1
00:00:19,612 --> 00:00:21,572
S1: All right. What I'm going to do today is we're

2
00:00:21,572 --> 00:00:24,652
S1: going to have a debate between myself and Marcus Hutchins.

3
00:00:24,892 --> 00:00:26,932
S1: If you don't know Marcus, he is probably the most

4
00:00:26,932 --> 00:00:30,932
S1: famous malware researcher in the world. He's been on the

5
00:00:30,932 --> 00:00:35,652
S1: front page of wired, and he has a dramatically different

6
00:00:35,652 --> 00:00:38,252
S1: view of AI than I do. So I am in

7
00:00:38,292 --> 00:00:41,012
S1: sort of the pro AI camp and believing that it

8
00:00:41,012 --> 00:00:46,412
S1: is net positive and that it is extremely useful today,

9
00:00:46,652 --> 00:00:50,251
S1: and also getting more useful and most importantly, that it's

10
00:00:50,251 --> 00:00:52,852
S1: going to start affecting jobs in a very serious way.

11
00:00:53,292 --> 00:00:58,092
S1: Marcus has called it basically, autocomplete does not believe that

12
00:00:58,092 --> 00:01:03,212
S1: it's intelligence, and believes that the impact is dramatically overplayed,

13
00:01:03,212 --> 00:01:06,412
S1: mostly by people trying to scam people and make money.

14
00:01:06,892 --> 00:01:11,132
S1: So again, we are on pretty much opposite sides of this,

15
00:01:11,572 --> 00:01:15,252
S1: but we are also both security people and, you know,

16
00:01:15,492 --> 00:01:19,422
S1: fairly well known in the security space, and we've both

17
00:01:19,462 --> 00:01:23,661
S1: done a lot and accomplished a lot. And this is

18
00:01:23,661 --> 00:01:26,581
S1: really showing that you can have two people who are

19
00:01:26,581 --> 00:01:29,942
S1: in kind of the same space, can have dramatically different opinions.

20
00:01:30,262 --> 00:01:33,101
S1: And what turned out to be fantastic here, because Marcus

21
00:01:33,102 --> 00:01:36,062
S1: and I have been friends for some amount of time now,

22
00:01:36,581 --> 00:01:39,222
S1: and we had this debate previously in text and it

23
00:01:39,222 --> 00:01:41,381
S1: was fine in text. I was a little worried that

24
00:01:41,862 --> 00:01:44,461
S1: we're both a bit snarky online, and I was worried

25
00:01:44,462 --> 00:01:46,542
S1: that we were going to be snarky to each other

26
00:01:46,542 --> 00:01:49,262
S1: in this debate that turned out not to happen. It

27
00:01:49,262 --> 00:01:52,222
S1: was a very positive experience, and I think we made

28
00:01:52,222 --> 00:01:55,102
S1: our points. Well, not only that, but I think I

29
00:01:55,102 --> 00:01:57,181
S1: made a number of points that he really enjoyed. He

30
00:01:57,182 --> 00:01:59,822
S1: made a number of points that I really enjoyed. I

31
00:01:59,822 --> 00:02:01,981
S1: feel like we both learned something from this, and we've

32
00:02:01,982 --> 00:02:04,822
S1: already committed to doing a follow up at some point

33
00:02:04,822 --> 00:02:08,822
S1: in the future, but this one is hopefully going to

34
00:02:08,822 --> 00:02:11,542
S1: be useful to anybody who is kind of agnostic or

35
00:02:11,542 --> 00:02:15,512
S1: really just kind of wants to hear strong points being

36
00:02:15,512 --> 00:02:18,672
S1: made on both sides of the conversation and also in

37
00:02:18,672 --> 00:02:21,232
S1: a civil way. So that's what we're going to do here.

38
00:02:21,512 --> 00:02:24,032
S1: I do apologize a little bit for the audio quality.

39
00:02:24,072 --> 00:02:26,352
S1: We actually did this on zoom, and we should have

40
00:02:26,352 --> 00:02:29,791
S1: done it on Riverside because you can't actually do 4K

41
00:02:30,392 --> 00:02:34,112
S1: in zoom yet. So anyway, the next one is going

42
00:02:34,112 --> 00:02:37,352
S1: to be higher quality. But um, let's jump into it.

43
00:02:37,352 --> 00:02:40,792
S1: This is a debate between myself and Marcus Hutchins on AI.

44
00:02:41,151 --> 00:02:43,592
S1: I think it's wrong to call what is happening now

45
00:02:43,591 --> 00:02:48,792
S1: with AI hype and to tell people not to worry. Um,

46
00:02:48,912 --> 00:02:51,312
S1: because I think it is a significant threat to people's

47
00:02:51,312 --> 00:02:58,112
S1: jobs and, um, overall, just like a massive thing, uh, in,

48
00:02:58,112 --> 00:02:59,872
S1: in much more of a way than crypto was, I

49
00:02:59,871 --> 00:03:03,112
S1: think crypto was a giant. Nothing. Um, I wasn't excited

50
00:03:03,112 --> 00:03:05,192
S1: about crypto at the time. I didn't see why why

51
00:03:05,192 --> 00:03:08,232
S1: it was like everyone was taking it so seriously. But

52
00:03:08,232 --> 00:03:11,552
S1: I see this as very different. Um, and so what?

53
00:03:11,552 --> 00:03:14,722
S1: I see people like yourself or someone like another person.

54
00:03:14,722 --> 00:03:17,362
S1: I have a lot of respect for, uh, Christopher Hoff,

55
00:03:17,962 --> 00:03:20,841
S1: who I've been friends with for a very long time, actually,

56
00:03:20,882 --> 00:03:24,522
S1: for him, like 15 years or something. 20 years. And

57
00:03:24,522 --> 00:03:28,442
S1: he's frequently putting out similar stuff to you, whereas like, look,

58
00:03:28,442 --> 00:03:30,522
S1: what are you guys talking about? This thing is like,

59
00:03:30,562 --> 00:03:32,681
S1: not anything. It's going to be just like crypto. It's

60
00:03:32,681 --> 00:03:37,282
S1: going to be gone soon. It's massively overhyped. So, um,

61
00:03:37,322 --> 00:03:39,722
S1: when I see you posting those things, I haven't responded

62
00:03:39,722 --> 00:03:42,162
S1: to you and I'm sure you're, like, itching to respond

63
00:03:42,162 --> 00:03:44,282
S1: to me as well. Uh, we haven't done that because

64
00:03:44,282 --> 00:03:46,321
S1: we planned on doing this. And this is a better,

65
00:03:46,362 --> 00:03:49,362
S1: I think, a better way to start that off. But, um,

66
00:03:49,362 --> 00:03:51,482
S1: I would respond to Chris Hoff and he would respond

67
00:03:51,482 --> 00:03:53,842
S1: back and it would get kind of like adversarial, snarky

68
00:03:53,842 --> 00:03:58,562
S1: or whatever. Um, but I mean, all that to say

69
00:03:58,562 --> 00:04:03,122
S1: that I, I think that position is steering people in a,

70
00:04:03,122 --> 00:04:05,522
S1: in a bad direction. It's telling people not to worry.

71
00:04:05,802 --> 00:04:09,041
S1: And I'm not saying like panic. I'm saying this thing

72
00:04:09,042 --> 00:04:10,442
S1: is real and we have to get ready for it.

73
00:04:10,442 --> 00:04:12,932
S1: I think it's wrong to call what is happening now

74
00:04:12,932 --> 00:04:18,132
S1: with AI hype and to tell people not to worry. Um,

75
00:04:18,172 --> 00:04:19,492
S1: because I think it is.

76
00:04:19,532 --> 00:04:23,972
S2: But when you say it's staring people in the direction of, uh,

77
00:04:24,212 --> 00:04:26,612
S2: I guess the, for a lack of a better term

78
00:04:26,612 --> 00:04:30,292
S2: is harmful. What do you think the harm is there?

79
00:04:30,572 --> 00:04:33,732
S2: So if I say now, this isn't my personal position,

80
00:04:33,731 --> 00:04:36,732
S2: but let's say I say AI is completely useless. Like

81
00:04:36,972 --> 00:04:41,851
S2: beyond useless. It does nothing of value whatsoever. What is

82
00:04:41,851 --> 00:04:44,652
S2: inherently bad about people believing that.

83
00:04:46,372 --> 00:04:49,252
S1: Uh, they will not, uh, think about their own skills

84
00:04:49,252 --> 00:04:53,051
S1: and their own capabilities and their own value and how

85
00:04:53,052 --> 00:04:56,292
S1: that is potentially going to be impacted by AI if

86
00:04:56,291 --> 00:04:58,731
S1: they don't see AI as a threat because it's a

87
00:04:58,731 --> 00:05:01,812
S1: nothing burger, and it's basically all hype and marketing and

88
00:05:01,812 --> 00:05:04,372
S1: a bunch of rich people trying to get rich, uh,

89
00:05:04,412 --> 00:05:08,252
S1: even more rich, then they will ignore it. So to

90
00:05:08,291 --> 00:05:11,822
S1: answer your question. The threat, the risk of that is

91
00:05:11,862 --> 00:05:15,222
S1: them ignoring something that is potentially harmful to their career.

