
Speaker 3: going to come from?

243
00:11:13,240 --> 00:11:13,400
Speaker 2: Right?

244
00:11:13,520 --> 00:11:15,640
Speaker 1: In the one article I read that you sent of

245
00:11:15,640 --> 00:11:17,760
Speaker 1: the seven hundred and sixty two, I was I saw

246
00:11:17,800 --> 00:11:19,760
Speaker 1: that I maybe read a couple of them, but I

247
00:11:19,800 --> 00:11:23,480
Speaker 1: saw that one data center is equal to one hundred

248
00:11:23,559 --> 00:11:25,240
Speaker 1: thousand households in power.

249
00:11:25,480 --> 00:11:27,800
Speaker 2: That's a small data center, but yes, a.

250
00:11:27,760 --> 00:11:30,200
Speaker 1: Big one is what ten x text easily.

251
00:11:29,840 --> 00:11:31,360
Speaker 2: Ten xc right.

252
00:11:31,440 --> 00:11:33,640
Speaker 3: So so so the problem is the grid's not as

253
00:11:33,679 --> 00:11:35,040
Speaker 3: Dick was saying the grid's not well set up to

254
00:11:35,080 --> 00:11:37,480
Speaker 3: add all that extra power. So you have these so

255
00:11:37,480 --> 00:11:39,720
Speaker 3: if you have these fanciful ideas, people will do there's

256
00:11:39,760 --> 00:11:41,520
Speaker 3: all kinds of great terms of our People will say, well,

257
00:11:41,520 --> 00:11:43,760
Speaker 3: I'll add power behind the meter and it's like, what

258
00:11:44,559 --> 00:11:45,000
Speaker 3: is that?

259
00:11:45,160 --> 00:11:45,880
Speaker 2: Right, Well, that's the.

260
00:11:45,840 --> 00:11:48,480
Speaker 3: Idea that I won't rely on the grid to provide

261
00:11:48,480 --> 00:11:51,040
Speaker 3: me with power because who knows when they'll add you know,

262
00:11:51,080 --> 00:11:54,280
Speaker 3: that nuclear plant that I needed, So I build my own, right,

263
00:11:54,320 --> 00:11:56,920
Speaker 3: So this is the idea increasingly in Texas, New Mexico.

264
00:11:57,080 --> 00:11:59,040
Speaker 1: Is this ferny and things like that.

265
00:11:59,200 --> 00:12:00,959
Speaker 3: Yeah, so that's an example of one of them. But

266
00:12:01,000 --> 00:12:03,000
Speaker 3: there's lots of others, Googles doing this, oracles doing this.

267
00:12:03,040 --> 00:12:04,600
Speaker 3: There's lots of people who are doing the idea of

268
00:12:04,640 --> 00:12:07,480
Speaker 3: adding power behind behind the meter because I don't want

269
00:12:07,480 --> 00:12:09,480
Speaker 3: to wait for the grid to do it. So now

270
00:12:09,480 --> 00:12:11,280
Speaker 3: you get into a secondary problem. So where's the money

271
00:12:11,280 --> 00:12:13,480
Speaker 3: coming from to build your own private natural gas plant?

272
00:12:13,520 --> 00:12:16,320
Speaker 3: Well that's more debt. So now you've got debt for

273
00:12:16,360 --> 00:12:20,240
Speaker 3: the natural gas plant being paid for by GPUs that

274
00:12:20,320 --> 00:12:22,840
Speaker 3: are only you're gonna have to return over every four years,

275
00:12:23,160 --> 00:12:25,760
Speaker 3: thirty year debt on a natural gas plant that probably

276
00:12:25,760 --> 00:12:28,599
Speaker 3: has a forty year lifespan. So that's what they call it,

277
00:12:28,720 --> 00:12:30,760
Speaker 3: like a like a mismatch in the debt markets. Like

278
00:12:30,800 --> 00:12:32,920
Speaker 3: you have a thing that's generating money for only a

279
00:12:32,920 --> 00:12:36,120
Speaker 3: short period of time paying for something an obligation, you

280
00:12:36,160 --> 00:12:37,560
Speaker 3: have the stretches up forty years.

281
00:12:37,800 --> 00:12:38,800
Speaker 2: This never works.

282
00:12:39,040 --> 00:12:41,160
Speaker 3: This goes very badly over and over and over again,

283
00:12:41,160 --> 00:12:43,200
Speaker 3: over like two hundred years of financial history.

284
00:12:43,440 --> 00:12:46,959
Speaker 1: So how does it play out in a like how

285
00:12:47,000 --> 00:12:49,320
Speaker 1: does it end? Do we have like you know, like

286
00:12:49,400 --> 00:12:52,400
Speaker 1: you drive up and down any city street today and

287
00:12:52,440 --> 00:12:54,920
Speaker 1: there's more four lease signs than there are Like, do

288
00:12:54,960 --> 00:12:57,560
Speaker 1: we just do dated centers become ice skating rings? Like

289
00:12:57,760 --> 00:13:01,280
Speaker 1: I say, laser tags, laser tag great rooms for late

290
00:13:01,679 --> 00:13:05,040
Speaker 1: Can Nick invest in a laser tag development business right now?

291
00:13:05,080 --> 00:13:06,760
Speaker 2: Yeah? Like that, but I think you're going to have

292
00:13:06,800 --> 00:13:10,079
Speaker 2: free space. One of the many things that pisses me

293
00:13:10,120 --> 00:13:11,320
Speaker 2: off about this subject.

294
00:13:11,120 --> 00:13:12,959
Speaker 3: Is people all want to say, like how does this end?

295
00:13:12,960 --> 00:13:14,960
Speaker 3: And they just want to pick one thing, and.

296
00:13:14,840 --> 00:13:15,559
Speaker 1: That's just lots.

297
00:13:16,040 --> 00:13:18,439
Speaker 3: There's like seventeen different ways that can break this. The

298
00:13:18,480 --> 00:13:21,559
Speaker 3: only the only thing that's inevitable is it's not sustainable.

299
00:13:21,600 --> 00:13:24,920
Speaker 3: We'll either have over capacity, we have a debt spiral

300
00:13:24,960 --> 00:13:27,360
Speaker 3: that cracks all of this stuff, or you know, you

301
00:13:27,400 --> 00:13:29,160
Speaker 3: can even dive under the hood and say most of

302
00:13:29,160 --> 00:13:32,560
Speaker 3: these data centers are being created backward looking, and the

303
00:13:32,559 --> 00:13:35,319
Speaker 3: reason why is that they all imagine that if I

304
00:13:35,480 --> 00:13:37,920
Speaker 3: just run my training cycles long enough on top of

305
00:13:37,960 --> 00:13:40,280
Speaker 3: these data centers, I can push open AIE out of

306
00:13:40,280 --> 00:13:42,920
Speaker 3: the market, which is a completely ridiculous idea. Right, We're

307
00:13:42,920 --> 00:13:46,000
Speaker 3: already seeing that the improvement in large language models has

308
00:13:46,040 --> 00:13:48,320
Speaker 3: really tailed off in the last two years. It's costs

309
00:13:48,360 --> 00:13:51,000
Speaker 3: more money, takes more time, and I can't tell the

310
00:13:51,000 --> 00:13:53,360
Speaker 3: difference from the last one. So GPT five versus four

311
00:13:53,440 --> 00:13:55,880
Speaker 3: is a good example. And so the idea that I

312
00:13:55,920 --> 00:13:58,320
Speaker 3: can train my way out of this, which is what

313
00:13:58,720 --> 00:14:00,839
Speaker 3: floats a lot of these data center or boats. It's

314
00:14:00,840 --> 00:14:03,880
Speaker 3: not about inference, it's about training new models. That's just

315
00:14:03,920 --> 00:14:06,400
Speaker 3: not true. It's all this backward thinking, this idea that

316
00:14:06,440 --> 00:14:09,440
Speaker 3: I can continue to run this old training model, and

317
00:14:09,480 --> 00:14:11,040
Speaker 3: it doesn't work that way. So not only are we

318
00:14:11,080 --> 00:14:13,480
Speaker 3: building too many, we're building it based on horrible logic.

319
00:14:14,520 --> 00:14:16,240
Speaker 1: I'm going to just push back for a second. I

320
00:14:16,240 --> 00:14:17,960
Speaker 1: have no idea what I'm talking about, but I'm going

321
00:14:18,000 --> 00:14:19,720
Speaker 1: to pretend I do for a minute, and Dick is

322
00:14:19,720 --> 00:14:20,520
Speaker 1: going to back me up.

323
00:14:20,560 --> 00:14:21,720
Speaker 3: It never stopped you before.

324
00:14:21,920 --> 00:14:26,120
Speaker 1: So when the housing bubble crashed, and how am I

325
00:14:26,120 --> 00:14:31,520
Speaker 1: going to back you up? I'm agree with me so

326
00:14:31,560 --> 00:14:32,600
Speaker 1: that we can argue with Paul.

327
00:14:32,640 --> 00:14:35,520
Speaker 2: I know you are, but what am I I do?

328
00:14:35,680 --> 00:14:36,320
Speaker 2: Get there eventually?

329
00:14:36,400 --> 00:14:41,040
Speaker 1: Yeah, okay, So the housing bubble when it crashed, it

330
00:14:41,160 --> 00:14:45,400
Speaker 1: was devastating for society because every people's homes and their investments,

331
00:14:45,440 --> 00:14:48,160
Speaker 1: and their money and and and their savings were all a.

332
00:14:48,200 --> 00:14:50,120
Speaker 3: Fine and the global banking system really and.

333
00:14:50,080 --> 00:14:52,960
Speaker 1: The global banking system, but it was, but it was

334
00:14:53,040 --> 00:14:57,440
Speaker 1: more so that it ended up affecting actual moms and

335
00:14:57,440 --> 00:15:00,680
Speaker 1: pops and people and so on. When the stock market

336
00:15:00,680 --> 00:15:02,800
Speaker 1: crash in ninety nine, the same thing, like we had

337
00:15:02,840 --> 00:15:04,800
Speaker 1: all put money into, or a lot of people would

338
00:15:04,800 --> 00:15:06,440
Speaker 1: put money into these companies the world.

339
00:15:06,440 --> 00:15:07,920
Speaker 2: I know exactly where this is going.

340
00:15:07,960 --> 00:15:09,640
Speaker 1: And then Dick's going to finish my question.

341
00:15:09,840 --> 00:15:13,600
Speaker 2: Where's going? So let's see if I can. Let me

342
00:15:13,640 --> 00:15:15,640
Speaker 2: see if I can finish it. But we are now

343
00:15:15,680 --> 00:15:18,120
Speaker 2: in I think I can be Nick Bilton for a

344
00:15:18,160 --> 00:15:22,520
Speaker 2: moment here, But we're now in a world where eighty

345
00:15:22,600 --> 00:15:27,520
Speaker 2: percent eighty five percent, eighty seven percent of the shareholders

346
00:15:27,560 --> 00:15:30,720
Speaker 2: of you know these stocks, the Man seven are only

347
00:15:30,800 --> 00:15:33,760
Speaker 2: the top twenty percent of the you know, ten percent

348
00:15:33,800 --> 00:15:38,400
Speaker 2: of the wealthy or already wealthy in America, So it's private.

