1
00:00:02,440 --> 00:00:03,520
Speaker 1: Also media.

2
00:00:05,880 --> 00:00:08,799
Speaker 2: Hello and welcome to Better Offline. I'm your host ed Zeitron.

3
00:00:20,800 --> 00:00:23,120
Speaker 2: Download a T shirt and subscribe to the newsletter that's

4
00:00:23,120 --> 00:00:25,799
Speaker 2: where you get your words. But today you're here for

5
00:00:25,840 --> 00:00:28,240
Speaker 2: the noises. And joining me today is the wonderful economist

6
00:00:28,280 --> 00:00:29,040
Speaker 2: Paul Kadrowski.

7
00:00:29,280 --> 00:00:30,240
Speaker 3: Paul, good to have you on.

8
00:00:31,000 --> 00:00:31,960
Speaker 1: Hey had good to be here.

9
00:00:32,600 --> 00:00:33,879
Speaker 3: So your recent work.

10
00:00:33,720 --> 00:00:36,279
Speaker 2: On AI, particularly the data center on economic side, it's

11
00:00:36,280 --> 00:00:38,239
Speaker 2: been great. One thing I really want to talk to

12
00:00:38,240 --> 00:00:42,040
Speaker 2: you about is what actual economic effects have you've seen

13
00:00:42,040 --> 00:00:46,520
Speaker 2: from AI, because it feels hard to get specific sometimes

14
00:00:46,600 --> 00:00:47,360
Speaker 2: if you know what I mean.

15
00:00:48,520 --> 00:00:52,120
Speaker 1: Yeah, so it's pretty easy to find that. I mean,

16
00:00:52,120 --> 00:00:53,760
Speaker 1: there's the old joke, which was the early days of

17
00:00:53,760 --> 00:00:56,240
Speaker 1: the technology industry that you could find technology everywhere except

18
00:00:56,280 --> 00:00:58,320
Speaker 1: for in the productivity data. And so that sort of

19
00:00:58,320 --> 00:01:01,000
Speaker 1: applies right here as well. So the answer you'll get

20
00:01:01,040 --> 00:01:03,640
Speaker 1: there's two answers. One is that it's too soon to tell,

21
00:01:03,720 --> 00:01:06,080
Speaker 1: which is fine, but it's a little bit of a sop, right,

22
00:01:07,400 --> 00:01:08,960
Speaker 1: So you'll get that, and then you get the second

23
00:01:08,959 --> 00:01:11,160
Speaker 1: answer is which I already see it. It's in. And

24
00:01:11,200 --> 00:01:13,680
Speaker 1: then someone will add hoc Terry picks some data and

25
00:01:13,720 --> 00:01:17,479
Speaker 1: say there goes Ai right there, And the answer, of course,

26
00:01:17,560 --> 00:01:19,120
Speaker 1: is is that it's nowhere to be seen yet in

27
00:01:19,160 --> 00:01:22,120
Speaker 1: any really meaningful productivity data. Anywhere you have some output

28
00:01:22,160 --> 00:01:26,800
Speaker 1: metrics like for example, you know this incredible number of

29
00:01:26,800 --> 00:01:29,520
Speaker 1: commits you'll see on GitHub using agentic code. But is

30
00:01:29,560 --> 00:01:32,640
Speaker 1: that productivity? I think it's largely masturbatory. I don't think

31
00:01:32,640 --> 00:01:37,040
Speaker 1: it's actually, I always productivity. And so most of what

32
00:01:37,120 --> 00:01:39,000
Speaker 1: got me interested was because you see it all on

33
00:01:39,040 --> 00:01:41,520
Speaker 1: the other side of things. So from an economic standpoint,

34
00:01:41,640 --> 00:01:44,200
Speaker 1: you actually see it in the economic data because of

35
00:01:44,240 --> 00:01:47,240
Speaker 1: how dominants become in a couple of statistics, like for example,

36
00:01:47,840 --> 00:01:51,800
Speaker 1: it's share of non residential fixed investments, which is at

37
00:01:51,880 --> 00:01:54,040
Speaker 1: levels we last saw with the railroad build out or

38
00:01:54,040 --> 00:01:55,880
Speaker 1: with rural electrification.

39
00:01:55,280 --> 00:01:59,120
Speaker 2: And there can you be specific, Well, non residential investments

40
00:01:59,120 --> 00:02:00,640
Speaker 2: may just so I'll get you.

41
00:02:01,440 --> 00:02:06,280
Speaker 1: So atoms that aren't in houses. So that's the short answer.

42
00:02:06,560 --> 00:02:11,519
Speaker 1: So so factories, manufacturing highways, Uh, you know, fixed equipment

43
00:02:11,560 --> 00:02:13,320
Speaker 1: you're putting in your in your in anything that you're

44
00:02:13,360 --> 00:02:17,200
Speaker 1: building for an economic purpose, it isn't residential, Okay, So

45
00:02:17,480 --> 00:02:21,200
Speaker 1: that usually is a fairly broad and diversified category. So

46
00:02:21,240 --> 00:02:23,920
Speaker 1: you've got people building out. You know, if the current

47
00:02:23,919 --> 00:02:27,720
Speaker 1: administration had its way, that statistic would currently be dominated

48
00:02:27,760 --> 00:02:30,640
Speaker 1: by fill in the blank manufacturing, right, because the attempt

49
00:02:30,680 --> 00:02:34,440
Speaker 1: is to onshore manufacturing, so that if policy was working

50
00:02:34,440 --> 00:02:37,600
Speaker 1: in that regard, you'd be seeing the dominant chunk of

51
00:02:37,639 --> 00:02:40,799
Speaker 1: that being the onshoring of manufacturing. As it came back

52
00:02:40,840 --> 00:02:43,800
Speaker 1: to this, to this blessed country. And so it's not

53
00:02:44,240 --> 00:02:47,600
Speaker 1: the largest chunk of non residential fixed investment currently, which

54
00:02:47,639 --> 00:02:50,720
Speaker 1: is data centers, which is a basket of things which

55
00:02:50,720 --> 00:02:54,080
Speaker 1: it's made up predominantly of GPUs, but also the build

56
00:02:54,120 --> 00:02:57,520
Speaker 1: out of the of the of the facilities, HVAC, heating, ventilating,

57
00:02:57,560 --> 00:03:00,400
Speaker 1: air conditioning, cooling, all those kinds of things. So what

58
00:03:00,520 --> 00:03:03,600
Speaker 1: got me interested originally was I couldn't see AI data

59
00:03:03,720 --> 00:03:07,440
Speaker 1: anywhere except for him these categories of fixed investment. And

60
00:03:07,440 --> 00:03:10,840
Speaker 1: then even more startling to me, which apparently I was

61
00:03:10,840 --> 00:03:13,560
Speaker 1: the only one initially startled, then I startled other people,

62
00:03:13,919 --> 00:03:16,880
Speaker 1: was the idea that it was the largest share of

63
00:03:17,040 --> 00:03:20,000
Speaker 1: US GDP growth for three or four quarters last year,

64
00:03:20,520 --> 00:03:24,520
Speaker 1: so that in the absence of this non residential fixed

65
00:03:24,560 --> 00:03:28,600
Speaker 1: investment wonder drug we call data centers, US would have

66
00:03:28,600 --> 00:03:31,519
Speaker 1: been in recession in the first quarter and the fourth quarter,

67
00:03:32,200 --> 00:03:34,800
Speaker 1: neutral in the second, and mildly positive in the third.

68
00:03:34,920 --> 00:03:38,480
Speaker 1: So it was the economic story of twenty twenty five.

69
00:03:38,520 --> 00:03:41,320
Speaker 1: And so the reason why this is important, and the

70
00:03:41,520 --> 00:03:44,920
Speaker 1: analogy I make all the time is my dog barks

71
00:03:45,120 --> 00:03:47,240
Speaker 1: when the mailman comes to the house, and the dog

72
00:03:47,320 --> 00:03:50,560
Speaker 1: keeps barking and the mailman goes away. Right, the dog

73
00:03:50,600 --> 00:03:51,320
Speaker 1: thinks he did it.

74
00:03:51,680 --> 00:03:52,680
Speaker 3: He didn't do it right.

75
00:03:52,960 --> 00:03:55,360
Speaker 1: He has a messed up He has a messed up

76
00:03:55,440 --> 00:03:58,080
Speaker 1: model of causality is what he has right? The dog

77
00:03:58,160 --> 00:04:01,520
Speaker 1: causality imply is that my barking made the male men

78
00:04:01,600 --> 00:04:03,320
Speaker 1: go away. No, the mail man goes away every time.

79
00:04:03,360 --> 00:04:05,800
Speaker 1: It's got other things to do. Right. So the same

80
00:04:05,840 --> 00:04:09,200
Speaker 1: thing is true with respect to fixed investment. If you

81
00:04:09,360 --> 00:04:12,520
Speaker 1: don't realize that the largest share of fixed investment and

82
00:04:12,560 --> 00:04:16,159
Speaker 1: the thing that's driving us GDP growth is this wonder

83
00:04:16,200 --> 00:04:18,960
Speaker 1: drug called data centers, you're the dog barking at the

84
00:04:18,960 --> 00:04:21,360
Speaker 1: mailman and the mailman goes away. Anyway, you don't actually

85
00:04:21,440 --> 00:04:24,800
Speaker 1: understand the things that are driving economic growth. So there's

86
00:04:25,160 --> 00:04:28,720
Speaker 1: the long answer is that's where you see data centers

87
00:04:28,760 --> 00:04:31,440
Speaker 1: in economic data, and it's in this very strange place

88
00:04:31,440 --> 00:04:33,320
Speaker 1: that was largely being missed for the longest time.

89
00:04:33,839 --> 00:04:38,200
Speaker 2: So with that number, that includes all of Nvidia's sales

90
00:04:38,440 --> 00:04:41,560
Speaker 2: all told, or just the ones in going to American

91
00:04:41,560 --> 00:04:43,839
Speaker 2: clients or is it just all of their sales?

92
00:04:44,839 --> 00:04:47,560
Speaker 1: No, So you parcel it out obviously so by geography.

93
00:04:47,600 --> 00:04:50,359
Speaker 1: So what matters immensely what's happening in the US? No, Granted,

94
00:04:50,360 --> 00:04:54,400
Speaker 1: the predominant share of in videos sales, just like the

95
00:04:54,520 --> 00:04:57,320
Speaker 1: largest sales of most of you know, transformers and everything,

96
00:04:57,360 --> 00:04:59,280
Speaker 1: are all happening in the US right as this build

97
00:04:59,279 --> 00:05:01,520
Speaker 1: out happens. Some like seventy to eighty percent of the

98
00:05:01,520 --> 00:05:04,119
Speaker 1: global data center build out is happening in the US,

99
00:05:04,160 --> 00:05:06,719
Speaker 1: and most of that is happening in northern Virginia, Texas.

100
00:05:07,040 --> 00:05:09,159
Speaker 1: So it's largely a US phenomenon in the first place.

101
00:05:09,200 --> 00:05:11,520
Speaker 1: But nevertheless, you have to parcel out the pieces of Burbrialley.

102
00:05:12,040 --> 00:05:13,159
Speaker 3: Yeah, I just meant it more.

103
00:05:13,279 --> 00:05:15,480
Speaker 2: Is when in video makes a doll that does that

104
00:05:15,520 --> 00:05:19,600
Speaker 2: count into this? Because yes, but see that's the thing

105
00:05:19,680 --> 00:05:23,000
Speaker 2: that feels like something that is just kind of almost

106
00:05:23,040 --> 00:05:27,359
Speaker 2: like a load bearing chip, a load bearing GPR. Like

107
00:05:27,440 --> 00:05:31,120
Speaker 2: if these sales go down, as bad for everyone.

