1
00:00:02,440 --> 00:00:08,320
Speaker 1: As media your witness to a great becoming. It's better

2
00:00:08,360 --> 00:00:23,520
Speaker 1: offline and I'm ed Zetron. Today we're joined by computer

3
00:00:23,600 --> 00:00:26,960
Speaker 1: science professor and tech writer Caln Newport for Hater Season.

4
00:00:27,040 --> 00:00:28,280
Speaker 1: Cal thank you for joining me.

5
00:00:29,240 --> 00:00:30,600
Speaker 2: Always happy to do some heding.

6
00:00:30,640 --> 00:00:34,200
Speaker 1: I suppose, well, you had an excellent YouTube that I'll

7
00:00:34,240 --> 00:00:37,239
Speaker 1: be linking in the show notes about the mistakes in

8
00:00:37,320 --> 00:00:40,440
Speaker 1: AI reporting. Though I would my hater in me says,

9
00:00:40,479 --> 00:00:43,720
Speaker 1: I don't think these are mistakes, but you really you touched.

10
00:00:43,720 --> 00:00:46,040
Speaker 1: One of the best videos I've seen on AI report

11
00:00:46,200 --> 00:00:48,120
Speaker 1: or AI in general was what you did. But it's

12
00:00:48,159 --> 00:00:51,479
Speaker 1: basically this thing of like the digital iic that these

13
00:00:51,520 --> 00:00:53,720
Speaker 1: stories are meant to make you feel uncomfortable, and of

14
00:00:53,760 --> 00:00:56,640
Speaker 1: course this full astonishment thing. You know what, Just run

15
00:00:56,680 --> 00:00:58,800
Speaker 1: the tape. Tell me a little bit about what the

16
00:00:58,840 --> 00:01:03,320
Speaker 1: bits that you found, because I watched it going yeah, yeah, yeah,

17
00:01:03,440 --> 00:01:04,679
Speaker 1: like an angry person.

18
00:01:05,160 --> 00:01:07,440
Speaker 2: Yeah I had that. I could imagine you. When I

19
00:01:07,480 --> 00:01:09,039
Speaker 2: was recording that video, I was like, I bet ed

20
00:01:09,160 --> 00:01:15,240
Speaker 2: is well now the great time exactly? Well, I mean,

21
00:01:15,319 --> 00:01:17,039
Speaker 2: let me just give the context, right, So I'm in

22
00:01:17,080 --> 00:01:19,600
Speaker 2: an interesting situation for observing this because you know, I

23
00:01:19,600 --> 00:01:22,840
Speaker 2: am a computer scientist, so I'm not afraid of the technologies.

24
00:01:22,880 --> 00:01:27,280
Speaker 2: I'm happy to talk about transformers and feed forward networks

25
00:01:27,319 --> 00:01:29,600
Speaker 2: and diffusion models and like, that's not that scary to me.

26
00:01:29,880 --> 00:01:33,120
Speaker 2: But I also write about technology. My main journalistic home

27
00:01:33,160 --> 00:01:34,720
Speaker 2: is to New Yorker, where I write a lot about

28
00:01:34,800 --> 00:01:37,720
Speaker 2: you know, I do AI journalism there, so I'm up

29
00:01:37,760 --> 00:01:41,200
Speaker 2: on that as well. And so I'm often noticing things

30
00:01:41,319 --> 00:01:47,200
Speaker 2: in journalism. When I see AI coverage that is faulty,

31
00:01:47,280 --> 00:01:49,360
Speaker 2: it really catches my attention because I have a foot

32
00:01:49,360 --> 00:01:51,360
Speaker 2: in both of these worlds. So there's a lot of

33
00:01:51,360 --> 00:01:54,040
Speaker 2: good AI report out there. There's also a lot of trash,

34
00:01:54,200 --> 00:01:56,200
Speaker 2: and I really wanted to help people figure out how

35
00:01:56,200 --> 00:01:57,920
Speaker 2: do you sort it? Like how do you figure out

36
00:01:58,400 --> 00:02:01,520
Speaker 2: if you're reading something? Should I pull the ripcord on

37
00:02:01,560 --> 00:02:03,440
Speaker 2: this article? Like this is not helping me, Like how

38
00:02:03,480 --> 00:02:04,400
Speaker 2: do you know if it's good or not? So I

39
00:02:04,440 --> 00:02:05,680
Speaker 2: was like, Okay, here's what I'll do is I'll come

40
00:02:05,760 --> 00:02:08,600
Speaker 2: up with the three most common traps I see in

41
00:02:08,680 --> 00:02:10,720
Speaker 2: AI reporting that makes me want to, you know, throw

42
00:02:10,760 --> 00:02:13,360
Speaker 2: my iPad at the wall. And I came up with

43
00:02:13,400 --> 00:02:15,040
Speaker 2: three and I'll just give you the three names. I

44
00:02:15,080 --> 00:02:17,880
Speaker 2: made up. All these names don't. I don't think they're great.

45
00:02:17,919 --> 00:02:19,960
Speaker 2: My producer thinks he loves them. But here we go

46
00:02:20,480 --> 00:02:24,280
Speaker 2: vibe reporting that's number one. So that is where you

47
00:02:25,360 --> 00:02:30,680
Speaker 2: will omit certain facts and put loosely related quotes next

48
00:02:30,680 --> 00:02:33,160
Speaker 2: to each other in a way that creates a general

49
00:02:33,320 --> 00:02:35,080
Speaker 2: vibe that you want to be true but it's not

50
00:02:35,160 --> 00:02:39,280
Speaker 2: quite true. So you don't actually make a concrete claim

51
00:02:39,320 --> 00:02:43,760
Speaker 2: that's not true, but you imply that claim by what

52
00:02:43,800 --> 00:02:45,800
Speaker 2: you omit or what you put in your story, or

53
00:02:45,800 --> 00:02:48,200
Speaker 2: what quotes you put next to it. So you'll put

54
00:02:48,200 --> 00:02:52,440
Speaker 2: a quote, for example, about layoffs at the gaming division

55
00:02:52,480 --> 00:02:56,160
Speaker 2: at Microsoft, next to an unrelated quote about concerns about

56
00:02:56,160 --> 00:03:00,320
Speaker 2: AI and it's impact on jobs. Now you have the vibe, oh, man,

57
00:03:00,360 --> 00:03:02,200
Speaker 2: all these people just got laid off because of AI.

58
00:03:02,320 --> 00:03:04,560
Speaker 2: AI is taking jobs, where in reality the layouts had

59
00:03:04,600 --> 00:03:06,040
Speaker 2: nothing to do with AI. But you put these things

60
00:03:06,040 --> 00:03:07,600
Speaker 2: next to each other, you give that sense. Then I

61
00:03:07,639 --> 00:03:11,480
Speaker 2: had a mining digital ick. So to me, that's any

62
00:03:11,520 --> 00:03:13,640
Speaker 2: AI story where you take an example from the edges

63
00:03:13,680 --> 00:03:16,000
Speaker 2: of AI, like something that wireheads and San Francisco are

64
00:03:16,080 --> 00:03:18,880
Speaker 2: up to, and you just tell a story that's unsettling

65
00:03:19,600 --> 00:03:22,520
Speaker 2: without talking about any of the technical details like well,

66
00:03:22,560 --> 00:03:24,880
Speaker 2: what's different here? Was there a technical innovation we need

67
00:03:24,919 --> 00:03:28,639
Speaker 2: to know about, and not discussing any concrete implications. Oh,

68
00:03:28,680 --> 00:03:30,680
Speaker 2: this means this is going to change in the future,

69
00:03:30,840 --> 00:03:32,400
Speaker 2: or it's going to have an impact on this sector.

70
00:03:32,440 --> 00:03:35,600
Speaker 2: You're just telling a story to unsettle, and I think

71
00:03:35,600 --> 00:03:38,520
Speaker 2: a lot of the coverage of molt Book and open

72
00:03:38,560 --> 00:03:40,880
Speaker 2: Claw fell into that. And then finally this astonishment, which

73
00:03:40,920 --> 00:03:44,040
Speaker 2: is more of a YouTube phenomenon than a print journalism phenomenon.

74
00:03:44,080 --> 00:03:46,360
Speaker 2: But that's where every single thing that happens in AI

75
00:03:46,720 --> 00:03:52,600
Speaker 2: is insane, amazing, terrifying. Everything is going to change, and

76
00:03:52,640 --> 00:03:57,720
Speaker 2: so you constantly create this atmosphere of something seismic just happened,

77
00:03:58,240 --> 00:04:00,680
Speaker 2: so that the consumer of the infra ends up in

78
00:04:00,680 --> 00:04:02,200
Speaker 2: a bit of a panic, like I can put my

79
00:04:02,200 --> 00:04:05,120
Speaker 2: finger on exactly what's terrible, but like everything terrible is happening.

80
00:04:05,520 --> 00:04:08,760
Speaker 2: Those three traps, to me, should be automatic rip chords

81
00:04:08,760 --> 00:04:09,840
Speaker 2: from what you're reading or watching.

82
00:04:11,040 --> 00:04:13,119
Speaker 1: The only disagreement I'll have is that you would say

83
00:04:13,120 --> 00:04:15,840
Speaker 1: that this isn't the majority of AI journalism, because I

84
00:04:15,840 --> 00:04:18,880
Speaker 1: actually argue it would be, especially that kind of the

85
00:04:19,640 --> 00:04:23,600
Speaker 1: beginning one. The vibe reporting is very common because The

86
00:04:23,640 --> 00:04:26,440
Speaker 1: big one right now I'm seeing is this AI software

87
00:04:26,760 --> 00:04:29,839
Speaker 1: AI is replacing software thing. If you are a reporter

88
00:04:30,040 --> 00:04:32,280
Speaker 1: and cal is not making this statement I am. If

89
00:04:32,279 --> 00:04:37,440
Speaker 1: you're a reporter and you bring up anthropic cowork around anything,

90
00:04:38,080 --> 00:04:40,880
Speaker 1: you are wrong. I was about to call you a name,

91
00:04:40,880 --> 00:04:43,800
Speaker 1: but I'm being nice today for some reason. All Right,

92
00:04:43,800 --> 00:04:46,120
Speaker 1: you're a dip shit, because you are a dip shit.

93
00:04:46,160 --> 00:04:48,560
Speaker 1: If you look at claude cowork, which is a thing

94
00:04:48,640 --> 00:04:50,719
Speaker 1: for fucking around on your desktop, and you say this

95
00:04:50,839 --> 00:04:53,240
Speaker 1: is going to compete with Salesforce, you just don't know

96
00:04:53,279 --> 00:04:56,320
Speaker 1: what you're talking about. You are wrong. But then one

97
00:04:56,360 --> 00:04:59,800
Speaker 1: abstraction higher is this idea that claud code is going

98
00:04:59,800 --> 00:05:02,800
Speaker 1: to destroy SaaS software as a service, and the idea

99
00:05:02,880 --> 00:05:05,040
Speaker 1: being that software as a service is this thing that

100
00:05:05,560 --> 00:05:08,080
Speaker 1: people are just going to stop building their own CRMs.

101
00:05:08,120 --> 00:05:11,240
Speaker 1: They're going to stop building their own per seat software things,

102
00:05:11,279 --> 00:05:13,720
Speaker 1: but they're going to build it internally. This just I

103
00:05:13,720 --> 00:05:15,640
Speaker 1: don't know if you've seen this, cow It just reflects

104
00:05:15,640 --> 00:05:19,520
Speaker 1: a complete lack of understanding of how software works. Yeah,

105
00:05:19,520 --> 00:05:24,000
Speaker 1: because you don't pay Microsoft for Microsoft three sixty five

106
00:05:24,040 --> 00:05:27,320
Speaker 1: teams because you can't build your own word or what

107
00:05:27,320 --> 00:05:29,080
Speaker 1: have you. I don't think you could. But nevertheless, it's

108
00:05:29,120 --> 00:05:31,320
Speaker 1: also because they maintain it, because they make sure it

109
00:05:31,320 --> 00:05:33,520
Speaker 1: doesn't break, or if it breaks, they fix the bit.

110
00:05:33,600 --> 00:05:36,320
Speaker 1: So they make sure that it stays up all the time.

111
00:05:37,080 --> 00:05:39,360
Speaker 1: They make sure it's accessible that has secure log in.

112
00:05:40,680 --> 00:05:43,320
Speaker 2: Well, look, there's a few things going on here. One

113
00:05:43,360 --> 00:05:45,800
Speaker 2: of these things I reported on last month. I did

114
00:05:45,839 --> 00:05:49,000
Speaker 2: a big New Yorker piece on agents. Right, so this

115
00:05:49,279 --> 00:05:50,960
Speaker 2: is relevant to the cloud code and how people are

116
00:05:50,960 --> 00:05:53,159
Speaker 2: thinking about the current future, and there was this basically,

117
00:05:53,240 --> 00:05:56,839
Speaker 2: here's what seems to have happened. Cloud code and these

118
00:05:56,839 --> 00:06:00,800
Speaker 2: other command line interface agents can do really cool things,

119
00:06:00,839 --> 00:06:03,600
Speaker 2: and I'm using the work cool here in a very

120
00:06:03,640 --> 00:06:05,160
Speaker 2: careful fine, yeah cool.

121
00:06:05,520 --> 00:06:07,719
Speaker 1: I would like you to be like completely like give

122
00:06:07,720 --> 00:06:09,160
Speaker 1: people the actual explanation here.

123
00:06:09,320 --> 00:06:12,400
Speaker 2: Yeah, so cool, meaning like Oculus. Right, you put on

124
00:06:12,520 --> 00:06:15,800
Speaker 2: the the Oculus visors for the first time. Everyone had

125
00:06:15,800 --> 00:06:19,520
Speaker 2: the same reaction. This is really cool. Now that's separate

126
00:06:19,560 --> 00:06:22,279
Speaker 2: from that's a trillion dollar business, Like, let's put a side,

127
00:06:22,279 --> 00:06:25,520
Speaker 2: but this is really cool. I'm seeing three D in

128
00:06:25,920 --> 00:06:28,200
Speaker 2: a world where it tracks my head. Cloud code and

129
00:06:28,240 --> 00:06:32,040
Speaker 2: other command line interface coding tools became like that. For programmers,

130
00:06:32,080 --> 00:06:35,159
Speaker 2: it was like really fun the watch it doing multi

131
00:06:35,320 --> 00:06:38,400
Speaker 2: step execution of the construction of demos or this or that,

132
00:06:38,800 --> 00:06:40,760
Speaker 2: and the reason why it could do those cool demos

133
00:06:40,800 --> 00:06:42,560
Speaker 2: it was sort of well suited for that world because

134
00:06:42,560 --> 00:06:45,080
Speaker 2: that's a world that exists only in text. So cloud

135
00:06:45,080 --> 00:06:47,480
Speaker 2: code works on a command line interface. It's all text based,

136
00:06:47,960 --> 00:06:50,680
Speaker 2: and it works with a filesystem. You can write files,

137
00:06:50,839 --> 00:06:54,159
Speaker 2: edit files, sind files to compilers. It's all exists in

138
00:06:54,160 --> 00:06:56,360
Speaker 2: a small number of commands on a command line interface.

139
00:06:56,400 --> 00:06:57,880
Speaker 2: It's all in a world of text, because that's where

140
00:06:57,880 --> 00:07:01,160
Speaker 2: computer programs are built. That's a perfect case for lms,

141
00:07:01,240 --> 00:07:03,200
Speaker 2: which love dealing with text, and they love dealing with

142
00:07:03,240 --> 00:07:07,359
Speaker 2: structure text like computer code. There was an extrapolation that

143
00:07:07,400 --> 00:07:12,200
Speaker 2: then happened, and really this caught on January of twenty five,

144
00:07:12,720 --> 00:07:15,680
Speaker 2: which is where you first began to get this sentiment

145
00:07:15,840 --> 00:07:19,320
Speaker 2: of oh, it's doing such cool things over in the

146
00:07:19,320 --> 00:07:23,080
Speaker 2: world of command line interfaces in code. Certainly these agents

147
00:07:23,120 --> 00:07:26,480
Speaker 2: can now soon do similar cool things and like all

148
00:07:26,520 --> 00:07:28,600
Speaker 2: different things we do on a computer, and that is

149
00:07:28,680 --> 00:07:32,800
Speaker 2: what laid this foundation of people were so impressed, programmers

150
00:07:32,800 --> 00:07:35,680
Speaker 2: were so impressed by the coolness of what was happening

151
00:07:35,720 --> 00:07:38,400
Speaker 2: with cloud code and the other command line interface tools

152
00:07:38,640 --> 00:07:43,640
Speaker 2: that they extrapolated that vibe over to other computer usage.

153
00:07:43,680 --> 00:07:46,880
Speaker 2: Is the point of that article I reported was, Oh,

154
00:07:46,920 --> 00:07:48,720
Speaker 2: it turns out like everything else we do on computer

155
00:07:48,800 --> 00:07:51,840
Speaker 2: is much harder. It's not six text commands on a

156
00:07:51,880 --> 00:07:54,520
Speaker 2: command line interface creating structure code that you can compile

157
00:07:54,560 --> 00:07:56,640
Speaker 2: and test to see if it compiled or not. It's

158
00:07:56,760 --> 00:08:00,840
Speaker 2: much messier. The interfaces are visual. I don't realize what

159
00:08:00,880 --> 00:08:03,400
Speaker 2: complexity goes into the things we click and select doing

160
00:08:03,400 --> 00:08:05,120
Speaker 2: something as simple as even trying to just book like

161
00:08:05,120 --> 00:08:08,520
Speaker 2: a hotel in a new city or something like this.

162
00:08:08,640 --> 00:08:11,000
Speaker 2: And if you use a language model as to underline

163
00:08:11,040 --> 00:08:13,520
Speaker 2: logic and decision engine of making actual actions in the

164
00:08:13,520 --> 00:08:16,000
Speaker 2: real world, well, the language models make things up and

165
00:08:16,000 --> 00:08:18,600
Speaker 2: get things wrong or a little bit wrong twenty percent

166
00:08:18,600 --> 00:08:19,040
Speaker 2: of the time.

167
00:08:19,360 --> 00:08:23,000
Speaker 1: And a little bit wrong in computer science means breaking everything.

168
00:08:23,080 --> 00:08:25,480
Speaker 2: Well, it means a lot like it's okay in code

169
00:08:25,640 --> 00:08:28,320
Speaker 2: because they say, oh, that didn't compile, let's try again.