349
00:15:38,520 --> 00:15:40,960
Speaker 2: If I were if I were Nick Bilton, I would

350
00:15:40,960 --> 00:15:44,280
Speaker 2: finish my thought by saying, so, yes, this is all

351
00:15:44,320 --> 00:15:46,520
Speaker 2: going to blow up, but it's just going to beat

352
00:15:46,600 --> 00:15:49,680
Speaker 2: up the billionaires and the rich people, which is the rich.

353
00:15:49,920 --> 00:15:52,280
Speaker 2: Thank you, the fat cats to go with my earlier

354
00:15:52,360 --> 00:15:55,040
Speaker 2: term of gussie, well, to go with the fat cats.

355
00:15:55,120 --> 00:15:58,280
Speaker 3: Let's bring out all the nineteen twenty so.

356
00:15:57,400 --> 00:15:59,640
Speaker 2: So it's just going to hurt the fat cats and

357
00:15:59,680 --> 00:16:03,280
Speaker 2: not mon pot kettle. Another good one.

358
00:16:03,320 --> 00:16:06,040
Speaker 1: Yeah, thanks, So he did it way better than I really, what.

359
00:16:06,120 --> 00:16:06,840
Speaker 3: Argument you're gonna make?

360
00:16:06,880 --> 00:16:09,880
Speaker 1: That was was going to thank you, Dick, thank you.

361
00:16:11,280 --> 00:16:12,840
Speaker 2: That wasn't Dick Costal though, by the way, that was

362
00:16:12,920 --> 00:16:13,440
Speaker 2: Nick Bilton.

363
00:16:13,480 --> 00:16:15,040
Speaker 1: Well would you not make that argument?

364
00:16:15,240 --> 00:16:18,400
Speaker 2: No? No, but why?

365
00:16:18,440 --> 00:16:20,720
Speaker 1: But what does it affect I don't understand. How does

366
00:16:20,760 --> 00:16:21,280
Speaker 1: it affect me?

367
00:16:21,400 --> 00:16:22,480
Speaker 3: Have you ever heard of this thing called the S

368
00:16:22,520 --> 00:16:23,280
Speaker 3: and P five hundred?

369
00:16:23,640 --> 00:16:24,680
Speaker 1: Yes, I've heard of the S and P.

370
00:16:25,120 --> 00:16:27,200
Speaker 3: Percentage of the S and P five hundred is mag

371
00:16:27,280 --> 00:16:32,120
Speaker 3: seven stocks. I don't meaning the largest AI textalks ten percent?

372
00:16:32,440 --> 00:16:33,760
Speaker 1: No, no, no, sorry, nine percent?

373
00:16:35,040 --> 00:16:35,880
Speaker 2: Did you say ninety?

374
00:16:36,080 --> 00:16:36,320
Speaker 1: Yeah?

375
00:16:36,640 --> 00:16:38,920
Speaker 2: No, do you get out of the house.

376
00:16:39,520 --> 00:16:42,120
Speaker 1: I'm not an economist. I'm just I'm ring together. I

377
00:16:42,240 --> 00:16:44,760
Speaker 1: read books, I do these things. I read Grendel by

378
00:16:44,760 --> 00:16:46,400
Speaker 1: the way, it was fantastic. But we'll come back to that.

379
00:16:46,880 --> 00:16:49,040
Speaker 2: Just give you a retort to why why is this

380
00:16:49,120 --> 00:16:52,440
Speaker 2: gonna Roughly, your dad thirty thirty five percent of the

381
00:16:52,480 --> 00:16:53,160
Speaker 2: S and P five.

382
00:16:53,160 --> 00:16:55,120
Speaker 1: I was in the middle when I said ten and ninety.

383
00:16:55,200 --> 00:16:57,120
Speaker 2: Yeah he was. He was narrowing.

384
00:16:57,160 --> 00:16:59,400
Speaker 3: He was mightily in. My head's in the oven, my

385
00:16:59,440 --> 00:17:00,640
Speaker 3: feet are in the on balance.

386
00:17:02,120 --> 00:17:03,120
Speaker 2: Yeah, yeah, no.

387
00:17:04,680 --> 00:17:09,000
Speaker 3: Anyways, so so roughly, what is gonna say next for

388
00:17:09,080 --> 00:17:12,119
Speaker 3: a single sector, Yes, to be that larger percentage of

389
00:17:12,160 --> 00:17:15,560
Speaker 3: the S and P five hundred is unprecedent. That doesn't happen, Okay, okay,

390
00:17:15,840 --> 00:17:18,120
Speaker 3: so at least in the last hundred years. So that's

391
00:17:18,160 --> 00:17:21,879
Speaker 3: incredibly important because it means that as these average jos

392
00:17:22,320 --> 00:17:24,680
Speaker 3: exposed to the market via passive index funde, which is

393
00:17:24,720 --> 00:17:27,320
Speaker 3: the largest chunk of the market now is passive funds,

394
00:17:27,359 --> 00:17:30,199
Speaker 3: and the largest piece of passive funds is index funds.

395
00:17:30,359 --> 00:17:33,439
Speaker 3: You're long AI. You're incredibly long. It's thirty percent of

396
00:17:33,440 --> 00:17:35,560
Speaker 3: your portfolio. If you're largely in passive fund.

397
00:17:35,440 --> 00:17:36,760
Speaker 1: Oh, I just invest in laser tag.

398
00:17:36,800 --> 00:17:40,320
Speaker 3: But yeah, yeah, So so the idea that somehow you

399
00:17:40,320 --> 00:17:43,760
Speaker 3: can dodge it, you're deeply exposed to this. You're incredibly

400
00:17:43,800 --> 00:17:45,440
Speaker 3: long AI by default.

401
00:17:45,400 --> 00:17:48,560
Speaker 2: Paul saying, everybody's iras that are just in the vanguard.

402
00:17:48,680 --> 00:17:53,040
Speaker 3: You know, of the gains in the last two years

403
00:17:53,119 --> 00:17:55,359
Speaker 3: were those seven stars. So it's not just that you

404
00:17:55,400 --> 00:17:58,360
Speaker 3: were exposed those unwined. Half your gains in the last

405
00:17:58,359 --> 00:18:00,880
Speaker 3: two years go away. Okay, so it's a double web.

406
00:18:00,920 --> 00:18:03,480
Speaker 3: So that's probably on the order of about eight hundred

407
00:18:03,520 --> 00:18:07,159
Speaker 3: billion dollars of capital be withdrawn that people think that

408
00:18:07,200 --> 00:18:09,919
Speaker 3: they're holding onto in these relatively safe conservative But that

409
00:18:10,000 --> 00:18:12,879
Speaker 3: doesn't stop there. It goes on like, for example, you

410
00:18:12,960 --> 00:18:14,840
Speaker 3: know exchange traded funds that are large holders of real

411
00:18:14,960 --> 00:18:18,879
Speaker 3: estate income trusts, which are essentially real estate funds. In

412
00:18:18,920 --> 00:18:22,520
Speaker 3: the ETF market, something like fifteen percent of the assets

413
00:18:22,560 --> 00:18:25,720
Speaker 3: under management at the largest ETFs are data centers.

414
00:18:26,560 --> 00:18:27,960
Speaker 2: Again another problem.

415
00:18:27,680 --> 00:18:30,600
Speaker 3: So who holds rates not people who are out there speculating,

416
00:18:30,600 --> 00:18:33,280
Speaker 3: and these are typically people who are retired, they want

417
00:18:33,359 --> 00:18:36,560
Speaker 3: income from some investment. They're some of the most conservative

418
00:18:36,600 --> 00:18:40,160
Speaker 3: investors in the market. They're incredibly long this stuff because

419
00:18:40,200 --> 00:18:43,240
Speaker 3: by default, these ETFs are increasingly holding a huge percentage

420
00:18:43,240 --> 00:18:45,760
Speaker 3: of data center So the idea that somehow this has

421
00:18:45,840 --> 00:18:49,280
Speaker 3: just to do with private equity is on the numbers ridiculous.

422
00:18:49,359 --> 00:18:51,920
Speaker 3: It's actually mostly retail investors who will take it on

423
00:18:52,000 --> 00:18:55,360
Speaker 3: the chin, and then secondarily it'll spiral to debt markets.

424
00:18:55,880 --> 00:18:58,399
Speaker 3: Within the debt markets in turn or collateral for other things,

425
00:18:58,400 --> 00:19:01,000
Speaker 3: so it'll contract spending, shrink the dock market, throw us

426
00:19:01,000 --> 00:19:03,240
Speaker 3: into a recession that's not dodgible.

427
00:19:27,680 --> 00:19:30,199
Speaker 1: Okay, So Dick's gonna ask the next question for me,

428
00:19:33,400 --> 00:19:38,639
Speaker 1: which is, Okay, here's the question. We've all talked about this,

429
00:19:38,680 --> 00:19:41,760
Speaker 1: we all know it. Everyone's saying it ten years, five years,

430
00:19:41,760 --> 00:19:44,440
Speaker 1: twenty years, whatever the fuck it is we are most

431
00:19:44,440 --> 00:19:47,400
Speaker 1: of us are going to be replaced by AI.

432
00:19:47,480 --> 00:19:47,560
Speaker 2: Oh.

433
00:19:47,560 --> 00:19:49,600
Speaker 3: I thought you're gonna make a different better We're gonna

434
00:19:49,720 --> 00:19:51,119
Speaker 3: I thought you're gonna make the argument that, well, it

435
00:19:51,119 --> 00:19:52,720
Speaker 3: really doesn't matter because in a couple of years, all

436
00:19:52,760 --> 00:19:55,400
Speaker 3: this stuff, no matter how much it all gets wasted,

437
00:19:55,560 --> 00:19:57,000
Speaker 3: some of it will turn out to be useful, like

438
00:19:57,000 --> 00:19:58,960
Speaker 3: the dot Com cycle. So even if it turns out

439
00:19:58,960 --> 00:20:01,720
Speaker 3: that all these data centers we've overbuilt them, it's pointless.

440
00:20:01,760 --> 00:20:04,040
Speaker 2: Yes, that's a big argument right now, and that's an

441
00:20:04,040 --> 00:20:07,280
Speaker 2: emerging argument. Yes, there will it will be overbuilt, and

442
00:20:07,359 --> 00:20:10,160
Speaker 2: yes we'll but what will come out the back end

443
00:20:10,200 --> 00:20:13,520
Speaker 2: will be all these amazing you know, technologies that improve

444
00:20:13,600 --> 00:20:14,000
Speaker 2: our lives.

445
00:20:14,080 --> 00:20:16,399
Speaker 3: Right, it's the dot Com example. You know, all that

446
00:20:16,400 --> 00:20:18,800
Speaker 3: stuff didn't work out then, but we laid the groundwork

447
00:20:18,880 --> 00:20:20,760
Speaker 3: with fiber and all these other kinds of things. It's

448
00:20:20,800 --> 00:20:23,639
Speaker 3: that argument, right, which is a really popular response to this.

449
00:20:23,680 --> 00:20:24,760
Speaker 2: It really doesn't matter that.

450
00:20:24,720 --> 00:20:26,200
Speaker 1: I wasn't making that argument, but go for it.

451
00:20:26,800 --> 00:20:29,520
Speaker 2: But it's a very it's a very sort of a

452
00:20:29,720 --> 00:20:32,800
Speaker 2: current argument. Do you like how I've gone from old

453
00:20:32,840 --> 00:20:36,480
Speaker 2: West to say yeah?

