1
00:00:02,800 --> 00:00:03,960
Speaker 1: Zone media.

2
00:00:05,320 --> 00:00:08,840
Speaker 2: Chosen by God, perfected by science. I'm ed Zetron. This

3
00:00:08,960 --> 00:00:23,400
Speaker 2: is better offline. And as I've written about many, many,

4
00:00:23,440 --> 00:00:26,119
Speaker 2: many many times and argued on this very podcast, just

5
00:00:26,160 --> 00:00:28,920
Speaker 2: as often, the large language models run by companies like

6
00:00:28,960 --> 00:00:33,120
Speaker 2: open Ai, Athropic, Google, and Meta are unprofitable and unsustainable,

7
00:00:33,280 --> 00:00:36,000
Speaker 2: and the transformer based architecture they run on as peaked.

8
00:00:36,159 --> 00:00:38,920
Speaker 2: They're running out of training data, and the actual capabilities

9
00:00:38,920 --> 00:00:40,840
Speaker 2: of these models were peaking as far back as March

10
00:00:40,880 --> 00:00:44,800
Speaker 2: twenty twenty four. Nevertheless, I'd assumed, incorrectly, by the way,

11
00:00:45,040 --> 00:00:47,320
Speaker 2: that there would be no way to make them more efficient,

12
00:00:47,360 --> 00:00:51,160
Speaker 2: because I had assumed, also incorrectly that the hyperscalers, along

13
00:00:51,200 --> 00:00:53,960
Speaker 2: with open Ai and Anthropic, would be constantly looking for

14
00:00:54,000 --> 00:00:56,680
Speaker 2: ways to bring down the ruinous cost of their services.

15
00:00:57,200 --> 00:00:59,920
Speaker 2: After all, open Ai lost five billion dollars last year,

16
00:01:00,000 --> 00:01:02,960
Speaker 2: and that's after three point seven billion dollars in revenue too.

17
00:01:03,160 --> 00:01:06,000
Speaker 2: An Anthropic lost just under three billion dollars in twenty

18
00:01:06,040 --> 00:01:08,600
Speaker 2: twenty four. And in the last episode I told you

19
00:01:08,640 --> 00:01:10,720
Speaker 2: a little bit about deep seek. By the way, in

20
00:01:10,800 --> 00:01:13,119
Speaker 2: this one we're going to get into well how fucked

21
00:01:13,120 --> 00:01:16,280
Speaker 2: things might actually be. But what I didn't wager was

22
00:01:16,280 --> 00:01:19,559
Speaker 2: that potentially nobody was actually trying to make these models

23
00:01:19,640 --> 00:01:22,640
Speaker 2: more efficient. My mistake was if you can believe this

24
00:01:23,240 --> 00:01:26,160
Speaker 2: being too generous to the AI companies, assuming that they

25
00:01:26,200 --> 00:01:29,600
Speaker 2: didn't pursue efficiency because they couldn't and not because they

26
00:01:29,600 --> 00:01:33,240
Speaker 2: couldn't be bothered. But then, as I just hinted at,

27
00:01:33,440 --> 00:01:36,080
Speaker 2: a little known Chinese company released a product that was

28
00:01:36,120 --> 00:01:39,800
Speaker 2: broadly equivalent to open AI's latest reasoning models, but cost

29
00:01:39,800 --> 00:01:42,160
Speaker 2: a fraction of the cost to train and run. And

30
00:01:42,200 --> 00:01:45,080
Speaker 2: now the conventional understanding of how generative AI should work

31
00:01:45,120 --> 00:01:49,120
Speaker 2: has been fundamentally upended. You see, the pre Deep Sea

32
00:01:49,360 --> 00:01:52,120
Speaker 2: status quo was one where several truths, and I'd say

33
00:01:52,120 --> 00:01:54,360
Speaker 2: that in the loosest sense of the word, allowed the

34
00:01:54,400 --> 00:01:56,960
Speaker 2: party to keep going. So the first one is that

35
00:01:56,960 --> 00:02:00,160
Speaker 2: these models were incredibly expensive to train. GPT four morrow

36
00:02:00,240 --> 00:02:02,240
Speaker 2: cost one hundred million dollars in the middle of twenty

37
00:02:02,240 --> 00:02:05,960
Speaker 2: twenty four, and future models, according Dario Ama Day of Anthropic,

38
00:02:06,040 --> 00:02:08,480
Speaker 2: might cost as much as one billion or more to train.

39
00:02:09,160 --> 00:02:11,880
Speaker 2: And training future models, by the way, as a result

40
00:02:11,919 --> 00:02:14,560
Speaker 2: of this, would necessitate spending billions of dollars on both

41
00:02:14,639 --> 00:02:17,520
Speaker 2: data centers and the GPUs necessary to keep training these

42
00:02:17,560 --> 00:02:21,120
Speaker 2: bigger huger models. Now, another thing was these models had

43
00:02:21,160 --> 00:02:23,720
Speaker 2: to be large because making them large, pumping them full

44
00:02:23,720 --> 00:02:26,799
Speaker 2: of training data, and throwing masses of compute at them

45
00:02:26,880 --> 00:02:29,919
Speaker 2: would unlock new features such as an AI that helps

46
00:02:29,960 --> 00:02:32,560
Speaker 2: us accomplish much more than we ever could without AI,

47
00:02:32,680 --> 00:02:36,280
Speaker 2: which is Sam Altman quote. And in the words of

48
00:02:36,320 --> 00:02:39,320
Speaker 2: Sam again, you'd be able to get a personal AI

49
00:02:39,360 --> 00:02:42,120
Speaker 2: team full of virtual experts in different areas working together

50
00:02:42,200 --> 00:02:44,960
Speaker 2: to create almost anything we can imagine. I don't know, mate,

51
00:02:45,000 --> 00:02:48,880
Speaker 2: you ever try creating a functional fucking business. Dipshit. Anyway,

52
00:02:49,040 --> 00:02:52,680
Speaker 2: here's another one. These models were incredibly expensive to run.

53
00:02:52,760 --> 00:02:54,480
Speaker 2: There has to be this way, but it was all

54
00:02:54,480 --> 00:02:57,440
Speaker 2: worth it because making these models powerful was way more

55
00:02:57,480 --> 00:02:59,919
Speaker 2: important than making them efficient. Because once the price of

56
00:03:00,080 --> 00:03:02,400
Speaker 2: silicon comes down, and this is a refrain I've heard

57
00:03:02,440 --> 00:03:04,600
Speaker 2: from multiple different people as a defense of the costs

58
00:03:04,600 --> 00:03:07,720
Speaker 2: of generative AI, we would then have these powerful models

59
00:03:07,720 --> 00:03:13,959
Speaker 2: that were cheaper somehow because of silicon. Now you may think,

60
00:03:14,120 --> 00:03:16,640
Speaker 2: ed that sounds like a not a real argument. That

61
00:03:16,720 --> 00:03:19,120
Speaker 2: just sounds like something someone said once, and it is.

62
00:03:19,200 --> 00:03:21,680
Speaker 2: It is something someone said once. Anyone who knows anything

63
00:03:21,680 --> 00:03:23,640
Speaker 2: about chips know how hard it is to make a

64
00:03:23,680 --> 00:03:26,520
Speaker 2: new chip. And remember one of the CES episodes when

65
00:03:26,520 --> 00:03:28,840
Speaker 2: I asked Max Cherney about this, should go back and

66
00:03:28,880 --> 00:03:33,040
Speaker 2: listen to him. Anyway, another thing, another part of this

67
00:03:33,919 --> 00:03:36,760
Speaker 2: was that as a result of this need to make bigger, huger,

68
00:03:36,920 --> 00:03:40,440
Speaker 2: even bigger models, the most powerful ones, these big beautiful

69
00:03:40,480 --> 00:03:42,400
Speaker 2: models that we love them. We look at the big

70
00:03:42,440 --> 00:03:45,880
Speaker 2: beautiful models, we would of course need to keep buying bigger,

71
00:03:46,000 --> 00:03:49,920
Speaker 2: more powerful GPUs, which would continue the American excellence of

72
00:03:49,960 --> 00:03:52,800
Speaker 2: burning a bunch of money on nothing. And by following

73
00:03:52,840 --> 00:03:56,640
Speaker 2: this roadmap, everybody wins. The hyperscalers get the justification they

74
00:03:56,680 --> 00:03:59,920
Speaker 2: needed to create more sprawling data centers and spend massive

75
00:04:00,000 --> 00:04:02,560
Speaker 2: amounts of money and open AI and their alk get

76
00:04:02,640 --> 00:04:06,040
Speaker 2: to continue building powerful models, and also in video continues

77
00:04:06,080 --> 00:04:09,440
Speaker 2: to make money selling GPUs. Remember I've said in the

78
00:04:09,480 --> 00:04:11,400
Speaker 2: past that things were kind of a death cull. This

79
00:04:11,520 --> 00:04:13,760
Speaker 2: is what this is. It's a capitalist death cull. It

80
00:04:13,840 --> 00:04:17,599
Speaker 2: runs on plagiarism and hubris and the assumption that at

81
00:04:17,640 --> 00:04:21,440
Speaker 2: some point all of this would turn into something meaningful. Now,

82
00:04:21,440 --> 00:04:23,640
Speaker 2: I've argued for a while that the latter part of

83
00:04:23,680 --> 00:04:25,800
Speaker 2: the plan was insane, that there was no profitability for

84
00:04:25,839 --> 00:04:28,320
Speaker 2: these large language models. As I believe, there simply wasn't

85
00:04:28,360 --> 00:04:30,839
Speaker 2: a way to make these models more efficient in a

86
00:04:30,880 --> 00:04:33,919
Speaker 2: way I was raimed. The current models developed by both

87
00:04:33,960 --> 00:04:37,200
Speaker 2: the hyperscalers, so Gemini from Google or Lama from Meta,

88
00:04:37,240 --> 00:04:39,240
Speaker 2: and so on and so forth, and the multi billion

89
00:04:39,279 --> 00:04:41,480
Speaker 2: dollar startups, if you can even fucking call them that,

90
00:04:41,600 --> 00:04:45,159
Speaker 2: Open Ai and Anthropic, they're horribly inefficient. And I just

91
00:04:45,240 --> 00:04:47,200
Speaker 2: made the mistake of assuming that they tried to make

92
00:04:47,200 --> 00:04:50,880
Speaker 2: them more efficient, and they couldn't. Now, but what we're

93
00:04:50,920 --> 00:04:56,839
Speaker 2: witnessing right now isn't some sort of weird China situation.