170
00:08:28,760 --> 00:08:31,080
Speaker 2: But when you're booking a hotel room, as I sort

171
00:08:31,080 --> 00:08:33,360
Speaker 2: of detailed that article, it means like you ended up

172
00:08:33,360 --> 00:08:35,920
Speaker 2: in the wrong city two years from now, and the

173
00:08:36,000 --> 00:08:38,439
Speaker 2: room costs six thousand dollars, right, and so it just

174
00:08:38,480 --> 00:08:40,839
Speaker 2: didn't work. So twenty twenty five was supposed to be

175
00:08:40,840 --> 00:08:42,520
Speaker 2: the year of the Agents and it just didn't work

176
00:08:42,559 --> 00:08:44,920
Speaker 2: and they don't really know how to fix it. But

177
00:08:44,960 --> 00:08:48,080
Speaker 2: we're still vibes. This is Vibe reporting. Yeah, but cloud

178
00:08:48,160 --> 00:08:51,240
Speaker 2: code is like really exciting, and so why can't we

179
00:08:51,280 --> 00:08:53,240
Speaker 2: do that with everything else in the computer. It's actually

180
00:08:53,240 --> 00:08:54,680
Speaker 2: a much harder problem than they were letting on.

181
00:08:55,360 --> 00:08:57,560
Speaker 1: Well. The thing is I've seen, especially like in twenty

182
00:08:57,600 --> 00:08:59,800
Speaker 1: twenty six, I've seen a lot more cloud code stuff.

183
00:09:00,240 --> 00:09:04,440
Speaker 1: Was there's been a very big consent manufacturing operation going

184
00:09:04,480 --> 00:09:07,760
Speaker 1: on right now. Wall Street Journal, Atlantic, CNBC Didrew Bosa.

185
00:09:07,960 --> 00:09:09,880
Speaker 1: This is a statement from me, not cow did you

186
00:09:09,920 --> 00:09:12,440
Speaker 1: both from CNBC should be fucking ashamed of herself going

187
00:09:12,440 --> 00:09:15,240
Speaker 1: on CNBC every day just going cord Code's got a

188
00:09:15,320 --> 00:09:18,440
Speaker 1: destryal software. She's on Twitter because she was able to

189
00:09:18,640 --> 00:09:23,240
Speaker 1: vibe code a some sort of Monday clone, which is

190
00:09:23,280 --> 00:09:26,600
Speaker 1: just like a it's a project management tool. It's like

191
00:09:26,640 --> 00:09:29,840
Speaker 1: I made some software that worked. This is everything now.

192
00:09:30,080 --> 00:09:31,720
Speaker 1: But that's kind of what you've been talking about with

193
00:09:31,720 --> 00:09:35,360
Speaker 1: this Vibe reporting, where it's like I did a small thing.

194
00:09:36,000 --> 00:09:38,720
Speaker 1: Now all things will be done in this manner, whether

195
00:09:38,760 --> 00:09:42,360
Speaker 1: it's possible, God no, God no. But she she, like

196
00:09:42,440 --> 00:09:44,720
Speaker 1: many reporters, are able to find a lot of people

197
00:09:44,720 --> 00:09:47,480
Speaker 1: who are invested in AI, who will absolutely go on

198
00:09:47,559 --> 00:09:51,200
Speaker 1: TV and say, Yep, it's completely true, that's gonna happen

199
00:09:51,360 --> 00:09:55,800
Speaker 1: one hundred percent. It's just so strange because it makes

200
00:09:55,840 --> 00:09:59,839
Speaker 1: me feel paranoid and kind of conspiratle when you look

201
00:09:59,840 --> 00:10:02,440
Speaker 1: at so the majority of news about AI, and it

202
00:10:02,520 --> 00:10:06,679
Speaker 1: is this viper POI. It's these vast extrapolations from sixteen

203
00:10:06,720 --> 00:10:12,160
Speaker 1: thousand job losses Amazon. They mentioned AI. This plus this

204
00:10:12,280 --> 00:10:16,240
Speaker 1: equals the AI is replacing people. It's just so it

205
00:10:16,280 --> 00:10:19,160
Speaker 1: makes me feel like uncomfortable with the world.

206
00:10:19,240 --> 00:10:22,199
Speaker 2: Yeah, well, can we sit for a moment on that

207
00:10:22,240 --> 00:10:24,080
Speaker 2: Amazon example, because I think it's a great please please,

208
00:10:24,120 --> 00:10:27,520
Speaker 2: it frustrates me that one frustling. All right, So Amazon

209
00:10:27,600 --> 00:10:31,760
Speaker 2: lays off sixteen thousand people, right, all right, it's covered

210
00:10:32,640 --> 00:10:35,559
Speaker 2: in two different ways, so the vibe reporting way it's

211
00:10:35,600 --> 00:10:37,839
Speaker 2: covered I in my newsletter, in my podcast, I looked

212
00:10:37,880 --> 00:10:41,480
Speaker 2: an example from Quartz and it was covered as clearly

213
00:10:41,600 --> 00:10:45,320
Speaker 2: intended to imply Amazon laid off sixteen thousand people because

214
00:10:45,360 --> 00:10:48,600
Speaker 2: of AI. They're being replaced by AI. They put the

215
00:10:48,640 --> 00:10:52,680
Speaker 2: subhead of the article was the CEO of Amazon talking

216
00:10:52,720 --> 00:10:57,319
Speaker 2: about how AI is going to increasingly disrupt the job force,

217
00:10:57,960 --> 00:11:02,080
Speaker 2: and then the article itself, no alternative explanations are given

218
00:11:02,120 --> 00:11:03,920
Speaker 2: for these layoffs. They kind of just give the details

219
00:11:03,920 --> 00:11:05,400
Speaker 2: of like here's how many people are laid off and

220
00:11:05,400 --> 00:11:06,839
Speaker 2: here's where they figured it out, and then they put

221
00:11:06,840 --> 00:11:08,440
Speaker 2: a couple of quotes in there about AI being very

222
00:11:08,480 --> 00:11:11,640
Speaker 2: disruptive and being able to automate jobs. It turns out

223
00:11:11,640 --> 00:11:13,679
Speaker 2: those layoffs had nothing to do with AI, and you

224
00:11:13,720 --> 00:11:16,480
Speaker 2: can find other reporting that focused because it was reporting

225
00:11:16,480 --> 00:11:18,040
Speaker 2: that was for the financial market, so it was trying

226
00:11:18,040 --> 00:11:20,240
Speaker 2: to focus on what the hell is actually happening, and

227
00:11:20,520 --> 00:11:22,560
Speaker 2: the deeper reporting was like, yeah, they laid off a

228
00:11:22,559 --> 00:11:25,040
Speaker 2: bunch of managers because like a lot of tech companies,

229
00:11:25,080 --> 00:11:28,640
Speaker 2: they over hired during the pandemic because cloud computing became

230
00:11:28,720 --> 00:11:30,480
Speaker 2: much in demand during the pandemics, So a lot of

231
00:11:30,520 --> 00:11:32,679
Speaker 2: tech companies over hired during the pandemic and they're all

232
00:11:32,679 --> 00:11:35,760
Speaker 2: shedding those jobs again. And Amazon's pretty ruthless about this, right,

233
00:11:35,920 --> 00:11:38,520
Speaker 2: They're always looking for excuses to fire people, and they said,

234
00:11:38,520 --> 00:11:41,880
Speaker 2: we have too much bureaucratic bloat, there's too many managers.

235
00:11:42,200 --> 00:11:43,920
Speaker 2: We're going to fire a bunch of these managers we

236
00:11:44,000 --> 00:11:46,280
Speaker 2: hired so that we can be more lean. Again, that

237
00:11:46,320 --> 00:11:48,200
Speaker 2: has nothing to do with AI. In fact, this is

238
00:11:48,240 --> 00:11:51,360
Speaker 2: the second or third round of these firings that they've

239
00:11:51,360 --> 00:11:54,880
Speaker 2: planned to do, the first round occurring before chat GPT

240
00:11:55,040 --> 00:11:58,040
Speaker 2: was even released. Like, this has nothing to do with

241
00:11:58,200 --> 00:12:00,640
Speaker 2: jobs being replaced by AI. But then you can vibe

242
00:12:00,640 --> 00:12:04,160
Speaker 2: report it because like, well, technically speaking, Amazon also is

243
00:12:04,200 --> 00:12:09,040
Speaker 2: investing money in AI products, So technically speaking, money saved

244
00:12:09,080 --> 00:12:12,320
Speaker 2: by firing these managers could be respent on AI. So

245
00:12:12,360 --> 00:12:15,000
Speaker 2: you can say with a like a semi straight face,

246
00:12:15,800 --> 00:12:18,560
Speaker 2: they fired people because of AI. But clearly, you know,

247
00:12:18,640 --> 00:12:20,440
Speaker 2: the impression that you're giving to the reader is that

248
00:12:20,440 --> 00:12:22,800
Speaker 2: they were replaced by AI and it had nothing to

249
00:12:22,800 --> 00:12:23,960
Speaker 2: do with it. And I heard, by the way, so

250
00:12:24,080 --> 00:12:25,960
Speaker 2: I wrote about that. I put in that video. I've

251
00:12:25,960 --> 00:12:29,240
Speaker 2: heard on background from multiple Amazon executives since they were like,

252
00:12:29,320 --> 00:12:33,160
Speaker 2: we were completely baffled by this coverage that was implying

253
00:12:33,280 --> 00:12:35,720
Speaker 2: that people were being fired here to do with AI.

254
00:12:35,880 --> 00:12:37,920
Speaker 2: This is just what we do at Amazon. We're ruthless,

255
00:12:37,960 --> 00:12:39,679
Speaker 2: Like if you're not earning your keep, we fire you.

256
00:12:40,040 --> 00:12:42,600
Speaker 2: They were completely baffled by that coverage and like things

257
00:12:42,600 --> 00:12:43,880
Speaker 2: for pointing it out, like what.

258
00:12:43,880 --> 00:12:46,360
Speaker 1: I go to be honest, cal if I got an

259
00:12:46,360 --> 00:12:49,079
Speaker 1: email like that from an Amazon executive, A tont go

260
00:12:49,200 --> 00:12:52,560
Speaker 1: fuck themselves? And I mean this nicely because Andy Jesse

261
00:12:52,840 --> 00:12:56,120
Speaker 1: last year June seventeen, twenty twenty five, the whole thing

262
00:12:56,160 --> 00:12:57,839
Speaker 1: about today and virtually every corner of the company we

263
00:12:58,000 --> 00:12:59,880
Speaker 1: use in generate if I had to make customers lives better.

264
00:13:00,080 --> 00:13:03,319
Speaker 1: I believe Amazon benefits from that obfuscation, and I think

265
00:13:03,320 --> 00:13:06,120
Speaker 1: they deliberately fuel it. Now that may be executives who

266
00:13:06,120 --> 00:13:06,719
Speaker 1: disagree with.

267
00:13:06,640 --> 00:13:08,880
Speaker 2: It, well, these were lower level managers, right, so not executs.

268
00:13:08,880 --> 00:13:11,640
Speaker 2: Shouldn't say those like people who work there that were

269
00:13:11,679 --> 00:13:14,480
Speaker 2: like oh yeah, no, no, no, they're not firing people

270
00:13:14,520 --> 00:13:17,000
Speaker 2: because like AI, they're just being brutal.

271
00:13:16,800 --> 00:13:18,440
Speaker 1: Right, right, they're the people who tell the truth.

272
00:13:18,640 --> 00:13:21,480
Speaker 2: Yeah, exactly exactly, But no, you are right. I think

273
00:13:21,480 --> 00:13:23,640
Speaker 2: for the there has been a lot of vibe reporting.

274
00:13:23,679 --> 00:13:25,400
Speaker 2: Basically the end I've been covering this for the last

275
00:13:25,440 --> 00:13:30,360
Speaker 2: two years, the entire sequence of job reductions that were

276
00:13:30,360 --> 00:13:33,200
Speaker 2: post pandemic corrections, so the entire tech industry over hired

277
00:13:33,240 --> 00:13:36,480
Speaker 2: during the pandemic. The entire tech industry cuts ups in

278
00:13:36,480 --> 00:13:38,480
Speaker 2: the last few years because they hire too many people

279
00:13:38,559 --> 00:13:41,520
Speaker 2: and now they have to correct back to where they

280
00:13:41,559 --> 00:13:44,840
Speaker 2: were pre pandemic. Consistently, across the board, these cuts are

281
00:13:44,880 --> 00:13:48,400
Speaker 2: vibe reported is due to AI consistently. You see exactly

282
00:13:48,400 --> 00:13:50,640
Speaker 2: that story. We see it, and I'm a computer scientist.

283
00:13:50,679 --> 00:13:54,120
Speaker 2: We see this in coverage of computer science majors as well.

284
00:13:54,160 --> 00:13:58,000
Speaker 2: Same idea. Computer science majors historically directly tied to the

285
00:13:58,000 --> 00:14:01,000
Speaker 2: tech industry. If they're cutting in a downside, majors go down.

286
00:14:01,360 --> 00:14:03,280
Speaker 2: If they're hiring. I mean, it's just economic. It's not

287
00:14:03,720 --> 00:14:06,719
Speaker 2: surprising when the technistry is booming, we get a lot

288
00:14:06,720 --> 00:14:09,320
Speaker 2: more majors because they're good jobs, right, And so computer

289
00:14:09,360 --> 00:14:11,560
Speaker 2: science majors went down in the last year or two

290
00:14:11,960 --> 00:14:15,000
Speaker 2: as the tech company started cutting. That was reported in

291
00:14:15,000 --> 00:14:19,480
Speaker 2: the Atlantic as kids are not majoring in AI because

292
00:14:19,520 --> 00:14:21,360
Speaker 2: AI makes the the.

293
00:14:21,400 --> 00:14:22,680
Speaker 1: Surprise majoring and kompsa.

294
00:14:22,760 --> 00:14:25,120
Speaker 2: You mean yeah, because AI is going to do all

295
00:14:25,120 --> 00:14:28,840
Speaker 2: these jobs. Because we've had these cycles every five years,

296
00:14:28,840 --> 00:14:32,000
Speaker 2: we have this cycle there, this is nothing anyways.

297
00:14:31,880 --> 00:14:34,600
Speaker 1: No, no, I'm and the Atlantic has just done a

298
00:14:34,640 --> 00:14:37,520
Speaker 1: piss poor job with this stuff. I'm I'm heighten and

299
00:14:37,840 --> 00:14:41,560
Speaker 1: I'm heighten on them because they had an awful claud

300
00:14:41,600 --> 00:14:43,640
Speaker 1: code thing. They just you know, I'm gonna bring it

301
00:14:43,680 --> 00:14:45,320
Speaker 1: up and read the title. I'm not gonna say the

302
00:14:45,360 --> 00:14:48,760
Speaker 1: reporter's name because I think that that's that's main. But

303
00:14:48,840 --> 00:14:51,720
Speaker 1: let's look at this move over chat GPT. You're about

304
00:14:51,760 --> 00:14:53,680
Speaker 1: to hear a lot more about Claude coding. Know why

305
00:14:53,680 --> 00:14:55,960
Speaker 1: you're gonna hear about that? The fucking Atlantic is writing

306
00:14:56,000 --> 00:14:59,520
Speaker 1: about it, and it's just over the whole days. Alex

307
00:14:59,560 --> 00:15:02,160
Speaker 1: lieberm And had an idea what if he could create

308
00:15:02,200 --> 00:15:04,840
Speaker 1: Spotify rapped for his text messages without writing a single

309
00:15:04,880 --> 00:15:07,360
Speaker 1: line of code. Liberman, co founder of the media outlet

310
00:15:07,360 --> 00:15:09,920
Speaker 1: morning Brew, created I message wrapped. I just want to

311
00:15:09,960 --> 00:15:12,320
Speaker 1: start here and say that guy doesn't do any fucking work,

312
00:15:12,360 --> 00:15:14,720
Speaker 1: and I know people who work there like I just

313
00:15:14,720 --> 00:15:17,400
Speaker 1: want to start by saying, if that's the best you've got,

314
00:15:17,880 --> 00:15:20,440
Speaker 1: and that probably involves throwing a giraffe and the entire

315
00:15:20,480 --> 00:15:23,880
Speaker 1: zoo into a vat of acid to make the GPUs move.

316
00:15:25,360 --> 00:15:28,200
Speaker 1: You're meant to read this, and the deliberate effect you're

317
00:15:28,200 --> 00:15:31,120
Speaker 1: meant to have is you're meant to be scared and

318
00:15:31,200 --> 00:15:33,880
Speaker 1: excited like it is a kind of a mishmash between

319
00:15:33,920 --> 00:15:38,040
Speaker 1: the vibe reporting and the I guess it's the folksting.

320
00:15:38,120 --> 00:15:40,960
Speaker 1: In fact, this might be a triple score because this

321
00:15:41,000 --> 00:15:43,479
Speaker 1: is meant to feel you, make you feel with discomfort,

322
00:15:43,560 --> 00:15:46,240
Speaker 1: but also make you freak up but also be excited,

323
00:15:46,240 --> 00:15:49,880
Speaker 1: a rare triple score. It's just stories like these pissed

324
00:15:49,920 --> 00:15:52,440
Speaker 1: me off because I'm fine with people going this could

325
00:15:52,520 --> 00:15:56,840
Speaker 1: do this. I don't mind it happens, But when it's

326
00:15:56,880 --> 00:16:00,600
Speaker 1: just like, hmm, feed me the slop right into my mouth, asshole,

327
00:16:00,760 --> 00:16:03,720
Speaker 1: make me feel, make me feel all the marketing things

328
00:16:03,760 --> 00:16:08,240
Speaker 1: all at once. First of all, I'm not impressed by

329
00:16:08,240 --> 00:16:11,000
Speaker 1: the idea of doing I message s grapped. That's not

330
00:16:11,320 --> 00:16:13,520
Speaker 1: That's not a thing a human being with that with

331
00:16:13,640 --> 00:16:18,760
Speaker 1: like friends and hobbies does. I've been I've been watching

332
00:16:18,760 --> 00:16:21,240
Speaker 1: the first season The True Detective. I've got many more

333
00:16:21,280 --> 00:16:24,280
Speaker 1: shows I could watch. Never once would I think what

334
00:16:24,360 --> 00:16:26,600
Speaker 1: if I could get it wrapped of all my messages?

335
00:16:26,720 --> 00:16:29,880
Speaker 1: What a fucking psychotic thing to do. But when you

336
00:16:29,920 --> 00:16:33,520
Speaker 1: read this, it's like yours. You spent your holidays with

337
00:16:33,560 --> 00:16:36,560
Speaker 1: your family, wrote one tech Polticy expert. That's nice. I

338
00:16:36,600 --> 00:16:38,400
Speaker 1: spent my holidays with Claude code.

339
00:16:38,800 --> 00:16:39,520
Speaker 2: Well, it's fun.

340
00:16:40,760 --> 00:16:41,280
Speaker 1: That's cool.

341
00:16:41,640 --> 00:16:45,320
Speaker 2: It's fun to create, uh these sort of demo apps

342
00:16:45,440 --> 00:16:48,120
Speaker 2: if you're like engineering minded, So in that sense, they're

343
00:16:48,120 --> 00:16:51,840
Speaker 2: like most Yeah, the most cynical analysis here is that

344
00:16:52,640 --> 00:16:56,960
Speaker 2: cloud code is like model trains for engineering minded people.