108
00:05:31,360 --> 00:05:33,240
Speaker 1: Right and there's been some tremendous piece the Wall Street

109
00:05:33,279 --> 00:05:35,120
Speaker 1: journally at a piece last week, I think, and I

110
00:05:35,200 --> 00:05:37,560
Speaker 1: said some salina and scendiary things in it, but it

111
00:05:37,640 --> 00:05:41,200
Speaker 1: was about how in Vidia kind of sits at the

112
00:05:41,240 --> 00:05:43,919
Speaker 1: center of this, like you know, Don Carleoni and sitting

113
00:05:44,040 --> 00:05:46,159
Speaker 1: with everything happening and everyone coming to the table, and

114
00:05:46,160 --> 00:05:49,800
Speaker 1: they're this you know, mafia Don, who is investing in things, right,

115
00:05:49,839 --> 00:05:52,120
Speaker 1: So they play the role of investor, they play the

116
00:05:52,200 --> 00:05:55,799
Speaker 1: role of acquirer, they play the role of you know, vendor.

117
00:05:56,000 --> 00:05:59,120
Speaker 1: So they have this this incredible hub role. So each

118
00:05:59,200 --> 00:06:01,159
Speaker 1: dollar that they're putting out there, in a sense, is

119
00:06:01,320 --> 00:06:03,560
Speaker 1: vastly more important because in a sense they are the

120
00:06:03,600 --> 00:06:05,720
Speaker 1: load bearing beam in the middle of all of this.

121
00:06:06,000 --> 00:06:08,479
Speaker 1: They are for now and it's changing rapidly, but nevertheless

122
00:06:08,520 --> 00:06:09,080
Speaker 1: they are for now.

123
00:06:09,560 --> 00:06:12,080
Speaker 2: The thing is it that just feels very unstable to

124
00:06:12,120 --> 00:06:15,599
Speaker 2: me because in video, from what I've worked out for

125
00:06:15,680 --> 00:06:18,040
Speaker 2: them to keep growing at their current right, they're going

126
00:06:18,080 --> 00:06:19,960
Speaker 2: to be selling one hundred and twenty billion dollars of

127
00:06:20,000 --> 00:06:23,080
Speaker 2: GPUs by in a year and then in like Q

128
00:06:23,160 --> 00:06:25,159
Speaker 2: two f y twenty eight, I think it will be

129
00:06:25,200 --> 00:06:28,400
Speaker 2: next year. Yeah, And that's just yeah, that feels impractical,

130
00:06:28,640 --> 00:06:31,600
Speaker 2: like just on a on a like an economic level,

131
00:06:31,640 --> 00:06:33,560
Speaker 2: for any country or anyone investing.

132
00:06:34,560 --> 00:06:36,359
Speaker 1: I have a tendency to make that argument, and I

133
00:06:36,360 --> 00:06:38,400
Speaker 1: hate when I do it because and this is so

134
00:06:38,440 --> 00:06:40,280
Speaker 1: we both have fault for this is this is Paul's

135
00:06:40,360 --> 00:06:43,640
Speaker 1: argument from personal incredulity. I don't think it can happen,

136
00:06:43,640 --> 00:06:45,839
Speaker 1: Therefore it can happen. So let's take them at phase

137
00:06:46,279 --> 00:06:47,960
Speaker 1: So let's take them at phase value. Let's say they're

138
00:06:48,040 --> 00:06:49,720
Speaker 1: right right, say this is what's going to happen. You

139
00:06:49,720 --> 00:06:52,479
Speaker 1: have to look at the dynamics. And they conceded this

140
00:06:52,560 --> 00:06:56,160
Speaker 1: at GtC their most recent conference. The dynamics are changing

141
00:06:56,360 --> 00:06:59,719
Speaker 1: quickly in the marketplace. That's driving two things, sales growth

142
00:06:59,760 --> 00:07:04,039
Speaker 1: in margins. So Nvidia has these anomalous GPU margins in

143
00:07:04,080 --> 00:07:06,839
Speaker 1: excess of seventy percent gross margins, which are ridiculous.

144
00:07:07,000 --> 00:07:10,160
Speaker 3: And that's the serious that's right exactly.

145
00:07:10,480 --> 00:07:13,360
Speaker 1: And so so their margins are very high. But they're

146
00:07:13,360 --> 00:07:17,240
Speaker 1: in the middle of this transition, this admitted transition from

147
00:07:17,440 --> 00:07:20,200
Speaker 1: what's euphemistically called training, which is kind of a misnomer,

148
00:07:20,400 --> 00:07:23,160
Speaker 1: to this thing that's called inference, which is probably more accurate.

149
00:07:23,200 --> 00:07:26,520
Speaker 1: So from training models to answering prompts right, and the

150
00:07:26,760 --> 00:07:30,000
Speaker 1: margins on inference are going to change dramatically because the

151
00:07:30,000 --> 00:07:31,960
Speaker 1: things that gave them and this is one of these

152
00:07:31,960 --> 00:07:34,800
Speaker 1: Silicon Valley silly words, but gave them a moat, the

153
00:07:34,800 --> 00:07:36,600
Speaker 1: things that gave them a moat in the world of

154
00:07:36,640 --> 00:07:39,720
Speaker 1: training are far less important in the world of inference.

155
00:07:39,760 --> 00:07:44,360
Speaker 1: And so you're saying this proliferation of new chip companies

156
00:07:44,400 --> 00:07:48,120
Speaker 1: coming to market incumbents with new products, so they're facing

157
00:07:48,200 --> 00:07:50,520
Speaker 1: much more competition in a world of inference. So even

158
00:07:50,520 --> 00:07:52,760
Speaker 1: if you grant that the market's going to grow as

159
00:07:52,840 --> 00:07:55,480
Speaker 1: large as it once did, which is whatever, it's not

160
00:07:55,520 --> 00:07:57,080
Speaker 1: going to all go to them. They're not in the

161
00:07:57,120 --> 00:07:59,600
Speaker 1: same position to accrue all that benefit that they did

162
00:08:00,080 --> 00:08:02,360
Speaker 1: almost accidentally in the world of training, which is really

163
00:08:02,400 --> 00:08:03,480
Speaker 1: an important point as well.

164
00:08:04,160 --> 00:08:06,960
Speaker 2: See that's the thing. I'm also questioning the demand, and

165
00:08:06,960 --> 00:08:09,800
Speaker 2: I also question whether they're done with training because training

166
00:08:10,080 --> 00:08:12,680
Speaker 2: you correctly said, it's a misnomer because that can mean

167
00:08:12,720 --> 00:08:15,720
Speaker 2: everything from a big pre training run to the post

168
00:08:15,760 --> 00:08:18,440
Speaker 2: training that's necessary to make these things work. And it's

169
00:08:18,720 --> 00:08:21,200
Speaker 2: kind of confusing at the moment because like last year,

170
00:08:21,200 --> 00:08:23,440
Speaker 2: they were saying it's all inference all the time. This

171
00:08:23,560 --> 00:08:26,560
Speaker 2: year they're kind of talking about open Claw. Kind of

172
00:08:26,560 --> 00:08:28,840
Speaker 2: feels like they're a bit lost, which I mean is

173
00:08:28,920 --> 00:08:31,400
Speaker 2: kind of the AI industry at large.

174
00:08:32,200 --> 00:08:36,120
Speaker 1: It's just it's so the way I look at it

175
00:08:36,160 --> 00:08:39,800
Speaker 1: is I always whenever someone says in video, I say

176
00:08:39,800 --> 00:08:43,600
Speaker 1: Saudi Arabia, and when they say tokens, I say humpies. Right,

177
00:08:43,880 --> 00:08:47,120
Speaker 1: So their goal is to get more people. Their goal

178
00:08:47,200 --> 00:08:49,760
Speaker 1: is to get more purchase for people purchasing humbies because

179
00:08:49,760 --> 00:08:54,200
Speaker 1: it's good for them because it consumes the thing they, however,

180
00:08:54,280 --> 00:08:56,959
Speaker 1: indirectly produce, which is to say, these things called tokens,

181
00:08:57,080 --> 00:08:59,679
Speaker 1: not the crick Right. So if you think about it

182
00:08:59,720 --> 00:09:03,240
Speaker 1: in the terms and translate Jensen's you know, GtC talk

183
00:09:03,280 --> 00:09:05,600
Speaker 1: and most of the things he says, it doesn't read

184
00:09:05,640 --> 00:09:09,760
Speaker 1: that differently from a random industrial minister in Saudi Arabia saying,

185
00:09:09,800 --> 00:09:13,160
Speaker 1: you know this oil stuff, If you guys, just back

186
00:09:13,200 --> 00:09:15,560
Speaker 1: away from the evs before it hurts, you get out

187
00:09:15,559 --> 00:09:17,559
Speaker 1: there in the humvies so you're safe on the freeways.

188
00:09:17,720 --> 00:09:18,960
Speaker 1: It can kind of feel equivalent.

189
00:09:19,240 --> 00:09:22,199
Speaker 2: And did that actually was that something that happened? I

190
00:09:22,320 --> 00:09:24,040
Speaker 2: was that like two thousand and eight. I remember there

191
00:09:24,160 --> 00:09:27,840
Speaker 2: was like the stories about empty like parking lots full

192
00:09:27,880 --> 00:09:31,280
Speaker 2: of just unsold cause yeah yeah right here right.

193
00:09:31,280 --> 00:09:34,240
Speaker 1: Yeah, no, no, no exactly so So so I think

194
00:09:34,280 --> 00:09:37,319
Speaker 1: you have to translate a lot of Jensen speak into

195
00:09:37,360 --> 00:09:41,040
Speaker 1: this kind of idea that there is this new commodity emerging,

196
00:09:41,320 --> 00:09:43,320
Speaker 1: no different than oil, no different than I don't know,

197
00:09:43,360 --> 00:09:45,320
Speaker 1: copper or whatever else, and this new token is this

198
00:09:45,440 --> 00:09:48,160
Speaker 1: or this new commodity is this thing called tokens, and

199
00:09:48,200 --> 00:09:52,160
Speaker 1: he's doing what he can as a diligent ambassador for

200
00:09:52,160 --> 00:09:54,560
Speaker 1: the for this commodity called tokens to make sure people

201
00:09:54,640 --> 00:09:57,920
Speaker 1: use as much as possible, like, for example, endorsing this

202
00:09:58,200 --> 00:10:02,080
Speaker 1: completely half hour, wild eyed thing called open claw, which

203
00:10:02,120 --> 00:10:04,079
Speaker 1: is a ridiculous idea to suggest that people should be

204
00:10:04,160 --> 00:10:05,679
Speaker 1: using this in their house. It's like I should have

205
00:10:05,800 --> 00:10:08,160
Speaker 1: you know, my own sort of you know home, uh,

206
00:10:08,520 --> 00:10:11,760
Speaker 1: nuclear reactor. It's it's it's a ridiculously dangerous technology for

207
00:10:11,840 --> 00:10:12,960
Speaker 1: most normals to be using.

208
00:10:13,559 --> 00:10:16,839
Speaker 2: Yeah, I just I also feel like it's a signed there,

209
00:10:16,840 --> 00:10:19,240
Speaker 2: a little washed when it's you've got a three D

210
00:10:19,400 --> 00:10:23,200
Speaker 2: AI generated picture of Jensen Huang, the CEO of a

211
00:10:23,200 --> 00:10:26,600
Speaker 2: company with a multi trillion dollar market cap, with crab

212
00:10:26,640 --> 00:10:31,360
Speaker 2: claws like the indignity of it. He wears seven dollars jackets. Man,

213
00:10:31,480 --> 00:10:33,520
Speaker 2: one man should have a little more swag than claws.

214
00:10:33,920 --> 00:10:37,280
Speaker 2: It's just tremendous jackets though, oh trum, they are trump

215
00:10:37,280 --> 00:10:39,880
Speaker 2: at the men'swag guy on a few years amazing jackets

216
00:10:39,920 --> 00:10:40,440
Speaker 2: the man does.

217
00:10:40,960 --> 00:10:44,280
Speaker 3: Here's a good tailor as well. But any who, So.