94
00:04:57,000 --> 00:05:01,360
Speaker 2: This isn't China being Chinese and doing scary Chinese things

95
00:05:01,400 --> 00:05:03,960
Speaker 2: to us. Now, what we're witnessing is the American tech

96
00:05:04,000 --> 00:05:08,120
Speaker 2: industry's greatest act of hubris. It's a monument to the

97
00:05:08,160 --> 00:05:12,120
Speaker 2: barely conscious stewards of so called innovation who are incapable

98
00:05:12,120 --> 00:05:14,920
Speaker 2: of breaking the k fab of the fake competition where

99
00:05:14,960 --> 00:05:18,080
Speaker 2: everybody makes the same products, charges about the same amount

100
00:05:18,080 --> 00:05:22,640
Speaker 2: of money, and mostly innovates in the same direction. Somehow, nobody,

101
00:05:22,920 --> 00:05:26,440
Speaker 2: not Google, not Microsoft, not open Ai, not Meta, not Amazon,

102
00:05:26,480 --> 00:05:29,640
Speaker 2: no Oracle thought to try or was capable of creating

103
00:05:29,720 --> 00:05:33,000
Speaker 2: something like deep Seek, Which doesn't mean that deep Seek's

104
00:05:33,040 --> 00:05:36,480
Speaker 2: team is particularly remarkable or found anything super new, but

105
00:05:36,720 --> 00:05:39,880
Speaker 2: that for all the talent, trillions of dollars of market capitalization,

106
00:05:39,920 --> 00:05:44,120
Speaker 2: and supposed expertise in Americans American tech hologarchs, not one

107
00:05:44,160 --> 00:05:46,799
Speaker 2: bright spark thought to try things that deep Seak had tried,

108
00:05:46,960 --> 00:05:48,880
Speaker 2: which appeared to be what if we didn't use as

109
00:05:48,960 --> 00:05:52,400
Speaker 2: much memory, and what if we tried synthetic data? And

110
00:05:52,440 --> 00:05:55,400
Speaker 2: because the cost of model development and inference was so

111
00:05:55,480 --> 00:05:58,520
Speaker 2: astronomical in the case of American models, they never assumed

112
00:05:58,520 --> 00:06:01,800
Speaker 2: that anyone would try to use their position. This is

113
00:06:01,960 --> 00:06:04,839
Speaker 2: especially bad considering that China's folks and AI as a

114
00:06:04,839 --> 00:06:08,200
Speaker 2: strategic part of its industrial priority was really no secret,

115
00:06:08,560 --> 00:06:11,840
Speaker 2: even if the ways it supported domestic companies kind of is.

116
00:06:12,400 --> 00:06:15,000
Speaker 2: In the same way that the automotive industry was blindsided

117
00:06:15,000 --> 00:06:18,760
Speaker 2: by China's EV manufacturers, the same is happening with AI. Fat,

118
00:06:19,040 --> 00:06:22,200
Speaker 2: happy and lazy, and most of all oblivious. America's most

119
00:06:22,279 --> 00:06:25,840
Speaker 2: powerful tech companies sacked back and built bigger, messier models

120
00:06:25,839 --> 00:06:28,520
Speaker 2: powered by sprawling data centers of billions of dollars of

121
00:06:28,560 --> 00:06:31,800
Speaker 2: GPUs from Nvidia, A bacchanalia are spending that strains our

122
00:06:31,920 --> 00:06:35,000
Speaker 2: energy grid and depletes our fucking water reserves. Without it

123
00:06:35,080 --> 00:06:39,359
Speaker 2: appears much consideration of whether an alternative was possible. I

124
00:06:39,560 --> 00:06:42,400
Speaker 2: refuse to believe that none of these companies could have

125
00:06:42,480 --> 00:06:44,760
Speaker 2: done what Deep Seek has done, which means that they

126
00:06:44,800 --> 00:06:47,880
Speaker 2: either chose not to or they were so utterly myopic,

127
00:06:48,080 --> 00:06:50,720
Speaker 2: so excited to burn so much money on so many

128
00:06:50,800 --> 00:06:53,880
Speaker 2: parts of burning the earth, boiling lakes, and stealing from

129
00:06:53,920 --> 00:06:56,719
Speaker 2: people in pursuit of further growth, that they didn't think

130
00:06:56,760 --> 00:07:00,719
Speaker 2: to try. This isn't about China. It's so much fucking

131
00:07:00,839 --> 00:07:03,960
Speaker 2: easier if we let it be about China. No, no, no, no.

132
00:07:04,160 --> 00:07:09,240
Speaker 2: It's about how the American tech industry is incurious, lazy, entitled, directionless,

133
00:07:09,240 --> 00:07:14,240
Speaker 2: and irresponsible open an anthropic of the antithesis of Silicon Valley.

134
00:07:14,400 --> 00:07:18,040
Speaker 2: They're incumbents, public companies wearing startup suits. I'm willing to

135
00:07:18,040 --> 00:07:21,240
Speaker 2: take on real challenges more focused on optics and marketing

136
00:07:21,280 --> 00:07:24,280
Speaker 2: than they are in solving actual fucking problems, even the

137
00:07:24,320 --> 00:07:27,560
Speaker 2: problems that they themselves created with their large language models.

138
00:07:28,200 --> 00:07:31,280
Speaker 2: By making this about China, we ignore the root of

139
00:07:31,320 --> 00:07:34,280
Speaker 2: the problem that the American tech industry is no longer

140
00:07:34,280 --> 00:07:38,360
Speaker 2: interested in making good software that actually helps people. Deep

141
00:07:38,400 --> 00:07:41,480
Speaker 2: seaks shouldn't be scary to Silicon Valley because they that

142
00:07:41,560 --> 00:07:43,920
Speaker 2: Silicon Valley should have come up with this first. It

143
00:07:44,040 --> 00:07:46,880
Speaker 2: uses less memory, fewer resources, and uses several kind of

144
00:07:46,920 --> 00:07:50,280
Speaker 2: quirky workarounds to adapt to the limited compute resources available.

145
00:07:50,440 --> 00:07:54,560
Speaker 2: All things that you'd previously associate with Silicon Valley, except now.

146
00:07:54,600 --> 00:07:57,040
Speaker 2: Silicon Valley's only interest that the rest of the American

147
00:07:57,040 --> 00:08:01,000
Speaker 2: tech industry is the rot economy. It only cares about growing,

148
00:08:01,280 --> 00:08:04,360
Speaker 2: growing at all costs, even if said costs were really

149
00:08:04,800 --> 00:08:07,360
Speaker 2: things you could mitigate, or if the costs themselves were

150
00:08:07,400 --> 00:08:10,800
Speaker 2: self defeating. To be clear, if the alternative is that

151
00:08:10,880 --> 00:08:13,200
Speaker 2: all of these companies simply did not come up with

152
00:08:13,200 --> 00:08:15,960
Speaker 2: this idea that in and of itself is a damning

153
00:08:16,080 --> 00:08:19,080
Speaker 2: indictment of the Valley. Was nobody thinking of this stuff?

154
00:08:19,240 --> 00:08:21,760
Speaker 2: If they were, why didn't Sam Altman or Dario Amadae

155
00:08:21,840 --> 00:08:24,360
Speaker 2: or sach In Adela or anyone else put serious resources

156
00:08:24,360 --> 00:08:26,920
Speaker 2: into efficiency? Was it because there was no reason to?

157
00:08:27,320 --> 00:08:29,600
Speaker 2: Was it because there was, if we're honest, no real

158
00:08:29,640 --> 00:08:33,800
Speaker 2: competition between any of these companies? Did anybody try anything

159
00:08:33,920 --> 00:08:36,679
Speaker 2: other than throwing as much compute and training data at

160
00:08:36,679 --> 00:08:40,640
Speaker 2: the model as possible. It's all just so cynical and

161
00:08:40,679 --> 00:08:44,440
Speaker 2: antithetical to innovation itself. Surely, if any of this shit mattered,

162
00:08:44,559 --> 00:08:47,440
Speaker 2: if generative AI truly was valid and viable in the

163
00:08:47,480 --> 00:08:50,280
Speaker 2: eyes of these companies, they would have actively worked to

164
00:08:50,320 --> 00:08:53,760
Speaker 2: do something like Deep Seeker has done. Don't get me wrong,

165
00:08:54,200 --> 00:08:56,840
Speaker 2: it appears Deep Seak employed all sorts of weird tricks

166
00:08:56,880 --> 00:08:59,640
Speaker 2: to make this work, including taking advantage of distinct parts

167
00:08:59,640 --> 00:09:02,400
Speaker 2: of both CPUs and GPUs to create something called a

168
00:09:02,400 --> 00:09:06,280
Speaker 2: digital processing unit, essentially redefining how data is communicated within

169
00:09:06,320 --> 00:09:09,080
Speaker 2: the servers running training and inference. And just this reminder,

170
00:09:09,120 --> 00:09:11,800
Speaker 2: inference is the thing where when you type something in

171
00:09:12,120 --> 00:09:15,440
Speaker 2: that's it infers the meaning. Just could have specified that earlier.

172
00:09:16,040 --> 00:09:18,640
Speaker 2: Deepseak had to do things that a company with unrestrained

173
00:09:18,679 --> 00:09:21,200
Speaker 2: access to capital and equipment wouldn't have to do, and

174
00:09:21,240 --> 00:09:25,439
Speaker 2: it often used impractical and quirky methods to do so. Nevertheless,

175
00:09:25,520 --> 00:09:28,160
Speaker 2: Open AI and Anthropic both have enough money and hiring

176
00:09:28,200 --> 00:09:30,959
Speaker 2: power to have tried and succeeded in creating a model

177
00:09:31,280 --> 00:09:34,880
Speaker 2: this efficient and capable of running on older GPUs, except

178
00:09:34,920 --> 00:09:38,199
Speaker 2: what they wanted, what they actually wanted was more goddamn

179
00:09:38,280 --> 00:09:41,320
Speaker 2: growth and the chance to build even bigger data centers

180
00:09:41,360 --> 00:09:44,560
Speaker 2: with even more compute that they would own. Open ai

181
00:09:44,760 --> 00:09:48,559
Speaker 2: is as much a lazy, cumbersome incumbent as Google or Microsoft,

182
00:09:48,679 --> 00:09:51,640
Speaker 2: and it's about as innovative too. The launch of its

183
00:09:51,679 --> 00:09:55,640
Speaker 2: Operator agent was a joke, a barely functional product that's

184
00:09:55,640 --> 00:09:58,720
Speaker 2: allegedly meant to control your computer and take distinct actions

185
00:09:58,760 --> 00:10:01,679
Speaker 2: like ordering stuff of instacark, you know, things you could

186
00:10:01,720 --> 00:10:04,080
Speaker 2: do with your hands, but just to be clear, it

187
00:10:04,080 --> 00:10:07,320
Speaker 2: doesn't work. You'll never guess who was really into it.