345
00:16:57,040 --> 00:16:59,040
Speaker 2: It's a fun hobby I can put. Look at this

346
00:16:59,200 --> 00:17:02,120
Speaker 2: like I made a thing that can Like I was

347
00:17:02,120 --> 00:17:04,600
Speaker 2: talking to a friend of mine yesterday, computer scientist, and

348
00:17:04,640 --> 00:17:06,360
Speaker 2: you're like, oh, I built a thing where I can,

349
00:17:06,440 --> 00:17:10,639
Speaker 2: you know whatever, email an appointment to a thing that

350
00:17:11,000 --> 00:17:13,840
Speaker 2: gets parsed by an LM and then goes to my calendar.

351
00:17:14,600 --> 00:17:16,520
Speaker 2: That's just fun for him, in the same way that

352
00:17:16,560 --> 00:17:18,600
Speaker 2: someone else might be like, hey, I built a cabinet

353
00:17:18,800 --> 00:17:20,560
Speaker 2: and like it's really nice, Like look, I got the

354
00:17:20,560 --> 00:17:22,440
Speaker 2: wood to go together, and like I'm proud. I could

355
00:17:22,480 --> 00:17:23,760
Speaker 2: just on a cabinet or whatever.

356
00:17:23,560 --> 00:17:27,800
Speaker 1: Except you build, Except you built something with your hands,

357
00:17:28,280 --> 00:17:30,560
Speaker 1: and a cabinet can have things put in it. It's

358
00:17:30,880 --> 00:17:35,280
Speaker 1: I just it feels like Lego almost, Like it's more

359
00:17:35,320 --> 00:17:39,639
Speaker 1: like it's toy software that has some functionality. Because the

360
00:17:39,680 --> 00:17:42,360
Speaker 1: big thing I'm waiting for with the vibe coding stuff

361
00:17:42,440 --> 00:17:45,920
Speaker 1: is an actual product, you know what I mean.

362
00:17:46,200 --> 00:17:48,400
Speaker 2: But you know that doesn't go well. I mean, look,

363
00:17:48,440 --> 00:17:50,840
Speaker 2: this is the story of This is the story of maltbook,

364
00:17:51,280 --> 00:17:52,040
Speaker 2: right because.

365
00:17:52,000 --> 00:17:54,280
Speaker 1: Oh yeah, tell me more about moltbook. We were just

366
00:17:54,320 --> 00:17:54,800
Speaker 1: talked about this.

367
00:17:55,560 --> 00:17:57,000
Speaker 2: The whole can of worms to open there, I'm just

368
00:17:57,040 --> 00:17:58,760
Speaker 2: going to open like the top of the nearest can,

369
00:17:58,920 --> 00:18:00,800
Speaker 2: which is before even getting the what molt book is

370
00:18:00,840 --> 00:18:02,960
Speaker 2: and is not. You know, it was in the news

371
00:18:02,960 --> 00:18:06,960
Speaker 2: all the time. It was vibe coded and immediately was

372
00:18:07,040 --> 00:18:10,480
Speaker 2: just full of terrible security holes because it was vibe coded.

373
00:18:10,520 --> 00:18:13,040
Speaker 2: And it turned out that you could get the API key,

374
00:18:13,160 --> 00:18:16,760
Speaker 2: so the key you use when you access the paid

375
00:18:16,800 --> 00:18:19,520
Speaker 2: service to get used in LLLM, you have an API key,

376
00:18:19,520 --> 00:18:21,120
Speaker 2: so they know who to charge. You could just steal

377
00:18:21,160 --> 00:18:23,160
Speaker 2: everyone's API key who was using it because the guy

378
00:18:23,280 --> 00:18:25,720
Speaker 2: just vibe coded it and so no one was actually

379
00:18:25,720 --> 00:18:28,439
Speaker 2: there looking at the code. But what but what I

380
00:18:28,440 --> 00:18:30,200
Speaker 2: think is going on? Okay, So here's what I would

381
00:18:30,200 --> 00:18:31,920
Speaker 2: like to see. I would like to see more reporting.

382
00:18:32,080 --> 00:18:35,119
Speaker 2: That would be how are people using X Like to me,

383
00:18:35,160 --> 00:18:36,880
Speaker 2: that's very interesting, yeah, right, how are.

384
00:18:36,760 --> 00:18:37,720
Speaker 1: People using agree?

385
00:18:37,960 --> 00:18:41,000
Speaker 2: Yeah? And the problem is those answers right now. And

386
00:18:41,040 --> 00:18:43,159
Speaker 2: this is confounding. I think that people who are very

387
00:18:43,160 --> 00:18:46,960
Speaker 2: excited by the potential and coolness of these things in isolation,

388
00:18:47,480 --> 00:18:49,760
Speaker 2: the answer to how are people using X is often

389
00:18:49,800 --> 00:18:52,760
Speaker 2: not nearly as exciting as you would guess. And I

390
00:18:52,760 --> 00:18:54,879
Speaker 2: think the reality with I mean, I go quite have

391
00:18:54,920 --> 00:18:56,600
Speaker 2: my arms around cloud code. I do know there's a

392
00:18:56,600 --> 00:18:58,600
Speaker 2: lot of people who are building kind of like internal

393
00:18:58,640 --> 00:19:01,600
Speaker 2: tools or personal tools with it, which I think is cool.

394
00:19:02,280 --> 00:19:04,600
Speaker 2: Most people aren't interested in that, but for some people

395
00:19:04,640 --> 00:19:07,040
Speaker 2: they are. And you know, I think it's fun. Computer

396
00:19:07,080 --> 00:19:09,360
Speaker 2: programming is fun, right, so like the ability to make

397
00:19:09,359 --> 00:19:11,840
Speaker 2: a program work is and some of those tools are useful.

398
00:19:11,840 --> 00:19:13,800
Speaker 2: But this is that's not a major industry on its own.

399
00:19:13,840 --> 00:19:16,160
Speaker 2: You know, I'm designing a tool for my small company

400
00:19:16,200 --> 00:19:19,000
Speaker 2: that makes it easier for us to do whatever. That's cool,

401
00:19:19,040 --> 00:19:21,880
Speaker 2: but that's not like a trillion dollar industry. I can't

402
00:19:21,880 --> 00:19:25,119
Speaker 2: get my arms around yet exactly how professional computer programmers

403
00:19:25,119 --> 00:19:27,480
Speaker 2: are using it. They're all talking about it, really yeah,

404
00:19:27,800 --> 00:19:30,240
Speaker 2: but I can't get I can't get my arms around

405
00:19:30,240 --> 00:19:32,760
Speaker 2: it yet. Yeah, what do you what's what's your sense of.

406
00:19:34,320 --> 00:19:36,840
Speaker 1: I'm gonna be honest, I was going to ask you.

407
00:19:37,160 --> 00:19:39,520
Speaker 1: I was literally going to ask have you heard people

408
00:19:39,640 --> 00:19:42,480
Speaker 1: using it? Because if you go on X, the Everything

409
00:19:42,480 --> 00:19:47,520
Speaker 1: app if you put on pulhasmat suit, you go on

410
00:19:47,800 --> 00:19:49,159
Speaker 1: X and you go and look. And the way that

411
00:19:49,200 --> 00:19:53,320
Speaker 1: people talk about this is like they have connected into

412
00:19:53,359 --> 00:19:57,040
Speaker 1: the matrix that they are now a thousand X engineer.

413
00:19:57,240 --> 00:19:58,560
Speaker 1: But when you go and look at what they do,

414
00:19:58,600 --> 00:20:01,840
Speaker 1: no one actually says. There's always these these kind of

415
00:20:01,920 --> 00:20:04,800
Speaker 1: vibe store, the vibe tales, the mythology where it's like

416
00:20:05,280 --> 00:20:08,919
Speaker 1: I had a problem vaguely that would have taken me

417
00:20:09,480 --> 00:20:11,960
Speaker 1: X number of hours, but when I used claud code,

418
00:20:12,040 --> 00:20:14,439
Speaker 1: it solved it immediately and caught two bugs that I

419
00:20:14,520 --> 00:20:18,520
Speaker 1: didn't know about. It's very Marine Todd's style. Then everybody

420
00:20:18,600 --> 00:20:23,480
Speaker 1: clapped yeah, and it's you were a computer science teacher

421
00:20:23,720 --> 00:20:26,400
Speaker 1: and you you don't know either, which makes me think

422
00:20:26,440 --> 00:20:29,639
Speaker 1: it's not as big a deal as people are saying.

423
00:20:29,920 --> 00:20:32,919
Speaker 1: My other bit of evidence is that at the end

424
00:20:32,960 --> 00:20:36,160
Speaker 1: of last year, Anthropics claud Code revenue was one point

425
00:20:36,200 --> 00:20:40,200
Speaker 1: one billion annualized, so about ninety to one hundred million

426
00:20:40,240 --> 00:20:46,120
Speaker 1: a month for a revolution. That feels low somehow.

427
00:20:57,480 --> 00:20:59,240
Speaker 2: Yeah, I mean what I know as a computer science

428
00:20:59,320 --> 00:21:03,400
Speaker 2: you can't rate performance oriented code. You can't write safe code.

429
00:21:04,480 --> 00:21:07,960
Speaker 2: You can't write code it has to sort of juggle

430
00:21:08,080 --> 00:21:10,359
Speaker 2: sort of complex out of scenarios. I mean, you just

431
00:21:10,440 --> 00:21:13,560
Speaker 2: need good programmers' eyes on it building this code.

432
00:21:13,600 --> 00:21:15,600
Speaker 1: Why is that? Is there a way of explaining to

433
00:21:15,680 --> 00:21:17,000
Speaker 1: a known code of why.

434
00:21:16,840 --> 00:21:20,520
Speaker 2: That is code? There's like a poetic element to it,

435
00:21:20,960 --> 00:21:24,480
Speaker 2: you know, it's Writing good computer code is difficult. You're

436
00:21:24,480 --> 00:21:28,200
Speaker 2: often you're dealing with, you know, what is my problem here?

437
00:21:28,240 --> 00:21:30,320
Speaker 2: And I want an elegant way of sort of satisfying

438
00:21:30,320 --> 00:21:34,680
Speaker 2: this problem. You're often drawing from pretty nuanced algorithms and

439
00:21:34,760 --> 00:21:37,159
Speaker 2: data structures to try to figure out how am I

440
00:21:37,200 --> 00:21:41,280
Speaker 2: going to organize information and efficiently access it? When performance

441
00:21:41,320 --> 00:21:43,360
Speaker 2: comes into play, there's a lot of really subtle decisions

442
00:21:43,400 --> 00:21:45,280
Speaker 2: to make about, you know, how am I going to

443
00:21:45,320 --> 00:21:47,040
Speaker 2: store use things in such a way that we don't

444
00:21:47,040 --> 00:21:49,840
Speaker 2: get bogged down when we're trying to execute things. It

445
00:21:49,920 --> 00:21:52,280
Speaker 2: just becomes I'm I don't code as much anymore. Exo'm

446
00:21:52,280 --> 00:21:54,000
Speaker 2: a theoretician, but I did my whole life since I

447
00:21:54,119 --> 00:21:57,160
Speaker 2: was you know, seven, And it's a there's a it's

448
00:21:57,160 --> 00:22:00,160
Speaker 2: an art form. Right When you're using cloud code, you're

449
00:22:00,160 --> 00:22:01,720
Speaker 2: not really supposed to look at the code, and so

450
00:22:01,760 --> 00:22:03,600
Speaker 2: I think that takes a lot of uses probably off

451
00:22:03,640 --> 00:22:05,800
Speaker 2: the table. It's like the use cases. You're supposed to

452
00:22:05,840 --> 00:22:08,320
Speaker 2: have these different instances of cloud code running, and this

453
00:22:08,400 --> 00:22:10,239
Speaker 2: one's going to write code, and then this one's going

454
00:22:10,280 --> 00:22:12,720
Speaker 2: to write test for that code, and then the cloud

455
00:22:12,760 --> 00:22:14,440
Speaker 2: code is going to run the test and then try

456
00:22:14,440 --> 00:22:16,480
Speaker 2: to fix the code if it doesn't matches the test.

457
00:22:16,960 --> 00:22:19,480
Speaker 2: But your eyes aren't on the code at all. And

458
00:22:19,520 --> 00:22:21,440
Speaker 2: there's I mean, obviously for a lot of programming, that's

459
00:22:21,440 --> 00:22:23,040
Speaker 2: an issue. But then the other thing I've heard about

460
00:22:23,040 --> 00:22:26,080
Speaker 2: the computer programming industry is it's very stratified. There's a

461
00:22:26,080 --> 00:22:29,040
Speaker 2: smaller number of like really good, serious programmers that produce

462
00:22:29,240 --> 00:22:33,280
Speaker 2: like ninety percent of the really important, valuable code on

463
00:22:33,320 --> 00:22:35,960
Speaker 2: which everything runs. I don't think they would touch cloud

464
00:22:36,000 --> 00:22:38,159
Speaker 2: code with a ten foot pole, like they're good at

465
00:22:38,160 --> 00:22:40,120
Speaker 2: what they do. And then you have these like huge

466
00:22:40,160 --> 00:22:44,840
Speaker 2: strata of people writing like JavaScript and sort of hacking

467
00:22:44,840 --> 00:22:47,719
Speaker 2: together Python, and you know it's like not very good code.

468
00:22:47,800 --> 00:22:49,760
Speaker 2: And then it's I guess you could replace some of

469
00:22:49,760 --> 00:22:52,399
Speaker 2: the functional it's functional enough, but I can Yeah, I

470
00:22:52,400 --> 00:22:53,840
Speaker 2: can't get my arms around it yet. But I do

471
00:22:53,840 --> 00:22:55,040
Speaker 2: get you know, I have a lot of sources, so

472
00:22:55,160 --> 00:22:57,240
Speaker 2: I hear from you know, I do hear from people

473
00:22:57,320 --> 00:23:00,520
Speaker 2: that are talking about how cool cloud code is. Is cool.

474
00:23:00,640 --> 00:23:02,720
Speaker 2: I hear from a lot of professional programmers that are like, yeah,

475
00:23:02,720 --> 00:23:04,840
Speaker 2: we don't use we can't use this. We're trying to

476
00:23:04,840 --> 00:23:07,159
Speaker 2: write serious programs, like we have to sell this software,

477
00:23:07,280 --> 00:23:09,879
Speaker 2: like this is this is not solving a problem we

478
00:23:09,960 --> 00:23:13,240
Speaker 2: have or you know we so you know, I don't know,

479
00:23:13,280 --> 00:23:15,400
Speaker 2: But the problem is that that's what the reporting should be. Hey,

480
00:23:15,480 --> 00:23:19,400
Speaker 2: here we are at this company looking over the shoulder

481
00:23:19,440 --> 00:23:22,159
Speaker 2: of people. What are they doing. Let's talk to the

482
00:23:22,200 --> 00:23:25,480
Speaker 2: engineering teams on background of this tech company exactly what

483
00:23:25,680 --> 00:23:26,960
Speaker 2: role is going on here? And I think what a

484
00:23:27,000 --> 00:23:30,480
Speaker 2: lot of reporting has become on AI is your hype laundering.

485
00:23:30,640 --> 00:23:35,000
Speaker 2: So you look at the discussion about the technology happening

486
00:23:35,000 --> 00:23:37,680
Speaker 2: from more engineering minded people, you convince yourself as a reporter,

487
00:23:37,800 --> 00:23:39,919
Speaker 2: I can't understand the engineering, but I will trust the

488
00:23:39,920 --> 00:23:42,639
Speaker 2: people who do, and then you launder what you're sensing

489
00:23:42,680 --> 00:23:46,119
Speaker 2: from that hype into your articles, not realizing that like

490
00:23:46,240 --> 00:23:48,960
Speaker 2: nerds like me, we get hyped up about stuff, and

491
00:23:49,000 --> 00:23:51,520
Speaker 2: we get super excited about stuff and we go create like, yeah,

492
00:23:51,560 --> 00:23:54,960
Speaker 2: you can't just launder our hype into this is what's happening.

493
00:23:55,320 --> 00:23:58,199
Speaker 2: And so it's like reporting on a war where you

494
00:23:58,280 --> 00:24:01,000
Speaker 2: have no one embedded, there's no one actually on the

495
00:24:01,080 --> 00:24:05,080
Speaker 2: ground where the battles are happening. You're just responding to

496
00:24:05,119 --> 00:24:07,800
Speaker 2: the press conferences that the generals are holding, you know,

497
00:24:08,040 --> 00:24:10,000
Speaker 2: back in the Pentagon. Like it's not a way to

498
00:24:10,000 --> 00:24:11,280
Speaker 2: report on what's actually going on.

499
00:24:12,200 --> 00:24:14,280
Speaker 1: And I think the other thing as well is there

500
00:24:14,359 --> 00:24:17,560
Speaker 1: is a if you don't do this hype laundering. I

501
00:24:18,480 --> 00:24:20,480
Speaker 1: don't know how these If you're a report listening to

502
00:24:20,520 --> 00:24:22,040
Speaker 1: this and you have a thought about this, send it

503
00:24:22,040 --> 00:24:25,440
Speaker 1: to me anonymously. Is it trod Thatt seventy six on signal.

504
00:24:25,840 --> 00:24:28,240
Speaker 1: But my thought is as well, is there's probably a

505
00:24:28,280 --> 00:24:30,879
Speaker 1: problem with rationalization as well, because if you look at

506
00:24:30,920 --> 00:24:34,879
Speaker 1: this and you say, Okay, well it's kind of cool,

507
00:24:34,920 --> 00:24:38,320
Speaker 1: it's fun in whatever indeterminate way, it doesn't seem like

508
00:24:38,840 --> 00:24:43,320
Speaker 1: serious software, like actual real deal software is being made

509
00:24:43,320 --> 00:24:45,760
Speaker 1: with it. But then the CEO of Google says thirty

510
00:24:45,760 --> 00:24:48,760
Speaker 1: percent of code is written in AI, which is bullshit.

511
00:24:49,119 --> 00:24:52,480
Speaker 1: And I've heard from so many software engineers, well, it

512
00:24:52,520 --> 00:24:55,679
Speaker 1: couldn't be that everyone's just wrong, right, It couldn't be

513
00:24:55,680 --> 00:24:58,600
Speaker 1: a case of that everyone is making the most egregious

514
00:24:58,640 --> 00:25:02,080
Speaker 1: capital expenditure fuck up of all time. This will be historic,

515
00:25:02,400 --> 00:25:06,640
Speaker 1: I believe, worse than railways, digital beanie babies, but done

516
00:25:06,680 --> 00:25:09,560
Speaker 1: at the scale of laser tag arenas. Now, that can't

517
00:25:09,560 --> 00:25:11,919
Speaker 1: possibly be that, because everyone else is saying this is

518
00:25:11,960 --> 00:25:16,040
Speaker 1: exciting and good. And at that point they choose instead

519
00:25:16,040 --> 00:25:20,160
Speaker 1: of being like worried, instead of being a bit anxious

520
00:25:20,200 --> 00:25:23,359
Speaker 1: about this, they say, well, Amazon Web Services spent a

521
00:25:23,400 --> 00:25:26,080
Speaker 1: lot of money, so this spends a lot of money

522
00:25:26,240 --> 00:25:31,400
Speaker 1: money too, so this is actually good. It's actually good.