218
00:10:44,320 --> 00:10:46,360
Speaker 1: Yeah, so I think that's that's a really important point

219
00:10:46,360 --> 00:10:48,360
Speaker 1: you make though that I think you have to when

220
00:10:48,360 --> 00:10:50,959
Speaker 1: you're The read through on open Claw isn't just some

221
00:10:51,000 --> 00:10:53,800
Speaker 1: confusion about where the market's going, but the read through

222
00:10:53,920 --> 00:10:57,640
Speaker 1: is the promotion in almost an industrial affairs level of

223
00:10:57,679 --> 00:10:59,960
Speaker 1: this commodity called token to how can I get people

224
00:11:00,080 --> 00:11:01,800
Speaker 1: to use more of this? So that's why you hear

225
00:11:02,160 --> 00:11:06,040
Speaker 1: song and dance acts about open Claw, these incredible hyperbole

226
00:11:06,200 --> 00:11:08,679
Speaker 1: about cloud code and how cloud code is going to

227
00:11:08,800 --> 00:11:12,000
Speaker 1: rapidly migrate from the world of software into all of

228
00:11:12,040 --> 00:11:14,840
Speaker 1: white collar work, which is agam and error. It's not

229
00:11:14,840 --> 00:11:17,400
Speaker 1: to say cloud code isn't a really interesting and important

230
00:11:17,400 --> 00:11:21,000
Speaker 1: piece of technology, but people are very misguided about how

231
00:11:21,000 --> 00:11:23,319
Speaker 1: these technologies are going to move or can move, or

232
00:11:23,360 --> 00:11:26,360
Speaker 1: will move from a world of software, which is wildly

233
00:11:26,400 --> 00:11:29,640
Speaker 1: anomalous in terms of both the amount of tokens it produces,

234
00:11:30,160 --> 00:11:32,679
Speaker 1: but also in terms of whether you can leave it alone.

235
00:11:32,760 --> 00:11:34,400
Speaker 1: So think about it by the way I sometimes think

236
00:11:34,400 --> 00:11:36,120
Speaker 1: about it is in terms of the idea of like

237
00:11:36,120 --> 00:11:38,760
Speaker 1: a ground truth. I can look at my software and

238
00:11:38,800 --> 00:11:42,120
Speaker 1: I put create some code and change it and it breaks.

239
00:11:41,520 --> 00:11:45,079
Speaker 1: In AI terms, that's a really tight gradient descent, meaning

240
00:11:45,080 --> 00:11:48,360
Speaker 1: that obviously I've just learned something really quickly. Changing this

241
00:11:48,480 --> 00:11:51,439
Speaker 1: to this doesn't work in most of white collar work.

242
00:11:51,840 --> 00:11:55,320
Speaker 1: That's not true. What matters is I create a PowerPoint presentation,

243
00:11:55,400 --> 00:11:58,280
Speaker 1: does my boss like it? Tell me the gradient descent there?

244
00:11:58,320 --> 00:12:00,360
Speaker 1: There's no gradient to set. It's very subject if it's

245
00:12:00,400 --> 00:12:03,360
Speaker 1: almost an aesthetic answer. There are parts of white collar

246
00:12:03,400 --> 00:12:05,360
Speaker 1: work where that's not true. But much of white color

247
00:12:05,400 --> 00:12:09,000
Speaker 1: work doesn't have the same characteristics as software. So if

248
00:12:09,040 --> 00:12:11,400
Speaker 1: you want to project the kind of growth that people

249
00:12:11,440 --> 00:12:15,360
Speaker 1: like Jensen are projecting, you need to believe that these

250
00:12:15,559 --> 00:12:18,000
Speaker 1: harnesses cloude code, codex blah blah blah.

251
00:12:18,080 --> 00:12:21,080
Speaker 3: Yeah light that you've used with the models, Yeah.

252
00:12:21,000 --> 00:12:23,480
Speaker 1: Yeah, yeah. You have to believe that, like the Veloci

253
00:12:23,559 --> 00:12:26,040
Speaker 1: raptors in Jurassic Park, that they can escape containment, that

254
00:12:26,080 --> 00:12:28,360
Speaker 1: they're going to escape containment, and they're going to get

255
00:12:28,400 --> 00:12:30,800
Speaker 1: out of this corral that we call software and they're

256
00:12:30,840 --> 00:12:32,520
Speaker 1: going to be everywhere. And not only are they going

257
00:12:32,520 --> 00:12:35,280
Speaker 1: to be everywhere, which is happening to a degree, they

258
00:12:35,280 --> 00:12:37,520
Speaker 1: will act in the same way, which is to say,

259
00:12:37,679 --> 00:12:39,800
Speaker 1: they will produce huge amounts of code for tiny or

260
00:12:39,880 --> 00:12:42,320
Speaker 1: huge amounts of output for tiny input. And they can

261
00:12:42,360 --> 00:12:45,040
Speaker 1: be left alone because there is this tight gradient descent

262
00:12:45,080 --> 00:12:47,480
Speaker 1: that tells them whether what they're doing is working or not.

263
00:12:48,000 --> 00:12:50,240
Speaker 3: That's things they don't know. If that like, they don't

264
00:12:50,280 --> 00:12:52,200
Speaker 3: know anything, So you don't guarantee that.

265
00:12:53,040 --> 00:12:55,400
Speaker 1: Because there's no ground truth, right, So this gradia descent

266
00:12:55,440 --> 00:12:57,720
Speaker 1: doesn't work. It doesn't work in almost any domain outside

267
00:12:57,760 --> 00:12:59,640
Speaker 1: of software. So the weird thing is is we've actually

268
00:12:59,679 --> 00:13:02,640
Speaker 1: started off in the nearly perfect domain to give a

269
00:13:02,640 --> 00:13:05,880
Speaker 1: completely unrepresentative example of what the future looks like by

270
00:13:05,960 --> 00:13:08,040
Speaker 1: start off in coding, which is why coding or the

271
00:13:08,120 --> 00:13:10,160
Speaker 1: coders and developers are some of the biggest sort of

272
00:13:10,160 --> 00:13:13,200
Speaker 1: flag waving ambassadors for what's going on. It's like, just

273
00:13:13,240 --> 00:13:15,719
Speaker 1: wait till this shows up in you know, I don't

274
00:13:15,760 --> 00:13:17,920
Speaker 1: know it. Pick your domain and why client work, and

275
00:13:18,240 --> 00:13:20,000
Speaker 1: the reality is those domains are very different.

276
00:13:30,440 --> 00:13:32,120
Speaker 3: It is strange as well. Like the.

277
00:13:33,640 --> 00:13:36,040
Speaker 2: You get very few people who are just like, yeah,

278
00:13:36,080 --> 00:13:39,040
Speaker 2: I kind of like this. It's either the pure hatred,

279
00:13:39,080 --> 00:13:42,040
Speaker 2: which I name my listeners is definitely or there's just

280
00:13:42,080 --> 00:13:46,120
Speaker 2: this cult like thing around it. I've never I've been

281
00:13:46,160 --> 00:13:49,160
Speaker 2: around tech for a while. You've probably been around it longer. Yeah,

282
00:13:49,200 --> 00:13:51,440
Speaker 2: I've never seen anything like this. And I've been on

283
00:13:51,520 --> 00:13:53,199
Speaker 2: web forums or games consoles.

284
00:13:53,280 --> 00:13:53,800
Speaker 3: I've been on.

285
00:13:54,320 --> 00:13:57,320
Speaker 2: I've been on I've used to because I'm a strange person,

286
00:13:57,400 --> 00:14:01,200
Speaker 2: read bodybuilding forums and BMW forums, and the arguments on

287
00:14:01,280 --> 00:14:03,160
Speaker 2: there were the same, like if you don't drive an

288
00:14:03,240 --> 00:14:05,240
Speaker 2: M three, I will run you over with mine. Then

289
00:14:05,240 --> 00:14:08,439
Speaker 2: that kind of thing. I've never seen anything like this happen.

290
00:14:08,520 --> 00:14:12,000
Speaker 2: Like I maybe in stocks, yeah, get.

291
00:14:12,040 --> 00:14:14,480
Speaker 1: In stocks to a limited degree. Right, So there's these

292
00:14:14,480 --> 00:14:17,520
Speaker 1: two overlapping phenomena going on. I'm convinced one this is

293
00:14:17,520 --> 00:14:21,160
Speaker 1: a tribal signifier. I show what tribe I'm in by

294
00:14:21,840 --> 00:14:26,560
Speaker 1: by being as wild eyed as possible about my support

295
00:14:26,600 --> 00:14:28,840
Speaker 1: for this shows I'm in the in group or the outgroup.

296
00:14:28,880 --> 00:14:31,280
Speaker 1: And you see that, right, this sort of tribal signifier stuff.

297
00:14:31,520 --> 00:14:33,080
Speaker 1: And the other one is and this is the one

298
00:14:33,080 --> 00:14:36,640
Speaker 1: that I find most fascinating is this kind of frantic

299
00:14:36,720 --> 00:14:39,400
Speaker 1: quest to believe that if we work hard enough and

300
00:14:39,440 --> 00:14:41,400
Speaker 1: fast enough, that I'll never have to work again. And

301
00:14:41,440 --> 00:14:44,000
Speaker 1: you hear this from the people in the Acceleration on,

302
00:14:44,040 --> 00:14:47,560
Speaker 1: the Acceleration camp, the Singularity camp, that when you penetrate

303
00:14:47,600 --> 00:14:50,920
Speaker 1: deeply enough, it's really just I think if I'm kind

304
00:14:50,960 --> 00:14:54,200
Speaker 1: of creating, think of me as as a bomber and

305
00:14:54,240 --> 00:14:56,080
Speaker 1: I've got a vest, I've got a vest attached with

306
00:14:56,120 --> 00:14:58,480
Speaker 1: all these explosives, and I'm threatened to walk into your

307
00:14:58,520 --> 00:15:01,120
Speaker 1: economy and blow it all up, right, and just try

308
00:15:01,160 --> 00:15:03,120
Speaker 1: and stop me, because there's hundreds of us doing this,

309
00:15:03,200 --> 00:15:04,520
Speaker 1: and we're all going to come and we're all going

310
00:15:04,560 --> 00:15:05,960
Speaker 1: to blow up your economy, and what are you gonna

311
00:15:06,000 --> 00:15:06,400
Speaker 1: do about it?

312
00:15:06,520 --> 00:15:06,680
Speaker 2: Right?

313
00:15:07,360 --> 00:15:09,040
Speaker 1: And so that's what I think a lot of this

314
00:15:09,160 --> 00:15:10,840
Speaker 1: is because they feel like if I do that, then

315
00:15:10,840 --> 00:15:13,000
Speaker 1: the government has no option but tends to do some

316
00:15:13,120 --> 00:15:15,480
Speaker 1: kind of broad policy of support where diet I can

317
00:15:15,520 --> 00:15:17,800
Speaker 1: then go out and you know, make figurines all day

318
00:15:17,840 --> 00:15:18,720
Speaker 1: or whatever it is I want to do.

319
00:15:19,400 --> 00:15:22,000
Speaker 2: Yeah. But I will also say that the people who

320
00:15:22,000 --> 00:15:23,800
Speaker 2: are most excited about a I don't seem to have

321
00:15:23,880 --> 00:15:24,520
Speaker 2: other hobbies.

322
00:15:26,160 --> 00:15:28,520
Speaker 1: That's always the thing of fun. It's totally true.

323
00:15:28,760 --> 00:15:30,400
Speaker 3: What do you plan to do with spare time?

324
00:15:30,840 --> 00:15:33,280
Speaker 2: I'm learning piano right now, I'm learning to code for fun.

325
00:15:33,320 --> 00:15:35,200
Speaker 2: And it's like I look at these people and they're.

326
00:15:35,000 --> 00:15:37,360
Speaker 3: Like, what do you do? I run seventeen sub agents

327
00:15:37,360 --> 00:15:37,800
Speaker 3: an hour.