188
00:10:07,360 --> 00:10:10,800
Speaker 2: Though his name is Casey Newton. He writes a blog

189
00:10:10,840 --> 00:10:14,080
Speaker 2: called Platformer, and he's a man so gratingly credulous that

190
00:10:14,080 --> 00:10:17,000
Speaker 2: it makes me want to fucking scream. And of course

191
00:10:17,200 --> 00:10:19,839
Speaker 2: he wrote the Operator when he used it was a

192
00:10:19,880 --> 00:10:24,720
Speaker 2: compelling demonstration that represented an extraordinary technological achievement. Also, somehow

193
00:10:24,800 --> 00:10:27,960
Speaker 2: was significantly slower, more frustrating, and more expensive than simply

194
00:10:28,000 --> 00:10:32,040
Speaker 2: doing any of these tasks yourself. Casey, of course, not

195
00:10:32,160 --> 00:10:35,280
Speaker 2: to worry, had some extra thoughts about Deep Seek, that

196
00:10:35,320 --> 00:10:38,079
Speaker 2: there were reasons to be worried, but that American AI

197
00:10:38,200 --> 00:10:40,600
Speaker 2: lads were still in the lead, saying that deep seak

198
00:10:40,679 --> 00:10:43,360
Speaker 2: was only optimizing technology that open ai and others had

199
00:10:43,360 --> 00:10:46,760
Speaker 2: invented first, before saying that deep seek was only last

200
00:10:46,760 --> 00:10:49,120
Speaker 2: week that open ai made available to pro plan used

201
00:10:49,120 --> 00:10:51,520
Speaker 2: as a computer that can use itself. This statement is

202
00:10:51,559 --> 00:10:54,760
Speaker 2: bordering on factually incorrect. It is fucking insane that Casey

203
00:10:54,840 --> 00:10:57,040
Speaker 2: is still doing this. I do not want to I

204
00:10:57,040 --> 00:10:58,800
Speaker 2: don't know what to do with this guy. This guy

205
00:10:59,600 --> 00:11:03,320
Speaker 2: just like that. That's a fucking lie. The computer can't

206
00:11:03,400 --> 00:11:06,680
Speaker 2: use itself, This shit can't just to explain what operator is,

207
00:11:06,720 --> 00:11:09,040
Speaker 2: You're meant to type in something like, hey, order me

208
00:11:09,120 --> 00:11:12,200
Speaker 2: some milk, Order me some milk off of vistacar. And

209
00:11:12,480 --> 00:11:14,719
Speaker 2: when Casey tried this, he tried to find milk and

210
00:11:14,760 --> 00:11:20,120
Speaker 2: de moine iowa. Just fucking insane. Just this is how

211
00:11:20,160 --> 00:11:22,880
Speaker 2: these companies have got big. It's people like cases people

212
00:11:22,880 --> 00:11:25,280
Speaker 2: are Casey, who are just like anything they show tore

213
00:11:25,280 --> 00:11:27,280
Speaker 2: like God, damn, that's the most impressive thing I've seen

214
00:11:27,320 --> 00:11:30,360
Speaker 2: in my life. It's a fucking fast. But let's be frank,

215
00:11:30,840 --> 00:11:34,240
Speaker 2: these companies aren't building shit. Open Air and Anthropic are

216
00:11:34,240 --> 00:11:37,080
Speaker 2: both limply throwing around the idea that agents are possible

217
00:11:37,240 --> 00:11:39,200
Speaker 2: in an attempt to raise more money to burn, and

218
00:11:39,240 --> 00:11:41,199
Speaker 2: after the launch of deep Seek, I have to wonder

219
00:11:41,200 --> 00:11:44,240
Speaker 2: what any investor thinks they're investing in other than certain

220
00:11:44,280 --> 00:11:46,720
Speaker 2: ones I'll get into in a bit. And to be clear,

221
00:11:46,760 --> 00:11:49,800
Speaker 2: an agent is meant to be this autonomous thing which

222
00:11:49,800 --> 00:11:52,239
Speaker 2: you say, hey, go and do this action, go and

223
00:11:52,360 --> 00:11:55,120
Speaker 2: sell things for me, and go and email people for me.

224
00:11:55,320 --> 00:11:58,040
Speaker 2: They don't really work. There are some that kind of

225
00:11:58,080 --> 00:12:00,520
Speaker 2: do that are really expensive, but large Lane which model

226
00:12:00,559 --> 00:12:03,600
Speaker 2: is not built for this kind of thing. But let's

227
00:12:03,640 --> 00:12:06,360
Speaker 2: be honest, Deep Seek and as I said in the

228
00:12:06,440 --> 00:12:09,680
Speaker 2: last episode, they've built a more efficient reasoning model. So

229
00:12:09,800 --> 00:12:14,080
Speaker 2: like open aiy zo one And you'd think, well, okay,

230
00:12:14,360 --> 00:12:17,600
Speaker 2: couldn't open Ai simply add on deep Seek to its models.

231
00:12:18,280 --> 00:12:21,360
Speaker 2: Not really. First of all, with the way these models work,

232
00:12:21,400 --> 00:12:22,960
Speaker 2: you can't just like plug it in. It's just not

233
00:12:23,040 --> 00:12:25,480
Speaker 2: how it works. They could train a new model using

234
00:12:25,520 --> 00:12:29,680
Speaker 2: deep Seek's techniques, but the optics of that aren't brilliant.

235
00:12:29,760 --> 00:12:32,040
Speaker 2: It would be a concession. It'd be in the middle

236
00:12:32,080 --> 00:12:35,120
Speaker 2: that open aiyes slipped and needs to catch up, and

237
00:12:35,120 --> 00:12:38,120
Speaker 2: not to its main rival pretend rival, I mean anthropic

238
00:12:38,440 --> 00:12:40,840
Speaker 2: or to like another big tech firm, but to an

239
00:12:40,960 --> 00:12:44,560
Speaker 2: outgrowth of a hedge fund in China, a company that

240
00:12:44,679 --> 00:12:49,600
Speaker 2: few had heard of before December and the like. Really

241
00:12:49,760 --> 00:12:52,720
Speaker 2: not that many people had heard before January twenty fifth.

242
00:12:53,320 --> 00:12:55,920
Speaker 2: It's very embarrassing, and this, in turn, I think will

243
00:12:55,960 --> 00:12:59,319
Speaker 2: make any serious investor think twice about writing the company

244
00:12:59,360 --> 00:13:01,560
Speaker 2: a blank chet. They're going to have to dip into

245
00:13:01,600 --> 00:13:06,160
Speaker 2: some very bothersome pockets, and as I've said ad nauseum,

246
00:13:06,440 --> 00:13:09,359
Speaker 2: this is potentially fatal as open ai needs to continually

247
00:13:09,440 --> 00:13:12,480
Speaker 2: raise money, more money than any startup has ever raised

248
00:13:12,480 --> 00:13:14,760
Speaker 2: in the history of anything, and it really doesn't have

249
00:13:14,800 --> 00:13:17,720
Speaker 2: a path to breaking even even if they copy what

250
00:13:17,760 --> 00:13:20,960
Speaker 2: Deep Sea did, because we still right now, though deep

251
00:13:20,960 --> 00:13:24,720
Speaker 2: Seek is thirty times cheaper than one, we don't know

252
00:13:24,760 --> 00:13:28,120
Speaker 2: if that's profit or we don't know. We haven't found out.

253
00:13:28,960 --> 00:13:31,120
Speaker 2: And if open ai wants to do its own, cheaper,

254
00:13:31,160 --> 00:13:33,319
Speaker 2: more efficient model, it's likely to have to create it

255
00:13:33,360 --> 00:13:35,240
Speaker 2: from scratch, like I said, And while it could do

256
00:13:35,360 --> 00:13:38,000
Speaker 2: distillation to make it kind of more like open ai

257
00:13:38,200 --> 00:13:40,720
Speaker 2: using their own models. By the way, deep Sea taught

258
00:13:40,760 --> 00:13:43,880
Speaker 2: itself using open aies outputs. Like I mentioned in the

259
00:13:43,920 --> 00:13:46,600
Speaker 2: last episode, it's kind of what deep Seak already did.

260
00:13:46,640 --> 00:13:49,840
Speaker 2: It already has been fed open ai bullshit. Even with

261
00:13:50,000 --> 00:13:53,280
Speaker 2: open AI's much larger team and more powerful hardware, it's

262
00:13:53,280 --> 00:13:55,480
Speaker 2: hard to see how creating a smaller, more efficient, and

263
00:13:55,520 --> 00:13:57,840
Speaker 2: almost as powerful version of OH one benefits them in

264
00:13:57,880 --> 00:14:01,480
Speaker 2: any way, because said version has well already been beaten

265
00:14:01,520 --> 00:14:03,559
Speaker 2: to market by deep Seek, and thanks to deep Seek,

266
00:14:03,800 --> 00:14:06,120
Speaker 2: will almost certainly have a great deal of competition for

267
00:14:06,160 --> 00:14:08,600
Speaker 2: a product that to this day lacks any killer apps.

268
00:14:08,600 --> 00:14:24,440
Speaker 1: Anyway, It's just it's very frustrating to me. It's very

269
00:14:24,440 --> 00:14:27,200
Speaker 1: frustrating to me. It drives me a little insane. Reading

270
00:14:27,200 --> 00:14:29,200
Speaker 1: all this stuff makes me feel crazy. I think you

271
00:14:29,240 --> 00:14:32,440
Speaker 1: hear it in my voice, you hear the sanity stripping away.

272
00:14:32,480 --> 00:14:36,360
Speaker 1: But I'm here to podcast and don't worry. But seriously, though,

273
00:14:37,240 --> 00:14:39,400
Speaker 1: anyone can build on top of what deep seek has

274
00:14:39,400 --> 00:14:44,480
Speaker 1: already built. Where is open Aiy's mote exactly? And where's anthropics?

275
00:14:44,720 --> 00:14:47,040
Speaker 1: What are the things that make these companies worth sixty

276
00:14:47,080 --> 00:14:49,960
Speaker 1: billion or one hundred and fifty billion, or oh my god,

277
00:14:50,000 --> 00:14:51,840
Speaker 1: as we'll discuss in a bit, three hundred and forty

278
00:14:51,880 --> 00:14:52,560
Speaker 1: billion dollars?

279
00:14:53,520 --> 00:14:56,000
Speaker 2: What is the technology they own or the talent they

280
00:14:56,040 --> 00:14:59,040
Speaker 2: have that justifies these valuations? Because it's kind of hard

281
00:14:59,080 --> 00:15:04,320
Speaker 2: to argue that their model particularly valuable anymore. Celebrity celebrity

282
00:15:04,400 --> 00:15:09,840
Speaker 2: them CULTI personality. Sam Orman, He's an artful bullshitter, and

283
00:15:09,880 --> 00:15:11,840
Speaker 2: he's built a career out of being in the right

284
00:15:11,920 --> 00:15:14,080
Speaker 2: places at the right times, have in the right connections,

285
00:15:14,160 --> 00:15:17,880
Speaker 2: knowing exactly what to say, especially to credulous tech media

286
00:15:18,120 --> 00:15:20,640
Speaker 2: ponces without the spine or inclination to push back on

287
00:15:20,680 --> 00:15:24,240
Speaker 2: his more stupid claims. And already Aortman has tried to

288
00:15:24,280 --> 00:15:27,760
Speaker 2: shrug off Deep Seaks Rise, admitting that while Deep Seeks

289
00:15:27,760 --> 00:15:30,520
Speaker 2: are one model is impressive, particularly when it comes to

290
00:15:30,560 --> 00:15:33,800
Speaker 2: its efficiency, open Ai will obviously deliver much better models,

291
00:15:33,800 --> 00:15:36,720
Speaker 2: and also it's legit invigorating for open Ai to have

292
00:15:36,760 --> 00:15:39,960
Speaker 2: a new competitor. Yeah, mate, sure, I'll bet your love

293
00:15:40,000 --> 00:15:43,080
Speaker 2: in this. Aortmand ended his tweet where that came from

294
00:15:43,120 --> 00:15:46,880
Speaker 2: with look forward to bringing you all AGI and beyond,

295
00:15:47,200 --> 00:15:50,000
Speaker 2: something which I add has always been close on the

296
00:15:50,040 --> 00:15:53,800
Speaker 2: horizon in Altman's world, but it's never really materialized. Then

297
00:15:53,840 --> 00:15:56,680
Speaker 2: the timeline keeps moving and there's no actual proof they

298
00:15:56,720 --> 00:15:59,720
Speaker 2: can do it. AGI is not fucking happening, And if

299
00:15:59,720 --> 00:16:02,240
Speaker 2: it's possible in any way, it's not coming out of

300
00:16:02,280 --> 00:16:05,840
Speaker 2: probabilistic models. I'm fucking sick of this and Opening I

301
00:16:06,080 --> 00:16:09,240
Speaker 2: can't even lean on its relationship with Microsoft, which on Wednesday,

302
00:16:09,320 --> 00:16:12,320
Speaker 2: January twenty ninth started offering deep Seek's models through its

303
00:16:12,320 --> 00:16:16,440
Speaker 2: own cloud services. Opening I hasn't got shit. Deep Seek

304
00:16:16,440 --> 00:16:19,800
Speaker 2: has commoditized the large language model, publishing both the source

305
00:16:19,840 --> 00:16:22,240
Speaker 2: code and the guide to building your own. Whether or

306
00:16:22,320 --> 00:16:25,040
Speaker 2: not someone chooses to pay deep Seek is largely irrelevant.