454
00:20:36,520 --> 00:20:39,520
Speaker 3: So that argument, though, is like a very common response

455
00:20:39,560 --> 00:20:42,199
Speaker 3: from the tech the technirati that this is you just

456
00:20:42,240 --> 00:20:46,000
Speaker 3: don't understand the dynamics of how technologies work in these

457
00:20:46,040 --> 00:20:46,879
Speaker 3: bubble cycles and.

458
00:20:47,960 --> 00:20:51,120
Speaker 2: Everything has a certain response to that.

459
00:20:52,160 --> 00:20:53,199
Speaker 1: Do you agree with all of this?

460
00:20:54,000 --> 00:20:56,840
Speaker 2: He's yes, I do agree with everything, saying no, I mean,

461
00:20:56,880 --> 00:20:58,840
Speaker 2: I mean sure there was some one or two things

462
00:20:58,840 --> 00:21:00,919
Speaker 2: in there, I disagreed with, but I generally agree with

463
00:21:00,960 --> 00:21:01,640
Speaker 2: everything you're say.

464
00:21:03,160 --> 00:21:07,400
Speaker 3: Okay, all right, Well, so the response to that point, though,

465
00:21:07,440 --> 00:21:10,800
Speaker 3: the ridiculous point about this the cycles is one you're

466
00:21:10,840 --> 00:21:12,879
Speaker 3: operating from a data stet of about three two or

467
00:21:12,880 --> 00:21:16,320
Speaker 3: three historical examples, possibly the dot Com period. Yeah, maybe

468
00:21:16,359 --> 00:21:18,520
Speaker 3: you can go back to like the PCs. Maybe it

469
00:21:18,560 --> 00:21:22,440
Speaker 3: wasn't nearly as big. Possibly rural electrification in the nineteen twenties,

470
00:21:22,760 --> 00:21:24,560
Speaker 3: maybe railroads in the mid nineteen SEO.

471
00:21:24,600 --> 00:21:26,680
Speaker 1: We're going to go to rural electrication in nineteen That's.

472
00:21:26,560 --> 00:21:27,520
Speaker 2: All you got, my friend.

473
00:21:27,720 --> 00:21:29,880
Speaker 3: So I was to launch a drug on that basis,

474
00:21:29,960 --> 00:21:32,280
Speaker 3: I wouldn't get approved. So the notion that that should

475
00:21:32,280 --> 00:21:34,920
Speaker 3: somehow convince me that this is some kind of law

476
00:21:35,000 --> 00:21:38,360
Speaker 3: of economic life for starters is just ridiculous on the face, right,

477
00:21:38,359 --> 00:21:41,680
Speaker 3: because we have a couple of anecdotes we like to tell,

478
00:21:41,720 --> 00:21:43,200
Speaker 3: and then that's all we have. Because even if you

479
00:21:43,280 --> 00:21:45,120
Speaker 3: take it back to the Industrial Revolution and say, well,

480
00:21:45,119 --> 00:21:47,800
Speaker 3: you know it all worked out, it took forty years

481
00:21:48,200 --> 00:21:50,040
Speaker 3: for the average person to get back to the same

482
00:21:50,080 --> 00:21:52,560
Speaker 3: standard of living as they had before the industrial society

483
00:21:52,600 --> 00:21:55,720
Speaker 3: got wealthier. Most people got poor for forty years. Try

484
00:21:55,760 --> 00:21:57,560
Speaker 3: that in the United States today and see how that goes.

485
00:21:57,680 --> 00:21:59,399
Speaker 1: And that's going to be in two years, not forty.

486
00:21:59,800 --> 00:22:04,400
Speaker 3: All get six months faster, more widespread, more consequential.

487
00:22:03,840 --> 00:22:05,840
Speaker 2: And what's your sense about quickly this starts. I mean,

488
00:22:05,880 --> 00:22:08,000
Speaker 2: it's kind of starting to on one right now, people

489
00:22:08,160 --> 00:22:09,080
Speaker 2: realize I think.

490
00:22:09,600 --> 00:22:12,520
Speaker 3: I always remind like I remind myself that back in

491
00:22:12,560 --> 00:22:15,320
Speaker 3: the eighties or nineties, whenever the dot com thing come

492
00:22:15,400 --> 00:22:17,720
Speaker 3: was beginning to blow up. So six four years before

493
00:22:17,760 --> 00:22:22,119
Speaker 3: it blew up, greasepads out there warning people about irrational exuberance.

494
00:22:22,320 --> 00:22:24,600
Speaker 3: So it was four years later people forget he was

495
00:22:24,640 --> 00:22:26,560
Speaker 3: warning people and saying, oh, you know, this is out

496
00:22:26,600 --> 00:22:28,560
Speaker 3: of control. I actually think it's going to saw tooth

497
00:22:28,600 --> 00:22:30,960
Speaker 3: considerably higher before it blows up. And at each stage

498
00:22:31,000 --> 00:22:33,080
Speaker 3: people will think, well that's the end, and it won't

499
00:22:33,080 --> 00:22:34,720
Speaker 3: be the end, and it'll be the end, you know,

500
00:22:34,800 --> 00:22:37,919
Speaker 3: so probably, you know, look out two to four years

501
00:22:37,960 --> 00:22:40,920
Speaker 3: is a good guess, but it'll saw tooth higher some

502
00:22:41,400 --> 00:22:43,440
Speaker 3: you know, catching people at each step all the way.

503
00:22:43,520 --> 00:22:45,200
Speaker 2: The very end of these things tends to be the

504
00:22:45,560 --> 00:22:48,199
Speaker 2: most juice. That's sort of the last six months, if

505
00:22:48,240 --> 00:22:49,800
Speaker 2: they will, it doubles again It's like.

506
00:22:49,760 --> 00:22:52,119
Speaker 3: The drunken Miller moment, the famously stand drunk and Miller

507
00:22:52,280 --> 00:22:55,159
Speaker 3: well on hedge fund manager was short this stuff, the

508
00:22:55,200 --> 00:22:57,639
Speaker 3: dot com stuff, all the way up, and then about

509
00:22:57,680 --> 00:22:59,919
Speaker 3: six weeks before at the top hit he.

510
00:23:00,119 --> 00:23:03,119
Speaker 2: Went like February twenty, February two.

511
00:23:02,760 --> 00:23:05,239
Speaker 3: Thousands, two thousand, he sold everything else, bought it all

512
00:23:05,240 --> 00:23:07,399
Speaker 3: and immediately blew up and the whole thing ended. And

513
00:23:07,440 --> 00:23:09,159
Speaker 3: he's a very smart guy. But the point is is

514
00:23:09,200 --> 00:23:11,840
Speaker 3: that you reach this point of maximum pain where people say,

515
00:23:11,840 --> 00:23:13,520
Speaker 3: you know what, this is all ridiculous, but I can't

516
00:23:13,520 --> 00:23:15,760
Speaker 3: take being on the other side of this anymore. I

517
00:23:15,800 --> 00:23:18,480
Speaker 3: need to just go all in whatever I need to

518
00:23:18,480 --> 00:23:20,840
Speaker 3: do in terms of lending or anything else. And that's

519
00:23:20,880 --> 00:23:23,639
Speaker 3: the moment when things go, you know, catastrophically badly.

520
00:23:24,320 --> 00:23:25,360
Speaker 2: I mean, we're not there yet.

521
00:23:25,480 --> 00:23:28,080
Speaker 1: Let me ask you a logistical question. How many of

522
00:23:28,160 --> 00:23:31,959
Speaker 1: all the data centers, how many are like an AWS

523
00:23:32,040 --> 00:23:36,639
Speaker 1: type situation where people are renting space, GPUs and so

524
00:23:36,680 --> 00:23:38,960
Speaker 1: on and so forth, versus Opening Eye and Facebook, and

525
00:23:39,000 --> 00:23:40,320
Speaker 1: those guys building their own.

526
00:23:40,520 --> 00:23:42,119
Speaker 3: Even when they build their own, they don't build their

527
00:23:42,119 --> 00:23:44,320
Speaker 3: own obviously, right, I mean, these guys are not in

528
00:23:44,359 --> 00:23:48,040
Speaker 3: the construction business. So they're doing this in conjunction with

529
00:23:48,119 --> 00:23:51,159
Speaker 3: construction companies and special in the context of these special

530
00:23:51,520 --> 00:23:54,440
Speaker 3: purpose vehicles and others like blue out and private credit

531
00:23:54,520 --> 00:23:56,040
Speaker 3: and what have you. So the tend to still have

532
00:23:56,560 --> 00:23:58,920
Speaker 3: there's operators and sponsors in the middle of it. They

533
00:23:59,040 --> 00:24:02,320
Speaker 3: just become the primary lease ore of the data center.

534
00:24:02,760 --> 00:24:05,359
Speaker 1: But given that, what are your what's your blood pressure

535
00:24:05,359 --> 00:24:05,720
Speaker 1: at right now?

536
00:24:05,760 --> 00:24:05,960
Speaker 2: Did it?

537
00:24:06,000 --> 00:24:07,280
Speaker 1: Is it up to like one sixty over?

538
00:24:07,440 --> 00:24:11,280
Speaker 2: Actually I find it soon and he's getting it off

539
00:24:11,280 --> 00:24:13,280
Speaker 2: his chessfest so much.

540
00:24:14,359 --> 00:24:16,480
Speaker 1: Let me do this little therapy session.

541
00:24:16,600 --> 00:24:17,200
Speaker 2: I'm all good.

542
00:24:17,400 --> 00:24:21,399
Speaker 1: No, but some of these guys, so we're all going

543
00:24:21,480 --> 00:24:23,720
Speaker 1: to be replaced by AI. This was gonna be my question, right,

544
00:24:24,119 --> 00:24:26,240
Speaker 1: everyone's saying this. It's going to happen. I I know

545
00:24:26,359 --> 00:24:28,040
Speaker 1: for a fact that I'm going to be right when.

546
00:24:27,920 --> 00:24:29,600
Speaker 2: You say everyone's saying this. Do you meet it in

547
00:24:29,640 --> 00:24:30,320
Speaker 2: the Donald Trump?

548
00:24:30,440 --> 00:24:35,000
Speaker 1: Everybody says we or you Actually the other day No,

549
00:24:35,080 --> 00:24:38,639
Speaker 1: but you're hearing it's it's I mean, it's not it's

550
00:24:38,720 --> 00:24:42,359
Speaker 1: so obvious. Yes, it may take a while, maybe it's

551
00:24:42,560 --> 00:24:45,760
Speaker 1: a year, maybe it's twenty, but we are going to

552
00:24:46,040 --> 00:24:48,920
Speaker 1: most people in society are going to be replaced by AI.

553
00:24:49,520 --> 00:24:52,120
Speaker 1: If that is the let's just hypothetically for one minute,

554
00:24:52,160 --> 00:24:53,200
Speaker 1: all agree that's going to happen.

555
00:24:53,200 --> 00:24:54,879
Speaker 3: And to see where this is going. What does this

556
00:24:54,920 --> 00:24:56,680
Speaker 3: have to do with the current financial problem?

557
00:24:56,800 --> 00:24:59,959
Speaker 1: You are going to need these data centers. You're not

558
00:25:00,160 --> 00:25:01,960
Speaker 1: to need data centers when AI runs the world and

559
00:25:02,040 --> 00:25:06,520
Speaker 1: humans are just taking their butts watching AI created.