92
00:05:16,862 --> 00:05:19,862
S2: But I think you could make that argument for, uh,

93
00:05:19,902 --> 00:05:23,742
S2: we use the the example of blockchain or crypto specifically.

94
00:05:24,142 --> 00:05:26,702
S2: You could have made that same argument for crypto. We

95
00:05:26,702 --> 00:05:29,942
S2: could have said, uh, this is going to be this

96
00:05:29,942 --> 00:05:33,541
S2: future big, impactful technology, which is something many people did say.

97
00:05:33,541 --> 00:05:36,622
S2: They were like, everything is going to be blockchain, crypto.

98
00:05:36,862 --> 00:05:39,062
S2: There were people arguing that Doge was going to replace

99
00:05:39,062 --> 00:05:41,862
S2: the US currency. And you could have made the argument

100
00:05:41,862 --> 00:05:45,222
S2: back then that if you don't go out right now

101
00:05:45,222 --> 00:05:47,822
S2: and spend hours and hours of your time learning blockchain,

102
00:05:48,142 --> 00:05:50,541
S2: you're going to be unemployed. You're not going to be

103
00:05:50,582 --> 00:05:53,822
S2: able to, uh, make it in the future. So we're

104
00:05:53,862 --> 00:05:57,061
S2: sort of trying to come up with this reason for

105
00:05:57,342 --> 00:06:00,182
S2: why AI is different from blockchain. Now, I'm not saying

106
00:06:00,182 --> 00:06:03,582
S2: it's similar. I, I think most people will agree large

107
00:06:03,582 --> 00:06:08,462
S2: language models have infinitely more usefulness than blockchain did. But

108
00:06:08,462 --> 00:06:12,262
S2: then it still comes down to what are we thinking

109
00:06:12,262 --> 00:06:15,262
S2: that where do we think that this technology is going

110
00:06:15,502 --> 00:06:18,262
S2: that is going to have such a big impact that

111
00:06:18,262 --> 00:06:21,062
S2: people need to be paying as much attention as possible?

112
00:06:21,101 --> 00:06:23,222
S2: And like, what even is the level of attention people

113
00:06:23,222 --> 00:06:25,702
S2: need to pay? Like, do I just need to know, okay,

114
00:06:25,742 --> 00:06:27,861
S2: ChatGPT is a thing and here's what it can do.

115
00:06:28,101 --> 00:06:29,702
S2: Or do I need to spend hours and hours of

116
00:06:29,702 --> 00:06:32,662
S2: my day learning how to use ChatGPT in order to

117
00:06:32,702 --> 00:06:35,382
S2: avert whatever problem we're going to say is going to

118
00:06:35,382 --> 00:06:37,022
S2: be the thing in the future.

119
00:06:38,782 --> 00:06:44,301
S1: Yeah, sure. I mean. Yeah. Well, first of all, I mean,

120
00:06:44,342 --> 00:06:47,422
S1: ChatGPT is just one app from one group, but I

121
00:06:47,422 --> 00:06:49,182
S1: think I think I see what you mean, just AI

122
00:06:49,182 --> 00:06:52,822
S1: in general. Um, no, I mean, what I advocate is

123
00:06:52,822 --> 00:06:56,982
S1: people focus more on first principles thinking and like, understanding

124
00:06:56,981 --> 00:07:00,822
S1: their own opinions and forming their own opinions. Uh, because

125
00:07:00,822 --> 00:07:02,342
S1: what I'm trying to get to is like, what comes

126
00:07:02,342 --> 00:07:04,582
S1: after the job replacement, which is like more human to

127
00:07:04,622 --> 00:07:09,752
S1: human interaction. So I think people need to be more, um,

128
00:07:09,792 --> 00:07:12,672
S1: thoughtful internally about who they are and what they are

129
00:07:12,712 --> 00:07:16,752
S1: and like, what they're about actually having opinions. I think

130
00:07:16,752 --> 00:07:19,272
S1: too many people who are in danger of being replaced

131
00:07:19,272 --> 00:07:24,632
S1: by this technology, um, simply had a job executing tasks

132
00:07:24,632 --> 00:07:28,352
S1: that someone else gave them, and that didn't require that

133
00:07:28,352 --> 00:07:32,072
S1: much intellect, but it required enough intelligence. And this is

134
00:07:32,072 --> 00:07:36,192
S1: another point we can debate. Like, is AI actually intelligence? Um,

135
00:07:36,632 --> 00:07:40,432
S1: but it has enough intelligence for all these hundreds of

136
00:07:40,432 --> 00:07:44,752
S1: millions of knowledge worker jobs. It has enough intelligence required

137
00:07:44,952 --> 00:07:47,272
S1: that you couldn't automate it. So which is why they

138
00:07:47,272 --> 00:07:51,232
S1: have the job today, and that if the bar is

139
00:07:51,272 --> 00:07:54,152
S1: getting to your question, if the bar for AI raises

140
00:07:54,192 --> 00:07:57,792
S1: above that level. So we could do basic intelligence tasks,

141
00:07:58,112 --> 00:08:02,472
S1: which is, you know, 90% or 99% of knowledge worker, uh,

142
00:08:02,562 --> 00:08:06,722
S1: Capabilities is included in that. Then those jobs go away.

143
00:08:06,762 --> 00:08:09,522
S1: Hundreds of millions. I don't know the actual numbers for

144
00:08:09,522 --> 00:08:13,522
S1: knowledge worker total employment, but I would say that vast,

145
00:08:13,522 --> 00:08:16,842
S1: vast majority of those jobs go away and they are

146
00:08:16,842 --> 00:08:19,602
S1: left not doing anything. And it's not like they could

147
00:08:19,602 --> 00:08:24,642
S1: just pivot to the next thing, like previous technology trends, because, um, those,

148
00:08:24,842 --> 00:08:28,122
S1: those pivots that they will make likely are also taken

149
00:08:28,122 --> 00:08:29,202
S1: over by that.

150
00:08:29,202 --> 00:08:34,282
S2: I, I understand that line of thinking of, okay, it's

151
00:08:34,282 --> 00:08:38,042
S2: not like the printing press or, uh, early machine learning

152
00:08:38,042 --> 00:08:40,762
S2: where I can just go into something that is not

153
00:08:41,282 --> 00:08:44,562
S2: covered by the current models. But I think the idea

154
00:08:44,562 --> 00:08:48,202
S2: of this just sweeping replacement, where llms are just going

155
00:08:48,242 --> 00:08:52,002
S2: to take out 90% of the workforce, I don't personally

156
00:08:52,002 --> 00:08:55,641
S2: believe that that is rooted in reality. I think, uh,

157
00:08:55,642 --> 00:09:00,482
S2: we're basically extrapolating with, uh, it's almost like a new car, right?

158
00:09:00,522 --> 00:09:03,651
S2: If I, If I've just invented the car, it only

159
00:09:03,652 --> 00:09:07,252
S2: goes 15mph. And then if I do some tweaks to it,

160
00:09:07,252 --> 00:09:10,132
S2: I spend like a couple of months doing some tweaks.

161
00:09:10,131 --> 00:09:12,372
S2: It goes a little bit faster. And now we could

162
00:09:12,372 --> 00:09:14,972
S2: say that if I doubled the speed in three months

163
00:09:15,292 --> 00:09:17,732
S2: than in six months, I can double the speed again.

164
00:09:17,892 --> 00:09:20,651
S2: And then you can just keep extrapolating, extrapolating, extrapolating. So

165
00:09:20,692 --> 00:09:25,372
S2: at some point this car will break the speed of light. Um,

166
00:09:25,412 --> 00:09:28,052
S2: whereas I think that's what we've been seeing with Llms

167
00:09:28,052 --> 00:09:30,572
S2: is the LLM technology was very new and it had

168
00:09:30,572 --> 00:09:33,292
S2: a lot of issues in the beginning. And that gave

169
00:09:33,292 --> 00:09:35,771
S2: the illusion of a lot of very fast progress. We

170
00:09:35,772 --> 00:09:39,131
S2: weren't actually seeing these models rapidly getting like better and

171
00:09:39,131 --> 00:09:41,771
S2: better and better. All that was happening is we made

172
00:09:41,772 --> 00:09:44,652
S2: a new technology that was a little bit, uh, like finicky.