307
00:16:25,280 --> 00:16:27,600
Speaker 2: Someone else will take what they have created and build

308
00:16:27,640 --> 00:16:29,760
Speaker 2: their own, or people will start running their own deep

309
00:16:29,840 --> 00:16:33,200
Speaker 2: Seek instances, renting GPUs from one of the various cloud

310
00:16:33,200 --> 00:16:35,520
Speaker 2: computing firms. They don't give a shit. They'll take the

311
00:16:35,600 --> 00:16:38,280
Speaker 2: money and while in video, will always find other ways

312
00:16:38,280 --> 00:16:41,120
Speaker 2: to make money. Jensen Wong is amazing at this. It's

313
00:16:41,160 --> 00:16:43,240
Speaker 2: going to be a hard sell for any hyperscala to

314
00:16:43,320 --> 00:16:46,360
Speaker 2: justify spending billions more on GPUs to markets that now

315
00:16:46,440 --> 00:16:48,560
Speaker 2: know that near identical models can be built for a

316
00:16:48,600 --> 00:16:51,400
Speaker 2: fraction of the cost with older hardware. Why do you

317
00:16:51,440 --> 00:16:55,120
Speaker 2: need Blackwell, which is the latest Nvidia GPU. The narrative

318
00:16:55,160 --> 00:16:57,720
Speaker 2: of this is the only way to build powerful models.

319
00:16:57,920 --> 00:17:00,800
Speaker 2: It doesn't really hold water anymore, and the only other

320
00:17:00,880 --> 00:17:04,560
Speaker 2: selling point is that what if China does something. Well,

321
00:17:04,680 --> 00:17:07,480
Speaker 2: the Chinese did something, and they've now proven that they

322
00:17:07,560 --> 00:17:10,960
Speaker 2: cannot only compete with American AI companies, but doing so

323
00:17:11,720 --> 00:17:14,440
Speaker 2: is possible in an efficient way that can effectively crash

324
00:17:14,480 --> 00:17:17,800
Speaker 2: the market while there's being a recovery. This is still

325
00:17:17,920 --> 00:17:21,480
Speaker 2: very worrying. I also want to address something real quick.

326
00:17:21,880 --> 00:17:24,160
Speaker 2: A few people on Twitter have been suggesting that talking

327
00:17:24,160 --> 00:17:26,600
Speaker 2: about deep Seek positively in any way is some sort

328
00:17:26,640 --> 00:17:30,120
Speaker 2: of Chinese op. If you believe this, you're a fucking moron.

329
00:17:30,600 --> 00:17:33,960
Speaker 2: I really must be clear, take your weirdsenophobia and go

330
00:17:34,119 --> 00:17:36,840
Speaker 2: eat your own shit. I don't fucking care anymore. Yes,

331
00:17:36,880 --> 00:17:39,920
Speaker 2: there are problems with China. Yes, China does something to America.

332
00:17:40,080 --> 00:17:43,240
Speaker 2: This is an open source thing. This is an open

333
00:17:43,280 --> 00:17:46,000
Speaker 2: source thing, and you can remove China from the equation.

334
00:17:46,080 --> 00:17:48,520
Speaker 2: Because it's open source, someone else is going to use this.

335
00:17:48,800 --> 00:17:52,320
Speaker 2: If your only defense is the sneaky Chinese are doing something,

336
00:17:52,480 --> 00:17:55,000
Speaker 2: go to therapy and talk to the therapist about you

337
00:17:55,119 --> 00:17:57,960
Speaker 2: being paranoid or racist, because it's one of them. I

338
00:17:57,960 --> 00:18:01,119
Speaker 2: should also be clear, concerns about China very realistic. The

339
00:18:01,200 --> 00:18:04,359
Speaker 2: Chinese government has tried to interfere with America. It's happened

340
00:18:04,359 --> 00:18:09,000
Speaker 2: many times. Even if China is funding Deep Seeks models,

341
00:18:09,920 --> 00:18:12,760
Speaker 2: the fact that they are open sourced means that anyone

342
00:18:12,800 --> 00:18:15,520
Speaker 2: can run them and anyone can build their own. They

343
00:18:15,520 --> 00:18:17,960
Speaker 2: can look under the hood. We don't have the training data,

344
00:18:18,000 --> 00:18:21,400
Speaker 2: but that is it. You cannot win on a senophobic

345
00:18:21,520 --> 00:18:25,800
Speaker 2: argument here. You can have realistic concerns about another foreign power.

346
00:18:26,040 --> 00:18:29,080
Speaker 2: I'm not saying not to, but what I am saying

347
00:18:29,520 --> 00:18:32,040
Speaker 2: is that you have to look at this realistically, and

348
00:18:32,040 --> 00:18:34,800
Speaker 2: you have to take this seriously. And dismissing this as

349
00:18:34,920 --> 00:18:41,919
Speaker 2: Chinese magic is stupid. It's very goddamn stupid. But like

350
00:18:41,960 --> 00:18:44,720
Speaker 2: I said earlier, it also isn't clear if these models

351
00:18:44,720 --> 00:18:47,480
Speaker 2: are actually going to be profitable. It's unclear who funds

352
00:18:47,520 --> 00:18:49,639
Speaker 2: Deep seg like I just said, and whether it's current

353
00:18:49,680 --> 00:18:53,119
Speaker 2: pricing is actually sustainable. But they're likely going to be

354
00:18:53,160 --> 00:18:56,040
Speaker 2: a damn site more profitable than anything open ai is

355
00:18:56,080 --> 00:18:59,280
Speaker 2: currently selling. After all, open ai loses money on every

356
00:18:59,280 --> 00:19:02,360
Speaker 2: single transaction, even their two hundred dollars a month Chat

357
00:19:02,480 --> 00:19:05,920
Speaker 2: GPT Pro subscription, And if open ai cuts its prices

358
00:19:05,920 --> 00:19:08,479
Speaker 2: to compete with deep sek, its losses are only going

359
00:19:08,520 --> 00:19:11,040
Speaker 2: to deepen. And as I've said again and again this

360
00:19:11,160 --> 00:19:13,960
Speaker 2: is also deeply cynical, because it's obvious that none of

361
00:19:14,000 --> 00:19:17,119
Speaker 2: this was ever about the proliferation of generative AI or

362
00:19:17,160 --> 00:19:21,040
Speaker 2: making sure that generative AI was accessible. Putting aside my

363
00:19:21,400 --> 00:19:24,520
Speaker 2: very obvious personal beliefs for a second, it's fairly obvious

364
00:19:24,520 --> 00:19:28,000
Speaker 2: why these companies, the big hyperscalers and open ai and

365
00:19:28,080 --> 00:19:31,680
Speaker 2: Anthropic wouldn't want to create something like deep seek because

366
00:19:31,760 --> 00:19:34,560
Speaker 2: creating an open source model that uses less resources means

367
00:19:34,600 --> 00:19:38,960
Speaker 2: that open Ai, Anthropic and their associated hyperscalar findom clients

368
00:19:39,200 --> 00:19:42,600
Speaker 2: would lose their soft monopoly on large language models. Now

369
00:19:42,600 --> 00:19:46,800
Speaker 2: what does that mean, I'll explain before deep seek to

370
00:19:46,840 --> 00:19:50,200
Speaker 2: make a competitive large language model like GPT four zero,

371
00:19:50,680 --> 00:19:53,720
Speaker 2: as in one that you can actually commercialize, required exceedingly

372
00:19:53,800 --> 00:19:56,840
Speaker 2: large amounts of capital, and to make larger ones effectively

373
00:19:56,880 --> 00:19:59,080
Speaker 2: required you to kiss the ring or the ass of Microsoft,

374
00:19:59,160 --> 00:20:01,600
Speaker 2: Google or Amazon. While it isn't clear what it cost

375
00:20:01,640 --> 00:20:04,239
Speaker 2: to train open aizo one reasoning model, we know that

376
00:20:04,280 --> 00:20:06,560
Speaker 2: GPT four to ZH cost in excess of one hundred

377
00:20:06,640 --> 00:20:09,200
Speaker 2: million dollars, and oh one is a more complex model

378
00:20:09,240 --> 00:20:12,280
Speaker 2: would likely cost even more. We also know that open

379
00:20:12,320 --> 00:20:14,719
Speaker 2: AI's training and inference costs in twenty twenty four were

380
00:20:14,760 --> 00:20:18,840
Speaker 2: around seven billion dollars, meaning that either refining current models

381
00:20:18,920 --> 00:20:22,480
Speaker 2: or building new ones is quite costly. The mythology of

382
00:20:22,520 --> 00:20:25,000
Speaker 2: both open AI and Anthropic is that these large amounts

383
00:20:25,040 --> 00:20:27,639
Speaker 2: of capital weren't just necessary, but the only way to

384
00:20:27,720 --> 00:20:31,280
Speaker 2: do this. While these companies ostensibly compete, neither of them

385
00:20:31,320 --> 00:20:34,200
Speaker 2: seemed concerned about doing so. As actual businesses that made

386
00:20:34,240 --> 00:20:37,760
Speaker 2: products that were say, cheaper and more efficient to run,

387
00:20:37,800 --> 00:20:40,600
Speaker 2: you know, they made more money than they cost because

388
00:20:40,600 --> 00:20:43,320
Speaker 2: in doing so, these companies would break the illusion that

389
00:20:43,359 --> 00:20:46,520
Speaker 2: the only way to create powerful artificial intelligence was to

390
00:20:46,560 --> 00:20:49,119
Speaker 2: hand billions of dollars to one of two companies and

391
00:20:49,160 --> 00:20:52,480
Speaker 2: build giant data centers to build even larger language models.