523
00:25:31,400 --> 00:25:34,040
Speaker 1: And indeed these people seem emphatic and excited, and I,

524
00:25:34,160 --> 00:25:38,199
Speaker 1: as a non coder, can build a fudged CRM that

525
00:25:38,359 --> 00:25:42,480
Speaker 1: probably would not withstand even the laziest haacker. I can

526
00:25:42,520 --> 00:25:44,960
Speaker 1: do this, and thus it will extrapolate further from there.

527
00:25:45,400 --> 00:25:47,760
Speaker 1: And what sucks is what the people that I believe

528
00:25:47,800 --> 00:25:52,000
Speaker 1: actually will be hurt by this are retail investors. I think,

529
00:25:52,119 --> 00:25:55,240
Speaker 1: regular people buying stocks in these companies or selling software

530
00:25:55,240 --> 00:25:58,520
Speaker 1: stocks because they believe that color code will replace them.

531
00:25:58,640 --> 00:26:01,120
Speaker 1: And altomately, I think it's just gonna be a blood bath.

532
00:26:01,480 --> 00:26:03,720
Speaker 1: For people's furrow. One case that could have been avoided,

533
00:26:03,840 --> 00:26:07,960
Speaker 1: except it would require reports to do something uncomfortable, and

534
00:26:08,000 --> 00:26:12,080
Speaker 1: I don't think they want to do that. Ema.

535
00:26:12,760 --> 00:26:14,480
Speaker 2: I think there's two things going on. This is what

536
00:26:14,480 --> 00:26:18,400
Speaker 2: I've decided with reporting. First of all, it's an asymmetric risk.

537
00:26:18,960 --> 00:26:20,600
Speaker 2: So a lot of reporters are like, look, there's not

538
00:26:20,640 --> 00:26:24,040
Speaker 2: a major risk if I'm excited about this and it

539
00:26:24,080 --> 00:26:27,080
Speaker 2: doesn't pan out, because we could be like, yeah, surprisingly

540
00:26:27,160 --> 00:26:29,400
Speaker 2: this didn't pan out and with some factors we couldn't see.

541
00:26:29,400 --> 00:26:30,840
Speaker 2: But they do feel like there's a huge amount of

542
00:26:30,920 --> 00:26:33,359
Speaker 2: risk of saying this is not a big deal and

543
00:26:33,359 --> 00:26:36,040
Speaker 2: then it is, and so it's definitely an asymmetric risk.

544
00:26:36,080 --> 00:26:38,000
Speaker 2: We saw a lot of this during like COVID as well.

545
00:26:38,680 --> 00:26:41,520
Speaker 2: Right is like you wanted it was less There would

546
00:26:41,520 --> 00:26:45,479
Speaker 2: be less harm if you were too alarmist about something,

547
00:26:45,560 --> 00:26:47,480
Speaker 2: but there could be a lot of harm. They felt

548
00:26:47,560 --> 00:26:49,680
Speaker 2: like reputationally, I feel like this is not a big deal,

549
00:26:49,680 --> 00:26:52,440
Speaker 2: and it was, and so there's definitely an asymmetric risk profile.

550
00:26:52,680 --> 00:26:56,040
Speaker 2: There's also like a meaning defining profile. It's just really

551
00:26:56,080 --> 00:26:59,480
Speaker 2: exciting to think everything is going to be disrupted in

552
00:26:59,600 --> 00:27:04,639
Speaker 2: chain unrecognizably. It sort of gives a focus and meaning

553
00:27:04,760 --> 00:27:09,439
Speaker 2: to like an otherwise somewhat chaotic and disrupted world that

554
00:27:09,480 --> 00:27:12,560
Speaker 2: we're in right now. And so there's that aspect too,

555
00:27:12,800 --> 00:27:17,080
Speaker 2: is that people want to believe there is something massive

556
00:27:17,119 --> 00:27:19,920
Speaker 2: about to happen because in some sense it wipes away

557
00:27:19,960 --> 00:27:22,000
Speaker 2: all the bad stuff that's happening. Who cares, none of

558
00:27:22,000 --> 00:27:25,240
Speaker 2: this is relevant because this much bigger thing is coming.

559
00:27:25,280 --> 00:27:27,119
Speaker 2: So I think that gets wrapped into it as well.

560
00:27:27,160 --> 00:27:29,800
Speaker 2: I think the economic reporters are more on this because

561
00:27:29,840 --> 00:27:32,080
Speaker 2: like their whole job is to try to I mean,

562
00:27:32,280 --> 00:27:34,320
Speaker 2: they're not on it. I think after your work they're

563
00:27:34,320 --> 00:27:34,880
Speaker 2: on it more.

564
00:27:34,920 --> 00:27:37,840
Speaker 1: But like, oh no, if I agree, I am reading

565
00:27:37,920 --> 00:27:41,240
Speaker 1: Big Series. There's an article in Bloomberg I'm trying to

566
00:27:41,320 --> 00:27:45,320
Speaker 1: shove through archive dot is because they were so main

567
00:27:45,400 --> 00:27:47,680
Speaker 1: and unfit to my friend Steve Burke at Games Next

568
00:27:47,680 --> 00:27:50,679
Speaker 1: as that won't pay them. But it's more shit about

569
00:27:50,720 --> 00:27:54,600
Speaker 1: like the software narrative, the fact that software is being

570
00:27:54,600 --> 00:27:57,760
Speaker 1: disrupted by Claude code. You get the same pallid reporting

571
00:27:57,960 --> 00:28:01,600
Speaker 1: from Bloomberg and even the Financial Time about Anthropic and

572
00:28:01,640 --> 00:28:03,760
Speaker 1: the fd is generally pretty good. Well, they're like, yeah,

573
00:28:03,800 --> 00:28:07,160
Speaker 1: they're just gonna make thirty billion dollars next year. It's

574
00:28:07,200 --> 00:28:12,399
Speaker 1: just fanciful. It's there's the skepticism, the cynicism doesn't exist.

575
00:28:12,440 --> 00:28:14,359
Speaker 1: And I get I agree on that there is a

576
00:28:14,359 --> 00:28:15,160
Speaker 1: bubble reporting.

577
00:28:15,200 --> 00:28:18,879
Speaker 2: The last fall, there was a period where everyone did

578
00:28:18,920 --> 00:28:21,120
Speaker 2: the bubble reporting. After GPG five, you had a two

579
00:28:21,119 --> 00:28:25,800
Speaker 2: month period where every major publication did bubble reportings. But you're, yeah,

580
00:28:25,800 --> 00:28:27,160
Speaker 2: I guess that did kind of die off.

581
00:28:28,040 --> 00:28:31,880
Speaker 1: It died off because they hear one nice thing from

582
00:28:31,960 --> 00:28:35,320
Speaker 1: Jensen Wong and they're like, well, I'm sold.

583
00:28:35,640 --> 00:28:38,320
Speaker 2: Like that jacket that jackets looking pretty sharp. I mean,

584
00:28:38,440 --> 00:28:41,200
Speaker 2: someone who wears that jacket. How could they be wrong?

585
00:28:41,520 --> 00:28:44,640
Speaker 1: Would a man that has a shiny jacket like the

586
00:28:44,680 --> 00:28:48,480
Speaker 1: Gleat singer of Corn wears at concerts? Would he lie?

587
00:28:48,680 --> 00:28:52,560
Speaker 1: And it's just what really bothers me. As far as

588
00:28:52,600 --> 00:28:56,240
Speaker 1: the economic reporting though, is the Oracle, because it's like

589
00:28:56,320 --> 00:28:59,040
Speaker 1: Oracle needs to prepay three hundred billion dollars over five

590
00:28:59,120 --> 00:29:02,440
Speaker 1: years by op AI, a company that, if we're to

591
00:29:02,480 --> 00:29:06,440
Speaker 1: believe reporting, which I do not, they made thirteen billion

592
00:29:06,480 --> 00:29:09,880
Speaker 1: dollars last year in revenue and lost. Sorry they've made

593
00:29:09,960 --> 00:29:13,120
Speaker 1: thirteen billion and lost. They claim nine. I think it's

594
00:29:13,120 --> 00:29:15,960
Speaker 1: probably higher. How are they meant to pay thirty to

595
00:29:16,000 --> 00:29:19,719
Speaker 1: sixty billion dollars a year in a year, And everyone's like, well,

596
00:29:20,840 --> 00:29:23,520
Speaker 1: they'll work it out. How's Oracle meant to build those

597
00:29:23,600 --> 00:29:25,680
Speaker 1: data centers? Those data centers are gonna cost one hundred

598
00:29:25,680 --> 00:29:29,680
Speaker 1: and eighty nine billion dollars. They've raised one hundred billion, Okay,

599
00:29:29,760 --> 00:29:31,360
Speaker 1: how they meant to do that? And everyone's just like, oh,

600
00:29:31,360 --> 00:29:34,760
Speaker 1: they'll work it out. If I wish I could do

601
00:29:34,800 --> 00:29:39,000
Speaker 1: this with the fucking bank, I need a five hundred

602
00:29:39,000 --> 00:29:41,480
Speaker 1: million dollar house, I'll pay for it at some point,

603
00:29:41,840 --> 00:29:44,040
Speaker 1: all right, it's fine, and the news would run articles

604
00:29:44,080 --> 00:29:47,960
Speaker 1: about my genius housing purchases. It's just it's one of

605
00:29:48,000 --> 00:29:51,720
Speaker 1: those things where and you said, well the asymmetric risk

606
00:29:52,240 --> 00:29:54,920
Speaker 1: doesn't exist, it does for some of these reporters because

607
00:29:54,920 --> 00:29:57,720
Speaker 1: I've been saving their their bylines for years. Because I

608
00:29:57,760 --> 00:29:59,640
Speaker 1: actually think that there needs to be some sort of

609
00:29:59,640 --> 00:30:02,840
Speaker 1: reckoning with this, because if you look back, there are

610
00:30:02,920 --> 00:30:06,200
Speaker 1: major financial outlets that did the same, that were literally

611
00:30:06,480 --> 00:30:10,880
Speaker 1: propping up Sam Bankman freed two weeks before FTX collapsed,

612
00:30:11,240 --> 00:30:15,240
Speaker 1: who then went on immediately to cover AI fucking and

613
00:30:15,320 --> 00:30:19,120
Speaker 1: now they're peddling bullshit for anthropic and sorry, I'm kind

614
00:30:19,120 --> 00:30:21,200
Speaker 1: of hating. Well, I guess it's hated season, so I can.

615
00:30:21,360 --> 00:30:25,600
Speaker 1: It just frustrates me because regular people are being scared.

616
00:30:26,200 --> 00:30:29,840
Speaker 1: They're being scared by the kind of astonishment which I

617
00:30:29,840 --> 00:30:33,280
Speaker 1: actually loved the astonishment reporting of like, oh well, open

618
00:30:33,320 --> 00:30:36,840
Speaker 1: clawd has proven open claw has proven that ai agi

619
00:30:36,960 --> 00:30:39,560
Speaker 1: is here, or they've built their own social networks that

620
00:30:39,560 --> 00:30:46,920
Speaker 1: we should be scared, scared insane did Yeah, And it's like,

621
00:30:47,320 --> 00:30:51,000
Speaker 1: I assume you saw that for the listeners. By the way,

622
00:30:51,120 --> 00:30:54,960
Speaker 1: open Claw had this this fucking claude bot whatever it is.

623
00:30:55,000 --> 00:30:57,840
Speaker 1: They had their own social network where the quote unquote

624
00:30:57,880 --> 00:31:00,640
Speaker 1: lms would post, but it turns out that those posts

625
00:31:00,640 --> 00:31:02,040
Speaker 1: would just make by their owners.

626
00:31:02,320 --> 00:31:04,280
Speaker 2: Yeah, and this is just I mean, this is a

627
00:31:04,280 --> 00:31:04,880
Speaker 2: good case study.

628
00:31:04,920 --> 00:31:05,040
Speaker 1: Right.

629
00:31:05,080 --> 00:31:07,239
Speaker 2: This one bothered me because like writer friends I know

630
00:31:07,440 --> 00:31:09,920
Speaker 2: who are not technology related, we're texting this to me.

631
00:31:10,480 --> 00:31:13,120
Speaker 2: They were worried, right, They're texting me these articles like

632
00:31:13,120 --> 00:31:15,480
Speaker 2: this seems bad, Like this seems like something really bad.

633
00:31:15,520 --> 00:31:18,400
Speaker 2: They were really getting the digital ic really strongly off

634
00:31:18,400 --> 00:31:21,200
Speaker 2: of these articles. And I would start reading these articles

635
00:31:21,200 --> 00:31:24,520
Speaker 2: and I say, well, there's no discussion of what is

636
00:31:24,560 --> 00:31:28,440
Speaker 2: the technical breakthrough here? And what are the concrete implications,

637
00:31:28,480 --> 00:31:31,720
Speaker 2: Because it turns out there was zero technical breakthroughs. There

638
00:31:31,800 --> 00:31:36,760
Speaker 2: is no new AI technology connected to open claw, which

639
00:31:36,840 --> 00:31:40,880
Speaker 2: used to be called molt whatever mold or open clawed

640
00:31:40,960 --> 00:31:43,400
Speaker 2: or whatever open claw is, I think where they ended

641
00:31:43,480 --> 00:31:46,800
Speaker 2: up right, which is an open source library or framework

642
00:31:46,840 --> 00:31:50,320
Speaker 2: for building AI agents powered by elims. There's no new

643
00:31:50,360 --> 00:31:53,360
Speaker 2: AI technology involved in this at all. The agents you

644
00:31:53,360 --> 00:31:56,520
Speaker 2: build are just accessing off the shelf elms that we're

645
00:31:56,520 --> 00:31:58,200
Speaker 2: all using for chatbots anyways. You can aim it at

646
00:31:58,240 --> 00:32:02,640
Speaker 2: whatever commercial chatpot you want. There's no new framework for

647
00:32:02,680 --> 00:32:04,800
Speaker 2: how the agents work. It's the same sort of react

648
00:32:04,800 --> 00:32:06,240
Speaker 2: loop that we've been trying for the last two or

649
00:32:06,240 --> 00:32:08,040
Speaker 2: three years, where you just you have a program. It's

650
00:32:08,080 --> 00:32:11,520
Speaker 2: like a bit of Python code that asks an LM, hey,

651
00:32:11,520 --> 00:32:13,120
Speaker 2: here's what I want to do, here's the tools you

652
00:32:13,120 --> 00:32:15,000
Speaker 2: have available. Come up with a plan, and then it

653
00:32:15,040 --> 00:32:18,160
Speaker 2: sends it back a plan, and then the Python code

654
00:32:18,600 --> 00:32:21,040
Speaker 2: takes the first step out of that text and says, okay,

655
00:32:22,120 --> 00:32:25,160
Speaker 2: here's the tools available. What should I do to execute

656
00:32:25,160 --> 00:32:27,480
Speaker 2: this step? And then like the LM will give you

657
00:32:27,480 --> 00:32:30,880
Speaker 2: some steps and then the Python code runs those steps

658
00:32:30,880 --> 00:32:33,240
Speaker 2: and then you just go back and forth right. That's

659
00:32:33,320 --> 00:32:34,840
Speaker 2: like this basic react, which.

660
00:32:34,640 --> 00:32:37,040
Speaker 1: Is exactly how Manus worked as well.

661
00:32:37,200 --> 00:32:38,440
Speaker 2: Everything there's nothing new about this.

662
00:32:38,840 --> 00:32:42,320
Speaker 1: Manus was this AI agent that Facebook might be Meta

663
00:32:42,440 --> 00:32:45,040
Speaker 1: might be acquiring. They literally when you use it, it

664
00:32:45,120 --> 00:32:47,000
Speaker 1: just writes prithon code for every step.

665
00:32:47,360 --> 00:32:49,400
Speaker 2: Well, so it's yeah, so this, I mean, there's nothing

666
00:32:49,440 --> 00:32:51,560
Speaker 2: new here technologically. The only thing that was new about

667
00:32:51,560 --> 00:32:55,840
Speaker 2: openclaw is it was open source, so it made it

668
00:32:55,880 --> 00:32:58,960
Speaker 2: easy for anyone to write bad agents. And the other

669
00:32:59,000 --> 00:33:01,320
Speaker 2: thing that made it interesting to wireheads in San Francisco

670
00:33:01,440 --> 00:33:03,360
Speaker 2: is that the commercial products where people are trying to

671
00:33:03,400 --> 00:33:07,560
Speaker 2: build these companies, they have common sense security, like, well,

672
00:33:07,560 --> 00:33:10,520
Speaker 2: probably if it's just a Python program blindly doing what

673
00:33:10,560 --> 00:33:12,800
Speaker 2: an LLM tells it to do, like it probably shouldn't

674
00:33:12,800 --> 00:33:15,040
Speaker 2: have access to like your credit cards or to your

675
00:33:15,040 --> 00:33:17,400
Speaker 2: hard drive or whatever. At open cloud like you can

676
00:33:17,440 --> 00:33:19,280
Speaker 2: just you can give it access to anything you want

677
00:33:19,320 --> 00:33:21,640
Speaker 2: on your computer, and so you could build really cool

678
00:33:21,640 --> 00:33:25,160
Speaker 2: demos that are also like incredibly insecure and unsafe. And

679
00:33:25,240 --> 00:33:27,239
Speaker 2: so it was fun for hobbyists, but there was no

680
00:33:27,320 --> 00:33:30,280
Speaker 2: new technology there. Nothing was new, So it was just

681
00:33:30,360 --> 00:33:33,800
Speaker 2: a way for other people like hobbyists to build their

682
00:33:33,840 --> 00:33:37,120
Speaker 2: own agents. That were like less safe than the more

683
00:33:37,160 --> 00:33:39,600
Speaker 2: carefully built Like what people don't here's a story people

684
00:33:39,640 --> 00:33:42,320
Speaker 2: don't know is like in the immediate aftermath of chat GPT.