173
00:09:44,692 --> 00:09:47,532
S2: It was it had problems. And we slowly ironed out

174
00:09:47,532 --> 00:09:49,972
S2: some of those problems, which made it look like it

175
00:09:49,972 --> 00:09:53,252
S2: was advancing at this rapid pace. But it seems now

176
00:09:53,292 --> 00:09:56,012
S2: that it is sort of capped out. Um, I would

177
00:09:56,011 --> 00:09:59,212
S2: say the leap from GPT two to GPT three and

178
00:09:59,302 --> 00:10:03,582
S2: 3.5 was pretty groundbreaking. But then as I look at

179
00:10:03,582 --> 00:10:06,782
S2: the newer models, when we go from 3.5 to whatever

180
00:10:06,782 --> 00:10:10,542
S2: they've called it now, like 001, I don't even know

181
00:10:10,542 --> 00:10:13,502
S2: what the naming convention is anymore. Um, we haven't really

182
00:10:13,502 --> 00:10:17,902
S2: seen any improvements. We're just seeing very slight increases in, oh,

183
00:10:17,942 --> 00:10:20,102
S2: it's better at coding now. It's a little bit better

184
00:10:20,102 --> 00:10:22,542
S2: at math. And what it looks like is happening is

185
00:10:22,582 --> 00:10:26,421
S2: they've made the base large language model, and they've refined

186
00:10:26,422 --> 00:10:29,822
S2: that almost as much as they can. And then they're

187
00:10:29,822 --> 00:10:32,702
S2: building these sub models to then make up the slack

188
00:10:32,702 --> 00:10:35,862
S2: where it can't do math or it can't do programming

189
00:10:36,102 --> 00:10:39,622
S2: or whatever other areas it struggles. So I think what

190
00:10:39,622 --> 00:10:42,182
S2: we're actually going to see is we're going to see

191
00:10:42,502 --> 00:10:45,022
S2: large language models as a technology just cap out like

192
00:10:45,022 --> 00:10:46,942
S2: they're just going to hit a wall. And then what

193
00:10:46,942 --> 00:10:49,382
S2: we're going to see is them start building out sub

194
00:10:49,381 --> 00:10:54,102
S2: models that make them better at a specific job. Now

195
00:10:54,142 --> 00:10:57,872
S2: that would just be normal technological evolution like I as

196
00:10:57,912 --> 00:11:01,631
S2: a malware researcher, I got into malware analysis when it

197
00:11:01,631 --> 00:11:05,232
S2: was already machine learning automated. There was still enough space

198
00:11:05,232 --> 00:11:08,712
S2: in my industry where I didn't need to become an

199
00:11:08,712 --> 00:11:10,712
S2: expert in machine learning. I didn't even need to touch

200
00:11:10,712 --> 00:11:13,271
S2: those models. I still to this day, have no idea

201
00:11:13,552 --> 00:11:17,872
S2: what the virus classification machine learning models do, because there

202
00:11:17,872 --> 00:11:20,032
S2: was still enough space for me to operate in that

203
00:11:20,032 --> 00:11:23,272
S2: field as a human, and I don't see large language

204
00:11:23,272 --> 00:11:25,792
S2: models as any different from that. I think what we're

205
00:11:25,792 --> 00:11:29,391
S2: going to just see is people making sub models that make, say,

206
00:11:29,432 --> 00:11:32,152
S2: malware analysis better. And it's like, okay, maybe I can't

207
00:11:32,152 --> 00:11:35,392
S2: do manual virus classification anymore, but I can do this

208
00:11:35,392 --> 00:11:38,352
S2: other thing and I think it's going to move slowly

209
00:11:38,352 --> 00:11:41,232
S2: enough as they do that, that people can just pivot

210
00:11:41,511 --> 00:11:45,392
S2: into not even different fields, but into different, different like

211
00:11:45,392 --> 00:11:47,912
S2: roles in the same field where a lot of their

212
00:11:47,912 --> 00:11:51,472
S2: skills are transferable. But rather than doing something that maybe

213
00:11:51,472 --> 00:11:54,802
S2: is very pattern recognition and is very, uh, lm Them

214
00:11:54,842 --> 00:11:57,842
S2: automatable they could move into something that takes the same

215
00:11:57,842 --> 00:12:00,042
S2: skills but actually requires more critical thinking.

216
00:12:01,602 --> 00:12:04,482
S1: Yeah, yeah. I mean, I, I definitely agree there's going

217
00:12:04,522 --> 00:12:08,642
S1: to be pivots. Um, still possible. Unfortunately, I think those

218
00:12:08,642 --> 00:12:12,602
S1: pivots are going to be mostly available to the really sharp,

219
00:12:13,042 --> 00:12:16,682
S1: exceptional people who can pivot up to something much higher

220
00:12:16,682 --> 00:12:19,041
S1: up in the stack, uh, which requires a lot more

221
00:12:19,042 --> 00:12:24,002
S1: critical thinking that the AI can't do yet. Um, you

222
00:12:24,002 --> 00:12:26,521
S1: had the analogy there of speed and approaching the speed

223
00:12:26,522 --> 00:12:29,282
S1: of light, uh, a car approaching the speed of light.

224
00:12:29,761 --> 00:12:33,041
S1: So here, here's what I think might be a fundamental difference, uh,

225
00:12:33,042 --> 00:12:36,602
S1: in our opinions here, um, the speed of light that

226
00:12:36,602 --> 00:12:39,002
S1: you're talking about, I agree it would be hard to say.

227
00:12:39,202 --> 00:12:42,962
S1: Went from 30mph to 35mph. And next year we're going

228
00:12:42,962 --> 00:12:45,242
S1: to be at the speed of light. That's a crazy statement.

229
00:12:46,522 --> 00:12:48,962
S1: The difference is the speed of light that we're trying

230
00:12:48,962 --> 00:12:53,202
S1: to get here, um, that I'm talking about is actually approaching.

231
00:12:53,572 --> 00:12:56,532
S1: Cannot do the work of a knowledge worker. An average

232
00:12:56,932 --> 00:13:01,131
S1: knowledge worker on earth. Okay. And so my argument to

233
00:13:01,172 --> 00:13:05,012
S1: you is that that bar is not high at all.

234
00:13:05,772 --> 00:13:08,651
S1: So it is basically and this is something we talked

235
00:13:08,652 --> 00:13:12,212
S1: about in text. Um, right in the previous version of this,

236
00:13:12,212 --> 00:13:16,252
S1: this discussion, it's like the average knowledge worker. They, they

237
00:13:16,292 --> 00:13:19,892
S1: clock in and clock out. And I'm talking about mean,

238
00:13:19,932 --> 00:13:23,372
S1: you know, like, uh, average just average across the globe.

239
00:13:23,372 --> 00:13:26,972
S1: It's like, hey, got some new emails, respond to the email. Uh,

240
00:13:26,972 --> 00:13:29,292
S1: the email says, hey, go, um, make a list of

241
00:13:29,292 --> 00:13:32,292
S1: all the documents that are related to this project. Okay, cool.

242
00:13:32,332 --> 00:13:36,652
S1: Here's your list. They send out the list. Um, someone says, hey,

243
00:13:36,892 --> 00:13:38,772
S1: I need you to write a report on how our

244
00:13:38,772 --> 00:13:41,372
S1: finances are doing. So what do they do? Is they

245
00:13:41,372 --> 00:13:46,052
S1: pull a wiki, uh, article, they review this other document,

246
00:13:46,052 --> 00:13:49,012
S1: they talk to Sarah in this other department that takes

247
00:13:49,011 --> 00:13:51,011
S1: them 4 hours or 5 hours or two days or

248
00:13:51,011 --> 00:13:53,941
S1: however long it takes, and they write a four page report,

249
00:13:54,302 --> 00:13:55,862
S1: and then they take that report and they put it

250
00:13:55,862 --> 00:13:57,582
S1: in an email and they send it out to the company.

251
00:13:57,782 --> 00:14:01,902
S1: They are doing work. They're being paid 50 grand, 100

252
00:14:01,902 --> 00:14:05,782
S1: grand for this work. And I'm claiming to you that

253
00:14:05,782 --> 00:14:09,462
S1: this is a low bar, not for I in the future,

254
00:14:09,462 --> 00:14:12,302
S1: but for I, we already had a year ago, in

255
00:14:12,302 --> 00:14:15,662
S1: some cases two years ago. So my concern here is

256
00:14:15,662 --> 00:14:18,702
S1: not about the technical definition of I. I don't give

257
00:14:18,742 --> 00:14:21,622
S1: two shits about it. People are going to argue about, oh,

258
00:14:21,622 --> 00:14:25,342
S1: AGI means this technically, specifically, I don't care. All I

259
00:14:25,382 --> 00:14:29,022
S1: care about is impact to humans. So I'm claiming that

260
00:14:29,022 --> 00:14:31,862
S1: the AI that we have a year ago already could

261
00:14:31,862 --> 00:14:35,542
S1: do if it were properly orchestrated. A lot of the

262
00:14:35,542 --> 00:14:39,061
S1: jobs that I just described. So customer service, um, responding

263
00:14:39,102 --> 00:14:43,102
S1: to emails, setting up like a lunch meeting, just just

264
00:14:43,142 --> 00:14:45,582
S1: basic tasks like that that could not be automated with

265
00:14:45,582 --> 00:14:48,542
S1: regular tech, but can be automated with with this tech.