392
00:20:53,119 --> 00:20:57,120
Speaker 2: This is aisrot economy, two lumbering companies claiming that their

393
00:20:57,160 --> 00:20:59,520
Speaker 2: startups creating a narrative that the only way to build

394
00:20:59,520 --> 00:21:02,359
Speaker 2: the future to keep growing, to build more data centers,

395
00:21:02,359 --> 00:21:05,359
Speaker 2: to build larger language models, to consume more training data

396
00:21:05,400 --> 00:21:08,760
Speaker 2: with each infusion of capital, GPU purchase, and data center

397
00:21:08,800 --> 00:21:11,960
Speaker 2: build out, creating an infrastructural mote that always leads back

398
00:21:12,000 --> 00:21:16,159
Speaker 2: to one of a few tech hyperscalers. Open an anthropic

399
00:21:16,320 --> 00:21:18,920
Speaker 2: need the narrative to say, buy more GPUs and build

400
00:21:19,000 --> 00:21:21,480
Speaker 2: more data centers, because in doing so they create the

401
00:21:21,480 --> 00:21:25,720
Speaker 2: conditions of that infrastructural monopoly. Because the terms forget about

402
00:21:25,720 --> 00:21:28,880
Speaker 2: building software that does stuff for a second, were implicitly

403
00:21:29,160 --> 00:21:32,080
Speaker 2: that smaller players cannot enter the market because the market

404
00:21:32,119 --> 00:21:34,679
Speaker 2: is defined as large language models that cost hundreds of

405
00:21:34,680 --> 00:21:38,000
Speaker 2: millions of dollars and require access to more compute than

406
00:21:38,040 --> 00:21:41,480
Speaker 2: any startup could ever reasonably access without the infrastructure that

407
00:21:41,480 --> 00:21:47,160
Speaker 2: a public tech company delivers. Remember, neither Anthropic nor open

408
00:21:47,200 --> 00:21:50,080
Speaker 2: AI has ever marketed themselves based on the products they

409
00:21:50,119 --> 00:21:53,960
Speaker 2: actually build. Large language models are in and of themselves,

410
00:21:54,040 --> 00:21:57,040
Speaker 2: fairly bland software products, which is why we've yet to

411
00:21:57,040 --> 00:21:59,920
Speaker 2: see any killer apps. This isn't a particularly exciting pick

412
00:22:00,200 --> 00:22:03,200
Speaker 2: to investors or the public markets because there's no product, innovation,

413
00:22:03,280 --> 00:22:05,560
Speaker 2: or business model to point to, and if they'd actually

414
00:22:05,560 --> 00:22:08,080
Speaker 2: tried to productize it and turn it into a business,

415
00:22:08,240 --> 00:22:10,360
Speaker 2: it's quite obvious at this point that there really isn't

416
00:22:10,400 --> 00:22:13,800
Speaker 2: a multi trillion dollar industry for generative AI. Look at

417
00:22:13,840 --> 00:22:17,840
Speaker 2: Microsoft and their attempts to strong arm Copilot into Microsoft

418
00:22:17,880 --> 00:22:21,640
Speaker 2: three sixty five. Both personally and commercially, Nobody said Wow,

419
00:22:21,680 --> 00:22:24,800
Speaker 2: this is great. When they demanded use copilot in word,

420
00:22:25,119 --> 00:22:27,720
Speaker 2: Lots of people, however, asked why am I being charged

421
00:22:27,720 --> 00:22:30,040
Speaker 2: significantly more for a product that I don't want or

422
00:22:30,080 --> 00:22:33,639
Speaker 2: care about. Open Ai only makes twenty seven percent of

423
00:22:33,680 --> 00:22:36,359
Speaker 2: its revenue from selling access to its models, so allowing

424
00:22:36,400 --> 00:22:38,960
Speaker 2: people to use their models to build products around a

425
00:22:39,000 --> 00:22:41,560
Speaker 2: billion dollars of annual recurring revenue, by the way, with

426
00:22:41,640 --> 00:22:43,800
Speaker 2: the rest of their money coming in about two point

427
00:22:43,800 --> 00:22:47,600
Speaker 2: seven billion dollars last year coming from subscriptions to chat GPT.

428
00:22:48,280 --> 00:22:50,719
Speaker 2: If you ignore the hype, open Air and Anthropic are

429
00:22:50,720 --> 00:22:54,800
Speaker 2: actually deeply boring software businesses with unprofitable, unreliable products prone

430
00:22:54,840 --> 00:22:58,120
Speaker 2: to hallucinations, and then new products such as open Aisaura

431
00:22:58,400 --> 00:23:00,440
Speaker 2: cost way too much money to both run and trained

432
00:23:00,440 --> 00:23:03,720
Speaker 2: to get the results that well, they suck. They're not good.

433
00:23:04,200 --> 00:23:06,600
Speaker 2: Even open a Eyes pushing to the federal government with

434
00:23:06,640 --> 00:23:09,600
Speaker 2: the release of chat GPT, GOV is unlikely to reverse

435
00:23:09,600 --> 00:23:13,000
Speaker 2: its dismal fortunes. Seriously, think about it. I'm sure some

436
00:23:13,040 --> 00:23:14,320
Speaker 2: of you are going to say, well, Trump will just

437
00:23:14,359 --> 00:23:17,320
Speaker 2: give them money. These motherfuckers need way more money than

438
00:23:17,359 --> 00:23:19,000
Speaker 2: Trump is going to give them, and why would Trump

439
00:23:19,040 --> 00:23:22,560
Speaker 2: bet on a loser? Why would Trump be like, oh, yeah,

440
00:23:22,600 --> 00:23:24,159
Speaker 2: I'm going to give more money to this company that

441
00:23:24,200 --> 00:23:26,520
Speaker 2: does the same thing. He doesn't understand it this shit,

442
00:23:26,560 --> 00:23:28,520
Speaker 2: and he probably just looks at Samut and goes, now,

443
00:23:28,560 --> 00:23:31,960
Speaker 2: that's a kind of new money I don't like. But

444
00:23:32,920 --> 00:23:35,920
Speaker 2: to make this more than a deeply boring software business,

445
00:23:36,200 --> 00:23:40,360
Speaker 2: Open iron Anthropic needed larger models, and they needed them

446
00:23:40,400 --> 00:23:44,320
Speaker 2: to get larger generally in perpetuity, and for the story

447
00:23:44,400 --> 00:23:46,480
Speaker 2: to always be that there was only one way to

448
00:23:46,520 --> 00:23:49,159
Speaker 2: build the future, and that the future cost hundreds of

449
00:23:49,200 --> 00:23:52,280
Speaker 2: billions of dollars, and that only the biggest geniuses, who

450
00:23:52,320 --> 00:23:54,880
Speaker 2: all happened to work in the same two or three places,

451
00:23:55,000 --> 00:23:58,080
Speaker 2: were capable of doing it. Post deep Seek, there really

452
00:23:58,080 --> 00:24:00,760
Speaker 2: isn't a compelling argument for investing hundreds of billions of

453
00:24:00,800 --> 00:24:03,840
Speaker 2: dollars of capex in data centers, or buying new GPUs,

454
00:24:03,960 --> 00:24:06,760
Speaker 2: or even pursuing large language models as they currently stand.

455
00:24:07,040 --> 00:24:10,000
Speaker 2: It's possible, and deep Seek, through its research papers, explained

456
00:24:10,040 --> 00:24:13,280
Speaker 2: in detail how to build models competitive with both of

457
00:24:13,359 --> 00:24:16,240
Speaker 2: open AI's leading models, and that's assuming you don't simply

458
00:24:16,280 --> 00:24:18,760
Speaker 2: build on top of the ones that deep Seek released.

459
00:24:19,440 --> 00:24:22,120
Speaker 2: It also seriously calls into question what it is you're

460
00:24:22,160 --> 00:24:25,359
Speaker 2: paying open Ai for in its various subscriptions, most of

461
00:24:25,359 --> 00:24:28,040
Speaker 2: which other than the two hundred dollars a month pro subscription,

462
00:24:28,240 --> 00:24:30,360
Speaker 2: have hard limits on how much you can use their

463
00:24:30,400 --> 00:24:34,320
Speaker 2: most advanced reasoning models. One thing we do know, though,

464
00:24:34,560 --> 00:24:36,840
Speaker 2: is that open ai and Anthropic will now have to

465
00:24:36,880 --> 00:24:40,160
Speaker 2: either drop the price of accessing their models and potentially

466
00:24:40,200 --> 00:24:43,399
Speaker 2: even the cost of their subscriptions too. I'd argue that

467
00:24:43,440 --> 00:24:46,240
Speaker 2: despite the significant price difference between O one and deep

468
00:24:46,240 --> 00:24:49,000
Speaker 2: Seek's are one reasoning model, the real danger to both

469
00:24:49,000 --> 00:24:51,680
Speaker 2: open ai and Anthropic is deep Seek V three, which

470
00:24:51,720 --> 00:24:55,280
Speaker 2: competes with GPT four h, which is their general purpose model.

471
00:24:55,320 --> 00:24:57,840
Speaker 2: By the way, and as on recording this episode, by

472
00:24:57,880 --> 00:25:00,240
Speaker 2: the way, news broke the Alibaba, a bohemoth thie out

473
00:25:00,240 --> 00:25:02,560
Speaker 2: of China in its own right, has created its own

474
00:25:02,560 --> 00:25:05,840
Speaker 2: model that outsperforms Deep Seek and yet definitely dive into it.

475
00:25:05,880 --> 00:25:07,600
Speaker 2: But if it's true, it's only going to pile on

476
00:25:07,600 --> 00:25:10,639
Speaker 2: the price pressure, though I kind of wonder what they

477
00:25:10,640 --> 00:25:13,359
Speaker 2: could possibly do. Is it going to be cheaper, because

478
00:25:13,359 --> 00:25:16,280
Speaker 2: if it's more powerful, that's just like it doesn't really

479
00:25:16,359 --> 00:25:21,040
Speaker 2: change shit anyway. Though Deep Seek's narrative shift isn't just

480
00:25:21,080 --> 00:25:25,160
Speaker 2: about commoditizing lllms at large, but commoditizing the most expensive

481
00:25:25,200 --> 00:25:29,240
Speaker 2: ones run by two monopolists backed by three other monopolists.

482
00:25:30,200 --> 00:25:34,320
Speaker 2: I mean, the magic's died. There's no halo around Sam

483
00:25:34,359 --> 00:25:37,200
Speaker 2: Moltman or Dario ama Day's head anymore, as their only

484
00:25:37,280 --> 00:25:39,560
Speaker 2: real argument was we're the only ones that can do this,

485
00:25:39,720 --> 00:25:42,320
Speaker 2: something that nobody should have believed in the first place.

486
00:25:43,160 --> 00:25:45,440
Speaker 2: Up until this point. People believe that the reason these

487
00:25:45,440 --> 00:25:47,879
Speaker 2: models were so expensive was because they had to be,

488
00:25:48,160 --> 00:25:50,199
Speaker 2: and that we had to build more data centers and

489
00:25:50,240 --> 00:25:53,560
Speaker 2: buy more silicon because that's just how things worked. They

490
00:25:53,600 --> 00:25:56,639
Speaker 2: believe that reasoning models were the future, even if members

491
00:25:56,640 --> 00:25:59,080
Speaker 2: of the media didn't really seem to understand what reasoning

492
00:25:59,080 --> 00:26:01,760
Speaker 2: models did or why they mattered, and that as a result,

493
00:26:01,840 --> 00:26:04,120
Speaker 2: they had to be expensive because open Ai and their

494
00:26:04,119 --> 00:26:06,879
Speaker 2: elk were just so fucking smart, even if it wasn't

495
00:26:07,080 --> 00:26:09,600
Speaker 2: obvious what reasoning men or what it allowed you to do,

496
00:26:09,680 --> 00:26:13,639
Speaker 2: or what the products were. It's just very annoying, and

497
00:26:13,680 --> 00:26:15,359
Speaker 2: now we're going to find out, by the way, because

498
00:26:15,400 --> 00:26:18,600
Speaker 2: reasoning is now commoditized along with large language models in general.