685
00:33:42,400 --> 00:33:44,720
Speaker 2: So we're in early twenty twenty three, right, so chatchepts

686
00:33:44,760 --> 00:33:47,600
Speaker 2: come out. Open AI goes on a road tour of

687
00:33:47,720 --> 00:33:51,040
Speaker 2: major publications, right, and they're to try to Hey, here's

688
00:33:51,040 --> 00:33:53,040
Speaker 2: what's going on. You need to write about this. In

689
00:33:53,080 --> 00:33:55,160
Speaker 2: early twenty twenty three, they're like, the next thing we're

690
00:33:55,200 --> 00:33:57,120
Speaker 2: going to offer is something like these agents. They call

691
00:33:57,160 --> 00:34:00,960
Speaker 2: them plugins, and you can install plugins that basically can

692
00:34:02,280 --> 00:34:05,760
Speaker 2: do actions on behalf of your LLLM queries. You could

693
00:34:05,760 --> 00:34:08,560
Speaker 2: have like a book an airline flight plug in and say, hey,

694
00:34:08,640 --> 00:34:11,560
Speaker 2: chatchipt book me a flight, and a language model can't

695
00:34:11,560 --> 00:34:13,359
Speaker 2: do anything but produce text, but then the plugin could

696
00:34:13,400 --> 00:34:15,479
Speaker 2: take the text and go and book you a flight.

697
00:34:15,560 --> 00:34:17,839
Speaker 2: And they're like, this is the future, obviously, and that

698
00:34:17,880 --> 00:34:22,400
Speaker 2: project disappeared because oh, it's incredibly unsafe and unreliable to

699
00:34:22,760 --> 00:34:25,279
Speaker 2: have code that could interact with the real world that's

700
00:34:25,640 --> 00:34:28,840
Speaker 2: following commands from an unreliable who's stayed an LLM like

701
00:34:28,880 --> 00:34:32,120
Speaker 2: that went away not because the technology was hard, but

702
00:34:32,160 --> 00:34:35,480
Speaker 2: because it's not safe. So there's nothing new. My whole

703
00:34:35,600 --> 00:34:37,959
Speaker 2: article on agents, they tried this all at twenty twenty five,

704
00:34:38,400 --> 00:34:40,360
Speaker 2: and we're failing to get this type of agent to

705
00:34:40,400 --> 00:34:43,200
Speaker 2: work for anything outside of basically like producing computer code.

706
00:34:43,320 --> 00:34:45,960
Speaker 2: So open clause is nothing new. It was just a

707
00:34:46,000 --> 00:34:48,000
Speaker 2: way for other people to build these things. The only

708
00:34:48,040 --> 00:34:51,600
Speaker 2: interesting stories were security whole stories. But it was reported. Man,

709
00:34:51,640 --> 00:34:53,839
Speaker 2: I was listening to the All In podcast and they

710
00:34:53,840 --> 00:34:57,799
Speaker 2: were tasing why this is the future of AI, Like

711
00:34:57,840 --> 00:35:00,919
Speaker 2: this is it. Everyone is going to have of an

712
00:35:00,960 --> 00:35:04,400
Speaker 2: open claw agent as like a personal assistant, and this

713
00:35:04,560 --> 00:35:06,680
Speaker 2: is like the biggest thing in open AI. And I

714
00:35:06,719 --> 00:35:08,879
Speaker 2: don't know who it was. One of those guys was like, yeah,

715
00:35:08,880 --> 00:35:13,360
Speaker 2: we replaced our podcast producer, you know, with this agent

716
00:35:13,360 --> 00:35:15,479
Speaker 2: who can like email guests on our behalf and book

717
00:35:15,480 --> 00:35:17,080
Speaker 2: it on our calendar. And well it's costing about a

718
00:35:17,120 --> 00:35:19,400
Speaker 2: thousand dollars a day right now to run it. But like,

719
00:35:19,880 --> 00:35:22,560
Speaker 2: I don't know why we're going to need employees in

720
00:35:22,600 --> 00:35:23,040
Speaker 2: the future.

721
00:35:23,120 --> 00:35:24,520
Speaker 1: Oh yeah brother, Yeah.

722
00:35:25,040 --> 00:35:27,160
Speaker 2: So anyways, like it's a non if you said what's

723
00:35:27,160 --> 00:35:29,040
Speaker 2: the technical implications, not if you say, what are the

724
00:35:29,040 --> 00:35:31,400
Speaker 2: concrete implications? The best story they can have is like

725
00:35:32,080 --> 00:35:34,680
Speaker 2: maybe when people I don't know what it is. All

726
00:35:34,719 --> 00:35:36,920
Speaker 2: the companies will try to build these things for years,

727
00:35:36,960 --> 00:35:41,120
Speaker 2: so I don't know, but it was reported like just

728
00:35:41,200 --> 00:35:44,719
Speaker 2: something icky is happening, and that molt book was an

729
00:35:44,719 --> 00:35:48,000
Speaker 2: application that was built for these agents to communicate on

730
00:35:48,080 --> 00:35:52,319
Speaker 2: like a Reddit style social network. They vibe coded that

731
00:35:52,400 --> 00:35:54,239
Speaker 2: framework and it was full of security holes as we

732
00:35:54,280 --> 00:35:57,239
Speaker 2: mentioned before, and there was like a small number of

733
00:35:57,320 --> 00:35:59,080
Speaker 2: users that created a huge amount of agents, and they

734
00:35:59,080 --> 00:36:01,160
Speaker 2: were just kind of like prompting and prodding their agents

735
00:36:01,160 --> 00:36:03,600
Speaker 2: to produce, like let's talk about creating around church or

736
00:36:03,640 --> 00:36:06,840
Speaker 2: killing humanity. It's like, guys, this you have Python code

737
00:36:06,880 --> 00:36:10,160
Speaker 2: asking lllms to like write text in the style of

738
00:36:10,200 --> 00:36:13,200
Speaker 2: like the matrix, and you're posting it on a fake

739
00:36:13,280 --> 00:36:17,239
Speaker 2: social network. The real story here should be where they

740
00:36:17,280 --> 00:36:20,000
Speaker 2: don't these people have things to do? Don't they have jobs?

741
00:36:20,040 --> 00:36:20,680
Speaker 1: Like what is y? Yeah?

742
00:36:20,719 --> 00:36:24,040
Speaker 2: Yeah, yeah, really brilliant my model train, I added a

743
00:36:24,040 --> 00:36:26,279
Speaker 2: tunnel so that one really got no.

744
00:36:26,600 --> 00:36:29,160
Speaker 1: I actually will push back on that. Mudile train listeners

745
00:36:29,400 --> 00:36:32,480
Speaker 1: I think make up twenty percent of the listenership of

746
00:36:32,560 --> 00:36:36,719
Speaker 1: this podcast course. Also, modile trains are cool. Signaling to

747
00:36:36,760 --> 00:36:40,040
Speaker 1: my fellow autists out there. But you know, mudtile trains

748
00:36:40,040 --> 00:36:41,880
Speaker 1: are a physical thing which you build it, you build

749
00:36:41,880 --> 00:36:44,880
Speaker 1: a little a little city. I think that delightful. Respect

750
00:36:44,920 --> 00:36:47,200
Speaker 1: to those with this. It's like if they and you

751
00:36:47,239 --> 00:36:48,920
Speaker 1: know what, if they were just saying I have been

752
00:36:49,000 --> 00:36:51,120
Speaker 1: fucking around with some software and it did something cool,

753
00:36:51,239 --> 00:36:53,160
Speaker 1: I'd respect the shit out. Then I'd be like, yeah,

754
00:36:53,320 --> 00:36:56,359
Speaker 1: enjoy yourself. It's going away. This is a little subsidized rites,

755
00:36:56,400 --> 00:36:58,160
Speaker 1: but have fun with that. I don't know why you

756
00:36:58,160 --> 00:37:01,160
Speaker 1: need a max studio, but good luck. No, they are like,

757
00:37:01,360 --> 00:37:04,480
Speaker 1: this is the future. But I'm gonna be honest, cal

758
00:37:04,840 --> 00:37:08,320
Speaker 1: My real question is what does openclaw actually do? Because

759
00:37:08,320 --> 00:37:11,120
Speaker 1: when I went and read what it actually I read

760
00:37:11,200 --> 00:37:14,719
Speaker 1: so many posts you said it books callen does and

761
00:37:14,760 --> 00:37:17,799
Speaker 1: does this. I could find no proof that anyone successfully

762
00:37:17,800 --> 00:37:18,320
Speaker 1: did that.

763
00:37:18,200 --> 00:37:20,800
Speaker 2: And it doesn't. It's it's just a series of library

764
00:37:20,840 --> 00:37:23,800
Speaker 2: calls that you can use to build your own program

765
00:37:23,880 --> 00:37:26,360
Speaker 2: in theory that would do that. So open claw itself

766
00:37:26,480 --> 00:37:28,840
Speaker 2: is just like a series of like interfaces and hooks

767
00:37:28,840 --> 00:37:31,840
Speaker 2: that makes it easy to write a program that talks

768
00:37:31,840 --> 00:37:34,640
Speaker 2: to other services basically and makes calls to LLLM. So

769
00:37:34,680 --> 00:37:37,200
Speaker 2: it's just like a wrapper in which you can write

770
00:37:37,239 --> 00:37:40,160
Speaker 2: your own code, and some people are trying, Yeah, like

771
00:37:40,200 --> 00:37:43,040
Speaker 2: you can you can write, you can have it talk

772
00:37:43,120 --> 00:37:45,080
Speaker 2: to your calendar, you can have it talk to your

773
00:37:45,120 --> 00:37:46,160
Speaker 2: email and theory and.

774
00:37:46,200 --> 00:37:48,239
Speaker 1: Look, but does it doesn't work?

775
00:37:48,360 --> 00:37:50,840
Speaker 2: Well, it's it's just asking an lell you can simulate.

776
00:37:51,040 --> 00:37:52,759
Speaker 2: I mean, I did this for my New Yorker piece.

777
00:37:52,800 --> 00:37:54,720
Speaker 2: I was like, look, I can just simulate being an agent.

778
00:37:54,800 --> 00:37:57,560
Speaker 2: Just ask any LLM, here's what I'm trying to do.

779
00:37:57,640 --> 00:37:59,640
Speaker 2: Give me like the steps for doing this or whatever,

780
00:37:59,800 --> 00:38:02,840
Speaker 2: and like anything else you ask of an LOLLM, it

781
00:38:02,880 --> 00:38:05,480
Speaker 2: will give you answers that sound very reasonable in general

782
00:38:05,520 --> 00:38:07,200
Speaker 2: and then have like a lot of issues in the details.

783
00:38:07,200 --> 00:38:10,520
Speaker 2: Which is why agents based on llms have struggled because

784
00:38:10,560 --> 00:38:12,719
Speaker 2: if you it's fine if you as a human are

785
00:38:12,719 --> 00:38:14,839
Speaker 2: talking to an M, because you don't realize how much

786
00:38:14,840 --> 00:38:17,080
Speaker 2: filtering and tweaking. You're like, well, that's kind of ignore

787
00:38:17,120 --> 00:38:18,600
Speaker 2: that and this is good, or let me ask you

788
00:38:18,640 --> 00:38:20,759
Speaker 2: to redo it. It's really an issue if you write

789
00:38:20,760 --> 00:38:23,759
Speaker 2: a program that just says I will do whatever. I'll

790
00:38:23,800 --> 00:38:25,840
Speaker 2: ask the LLM for a plan and then just do

791
00:38:26,000 --> 00:38:30,440
Speaker 2: whatever it says because it doesn't know that doesn't really

792
00:38:30,520 --> 00:38:33,960
Speaker 2: make sen or like this sounds generally generally reasonable, but

793
00:38:34,040 --> 00:38:36,480
Speaker 2: with like issues in the details. Does it work out

794
00:38:36,560 --> 00:38:39,680
Speaker 2: when you execute it? I mean, in a practical sense,

795
00:38:40,160 --> 00:38:43,160
Speaker 2: can it? They said only all in podcasts, three dumb

796
00:38:43,160 --> 00:38:47,400
Speaker 2: bitches saying exactly, that's the main reference. Don't use that

797
00:38:47,440 --> 00:38:50,960
Speaker 2: word usually those people sitting around going blah blah blah.

798
00:38:51,239 --> 00:38:54,400
Speaker 1: Oh we've made it and replaced our podcast producer with

799
00:38:54,520 --> 00:38:59,760
Speaker 1: an LM that can make appointments and send emails in practice.

800
00:38:59,840 --> 00:39:03,560
Speaker 1: Is that true? I severely doubt that. I just I Also,

801
00:39:03,760 --> 00:39:05,919
Speaker 1: you're spending on thousand dollars a day, so you're paying

802
00:39:05,920 --> 00:39:09,640
Speaker 1: three hundred and sixty five thousand dollars a year for this, right?

803
00:39:11,080 --> 00:39:14,160
Speaker 1: Are you really or are you just? Have you? If

804
00:39:14,200 --> 00:39:17,040
Speaker 1: it is it completely it probably isn't. And that's the

805
00:39:17,120 --> 00:39:20,320
Speaker 1: thing I don't I get why they all in guys

806
00:39:20,360 --> 00:39:23,440
Speaker 1: do it because they probably have investments and they're boosters TVPN,

807
00:39:24,080 --> 00:39:28,239
Speaker 1: same fucking deal. It really rules that the two largest

808
00:39:28,440 --> 00:39:32,240
Speaker 1: like tech things in the valley are just state media

809
00:39:32,360 --> 00:39:36,560
Speaker 1: book for Silicon Valley. What gets me is when you

810
00:39:36,600 --> 00:39:39,759
Speaker 1: get like the Atlantic CNBC Business Insider in places like

811
00:39:39,800 --> 00:39:41,880
Speaker 1: that doing stuff like this, and it bothers me because

812
00:39:42,160 --> 00:39:44,319
Speaker 1: they don't even need the hate on it. They could

813
00:39:44,440 --> 00:39:47,120
Speaker 1: just say, yeah, I could do this, it's pretty cool,

814
00:39:47,200 --> 00:39:51,320
Speaker 1: right yeah, But I guess that that doesn't get the clear.

815
00:39:51,480 --> 00:39:53,640
Speaker 2: It's not an interesting story. That's not the real story

816
00:39:53,719 --> 00:39:55,759
Speaker 2: is not always that interesting. It's like like hobbyists are

817
00:39:55,760 --> 00:39:57,920
Speaker 2: building these tools that kind of do cool things but

818
00:39:58,040 --> 00:40:00,640
Speaker 2: make mistakes, and it's a little expensive. Like the most

819
00:40:00,680 --> 00:40:04,040
Speaker 2: interesting story out of open claw is the only thing

820
00:40:04,080 --> 00:40:05,880
Speaker 2: I think is going to be impactful out of it

821
00:40:05,920 --> 00:40:08,520
Speaker 2: is because it was so expensive, right, like that thousand

822
00:40:08,520 --> 00:40:11,680
Speaker 2: dollars a day. What it's forcing people, these hobbyists to

823
00:40:11,760 --> 00:40:15,440
Speaker 2: do is to turn to like cheaper alternatives for models,

824
00:40:15,880 --> 00:40:18,600
Speaker 2: and that I do think is significant, this idea that

825
00:40:19,040 --> 00:40:23,480
Speaker 2: there's open source models out there, they're significantly cheaper than

826
00:40:23,520 --> 00:40:25,719
Speaker 2: trying to use like open ai or claud. More and

827
00:40:25,719 --> 00:40:27,960
Speaker 2: more people are running their own local models because it

828
00:40:28,000 --> 00:40:30,600
Speaker 2: turns out, for you know, most specific uses you might

829
00:40:30,680 --> 00:40:33,640
Speaker 2: use in LLM, you don't need like a super fine

830
00:40:33,680 --> 00:40:36,120
Speaker 2: tune trillion parameter beast of a model running in some

831
00:40:36,239 --> 00:40:38,440
Speaker 2: data center somewhere. It's like, you know what, I'm parsing

832
00:40:38,520 --> 00:40:42,520
Speaker 2: my email that try to extract like suggested times. I'm

833
00:40:42,600 --> 00:40:45,319
Speaker 2: fine with like a twenty billion parameter model that can

834
00:40:45,400 --> 00:40:48,000
Speaker 2: easily fit in a single GPU on a thing I have,

835
00:40:48,160 --> 00:40:50,600
Speaker 2: you know, we share in our office or something like that.

836
00:40:50,760 --> 00:40:52,680
Speaker 2: So to me, that's the most interesting story out of

837
00:40:52,680 --> 00:40:55,120
Speaker 2: open claw is the bad news it could be for

838
00:40:55,239 --> 00:40:59,440
Speaker 2: the big companies as people get more comfortable with we

839
00:40:59,520 --> 00:41:04,040
Speaker 2: don't needed these formula one car versions of language models

840
00:41:04,080 --> 00:41:07,600
Speaker 2: for the stuff we're actually doing. We're fine with the

841
00:41:07,640 --> 00:41:10,279
Speaker 2: forward focus, right, And I think that's a transition that's

842
00:41:10,320 --> 00:41:12,680
Speaker 2: going to That's a transition that to me is more interest.

843
00:41:12,680 --> 00:41:14,200
Speaker 2: That's what we should be writing about. Like, that's an

844
00:41:14,200 --> 00:41:16,399
Speaker 2: interesting idea to me is that you have these huge,

845
00:41:16,440 --> 00:41:19,240
Speaker 2: high valuation companies, but you also have all these open

846
00:41:19,239 --> 00:41:21,279
Speaker 2: source models, like the weights are just out there in

847
00:41:21,320 --> 00:41:23,319
Speaker 2: the public domain that can do ninety eight percent of

848
00:41:23,320 --> 00:41:25,480
Speaker 2: what people care about, and you're beginning to get low

849
00:41:25,520 --> 00:41:28,239
Speaker 2: cost competitive services where people can just spin those up

850
00:41:28,280 --> 00:41:31,120
Speaker 2: in cheaper data centers. That's interesting to me. That's an

851
00:41:31,120 --> 00:41:34,319
Speaker 2: economic story. AI creating its own church is not as

852
00:41:34,480 --> 00:41:35,400
Speaker 2: not as relevant to me.

853
00:41:50,040 --> 00:41:53,239
Speaker 1: I'm already working on a story full this Friday in

854
00:41:53,320 --> 00:41:56,800
Speaker 1: fact oh my premium newsletter about the fact that actually

855
00:41:56,840 --> 00:42:00,120
Speaker 1: the margins of serving GPU compute are dog shit like

856
00:42:00,160 --> 00:42:02,400
Speaker 1: the best of the best of like a co location

857
00:42:02,600 --> 00:42:05,960
Speaker 1: place like Applied Digital, it's like twenty seven percent gross

858
00:42:06,000 --> 00:42:09,320
Speaker 1: margins and that's at one hundred percent utilization. Anything below

859
00:42:09,320 --> 00:42:12,799
Speaker 1: point seven they're burning cash. I hear this date center

860
00:42:12,800 --> 00:42:14,759
Speaker 1: out in North Dakota losing a million dollars a day.

861
00:42:15,040 --> 00:42:17,919
Speaker 1: That's the thing. This is even with these low these

862
00:42:17,960 --> 00:42:20,680
Speaker 1: lower cost models on device could be interesting.

863
00:42:20,680 --> 00:42:22,600
Speaker 2: That's what's going to happen. I think, I think undevice

864
00:42:22,680 --> 00:42:23,480
Speaker 2: is what's going to happen.