560
00:25:06,280 --> 00:25:08,959
Speaker 3: Slop again back for I was saying earlier, the current

561
00:25:09,560 --> 00:25:13,040
Speaker 3: buildout of data centers is predicated on a terrible backward

562
00:25:13,080 --> 00:25:16,200
Speaker 3: looking idea, which is that right now, sixty to seventy

563
00:25:16,200 --> 00:25:19,879
Speaker 3: percent of data center usage is not for making my

564
00:25:19,960 --> 00:25:21,800
Speaker 3: fake nick do fake nick things.

565
00:25:22,000 --> 00:25:23,480
Speaker 2: It's for training.

566
00:25:23,200 --> 00:25:24,040
Speaker 1: Coslos for that.

567
00:25:24,160 --> 00:25:26,600
Speaker 3: Yeah, yeah, that's true. We have here here for that.

568
00:25:26,840 --> 00:25:29,640
Speaker 3: It's for training a new AI's training new frontier models.

569
00:25:29,880 --> 00:25:31,920
Speaker 3: If we're not training a new frontier models. And we

570
00:25:31,960 --> 00:25:33,600
Speaker 3: can come back to why that's not going to happen

571
00:25:34,880 --> 00:25:37,280
Speaker 3: if you're not training new frontier models. We arguably have

572
00:25:37,400 --> 00:25:40,879
Speaker 3: enough capacity now in data centers to take us to

573
00:25:40,920 --> 00:25:44,400
Speaker 3: handle inference through about twenty thirty two. Like we're already

574
00:25:44,440 --> 00:25:49,320
Speaker 3: overbuilt for doing fake nicks. The whole model is built

575
00:25:49,320 --> 00:25:51,040
Speaker 3: on a broken assumption, which is that we're going to

576
00:25:51,040 --> 00:25:54,960
Speaker 3: continue training frontier models for the next decade, and we're

577
00:25:54,960 --> 00:25:59,000
Speaker 3: already at a point where the deltas between models are collapsing.

578
00:25:59,160 --> 00:26:01,879
Speaker 3: The models from different vendors are all converging. I defy

579
00:26:01,960 --> 00:26:04,280
Speaker 3: you to tell the difference between like Anthropic and Claw

580
00:26:04,440 --> 00:26:07,760
Speaker 3: or Cloud and open TPT or whatever. They're all the same.

581
00:26:07,800 --> 00:26:09,120
Speaker 3: They all feel exactly the same.

582
00:26:08,920 --> 00:26:13,240
Speaker 2: For general purpose frontier models. You can even see in

583
00:26:13,280 --> 00:26:17,240
Speaker 2: the data from these there are companies that are like, hey,

584
00:26:17,480 --> 00:26:19,879
Speaker 2: plug into us, and then we'll route your request to

585
00:26:19,920 --> 00:26:24,119
Speaker 2: whatever the cheap open router yes, for example, Yeah, plug

586
00:26:24,119 --> 00:26:26,600
Speaker 2: in your request US, and we'll route the request to

587
00:26:26,640 --> 00:26:30,200
Speaker 2: the which whatever each expensive provider is right now, and

588
00:26:30,840 --> 00:26:33,760
Speaker 2: people don't care. What you quickly see from looking at

589
00:26:33,760 --> 00:26:36,640
Speaker 2: a month to month graph of open router API API

590
00:26:36,800 --> 00:26:39,600
Speaker 2: calls is that for general purpose requests, nobody cares what's

591
00:26:39,720 --> 00:26:42,880
Speaker 2: what it's which frontier models being routed to, which tells

592
00:26:42,920 --> 00:26:46,760
Speaker 2: you that there and in fact, nobody owns more than

593
00:26:46,840 --> 00:26:48,960
Speaker 2: like x percent of that open router share, And it's

594
00:26:49,000 --> 00:26:51,280
Speaker 2: very fluid, and it's very fluid changes, so you can

595
00:26:51,359 --> 00:26:54,480
Speaker 2: just sort of see in the graph that there's no

596
00:26:55,359 --> 00:26:57,840
Speaker 2: I have to route it to this one right now.

597
00:26:58,280 --> 00:27:00,480
Speaker 2: Because it's so far ahead of these other ones. Yeah,

598
00:27:00,520 --> 00:27:03,960
Speaker 2: and these things can change constantly, and saying which shows you,

599
00:27:04,040 --> 00:27:07,359
Speaker 2: I mean it's basically it's a commoditization ground. It's a

600
00:27:07,400 --> 00:27:09,000
Speaker 2: graph that highlights commoditization.

601
00:27:09,160 --> 00:27:11,159
Speaker 3: Totally shows it now. And I mean in saying that

602
00:27:11,320 --> 00:27:12,919
Speaker 3: you know we're going to automate you out of existence

603
00:27:12,960 --> 00:27:13,680
Speaker 3: or whatever we're going.

604
00:27:13,560 --> 00:27:14,760
Speaker 1: To do to you, it's going to happen.

605
00:27:15,119 --> 00:27:16,800
Speaker 2: I mean, we are going to automate you out of

606
00:27:17,720 --> 00:27:19,960
Speaker 2: a specifically as soon as par see.

607
00:27:20,000 --> 00:27:21,720
Speaker 3: One of the mistakes people make, I think, in talking

608
00:27:21,760 --> 00:27:24,000
Speaker 3: about this stuff is they say, if you say there's

609
00:27:24,040 --> 00:27:27,119
Speaker 3: an AI capex bubble, it's somehow that becomes turned into

610
00:27:27,320 --> 00:27:30,440
Speaker 3: you think AI is nonsense. I think the exact opposite.

611
00:27:30,480 --> 00:27:33,320
Speaker 3: I think is an incredibly important technology. It's like saying

612
00:27:33,800 --> 00:27:35,919
Speaker 3: I thought there was a global financial crisis about to

613
00:27:35,920 --> 00:27:38,320
Speaker 3: happen back in two thousand and seven, two thousand and eight. Oh,

614
00:27:38,359 --> 00:27:41,320
Speaker 3: you're against houses if I'm not against houses.

615
00:27:42,760 --> 00:27:46,720
Speaker 2: Specifically, so you're against like retail investors getting in on

616
00:27:46,760 --> 00:27:49,200
Speaker 2: the housing or even people owning houses.

617
00:27:49,640 --> 00:27:52,640
Speaker 3: Somehow I'm dismissing the whole technology of houses.

618
00:27:52,760 --> 00:27:52,920
Speaker 2: Right.

619
00:27:52,960 --> 00:27:55,200
Speaker 3: Oh, let's go back to lean to's or something like this,

620
00:27:55,720 --> 00:27:56,320
Speaker 3: and so this.

621
00:27:56,320 --> 00:27:56,960
Speaker 2: Is the same for me.

622
00:27:57,000 --> 00:27:59,120
Speaker 3: I find it a ridiculous argument that if you think

623
00:27:59,160 --> 00:28:01,679
Speaker 3: that there's a CAPEX bubble, that must mean that somehow

624
00:28:01,760 --> 00:28:04,320
Speaker 3: AI therefore isn't going to do all of this damage

625
00:28:04,359 --> 00:28:06,080
Speaker 3: in the economy. No, it's going to do men's damage

626
00:28:06,080 --> 00:28:08,680
Speaker 3: in the economy. But that's a different point. Right, We're

627
00:28:08,720 --> 00:28:10,480
Speaker 3: going to have two things happen at once. We're going

628
00:28:10,560 --> 00:28:13,840
Speaker 3: to have this incredible overbuilding of AI fueled by debt,

629
00:28:13,840 --> 00:28:16,159
Speaker 3: which will all collapse, which will have consequences for this

630
00:28:16,200 --> 00:28:18,800
Speaker 3: stock market, for the economy. And then we'll also, at

631
00:28:18,800 --> 00:28:20,920
Speaker 3: the same time, at the other end of this process,

632
00:28:21,119 --> 00:28:23,640
Speaker 3: we'll still have data centers to putting out cheap tokens,

633
00:28:23,680 --> 00:28:26,640
Speaker 3: probably even cheaper, because they're all going to be excess capacity.

634
00:28:26,680 --> 00:28:28,640
Speaker 3: It's going to be like you know, opeck just put

635
00:28:28,640 --> 00:28:30,840
Speaker 3: pumping the stuff out, and that's going to make it

636
00:28:30,840 --> 00:28:31,679
Speaker 3: easier automating it.

637
00:28:32,119 --> 00:28:35,720
Speaker 1: Okay, So let's just say that you're correct with everything, right,

638
00:28:35,840 --> 00:28:38,400
Speaker 1: Let's just say I'm correct. Let's just say that you're right.

639
00:28:39,000 --> 00:28:39,360
Speaker 2: Okay.

640
00:28:39,600 --> 00:28:44,200
Speaker 1: The question I have is for us to reach this.

641
00:28:44,280 --> 00:28:45,960
Speaker 3: If you say, what's the trade, I'm going to turn

642
00:28:45,960 --> 00:28:47,080
Speaker 3: off my microphone.

643
00:28:48,520 --> 00:28:53,240
Speaker 1: That's not me. The question I have is for this,

644
00:28:54,200 --> 00:28:56,600
Speaker 1: so we're going to get to this point where I'm replaced. Okay,

645
00:28:56,960 --> 00:28:58,200
Speaker 1: let's just pretend that that's true.

646
00:28:58,280 --> 00:29:00,800
Speaker 3: Okay, just to be clear, this is a sum narcissistic

647
00:29:00,920 --> 00:29:01,520
Speaker 3: view of the world.

648
00:29:01,680 --> 00:29:05,560
Speaker 1: No, it's not. I'm just saying. I'm saying most people,

649
00:29:06,080 --> 00:29:09,560
Speaker 1: me being one of them. Okay, most jobs. Let's can

650
00:29:09,600 --> 00:29:11,640
Speaker 1: we can we finally agree on one thing on this

651
00:29:11,680 --> 00:29:18,760
Speaker 1: podcast that a lot of jobs are going to vanish? Yes, okay, great, Okay,

652
00:29:18,840 --> 00:29:22,280
Speaker 1: so if that happens, I literally forgot my question.

653
00:29:22,600 --> 00:29:23,760
Speaker 3: Yes, it probably wasn't.

654
00:29:24,920 --> 00:29:29,360
Speaker 1: Okay, if that happens, so we have to the models

655
00:29:29,360 --> 00:29:31,840
Speaker 1: have to get better for that to happen, right, They

656
00:29:31,880 --> 00:29:34,600
Speaker 1: don't know because the models, they feel like they're kind

657
00:29:34,640 --> 00:29:36,240
Speaker 1: of going back. Was you said this the other day?

658
00:29:36,360 --> 00:29:37,280
Speaker 1: Like you said the check.

659
00:29:37,160 --> 00:29:39,360
Speaker 3: But that's a different problem. But half of people are

660
00:29:39,400 --> 00:29:41,000
Speaker 3: dumber than average.

661
00:29:41,080 --> 00:29:43,720
Speaker 1: Okay, but the models still have to the models still

662
00:29:43,720 --> 00:29:49,440
Speaker 1: have to get better to replace most people. So the

663
00:29:49,560 --> 00:29:53,240
Speaker 1: question I have, that's okay, but let's just pang with

664
00:29:53,280 --> 00:29:57,320
Speaker 1: me for one second. The question I have is what

665
00:29:57,320 --> 00:30:00,760
Speaker 1: what needs to happen with the technology to get there,

666
00:30:01,280 --> 00:30:05,480
Speaker 1: and don't the data centers need to exist to get

667
00:30:05,560 --> 00:30:06,360
Speaker 1: us to that point.