266
00:14:51,872 --> 00:14:54,352
S2: Uh. Well, you see, I kind of disagree with that

267
00:14:54,352 --> 00:14:58,792
S2: because I've unfortunately had to interface with LM based customer

268
00:14:58,792 --> 00:15:03,512
S2: service agents. And they do have the the flaws that

269
00:15:03,512 --> 00:15:07,192
S2: both of us are familiar with with LMS. They are

270
00:15:07,232 --> 00:15:11,472
S2: not deterministic. If I give it the same question 50 times,

271
00:15:11,472 --> 00:15:14,392
S2: it does not have the same response 50 times. Like

272
00:15:14,592 --> 00:15:19,272
S2: depending on, uh, I don't know, the, the star alignment,

273
00:15:19,472 --> 00:15:22,312
S2: how the electricity is going that day, who knows. It

274
00:15:22,312 --> 00:15:25,192
S2: comes out with a different response to each, um, to

275
00:15:25,232 --> 00:15:28,952
S2: each input. And that is fundamentally a problem because if

276
00:15:28,952 --> 00:15:32,232
S2: you need a machine to do a deterministic task like

277
00:15:32,232 --> 00:15:36,192
S2: finance reports or customer service, you cannot have a model

278
00:15:36,512 --> 00:15:39,512
S2: that comes out with completely different answers every single time

279
00:15:39,512 --> 00:15:42,072
S2: you ask the same question. And that is one of

280
00:15:42,072 --> 00:15:45,832
S2: the major limitations that has been, uh, currently holding back

281
00:15:45,832 --> 00:15:49,442
S2: large language models is that by their very nature they

282
00:15:49,442 --> 00:15:52,602
S2: are not deterministic. So all of those tasks you are

283
00:15:52,602 --> 00:15:56,882
S2: talking about are tasks which, uh, when you say knowledge task, you,

284
00:15:56,882 --> 00:15:59,842
S2: you're kind of, I think, sort of talking about tasks

285
00:15:59,842 --> 00:16:03,322
S2: that don't require critical thinking. They don't really require much intelligence.

286
00:16:03,522 --> 00:16:06,962
S2: They're just doing like tasks that will be the same

287
00:16:06,962 --> 00:16:09,642
S2: every time you do them. And if the large language

288
00:16:09,642 --> 00:16:13,402
S2: model cannot even do tasks the same way multiple times,

289
00:16:13,682 --> 00:16:16,522
S2: how can we then go and automate all of those jobs?

290
00:16:17,642 --> 00:16:20,602
S1: Yeah, I think I think it's fairly easy. I mean,

291
00:16:20,882 --> 00:16:23,322
S1: I've been doing this for a long time. Everyone I

292
00:16:23,322 --> 00:16:26,402
S1: know who's building apps here, it's fairly easy to actually

293
00:16:26,402 --> 00:16:29,562
S1: give somebody, uh, give an AI a list and say,

294
00:16:29,562 --> 00:16:31,842
S1: you can only pick from these, like, this is the

295
00:16:31,842 --> 00:16:33,802
S1: thing that was solved, you know, two and a half

296
00:16:33,842 --> 00:16:37,882
S1: years ago, probably before that. Um, but you can make

297
00:16:37,922 --> 00:16:42,562
S1: a system fairly deterministic. Um, and keep in mind, like

298
00:16:42,562 --> 00:16:46,682
S1: customer service is already horrible. Everyone already hated customer service

299
00:16:46,682 --> 00:16:49,402
S1: because it was so bad. Before I came out. Right.

300
00:16:49,442 --> 00:16:52,122
S1: So it's not like customer service is the high bar,

301
00:16:52,242 --> 00:16:54,082
S1: and we're seeing if we can get there. No, it's

302
00:16:54,082 --> 00:16:57,802
S1: a horrible bar. Like, everyone hates it already. And oftentimes

303
00:16:58,082 --> 00:17:00,482
S1: it has the exact same characteristics you were talking about,

304
00:17:00,482 --> 00:17:03,442
S1: where it's like you're not getting good answers. The good

305
00:17:03,442 --> 00:17:07,202
S1: answer is actually somewhere there in the knowledge base. But

306
00:17:07,202 --> 00:17:09,802
S1: somehow this tier one or tier two person you're talking to,

307
00:17:09,842 --> 00:17:13,962
S1: they didn't find the answer. Um, so a couple of,

308
00:17:14,002 --> 00:17:18,602
S1: you know, objections to that. I don't think, um, current

309
00:17:18,602 --> 00:17:21,522
S1: execution of a lot of these jobs is at a

310
00:17:21,522 --> 00:17:24,202
S1: high level. And that's why I think it's possible to

311
00:17:24,242 --> 00:17:28,962
S1: beat it. The other thing is, um, you're not actually

312
00:17:29,002 --> 00:17:32,002
S1: when your boss says, hey, go make a write a

313
00:17:32,002 --> 00:17:35,322
S1: report for this thing that is fundamentally a non-deterministic thing

314
00:17:35,362 --> 00:17:38,482
S1: because the problem is always different that the human is given.

315
00:17:38,482 --> 00:17:41,802
S1: So it definitely does require intelligence, right? You can't give

316
00:17:41,802 --> 00:17:43,842
S1: that to a script because the script doesn't have like

317
00:17:43,882 --> 00:17:46,692
S1: an if, then it doesn't have a deterministic thing to

318
00:17:46,732 --> 00:17:49,492
S1: go look up the thing because the problem is always different.

319
00:17:49,492 --> 00:17:51,292
S1: And this this is another thing that you and I

320
00:17:51,652 --> 00:17:54,892
S1: talked about before, which I thought was super interesting, is like,

321
00:17:55,892 --> 00:18:00,172
S1: what actually is intelligence and how much deviation from a

322
00:18:00,172 --> 00:18:03,611
S1: standard slotted thing does it require to be called intelligence.

323
00:18:03,892 --> 00:18:06,612
S1: So I would argue if someone is thrown a giant

324
00:18:06,612 --> 00:18:09,692
S1: batch of emails and a giant batch of documents and

325
00:18:09,692 --> 00:18:12,292
S1: they're like, hey, Chris. Hey, Sarah, um, go write a

326
00:18:12,292 --> 00:18:15,772
S1: report on this. And I need it by today at 4:00.

327
00:18:17,372 --> 00:18:21,091
S1: No one person or two people or no combination of

328
00:18:21,092 --> 00:18:23,692
S1: humans is going to make the same report. They're going

329
00:18:23,692 --> 00:18:26,612
S1: to make vastly different reports based on how they choose

330
00:18:26,612 --> 00:18:28,852
S1: to go through the documents, based on how they choose

331
00:18:28,852 --> 00:18:30,652
S1: to form the thing. How long is the report going

332
00:18:30,692 --> 00:18:32,852
S1: to be? If you ask a thousand people to make

333
00:18:32,852 --> 00:18:35,292
S1: a report, they're all going to be different, just like

334
00:18:35,292 --> 00:18:38,212
S1: with an LLM. So and the other thing is just

335
00:18:38,212 --> 00:18:42,581
S1: because that doesn't require like PhD level invention or creativity

336
00:18:43,182 --> 00:18:46,862
S1: that still requires human intelligence. No, no tech prior to

337
00:18:46,902 --> 00:18:48,582
S1: AI could do that job.

338
00:18:50,462 --> 00:18:51,982
S2: So I'd actually would like to go back to the

339
00:18:51,982 --> 00:18:56,262
S2: customer service example, because I think I would argue customer

340
00:18:56,262 --> 00:18:59,861
S2: service is horrible for the exact same reason that the

341
00:18:59,862 --> 00:19:03,662
S2: solution you proposed would be horrible. They are given these

342
00:19:03,662 --> 00:19:07,662
S2: limited scripts of answers where if customer asks this, then

343
00:19:07,662 --> 00:19:10,742
S2: you must answer that and they cannot go beyond their script.

344
00:19:10,942 --> 00:19:12,782
S2: So if I come up with a unique problem that

345
00:19:12,782 --> 00:19:16,502
S2: requires some kind of critical thinking, although the human on

346
00:19:16,502 --> 00:19:20,542
S2: the phone has intelligence and is capable of critical thinking,

347
00:19:20,782 --> 00:19:23,182
S2: they are not allowed to go outside of my script.

348
00:19:23,422 --> 00:19:25,942
S2: And essentially you would run into the same problem with

349
00:19:25,942 --> 00:19:28,182
S2: trying to do that with a large language model is

350
00:19:28,182 --> 00:19:30,302
S2: you've just given it a list of canned answers it's

351
00:19:30,302 --> 00:19:32,381
S2: not allowed to go outside of. And you're going to

352
00:19:32,382 --> 00:19:36,142
S2: get the exact same results where there is no critical thinking.

353
00:19:36,302 --> 00:19:38,262
S2: We can get the argument that I don't think I

354
00:19:38,302 --> 00:19:42,112
S2: can think like full stop. But you're essentially just recreating

355
00:19:42,112 --> 00:19:46,352
S2: the same awful system where it's just it's just parroting

356
00:19:46,352 --> 00:19:49,712
S2: back canned responses. And I'm trying to ask my question,

357
00:19:49,872 --> 00:19:52,312
S2: and it doesn't match up to some canned response that

358
00:19:52,311 --> 00:19:55,552
S2: they have. So they're just repeating the same closest line

359
00:19:55,552 --> 00:19:57,872
S2: that they have in their script. And then we both

360
00:19:57,872 --> 00:20:02,232
S2: leave the interaction completely annoyed. And, um, I think that's

361
00:20:02,232 --> 00:20:04,152
S2: exactly what is going to happen with large language models.

362
00:20:04,152 --> 00:20:07,671
S2: And it's already happening. They're just feeding their dogs or

363
00:20:07,672 --> 00:20:11,872
S2: their scripts into the LLM, and it's just making a

364
00:20:11,912 --> 00:20:15,752
S2: worse version of a customer service representative, because not only

365
00:20:15,752 --> 00:20:18,631
S2: does it still have the canned responses, but it doesn't

366
00:20:18,632 --> 00:20:20,672
S2: have the critical thinking to go, oh, maybe I should

367
00:20:20,672 --> 00:20:21,952
S2: escalate this to my boss.