865
00:42:24,080 --> 00:42:26,880
Speaker 1: I just remembered something. This is a classic bullshit story

866
00:42:26,880 --> 00:42:29,239
Speaker 1: that I see every so often. So have you read

867
00:42:29,239 --> 00:42:32,040
Speaker 1: any of the stories that are like, yeah, Claude can

868
00:42:32,080 --> 00:42:35,439
Speaker 1: now work for hours uninterrupted? Have you read about these?

869
00:42:35,719 --> 00:42:36,839
Speaker 1: You heard this? You've seen this.

870
00:42:37,320 --> 00:42:39,720
Speaker 2: Work for our Oh yeah, yeah, we were talking about

871
00:42:39,760 --> 00:42:42,400
Speaker 2: the Yeah, the multi step agenda execution.

872
00:42:42,520 --> 00:42:43,719
Speaker 1: Well, no way, it keeps coming back.

873
00:42:44,080 --> 00:42:45,759
Speaker 2: Yeah, yeah.

874
00:42:45,840 --> 00:42:48,439
Speaker 1: AI on the CNBC AI on the verge of eight

875
00:42:48,480 --> 00:42:51,000
Speaker 1: hour job shift without burnout or break is it is

876
00:42:51,040 --> 00:42:55,440
Speaker 1: twenty four hour our AI workday. Next going to just

877
00:42:55,600 --> 00:42:58,120
Speaker 1: censor myself what I was going to say there, because

878
00:42:58,160 --> 00:43:02,440
Speaker 1: it's just like yeah, it can work for hours. Is

879
00:43:02,480 --> 00:43:03,520
Speaker 1: the output good?

880
00:43:03,760 --> 00:43:05,319
Speaker 2: Does it a problem? Yeah, just a lose?

881
00:43:05,360 --> 00:43:05,719
Speaker 1: Does it? Will?

882
00:43:05,800 --> 00:43:08,200
Speaker 2: Just keep reading? Call it? Recall it? Recall it like

883
00:43:08,239 --> 00:43:11,160
Speaker 2: I can. I can write a Python program that calls

884
00:43:11,160 --> 00:43:12,560
Speaker 2: an hour for twenty four hours.

885
00:43:12,760 --> 00:43:15,640
Speaker 1: If you need me to burn something for hours on end,

886
00:43:15,719 --> 00:43:17,960
Speaker 1: I just give me some gasoline. I got your baby.

887
00:43:18,200 --> 00:43:20,759
Speaker 1: I've sort this right out. Be much cheaper, but it's

888
00:43:20,760 --> 00:43:23,680
Speaker 1: like yeah. By September twenty twenty five, Claudes on It

889
00:43:23,760 --> 00:43:26,000
Speaker 1: four point five is reported to run autonomously for up

890
00:43:26,000 --> 00:43:29,840
Speaker 1: to thirty hours, reported with com I mean, but that's

891
00:43:30,280 --> 00:43:34,000
Speaker 1: more vibe reporting. It's you're meant to read that and go, wow,

892
00:43:34,200 --> 00:43:38,919
Speaker 1: this thing is performing tasks that are useful, that execute code,

893
00:43:38,920 --> 00:43:41,520
Speaker 1: that do something, and in that thirty hours that is

894
00:43:41,560 --> 00:43:45,080
Speaker 1: equivalent to thirty human working hours versus they sat there

895
00:43:45,120 --> 00:43:48,160
Speaker 1: and pissed their pants for like twenty nine point five

896
00:43:48,280 --> 00:43:50,080
Speaker 1: of those at least well.

897
00:43:50,120 --> 00:43:52,719
Speaker 2: And also that those are like specialized TSK typically that

898
00:43:54,480 --> 00:43:56,920
Speaker 2: have clear milestones and testing, so it's like it can

899
00:43:56,960 --> 00:44:00,640
Speaker 2: do something. The loop can try and try until it

900
00:44:00,640 --> 00:44:02,520
Speaker 2: passes the test, and then you know that's done, and

901
00:44:02,520 --> 00:44:03,920
Speaker 2: then you can move on to the next step. And

902
00:44:03,920 --> 00:44:06,280
Speaker 2: then you can keep calling to l M and executing

903
00:44:06,360 --> 00:44:08,200
Speaker 2: until it passes the next test and you can move on.

904
00:44:08,560 --> 00:44:12,720
Speaker 2: And in theory like it's yeah, they're avoiding error cascade

905
00:44:12,719 --> 00:44:14,959
Speaker 2: because it's testable and they can keep retrying or something

906
00:44:15,000 --> 00:44:16,520
Speaker 2: like that. I mean, this was like the meter had

907
00:44:16,560 --> 00:44:19,880
Speaker 2: this problem with their graph of how long of tasks

908
00:44:19,960 --> 00:44:22,719
Speaker 2: AI can now do, and it was like the sort

909
00:44:22,719 --> 00:44:25,440
Speaker 2: of like super arbitrary decision of like, oh, here's a

910
00:44:25,480 --> 00:44:27,719
Speaker 2: test that takes a human five hours, Now AI can

911
00:44:27,760 --> 00:44:30,440
Speaker 2: do that. Here's a task, whereas really more about like

912
00:44:30,600 --> 00:44:33,560
Speaker 2: how many things in a row can you do without

913
00:44:33,600 --> 00:44:36,040
Speaker 2: the errors cascading out of control or something like that.

914
00:44:36,440 --> 00:44:37,400
Speaker 2: It's very different.

915
00:44:38,280 --> 00:44:41,360
Speaker 1: You're being way nicer than you should be, in my opinion,

916
00:44:41,440 --> 00:44:45,799
Speaker 1: just because they don't even explain. They don't even say like, yeah,

917
00:44:45,840 --> 00:44:47,560
Speaker 1: and we managed to make it through it and let's

918
00:44:47,600 --> 00:44:51,480
Speaker 1: go bus right hours c NBC just fucking woo just

919
00:44:51,520 --> 00:44:53,520
Speaker 1: for tearing their shirts off and screaming.

920
00:44:54,080 --> 00:44:56,160
Speaker 2: But it goes back to the two things. We need

921
00:44:56,160 --> 00:44:59,000
Speaker 2: this report and you need technical the details of the

922
00:44:59,080 --> 00:45:02,800
Speaker 2: relevant technical innovation, and a discussion of concrete implications in

923
00:45:02,840 --> 00:45:06,000
Speaker 2: your future implications of that breakthrough. That's what I'm always

924
00:45:06,000 --> 00:45:07,400
Speaker 2: looking for, So, like, if you want to cover a

925
00:45:07,440 --> 00:45:09,600
Speaker 2: story like that, like, well, what does that mean? What

926
00:45:09,760 --> 00:45:12,800
Speaker 2: is the technical breakthrough? Right? What does this mean technically

927
00:45:12,800 --> 00:45:14,720
Speaker 2: you can do thirty hours of work or eight hours?

928
00:45:14,840 --> 00:45:17,360
Speaker 2: What is the work? What was changed? What did they figure?

929
00:45:17,560 --> 00:45:21,279
Speaker 2: What was happening before? What technical breakthrough made this possible now?

930
00:45:21,400 --> 00:45:24,280
Speaker 2: And then what are the concrete implications? What specific things now?

931
00:45:24,600 --> 00:45:26,920
Speaker 2: Can we directly this will now allow us to do? What?

932
00:45:26,960 --> 00:45:28,640
Speaker 2: Tell us what jobs are going away? Tell us what

933
00:45:28,719 --> 00:45:30,520
Speaker 2: tool we're going to see, like you have to, but

934
00:45:31,280 --> 00:45:34,080
Speaker 2: we avoid that, right because then you're putting your chips

935
00:45:34,080 --> 00:45:37,239
Speaker 2: down and then those things don't come along, And so

936
00:45:37,320 --> 00:45:39,400
Speaker 2: I'm always looking for that. What's the technical innovation and

937
00:45:39,400 --> 00:45:41,239
Speaker 2: what are the concrete implications? If you don't have that,

938
00:45:41,280 --> 00:45:43,400
Speaker 2: you're mining emotions.

939
00:45:43,600 --> 00:45:46,279
Speaker 1: You're mining emotions or just helping boost stook. That's what

940
00:45:46,360 --> 00:45:50,400
Speaker 1: really bothers me because I hate that they're scaring people.

941
00:45:50,840 --> 00:45:54,880
Speaker 1: I really really hate that they're doing marketing like they're

942
00:45:54,920 --> 00:45:59,320
Speaker 1: just doing like I run a PIAFA. I've dealt with

943
00:45:59,400 --> 00:46:04,080
Speaker 1: ether these states startups for like fifteen years. There isn't

944
00:46:04,120 --> 00:46:06,879
Speaker 1: a single early stage startup that would get a percentage

945
00:46:06,920 --> 00:46:11,319
Speaker 1: point of this bullshit, Like you email the reporter about

946
00:46:11,360 --> 00:46:14,160
Speaker 1: like a series a start. They're like, all right, all right, motherfucker,

947
00:46:15,120 --> 00:46:18,160
Speaker 1: how much revenue are they profitable? Yet? Why not? Why

948
00:46:18,200 --> 00:46:21,359
Speaker 1: not explain to me right now? Anthropics like we're gonna

949
00:46:21,440 --> 00:46:24,920
Speaker 1: gonna burn a hundred billion on training. I guess what

950
00:46:25,000 --> 00:46:27,600
Speaker 1: do you think? And they're like, yeah, I love it.

951
00:46:27,640 --> 00:46:29,680
Speaker 1: I actually think that's huge, that's great. I'm not And

952
00:46:29,719 --> 00:46:32,200
Speaker 1: to be clear, I think that this scrutiny should be

953
00:46:32,520 --> 00:46:34,920
Speaker 1: from the beginning to the end. I think that everyone

954
00:46:34,920 --> 00:46:38,200
Speaker 1: should face this scrutiny. Yeah, I'm not saying that I

955
00:46:38,200 --> 00:46:41,840
Speaker 1: should get an easier I'm saying, actually everyone should. Anthropics

956
00:46:41,840 --> 00:46:47,160
Speaker 1: should face the most brutal scrutiny, and I guess that

957
00:46:47,239 --> 00:46:50,920
Speaker 1: they don't because they want access. It's just very dull

958
00:46:51,320 --> 00:46:54,480
Speaker 1: and annoying, and it's just helping already rich guys like

959
00:46:54,520 --> 00:46:58,560
Speaker 1: Wario Ama Day, who should not have more money. Listen

960
00:46:58,600 --> 00:47:01,360
Speaker 1: to him speak, he needs he needs to face some stress.

961
00:47:02,520 --> 00:47:05,239
Speaker 1: I think some stress would help him grow. But caw,

962
00:47:05,320 --> 00:47:07,200
Speaker 1: as we wrap up, I did actually have like a

963
00:47:07,480 --> 00:47:10,839
Speaker 1: technical thing that wants an idea of being perclaying. So

964
00:47:11,520 --> 00:47:15,440
Speaker 1: I think that this term training with models is being

965
00:47:15,480 --> 00:47:18,319
Speaker 1: misused and used in a way that is kind of

966
00:47:18,360 --> 00:47:21,880
Speaker 1: vibe reporting style, which is they use training as a

967
00:47:21,920 --> 00:47:24,440
Speaker 1: word that suggests that it will stop, that they will

968
00:47:24,440 --> 00:47:27,160
Speaker 1: stop training these models. But from correct me if I'm wrong.

969
00:47:27,320 --> 00:47:30,640
Speaker 1: Training is everything from building a new model to updating

970
00:47:30,640 --> 00:47:32,480
Speaker 1: and models current parameters. Correct.

971
00:47:32,880 --> 00:47:35,640
Speaker 2: Yeah, there's pre training and post training is the right

972
00:47:35,640 --> 00:47:38,080
Speaker 2: way to think about it. So the pre training is unsupervised.

973
00:47:38,120 --> 00:47:40,120
Speaker 2: That's where you take all the text that's ever been written,

974
00:47:40,600 --> 00:47:42,520
Speaker 2: and you will take a real piece of text written

975
00:47:42,520 --> 00:47:44,360
Speaker 2: by a human, and you'll cut it off at an

976
00:47:44,440 --> 00:47:47,840
Speaker 2: arbitrary point, and you'll tell the language model your guess

977
00:47:47,920 --> 00:47:50,000
Speaker 2: the word that came next. There's a real word that

978
00:47:50,000 --> 00:47:52,359
Speaker 2: came next. This is a real writing Guess the word

979
00:47:52,400 --> 00:47:54,239
Speaker 2: that came next, and then it guesses, and then you

980
00:47:54,520 --> 00:47:56,399
Speaker 2: adjust the weight so it gets closer to the right answer.

981
00:47:56,400 --> 00:47:58,359
Speaker 2: That's prehen you adjust the waits. What are you doing,

982
00:47:58,760 --> 00:48:02,759
Speaker 2: You're just so you're a training algorithm called back propagation.

983
00:48:02,840 --> 00:48:05,399
Speaker 2: This is a Jeff Hinton, you know innovation where you're

984
00:48:05,440 --> 00:48:07,640
Speaker 2: you're going through and you're adjusting the weights all the

985
00:48:07,640 --> 00:48:09,880
Speaker 2: way through the layers in such a way that the

986
00:48:09,960 --> 00:48:12,839
Speaker 2: answer it gives for this particular test gets a little

987
00:48:12,840 --> 00:48:14,400
Speaker 2: bit closer to the right answer, because you know the

988
00:48:14,440 --> 00:48:17,040
Speaker 2: right answer that's pre training, that's pre trains unsupervised. So

989
00:48:17,080 --> 00:48:19,879
Speaker 2: you take Hamlet or you take Dickens like the best

990
00:48:19,880 --> 00:48:21,680
Speaker 2: of times it was the worst of and then you

991
00:48:21,719 --> 00:48:23,840
Speaker 2: give that to the thing what were to come next?

992
00:48:24,360 --> 00:48:26,640
Speaker 2: And you know it says Bacon like, we're going to

993
00:48:26,719 --> 00:48:28,560
Speaker 2: adjust these weights now in a way that like your

994
00:48:28,560 --> 00:48:31,319
Speaker 2: answer gets a little bit closer towards times. Right. Okay,

995
00:48:31,320 --> 00:48:34,360
Speaker 2: that's pre training. Then you get post training, where you

996
00:48:34,400 --> 00:48:36,319
Speaker 2: already have trained, so you have this network. All the

997
00:48:36,320 --> 00:48:38,960
Speaker 2: weights have been set through this massive multi month you know,

998
00:48:39,160 --> 00:48:42,480
Speaker 2: billion dollar pre training, and now you want to go

999
00:48:42,520 --> 00:48:44,800
Speaker 2: through and you want to tweak this to avoid certain

1000
00:48:44,840 --> 00:48:47,920
Speaker 2: types of behaviors or to influence towards certain types of behaviors.

1001
00:48:47,920 --> 00:48:51,440
Speaker 2: So for post training, it's almost always based off of

1002
00:48:51,520 --> 00:48:55,960
Speaker 2: you have inputs like prompts and correct answers. This is

1003
00:48:56,000 --> 00:48:58,600
Speaker 2: the right way to respond to this question, right, So

1004
00:48:58,760 --> 00:49:02,680
Speaker 2: you've you've you have of questions and answers. You give

1005
00:49:02,680 --> 00:49:05,800
Speaker 2: the prompt to the LLM, it spits out some answer,

1006
00:49:05,960 --> 00:49:08,480
Speaker 2: and now what you're doing you're using is called reinforcement learning.

1007
00:49:08,480 --> 00:49:10,920
Speaker 2: It's a general technology, but you're using techniques from reinforcement

1008
00:49:11,000 --> 00:49:12,719
Speaker 2: learning the sort of like zap it like you would

1009
00:49:12,760 --> 00:49:16,000
Speaker 2: zap a dog when you're dog training it to if

1010
00:49:16,000 --> 00:49:18,239
Speaker 2: it's a bad answer, you zap it, so you get

1011
00:49:18,280 --> 00:49:20,040
Speaker 2: those weights away from that answer, and if it's a

1012
00:49:20,040 --> 00:49:22,360
Speaker 2: good answer, you give it a treat like Okay, this

1013
00:49:22,480 --> 00:49:25,520
Speaker 2: is post this is post training. And so that's where

1014
00:49:26,280 --> 00:49:29,640
Speaker 2: you've moved past the word guessing game, which is where

1015
00:49:29,640 --> 00:49:32,000
Speaker 2: all of the sort of general smarts comes from these models,

1016
00:49:32,040 --> 00:49:35,400
Speaker 2: And now you're you're you're doing the sort of zapping

1017
00:49:35,440 --> 00:49:37,960
Speaker 2: and treat training around very specific things, right, so this

1018
00:49:38,000 --> 00:49:41,759
Speaker 2: is where you keep going. Yeah, So like so you'll

1019
00:49:41,760 --> 00:49:43,840
Speaker 2: go through and like ask it questions where the answers

1020
00:49:43,920 --> 00:49:46,600
Speaker 2: might be like about building bombs or whatever, and every

1021
00:49:46,640 --> 00:49:48,759
Speaker 2: time it spits out an answer about a bomb, you

1022
00:49:48,800 --> 00:49:51,400
Speaker 2: give it like a really bad negative shock, and you

1023
00:49:51,520 --> 00:49:54,319
Speaker 2: like definitely turn off those circuits, like we don't want

1024
00:49:54,360 --> 00:49:56,000
Speaker 2: you to spit out answers by bombs, Like that's where

1025
00:49:56,000 --> 00:49:58,279
Speaker 2: all the guardrails come from. Or if you want it

1026
00:49:58,320 --> 00:50:00,640
Speaker 2: to get better at doing like a particular math exam,

1027
00:50:00,680 --> 00:50:03,279
Speaker 2: you can give it like lots of questions from that

1028
00:50:03,520 --> 00:50:06,040
Speaker 2: math exam and then you have the right answer and

1029
00:50:06,040 --> 00:50:07,880
Speaker 2: you can kind of zap it to move it towards

1030
00:50:08,239 --> 00:50:10,480
Speaker 2: what the answers look like on this math exam. So

1031
00:50:10,480 --> 00:50:14,880
Speaker 2: so post training is more focused. You have particular types

1032
00:50:14,920 --> 00:50:17,360
Speaker 2: of behavior you're trying to sort of instill in this

1033
00:50:17,480 --> 00:50:21,240
Speaker 2: already pre trained massive network. That's mainly where the focus

1034
00:50:21,320 --> 00:50:24,400
Speaker 2: turned after GPT four. So GPT four was like the

1035
00:50:24,480 --> 00:50:28,959
Speaker 2: extent of pre training making it smarter. After GPT four,

1036
00:50:29,080 --> 00:50:32,080
Speaker 2: trying to make those models even bigger and pre train

1037
00:50:32,160 --> 00:50:35,080
Speaker 2: them longer didn't lead to much performance increases. So everything

1038
00:50:35,080 --> 00:50:37,440
Speaker 2: we got between four in the lead up to five

1039
00:50:38,360 --> 00:50:40,680
Speaker 2: was post training, and that's when they began focusing on

1040
00:50:40,800 --> 00:50:43,879
Speaker 2: metrics because if I have a particular metric, I can

1041
00:50:43,960 --> 00:50:46,200
Speaker 2: post train a model to do well on that test.