668
00:30:06,560 --> 00:30:08,640
Speaker 3: No, it's back to the same argument, which is to

669
00:30:08,640 --> 00:30:11,640
Speaker 3: say that, yes, if we continue doing what we're doing

670
00:30:11,680 --> 00:30:14,840
Speaker 3: pre twenty twenty two in terms of these large data

671
00:30:14,880 --> 00:30:18,160
Speaker 3: sets and training in multiple cycles, these incredibly huge, these

672
00:30:18,200 --> 00:30:20,920
Speaker 3: huge frontier models based on you know, exabytes of data,

673
00:30:20,960 --> 00:30:24,760
Speaker 3: blah blah blah blah. Absolutely, but again that's like saying,

674
00:30:24,800 --> 00:30:27,120
Speaker 3: you know, well, there's only room for like, you know,

675
00:30:27,200 --> 00:30:29,560
Speaker 3: four computers in the world. Was the IBM argument in

676
00:30:29,560 --> 00:30:31,040
Speaker 3: the fifties, because you know, these things are going to

677
00:30:31,040 --> 00:30:32,320
Speaker 3: be so big and take.

678
00:30:32,240 --> 00:30:32,960
Speaker 2: So much power.

679
00:30:33,000 --> 00:30:35,600
Speaker 3: It just doesn't That's an extrapolative way of thinking about

680
00:30:35,600 --> 00:30:38,160
Speaker 3: what's ahead based on this really brief period where we

681
00:30:38,200 --> 00:30:40,440
Speaker 3: had all this free data to train on and we've

682
00:30:40,440 --> 00:30:42,400
Speaker 3: already exhausted it. I mean, I can make it just

683
00:30:42,440 --> 00:30:44,000
Speaker 3: like the Saudi Arabia of data and it's gone.

684
00:30:44,120 --> 00:30:47,480
Speaker 2: Yeah. What Paul is saying is they are these they're

685
00:30:47,520 --> 00:30:52,600
Speaker 2: the diminishing returns on training now based on we've we've

686
00:30:52,840 --> 00:30:56,120
Speaker 2: consumed the web, the models have consumed the web, and

687
00:30:56,160 --> 00:30:58,720
Speaker 2: the diminishing returns on training data and the amount of

688
00:30:58,720 --> 00:31:03,520
Speaker 2: money the incredibly more vast amount of money you have

689
00:31:03,600 --> 00:31:08,560
Speaker 2: to pour in for the incremental and bit of training.

690
00:31:08,880 --> 00:31:11,560
Speaker 3: Is again to get only marginal gains.

691
00:31:11,360 --> 00:31:13,280
Speaker 2: To get only marginal little gains in frontier in the

692
00:31:13,280 --> 00:31:13,960
Speaker 2: front term models.

693
00:31:14,000 --> 00:31:15,800
Speaker 3: It's on so so to take it the one x

694
00:31:15,840 --> 00:31:18,400
Speaker 3: step further so the notion that therefore I need even

695
00:31:18,440 --> 00:31:20,760
Speaker 3: more data centers because to make these things even more smart,

696
00:31:20,880 --> 00:31:25,040
Speaker 3: even smarter. No, we've basically we've we're frozen in place.

697
00:31:25,080 --> 00:31:26,840
Speaker 3: We're gonna kind of like go, you know, what's the

698
00:31:26,880 --> 00:31:31,160
Speaker 3: French term, cul de sac. We're in a dead end

699
00:31:31,240 --> 00:31:33,520
Speaker 3: with respect of this stuff in terms of progress right

700
00:31:33,560 --> 00:31:36,640
Speaker 3: now because of the architectural limitations of large language models.

701
00:31:37,600 --> 00:31:38,640
Speaker 2: That's incredibly important.

702
00:31:38,640 --> 00:31:41,720
Speaker 3: The architectural limitations of large language models, in conjunction with

703
00:31:41,760 --> 00:31:45,280
Speaker 3: the the end of the Saudi Arabia of data leaves

704
00:31:45,320 --> 00:31:48,880
Speaker 3: us more or less where we are very expensively, you.

705
00:31:48,840 --> 00:31:49,760
Speaker 2: Know, and so so.

706
00:31:51,160 --> 00:31:54,320
Speaker 3: Things like no, it's but if you think about like

707
00:31:54,360 --> 00:31:56,280
Speaker 3: this whole you know, what was the outrage at the

708
00:31:56,360 --> 00:31:58,480
Speaker 3: end of the summer about like sick Offencian models, the

709
00:31:58,480 --> 00:31:59,920
Speaker 3: idea that they kiss asshole the time.

710
00:32:00,200 --> 00:32:02,200
Speaker 2: Yeah, people treat it as this really.

711
00:32:02,040 --> 00:32:04,280
Speaker 3: Superficial thing about oh you know, they're just trying to

712
00:32:04,520 --> 00:32:07,640
Speaker 3: create more engagement. Now, that was totally predictable from what's

713
00:32:07,640 --> 00:32:10,120
Speaker 3: happening right now. And why is it totally predictable Because

714
00:32:10,560 --> 00:32:12,840
Speaker 3: if you can't get gains from what's called pre training,

715
00:32:13,000 --> 00:32:15,080
Speaker 3: which is what happens whenever you bring in new data

716
00:32:15,200 --> 00:32:17,320
Speaker 3: and you get smarter and smarter models, what do you

717
00:32:17,360 --> 00:32:19,880
Speaker 3: do to improve your model? You do what's called post training,

718
00:32:19,880 --> 00:32:22,120
Speaker 3: which happens after the frontier model has been created. You

719
00:32:22,160 --> 00:32:25,400
Speaker 3: do these reinforcement learning and mixture of expert stuff. But

720
00:32:25,440 --> 00:32:27,440
Speaker 3: what does that lead to? Well, that leads to people

721
00:32:27,480 --> 00:32:29,400
Speaker 3: giving you feedback on your model and say, oh, I

722
00:32:29,440 --> 00:32:31,560
Speaker 3: like when it does that, I evaluated more. So you

723
00:32:31,640 --> 00:32:34,840
Speaker 3: reward the models for doing things that make me happy. Well,

724
00:32:34,840 --> 00:32:38,200
Speaker 3: guess what. That leads to sycophancy. So sycophancy is a

725
00:32:38,280 --> 00:32:42,080
Speaker 3: natural outcome from an exhausted data set in models that

726
00:32:42,120 --> 00:32:44,400
Speaker 3: are no longer improving. It's totally predictable. So you just

727
00:32:44,440 --> 00:32:45,800
Speaker 3: get more and more sick of fans, and then you

728
00:32:45,840 --> 00:32:47,640
Speaker 3: can even go off the cliff, which is what we're

729
00:32:47,640 --> 00:32:50,720
Speaker 3: seeing now. Models actually be starting to become worse because

730
00:32:50,720 --> 00:32:52,920
Speaker 3: there's too much post training, and it's suddenly it's like,

731
00:32:52,960 --> 00:32:55,000
Speaker 3: wait a minute, you're dumber than you were six months ago.

732
00:32:56,400 --> 00:32:57,080
Speaker 2: That's why.

733
00:32:57,320 --> 00:32:59,920
Speaker 3: But people, you got to go on. If you go underneath,

734
00:33:00,120 --> 00:33:02,520
Speaker 3: this is what's happening architecturally, that's what you're seeing. What

735
00:33:02,520 --> 00:33:04,640
Speaker 3: you're really seeing is that if only these guys were

736
00:33:04,640 --> 00:33:07,320
Speaker 3: good enough to actually make the models into trusick. That's

737
00:33:07,360 --> 00:33:10,200
Speaker 3: not what's happening. It's this attempt to use post training

738
00:33:10,200 --> 00:33:11,120
Speaker 3: to improve their models.

739
00:33:11,200 --> 00:33:12,160
Speaker 2: So the model.

740
00:33:11,920 --> 00:33:14,040
Speaker 1: Is choosing to become a sick event.

741
00:33:14,440 --> 00:33:18,880
Speaker 3: No, it's with reinforcement learning. You take you give people

742
00:33:18,920 --> 00:33:23,280
Speaker 3: the opportunity to respond and then incorporate the reward function

743
00:33:23,440 --> 00:33:25,440
Speaker 3: in terms of how oh I like that, I don't

744
00:33:25,480 --> 00:33:27,960
Speaker 3: like that, and that in turn becomes the launch product.

745
00:33:28,280 --> 00:33:31,880
Speaker 3: But it post training inherently has this problem of leading

746
00:33:31,880 --> 00:33:35,200
Speaker 3: to these what we've been calling the sycophantic models, but

747
00:33:35,200 --> 00:33:38,040
Speaker 3: more importantly, just leads to a performance collapse because there's

748
00:33:38,040 --> 00:33:40,560
Speaker 3: no new data. You're just over optimizing and I think

749
00:33:40,760 --> 00:33:43,440
Speaker 3: the thing that makes users happiest, which isn't the same thing.

750
00:33:43,520 --> 00:33:45,920
Speaker 2: At what point does someone like Saudi Arabia come in

751
00:33:45,960 --> 00:33:49,720
Speaker 2: and go, hey, great news, everybody, we have super cheap

752
00:33:50,400 --> 00:33:54,600
Speaker 2: inference compute here for you. With the tank The tankers

753
00:33:54,600 --> 00:33:58,080
Speaker 2: are hooked right up to the data center. We use ours.

754
00:33:58,160 --> 00:34:00,640
Speaker 2: But that's that's already happening. It's already happened.

755
00:34:00,640 --> 00:34:03,520
Speaker 3: It's really interesting when you talk to people on that

756
00:34:03,600 --> 00:34:05,680
Speaker 3: side of things. They're like, here we are, we've got

757
00:34:05,680 --> 00:34:08,400
Speaker 3: all this energy. We've got the energy leaving aside the cooling.

758
00:34:08,440 --> 00:34:10,000
Speaker 3: We can fuel it though, we'll make it all happen.

759
00:34:10,040 --> 00:34:12,959
Speaker 3: That's already happening. Like they would cheerfully absorb a huge

760
00:34:13,000 --> 00:34:15,600
Speaker 3: fraction of the workloads we're seeing here. And so you know,

761
00:34:15,680 --> 00:34:18,359
Speaker 3: my expectation is, you know, a huge fraction of what's

762
00:34:18,360 --> 00:34:19,279
Speaker 3: going on in the US will.

763
00:34:19,200 --> 00:34:19,880
Speaker 2: All move offshore.

764
00:34:20,360 --> 00:34:22,680
Speaker 3: It's going to move to it's gonna move to places

765
00:34:22,719 --> 00:34:24,160
Speaker 3: like China, and it's going to move to the Middle

766
00:34:24,160 --> 00:34:24,880
Speaker 3: East places.

767
00:34:25,280 --> 00:34:27,680
Speaker 1: I saw someone talking about building it in space. Thanks.

768
00:34:27,760 --> 00:34:29,920
Speaker 3: Yeah, yeah, yeah sure.

769
00:34:30,080 --> 00:34:34,000
Speaker 1: Sure. People have their questions, is that you want to

770
00:34:34,000 --> 00:34:37,400
Speaker 1: move on to duck ragu or is working on a

771
00:34:37,440 --> 00:34:38,760
Speaker 1: really great ducks.