368
00:20:23,071 --> 00:20:28,912
S1: Yeah, no, it's an interesting point. I, I like that argument. Um,

369
00:20:29,472 --> 00:20:31,552
S1: I would say the best argument I have against that

370
00:20:31,552 --> 00:20:35,911
S1: is that, um, companies have been trying this customer service

371
00:20:35,952 --> 00:20:40,722
S1: thing for decades, Like, uh, human customer service with the

372
00:20:40,722 --> 00:20:45,362
S1: different levels. And they have tried like millions of different things.

373
00:20:45,482 --> 00:20:47,722
S1: They've tried to give people autonomy. They've tried to not

374
00:20:47,722 --> 00:20:51,082
S1: give people autonomy. The system they came up with is, look,

375
00:20:51,482 --> 00:20:55,121
S1: follow this script, use this document. So I would argue

376
00:20:55,122 --> 00:20:56,881
S1: to you that the reason that we see this in

377
00:20:56,882 --> 00:20:59,442
S1: the industry is because it's the thing that works best.

378
00:20:59,882 --> 00:21:03,522
S1: If what worked best was to just hire people who

379
00:21:03,522 --> 00:21:07,362
S1: are somewhat smart and give them free rein, then that's

380
00:21:07,362 --> 00:21:09,722
S1: what companies would be doing. Uh, and I don't think

381
00:21:09,722 --> 00:21:12,042
S1: they're doing that because it doesn't work. Now, I would

382
00:21:12,042 --> 00:21:14,602
S1: argue that as you move into tier three, tier four

383
00:21:14,602 --> 00:21:18,561
S1: or whatever, whatever the structure is for these, um, like

384
00:21:18,602 --> 00:21:21,562
S1: more senior customer service people, I feel like that thing

385
00:21:21,561 --> 00:21:24,762
S1: you're describing is exactly what's happening. They are given free

386
00:21:24,762 --> 00:21:28,401
S1: rein because all the known answers have already been tried,

387
00:21:28,882 --> 00:21:32,321
S1: and now they're, um, now now they're going to have

388
00:21:32,321 --> 00:21:33,362
S1: to use critical thinking.

389
00:21:36,122 --> 00:21:39,611
S2: But I think a lot more jobs than you maybe

390
00:21:39,612 --> 00:21:43,372
S2: estimate are they fall into that category of like, sure,

391
00:21:43,372 --> 00:21:46,012
S2: a lot of it is automatable, but the reason they

392
00:21:46,012 --> 00:21:48,252
S2: need a human in that role and not a script

393
00:21:48,252 --> 00:21:51,492
S2: or a machine or whatever we want to call it,

394
00:21:51,772 --> 00:21:57,052
S2: is because there is some aspect that requires critical thinking, um, uh,

395
00:21:57,052 --> 00:21:59,692
S2: like the one I will push back against endlessly until

396
00:21:59,692 --> 00:22:02,932
S2: my last breath is software engineers. I think there is

397
00:22:02,932 --> 00:22:06,812
S2: no point in any of the AI's evolution, at which

398
00:22:06,811 --> 00:22:09,572
S2: point you will be able to replace software engineers. And

399
00:22:09,571 --> 00:22:12,891
S2: I think the disconnect that happens is a lot of

400
00:22:12,892 --> 00:22:16,892
S2: the non software engineers or even the coders, the people

401
00:22:16,892 --> 00:22:20,532
S2: who just don't fully understand software engineering don't understand the

402
00:22:20,532 --> 00:22:24,932
S2: difference between writing code and designing software. So they're looking

403
00:22:24,932 --> 00:22:30,371
S2: at ChatGPT and Cursor and Gemini or whatever, and they're looking, oh,

404
00:22:30,372 --> 00:22:31,851
S2: I can put in a query and it spits out

405
00:22:31,852 --> 00:22:34,252
S2: all of this code. And they're like, that is what

406
00:22:34,342 --> 00:22:38,422
S2: coding is. So like that very easily could replace software engineering.

407
00:22:38,422 --> 00:22:41,422
S2: But because they're not a software engineer, they don't understand

408
00:22:41,422 --> 00:22:44,381
S2: the amount of critical thinking and design decisions that actually

409
00:22:44,382 --> 00:22:47,222
S2: go into the software. I think, uh, a lot of

410
00:22:47,222 --> 00:22:50,902
S2: software engineers will say that writing code is like 10%

411
00:22:50,902 --> 00:22:53,942
S2: of software engineering. But then if you're not a software engineer,

412
00:22:53,942 --> 00:22:58,061
S2: you're looking at that, that a small percentage that the AI, okay,

413
00:22:58,302 --> 00:23:01,262
S2: sort of can do that and then ignoring the rest.

414
00:23:01,462 --> 00:23:04,342
S2: And I think the same is true for every role. Um,

415
00:23:04,342 --> 00:23:07,582
S2: until you actually have like spent a year in any

416
00:23:07,582 --> 00:23:11,262
S2: given job, you don't really understand how much is automatable

417
00:23:11,262 --> 00:23:15,102
S2: versus how much actually does require a human with critical thinking.

418
00:23:15,342 --> 00:23:19,502
S2: So my personal belief is, I think for the large,

419
00:23:19,662 --> 00:23:23,782
S2: like the overwhelming majority of all roles, there will never,

420
00:23:23,821 --> 00:23:26,702
S2: ever be a point in which AI replaces that person.

421
00:23:27,102 --> 00:23:29,381
S2: It might be able to make the person faster. You

422
00:23:29,382 --> 00:23:31,982
S2: might be able to take one person and have them

423
00:23:32,022 --> 00:23:35,272
S2: do work at a slightly more productive rate, but I

424
00:23:35,272 --> 00:23:37,832
S2: don't think there is any point in which I actually

425
00:23:37,832 --> 00:23:42,032
S2: starts replacing like, like large swaths of the job market.

426
00:23:43,512 --> 00:23:47,392
S1: Yeah. Interesting. Um, so let me give you a counterpoint.

427
00:23:47,392 --> 00:23:50,552
S1: And this this is anecdotal, of course, but I have

428
00:23:50,592 --> 00:23:54,272
S1: a close friend who is a cardiologist, like actual MD

429
00:23:54,352 --> 00:23:56,912
S1: actually sits in the, you know, the doctor's office every day.

430
00:23:57,272 --> 00:24:00,832
S1: He also happens to be a bug bounty hunter and like,

431
00:24:00,872 --> 00:24:05,912
S1: a programmer. So he he is, uh, you know, using

432
00:24:05,912 --> 00:24:09,792
S1: whatever tools are approved at his particular job. And, um,

433
00:24:09,792 --> 00:24:12,672
S1: he is now finding as of like this is this

434
00:24:12,672 --> 00:24:15,512
S1: is six months ago. This is a year ago. Um,

435
00:24:15,952 --> 00:24:19,392
S1: the analysis that this, this thing is doing is as

436
00:24:19,392 --> 00:24:23,152
S1: good or better in some cases than the thing that

437
00:24:23,152 --> 00:24:26,191
S1: he was going to write. So this is at the

438
00:24:26,192 --> 00:24:29,472
S1: cardiologist level. Not not even like a nursing level. This

439
00:24:29,472 --> 00:24:32,322
S1: is really, really advanced stuff. So it's it's taking all

440
00:24:32,321 --> 00:24:35,482
S1: the inputs that came from like the patient and also

441
00:24:35,802 --> 00:24:39,522
S1: from his notes. It's combining all that together with, of course,

442
00:24:39,522 --> 00:24:42,802
S1: the knowledge of the model itself. And then, you know,

443
00:24:42,842 --> 00:24:47,682
S1: giving out recommendations, recommending the actual drugs. And so my

444
00:24:47,722 --> 00:24:49,841
S1: thing is like if this thing is actually doing the

445
00:24:49,842 --> 00:24:53,482
S1: job and it's not just one person, right? A lot

446
00:24:53,522 --> 00:24:56,562
S1: of doctors are actually saying that the output is matching

447
00:24:56,561 --> 00:25:00,282
S1: their capabilities as well. We've already seen this with analyzing moles.

448
00:25:00,282 --> 00:25:03,762
S1: We've seen this like in lots of different places. Um,

449
00:25:03,802 --> 00:25:05,922
S1: and that brings me back to this one other point

450
00:25:05,922 --> 00:25:09,002
S1: that you made earlier, which I thought was really interesting, um,

451
00:25:09,402 --> 00:25:11,561
S1: how they're going to hit a wall. The Elon's hit

452
00:25:11,602 --> 00:25:14,562
S1: a wall and they basically start building subsystems. This is

453
00:25:14,561 --> 00:25:18,162
S1: exactly how I define AGI. I don't think AGI is

454
00:25:18,162 --> 00:25:22,242
S1: going to be some breakthrough, generally intelligent thing. I think

455
00:25:22,242 --> 00:25:26,122
S1: that is more of an AI breakthrough than we've actually seen.

456
00:25:26,162 --> 00:25:29,172
S1: I don't think it's happened yet. What I'm arguing is

457
00:25:29,172 --> 00:25:32,372
S1: going to replace the average knowledge worker is actually one

458
00:25:32,372 --> 00:25:37,812
S1: of these, uh, Frankenstein systems where the LLM is like

459
00:25:37,811 --> 00:25:42,252
S1: an orchestrator, and it's basically spinning up these individual things

460
00:25:42,252 --> 00:25:45,452
S1: to go and accomplish individual tasks like write the email,

461
00:25:46,012 --> 00:25:50,492
S1: organize the events, organize the conference, uh, summarize this or whatever.