328
00:15:38,360 --> 00:15:41,040
Speaker 2: If these things, if these things don't create my special

329
00:15:41,080 --> 00:15:43,160
Speaker 2: project that I'll never show you, I will die.

330
00:15:43,480 --> 00:15:46,040
Speaker 1: It's like to gray out the power in my neighborhood

331
00:15:46,040 --> 00:15:49,280
Speaker 1: by running as many subprocesses as possible. Yeah, there great hobbies.

332
00:15:49,400 --> 00:15:49,800
Speaker 1: Get on that.

333
00:15:50,600 --> 00:15:54,280
Speaker 2: Just I also think that people I've just did an

334
00:15:54,320 --> 00:15:56,920
Speaker 2: episode about this, Actually, I think people also assume some

335
00:15:57,040 --> 00:15:59,320
Speaker 2: too big to fail thing will happen, even though that's

336
00:15:59,440 --> 00:16:02,480
Speaker 2: just too big to vote. And the Great Financial Crisis

337
00:16:02,520 --> 00:16:04,840
Speaker 2: was so much bigger and so much worse, and also

338
00:16:05,080 --> 00:16:06,080
Speaker 2: very different.

339
00:16:06,760 --> 00:16:09,360
Speaker 1: Like no, the financial scale was very different, which is

340
00:16:09,360 --> 00:16:11,000
Speaker 1: funny because you know, early on in this one of

341
00:16:11,040 --> 00:16:13,840
Speaker 1: the things that I found most striking was how quickly,

342
00:16:13,960 --> 00:16:17,080
Speaker 1: so initially, people were arguing to me that Paul, don't

343
00:16:17,080 --> 00:16:19,360
Speaker 1: worry your pretty little ahead about this, because these are

344
00:16:19,400 --> 00:16:21,240
Speaker 1: big boys and girls who are spending all this money.

345
00:16:21,280 --> 00:16:24,600
Speaker 1: The hyperscalers, the Microsoft's Amazon's Google and everything else. What

346
00:16:24,640 --> 00:16:27,040
Speaker 1: they spend on their money on should be no concern

347
00:16:27,120 --> 00:16:29,320
Speaker 1: to you, because these are big and profitable companies. Well

348
00:16:29,360 --> 00:16:31,400
Speaker 1: leaving a side open air on anthropic, but that are

349
00:16:31,400 --> 00:16:33,640
Speaker 1: big and profitable companies. How are you to tell them

350
00:16:33,640 --> 00:16:35,960
Speaker 1: what to do? And then I was so I said,

351
00:16:35,960 --> 00:16:40,000
Speaker 1: you know, fine, it's so much much. It's my job,

352
00:16:40,080 --> 00:16:42,480
Speaker 1: but fine. But then it ripped rapidly changed, right because

353
00:16:42,520 --> 00:16:45,160
Speaker 1: halfway through last year, maybe in the second quarter of

354
00:16:45,280 --> 00:16:48,480
Speaker 1: last year, suddenly the cash needs of building out these

355
00:16:48,600 --> 00:16:51,720
Speaker 1: data centers exceeded the cash flow I should say, the

356
00:16:51,800 --> 00:16:54,680
Speaker 1: unencumbered cash flow of these large and profitable companies, because

357
00:16:54,680 --> 00:16:56,600
Speaker 1: they have other uses for cash litle. So you started

358
00:16:57,320 --> 00:17:00,680
Speaker 1: moving more and more towards external sources of capital, private credit,

359
00:17:00,760 --> 00:17:03,920
Speaker 1: most notoriously, but also a host of other special purpose

360
00:17:04,000 --> 00:17:07,119
Speaker 1: vehicles and other off balance sheet financing structures, to the

361
00:17:07,160 --> 00:17:09,400
Speaker 1: point that by the end of the year of twenty

362
00:17:09,440 --> 00:17:12,960
Speaker 1: twenty five, a data center related debt was the largest

363
00:17:13,080 --> 00:17:15,320
Speaker 1: chunk of investment grade debt issued in the US, and

364
00:17:15,320 --> 00:17:19,159
Speaker 1: twenty twenty five tech used to be the no debt sector.

365
00:17:19,440 --> 00:17:22,359
Speaker 1: They went from no debt to the largest issuer of

366
00:17:22,480 --> 00:17:25,560
Speaker 1: investment grade and non investment grade debt in the United

367
00:17:25,600 --> 00:17:28,280
Speaker 1: States last year, and of course all the way along

368
00:17:28,359 --> 00:17:30,680
Speaker 1: people normalize it and say, you know, it's fine, they're

369
00:17:30,800 --> 00:17:32,720
Speaker 1: very profitable. And then when it turned out the profits

370
00:17:32,760 --> 00:17:35,040
Speaker 1: weren't enough to pay for all the data centers, that's

371
00:17:35,080 --> 00:17:37,399
Speaker 1: also fine, right, So we buill did these things end

372
00:17:37,520 --> 00:17:39,400
Speaker 1: up being fine? And then of course it all came

373
00:17:39,440 --> 00:17:40,840
Speaker 1: home to roost to a degree at the end of

374
00:17:40,880 --> 00:17:43,600
Speaker 1: the year as private credit started sort of got attack

375
00:17:43,680 --> 00:17:46,000
Speaker 1: from through the side door because of this problems with

376
00:17:46,200 --> 00:17:48,600
Speaker 1: these very large positions they held in software as the

377
00:17:48,640 --> 00:17:50,840
Speaker 1: service companies, which it turned out were at least seen

378
00:17:50,960 --> 00:17:53,560
Speaker 1: is threatened by AI. So that's the first shoe to

379
00:17:53,640 --> 00:17:55,760
Speaker 1: drop on this stuff. But there's still another shoe to drop,

380
00:17:55,800 --> 00:17:59,600
Speaker 1: which is the overexposure of private credit to data center

381
00:17:59,640 --> 00:18:02,399
Speaker 1: related debt, because if you think about it, it's not

382
00:18:02,560 --> 00:18:04,960
Speaker 1: the consequentiality of it. It is something like the global

383
00:18:05,040 --> 00:18:08,840
Speaker 1: financial crisis. The more entertaining part of this is they're

384
00:18:08,960 --> 00:18:12,040
Speaker 1: treating data centers as real estate. They look at data

385
00:18:12,080 --> 00:18:14,600
Speaker 1: centers as being like apartment buildings with who the hell

386
00:18:14,720 --> 00:18:17,280
Speaker 1: knows what's going on inside, but they're good for the rent, right,

387
00:18:17,520 --> 00:18:18,760
Speaker 1: So this is the way they look at at this

388
00:18:18,840 --> 00:18:22,119
Speaker 1: way look at data centers. But the problem is the

389
00:18:22,320 --> 00:18:28,200
Speaker 1: thing that's generating the income is inherently deflationary, hyperdflationary, falling

390
00:18:28,240 --> 00:18:30,480
Speaker 1: seventy to eighty percent year over year. These are tokens.

391
00:18:30,840 --> 00:18:33,840
Speaker 1: So the idea that you can you're having to pay

392
00:18:33,880 --> 00:18:37,280
Speaker 1: a fixed obligation these notes that have been issued with

393
00:18:37,320 --> 00:18:39,720
Speaker 1: respect to the debt to finance data centers with a

394
00:18:39,880 --> 00:18:42,560
Speaker 1: thing that's following seventy to eighty percent year where year

395
00:18:42,560 --> 00:18:44,840
Speaker 1: in price. Just try and run an auto company that

396
00:18:44,920 --> 00:18:46,400
Speaker 1: way with a significant depation.

397
00:18:47,160 --> 00:18:50,640
Speaker 2: But the price of tokens coming down wouldn't affect GPU

398
00:18:50,800 --> 00:18:53,760
Speaker 2: compute in that way though. Also that price coming down

399
00:18:53,880 --> 00:18:57,280
Speaker 2: isn't necessarily a result of cost savings, but it's a

400
00:18:57,320 --> 00:19:00,760
Speaker 2: result of the company's cutting the prices. Isn't the problem

401
00:19:00,840 --> 00:19:03,800
Speaker 2: that they're also full of these depreciating GPUs as well?

402
00:19:04,480 --> 00:19:06,800
Speaker 1: Yeah, yeah, So there's a double whammy. There's both sides

403
00:19:06,800 --> 00:19:10,120
Speaker 1: of it, right, So to a degree, if you're working

404
00:19:10,160 --> 00:19:12,560
Speaker 1: with a model directly through the API, which the largest

405
00:19:12,600 --> 00:19:15,639
Speaker 1: issuers are, or at least if you're in a production position,

406
00:19:15,880 --> 00:19:18,440
Speaker 1: you're actually paying an API price ten which is metered

407
00:19:18,480 --> 00:19:20,400
Speaker 1: at the token level, so you do see those token

408
00:19:20,440 --> 00:19:23,359
Speaker 1: prices directly. And then secondarily you have the problem that

409
00:19:23,840 --> 00:19:28,280
Speaker 1: the GPUs themselves, insofar as they are capital investment around

410
00:19:28,320 --> 00:19:32,119
Speaker 1: which the investments predicated, the depreciation of those items is

411
00:19:32,160 --> 00:19:34,359
Speaker 1: relatively rapid to it. And the analogy I always make

412
00:19:34,400 --> 00:19:36,679
Speaker 1: to people is that it really depends on what they

413
00:19:36,720 --> 00:19:38,920
Speaker 1: were used for historically. So if they're just being used

414
00:19:38,960 --> 00:19:41,880
Speaker 1: for inference to a degree, they have a longer lifespan.

415
00:19:42,280 --> 00:19:44,320
Speaker 1: But if they were ever used for training, which is

416
00:19:44,400 --> 00:19:47,359
Speaker 1: like flat out pedal to the floor twenty four to seven,

417
00:19:47,640 --> 00:19:50,199
Speaker 1: huge workload, It's kind of like, you know, you had

418
00:19:50,240 --> 00:19:52,520
Speaker 1: a car that was only driven to church on Sundays

419
00:19:52,960 --> 00:19:55,159
Speaker 1: and a car that was raised at Lamon's one weekend.

420
00:19:55,280 --> 00:19:57,240
Speaker 1: They have the same number of miles. I know which

421
00:19:57,280 --> 00:19:59,800
Speaker 1: car I want. The same thing applies to GPUs.

422
00:20:00,840 --> 00:20:03,280
Speaker 2: Yeah. And the other thing as well is I've really

423
00:20:03,320 --> 00:20:05,399
Speaker 2: been looking at this. So you've you brought up the

424
00:20:05,440 --> 00:20:08,199
Speaker 2: wood Mac study, which was awesome. I don't know if

425
00:20:08,240 --> 00:20:10,240
Speaker 2: you saw the Siteline climate one where it was like,

426
00:20:10,680 --> 00:20:13,320
Speaker 2: of the sixteen gigabs that were meant to come online

427
00:20:13,359 --> 00:20:17,000
Speaker 2: this year, only five are actually under construction. I think

428
00:20:17,080 --> 00:20:19,800
Speaker 2: that there's a big problem with just the speed of

429
00:20:19,880 --> 00:20:23,280
Speaker 2: the rollout and the upgrade cycle, because we are still

430
00:20:23,359 --> 00:20:26,440
Speaker 2: going to be installing Blackwell GPUs into twenty twenty seven,

431
00:20:26,760 --> 00:20:29,960
Speaker 2: if not twenty twenty eight. That's insane that like that

432
00:20:30,359 --> 00:20:33,240
Speaker 2: it I worked it out as it's like it takes

433
00:20:33,880 --> 00:20:36,680
Speaker 2: six months to install a single quarters worth of GPUs.

434
00:20:37,160 --> 00:20:40,639
Speaker 2: But it's like, how at some point Nvidio has to

435
00:20:40,720 --> 00:20:43,280
Speaker 2: slow not even because of me wanting it to or not,

436
00:20:43,760 --> 00:20:44,600
Speaker 2: but because.

437
00:20:45,359 --> 00:20:48,160
Speaker 3: Where are they going? Like where are we putting these things?