772
00:34:39,719 --> 00:34:41,359
Speaker 3: I just think the thing that people have to keep

773
00:34:41,360 --> 00:34:43,800
Speaker 3: in mind is this idea that it's not just another

774
00:34:43,840 --> 00:34:45,960
Speaker 3: tech bubble, right, It's really not.

775
00:34:46,160 --> 00:34:46,960
Speaker 2: It's the thing that's.

776
00:34:46,840 --> 00:34:50,200
Speaker 3: Super unusual is that you have all the pieces of everything,

777
00:34:50,600 --> 00:34:52,160
Speaker 3: all four combined on one place.

778
00:34:52,239 --> 00:34:53,560
Speaker 1: Have we ever had that before in history?

779
00:34:53,719 --> 00:34:53,879
Speaker 2: Never?

780
00:34:54,040 --> 00:34:54,239
Speaker 1: Never?

781
00:34:54,480 --> 00:34:54,719
Speaker 2: Never.

782
00:34:55,320 --> 00:34:57,080
Speaker 3: That's the thing that's really interesting about this one. And

783
00:34:57,120 --> 00:34:58,799
Speaker 3: so the idea that it's like, oh, you know, I

784
00:34:58,840 --> 00:35:00,959
Speaker 3: know how this plays out, that's like, no, you don't.

785
00:35:01,040 --> 00:35:03,200
Speaker 1: But we've never had it before in history. But we've

786
00:35:03,239 --> 00:35:08,319
Speaker 1: also never had a moment in time where humans are

787
00:35:08,320 --> 00:35:09,640
Speaker 1: going to be replaced in the way that they.

788
00:35:09,600 --> 00:35:10,719
Speaker 2: Will, yes, just make things worse.

789
00:35:10,760 --> 00:35:14,440
Speaker 3: Though, that's just like an accelerant, right, I mean that

790
00:35:15,239 --> 00:35:16,920
Speaker 3: if you look at the other side of this, what

791
00:35:16,960 --> 00:35:19,200
Speaker 3: are we going to have lots of underused data set

792
00:35:19,360 --> 00:35:22,520
Speaker 3: is pumping out tokens at very very low prices reasonably effectively,

793
00:35:22,760 --> 00:35:24,720
Speaker 3: that's going to make it much cheaper to automate people.

794
00:35:24,800 --> 00:35:27,280
Speaker 3: So the consequences on the other side are even larger,

795
00:35:27,280 --> 00:35:29,640
Speaker 3: and there's people wandering around making this argument better than

796
00:35:29,640 --> 00:35:31,080
Speaker 3: I can, and I mean, but that's I think that's

797
00:35:31,080 --> 00:35:33,439
Speaker 3: an important point that leaving aside what happens in terms

798
00:35:33,440 --> 00:35:36,120
Speaker 3: of the next two or three years of captex frenzy,

799
00:35:36,360 --> 00:35:38,480
Speaker 3: at the other end, it's really predictable. You have this

800
00:35:39,120 --> 00:35:42,320
Speaker 3: incredibly low cost token producing engines out there where everybody's

801
00:35:42,320 --> 00:35:44,319
Speaker 3: trying to cover their marginal costs because they got all

802
00:35:44,320 --> 00:35:46,040
Speaker 3: this debt behind it. And guess what they're going to do.

803
00:35:46,160 --> 00:35:47,960
Speaker 3: They're going to drop prices as low as they can

804
00:35:48,000 --> 00:35:50,399
Speaker 3: and pump tokens out like crazy while they can, which

805
00:35:50,400 --> 00:35:53,680
Speaker 3: will have implications for further automation of the workforce.

806
00:35:53,760 --> 00:35:58,239
Speaker 2: And make companies who are looking to replace jobs with

807
00:35:59,160 --> 00:36:02,080
Speaker 2: you know, and it will make it even more price

808
00:36:02,239 --> 00:36:02,920
Speaker 2: cost effective.

809
00:36:03,000 --> 00:36:05,759
Speaker 3: Yeah, I can just waste tokens with impunity. But and

810
00:36:05,840 --> 00:36:07,080
Speaker 3: so that's at the other end of this.

811
00:36:07,400 --> 00:36:10,640
Speaker 2: That's absolutely a different And if you think there's a

812
00:36:10,680 --> 00:36:15,040
Speaker 2: lot of video and music slop on Spotify and Instagram

813
00:36:15,080 --> 00:36:17,520
Speaker 2: and everywhere else, right now, just wait, just wait, just wait,

814
00:36:17,680 --> 00:36:19,439
Speaker 2: it's free. Just wait till it's all free. Right.

815
00:36:19,560 --> 00:36:22,760
Speaker 3: So, I just think the idea that people get stuck

816
00:36:22,760 --> 00:36:25,399
Speaker 3: on this thing of like, well it's just a ws

817
00:36:25,520 --> 00:36:28,000
Speaker 3: or it's just another tech bubble or whatever else, is

818
00:36:28,080 --> 00:36:29,880
Speaker 3: like completely lost.

819
00:36:29,960 --> 00:36:31,920
Speaker 1: I'm not going to ask you what the trade is,

820
00:36:32,400 --> 00:36:34,279
Speaker 1: but I do know people are listening and thinking, well,

821
00:36:34,320 --> 00:36:37,600
Speaker 1: what do I do in this situation? Dick, do you

822
00:36:37,680 --> 00:36:38,240
Speaker 1: have any answers?

823
00:36:38,360 --> 00:36:38,919
Speaker 2: Doge coin?

824
00:36:42,400 --> 00:36:45,160
Speaker 1: I should have said that absolutely all right, there you go.

825
00:36:45,400 --> 00:36:47,040
Speaker 1: You have it you heard it here first.

826
00:36:47,120 --> 00:36:49,839
Speaker 3: No, there's nothing it's it's this is like a planetary

827
00:36:50,120 --> 00:36:52,240
Speaker 3: object smashing into another planetary object.

828
00:36:52,280 --> 00:36:53,120
Speaker 2: There's nothing you can do.

829
00:36:53,560 --> 00:36:56,160
Speaker 3: Just it's like that scene in Hitchhiker's Gat to the Galaxy.

830
00:36:56,480 --> 00:36:57,680
Speaker 3: You know, what should we do right now?

831
00:36:58,360 --> 00:37:00,239
Speaker 2: Yah? Yeah? But it makes you feel better? Do you

832
00:37:00,280 --> 00:37:02,239
Speaker 2: think this is the Michael Burry so for people who

833
00:37:02,239 --> 00:37:04,200
Speaker 2: are listening, I don't know that is that from the

834
00:37:04,200 --> 00:37:08,680
Speaker 2: Big Short has been commenting a lot on this online lay.

835
00:37:09,239 --> 00:37:13,480
Speaker 2: One of his sort of cryptic tweets the other day was, sometimes,

836
00:37:13,520 --> 00:37:15,680
Speaker 2: you know, there are sometimes moments where you just see

837
00:37:15,680 --> 00:37:17,520
Speaker 2: it coming in. There's nothing you can do other than

838
00:37:17,560 --> 00:37:20,799
Speaker 2: not play the game. Yeah, yeah, And it seems like

839
00:37:20,840 --> 00:37:23,640
Speaker 2: he was referencing this. Yeah. I think so.

840
00:37:23,680 --> 00:37:25,040
Speaker 3: And I mean and I think in a lot of ways,

841
00:37:25,080 --> 00:37:27,279
Speaker 3: this current moment has, you know, echoes of some of

842
00:37:27,320 --> 00:37:29,279
Speaker 3: the stuff that he was that he did well during

843
00:37:29,960 --> 00:37:33,239
Speaker 3: that particular crisis in terms of recognizing this sort of

844
00:37:33,280 --> 00:37:35,400
Speaker 3: flywheel of debt creation, because we haven't talked about this

845
00:37:35,440 --> 00:37:37,440
Speaker 3: idea that one of the reasons why things are spiraling

846
00:37:37,440 --> 00:37:39,480
Speaker 3: out of control is because some of the people who

847
00:37:39,520 --> 00:37:42,319
Speaker 3: are providing the capital don't get a rats ask about

848
00:37:42,320 --> 00:37:44,480
Speaker 3: what's going on in AI. They just like the idea

849
00:37:44,600 --> 00:37:48,360
Speaker 3: that a really good credit like Google, Google or Amazon

850
00:37:48,400 --> 00:37:51,560
Speaker 3: whoever else, is going to pay money for renting access

851
00:37:51,600 --> 00:37:53,440
Speaker 3: to a data center. And then at the other end,

852
00:37:53,600 --> 00:37:55,799
Speaker 3: I can take that income stream and repackage it and

853
00:37:55,840 --> 00:37:58,279
Speaker 3: sell it. Whether there's AI going on in the middle

854
00:37:58,360 --> 00:38:00,360
Speaker 3: or hide and seek, I don't care.

855
00:38:01,400 --> 00:38:03,040
Speaker 2: It doesn't make any difference. And that's so good.

856
00:38:03,120 --> 00:38:05,320
Speaker 3: So that capital just keeps going around and round the

857
00:38:05,360 --> 00:38:07,560
Speaker 3: rond because you're perfectly happy to just keep repackaging it

858
00:38:07,800 --> 00:38:08,840
Speaker 3: and it doesn't matter what's going on.

859
00:38:08,880 --> 00:38:11,800
Speaker 2: So from their standpoint, and that's the pattern he's recognized.

860
00:38:11,800 --> 00:38:14,160
Speaker 3: That's the pattern he's recognized because from their standpoint, data

861
00:38:14,200 --> 00:38:17,720
Speaker 3: centers are basically just apartment buildings and a bunch of tenants.

862
00:38:17,800 --> 00:38:20,279
Speaker 3: They're all paying rent. Some of these tenants are really good.

863
00:38:20,880 --> 00:38:22,799
Speaker 3: I want more of that action because then I can

864
00:38:22,800 --> 00:38:25,600
Speaker 3: take the income stream that's produced and repackage and.

865
00:38:25,560 --> 00:38:28,719
Speaker 2: Relic of different pieces, resell it and resell it.

866
00:38:28,960 --> 00:38:31,040
Speaker 3: And so this is the mistake the tech techies make

867
00:38:31,080 --> 00:38:32,360
Speaker 3: all the time as they get hung up on the

868
00:38:32,400 --> 00:38:34,480
Speaker 3: idea of well it's AI. No, that has something to

869
00:38:34,520 --> 00:38:36,520
Speaker 3: do with the credit side of this. As I said,

870
00:38:36,520 --> 00:38:39,200
Speaker 3: it could literally be like you know, securitize rented hide

871
00:38:39,239 --> 00:38:41,759
Speaker 3: and seek competitions. It doesn't really matter. And that's why

872
00:38:41,800 --> 00:38:44,840
Speaker 3: you're going to have so much more data centers created,

873
00:38:45,239 --> 00:38:47,200
Speaker 3: is because that on the other side, they love the

874
00:38:47,239 --> 00:38:50,000
Speaker 3: income stream in particular, given that they're like, well, this

875
00:38:50,120 --> 00:38:53,120
Speaker 3: can't fail because look it's Google making payments on this

876
00:38:53,719 --> 00:38:54,600
Speaker 3: or Meta or whoever.