499
00:26:19,000 --> 00:26:21,080
Speaker 2: Funnily enough, the way that deep seek may have been

500
00:26:21,119 --> 00:26:24,800
Speaker 2: trained using at least in part synthetic data. Also pushes

501
00:26:24,840 --> 00:26:27,120
Speaker 2: against the paradigm that these companies even need to use

502
00:26:27,160 --> 00:26:30,200
Speaker 2: other people's training data, though their argument, of course will

503
00:26:30,240 --> 00:26:33,840
Speaker 2: be that they need more training data always. We also

504
00:26:33,880 --> 00:26:36,560
Speaker 2: don't know the environmental effects, by the way, with deep SEEK,

505
00:26:36,640 --> 00:26:39,360
Speaker 2: because even if it's cheaper, these models still require those

506
00:26:39,480 --> 00:26:42,560
Speaker 2: energy guzzling GPUs to run, and they're running at full tilt.

507
00:26:43,520 --> 00:26:45,320
Speaker 2: In any case, if I had to guess, the result

508
00:26:45,359 --> 00:26:46,880
Speaker 2: would be that the markets are going to be far

509
00:26:46,960 --> 00:26:50,160
Speaker 2: less tolerant of generative AI, and the idea that generative

510
00:26:50,160 --> 00:27:07,639
Speaker 2: AI is the future. Open AI and anthropic no longer

511
00:27:07,640 --> 00:27:13,200
Speaker 2: really have motes. Unless, well, there's another idea. What if

512
00:27:13,240 --> 00:27:16,520
Speaker 2: there was a huge fucking idiot with a lot of money.

513
00:27:17,200 --> 00:27:21,040
Speaker 2: How about billionaire dipshit Massioshi's son, the CEO of soft Bank,

514
00:27:21,040 --> 00:27:24,159
Speaker 2: a multinational investment firm that's rumored to be investing anywhere

515
00:27:24,160 --> 00:27:27,800
Speaker 2: from fifteen to twenty five billion dollars in Open AI

516
00:27:28,119 --> 00:27:30,960
Speaker 2: in a round that values the company and astonishing three

517
00:27:31,080 --> 00:27:34,119
Speaker 2: hundred and forty billion dollars as part of a round

518
00:27:34,280 --> 00:27:38,600
Speaker 2: of up to forty billion dollars. Now you may think, damn,

519
00:27:38,600 --> 00:27:40,080
Speaker 2: this is a sign that open ai is going to

520
00:27:40,160 --> 00:27:43,000
Speaker 2: make it. But I must remind you how bad soft

521
00:27:43,040 --> 00:27:46,280
Speaker 2: Bank is at investing. They put sixteen billion dollars into

522
00:27:46,320 --> 00:27:49,240
Speaker 2: famously awful real estate company we Work and managed to

523
00:27:49,240 --> 00:27:52,080
Speaker 2: lose I think eight hundred million dollars on the DoorDash

524
00:27:52,119 --> 00:27:54,960
Speaker 2: Ipo and one point eight billion dollars on their investment

525
00:27:55,000 --> 00:27:57,399
Speaker 2: in Uber, and in both cases did so because they

526
00:27:57,400 --> 00:28:00,879
Speaker 2: were desperate to bandwagon on to supposedly your fire bets,

527
00:28:01,160 --> 00:28:04,399
Speaker 2: either just before they crashed away after an investment made sense.

528
00:28:05,119 --> 00:28:07,639
Speaker 2: According to the Wall Street Journal, SoftBank would lead this

529
00:28:07,800 --> 00:28:10,560
Speaker 2: insane forty billion dollar round in the company and would

530
00:28:10,640 --> 00:28:13,560
Speaker 2: and I quote help assemble investors for the rest of

531
00:28:13,600 --> 00:28:16,600
Speaker 2: the round. In doing so, SoftBank would also become open

532
00:28:16,640 --> 00:28:20,760
Speaker 2: AI's largest investor, replacing Microsoft. And yes, SoftBank was the

533
00:28:20,880 --> 00:28:23,640
Speaker 2: largest investor in whe Work before it went tits up.

534
00:28:24,200 --> 00:28:26,520
Speaker 2: It's also important to remember the open ai is pledged

535
00:28:26,560 --> 00:28:28,879
Speaker 2: to put eighteen to nineteen billion dollars to fund the

536
00:28:28,920 --> 00:28:33,160
Speaker 2: Stargate data center project. Along with you guessed SoftBank will

537
00:28:33,160 --> 00:28:36,040
Speaker 2: be committing the same amount. This is on some level,

538
00:28:36,119 --> 00:28:38,800
Speaker 2: soft Bank handing money to itself to invest in data

539
00:28:38,840 --> 00:28:42,000
Speaker 2: centers to prop up an industry that's dying. Now. This

540
00:28:42,080 --> 00:28:44,280
Speaker 2: is a developing story, but it's hard to imagine any

541
00:28:44,320 --> 00:28:48,320
Speaker 2: serious contribution from any respectable investor at this point. My

542
00:28:48,440 --> 00:28:51,479
Speaker 2: money is on a few VC firms desperately scraping at

543
00:28:51,480 --> 00:28:54,080
Speaker 2: the bottom and the barrel quarter of a billion here,

544
00:28:54,240 --> 00:28:56,400
Speaker 2: quarter of a billion there. Maybe in Vidia chucks in

545
00:28:56,480 --> 00:29:00,120
Speaker 2: some and on the subject of barrels of stuff, I

546
00:29:00,120 --> 00:29:03,120
Speaker 2: also expect money from the Kingdom of Saudi Arabia and

547
00:29:03,160 --> 00:29:06,200
Speaker 2: its associated venture arms. You're going to see a few

548
00:29:06,320 --> 00:29:09,400
Speaker 2: maybe Andrews and Horowitz gets involved. They don't think much.

549
00:29:10,040 --> 00:29:12,520
Speaker 2: It's just very fucking silly, and I don't know how

550
00:29:12,560 --> 00:29:15,560
Speaker 2: this works out. Open AI burns money, and even if

551
00:29:15,560 --> 00:29:18,840
Speaker 2: they somehow make more efficient models, the actual total addressable

552
00:29:18,880 --> 00:29:22,440
Speaker 2: market of generative AI is actually pretty small. Microsoft said

553
00:29:22,440 --> 00:29:25,600
Speaker 2: in their recent earnings they made twelve to thirteen billion

554
00:29:25,600 --> 00:29:28,720
Speaker 2: dollars of arr on AI. Just to be clear, that's

555
00:29:28,720 --> 00:29:31,760
Speaker 2: not profit and that's not a business unit. There's no

556
00:29:31,800 --> 00:29:34,200
Speaker 2: AI business unit, which means that that is just spread

557
00:29:34,240 --> 00:29:39,080
Speaker 2: across delivering cloud compute for AI copilot on Microsoft three

558
00:29:39,120 --> 00:29:41,600
Speaker 2: sixty five products, which by the way, no one likes.

559
00:29:41,640 --> 00:29:46,160
Speaker 2: They're having trouble selling and other associated copilot products they sell.

560
00:29:47,080 --> 00:29:50,040
Speaker 2: I don't think and like twelve to thirteen billion dollars

561
00:29:50,080 --> 00:29:53,400
Speaker 2: across four quarters. That's not actually good at all. It's

562
00:29:53,440 --> 00:29:55,440
Speaker 2: just very silly. All of this is so silly, and

563
00:29:55,480 --> 00:29:57,000
Speaker 2: when I think about it, too hard to feel a

564
00:29:57,040 --> 00:30:01,640
Speaker 2: little crazy. Open Ai makes three seven billion dollars in revenue,

565
00:30:01,640 --> 00:30:03,960
Speaker 2: and they do so, as I mentioned, primarily from chad

566
00:30:04,000 --> 00:30:07,480
Speaker 2: GPT subscriptions. Even if that somehow and it won't by

567
00:30:07,520 --> 00:30:10,560
Speaker 2: the way, turns into three point seven billion dollars of

568
00:30:10,600 --> 00:30:12,960
Speaker 2: profit or even ten billion dollars of profit a year,

569
00:30:13,240 --> 00:30:15,880
Speaker 2: that's less than the profits in a single quarter of

570
00:30:15,880 --> 00:30:20,000
Speaker 2: any given hyperscaler. It would be respectable, sure, But three

571
00:30:20,040 --> 00:30:22,960
Speaker 2: hundred and forty billion dollar valuation, I guess makes sense

572
00:30:22,960 --> 00:30:25,760
Speaker 2: if it was profit. It doesn't make sense if it's

573
00:30:25,800 --> 00:30:29,200
Speaker 2: not profitable. Though it also isn't obvious how open ai

574
00:30:29,240 --> 00:30:32,280
Speaker 2: would actually provide any liquidity to investors, by which I

575
00:30:32,280 --> 00:30:36,280
Speaker 2: mean allow them to sell their stock beyond selling shares

576
00:30:36,720 --> 00:30:40,320
Speaker 2: of people that work at open ai, like people who

577
00:30:40,360 --> 00:30:42,719
Speaker 2: work there who have been given stock runts selling that

578
00:30:42,800 --> 00:30:46,880
Speaker 2: to another investor, a really dumb guy maybe, And as

579
00:30:46,880 --> 00:30:49,560
Speaker 2: in the side, SoftBank bought one point five billion dollars

580
00:30:49,560 --> 00:30:51,720
Speaker 2: of stock from open ai employees and the tender at

581
00:30:51,720 --> 00:30:53,680
Speaker 2: the end of November twenty twenty four. Just a note

582
00:30:53,680 --> 00:30:56,840
Speaker 2: for you. They could also take the company public, but

583
00:30:56,920 --> 00:31:00,440
Speaker 2: with the unit economics of this fucking company, which boil

584
00:31:00,520 --> 00:31:03,000
Speaker 2: down by the way to our products, lose billions of

585
00:31:03,000 --> 00:31:05,479
Speaker 2: dollars and are extremely compoditized. I'm not really sure what

586
00:31:05,480 --> 00:31:08,280
Speaker 2: the plan is here. The fundamental problems that open ai

587
00:31:08,400 --> 00:31:11,200
Speaker 2: has are not solved by throwing more money at the problem.

588
00:31:11,360 --> 00:31:13,560
Speaker 2: This hasn't worked before and it won't work this time.

589
00:31:13,640 --> 00:31:16,880
Speaker 2: They're burning cash and in soft banks case, it isn't

590
00:31:16,920 --> 00:31:19,920
Speaker 2: obvious what it is they're getting from open ai. Is

591
00:31:19,920 --> 00:31:22,120
Speaker 2: it the chance to continue an industry wide con the

592
00:31:22,200 --> 00:31:24,800
Speaker 2: chance to participate in a catalyst death cult. I don't know.