438
00:20:48,800 --> 00:20:52,240
Speaker 3: I mean Taiwan warehouses, I guess, yeah.

439
00:20:52,119 --> 00:20:53,000
Speaker 1: Tiewe warehouses.

440
00:20:53,400 --> 00:20:53,600
Speaker 3: Yeah.

441
00:20:53,680 --> 00:20:55,680
Speaker 1: And that's a big problem, is the build out. So

442
00:20:55,840 --> 00:21:01,280
Speaker 1: this this problem of and of the Northeastern I forgotten

443
00:21:01,280 --> 00:21:02,960
Speaker 1: WHI Utility in the Northeast just recently put out some

444
00:21:03,080 --> 00:21:05,479
Speaker 1: data on this showing that something like them data U suggests,

445
00:21:05,480 --> 00:21:07,320
Speaker 1: which is like twenty five percent of the total commit

446
00:21:07,800 --> 00:21:11,280
Speaker 1: was actually ever produced and likely will ever be built out.

447
00:21:11,640 --> 00:21:13,280
Speaker 1: And that's in part because a lot of these things

448
00:21:13,280 --> 00:21:15,760
Speaker 1: are speculative projects naming no names. There's a very large

449
00:21:15,800 --> 00:21:19,200
Speaker 1: Texas company that's doing this directly the career that is

450
00:21:19,240 --> 00:21:21,560
Speaker 1: a very speculative position. We're gonna we're going to power

451
00:21:21,600 --> 00:21:23,639
Speaker 1: it all behind the meter with nuclear reactors and all

452
00:21:23,680 --> 00:21:25,879
Speaker 1: these kinds of things. But this is a this is

453
00:21:25,960 --> 00:21:27,639
Speaker 1: this is a game we've seen back to I mean,

454
00:21:27,640 --> 00:21:29,600
Speaker 1: the analogy I make all the time is back to Chinatown.

455
00:21:29,680 --> 00:21:32,040
Speaker 1: This is like Chinatown right where I'm I'm buying up

456
00:21:32,080 --> 00:21:35,480
Speaker 1: real estate with numbered companies in hopes of securing water rights.

457
00:21:35,920 --> 00:21:36,359
Speaker 3: I saw this.

458
00:21:36,600 --> 00:21:38,480
Speaker 1: I saw this when Jack Nicholson was wandering around Ja

459
00:21:38,600 --> 00:21:44,040
Speaker 1: Eddies was wandering around in the deserts of California, Chinatown example.

460
00:21:44,560 --> 00:21:47,439
Speaker 1: So what's happening is increasingly a lot of what's going

461
00:21:47,480 --> 00:21:49,920
Speaker 1: on under the hood here that's creating the impression of

462
00:21:50,000 --> 00:21:52,480
Speaker 1: a build out that doesn't exist. Are these things called

463
00:21:52,520 --> 00:21:56,960
Speaker 1: powered land companies? So powered land companies are these speculators

464
00:21:56,960 --> 00:21:59,840
Speaker 1: they don't call themselves that, who look for strategic location.

465
00:22:00,240 --> 00:22:03,240
Speaker 1: We're using numbered companies. They can purchase real estate that

466
00:22:03,359 --> 00:22:06,479
Speaker 1: has access to peering points, so high speed interconnection.

467
00:22:06,000 --> 00:22:07,679
Speaker 3: To the what is a number of company as well?

468
00:22:07,680 --> 00:22:08,399
Speaker 3: I'm really sorry.

469
00:22:09,240 --> 00:22:11,560
Speaker 1: So a company that doesn't make it obvious to who

470
00:22:11,640 --> 00:22:13,840
Speaker 1: the actual direct owners are, so it's more clear at

471
00:22:13,840 --> 00:22:15,720
Speaker 1: what their purpose is. So I you know, think of

472
00:22:15,760 --> 00:22:16,840
Speaker 1: it as like Cayman Islands.

473
00:22:16,880 --> 00:22:17,359
Speaker 3: But it doesn't.

474
00:22:17,520 --> 00:22:19,560
Speaker 1: So the idea that you're trying, you're trying to at

475
00:22:19,640 --> 00:22:22,760
Speaker 1: least loosely obfuscate who the what the ownership and purpose are.

476
00:22:23,200 --> 00:22:25,840
Speaker 1: But even that's less important. The idea though, that they're

477
00:22:25,920 --> 00:22:29,240
Speaker 1: buying on a speculative what basis This tracts of land

478
00:22:29,880 --> 00:22:32,000
Speaker 1: could be you know, hundreds of acres in some location

479
00:22:32,440 --> 00:22:35,440
Speaker 1: that has access to power, access to water, and potentially

480
00:22:35,480 --> 00:22:37,880
Speaker 1: access to a peering point, a high speed interconnection point

481
00:22:37,920 --> 00:22:40,280
Speaker 1: to the broader Internet. And then they then they lock

482
00:22:40,400 --> 00:22:42,680
Speaker 1: that up. And then they go out and say, okay,

483
00:22:42,720 --> 00:22:44,760
Speaker 1: I've got this position. You guys need to build out

484
00:22:44,800 --> 00:22:47,360
Speaker 1: more data center capacity, talking to the hyperscalers.

485
00:22:47,760 --> 00:22:48,600
Speaker 3: Look at me, you got.

486
00:22:48,520 --> 00:22:50,520
Speaker 1: Nowhere else to go. I've locked all this up early on,

487
00:22:50,800 --> 00:22:53,119
Speaker 1: kind of like locking up the water for the orange grows?

488
00:22:53,640 --> 00:22:56,679
Speaker 3: Right? Is that is that widespread? Is that widespread?

489
00:22:56,800 --> 00:23:00,960
Speaker 2: Think that's so bad because I mean, this whole time

490
00:23:01,000 --> 00:23:03,480
Speaker 2: I've been looking for the speculative pop I will I

491
00:23:03,560 --> 00:23:06,359
Speaker 2: fully admit that's been it. If it's the land, that's

492
00:23:06,880 --> 00:23:09,679
Speaker 2: not great for anyone involved. But even then though gpu

493
00:23:09,800 --> 00:23:12,439
Speaker 2: is still being sold like that, that's the real.

494
00:23:12,400 --> 00:23:15,040
Speaker 1: So there's a game, right. But part of the problem is,

495
00:23:15,040 --> 00:23:17,159
Speaker 1: and there's was a great piece from trend Force, I

496
00:23:17,240 --> 00:23:18,680
Speaker 1: think it was the other day, one of the market

497
00:23:18,840 --> 00:23:21,680
Speaker 1: research firms. In this they're pointing out how how how

498
00:23:21,760 --> 00:23:24,720
Speaker 1: many firms are now buying LTA's long term purchase agreements

499
00:23:25,240 --> 00:23:28,000
Speaker 1: because they they're being told that if they don't lock

500
00:23:28,080 --> 00:23:30,960
Speaker 1: in demand now, lock in now, they won't get product

501
00:23:31,000 --> 00:23:33,760
Speaker 1: in two years. So this is the other layer that

502
00:23:33,840 --> 00:23:36,280
Speaker 1: a lot of what you're seeing is purchasing isn't completely

503
00:23:36,359 --> 00:23:39,560
Speaker 1: speculative by people who are worried that they don't lock

504
00:23:39,640 --> 00:23:41,520
Speaker 1: in a long term purchase agreement now, they will never

505
00:23:41,560 --> 00:23:42,919
Speaker 1: get supplied later, so.

506
00:23:42,960 --> 00:23:43,760
Speaker 3: That their approaches.

507
00:23:44,000 --> 00:23:46,399
Speaker 1: I'll worry, I'll worry my about that other stuff later on,

508
00:23:46,560 --> 00:23:48,040
Speaker 1: but for now I need to sign up. So that's

509
00:23:48,119 --> 00:23:50,000
Speaker 1: the other piece that a lot of people miss here

510
00:23:50,119 --> 00:23:52,120
Speaker 1: is a lot of the demand now is increasingly tied

511
00:23:52,160 --> 00:23:54,119
Speaker 1: into these sorts of long term purchase agreements which have

512
00:23:54,240 --> 00:23:55,920
Speaker 1: nothing to do with actual units being shipped.

513
00:23:56,480 --> 00:23:59,360
Speaker 3: The mob boss thing again, Yeah yeah, yeah.

514
00:23:59,720 --> 00:24:02,639
Speaker 1: But I was still getting out. I told someone this

515
00:24:02,720 --> 00:24:04,080
Speaker 1: the other day. Who was I talking to you? I

516
00:24:04,080 --> 00:24:05,359
Speaker 1: have the Wall three journal I was talking that was

517
00:24:05,359 --> 00:24:07,080
Speaker 1: saying like it's a it's kind of like the old

518
00:24:07,200 --> 00:24:10,080
Speaker 1: mafia threat, Like really nice AI market you have there

519
00:24:10,640 --> 00:24:11,600
Speaker 1: if something happened.

520
00:24:11,359 --> 00:24:14,920
Speaker 2: To it, Oh you want you don't, I guess you

521
00:24:15,040 --> 00:24:16,840
Speaker 2: don't want Vera Rubin anymore.

522
00:24:17,400 --> 00:24:18,480
Speaker 3: I guess we'll have to give.

523
00:24:19,720 --> 00:24:22,159
Speaker 1: It's exactly like that. So this is the thing. So

524
00:24:22,280 --> 00:24:25,040
Speaker 1: those two pieces are really important because it creates the

525
00:24:25,119 --> 00:24:28,560
Speaker 1: impression of unit growth where unit growth doesn't actually exist

526
00:24:28,600 --> 00:24:31,440
Speaker 1: because it's predicated on locking supply in the hopes of

527
00:24:31,520 --> 00:24:33,600
Speaker 1: something later. But at the other side of it, you've

528
00:24:33,600 --> 00:24:36,000
Speaker 1: also got the speculative lamb component. There was a company

529
00:24:36,080 --> 00:24:38,160
Speaker 1: the other day, there was a great Bloomberg story about

530
00:24:38,200 --> 00:24:42,119
Speaker 1: it who raised I think it was three billion dollars

531
00:24:42,200 --> 00:24:45,280
Speaker 1: in junk bonds for exactly this purpose. So this is

532
00:24:45,480 --> 00:24:49,400
Speaker 1: highly speculative stuff that's actually maybe no, no, not terrible.

533
00:24:49,440 --> 00:24:50,920
Speaker 1: So just literally last week, uh.

534
00:24:51,560 --> 00:24:53,000
Speaker 3: It was called tracked.

535
00:24:53,040 --> 00:24:54,399
Speaker 1: I think it was called t R A C T.

536
00:24:54,720 --> 00:24:58,200
Speaker 3: Yes, tracked. I remember this one. This is the this

537
00:24:58,359 --> 00:24:59,080
Speaker 3: is the insane thing.

538
00:24:59,160 --> 00:25:01,800
Speaker 2: They're still able to set like raise those bones though,

539
00:25:02,080 --> 00:25:03,760
Speaker 2: like they're still able to get the money.

540
00:25:04,400 --> 00:25:06,119
Speaker 1: Well that's because again this is this, this is the

541
00:25:06,200 --> 00:25:09,360
Speaker 1: insidious problem. Here is there's this idea that if I'm successful,

542
00:25:09,920 --> 00:25:13,960
Speaker 1: my counterparty, the counterparty in the data centers, they're good

543
00:25:14,040 --> 00:25:16,520
Speaker 1: for it because Microsoft Google. So instead of having a

544
00:25:16,600 --> 00:25:19,200
Speaker 1: bunch of dodgy you know, Florida strippers or something on

545
00:25:19,240 --> 00:25:21,280
Speaker 1: the other side of this, like in the financial crisis,

546
00:25:21,640 --> 00:25:24,160
Speaker 1: what I've got on the other side, my counterparty has

547
00:25:24,240 --> 00:25:27,159
Speaker 1: a high credit, very focused, very small group. It's the

548
00:25:27,240 --> 00:25:29,840
Speaker 1: Microsofts and Googles and others. So people are willing to

549
00:25:29,920 --> 00:25:33,800
Speaker 1: take much crazier risks because the counter party looks so

550
00:25:33,960 --> 00:25:36,680
Speaker 1: good from a credit standpoint, which is very different from

551
00:25:36,720 --> 00:25:38,520
Speaker 1: what happened in the financial crisis, where I wouldn't have

552
00:25:38,560 --> 00:25:40,920
Speaker 1: issued junk to create something that was going to be

553
00:25:40,960 --> 00:25:44,320
Speaker 1: purchased by you know who knows. But if it's Microsoft,

554
00:25:44,400 --> 00:25:46,159
Speaker 1: Google and the other hyper skills the other side, I'm like,

555
00:25:46,200 --> 00:25:48,680
Speaker 1: you know, what, if this works, they're good for it.