877
00:39:18,920 --> 00:39:22,319
Speaker 1: Well, given that my prediction was correct. When I said

878
00:39:22,320 --> 00:39:25,000
Speaker 1: the word data center to Paul, we got this response,

879
00:39:25,040 --> 00:39:27,319
Speaker 1: I'm gonna I think we should. We should come to

880
00:39:27,400 --> 00:39:30,480
Speaker 1: a close with this podcast by saying something to Dick

881
00:39:30,560 --> 00:39:34,080
Speaker 1: Costello that will equally get him upset. It is not

882
00:39:34,200 --> 00:39:36,160
Speaker 1: duck regu. It's not data centers.

883
00:39:36,200 --> 00:39:39,360
Speaker 2: It is I'm not making a duck ragu. Forget's a rabbit,

884
00:39:39,760 --> 00:39:41,720
Speaker 2: one's a bird, and one's a mammal.

885
00:39:42,000 --> 00:39:47,960
Speaker 1: It's not like, sorry, rabbit ragu. It is you ready,

886
00:39:47,960 --> 00:39:50,080
Speaker 1: Paul travel and TikTok.

887
00:39:50,640 --> 00:39:53,480
Speaker 2: Yeah, So it's really bad.

888
00:39:53,920 --> 00:39:55,160
Speaker 1: It's so fucking terrible.

889
00:39:55,400 --> 00:40:00,759
Speaker 2: So I was I was in uh Florence, Italy on

890
00:40:00,800 --> 00:40:02,799
Speaker 2: a cooking trip and.

891
00:40:02,719 --> 00:40:04,320
Speaker 1: To make duck regu or rabbit regue.

892
00:40:04,800 --> 00:40:06,640
Speaker 2: You know what, I'm just gonna pretend you're not here.

893
00:40:06,480 --> 00:40:08,480
Speaker 3: For a little while you're close enough to hit them.

894
00:40:08,719 --> 00:40:13,880
Speaker 2: Yeah, I know. And there are two things that I observed.

895
00:40:13,960 --> 00:40:21,000
Speaker 2: One is the the classic uh spots, you know, Michelangelo's David,

896
00:40:21,880 --> 00:40:28,160
Speaker 2: the Duomo. They're they're they're like the lines are insurmountable.

897
00:40:28,640 --> 00:40:32,000
Speaker 2: I mean the lines are the lines at this near

898
00:40:32,000 --> 00:40:36,400
Speaker 2: the Duomo in the center Florence. I mean, I'm guessing

899
00:40:37,200 --> 00:40:39,719
Speaker 2: three three and a half hours. They're they're hundreds of

900
00:40:40,040 --> 00:40:44,120
Speaker 2: yards long, winding around, and there are like you can't

901
00:40:44,160 --> 00:40:45,839
Speaker 2: fit an infinite you know, there are only so many

902
00:40:45,840 --> 00:40:47,479
Speaker 2: people that are gonna allow in the church at a time,

903
00:40:47,719 --> 00:40:49,239
Speaker 2: and those people want to hang out in there and

904
00:40:49,280 --> 00:40:54,400
Speaker 2: take pictures and post their pictures and and and you're thinking, like, okay,

905
00:40:54,440 --> 00:40:58,200
Speaker 2: you're you're in this town for some period of time

906
00:40:58,280 --> 00:41:02,239
Speaker 2: and you're going to spend four hours of it. Then

907
00:41:02,320 --> 00:41:04,880
Speaker 2: the second order effects of that are horrible. Second order

908
00:41:04,880 --> 00:41:09,640
Speaker 2: effects include one, the entire piazza is now Choschke salespeople

909
00:41:09,680 --> 00:41:14,240
Speaker 2: who are selling little trinkets, and you know, the beautiful

910
00:41:14,239 --> 00:41:16,880
Speaker 2: the little cafes and everything are being replaced by the

911
00:41:16,960 --> 00:41:20,759
Speaker 2: A the Choshke shops and b the places that just

912
00:41:20,840 --> 00:41:23,000
Speaker 2: all serve the same pasta because I have to go

913
00:41:23,080 --> 00:41:27,000
Speaker 2: to Florence and get you know, carbonar or whatever. The

914
00:41:27,040 --> 00:41:33,560
Speaker 2: one the number two bad derivative consequence of these horrible

915
00:41:33,800 --> 00:41:36,160
Speaker 2: you know when I talk about the other thing that

916
00:41:36,200 --> 00:41:39,319
Speaker 2: happens on on TikTok and Instagram, of course, is you know,

917
00:41:39,360 --> 00:41:41,719
Speaker 2: the influencer goes like, these are the top ten restaurants

918
00:41:41,719 --> 00:41:44,960
Speaker 2: in Laurence, Italy. We're obsessed. We went, you know, number three, Cystonza.

919
00:41:45,000 --> 00:41:49,600
Speaker 2: We're obsessed with the butter chicken. So you go to Cistanza,

920
00:41:50,200 --> 00:41:56,560
Speaker 2: which is this great little trotteria on Gobi Street, and

921
00:41:57,480 --> 00:42:00,560
Speaker 2: used to be like, you know, go there and walk in,

922
00:42:00,640 --> 00:42:02,400
Speaker 2: maybe at a table, maybe wait a little bit. And

923
00:42:02,400 --> 00:42:05,839
Speaker 2: now it's like you have a reservation or you don't

924
00:42:05,880 --> 00:42:10,480
Speaker 2: sit because it's you know, number three on the influencers,

925
00:42:11,440 --> 00:42:17,720
Speaker 2: and it's all it's all almost all Americans, and they

926
00:42:18,080 --> 00:42:22,239
Speaker 2: all only order the buttered chicken. So you go back

927
00:42:22,280 --> 00:42:24,400
Speaker 2: in the back of the restaurant and talk to the

928
00:42:24,480 --> 00:42:28,000
Speaker 2: guy who owns the place, and he's like, I go

929
00:42:28,080 --> 00:42:31,680
Speaker 2: through five and a half kilos of butter a day,

930
00:42:32,800 --> 00:42:36,640
Speaker 2: unpacked right now, which is great. People only order the

931
00:42:36,680 --> 00:42:39,439
Speaker 2: buttered chicken, and there's going to come and by the way,

932
00:42:39,600 --> 00:42:43,040
Speaker 2: all the locals now who can't come here anymore because

933
00:42:43,080 --> 00:42:45,000
Speaker 2: it's booked way up in advance, and so they don't

934
00:42:45,000 --> 00:42:47,560
Speaker 2: even think about it. There's obviously going to come a

935
00:42:47,600 --> 00:42:50,680
Speaker 2: point a week, a month, a year, whenever from now

936
00:42:51,160 --> 00:42:53,800
Speaker 2: when people stop going to Cistanza for the butter chicken,

937
00:42:54,360 --> 00:42:57,840
Speaker 2: and the place will be empty, plates will be empty,

938
00:42:58,280 --> 00:42:59,959
Speaker 2: and you know, he'll be like, I got to start

939
00:43:00,080 --> 00:43:02,279
Speaker 2: basically start over. I have this great little.

940
00:43:02,160 --> 00:43:05,680
Speaker 1: Does the owner have any like any like? Is he

941
00:43:05,719 --> 00:43:07,200
Speaker 1: annoyed by Yeah, but.

942
00:43:07,200 --> 00:43:10,680
Speaker 2: He's also like what am I gonna do? Yeah? Yeah,

943
00:43:10,680 --> 00:43:15,759
Speaker 2: what's my what's my chicken? Yeah? I'm not He's like

944
00:43:16,000 --> 00:43:17,920
Speaker 2: unpacked every night.

945
00:43:18,360 --> 00:43:23,280
Speaker 3: Yeah. But there's a can I abstract upper level please.

946
00:43:23,960 --> 00:43:25,480
Speaker 3: There's a tremendous book that.

947
00:43:25,480 --> 00:43:27,160
Speaker 1: You can't say the word data center though.

948
00:43:27,640 --> 00:43:31,680
Speaker 3: So in terms of data centers. Actually this is connected

949
00:43:31,680 --> 00:43:32,239
Speaker 3: with data centers.

950
00:43:32,280 --> 00:43:32,600
Speaker 2: What I was.

951
00:43:34,160 --> 00:43:36,440
Speaker 3: There's a tremendous book from like eighty eight or eighty

952
00:43:36,520 --> 00:43:39,000
Speaker 3: nine by a guy named Steve Strogatz who's a mathematician

953
00:43:39,040 --> 00:43:42,040
Speaker 3: at Cornell, which is about this idea of the spontaneous

954
00:43:42,080 --> 00:43:45,279
Speaker 3: emergence of synchrony in complex systems. So the idea that,

955
00:43:45,960 --> 00:43:50,240
Speaker 3: for example, like butterflies in Georgia's once one there's initially

956
00:43:50,239 --> 00:43:52,960
Speaker 3: start off very chaot not butterfly, sorry, fireflies, the one

957
00:43:53,000 --> 00:43:55,680
Speaker 3: will start off signaling and it's very chaotic, but the

958
00:43:55,719 --> 00:43:59,080
Speaker 3: system rapidly converges until the entire group of fireflies is

959
00:43:59,120 --> 00:44:01,600
Speaker 3: all pulsing in uni right, which is you see this

960
00:44:01,640 --> 00:44:04,640
Speaker 3: in natural systems all the time. And there's certain characteristics

961
00:44:04,640 --> 00:44:07,120
Speaker 3: of systems where that tends to happen. But among them

962
00:44:07,160 --> 00:44:09,480
Speaker 3: is this idea that everyone can observe each other. Right,

963
00:44:09,480 --> 00:44:12,319
Speaker 3: all the fireflies see what the other firefly is doing

964
00:44:12,360 --> 00:44:14,839
Speaker 3: and spontaneously changes to be in sync with each of them.

965
00:44:15,160 --> 00:44:16,840
Speaker 3: In human systems, for a long time that was very

966
00:44:16,880 --> 00:44:18,880
Speaker 3: hard to do because we couldn't observe each other at

967
00:44:18,920 --> 00:44:22,280
Speaker 3: large scale. But now with social media, we can observe

968
00:44:22,280 --> 00:44:24,719
Speaker 3: each other at large scale. So guess what happens. Much

969
00:44:24,800 --> 00:44:26,919
Speaker 3: like the fireflies pulsing in the forest.

970
00:44:26,640 --> 00:44:27,799
Speaker 2: We all ordered the butter chicken.

971
00:44:27,800 --> 00:44:31,040
Speaker 3: We all order the buttered chicken. So this synchrony emerges

972
00:44:31,080 --> 00:44:33,840
Speaker 3: in these systems that never happened before, and it becomes

973
00:44:33,880 --> 00:44:34,680
Speaker 3: even more.

974
00:44:36,160 --> 00:44:36,719
Speaker 2: Organized.

975
00:44:36,760 --> 00:44:39,680
Speaker 3: So you'll have the point masses of people converge on

976
00:44:39,719 --> 00:44:43,080
Speaker 3: a single plaza or on a single restaurant. But the

977
00:44:43,120 --> 00:44:46,520
Speaker 3: difference isn't just the existence of influencers. The existence is

978
00:44:46,600 --> 00:44:49,880
Speaker 3: these systems make possible something that was never possible before,

979
00:44:49,960 --> 00:44:52,319
Speaker 3: because now we can all observe each other in real time,

980
00:44:52,600 --> 00:44:55,680
Speaker 3: and even if we don't want to. Humans are hurting creatures.

981
00:44:55,680 --> 00:44:57,319
Speaker 3: They like to see what each other is doing and

982
00:44:57,360 --> 00:44:58,920
Speaker 3: kind of fit in with the tribe, and so they

983
00:44:58,960 --> 00:45:01,600
Speaker 3: all organize themselves in a way that causes exactly this.