593
00:31:25,120 --> 00:31:26,800
Speaker 2: Maybe it's the chance to burn money at a faster

594
00:31:26,880 --> 00:31:29,920
Speaker 2: rate than we work ever could dream of. Will this

595
00:31:30,000 --> 00:31:32,760
Speaker 2: be the time that Microsoft, Amazon, and Google just drop

596
00:31:32,800 --> 00:31:35,200
Speaker 2: open ai and Anthropic and make their own models based

597
00:31:35,240 --> 00:31:38,440
Speaker 2: on deep Seek's work. What incentive is there for the

598
00:31:38,520 --> 00:31:42,360
Speaker 2: hyperscalers to keep funding open ai and Anthropic. They hold

599
00:31:42,400 --> 00:31:44,800
Speaker 2: all the cards. The GPU is the infrastructure, and in

600
00:31:44,840 --> 00:31:48,200
Speaker 2: the case of Microsoft, non revocable licenses that permit them

601
00:31:48,440 --> 00:31:51,440
Speaker 2: unfettered use and access to open AI's tech, and there's

602
00:31:51,480 --> 00:31:54,240
Speaker 2: little stopping the hyperscalers from building their own models and

603
00:31:54,280 --> 00:31:58,640
Speaker 2: just dumping them entirely. In fact, Microsoft might actually be

604
00:31:58,680 --> 00:32:00,800
Speaker 2: a little glad to see soft bare become the biggest

605
00:32:00,800 --> 00:32:03,560
Speaker 2: investor and pick up the tab for open AI's expenses.

606
00:32:03,720 --> 00:32:06,520
Speaker 2: I can imagine satching Adella texting Semilt and being like, no,

607
00:32:06,640 --> 00:32:10,680
Speaker 2: don't take that money. Hell, well, don't do oh I'd

608
00:32:10,720 --> 00:32:13,440
Speaker 2: hate that. I'd hate if this was someone else's problem.

609
00:32:13,880 --> 00:32:16,720
Speaker 2: And the stargate thing, by the way, is an attempt

610
00:32:16,800 --> 00:32:18,880
Speaker 2: to that up to five hundred billion thing. It's just

611
00:32:19,000 --> 00:32:23,480
Speaker 2: bollocks whatever. It's an attempt to remove themselves from Microsoft.

612
00:32:23,760 --> 00:32:26,760
Speaker 2: And Microsoft actually allowed open ai to alter their deal

613
00:32:26,840 --> 00:32:29,360
Speaker 2: so that they could get cloud compute from others. Now

614
00:32:29,400 --> 00:32:32,880
Speaker 2: at the time, people were like, yeah, man, this is

615
00:32:32,880 --> 00:32:35,720
Speaker 2: a good thing. This shows open ai will be independent. No,

616
00:32:35,880 --> 00:32:38,400
Speaker 2: it doesn't. It just means that they're gonna be under

617
00:32:38,720 --> 00:32:42,440
Speaker 2: Masayoshi's son, now the funniest dumbest man in investing. I

618
00:32:42,480 --> 00:32:45,080
Speaker 2: love Massa Yoshi's son. I think it's nice that we

619
00:32:45,160 --> 00:32:49,400
Speaker 2: have an insane guy who isn't instantly murderous in our

620
00:32:49,440 --> 00:32:55,240
Speaker 2: lives anyway. Anyway, though, as I've said before, I believe

621
00:32:55,240 --> 00:32:58,280
Speaker 2: we're at PKI and now that generative AI has been commoditized,

622
00:32:58,320 --> 00:33:00,600
Speaker 2: the only thing that open AI and Anthropy have left

623
00:33:00,680 --> 00:33:03,960
Speaker 2: other than a pile of cash is their ability to innovate,

624
00:33:04,360 --> 00:33:07,920
Speaker 2: and I don't think they're capable of doing so. And

625
00:33:07,960 --> 00:33:10,080
Speaker 2: because we sit in the ruins of Silicon Valley with

626
00:33:10,120 --> 00:33:12,440
Speaker 2: our biggest startups all doing the same thing in the

627
00:33:12,520 --> 00:33:15,200
Speaker 2: least efficient way possible, living at the beck and call

628
00:33:15,320 --> 00:33:18,480
Speaker 2: of public companies with multi trillion dollar market caps, everyone

629
00:33:18,560 --> 00:33:21,080
Speaker 2: is trying to do the same thing in the same way,

630
00:33:21,120 --> 00:33:24,520
Speaker 2: based on the fantastical marketing nonsense of a succession of

631
00:33:24,560 --> 00:33:27,760
Speaker 2: directionless rich guys that all want to create Americas next

632
00:33:27,800 --> 00:33:31,080
Speaker 2: top monopoly. It's time to wake up and accept that

633
00:33:31,120 --> 00:33:33,840
Speaker 2: there was never any kind of AI arms race, and

634
00:33:33,880 --> 00:33:36,240
Speaker 2: that the only reason that Hyperscale has built so many

635
00:33:36,320 --> 00:33:39,120
Speaker 2: data centers and bought so many GPUs is because they're

636
00:33:39,200 --> 00:33:42,040
Speaker 2: run by people that don't experience real problems and thus

637
00:33:42,160 --> 00:33:46,320
Speaker 2: don't know what problems real people face. Generator AI does

638
00:33:46,360 --> 00:33:49,280
Speaker 2: not solve any trillion dollar problems, nor does it create

639
00:33:49,320 --> 00:33:53,320
Speaker 2: outcomes that are profitable for any particular business. Deep Seeks

640
00:33:53,360 --> 00:33:55,760
Speaker 2: models are cheaper to run. But the real magic trick

641
00:33:55,800 --> 00:33:58,400
Speaker 2: they pulled is that they showed how utterly replaceable a

642
00:33:58,440 --> 00:34:01,760
Speaker 2: company like open AI and buy extension any LLM company

643
00:34:01,840 --> 00:34:05,000
Speaker 2: really is. There Really isn't anything special about any of

644
00:34:05,040 --> 00:34:08,719
Speaker 2: these companies. They have no moat, their infrastructural advantages moot,

645
00:34:08,840 --> 00:34:12,120
Speaker 2: and their hordes of talent are relatively irrelevant. What Deep

646
00:34:12,160 --> 00:34:16,280
Speaker 2: Seek is proven isn't just technological, it's philosophical. It shows

647
00:34:16,280 --> 00:34:19,280
Speaker 2: that the scrappy spirit of Silicon Valley builders is dead,

648
00:34:19,600 --> 00:34:22,480
Speaker 2: replaced by a series of different management consultants that lead

649
00:34:22,520 --> 00:34:25,759
Speaker 2: teams of engineers to do things based on vibes. You

650
00:34:25,840 --> 00:34:28,840
Speaker 2: may ask if all of this means generative AI suddenly

651
00:34:28,880 --> 00:34:31,840
Speaker 2: gets more prevalent after all. Sech in Adella of Microsoft

652
00:34:31,920 --> 00:34:35,000
Speaker 2: quoted Yvonne's paradox, which posits that when resources are made

653
00:34:35,080 --> 00:34:39,240
Speaker 2: more efficient, their use increases. Sadly, I hypothesize that something

654
00:34:39,320 --> 00:34:42,120
Speaker 2: else happens right now. I do not believe that there

655
00:34:42,120 --> 00:34:45,040
Speaker 2: are companies that are stymied by the pricing the Open

656
00:34:45,080 --> 00:34:47,279
Speaker 2: AI and their ILK coffer, Nor do I think there

657
00:34:47,280 --> 00:34:49,919
Speaker 2: are many companies or use cases that don't exist because

658
00:34:50,000 --> 00:34:53,880
Speaker 2: large language models are too expensive. AI companies took up

659
00:34:53,880 --> 00:34:56,440
Speaker 2: a third of all venture capital funding last year, and

660
00:34:56,480 --> 00:34:59,200
Speaker 2: on top of that, it's fairly try reasoning models like

661
00:34:59,239 --> 00:35:01,680
Speaker 2: O one and make a proof of concept without having

662
00:35:01,760 --> 00:35:05,280
Speaker 2: to make an entire operational company. Sheer open ai barely

663
00:35:05,320 --> 00:35:07,640
Speaker 2: has one. I don't think anyone has been on the

664
00:35:07,719 --> 00:35:11,160
Speaker 2: sidelines of generative II due to costs, and remember, few

665
00:35:11,160 --> 00:35:12,600
Speaker 2: seem to be able to come up with great use

666
00:35:12,640 --> 00:35:15,359
Speaker 2: case for one or other reasoning models anyway, and deep

667
00:35:15,400 --> 00:35:19,640
Speaker 2: Seek's models, while cheaper, don't have any new functionality. As

668
00:35:19,680 --> 00:35:22,759
Speaker 2: a result, I don't really see anything changing beyond the

669
00:35:22,760 --> 00:35:25,399
Speaker 2: eventual collapse of the API market, which is the way

670
00:35:25,440 --> 00:35:28,360
Speaker 2: you plug these models into things. For companies like Anthropic

671
00:35:28,400 --> 00:35:32,560
Speaker 2: and open AI, large language models and reasoning models the niche.

672
00:35:32,719 --> 00:35:35,080
Speaker 2: The only reason that chat GPT became such a big

673
00:35:35,080 --> 00:35:37,680
Speaker 2: deal is because the tech industry is no other growth ideas,

674
00:35:37,880 --> 00:35:41,160
Speaker 2: and despite the entire industry and public markets screaming about it,

675
00:35:41,480 --> 00:35:44,520
Speaker 2: I can't think of any mass market product that really matters.

676
00:35:45,400 --> 00:35:48,160
Speaker 2: Even if deep Seek doesn't land the fatal blow, it

677
00:35:48,160 --> 00:35:51,200
Speaker 2: could set the foundations from another company to drag open

678
00:35:51,239 --> 00:35:53,480
Speaker 2: AI's carcass out behind the barn and hit it with

679
00:35:53,520 --> 00:35:56,760
Speaker 2: a big stick. One way in which this entire fast

680
00:35:56,760 --> 00:35:59,759
Speaker 2: could fall as if nasty Mark Zuckerberg decides he wants

681
00:35:59,760 --> 00:36:04,160
Speaker 2: to splay destroy the entire market for llms. Meta has

682
00:36:04,160 --> 00:36:07,399
Speaker 2: already formed four separate war rooms to break down how

683
00:36:07,440 --> 00:36:10,040
Speaker 2: Deep Seep did it and apparently to quote the information

684
00:36:10,440 --> 00:36:13,360
Speaker 2: in pursuing Lama, which is their large language model. CEO

685
00:36:13,440 --> 00:36:16,440
Speaker 2: Mark Zuckerberg wants to commoditize AI models so that the

686
00:36:16,480 --> 00:36:20,320
Speaker 2: applications that use such models, including metas, generate more money

687
00:36:20,320 --> 00:36:23,680
Speaker 2: than the sales of the AI models themselves. That could

688
00:36:23,760 --> 00:36:26,560
Speaker 2: hurt metas AI rivals such as open ai and Anthropic,

689
00:36:26,719 --> 00:36:28,799
Speaker 2: which were on pace to generate billions of dollars in

690
00:36:28,840 --> 00:36:32,840
Speaker 2: revenue from such sales and lose billions. Fucking hell. I

691
00:36:32,920 --> 00:36:35,680
Speaker 2: love the information, but can you add the most important bit?