462
00:25:50,612 --> 00:25:53,772
S1: So it's going to have at its disposal dozens or

463
00:25:53,811 --> 00:25:57,611
S1: hundreds of these little workers. But the problem is that

464
00:25:57,612 --> 00:25:59,692
S1: you're going to buy this thing, let's call it a

465
00:25:59,852 --> 00:26:03,012
S1: synth worker or whatever the stupid AI company is going

466
00:26:03,012 --> 00:26:06,532
S1: to be called. You buy, you buy synth worker, synth

467
00:26:06,532 --> 00:26:09,292
S1: worker shows up to work on Monday. It goes to

468
00:26:09,332 --> 00:26:12,452
S1: the onboarding meeting. It talks to the manager. It talks

469
00:26:12,452 --> 00:26:15,892
S1: to all its team members. It reads slack, it reads

470
00:26:15,892 --> 00:26:19,171
S1: the wiki. And it starts working just like a regular employee.

471
00:26:19,532 --> 00:26:21,932
S1: And it produces the work of a regular employee. And

472
00:26:21,932 --> 00:26:25,932
S1: I would argue in a lot of cases, even better. Um,

473
00:26:26,492 --> 00:26:29,142
S1: but if you look under the hood, it's not some

474
00:26:29,182 --> 00:26:32,862
S1: all intelligent AI system. It's actually a whole bunch of tricks.

475
00:26:32,862 --> 00:26:35,341
S1: It's like this big agent orchestrating a whole bunch of

476
00:26:35,342 --> 00:26:38,462
S1: other ones. But functionally, what you end up with is

477
00:26:38,462 --> 00:26:42,262
S1: this new AI named Julie actually does the work. And

478
00:26:42,262 --> 00:26:44,622
S1: guess what? We didn't hire a human to do that job.

479
00:26:46,702 --> 00:26:48,982
S2: Um, so I guess the first thing I'd say is

480
00:26:48,982 --> 00:26:52,821
S2: that is not the definition of AGI that most people

481
00:26:52,821 --> 00:26:56,462
S2: are going with. Um, but it is a fair point. Um,

482
00:26:56,502 --> 00:26:58,702
S2: a lot of people have this idea of, uh, if

483
00:26:58,702 --> 00:27:01,422
S2: we give the llms enough data, at some point, they

484
00:27:01,422 --> 00:27:04,061
S2: start learning for themselves, and at some point they can

485
00:27:04,102 --> 00:27:06,782
S2: do all of these things at the Submodels are doing, um,

486
00:27:06,821 --> 00:27:09,742
S2: which I think most likely we both agree that that's

487
00:27:09,742 --> 00:27:13,062
S2: that's horseshit. That's that's never happening. Um, but then the

488
00:27:13,061 --> 00:27:17,861
S2: problem is when it comes to customizing these models, um,

489
00:27:17,902 --> 00:27:20,662
S2: now you need a custom model for everything. You, uh,

490
00:27:20,662 --> 00:27:23,982
S2: you need your model to identify tumors. And that is

491
00:27:23,982 --> 00:27:26,542
S2: no different from where we've been like. That is what

492
00:27:26,542 --> 00:27:29,742
S2: we've been doing for the past. Uh, I don't know.

493
00:27:29,782 --> 00:27:32,782
S2: Probably longer than I've been alive, like 30 years, maybe more.

494
00:27:33,102 --> 00:27:36,502
S2: And that hasn't replaced humans. Um, I would go back

495
00:27:36,502 --> 00:27:39,422
S2: to my own personal example, which is machine learning based

496
00:27:39,422 --> 00:27:46,742
S2: malware detection. Those machines have massively improved, uh, malware detection. Like, uh,

497
00:27:46,742 --> 00:27:48,702
S2: back in the old days, you would have to open

498
00:27:48,702 --> 00:27:51,662
S2: up a virus. You would have to manually determine that

499
00:27:51,662 --> 00:27:55,062
S2: this is actually malicious code, and then enter a signature

500
00:27:55,061 --> 00:27:57,542
S2: into a database. And can you imagine, like a human

501
00:27:57,542 --> 00:28:00,262
S2: trying to do that with billions and billions of new

502
00:28:00,302 --> 00:28:04,142
S2: files every single day? It's impossible. Um, so we built

503
00:28:04,142 --> 00:28:07,662
S2: machine learning models to do that. But then the question is, well,

504
00:28:07,702 --> 00:28:10,341
S2: how many malware analysts got laid off as a result?

505
00:28:10,821 --> 00:28:13,702
S2: And the answer is there's actually more malware analysts today

506
00:28:13,902 --> 00:28:17,022
S2: than there was in those days before the machine learning

507
00:28:17,022 --> 00:28:20,582
S2: based virus classification. And the reason was the classification could

508
00:28:20,582 --> 00:28:23,022
S2: only do so much, and you still needed people to

509
00:28:23,152 --> 00:28:25,952
S2: check false positives. You still needed people to write reports,

510
00:28:26,071 --> 00:28:29,631
S2: and there was always a place for humans and it

511
00:28:29,632 --> 00:28:32,552
S2: wasn't even different skills like, uh, my job has been

512
00:28:32,552 --> 00:28:35,872
S2: automated three times now, and in none of those three

513
00:28:35,872 --> 00:28:38,872
S2: times have I had to, uh, even learn like, some

514
00:28:38,872 --> 00:28:41,472
S2: massively different skill set. I've always been able to pivot,

515
00:28:41,752 --> 00:28:44,152
S2: and I think that is all we're going to see

516
00:28:44,152 --> 00:28:48,432
S2: with that is like, okay, maybe our cardiologist, I can

517
00:28:48,472 --> 00:28:51,232
S2: automate a lot of that. So the cardiologists end up

518
00:28:51,232 --> 00:28:53,192
S2: doing something else. I can't say what else they might

519
00:28:53,192 --> 00:28:54,792
S2: end up doing because that's not my field and I

520
00:28:54,792 --> 00:28:57,432
S2: don't know anything about it. But I do know that

521
00:28:57,432 --> 00:29:00,832
S2: in my field, we have just been automating every single thing.

522
00:29:00,832 --> 00:29:03,952
S2: We've been building sub models for malware classification. We've built

523
00:29:03,952 --> 00:29:07,512
S2: some models for, uh, security alerts. We've built uh, sub

524
00:29:07,512 --> 00:29:10,911
S2: models for analyzing code. Um, and we built up all

525
00:29:10,912 --> 00:29:13,232
S2: of these sub models. But there's two problems there. The

526
00:29:13,232 --> 00:29:15,752
S2: first is there is no evidence of it ever actually

527
00:29:15,752 --> 00:29:19,552
S2: replacing humans. In fact, we've, uh, the industry has grown

528
00:29:19,552 --> 00:29:23,322
S2: by a factor of ten since these machine learning models

529
00:29:23,322 --> 00:29:26,162
S2: started coming out. And the second thing is they're very

530
00:29:26,162 --> 00:29:30,682
S2: expensive to produce and maintain. So you're not actually getting

531
00:29:30,682 --> 00:29:32,882
S2: the thing, which is the AI companies are trying to

532
00:29:32,882 --> 00:29:34,682
S2: sell people on, which is, oh, you can get rid

533
00:29:34,682 --> 00:29:37,722
S2: of employees, get rid of Greg, get rid of Simon,

534
00:29:37,722 --> 00:29:40,362
S2: get rid of Sandra, and just keep all their salaries.

535
00:29:40,362 --> 00:29:43,562
S2: Just pay $20 a month for this AI subscription. And

536
00:29:43,562 --> 00:29:46,242
S2: now you're saving all of this, all of this money.

537
00:29:46,602 --> 00:29:49,202
S2: But then when it comes to the actual who's maintaining

538
00:29:49,202 --> 00:29:53,082
S2: these models, who's updating them for the latest, uh, whatever

539
00:29:53,082 --> 00:29:56,802
S2: it is they're doing, because, um, like, obviously the technologies

540
00:29:56,802 --> 00:29:59,202
S2: that they're working on are always changing. So someone has

541
00:29:59,202 --> 00:30:03,122
S2: to maintain these models and they're phenomenally expensive to run.

542
00:30:03,402 --> 00:30:07,162
S2: Like any kind of pattern matching model requires insane amounts

543
00:30:07,162 --> 00:30:10,802
S2: of computational power. And what's actually just happening is you're

544
00:30:10,802 --> 00:30:13,842
S2: taking those salaries, and I don't think you're even cutting them.