1042
00:50:46,280 --> 00:50:48,920
Speaker 2: And so everything became about metrics and post now we

1043
00:50:48,920 --> 00:50:51,359
Speaker 2: can do this thing better or look at this thing

1044
00:50:51,440 --> 00:50:53,560
Speaker 2: we do better, And so that's kind of the game

1045
00:50:54,040 --> 00:50:56,719
Speaker 2: that's played now is we do lots of post training

1046
00:50:57,280 --> 00:50:59,879
Speaker 2: that requires like much more specific data because you need

1047
00:50:59,920 --> 00:51:03,600
Speaker 2: like right answers, pairs of props, and correct answers. So

1048
00:51:03,640 --> 00:51:05,719
Speaker 2: only certain things we can do this with. But that's

1049
00:51:05,719 --> 00:51:07,800
Speaker 2: the game we've been playing since like twenty that's one.

1050
00:51:08,400 --> 00:51:10,600
Speaker 1: Do they do that with models that exist now? So

1051
00:51:11,200 --> 00:51:12,799
Speaker 1: this is basically updates, right?

1052
00:51:12,920 --> 00:51:17,319
Speaker 2: Yeah? They do it on a semi regular basis. Yeah,

1053
00:51:17,320 --> 00:51:19,240
Speaker 2: but they usually there'll be a name change, like GPT

1054
00:51:19,320 --> 00:51:21,280
Speaker 2: five to two is different than five to one, different

1055
00:51:21,320 --> 00:51:21,760
Speaker 2: than five.

1056
00:51:21,960 --> 00:51:23,680
Speaker 1: But don't they update the current models?

1057
00:51:24,880 --> 00:51:27,440
Speaker 2: They don't repretrain them. That's too expensive.

1058
00:51:27,480 --> 00:51:29,359
Speaker 1: But they no, no, I'm not saying that. And say

1059
00:51:29,400 --> 00:51:31,560
Speaker 1: do they post train the current models to make them

1060
00:51:31,640 --> 00:51:32,320
Speaker 1: better at stuff?

1061
00:51:32,400 --> 00:51:32,640
Speaker 2: Yeah?

1062
00:51:32,680 --> 00:51:33,440
Speaker 1: To tweak things.

1063
00:51:33,520 --> 00:51:33,920
Speaker 2: Yeah.

1064
00:51:33,960 --> 00:51:36,560
Speaker 1: So that's This is a very long way to get

1065
00:51:36,600 --> 00:51:40,200
Speaker 1: to a point I'm kind of making. Which is one

1066
00:51:40,239 --> 00:51:42,880
Speaker 1: of the problems with VIBE reporting on this is training

1067
00:51:42,960 --> 00:51:46,319
Speaker 1: is framed as this thing like inference is framed as

1068
00:51:46,400 --> 00:51:49,560
Speaker 1: op X that is permanent that you cannot avoid influence

1069
00:51:49,600 --> 00:51:52,520
Speaker 1: being creating the outputs. Training is framed as this R

1070
00:51:52,600 --> 00:51:56,440
Speaker 1: and D mysticism, which is just out there and you're

1071
00:51:56,440 --> 00:51:59,280
Speaker 1: not like training. They never say this, but you hear training,

1072
00:51:59,320 --> 00:52:02,160
Speaker 1: you think, oh, you train and then you perform, and

1073
00:52:02,280 --> 00:52:06,080
Speaker 1: so training would end. But from what I understand, training

1074
00:52:06,160 --> 00:52:09,680
Speaker 1: is as common and necessary an expense as influence at

1075
00:52:09,680 --> 00:52:10,040
Speaker 1: this point.

1076
00:52:10,160 --> 00:52:12,719
Speaker 2: Yeah, it's the only way you improve or update things, right, Like, so,

1077
00:52:12,719 --> 00:52:16,000
Speaker 2: if you had a Microsoft through sixty five software, you're

1078
00:52:16,040 --> 00:52:19,120
Speaker 2: constantly sending updates and patches and whatever is you like

1079
00:52:19,160 --> 00:52:21,959
Speaker 2: to add new features or whatever. In the AI model world,

1080
00:52:22,040 --> 00:52:24,520
Speaker 2: it's it's yeah, it's posting more rounds of post training.

1081
00:52:24,520 --> 00:52:27,440
Speaker 2: It's the only way to make any sort of improvement

1082
00:52:27,600 --> 00:52:30,600
Speaker 2: or fixing bugs. Like you're like, oh, here's something that's

1083
00:52:30,640 --> 00:52:32,799
Speaker 2: saying that we're really upset about. Okay, let's go in

1084
00:52:32,840 --> 00:52:35,160
Speaker 2: and do some zapping. Let's get out to zapper, give

1085
00:52:35,200 --> 00:52:36,920
Speaker 2: it a bunch of examples of saying that bad thing,

1086
00:52:36,960 --> 00:52:38,640
Speaker 2: and let zap and say don't do that. And now

1087
00:52:38,680 --> 00:52:41,640
Speaker 2: it's like very unlikely to do that. So yeah, they're

1088
00:52:41,640 --> 00:52:44,879
Speaker 2: constantly they're constantly doing that. Otherwise, your a stasis, which

1089
00:52:44,920 --> 00:52:47,440
Speaker 2: is different than the way most people actually think of it. Differently,

1090
00:52:48,000 --> 00:52:50,640
Speaker 2: they think that the model is somehow like learning online

1091
00:52:51,239 --> 00:52:54,120
Speaker 2: that like, which as it's absolutely not true that as

1092
00:52:54,160 --> 00:52:57,359
Speaker 2: you talk to it, it's learning and it's getting uh,

1093
00:52:57,600 --> 00:53:00,320
Speaker 2: it's getting better. And they're like, but wait a second,

1094
00:53:00,440 --> 00:53:03,600
Speaker 2: something I talked about earlier, it must be getting smarter.

1095
00:53:03,800 --> 00:53:06,840
Speaker 2: The model doesn't care about you, just your local software.

1096
00:53:06,920 --> 00:53:10,600
Speaker 2: You don't realize this. It's including like huge bits of

1097
00:53:10,600 --> 00:53:12,719
Speaker 2: stuff you've talked to it before. In the prompt that's

1098
00:53:12,760 --> 00:53:14,480
Speaker 2: going to the model. It's like, here's a bunch of

1099
00:53:14,480 --> 00:53:17,480
Speaker 2: stuff that Ed has submitted to you in the past. Okay,

1100
00:53:17,480 --> 00:53:21,080
Speaker 2: now here's his current question. So it's the model hasn't changed.

1101
00:53:21,080 --> 00:53:23,160
Speaker 2: The model doesn't have memory a canunt of justice. It

1102
00:53:23,200 --> 00:53:25,800
Speaker 2: doesn't adjust its memory in real time. It's all static.

1103
00:53:25,840 --> 00:53:29,400
Speaker 2: There is no dynamic memory involved in these models. They're

1104
00:53:29,640 --> 00:53:31,520
Speaker 2: fixed until you go out and post train it and

1105
00:53:31,520 --> 00:53:33,279
Speaker 2: then you bring the new weights back in the data center.

1106
00:53:33,320 --> 00:53:34,640
Speaker 2: And that's the thing. Do an inference now.

1107
00:53:35,360 --> 00:53:37,960
Speaker 1: But that's the thing like this. The reason I bring

1108
00:53:38,000 --> 00:53:41,319
Speaker 1: it up as the kind of closing vibe story is

1109
00:53:41,360 --> 00:53:45,920
Speaker 1: because training is very clearly getting more expensive what they

1110
00:53:45,960 --> 00:53:48,919
Speaker 1: say profitable in inference, which again doesn't really make sense

1111
00:53:48,960 --> 00:53:52,040
Speaker 1: to me. But putting that aside, they're not. But even

1112
00:53:52,080 --> 00:53:56,040
Speaker 1: if they were, if training never stops, then who cares

1113
00:53:56,239 --> 00:53:58,920
Speaker 1: about profitable in inference? Like it just means that you

1114
00:53:59,520 --> 00:54:03,640
Speaker 1: will get more or expensive forever. Inference is just yeah, yeah,

1115
00:54:03,719 --> 00:54:08,239
Speaker 1: differences pooh, training is pooh. That's just that's I'm not

1116
00:54:08,280 --> 00:54:08,960
Speaker 1: putting that one in.

1117
00:54:09,000 --> 00:54:12,000
Speaker 2: A yeah, okay, it's an interesting question, right because it's

1118
00:54:12,000 --> 00:54:14,120
Speaker 2: also getting like the what's an altman who is making

1119
00:54:14,120 --> 00:54:16,680
Speaker 2: those comments about well, we just didn't have to pay

1120
00:54:16,719 --> 00:54:18,839
Speaker 2: for the training. This would be part I.

1121
00:54:18,760 --> 00:54:21,480
Speaker 1: Didn't have all these expenses, I would be so profitable.

1122
00:54:22,040 --> 00:54:24,800
Speaker 2: The one distinction that's maybe relevant there not to be

1123
00:54:24,840 --> 00:54:26,880
Speaker 2: an apologist, but the one one distinction that is relevant

1124
00:54:26,880 --> 00:54:31,480
Speaker 2: is pre training versus posts. So pre training is insanely expensive, right,

1125
00:54:31,600 --> 00:54:35,440
Speaker 2: because you're you're training something on all the all the

1126
00:54:35,480 --> 00:54:38,880
Speaker 2: words in the world, and it takes months. So it's

1127
00:54:38,960 --> 00:54:42,640
Speaker 2: it's just like you're running a data center that's going

1128
00:54:42,719 --> 00:54:47,440
Speaker 2: to have nowadays up to six digit GPUs uh running

1129
00:54:47,440 --> 00:54:50,719
Speaker 2: full time for months just to get that pre training done,

1130
00:54:50,760 --> 00:54:52,440
Speaker 2: and you have to pay for all of that, right,

1131
00:54:52,480 --> 00:54:54,920
Speaker 2: So that's all right time, you're not getting money, you're

1132
00:54:54,920 --> 00:54:57,279
Speaker 2: just paying for training. Post training is also expensive. It's

1133
00:54:57,320 --> 00:55:00,600
Speaker 2: not that expensive though, because as expensive as pre training,

1134
00:55:00,680 --> 00:55:02,759
Speaker 2: because each post training session it's a way, way, way

1135
00:55:02,760 --> 00:55:05,080
Speaker 2: way smaller data set that you're post training on. It's like,

1136
00:55:05,120 --> 00:55:08,240
Speaker 2: all right, we generated like ten thousand examples of people,

1137
00:55:08,440 --> 00:55:11,360
Speaker 2: you know, responding to questions in a racist way, and

1138
00:55:11,400 --> 00:55:15,200
Speaker 2: those ten thousand examples will use to reinforcement learn and

1139
00:55:15,280 --> 00:55:17,200
Speaker 2: try to move it away from answering those type of

1140
00:55:17,239 --> 00:55:19,680
Speaker 2: questions in a racist way. That's like not that big

1141
00:55:19,719 --> 00:55:22,600
Speaker 2: of a data set compared to we're going to train

1142
00:55:22,680 --> 00:55:25,840
Speaker 2: this on every word written that we have access to.

1143
00:55:25,920 --> 00:55:28,520
Speaker 2: So the post training is not as expensive as pre training.

1144
00:55:29,040 --> 00:55:30,520
Speaker 1: Unless you're doing it all the time.

1145
00:55:30,680 --> 00:55:31,600
Speaker 2: Yeah, that's true.

1146
00:55:31,960 --> 00:55:34,080
Speaker 1: That's the thing if you're doing it all the time

1147
00:55:35,000 --> 00:55:38,440
Speaker 1: and it takes months of pre training, but you're constantly

1148
00:55:38,440 --> 00:55:41,359
Speaker 1: doing something like post training for months. Yeah, I mean

1149
00:55:41,360 --> 00:55:43,640
Speaker 1: post training not functionally the same thing.

1150
00:55:43,640 --> 00:55:45,319
Speaker 2: It's the same thing. I think this is a fair point.

1151
00:55:45,360 --> 00:55:47,439
Speaker 2: When you when you have a particular example that you're

1152
00:55:47,440 --> 00:55:50,120
Speaker 2: post training on like one prompt with an answer, it's

1153
00:55:50,200 --> 00:55:52,080
Speaker 2: kind of like you're doing inference in reverse, right, So

1154
00:55:52,120 --> 00:55:55,080
Speaker 2: you're going from your back propagating from one side of

1155
00:55:55,120 --> 00:55:56,839
Speaker 2: the network to the beginning, as a post are going

1156
00:55:56,840 --> 00:55:58,800
Speaker 2: from the beginning to the end. Now, it's more expensive

1157
00:55:58,840 --> 00:56:02,759
Speaker 2: than that because when you're just doing inference, the fundamental

1158
00:56:02,800 --> 00:56:06,920
Speaker 2: operation is basic multiplication. You're just multiplying numbers in the

1159
00:56:06,920 --> 00:56:10,759
Speaker 2: big table backpropagation which you use to training. You're it

1160
00:56:10,800 --> 00:56:12,680
Speaker 2: is multiplication. When you do a bunch of it's a

1161
00:56:12,680 --> 00:56:16,360
Speaker 2: bunch of derivatives because you're constantly you need to calculate

1162
00:56:16,440 --> 00:56:19,879
Speaker 2: like the derivative of these. I mean like it doesn't

1163
00:56:19,920 --> 00:56:21,319
Speaker 2: really matter, but you kind of need that. You want

1164
00:56:21,320 --> 00:56:23,759
Speaker 2: the derivative because you want the gradient to scent to

1165
00:56:23,800 --> 00:56:26,440
Speaker 2: be towards like better and away from worse and derivatives.

1166
00:56:27,560 --> 00:56:31,000
Speaker 2: My understanding is this is like more expensive operations per

1167
00:56:31,239 --> 00:56:33,200
Speaker 2: weight that you're trying to change because you're not just

1168
00:56:33,280 --> 00:56:36,160
Speaker 2: multiplying a number, You're having to calculate derivatives and becomes

1169
00:56:36,160 --> 00:56:38,120
Speaker 2: a little bit more complicated. So yeah, it's like inference

1170
00:56:38,160 --> 00:56:40,719
Speaker 2: in reverse, but also like a little bit more expensive.

1171
00:56:40,800 --> 00:56:45,480
Speaker 2: So if you if you have ten thousand sample question responses,

1172
00:56:45,520 --> 00:56:47,319
Speaker 2: you're going to use the post train, it's kind of

1173
00:56:47,360 --> 00:56:51,600
Speaker 2: like ten thousand users sent prompts and there are particularly

1174
00:56:51,600 --> 00:56:53,520
Speaker 2: expensive prompts, and you had to pay for that instead

1175
00:56:53,520 --> 00:56:55,960
Speaker 2: of like them paying for it. So yeah, it's good

1176
00:56:55,960 --> 00:56:58,440
Speaker 2: to think of it as like inference and reverse.

1177
00:56:59,080 --> 00:57:04,080
Speaker 1: It's also an ongoing cost, Like it's everyone is leaning

1178
00:57:04,080 --> 00:57:07,320
Speaker 1: on this idea that this stuff will magically become profitable.

1179
00:57:07,320 --> 00:57:10,120
Speaker 1: I don't know if you've seen the cash flow diagrams

1180
00:57:10,120 --> 00:57:13,040
Speaker 1: of anthropic and open ai. But there is a mysterious

1181
00:57:13,080 --> 00:57:15,560
Speaker 1: math going on where year twenty twenty eight, twenty twenty

1182
00:57:15,680 --> 00:57:16,960
Speaker 1: nine they just become profitable.

1183
00:57:16,960 --> 00:57:19,560
Speaker 2: Well, I was going to ask you about this. Last month,

1184
00:57:19,920 --> 00:57:23,800
Speaker 2: there's this announcement that open ai had some massive increase

1185
00:57:23,960 --> 00:57:25,600
Speaker 2: in revenue.

1186
00:57:26,800 --> 00:57:29,400
Speaker 1: Yeah, well this is actually a great vibe reporting thing,

1187
00:57:29,440 --> 00:57:32,680
Speaker 1: and that's annualized revenue. They said they hit twenty billion

1188
00:57:32,920 --> 00:57:35,439
Speaker 1: in annualized revenue, which would mean one point sixty seven

1189
00:57:35,480 --> 00:57:39,960
Speaker 1: billion in a month. Now, important details. We don't know

1190
00:57:40,040 --> 00:57:42,800
Speaker 1: how they're defining a month. We don't know if they

1191
00:57:42,840 --> 00:57:45,560
Speaker 1: mean twenty eight days, twenty nine days, thirty days. We

1192
00:57:45,600 --> 00:57:47,920
Speaker 1: don't know if they mean a calendar month, so the

1193
00:57:47,960 --> 00:57:50,240
Speaker 1: month of November or December, or do they just bean

1194
00:57:50,400 --> 00:57:54,080
Speaker 1: any thirty days. We don't know if they're doing insane math,

1195
00:57:54,200 --> 00:57:57,880
Speaker 1: which happens very rarely, that this company feels like one

1196
00:57:57,880 --> 00:58:01,000
Speaker 1: that might do it. Just my gut instinct is they

1197
00:58:01,040 --> 00:58:03,320
Speaker 1: may be doing his seven day period and we're going

1198
00:58:03,360 --> 00:58:05,920
Speaker 1: to turn it into a month like they We don't

1199
00:58:05,920 --> 00:58:07,720
Speaker 1: know how they're doing this. And they also coupled this

1200
00:58:07,800 --> 00:58:11,640
Speaker 1: by saying that as compute grows, revenue grows. I don't

1201
00:58:11,640 --> 00:58:12,280
Speaker 1: know if you seen this.

1202
00:58:13,040 --> 00:58:15,520
Speaker 2: No, it's like, say, their formula, yeah, okay.

1203
00:58:15,400 --> 00:58:18,360
Speaker 1: Yeah, it's an insane formula that does not map to

1204
00:58:18,520 --> 00:58:21,680
Speaker 1: any economics. Like it's just it's the kind of thing

1205
00:58:21,760 --> 00:58:24,120
Speaker 1: that if we had a functioning business and tech press,

1206
00:58:24,880 --> 00:58:28,680
Speaker 1: that would just be scrutinized to the bone, that would

1207
00:58:28,720 --> 00:58:31,000
Speaker 1: just be ripped apart and say, what the fuck does

1208
00:58:31,040 --> 00:58:33,960
Speaker 1: that mean? Because if this were true, if you simply

1209
00:58:34,040 --> 00:58:38,120
Speaker 1: add more gigawattch at more revenue, then you would simply

1210
00:58:38,760 --> 00:58:40,880
Speaker 1: print more money, like it would be a money print

1211
00:58:40,920 --> 00:58:41,360
Speaker 1: of us.