556
00:25:49,400 --> 00:25:51,920
Speaker 1: But what if it doesn't work, well, then of course

557
00:25:51,960 --> 00:25:53,439
Speaker 1: you went up with a lot of room, a lot

558
00:25:53,480 --> 00:25:55,600
Speaker 1: of extra buildings for you know, laser Tag or something

559
00:25:55,680 --> 00:25:55,800
Speaker 1: like that.

560
00:25:56,320 --> 00:26:00,080
Speaker 2: I'd I'm excited about the Laser Tag arena arena in

561
00:26:00,119 --> 00:26:02,840
Speaker 2: the future, we have just just America's the Laser Tag

562
00:26:02,960 --> 00:26:04,400
Speaker 2: Capital of the world.

563
00:26:04,480 --> 00:26:06,680
Speaker 1: That's right, We've got a lot of extra space for

564
00:26:07,160 --> 00:26:23,800
Speaker 1: use storets and laser tag Because again, the only thing

565
00:26:23,840 --> 00:26:28,000
Speaker 1: we know for sure is that the the efficiency of inference.

566
00:26:28,080 --> 00:26:30,560
Speaker 1: If you buy the argument that we're transitioning rapidly to inference,

567
00:26:30,600 --> 00:26:33,440
Speaker 1: the efficiency the inference is going on is rising rapidly

568
00:26:33,560 --> 00:26:36,679
Speaker 1: because of things like distillation. Models are shrinking, chips are

569
00:26:36,720 --> 00:26:40,160
Speaker 1: becoming more efficient, there's less less memory.

570
00:26:39,920 --> 00:26:43,680
Speaker 2: Required, because that's something I've The people that have told

571
00:26:43,720 --> 00:26:46,440
Speaker 2: me inference is getting more efficient are usually referring to

572
00:26:46,560 --> 00:26:48,400
Speaker 2: Nvidia demos based entirely.

573
00:26:48,200 --> 00:26:50,160
Speaker 3: On d Yeah, that's a ridiculous question.

574
00:26:51,040 --> 00:26:55,679
Speaker 1: Look at companies like Fractile out of the UK, an

575
00:26:55,720 --> 00:26:59,400
Speaker 1: interesting company with I think some really groundbreaking inference technology

576
00:26:59,440 --> 00:27:01,160
Speaker 1: that will be ship this year. Look at a company

577
00:27:01,240 --> 00:27:06,919
Speaker 1: like Talus in Torontol. Again, so Talus is a good example.

578
00:27:07,000 --> 00:27:10,280
Speaker 1: They're doing some innovative stuff showing So a high speed

579
00:27:10,359 --> 00:27:13,280
Speaker 1: inference chip today might do a few hundred tokens per second,

580
00:27:13,359 --> 00:27:15,879
Speaker 1: that's so considered a relative leaving aside the wattage required,

581
00:27:16,200 --> 00:27:20,200
Speaker 1: that's a relatively high speed token UH producing chip. So

582
00:27:20,640 --> 00:27:23,760
Speaker 1: Talus is demoing sixteen thousand tokens per second so we're

583
00:27:23,800 --> 00:27:27,359
Speaker 1: seeing step function increases at lower power from some of

584
00:27:27,440 --> 00:27:30,840
Speaker 1: these next generations silicon vendors. Granted, they're not going to

585
00:27:30,880 --> 00:27:32,960
Speaker 1: dominate the marketplace, but the idea that we're not going

586
00:27:33,040 --> 00:27:36,320
Speaker 1: to see step function changes that we can project into

587
00:27:36,359 --> 00:27:38,000
Speaker 1: the future based on what we've seen in the recent

588
00:27:38,080 --> 00:27:39,400
Speaker 1: past is just dead rock.

589
00:27:40,119 --> 00:27:44,080
Speaker 2: But the thing is what models are they able to

590
00:27:44,160 --> 00:27:46,360
Speaker 2: do that with? Because again, a lot of I mean,

591
00:27:46,440 --> 00:27:48,880
Speaker 2: all of the benchmarks we see are based on open

592
00:27:48,920 --> 00:27:52,320
Speaker 2: source models, because open source models are open sourcest ones

593
00:27:52,359 --> 00:27:53,000
Speaker 2: they can test on.

594
00:27:53,440 --> 00:27:55,840
Speaker 3: I feel like everything every time I get any kind.

595
00:27:55,720 --> 00:27:59,960
Speaker 2: Of leak out of azore or AWS about these models,

596
00:28:00,320 --> 00:28:03,240
Speaker 2: it's I have a Microsoft personally topic.

597
00:28:03,280 --> 00:28:04,439
Speaker 3: It's like four to.

598
00:28:04,600 --> 00:28:08,800
Speaker 2: Six, sorry, four to twelve GPUs, one generation of the

599
00:28:09,760 --> 00:28:13,400
Speaker 2: smaller reasoning model. Oh four I think it was, Yeah,

600
00:28:13,480 --> 00:28:16,960
Speaker 2: And that's just for one generation over several minutes that

601
00:28:17,359 --> 00:28:20,640
Speaker 2: someone's doing a particularly different coding task. It doesn't feel

602
00:28:20,760 --> 00:28:24,600
Speaker 2: like these inference chain things will trickle down there, So

603
00:28:24,720 --> 00:28:26,920
Speaker 2: maybe it will be the future of llm's. Is this

604
00:28:27,040 --> 00:28:30,240
Speaker 2: small industry run on these smaller chips or something like this.

605
00:28:30,560 --> 00:28:32,480
Speaker 1: I think you're gonna see all of the above thing. Well,

606
00:28:32,520 --> 00:28:35,040
Speaker 1: let's say, for example, we're seeing inference happening inside of

607
00:28:35,320 --> 00:28:38,760
Speaker 1: evs where it's we're doing. I'm doing rapid ingestion of

608
00:28:38,840 --> 00:28:41,160
Speaker 1: video tokens for the purpose of deciding whether I'm going

609
00:28:41,200 --> 00:28:44,560
Speaker 1: to back into my neighbor's trash can and these kinds

610
00:28:44,600 --> 00:28:47,680
Speaker 1: of things. Language models, Oh, absolutely, because you can you

611
00:28:47,760 --> 00:28:50,480
Speaker 1: can imagine I'm ingesting all of that the video. I'm

612
00:28:50,520 --> 00:28:52,760
Speaker 1: gonna have a lot more granularity with respected the tokens

613
00:28:52,800 --> 00:28:54,280
Speaker 1: flowing back to me, and I can do more with

614
00:28:54,440 --> 00:28:56,560
Speaker 1: it and know more about what's happening in my environment.

615
00:28:56,880 --> 00:29:00,800
Speaker 1: I think most of video ingestion is gonna move towards tokens,

616
00:29:01,120 --> 00:29:04,160
Speaker 1: but done on small language models, very very small stuff.

617
00:29:04,880 --> 00:29:08,440
Speaker 2: We've already had edge compute AI vision stuff for like

618
00:29:08,960 --> 00:29:09,880
Speaker 2: a decade or more.

619
00:29:10,040 --> 00:29:11,320
Speaker 3: Like I was working on stuff like that in the

620
00:29:11,400 --> 00:29:12,160
Speaker 3: twenty fourteen.

621
00:29:12,600 --> 00:29:14,880
Speaker 2: It just feels like a lot of this is trying

622
00:29:14,960 --> 00:29:19,080
Speaker 2: to make tokens through stuff that we already kind of do.

623
00:29:20,400 --> 00:29:23,800
Speaker 1: It is but the thing I and again you know,

624
00:29:23,920 --> 00:29:26,360
Speaker 1: I'm deeply in this skeptical camp here. The thing I

625
00:29:26,440 --> 00:29:29,600
Speaker 1: will say in favor of tokens in terms of absorbing

626
00:29:29,600 --> 00:29:31,800
Speaker 1: a lot of this stuff is that it makes it

627
00:29:31,840 --> 00:29:34,600
Speaker 1: all less ad hoc because you now have this sort

628
00:29:34,600 --> 00:29:38,880
Speaker 1: of universal commodity for ingestion and production of information. That's

629
00:29:39,080 --> 00:29:41,280
Speaker 1: that makes things a little more interesting because I can

630
00:29:41,320 --> 00:29:45,040
Speaker 1: now abstract away some of the hard problems of video processing.

631
00:29:45,120 --> 00:29:47,080
Speaker 1: I can abstract away some of the hard problems with

632
00:29:47,160 --> 00:29:50,720
Speaker 1: speech synthesis because they all kind of disappear and become

633
00:29:51,040 --> 00:29:53,720
Speaker 1: in a sense of this universal token. And you know,

634
00:29:53,800 --> 00:29:55,680
Speaker 1: to a degree, I think that's true, and I think

635
00:29:55,720 --> 00:29:58,240
Speaker 1: it will lower the barriers to more people playing in

636
00:29:58,280 --> 00:30:00,520
Speaker 1: the worlds of you know, video and speech and other places.

637
00:30:00,560 --> 00:30:02,840
Speaker 1: But that doesn't create the kind of marketplace that people

638
00:30:02,840 --> 00:30:06,080
Speaker 1: who are pushing huge numbers of tokens run on frontier models.

639
00:30:06,120 --> 00:30:08,480
Speaker 3: Why this is just edge stuff, cheap and dirty.

640
00:30:08,720 --> 00:30:11,840
Speaker 2: So the problem with tokens as a commodity as well

641
00:30:12,080 --> 00:30:14,320
Speaker 2: is it's very hard to know, like one toe like

642
00:30:14,440 --> 00:30:19,040
Speaker 2: a million tokens per million tokens. It's impossible to actually

643
00:30:19,120 --> 00:30:23,760
Speaker 2: measure how longer tasks, how many tokens a task will

644
00:30:23,880 --> 00:30:27,440
Speaker 2: use because of the inherent I'm reliable of large language models,

645
00:30:27,520 --> 00:30:29,840
Speaker 2: and it's like so it's hard to weave it. Like

646
00:30:29,960 --> 00:30:31,760
Speaker 2: right now you're seeing with this, have you seen this

647
00:30:32,120 --> 00:30:34,360
Speaker 2: people complaining about anthropics rate limits.

648
00:30:34,120 --> 00:30:35,920
Speaker 3: For example, I'm not sure all the time I see it.

649
00:30:37,160 --> 00:30:38,920
Speaker 2: Well, right now people are mad that they can only

650
00:30:38,960 --> 00:30:40,760
Speaker 2: spend a thousand dollars on a two hundred dollars a

651
00:30:40,760 --> 00:30:44,120
Speaker 2: month plan. But the thing with that is people very

652
00:30:44,240 --> 00:30:46,840
Speaker 2: clearly do not know how much to how far a

653
00:30:46,920 --> 00:30:49,520
Speaker 2: token goes or a million tokens goes, like they it's

654
00:30:49,800 --> 00:30:53,440
Speaker 2: it's difficult to evaluate and measure that, which feels like

655
00:30:53,760 --> 00:30:55,880
Speaker 2: kind of economic poison at some point because if you

656
00:30:55,920 --> 00:30:58,880
Speaker 2: can't say how much a task will cost, you don't

657
00:30:58,920 --> 00:31:02,920
Speaker 2: have a miles per gallon, like sixteen thousand tokens per second. Well,

658
00:31:02,960 --> 00:31:06,320
Speaker 2: you could do inference fast, but if a customer can't

659
00:31:06,360 --> 00:31:10,320
Speaker 2: afford it, if a customer can't actually reliably say I'll

660
00:31:10,320 --> 00:31:12,960
Speaker 2: be able to use this in this way, what uses

661
00:31:13,040 --> 00:31:15,080
Speaker 2: tokens is a measure, but I mean I know what

662
00:31:15,160 --> 00:31:17,080
Speaker 2: they used for a measurement. But it's like you can't

663
00:31:17,120 --> 00:31:18,720
Speaker 2: say what even a million tokens might do.