545
00:30:13,842 --> 00:30:17,482
S2: You're actually spending more running these models or paying to

546
00:30:17,522 --> 00:30:20,732
S2: have the models run than you would be having employees

547
00:30:20,732 --> 00:30:24,212
S2: do it. It only makes sense when you're operating at scale,

548
00:30:24,772 --> 00:30:27,172
S2: where employees physically like there aren't enough employees to do

549
00:30:27,172 --> 00:30:30,132
S2: this thing. Like, imagine if you had to hire, uh,

550
00:30:30,172 --> 00:30:33,612
S2: humans to manually categorize a billion malware samples. Like, it

551
00:30:33,612 --> 00:30:36,652
S2: just wouldn't be possible. Uh, but when it actually comes to, hey,

552
00:30:36,652 --> 00:30:39,852
S2: let's replace a human with this AI, uh, it usually

553
00:30:39,852 --> 00:30:42,732
S2: ends up just working out more expensive. And the only

554
00:30:42,732 --> 00:30:46,892
S2: reason we haven't started seeing that with Llms yet is

555
00:30:46,892 --> 00:30:50,972
S2: because they're being subsidized by VCs. We don't see how much. Uh,

556
00:30:51,012 --> 00:30:52,932
S2: I'm just going to say ChatGPT, because that's the one

557
00:30:52,932 --> 00:30:56,092
S2: most people are familiar with. Um, we don't really see

558
00:30:56,132 --> 00:30:59,171
S2: how much ChatGPT costs to actually operate. We don't know

559
00:30:59,172 --> 00:31:02,212
S2: how much they're spending on acquiring the training data, training

560
00:31:02,212 --> 00:31:05,532
S2: the model, how much computational power is being used, how

561
00:31:05,532 --> 00:31:08,372
S2: much electricity is being used. And then when you think

562
00:31:08,372 --> 00:31:11,852
S2: about an actual human being, it's just some cells. They

563
00:31:11,852 --> 00:31:14,252
S2: need a little bit of water, a little bit of glucose,

564
00:31:14,532 --> 00:31:17,942
S2: the amount of money it costs to operate a actual

565
00:31:17,942 --> 00:31:21,142
S2: human being is almost zero. The only reason it's not

566
00:31:21,142 --> 00:31:24,661
S2: zero is because capitalism, we've made houses so expensive, so

567
00:31:24,662 --> 00:31:26,862
S2: now they've got to pay rent and we want to

568
00:31:26,902 --> 00:31:29,742
S2: make a profit selling food. So now they've got to buy, uh,

569
00:31:29,782 --> 00:31:33,662
S2: like this $0.01 banana is now a dollar. Um, so

570
00:31:33,662 --> 00:31:36,982
S2: we basically just increased the cost of running humans to

571
00:31:37,022 --> 00:31:40,702
S2: the point where these machines now look compatible. But then

572
00:31:40,702 --> 00:31:42,542
S2: when you try and actually replace the humans, you're just

573
00:31:42,542 --> 00:31:44,382
S2: going to run into the exact same problem, which is

574
00:31:44,382 --> 00:31:47,302
S2: these machines are expensive. They may not need to pay rent. Actually,

575
00:31:47,342 --> 00:31:48,822
S2: technically they do need to pay rent. You need a

576
00:31:48,822 --> 00:31:51,182
S2: data centre, but they may not need to eat. They

577
00:31:51,182 --> 00:31:53,942
S2: may not need water, they may not need plumbing. Uh,

578
00:31:53,942 --> 00:31:56,742
S2: but they need a billion, trillion dollars worth of GPUs,

579
00:31:56,782 --> 00:32:00,142
S2: a ton of electricity, a ton of cooling. And I

580
00:32:00,142 --> 00:32:03,262
S2: actually don't think we are going to reach a point

581
00:32:03,262 --> 00:32:08,062
S2: where we can make, uh, like computational based intelligence more

582
00:32:08,422 --> 00:32:12,102
S2: cheaper than the equivalent human intelligence. I think right now

583
00:32:12,142 --> 00:32:16,032
S2: we're in, uh, I call it sort of the Utopia era.

584
00:32:16,072 --> 00:32:17,912
S2: Even though I don't feel like we're in Utopia, I

585
00:32:17,912 --> 00:32:21,112
S2: think everything is terrible. I hate what's going on with AI.

586
00:32:21,472 --> 00:32:25,512
S2: It makes me miserable. But it reminds me of, um, uh,

587
00:32:25,512 --> 00:32:27,712
S2: in LA, we had this thing called, uh. Well, it's

588
00:32:27,752 --> 00:32:30,392
S2: actually it's quite widespread now, but we had this program

589
00:32:30,392 --> 00:32:32,992
S2: called the Bird Scooter, and it was actually the very

590
00:32:32,992 --> 00:32:35,832
S2: first bird scooter was in LA, and I was there

591
00:32:35,832 --> 00:32:39,072
S2: for the pilot program. It's like, okay, free public transport.

592
00:32:39,072 --> 00:32:41,112
S2: You pay like maybe a dollar at most. You can

593
00:32:41,112 --> 00:32:43,712
S2: go anywhere, you can leave the scooter anywhere you want.

594
00:32:43,992 --> 00:32:46,432
S2: And I was like, this is incredible. Like, this is

595
00:32:46,432 --> 00:32:50,432
S2: going to revolutionize transport. I can go from point A

596
00:32:50,472 --> 00:32:52,712
S2: to point B, I don't need to find parking. It

597
00:32:52,752 --> 00:32:55,152
S2: cost me a dollar. Like this is perfect. But what

598
00:32:55,152 --> 00:32:58,392
S2: was actually happening on the back end is VCs were

599
00:32:58,432 --> 00:33:00,752
S2: pouring billions and billions of dollars into this system to

600
00:33:00,792 --> 00:33:03,392
S2: try and, like, build out this company. And what was

601
00:33:03,392 --> 00:33:05,712
S2: happening is people were getting hit by cars. They were

602
00:33:05,712 --> 00:33:08,592
S2: throwing the scooters into rivers and oceans, and they were

603
00:33:08,632 --> 00:33:12,232
S2: actually losing more money, uh, to running this program than

604
00:33:12,232 --> 00:33:15,122
S2: they were ever going to make. And I think we're

605
00:33:15,122 --> 00:33:17,522
S2: seeing the same thing start to happen with AI is

606
00:33:17,522 --> 00:33:21,681
S2: it's being funded by VCs, but it's there's not really

607
00:33:21,682 --> 00:33:24,802
S2: yet any certain path where it's even going to be profitable.

608
00:33:25,002 --> 00:33:28,402
S2: Like they are currently operating at a loss, and they're

609
00:33:28,402 --> 00:33:31,082
S2: selling us on this idea that at some point you

610
00:33:31,082 --> 00:33:33,882
S2: will be able to run like a genetic AI or

611
00:33:33,882 --> 00:33:37,282
S2: whatever for less than the equivalent of an employee. But

612
00:33:37,322 --> 00:33:39,482
S2: I actually don't think that's going to happen. I don't

613
00:33:39,482 --> 00:33:42,642
S2: think we're going to reach a point where computers are

614
00:33:42,642 --> 00:33:46,242
S2: cheap enough to run, that we can replace humans with these, uh,

615
00:33:46,402 --> 00:33:49,402
S2: with these AI agents. And I think the more that

616
00:33:49,402 --> 00:33:52,322
S2: we start, uh, going from this idea, which is never

617
00:33:52,322 --> 00:33:55,322
S2: going to work of a single AI that is just

618
00:33:55,322 --> 00:33:58,282
S2: able to do everything to building out sub models, I

619
00:33:58,282 --> 00:34:01,442
S2: think that's just going to it's going to compound the cost. Uh,

620
00:34:01,442 --> 00:34:03,322
S2: we're just going to get more and more and more

621
00:34:03,322 --> 00:34:05,962
S2: cost to the point where eventually we'll come to the

622
00:34:06,002 --> 00:34:08,642
S2: realization that actually, it's cheaper to just hire humans.

623
00:34:09,722 --> 00:34:14,292
S1: Okay. I like this argument. I find that interesting. I

624
00:34:14,692 --> 00:34:18,132
S1: don't believe it's true. My intuition is that it's not true.

625
00:34:18,292 --> 00:34:22,492
S1: And the reason for that is just the the amount

626
00:34:22,492 --> 00:34:25,932
S1: that the, um, the costs are falling. And the fact

627
00:34:25,932 --> 00:34:28,372
S1: that I think we're so bad at what we're doing

628
00:34:28,372 --> 00:34:31,732
S1: right now with artificial intelligence. I think there's so much

629
00:34:31,732 --> 00:34:33,972
S1: slack in the rope that I think we're going to

630
00:34:33,972 --> 00:34:38,572
S1: end up getting, you know, 99% or whatever, some, some

631
00:34:38,572 --> 00:34:41,492
S1: dramatic amount of the cost that we're paying now just

632
00:34:41,492 --> 00:34:44,012
S1: will keep falling out and falling out. But I would

633
00:34:44,012 --> 00:34:46,732
S1: say that's an empirical question. And we'll just have to

634
00:34:46,772 --> 00:34:49,212
S1: like see how that works out. Because I do agree

635
00:34:49,212 --> 00:34:52,692
S1: with you that there is a the current state is

636
00:34:52,732 --> 00:34:55,332
S1: like a state of confusion because you don't know how

637
00:34:55,332 --> 00:34:58,292
S1: much it's being propped up. So so I like your argument,

638
00:34:58,532 --> 00:35:01,612
S1: and I definitely think the there will be an effect

639
00:35:01,612 --> 00:35:07,612
S1: from what you're saying. Um, I, um, I want to

640
00:35:07,612 --> 00:35:09,942
S1: I want to touch on this, this whole concept of

641
00:35:10,222 --> 00:35:16,142
S1: intelligence itself in like, because I feel like, um, we