692
00:36:36,920 --> 00:36:39,760
Speaker 2: But I could absolutely see Meta releasing its own version

693
00:36:39,760 --> 00:36:42,439
Speaker 2: of deep Seek's models. They've got the GPUs and Mark

694
00:36:42,480 --> 00:36:44,800
Speaker 2: Zuckerberg can never be fired, meaning that if he simply

695
00:36:44,840 --> 00:36:48,279
Speaker 2: decided to throw billions of dollars into specifically creating his

696
00:36:48,360 --> 00:36:51,640
Speaker 2: own deep discounted alms to wipe out open ai he

697
00:36:51,760 --> 00:36:56,160
Speaker 2: absolutely could. After all, a few weeks back, Mark Zuckerberg

698
00:36:56,239 --> 00:36:58,759
Speaker 2: said that Meta would spend between sixty and sixty five

699
00:36:58,840 --> 00:37:01,799
Speaker 2: billion dollars in CAPITALGICS spenditures in twenty twenty five. And

700
00:37:01,840 --> 00:37:04,560
Speaker 2: this was before the deep Seek situation. And I imagine

701
00:37:04,560 --> 00:37:07,000
Speaker 2: the markets would love a more modest proposal that involves

702
00:37:07,040 --> 00:37:09,640
Speaker 2: Meta offering a chat GPT be to simply to fuck

703
00:37:09,719 --> 00:37:13,640
Speaker 2: over Sam Altman. And that's the thing. Chat GPT is

704
00:37:13,680 --> 00:37:17,279
Speaker 2: big because everybody's talking about AI, and chat GPT is

705
00:37:17,360 --> 00:37:20,279
Speaker 2: the big brand in AI. It is not essential, and

706
00:37:20,320 --> 00:37:22,319
Speaker 2: it's only been treated as such because the media and

707
00:37:22,360 --> 00:37:25,200
Speaker 2: the markets ran away with a narrative that they barely understood.

708
00:37:25,800 --> 00:37:29,560
Speaker 2: Deep Seek pierced that narrative because believing said narrative also

709
00:37:29,640 --> 00:37:31,920
Speaker 2: required you to believe that Sam Moortman is a magician

710
00:37:32,239 --> 00:37:35,040
Speaker 2: versus an extremely shitty CEO that burned a bunch of money.

711
00:37:35,520 --> 00:37:40,160
Speaker 2: And I don't believe that, even before Deep Seek, that

712
00:37:40,280 --> 00:37:44,480
Speaker 2: Altman's peers really bought into the hype. Sure you can

713
00:37:44,600 --> 00:37:47,160
Speaker 2: argue that deep Seek just built on top of software

714
00:37:47,160 --> 00:37:50,120
Speaker 2: that already existed thanks to open Ai. Thank you Casey,

715
00:37:50,200 --> 00:37:53,480
Speaker 2: by the way, but this beg's a fairly obvious question

716
00:37:54,040 --> 00:37:57,719
Speaker 2: why didn't open ai build on top of software invented

717
00:37:57,760 --> 00:38:01,040
Speaker 2: by open ai? And here's another question, why is it

718
00:38:01,080 --> 00:38:05,800
Speaker 2: goddamn matter? In any case, the massive expense of running

719
00:38:05,880 --> 00:38:08,800
Speaker 2: generative models hasn't been the limiter on their deployment, of

720
00:38:08,840 --> 00:38:11,000
Speaker 2: their success. You can blame that on the fact that

721
00:38:11,040 --> 00:38:13,920
Speaker 2: they as a piece of technology and neither artificial intelligence

722
00:38:13,960 --> 00:38:16,680
Speaker 2: nor capable of providing the kind of meaningful outcomes that

723
00:38:16,719 --> 00:38:20,680
Speaker 2: would make them the next smartphone or cloud computing. Honestly,

724
00:38:20,719 --> 00:38:22,960
Speaker 2: it's all been a con It's been a painfully obvious

725
00:38:22,960 --> 00:38:25,400
Speaker 2: one one had been screaming about since February, since when

726
00:38:25,440 --> 00:38:28,600
Speaker 2: I started this podcast, trying to explain that beneath the

727
00:38:28,680 --> 00:38:31,920
Speaker 2: hype was an industry that provided monister at best outcomes

728
00:38:32,000 --> 00:38:35,600
Speaker 2: rather than any kind of next big thing. Without reasoning

729
00:38:35,719 --> 00:38:38,640
Speaker 2: is its magical new creation. Open ai really doesn't have

730
00:38:38,680 --> 00:38:42,520
Speaker 2: anything left. Agents aren't coming, large language models aren't going

731
00:38:42,560 --> 00:38:45,839
Speaker 2: to build them. AGI isn't coming either. There's no proof

732
00:38:45,840 --> 00:38:48,640
Speaker 2: it's possible. All of this is fucking flim flamm to

733
00:38:48,640 --> 00:38:51,840
Speaker 2: cover up how mediocre and unreliable the fundament of the

734
00:38:51,880 --> 00:38:56,799
Speaker 2: supposed AI revolution really was. All of this money, all

735
00:38:56,840 --> 00:38:59,719
Speaker 2: of this energy, and all of this talent was wasted

736
00:39:00,160 --> 00:39:03,239
Speaker 2: thanks to markets that don't actually do anything, markets that

737
00:39:03,360 --> 00:39:06,360
Speaker 2: don't make for good companies, just growth hogs, and a

738
00:39:06,440 --> 00:39:09,160
Speaker 2: media industry that fails to hold the powerful to account.

739
00:39:10,480 --> 00:39:13,120
Speaker 2: And it looks like everything got broken by some random

740
00:39:13,160 --> 00:39:16,759
Speaker 2: outgrowth of a Chinese hedge fund. It's so ridiculous, it's

741
00:39:16,760 --> 00:39:19,360
Speaker 2: so sickening. I can't believe it. While I can totally

742
00:39:19,400 --> 00:39:21,560
Speaker 2: believe it, I'm actually surprised they didn't come up with

743
00:39:21,600 --> 00:39:24,400
Speaker 2: this idea myself, Just the idea that someone could do

744
00:39:24,440 --> 00:39:27,120
Speaker 2: this cheaper. It makes me go insane. And what's more

745
00:39:27,160 --> 00:39:29,120
Speaker 2: insane is that Open Eye is still going to be

746
00:39:29,120 --> 00:39:31,760
Speaker 2: able to raise that round. But I think we're approaching

747
00:39:31,800 --> 00:39:33,360
Speaker 2: the end of days. I'm not calling the end of

748
00:39:33,360 --> 00:39:35,560
Speaker 2: the bubble yet. I refuse to do that. I'm not

749
00:39:35,600 --> 00:39:37,640
Speaker 2: going to do that. What I am going to say

750
00:39:37,680 --> 00:39:39,759
Speaker 2: is it's deflating, and I am going to say I

751
00:39:39,760 --> 00:39:42,560
Speaker 2: have no idea how they reinflate it. A bunch more

752
00:39:42,600 --> 00:39:45,520
Speaker 2: money isn't going to change anything. These companies are washed.

753
00:39:45,640 --> 00:39:49,880
Speaker 2: Sam Morton's washed. He's the Mark Sanchez of the tech industry,

754
00:39:51,160 --> 00:39:53,080
Speaker 2: and he's so sickening. All of them are so sickening.

755
00:39:53,160 --> 00:39:55,560
Speaker 2: Imagine if this money had gone anywhere else. Imagine if

756
00:39:55,600 --> 00:39:58,080
Speaker 2: it had gone into batteries. Imagine if it had gone

757
00:39:58,080 --> 00:40:02,120
Speaker 2: into climate stuff. Imagine if gone somewhere useful. Imagine if

758
00:40:02,120 --> 00:40:05,480
Speaker 2: instead of spending billions on this dog shit, they actually

759
00:40:05,560 --> 00:40:08,600
Speaker 2: fix their problems, they actually fix the products that they've

760
00:40:08,640 --> 00:40:11,600
Speaker 2: made worse. But the problem is that the rot economy

761
00:40:11,640 --> 00:40:14,560
Speaker 2: is in control. That the growth at all costs mindset

762
00:40:14,880 --> 00:40:18,759
Speaker 2: is all that you see in the tech industry, and

763
00:40:18,840 --> 00:40:22,400
Speaker 2: Silicon Valley needs to repent. Silicon Valley needs to change

764
00:40:22,440 --> 00:40:25,160
Speaker 2: its ways because when the bubble bursts, and I really

765
00:40:25,160 --> 00:40:29,239
Speaker 2: think it will, the destruction that follows will be horrifying

766
00:40:29,360 --> 00:40:32,000
Speaker 2: and it will hit workers. It will hit tens of

767
00:40:32,040 --> 00:40:35,320
Speaker 2: thousands of tech workers, and it will affect the markets.

768
00:40:36,040 --> 00:40:38,240
Speaker 2: And after that, the markets are going to realize something.

769
00:40:38,719 --> 00:40:40,759
Speaker 2: The tech industry doesn't have anything left, they don't have

770
00:40:40,840 --> 00:40:44,880
Speaker 2: another growth market. They're out. They're all out. And I

771
00:40:44,920 --> 00:40:47,279
Speaker 2: look forward to telling you how I look forward to

772
00:40:47,320 --> 00:40:49,840
Speaker 2: talking about it when it happens. And I'm so grateful

773
00:40:49,840 --> 00:41:00,160
Speaker 2: for you listening. Thank you for listening to Better Offline.

774
00:41:00,400 --> 00:41:02,799
Speaker 3: The editor and composer of the Better Offline theme song,

775
00:41:02,880 --> 00:41:05,520
Speaker 3: is Matasowski. You can check out more of his music

776
00:41:05,560 --> 00:41:09,200
Speaker 3: and audio projects at Matasowski dot com, M A T

777
00:41:09,200 --> 00:41:13,680
Speaker 3: T O S O W s ki dot com. You

778
00:41:13,680 --> 00:41:16,200
Speaker 3: can email me at easy at better offline dot com

779
00:41:16,280 --> 00:41:18,600
Speaker 3: or visit better offline dot com to find more podcast

780
00:41:18,640 --> 00:41:21,960
Speaker 3: links and of course, my newsletter. I also really recommend

781
00:41:22,000 --> 00:41:23,960
Speaker 3: you go to chat dot where's youreaed dot at to

782
00:41:24,040 --> 00:41:26,840
Speaker 3: visit the discord, and go to our slash Better Offline

783
00:41:26,880 --> 00:41:30,080
Speaker 3: to check out our reddit. Thank you so much for listening.

784
00:41:30,920 --> 00:41:33,600
Speaker 3: Better Offline is a production of cool Zone Media. For

785
00:41:33,719 --> 00:41:36,880
Speaker 3: more from cool Zone Media, visit our website cool Zonemedia

786
00:41:36,960 --> 00:41:39,799
Speaker 3: dot com, or check us out on the iHeartRadio app,

787
00:41:39,840 --> 00:41:42,520
Speaker 3: Apple Podcasts, or wherever you get your podcasts.