664
00:31:19,560 --> 00:31:21,200
Speaker 1: No, you can't. But that's kind of innate to the

665
00:31:21,240 --> 00:31:23,800
Speaker 1: world of like a true a true commodity. I don't

666
00:31:23,920 --> 00:31:26,080
Speaker 1: know how much you're going to put in your car.

667
00:31:26,320 --> 00:31:29,040
Speaker 1: I don't know it. Really that becomes an engineering decision

668
00:31:29,080 --> 00:31:30,680
Speaker 1: for you, not a production.

669
00:31:30,440 --> 00:31:33,640
Speaker 3: Decision for me. Right, So yeah, that makes that you

670
00:31:33,720 --> 00:31:34,800
Speaker 3: have to sumporate.

671
00:31:34,840 --> 00:31:37,800
Speaker 1: Those two pieces and so that I don't can't tell you,

672
00:31:37,960 --> 00:31:39,560
Speaker 1: here's so many tokens it will take for you to

673
00:31:39,640 --> 00:31:41,720
Speaker 1: do X. That's no more my job than it is

674
00:31:41,800 --> 00:31:43,120
Speaker 1: for me to tell us that, you know, a copper

675
00:31:43,160 --> 00:31:44,920
Speaker 1: mind how many how much copper it is going to take.

676
00:31:44,920 --> 00:31:47,920
Speaker 3: For g I'm in a particular car, so know how much? Yeah,

677
00:31:48,160 --> 00:31:49,719
Speaker 3: but they know how much. I get what you mean.

678
00:31:49,760 --> 00:31:52,400
Speaker 3: It's like they know how much copper they need to use.

679
00:31:52,480 --> 00:31:55,920
Speaker 1: But then and then they'll make constrained, you know, constraining

680
00:31:55,960 --> 00:31:57,680
Speaker 1: decisions whether they'll say, I only want to use this

681
00:31:57,800 --> 00:32:00,160
Speaker 1: much copper because copper is really freaking expensive, and so

682
00:32:00,280 --> 00:32:01,680
Speaker 1: you know, we're gonna like I just saw I think

683
00:32:01,720 --> 00:32:03,640
Speaker 1: it was really in the other day, said they'd cut

684
00:32:03,680 --> 00:32:06,120
Speaker 1: out like I don't know, like ten miles of electric

685
00:32:06,200 --> 00:32:08,960
Speaker 1: cable inside their cars, which seemed ridiculous to me, but

686
00:32:09,040 --> 00:32:11,360
Speaker 1: it was again because of the price of copper. So

687
00:32:11,440 --> 00:32:13,800
Speaker 1: there's yeah, so they turn into and these things turn

688
00:32:13,840 --> 00:32:16,160
Speaker 1: into engineering decisions. Right now, we're in this kind of

689
00:32:16,200 --> 00:32:18,720
Speaker 1: subsidized wild wild West where everyone thinks it's a land

690
00:32:18,760 --> 00:32:21,080
Speaker 1: grab and they're so they're subsidizing it to a degree,

691
00:32:21,560 --> 00:32:24,080
Speaker 1: and people are over using tokens. Oh, I've had a

692
00:32:24,120 --> 00:32:26,640
Speaker 1: model running for three days and it's doing all these

693
00:32:26,680 --> 00:32:28,280
Speaker 1: agentic things and it's like, well, what are you trying

694
00:32:28,320 --> 00:32:29,800
Speaker 1: to create out of it? And it's like, I don't know,

695
00:32:29,880 --> 00:32:32,960
Speaker 1: it's some nonsensical things. So it's people are being subsidized

696
00:32:33,000 --> 00:32:36,280
Speaker 1: to do non economic behaviors to an incredible degree right now.

697
00:32:36,480 --> 00:32:38,880
Speaker 1: And I find that remarkable, which is a statement about

698
00:32:39,600 --> 00:32:42,640
Speaker 1: this kind of land grab mentality among the frontier model vendors,

699
00:32:42,680 --> 00:32:44,880
Speaker 1: and and Dariol has been real. Darioeminad in the Tropic

700
00:32:44,920 --> 00:32:46,880
Speaker 1: has been very upfront about this. Is he believes that

701
00:32:47,280 --> 00:32:48,920
Speaker 1: we are in a land grab mode. There's only going

702
00:32:48,960 --> 00:32:51,479
Speaker 1: to be a couple of frontier models vendors left standing,

703
00:32:51,760 --> 00:32:53,560
Speaker 1: and so we need to make sure that you know,

704
00:32:53,600 --> 00:32:56,160
Speaker 1: we're the dominant provider, if you will, of you know,

705
00:32:56,280 --> 00:32:58,680
Speaker 1: coding harnesses in frontier models. Now, I think that's something

706
00:32:58,760 --> 00:33:00,160
Speaker 1: Snowmer too, but that's another.

707
00:33:00,160 --> 00:33:01,560
Speaker 3: Yeah, I think.

708
00:33:01,800 --> 00:33:03,960
Speaker 2: But my core economic I mean one of the many

709
00:33:04,000 --> 00:33:07,200
Speaker 2: core economic things, such as it's totally unprofitable, My thing

710
00:33:07,320 --> 00:33:10,120
Speaker 2: right now is these rate limit changes are more severe

711
00:33:10,160 --> 00:33:13,120
Speaker 2: on an economic level than people give credit for just

712
00:33:13,240 --> 00:33:15,440
Speaker 2: because of the economics, but because.

713
00:33:15,200 --> 00:33:15,880
Speaker 3: Of the habits.

714
00:33:16,760 --> 00:33:19,360
Speaker 2: If you believe, like the way I analogize it is

715
00:33:19,440 --> 00:33:22,120
Speaker 2: like if your car can drive fifteen miles and then

716
00:33:22,160 --> 00:33:24,200
Speaker 2: one day it can only drive three, can you get

717
00:33:24,240 --> 00:33:24,560
Speaker 2: to work?

718
00:33:25,120 --> 00:33:27,440
Speaker 3: Yeah? And is it because if you.

719
00:33:27,560 --> 00:33:30,240
Speaker 1: Bought a house predicated on being able to be drive

720
00:33:30,320 --> 00:33:32,040
Speaker 1: the fifteen miles to work, right, So it has it

721
00:33:32,120 --> 00:33:33,960
Speaker 1: has externalities. This is outside.

722
00:33:33,600 --> 00:33:36,680
Speaker 2: Consequence, yes, but the thing is you're paying. They've trained

723
00:33:36,800 --> 00:33:39,520
Speaker 2: everyone in this land grab to act in a way

724
00:33:39,600 --> 00:33:42,720
Speaker 2: that doesn't make sense long term. And I'm not I

725
00:33:42,760 --> 00:33:45,360
Speaker 2: don't think they can become profitable. I actually truly don't

726
00:33:45,400 --> 00:33:47,240
Speaker 2: think that there's an economic way for it to happen.

727
00:33:48,360 --> 00:33:50,640
Speaker 2: But they've trained people to use the product in a

728
00:33:50,680 --> 00:33:52,920
Speaker 2: way that doesn't make sense. Like it's not even a

729
00:33:53,360 --> 00:33:56,160
Speaker 2: oh they can charge more. Your habits are not built

730
00:33:56,240 --> 00:33:56,440
Speaker 2: for this.

731
00:33:56,720 --> 00:33:59,120
Speaker 1: That's exactly right. So I have a wild eyed theory.

732
00:33:59,160 --> 00:33:59,640
Speaker 1: Are you ready?

733
00:34:00,160 --> 00:34:00,600
Speaker 3: Don't go go?

734
00:34:02,000 --> 00:34:04,600
Speaker 1: So the IBI theory as the first of the first

735
00:34:04,640 --> 00:34:07,760
Speaker 1: frontier model company to abandon Frontier models wins.

736
00:34:08,960 --> 00:34:09,480
Speaker 2: What do you mean?

737
00:34:10,120 --> 00:34:12,919
Speaker 1: So my theory is that all of this stuff, most

738
00:34:12,960 --> 00:34:15,640
Speaker 1: people in a Pepsi Cooke challenge kind of way, can't

739
00:34:15,680 --> 00:34:17,919
Speaker 1: tell the difference. They claim they can, but they can't

740
00:34:17,920 --> 00:34:20,120
Speaker 1: actually tell the difference between most frontier models for a

741
00:34:20,200 --> 00:34:23,120
Speaker 1: typical test. Certainly normals can't. Coders claim they can, but

742
00:34:23,160 --> 00:34:24,839
Speaker 1: if you actually do it in a blind way, most

743
00:34:24,880 --> 00:34:26,719
Speaker 1: of them can't tell. This is just ego, and so

744
00:34:27,320 --> 00:34:30,640
Speaker 1: increasingly what people see instead is these coding harnesses, these

745
00:34:30,680 --> 00:34:32,840
Speaker 1: tools like quad code and codex and open code and

746
00:34:32,880 --> 00:34:34,560
Speaker 1: all these sorts of things. So most of the value

747
00:34:34,640 --> 00:34:36,360
Speaker 1: they see, and most of what they actually think of

748
00:34:36,440 --> 00:34:38,600
Speaker 1: as the model is just the harness, and that's where

749
00:34:38,640 --> 00:34:40,719
Speaker 1: most of the innovation is happening. So my argument is,

750
00:34:41,280 --> 00:34:43,399
Speaker 1: and that's why it was so dangerous Foraranthropic this week

751
00:34:43,440 --> 00:34:46,000
Speaker 1: when they accidentally did a whoops and released quad code.

752
00:34:46,320 --> 00:34:48,160
Speaker 1: Most of the values in the harness. And so the

753
00:34:48,239 --> 00:34:50,279
Speaker 1: first company to say, you know what, we don't need

754
00:34:50,360 --> 00:34:53,000
Speaker 1: to spend this kind of money anymore on training new models,

755
00:34:53,000 --> 00:34:55,200
Speaker 1: because we're going to just sit on top of models

756
00:34:55,239 --> 00:34:57,320
Speaker 1: from all of these loons who are out there spending

757
00:34:57,880 --> 00:35:00,440
Speaker 1: crazily on new frontier models that aren't improving very much

758
00:35:00,480 --> 00:35:02,160
Speaker 1: anymore now, certainly not like they were four or five

759
00:35:02,200 --> 00:35:05,200
Speaker 1: years ago, and that they will be rewarded for that,

760
00:35:05,360 --> 00:35:07,040
Speaker 1: no different than saying, you know, I've let go all

761
00:35:07,080 --> 00:35:09,359
Speaker 1: of my employees, or I've decided to stop spending money

762
00:35:09,400 --> 00:35:12,600
Speaker 1: on hydroelectric dams. You've cut cap X, you've cut You've

763
00:35:12,640 --> 00:35:15,520
Speaker 1: made your business more financially appealing by taking away the

764
00:35:15,560 --> 00:35:18,640
Speaker 1: single biggest piece of cost because you're recognizing that the

765
00:35:18,680 --> 00:35:21,680
Speaker 1: world has changed, that I'm not getting incrementally as much

766
00:35:21,800 --> 00:35:24,080
Speaker 1: value for dollar on training a model as I was

767
00:35:24,200 --> 00:35:26,399
Speaker 1: five years ago. And that's very clear in the data

768
00:35:26,440 --> 00:35:29,680
Speaker 1: if you look at any of the composite benchmark models

769
00:35:29,800 --> 00:35:31,840
Speaker 1: getting away from like can they solve math problems? But

770
00:35:31,960 --> 00:35:35,400
Speaker 1: literally real world composite models, it's been a sharp decline

771
00:35:35,480 --> 00:35:37,880
Speaker 1: from like, you know, eighteen percent year over year improvements

772
00:35:37,920 --> 00:35:40,040
Speaker 1: in models to like four or five percent at vastly

773
00:35:40,120 --> 00:35:42,960
Speaker 1: higher costs and more dollars and more time. So that

774
00:35:43,080 --> 00:35:46,760
Speaker 1: really matters, and so my completely will never happen. Theory

775
00:35:46,920 --> 00:35:48,560
Speaker 1: is that the winner here is the first one to

776
00:35:48,560 --> 00:35:49,120
Speaker 1: stop doing it.