984
00:45:01,719 --> 00:45:04,239
Speaker 3: And it's really it goes back to like strokeouts and

985
00:45:04,320 --> 00:45:06,560
Speaker 3: fireflies and all this stuff, and it's like, you can

986
00:45:06,560 --> 00:45:08,480
Speaker 3: only see it getting more dramatic with hert.

987
00:45:08,560 --> 00:45:13,600
Speaker 2: It'll be interesting to see what the reaction is from

988
00:45:13,880 --> 00:45:17,000
Speaker 2: towns that are become fed up with it, where it's

989
00:45:17,000 --> 00:45:20,040
Speaker 2: just I mean, there are certain streets in central floors

990
00:45:20,080 --> 00:45:22,480
Speaker 2: you just can't walk through. This is in October. It's

991
00:45:22,520 --> 00:45:25,600
Speaker 2: not like I remember you summer, you just can't. It's

992
00:45:25,640 --> 00:45:28,120
Speaker 2: literally playing Mario Kart, just trying to get out of

993
00:45:28,120 --> 00:45:31,319
Speaker 2: the neighborhood. And you wonder if there'll be like, hey,

994
00:45:31,360 --> 00:45:33,880
Speaker 2: we don't you know, restaurants they're like, we don't allow no,

995
00:45:34,000 --> 00:45:36,279
Speaker 2: can't like no phones in the restaurant. I mean you

996
00:45:36,320 --> 00:45:38,279
Speaker 2: have to imagine start to I.

997
00:45:38,160 --> 00:45:40,160
Speaker 3: Remember like frisk you to see if where they have

998
00:45:40,160 --> 00:45:41,279
Speaker 3: an Instagram on your phone.

999
00:45:41,320 --> 00:45:44,080
Speaker 1: It's like, dude, you're It's like there was that pushback

1000
00:45:44,080 --> 00:45:45,759
Speaker 1: where it was like probably ten years ago where people

1001
00:45:45,800 --> 00:45:48,319
Speaker 1: were like coffee shops were like no computers, you know,

1002
00:45:49,160 --> 00:45:51,759
Speaker 1: Wi Fi, no, whatever it is, and you have there

1003
00:45:51,760 --> 00:45:54,960
Speaker 1: has to be at some point because I've seen it's horrific.

1004
00:45:55,120 --> 00:45:57,440
Speaker 3: I think it actually gets even more dramatic because it

1005
00:45:57,480 --> 00:46:00,520
Speaker 3: extends across more systems, like you see it in financial markets, right,

1006
00:46:00,560 --> 00:46:02,680
Speaker 3: you see people synchronized because I can now observe other

1007
00:46:02,680 --> 00:46:05,000
Speaker 3: people's trades in Robin Hood, I get a real time

1008
00:46:05,040 --> 00:46:07,600
Speaker 3: sense of what's going on. That's the exact same phenomenons

1009
00:46:07,640 --> 00:46:08,600
Speaker 3: you get prediction markets.

1010
00:46:09,880 --> 00:46:10,319
Speaker 2: Right, all of a.

1011
00:46:10,320 --> 00:46:12,239
Speaker 3: Sudden, I can see what other people think is going

1012
00:46:12,239 --> 00:46:14,440
Speaker 3: to happen, synchronize it so you see it on Instagram.

1013
00:46:15,080 --> 00:46:18,520
Speaker 2: By the way, speaking speak speaking of it's always funny

1014
00:46:18,560 --> 00:46:20,960
Speaker 2: to me when people are shocked shocked to learn that

1015
00:46:21,080 --> 00:46:24,560
Speaker 2: some of the actual performers in the sport are gambling,

1016
00:46:24,760 --> 00:46:28,360
Speaker 2: Like meanwhile, you're watching the sport on TV and every

1017
00:46:28,400 --> 00:46:31,080
Speaker 2: ad is parlay this right now and lah blah blah

1018
00:46:31,120 --> 00:46:33,799
Speaker 2: if you want to on the second half. It's just

1019
00:46:34,960 --> 00:46:37,319
Speaker 2: it's just like begging people to do it. Yeah, of

1020
00:46:37,360 --> 00:46:39,080
Speaker 2: course they're gonna get. Of course you're going to be

1021
00:46:39,120 --> 00:46:41,200
Speaker 2: able to like, what's my edge, you're going to be

1022
00:46:41,200 --> 00:46:42,840
Speaker 2: in this particular market. I'm going to get one of

1023
00:46:42,840 --> 00:46:45,120
Speaker 2: the players or coaches or so and so I'll find somebody.

1024
00:46:45,200 --> 00:46:47,520
Speaker 3: Yeah, just the story back to like early boxing and

1025
00:46:47,560 --> 00:46:49,960
Speaker 3: so on. But this, but this, the synchrony thing, I

1026
00:46:49,960 --> 00:46:51,280
Speaker 3: think is incredibly important.

1027
00:46:51,360 --> 00:46:51,960
Speaker 1: No, it really is.

1028
00:46:52,000 --> 00:46:55,800
Speaker 3: I have it, and it is people don't understand the mechanics,

1029
00:46:55,840 --> 00:46:58,239
Speaker 3: so they don't realize in a sense, how they're being

1030
00:46:58,320 --> 00:47:00,880
Speaker 3: led into it via these top ten lists, which in

1031
00:47:00,880 --> 00:47:03,080
Speaker 3: the sense is kind of like observing other fireflies.

1032
00:47:03,320 --> 00:47:05,960
Speaker 1: Is this the last couple of questions and then we'll

1033
00:47:05,960 --> 00:47:08,359
Speaker 1: wrap up. But if there's anyone still listening, I don't know,

1034
00:47:08,400 --> 00:47:11,759
Speaker 1: but maybe an AI you know, uh listening from the

1035
00:47:11,840 --> 00:47:15,800
Speaker 1: data center. There was this study that was done recently,

1036
00:47:15,840 --> 00:47:17,319
Speaker 1: like the last couple of weeks where they gave all

1037
00:47:17,320 --> 00:47:19,560
Speaker 1: these different ais ten thousand dollars to trade and they

1038
00:47:19,600 --> 00:47:22,120
Speaker 1: all lost like nine hundred and you know, all of

1039
00:47:22,160 --> 00:47:24,800
Speaker 1: it pretty much is that because of the same thing.

1040
00:47:24,719 --> 00:47:26,200
Speaker 2: They blew it all on liquor and Horse.

1041
00:47:28,400 --> 00:47:31,840
Speaker 3: I did see that, which should come as no surprise.

1042
00:47:31,880 --> 00:47:32,840
Speaker 3: They were trained on Reddit.

1043
00:47:35,160 --> 00:47:38,480
Speaker 1: It's all fit, all right, you have any last things,

1044
00:47:38,560 --> 00:47:39,480
Speaker 1: Dick before we wrap.

1045
00:47:39,320 --> 00:47:42,719
Speaker 3: This though, Like even the stuff on the market, I'm just.

1046
00:47:42,719 --> 00:47:43,719
Speaker 2: Gonna sit here and keep talking.

1047
00:47:43,719 --> 00:47:47,600
Speaker 1: We can go to lunch now, please, stuff on the markets.

1048
00:47:47,320 --> 00:47:48,840
Speaker 3: But even the stuff on the markets, like, I just

1049
00:47:48,840 --> 00:47:51,759
Speaker 3: think it comes back to people misunderstanding like how these

1050
00:47:51,760 --> 00:47:54,080
Speaker 3: models are trained, right, I mean I always I joke

1051
00:47:54,080 --> 00:47:55,879
Speaker 3: all the time that it's basically instead of when someone

1052
00:47:55,880 --> 00:47:58,400
Speaker 3: says a large language model told me X, should just

1053
00:47:58,440 --> 00:47:59,920
Speaker 3: say like a thirty seven year old dude on red

1054
00:48:00,120 --> 00:48:01,239
Speaker 3: told me, yeah right, I mean.

1055
00:48:01,160 --> 00:48:04,000
Speaker 1: That's basically open eyes training.

1056
00:48:04,040 --> 00:48:07,399
Speaker 3: That's the median influence. So again, so now thinking about

1057
00:48:07,400 --> 00:48:09,920
Speaker 3: it in stock market terms, if some thirty seven year

1058
00:48:09,960 --> 00:48:11,719
Speaker 3: old dude on Reddit told me that I should be

1059
00:48:11,760 --> 00:48:14,759
Speaker 3: long whatever, would it surprise me to find out that

1060
00:48:14,880 --> 00:48:15,839
Speaker 3: terry doesn't work out?

1061
00:48:16,239 --> 00:48:19,160
Speaker 1: If the other fireflies blinking I'm an link too, Yeah, But.

1062
00:48:19,120 --> 00:48:20,719
Speaker 3: I just think I this is the idea of how

1063
00:48:20,760 --> 00:48:24,120
Speaker 3: badly people misunderstand this thing that they've committed so much

1064
00:48:24,120 --> 00:48:26,440
Speaker 3: time and energy to and then they're surprised at the outcome.

1065
00:48:26,920 --> 00:48:29,080
Speaker 3: Is really I think important because then it shows you

1066
00:48:29,120 --> 00:48:31,480
Speaker 3: why it's going to continue to synchronize behaviors and do

1067
00:48:31,520 --> 00:48:34,200
Speaker 3: all these other things that people will be completely surprised

1068
00:48:34,200 --> 00:48:36,520
Speaker 3: by because it's leading you to the median outcome all

1069
00:48:36,600 --> 00:48:36,960
Speaker 3: the time.

1070
00:48:37,080 --> 00:48:42,920
Speaker 2: And people anoint it with more capacity than it deserves

1071
00:48:43,040 --> 00:48:47,600
Speaker 2: because it's conversational, and so they think, like, obviously it's

1072
00:48:47,760 --> 00:48:51,360
Speaker 2: thinking about this because it's having this conversational tone with me.

1073
00:48:51,680 --> 00:48:55,239
Speaker 2: It's not just reposting. Thirty seven year old guy on

1074
00:48:55,280 --> 00:48:57,640
Speaker 2: Reddit says, these are the top five movies of the year.

1075
00:48:57,640 --> 00:49:00,440
Speaker 3: You know, it feels like someone I know it likes

1076
00:49:00,440 --> 00:49:02,160
Speaker 3: me because it's it will tell me things like that

1077
00:49:02,280 --> 00:49:04,279
Speaker 3: was a really sharp question, Like.

1078
00:49:04,360 --> 00:49:06,440
Speaker 1: I hate when it tells me anything good. I'm like,

1079
00:49:07,480 --> 00:49:09,480
Speaker 1: I'm like, tell me bad things. That's all I want

1080
00:49:09,520 --> 00:49:10,040
Speaker 1: to know it again.

1081
00:49:10,080 --> 00:49:11,319
Speaker 2: That's you can pay extra for that.

1082
00:49:14,160 --> 00:49:18,160
Speaker 1: Well, this has been a fascinating episode. Uh, The Nick

1083
00:49:18,239 --> 00:49:19,960
Speaker 1: dickm Polshow will be coming live to a bunch of

1084
00:49:20,040 --> 00:49:26,080
Speaker 1: data centers in the new year. And that conclusive episode.

1085
00:49:26,120 --> 00:49:26,840
Speaker 1: Thanks guys,