1212
00:58:41,440 --> 00:58:44,240
Speaker 2: Like my movie theater did well. Therefore, yeah, if I

1213
00:58:44,240 --> 00:58:46,200
Speaker 2: build one hundred thousand movie theaters, we're gonna make one

1214
00:58:46,240 --> 00:58:48,040
Speaker 2: hundred thousand more you know, times the amount of money.

1215
00:58:48,080 --> 00:58:49,880
Speaker 2: And it's like, well, wait that theater was in Manhattan

1216
00:58:50,000 --> 00:58:54,680
Speaker 2: and it was like really well run, and they're yeah,

1217
00:58:54,800 --> 00:58:58,520
Speaker 2: well okay, I assume I assumed as much trying to

1218
00:58:58,800 --> 00:59:01,520
Speaker 2: understand that story. Okay, yeah, the annualized revenue stuff, I

1219
00:59:01,560 --> 00:59:03,360
Speaker 2: think I even use that term talking to someone I

1220
00:59:03,360 --> 00:59:05,360
Speaker 2: credit you as like Trum would say, look out for

1221
00:59:05,400 --> 00:59:06,280
Speaker 2: annualized revenue.

1222
00:59:07,160 --> 00:59:09,320
Speaker 1: It is. The funny thing is with that as well,

1223
00:59:09,720 --> 00:59:16,280
Speaker 1: it's more viabe reporting because arr standard. It's very standard

1224
00:59:16,280 --> 00:59:19,680
Speaker 1: in SaaS companies that sell on a per seat basis,

1225
00:59:19,720 --> 00:59:22,680
Speaker 1: So a sales force would do a well even they

1226
00:59:22,720 --> 00:59:25,040
Speaker 1: are doing annualized now with the AI revenue, but it

1227
00:59:25,080 --> 00:59:27,760
Speaker 1: would be a software company has one hundred seats they

1228
00:59:27,760 --> 00:59:29,600
Speaker 1: sell to a company and they charge fifteen bucks a

1229
00:59:29,640 --> 00:59:33,480
Speaker 1: bit a seat per month, and they charge that annually,

1230
00:59:33,880 --> 00:59:37,200
Speaker 1: and the actual cost of each user is fairly measurable

1231
00:59:37,400 --> 00:59:41,000
Speaker 1: because they're doing CPU based stuff. Like it gets expensive

1232
00:59:41,000 --> 00:59:43,080
Speaker 1: at scale because there's a lot of people, but it

1233
00:59:43,120 --> 00:59:47,240
Speaker 1: doesn't get multiply multi I can't say the word. It

1234
00:59:47,280 --> 00:59:50,280
Speaker 1: doesn't get much more expensive as you grow with AI.

1235
00:59:50,720 --> 00:59:53,560
Speaker 1: It's actually because of the way large language models work,

1236
00:59:53,840 --> 00:59:56,640
Speaker 1: your most excited customers are your most expensive.

1237
00:59:56,880 --> 00:59:59,840
Speaker 2: Yeah, because they blow whatever whatever their monthly revenue is,

1238
01:00:00,000 --> 01:00:02,240
Speaker 2: whatever their cost is, they blow past it, so.

1239
01:00:02,360 --> 01:00:06,320
Speaker 1: You can't even do a percy revenue. It's just every

1240
01:00:06,400 --> 01:00:08,560
Speaker 1: all of these things. When I say them out loud,

1241
01:00:08,600 --> 01:00:11,280
Speaker 1: I'm like, I feel like this should be more obvious.

1242
01:00:11,280 --> 01:00:12,680
Speaker 1: It's why I loved your video because it was like

1243
01:00:12,720 --> 01:00:15,240
Speaker 1: thank you someone else, Well, here's the.

1244
01:00:15,280 --> 01:00:17,320
Speaker 2: Question, here's my here's my question, here's like the dangerous

1245
01:00:17,360 --> 01:00:19,160
Speaker 2: question I actually put out a video last week. It

1246
01:00:19,200 --> 01:00:24,080
Speaker 2: was like dangerous question. We're now on year three or

1247
01:00:24,120 --> 01:00:27,640
Speaker 2: four of like the I'm counting new year starting like

1248
01:00:27,640 --> 01:00:31,040
Speaker 2: New Year twenty twenty three or whatever of people saying,

1249
01:00:31,800 --> 01:00:35,200
Speaker 2: oh my god, this is so cool. These massive disruptions

1250
01:00:35,200 --> 01:00:38,439
Speaker 2: they're going to change everything is imminent, and year after

1251
01:00:38,520 --> 01:00:42,600
Speaker 2: year we've said that I'm not yet seeing the massive disruptions,

1252
01:00:43,720 --> 01:00:45,720
Speaker 2: like not the not the stories of what might be

1253
01:00:45,760 --> 01:00:48,360
Speaker 2: disrupted or the stories of what's different, but like, how

1254
01:00:48,400 --> 01:00:53,840
Speaker 2: many years do we have to go without industries crumbling

1255
01:00:54,400 --> 01:00:58,720
Speaker 2: or major new economic players that didn't exist before, or

1256
01:00:58,800 --> 01:01:02,680
Speaker 2: complete restructuring of huge companies around this technology? How many

1257
01:01:02,760 --> 01:01:05,160
Speaker 2: years we have to go? So I did a video,

1258
01:01:05,240 --> 01:01:07,160
Speaker 2: right I went, I found the Reddit thread where someone

1259
01:01:07,200 --> 01:01:08,880
Speaker 2: just asked this question. This was from like earlier in

1260
01:01:08,880 --> 01:01:12,080
Speaker 2: the month. They were like, outside of the vibe coding stuff,

1261
01:01:13,880 --> 01:01:16,240
Speaker 2: what what are the like, what are people? What are

1262
01:01:16,240 --> 01:01:18,200
Speaker 2: the big tools? Like, what are the big things that

1263
01:01:18,200 --> 01:01:20,240
Speaker 2: have come out of this technology? They're changing things? And

1264
01:01:20,280 --> 01:01:22,240
Speaker 2: I read through this whole thread and it was interesting,

1265
01:01:22,280 --> 01:01:24,160
Speaker 2: there's not people don't have much like well, you know,

1266
01:01:24,280 --> 01:01:26,960
Speaker 2: like it could these are like really small case studies.

1267
01:01:27,400 --> 01:01:30,080
Speaker 2: I used it to help, you know, gather clean up

1268
01:01:30,160 --> 01:01:32,840
Speaker 2: data that I got from whatever. I was like, this

1269
01:01:32,880 --> 01:01:35,480
Speaker 2: is like such a nerdy specific use case, and so

1270
01:01:35,560 --> 01:01:37,640
Speaker 2: I went through that thread in a video and this

1271
01:01:37,680 --> 01:01:39,240
Speaker 2: has kind of been my question. It's like, it is

1272
01:01:39,360 --> 01:01:41,480
Speaker 2: very cool technology, but how do we know where this

1273
01:01:41,520 --> 01:01:44,440
Speaker 2: is going to fall? Like to me, the scale is

1274
01:01:44,600 --> 01:01:50,880
Speaker 2: it would go like this, like blockchain software, then Oculus VR,

1275
01:01:51,680 --> 01:01:55,240
Speaker 2: then maybe internet to electricity. Right, so we're gonna have

1276
01:01:55,280 --> 01:01:56,200
Speaker 2: a scale of disruption.

1277
01:01:56,320 --> 01:01:56,400
Speaker 1: Right.

1278
01:01:56,480 --> 01:02:00,400
Speaker 2: Blockchain software is something where the premise it's made no

1279
01:02:00,520 --> 01:02:02,240
Speaker 2: sense and it was never going to get off the

1280
01:02:02,240 --> 01:02:04,680
Speaker 2: ground and there was going to be no impact on

1281
01:02:04,720 --> 01:02:07,400
Speaker 2: the world. And you know, because I'm a my training

1282
01:02:07,400 --> 01:02:09,560
Speaker 2: in CS is in distributed system theory. I was there

1283
01:02:09,560 --> 01:02:11,240
Speaker 2: in twenty twenty saying guys, let me just tell you

1284
01:02:11,600 --> 01:02:13,520
Speaker 2: this is nonsense. No Web three is not about to

1285
01:02:13,520 --> 01:02:15,000
Speaker 2: take off. None of this makes sense. And that was true.

1286
01:02:15,040 --> 01:02:18,840
Speaker 2: That did nothing. Then you have like Oculus VR. It

1287
01:02:18,920 --> 01:02:21,720
Speaker 2: really is cool, right you put on these things like

1288
01:02:21,760 --> 01:02:24,640
Speaker 2: that is awesome, Like I love this technology, but it's

1289
01:02:24,680 --> 01:02:26,600
Speaker 2: having a hard time having any real major impact because

1290
01:02:26,600 --> 01:02:27,280
Speaker 2: people aren't sure what.

1291
01:02:27,320 --> 01:02:31,440
Speaker 1: Times use of it. Also, most people don't necessarily have

1292
01:02:31,520 --> 01:02:35,520
Speaker 1: a great experience initially because it's extremely dependent on where

1293
01:02:35,560 --> 01:02:37,520
Speaker 1: you are, who you are, the size of your skull

1294
01:02:37,640 --> 01:02:38,840
Speaker 1: not kind of yeah so.

1295
01:02:38,960 --> 01:02:41,880
Speaker 2: Big right for a limited group of people, it's really cool,

1296
01:02:41,880 --> 01:02:43,960
Speaker 2: but it fails to come out. Then the Internet is

1297
01:02:44,040 --> 01:02:47,160
Speaker 2: like really disruptive, changed a lot of things more. It's

1298
01:02:47,160 --> 01:02:49,480
Speaker 2: not so much as like whole industries disappeared or whatever,

1299
01:02:49,560 --> 01:02:51,160
Speaker 2: but it like changed the way a lot of industries

1300
01:02:51,200 --> 01:02:53,160
Speaker 2: actually function. And then like electricity, you could say, like

1301
01:02:53,160 --> 01:02:56,360
Speaker 2: it just completely changed what day to day existence of.

1302
01:02:56,360 --> 01:02:59,000
Speaker 1: Business was, like how existence of exactly?

1303
01:02:59,080 --> 01:03:01,200
Speaker 2: Yeah, And so that the question like that everyone should

1304
01:03:01,200 --> 01:03:02,960
Speaker 2: be asking is where is the ENERVEI going to fall

1305
01:03:03,080 --> 01:03:06,400
Speaker 2: on this? And you know, I would say, right now,

1306
01:03:06,440 --> 01:03:08,040
Speaker 2: this is what gives me yelled at. I'm not saying

1307
01:03:08,040 --> 01:03:10,280
Speaker 2: this is the prediction necessarily going forward, but right now

1308
01:03:10,360 --> 01:03:12,760
Speaker 2: I don't think it's got passed much farther past the

1309
01:03:13,200 --> 01:03:16,760
Speaker 2: oculus part of that scale where it's really cool. There's

1310
01:03:16,960 --> 01:03:19,920
Speaker 2: very cool things. Chat Gipt is very cool. It's very

1311
01:03:19,960 --> 01:03:22,200
Speaker 2: cool that it can like have that comprehension and no

1312
01:03:22,240 --> 01:03:24,600
Speaker 2: one thought it could do that, But we haven't yet

1313
01:03:24,600 --> 01:03:25,160
Speaker 2: figured out what.

1314
01:03:27,400 --> 01:03:31,240
Speaker 1: Comprehension. It's no, it's a text I mean.

1315
01:03:31,440 --> 01:03:33,280
Speaker 2: We take it for granted, but for CS people, the

1316
01:03:33,320 --> 01:03:35,480
Speaker 2: ability that like I can add, hey, give me text

1317
01:03:35,680 --> 01:03:38,160
Speaker 2: that like whatever in the style of a poem that

1318
01:03:38,200 --> 01:03:42,480
Speaker 2: does whatever, and that includes a character from Star Wars,

1319
01:03:42,640 --> 01:03:44,280
Speaker 2: and then it can give you text it does that.

1320
01:03:44,280 --> 01:03:46,680
Speaker 2: That comprehension, like for computer scientist was like, oh, we

1321
01:03:46,720 --> 01:03:49,520
Speaker 2: didn't really know how to consistently one yeah. Yeah, that's

1322
01:03:49,520 --> 01:03:53,280
Speaker 2: like very cool. Yeah, but we haven't got past this

1323
01:03:53,440 --> 01:03:56,240
Speaker 2: is like the the surprising thing of this field is

1324
01:03:57,120 --> 01:03:59,760
Speaker 2: we're not really past the oculus stage yet, like where

1325
01:03:59,800 --> 01:04:02,320
Speaker 2: they're like, for certain this is vibe coding is really cool.

1326
01:04:02,600 --> 01:04:05,240
Speaker 2: The comprehension is cool, so like sora is weird, but

1327
01:04:05,280 --> 01:04:07,880
Speaker 2: like it's cool, you can do that, But the markets

1328
01:04:07,920 --> 01:04:10,920
Speaker 2: are not None of these have markets yet, right, Like that,

1329
01:04:10,960 --> 01:04:13,600
Speaker 2: there's not big markets in any of these yet. And

1330
01:04:13,640 --> 01:04:16,000
Speaker 2: will how far will it go from oculus to the Internet.

1331
01:04:16,040 --> 01:04:18,400
Speaker 2: To me, that is like the number one question, the

1332
01:04:18,440 --> 01:04:20,520
Speaker 2: number two question, the number three question of all reporting

1333
01:04:20,560 --> 01:04:23,600
Speaker 2: on this, and almost no one's talking about that. It's

1334
01:04:23,680 --> 01:04:28,280
Speaker 2: just hype laundering. We'll take this hype, will extrapolate, it

1335
01:04:28,280 --> 01:04:31,160
Speaker 2: will react to that extrapolation. That's kind of what reporting is, right,

1336
01:04:31,200 --> 01:04:33,919
Speaker 2: now in AI, Whereas to me, this is the hugest question.

1337
01:04:33,960 --> 01:04:36,600
Speaker 2: If this ends up oculus, retail investors are going to

1338
01:04:36,640 --> 01:04:39,400
Speaker 2: get screwed. If it ends up Internet, all right, that's like,

1339
01:04:39,640 --> 01:04:42,120
Speaker 2: that's like a really interesting significant story. If it ends

1340
01:04:42,160 --> 01:04:44,640
Speaker 2: up electricity, obviously that really matters. But like, I don't

1341
01:04:44,680 --> 01:04:47,240
Speaker 2: know anyone who actually thinks it's going to be that disruptive.

1342
01:04:47,320 --> 01:04:51,000
Speaker 2: Not the current technology. That's the story to me. Not

1343
01:04:52,120 --> 01:04:54,840
Speaker 2: let's like extrapolate us. You know, hey, what are these

1344
01:04:54,840 --> 01:04:57,080
Speaker 2: things are creating a church? Or let's let's let's hype

1345
01:04:57,120 --> 01:05:01,240
Speaker 2: launder extrapolate that and react to our extrapol That's not

1346
01:05:01,280 --> 01:05:04,920
Speaker 2: really reporting so much as speculative fiction writing. I guess

1347
01:05:04,960 --> 01:05:06,560
Speaker 2: I don't know quite what to call it. But this

1348
01:05:06,680 --> 01:05:10,560
Speaker 2: is the real question, where exactly are we now and

1349
01:05:10,600 --> 01:05:12,600
Speaker 2: what are the possibilities of like where this is going

1350
01:05:12,680 --> 01:05:14,560
Speaker 2: to go positive and negative? I don't we have enough

1351
01:05:14,560 --> 01:05:14,960
Speaker 2: talk on that.

1352
01:05:16,280 --> 01:05:18,520
Speaker 1: I fully agree, Cal, it's been such a pleasure in

1353
01:05:18,560 --> 01:05:19,840
Speaker 1: having you. Thank you for joining me.

1354
01:05:20,120 --> 01:05:22,640
Speaker 2: Always happy to talk shop, always happy to hate with you.

1355
01:05:22,720 --> 01:05:25,000
Speaker 2: I guess in your season, I like.

1356
01:05:24,960 --> 01:05:27,600
Speaker 1: It, Hey, season is the best. We will be back

1357
01:05:27,640 --> 01:05:29,800
Speaker 1: this week with either a monologu or an interview. I

1358
01:05:29,840 --> 01:05:34,040
Speaker 1: have not decided, because I've got wonderful Corey Quinn interview

1359
01:05:34,080 --> 01:05:36,160
Speaker 1: I just did, so I'm considering putting that in a monologue.

1360
01:05:36,160 --> 01:05:38,600
Speaker 1: You'll find out on Friday. Anyway, this has been Better

1361
01:05:38,600 --> 01:05:42,800
Speaker 1: off Line. I'm at zychron Subscribe to the premium, download

1362
01:05:42,840 --> 01:05:53,760
Speaker 1: a T shirt, whatever you desire. Thank you for listening

1363
01:05:53,800 --> 01:05:56,440
Speaker 1: to Better Offline. The editor and composer of the Better

1364
01:05:56,440 --> 01:05:59,440
Speaker 1: Offline theme song is Matosowski. You can check out more

1365
01:05:59,480 --> 01:06:03,000
Speaker 1: of his music and audio projects at Mattasowski dot com,

1366
01:06:03,120 --> 01:06:06,640
Speaker 1: m A T T O S O W s ki

1367
01:06:06,840 --> 01:06:09,760
Speaker 1: dot com. You can email me at easy at Better

1368
01:06:09,760 --> 01:06:12,200
Speaker 1: Offline dot com or visit Better Offline dot com to

1369
01:06:12,240 --> 01:06:15,080
Speaker 1: find more podcast links and of course, my newsletter. I

1370
01:06:15,120 --> 01:06:17,840
Speaker 1: also really recommend you go to chat dot Where's youreaed

1371
01:06:17,920 --> 01:06:20,160
Speaker 1: dot at to visit the discord, and go to our

1372
01:06:20,240 --> 01:06:23,560
Speaker 1: slash Better Offline to check out our reddit. Thank you

1373
01:06:23,600 --> 01:06:24,560
Speaker 1: so much for listening.

1374
01:06:25,400 --> 01:06:27,840
Speaker 2: Better Offline is a production of cool Zone Media.

1375
01:06:28,040 --> 01:06:30,880
Speaker 1: For more from cool Zone Media, visit our website cool

1376
01:06:30,920 --> 01:06:33,280
Speaker 1: Zonemedia dot com, or check us out on the

1377
01:06:33,360 --> 01:06:37,040
Speaker 2: iHeartRadio app, Apple Podcasts, or wherever you get your podcasts.