777
00:35:49,960 --> 00:35:51,680
Speaker 3: Isn't that just describing cursor?

778
00:35:53,239 --> 00:35:56,000
Speaker 1: So cursor is an interesting examp. So cursor doesn't actually

779
00:35:56,560 --> 00:36:00,480
Speaker 1: embrace the full idea of being a harness across all

780
00:36:00,520 --> 00:36:02,759
Speaker 1: of these white collar applications. They're still kind of trapped

781
00:36:02,800 --> 00:36:04,839
Speaker 1: in a coding world, and so the idea So if

782
00:36:04,840 --> 00:36:07,600
Speaker 1: you're right, so I just think people get trapped in

783
00:36:07,640 --> 00:36:09,560
Speaker 1: coding because it was the first place this stuff emerged.

784
00:36:09,600 --> 00:36:11,680
Speaker 1: So you have to think about, like Cowork is a

785
00:36:11,719 --> 00:36:14,200
Speaker 1: good example of at least an attempt to break again,

786
00:36:14,360 --> 00:36:16,160
Speaker 1: escape containment and get out of the world of coding

787
00:36:16,280 --> 00:36:18,640
Speaker 1: and say, okay, this is actually for all white collar workers.

788
00:36:19,000 --> 00:36:21,719
Speaker 1: Like I watched people struggle with Claude pardon me.

789
00:36:22,360 --> 00:36:23,160
Speaker 3: But it didn't work.

790
00:36:23,520 --> 00:36:26,600
Speaker 1: It doesn't work, but it's the least it's directionally the

791
00:36:26,719 --> 00:36:29,480
Speaker 1: right idea if you buy my theory that all of

792
00:36:29,520 --> 00:36:31,960
Speaker 1: this stuff is commoditizing so fast and there's a loser's

793
00:36:32,040 --> 00:36:35,440
Speaker 1: game financially, that maybe the right sort of game theoretic

794
00:36:35,520 --> 00:36:38,360
Speaker 1: strategy is to be the first frontier model company to

795
00:36:38,360 --> 00:36:40,640
Speaker 1: stop making frontier models. And in a weird way, Apple

796
00:36:40,719 --> 00:36:43,040
Speaker 1: kind of showed the way right because early on they

797
00:36:43,080 --> 00:36:45,839
Speaker 1: were getting pillaried for why is an Apple spending more

798
00:36:45,880 --> 00:36:47,480
Speaker 1: on AI? Why does Apple not have a model? And

799
00:36:47,560 --> 00:36:49,640
Speaker 1: now of course it's reverse where it's like look at Apple,

800
00:36:49,719 --> 00:36:50,799
Speaker 1: They're so smart.

801
00:36:50,640 --> 00:36:53,279
Speaker 3: They're gonna come smart lail without that TAPEX. I don't know.

802
00:36:53,480 --> 00:36:55,440
Speaker 2: I just feel like I feel like what you wore

803
00:36:55,520 --> 00:36:59,560
Speaker 2: describing as just AI modal wrapper companies. Yes, but I

804
00:37:00,080 --> 00:37:04,480
Speaker 2: so My whole thing is unless someone is able to

805
00:37:04,520 --> 00:37:07,000
Speaker 2: break out of coding, there isn't really a hope for

806
00:37:07,080 --> 00:37:10,680
Speaker 2: any of this because to this point, every time I

807
00:37:10,800 --> 00:37:13,600
Speaker 2: read about an integration with like a Goldman sax or somewhere, yeah,

808
00:37:13,640 --> 00:37:15,319
Speaker 2: I can never actually find out what it does.

809
00:37:15,520 --> 00:37:18,359
Speaker 3: I can never the further you get into the region,

810
00:37:18,440 --> 00:37:19,160
Speaker 3: I have a good one for you.

811
00:37:19,320 --> 00:37:21,200
Speaker 1: So I was, I was talking to a very large

812
00:37:21,239 --> 00:37:23,640
Speaker 1: investment bank the other day about how about they're you know,

813
00:37:23,920 --> 00:37:27,160
Speaker 1: prodigious AI integration efforts, and so they built it out.

814
00:37:27,160 --> 00:37:30,600
Speaker 1: They they told me across equity research, sales trading, and

815
00:37:30,719 --> 00:37:33,800
Speaker 1: investment banking, and they asked me which one do you

816
00:37:33,920 --> 00:37:36,600
Speaker 1: think has seen the greatest benefit? And I said, oh god,

817
00:37:36,640 --> 00:37:38,160
Speaker 1: I said, I used to work on the cell side.

818
00:37:38,320 --> 00:37:40,120
Speaker 1: I said, my first instinct is none, that they're all

819
00:37:40,160 --> 00:37:42,600
Speaker 1: lying to you. But my second instinct is, I'll say

820
00:37:42,640 --> 00:37:45,759
Speaker 1: equity research because you know, no one likes to build spreadsheets,

821
00:37:45,760 --> 00:37:47,640
Speaker 1: and maybe it helps them build spreadsheets, and they said no.

822
00:37:48,560 --> 00:37:50,879
Speaker 1: So the answer, of course was investment banking. And I said,

823
00:37:50,920 --> 00:37:54,399
Speaker 1: why investment banking? These are like knuckle dragging dinosaur. Okaye, man,

824
00:37:55,000 --> 00:37:56,759
Speaker 1: what are they doing? And he said, the answer, of

825
00:37:56,800 --> 00:38:01,480
Speaker 1: course is the main thing junior investment bankers do is build. Well,

826
00:38:01,520 --> 00:38:03,080
Speaker 1: they get shouted at, that's the main thing they do.

827
00:38:03,200 --> 00:38:05,000
Speaker 1: But the second thing they do after being shouted at

828
00:38:05,320 --> 00:38:07,359
Speaker 1: is they build pitch decks for companies that don't want them.

829
00:38:07,719 --> 00:38:09,719
Speaker 1: So they build a pitch deck because a partner wants

830
00:38:09,719 --> 00:38:13,040
Speaker 1: a pitch deck built for some random company somewhere. And

831
00:38:13,440 --> 00:38:15,640
Speaker 1: that's a pain in the ass. And so now that

832
00:38:15,800 --> 00:38:18,040
Speaker 1: used to cause all these sleepless nights and blah blah blah,

833
00:38:18,160 --> 00:38:21,200
Speaker 1: they're doing them all with AI. So junior investment bankers

834
00:38:21,280 --> 00:38:24,640
Speaker 1: love AI because it lets them do this completely unproductive,

835
00:38:25,200 --> 00:38:28,560
Speaker 1: largely and consequential task of building pitch decks for companies

836
00:38:28,600 --> 00:38:30,600
Speaker 1: that don't want the pitch deck. And so that's These

837
00:38:30,640 --> 00:38:34,080
Speaker 1: are the kinds of applications that have really minimal economic

838
00:38:34,200 --> 00:38:37,359
Speaker 1: value and yet sort of superficially appear really exciting because

839
00:38:37,360 --> 00:38:38,960
Speaker 1: if I'm someone who otherwise had to stay up all

840
00:38:39,000 --> 00:38:41,480
Speaker 1: weekend building a PowerPoint deck to pitch to some random

841
00:38:41,520 --> 00:38:44,040
Speaker 1: small cap company and I'm like, yeah, this is terrific.

842
00:38:44,600 --> 00:38:47,000
Speaker 1: So that was the answer. It was really interesting, Yes.

843
00:38:47,400 --> 00:38:50,600
Speaker 2: But that's also kind of worrying because like, the best

844
00:38:50,719 --> 00:38:53,680
Speaker 2: example we have for this thing that has taken over

845
00:38:53,760 --> 00:38:59,320
Speaker 2: everything at least optically, even though asn't in economic terms,

846
00:38:59,480 --> 00:39:02,680
Speaker 2: is like, we can do powerpoints kind of for we.

847
00:39:02,719 --> 00:39:05,400
Speaker 1: Can do pop power points for junk bond raisers for

848
00:39:05,520 --> 00:39:07,600
Speaker 1: microcap companies way better than we used to.

849
00:39:08,600 --> 00:39:10,640
Speaker 3: Wow. And it's like helping junior analysts.

850
00:39:10,719 --> 00:39:13,920
Speaker 2: So it's like, are you really the time they're saving

851
00:39:14,040 --> 00:39:17,000
Speaker 2: is just lowering their workdays from fifteen hours to eleven.

852
00:39:17,239 --> 00:39:19,359
Speaker 1: Well, and they're still being shouted at unfortunately for them,

853
00:39:19,400 --> 00:39:20,520
Speaker 1: but you know, that's the way it goes.

854
00:39:20,600 --> 00:39:21,799
Speaker 3: That's that at the bottom of the job.

855
00:39:22,400 --> 00:39:25,400
Speaker 2: Part of the Joel Paul, It's been such a pleasure

856
00:39:25,480 --> 00:39:26,840
Speaker 2: having you. Where can people find you.

857
00:39:27,719 --> 00:39:28,719
Speaker 1: Pokadrowski dot com.

858
00:39:29,719 --> 00:39:31,960
Speaker 2: Thank you for joining us, and yes we'll be back

859
00:39:32,000 --> 00:39:33,920
Speaker 2: with the monologue this week. I'm of course ed Zeitron.

860
00:39:34,160 --> 00:39:44,480
Speaker 2: Thank you everyone for listening. Thank you for listening to

861
00:39:44,520 --> 00:39:47,480
Speaker 2: Better Offline, The editor and composer of the Better Offline

862
00:39:47,520 --> 00:39:50,160
Speaker 2: theme song is Matasowski. You can check out more of

863
00:39:50,239 --> 00:39:53,840
Speaker 2: his music and audio projects at Matasowski dot com, M.

864
00:39:53,880 --> 00:39:57,960
Speaker 3: A T T O S O W s ki dot com.

865
00:39:58,719 --> 00:40:01,239
Speaker 2: You can email me an easy Better offline dot com

866
00:40:01,440 --> 00:40:03,720
Speaker 2: or visit Better Offline dot com to find more podcast

867
00:40:03,800 --> 00:40:07,120
Speaker 2: links and of course, my newsletter. I also really recommend

868
00:40:07,120 --> 00:40:09,080
Speaker 2: you go to chat dot Where's youread dot at to

869
00:40:09,160 --> 00:40:11,520
Speaker 2: visit the discord, and go to our slash.

870
00:40:11,239 --> 00:40:14,320
Speaker 3: Better Offline to check out I'll Reddit. Thank you so

871
00:40:14,480 --> 00:40:17,879
Speaker 3: much for listening. Better Offline is a production of cool

872
00:40:17,960 --> 00:40:18,520
Speaker 3: Zone Media.

873
00:40:18,680 --> 00:40:21,520
Speaker 2: For more from cool Zone Media, visit our website cool

874
00:40:21,560 --> 00:40:24,919
Speaker 2: Zonemedia dot com, or check us out on the iHeartRadio app,

875
00:40:25,000 --> 00:40:27,400
Speaker 2: Apple Podcasts, or wherever you get your podcasts.