642
00:35:16,142 --> 00:35:18,702
S1: were trying to define it before, uh, when we had

643
00:35:18,702 --> 00:35:22,142
S1: the previous conversation, we were trying to define like what, uh,

644
00:35:22,182 --> 00:35:25,302
S1: what exactly do we mean by it? So I think

645
00:35:25,542 --> 00:35:29,182
S1: the way that I defined it before was, um, or

646
00:35:29,182 --> 00:35:30,702
S1: at least the way that I wanted to define it now,

647
00:35:30,702 --> 00:35:33,782
S1: which which I think is roughly similar, the ability to

648
00:35:33,822 --> 00:35:38,982
S1: solve everyday human problems using knowledge about the world. So

649
00:35:39,022 --> 00:35:41,022
S1: I think that I try to hit it from a

650
00:35:41,022 --> 00:35:44,062
S1: bunch of angles, and I think that accounts for why

651
00:35:44,102 --> 00:35:48,982
S1: an average knowledge worker can't be automated. Um, and I

652
00:35:49,022 --> 00:35:51,782
S1: put every day in there because it's basic stuff like,

653
00:35:51,982 --> 00:35:54,902
S1: should I break up with my boyfriend? Um, and keep

654
00:35:54,902 --> 00:35:57,302
S1: in mind, just just the case of, like, how useful

655
00:35:57,302 --> 00:36:01,062
S1: this stuff is. I think the current numbers are OpenAI

656
00:36:01,182 --> 00:36:05,582
S1: or ChatGPT has like a billion users per day. So

657
00:36:05,582 --> 00:36:09,992
S1: I feel like for everyday problems, It's if we're using

658
00:36:09,992 --> 00:36:12,552
S1: that as a as a metric doesn't that to me

659
00:36:12,552 --> 00:36:15,512
S1: indicates that it's very useful. Um, but but I guess

660
00:36:15,512 --> 00:36:17,752
S1: that's a red herring because you're, you're already acknowledged that

661
00:36:17,752 --> 00:36:21,552
S1: it's useful. Um, but I've got some examples of, like,

662
00:36:22,152 --> 00:36:25,432
S1: things that ChatGPT is really good at. So write an

663
00:36:25,432 --> 00:36:27,712
S1: essay about a book that you read. It could already

664
00:36:27,712 --> 00:36:30,192
S1: write pretty good essays. Oh, and by the way, this

665
00:36:30,192 --> 00:36:33,032
S1: is unrelated to the debate, but we actually blew by

666
00:36:33,032 --> 00:36:36,112
S1: the Turing test because most people can't tell the difference

667
00:36:36,112 --> 00:36:39,592
S1: between AI and humans at this point, because it's already.

668
00:36:39,992 --> 00:36:41,992
S1: I just thought it was an interesting piece of trivia

669
00:36:41,992 --> 00:36:44,272
S1: that that used to be the gold standard for AI,

670
00:36:44,272 --> 00:36:47,712
S1: and now no one even cares that we passed it. Um, managing.

671
00:36:47,792 --> 00:36:52,672
S2: That one was actually passed the with the very early GPT. Um,

672
00:36:52,672 --> 00:36:53,312
S2: I thought I.

673
00:36:53,312 --> 00:36:53,792
S1: Think it might.

674
00:36:53,832 --> 00:36:58,272
S2: Have. Yeah. Because I mean, so as, like a, like

675
00:36:58,552 --> 00:37:01,192
S2: putting my, uh, like my philosophy hat on instead of

676
00:37:01,192 --> 00:37:04,592
S2: my scientist hat on. I have never, ever seen the

677
00:37:04,592 --> 00:37:09,162
S2: Turing test as a reliable metric for anything. Because essentially,

678
00:37:09,162 --> 00:37:11,642
S2: for anyone who doesn't know what that is, it is

679
00:37:11,642 --> 00:37:14,762
S2: just you put a machine and you put a person, uh,

680
00:37:15,162 --> 00:37:19,322
S2: like basically concealed from the person they're talking to. And

681
00:37:19,322 --> 00:37:22,922
S2: the question is, can that person distinguish between a machine

682
00:37:23,682 --> 00:37:26,762
S2: or a human conversation? But the bar for that is

683
00:37:26,762 --> 00:37:30,342
S2: very low, because your average person will spend like 30

684
00:37:30,342 --> 00:37:33,762
S2: hours arguing with a troll bot on Twitter, not realizing

685
00:37:33,762 --> 00:37:36,681
S2: that that is a Python script. Um, so I've always

686
00:37:36,682 --> 00:37:39,562
S2: felt that Turing test is more a test of humans

687
00:37:39,562 --> 00:37:44,042
S2: lack of ability to distinguish a real human, rather than, uh,

688
00:37:44,042 --> 00:37:47,322
S2: AI being successful at what it does. Now, granted, I

689
00:37:47,322 --> 00:37:51,482
S2: will concede that AI is, or at least Llms are very,

690
00:37:51,482 --> 00:37:55,562
S2: very good at imitating human like conversation. I will give

691
00:37:55,562 --> 00:37:59,322
S2: them that. But there is a huge difference between imitating

692
00:37:59,322 --> 00:38:03,202
S2: human like conversation and human intelligence, which is why I

693
00:38:03,202 --> 00:38:06,812
S2: believe that people aren't, uh, that they're not seeing it

694
00:38:06,812 --> 00:38:09,892
S2: passing the Turing test as this amazing feat because it's

695
00:38:09,892 --> 00:38:13,452
S2: not showing that the AI can mimic human intelligence. It's

696
00:38:13,452 --> 00:38:15,812
S2: showing that it can mimic human conversation, which is a

697
00:38:15,812 --> 00:38:16,812
S2: very different thing.

698
00:38:17,412 --> 00:38:20,732
S1: Yeah, that makes sense. I think we agree there. Um,

699
00:38:20,892 --> 00:38:23,972
S1: so what about this definition that I'm using? Um, and

700
00:38:23,972 --> 00:38:26,372
S1: I can't remember the exact definitions that we used before.

701
00:38:26,372 --> 00:38:28,572
S1: We had a slight disagreement there. But what do you

702
00:38:28,572 --> 00:38:31,452
S1: think about this? The ability to solve everyday human problems

703
00:38:31,452 --> 00:38:36,812
S1: using knowledge about the world. So it's kind of general. Right.

704
00:38:37,012 --> 00:38:39,012
S1: And it's general in the sense that a knowledge worker

705
00:38:39,012 --> 00:38:41,492
S1: could do it. It's general in the sense that AGI

706
00:38:41,532 --> 00:38:43,652
S1: would be able to do it. AC obviously would be

707
00:38:43,652 --> 00:38:46,892
S1: able to do it. But I've got examples here managing

708
00:38:46,892 --> 00:38:51,132
S1: someone's daily schedule, writing performance review for an employee, reading

709
00:38:51,132 --> 00:38:54,532
S1: and answering emails, writing code for an application at a company. Um,

710
00:38:54,532 --> 00:38:59,252
S1: by the way, I completely agreed with, um, your your

711
00:38:59,252 --> 00:39:03,892
S1: differentiation between writing code and doing software engineering or software architecture.

712
00:39:03,892 --> 00:39:06,132
S1: Completely agree with you there. I think that's a big

713
00:39:06,132 --> 00:39:09,132
S1: thing that coding is missing out on. Um, writing a

714
00:39:09,132 --> 00:39:12,692
S1: requirements document based on conversations, writing a work status update

715
00:39:12,692 --> 00:39:16,172
S1: for your boss plan, a four day trip to Switzerland.

716
00:39:16,172 --> 00:39:19,572
S1: So these are my examples of like everyday human problems

717
00:39:19,852 --> 00:39:24,532
S1: that are nowhere near free, like automation could possibly even

718
00:39:24,572 --> 00:39:28,412
S1: try to do versus, um, with AI, these are kind

719
00:39:28,412 --> 00:39:30,532
S1: of like trivial. And I would argue that these are

720
00:39:30,532 --> 00:39:33,212
S1: the types of things that are going to make up

721
00:39:33,692 --> 00:39:36,412
S1: a human replacing AI in the workforce.

722
00:39:38,292 --> 00:39:43,532
S2: I'm not sure I agree because, um, uh, the the

723
00:39:43,612 --> 00:39:46,732
S2: point on, uh, I mean, it goes back to your

724
00:39:46,732 --> 00:39:50,972
S2: general knowledge worker point, the question of can it replace

725
00:39:51,012 --> 00:39:54,612
S2: a average worker? Um, and we've talked about that already. Uh,

726
00:39:54,612 --> 00:39:57,292
S2: but I don't think that is the definition of intelligence.

727
00:39:57,292 --> 00:40:00,582
S2: For me, the definition of intelligence is not the ability

728
00:40:00,582 --> 00:40:04,742
S2: to perform a task that wasn't performable by previous automation.

729
00:40:04,942 --> 00:40:08,782
S2: It is the ability to work with incomplete information. Like

730
00:40:08,782 --> 00:40:12,502
S2: if I were to, I don't know. Let's say I

731
00:40:12,542 --> 00:40:14,702
S2: build you a puzzle and I take away some of

732
00:40:14,702 --> 00:40:17,302
S2: the pieces and let's say I take away five of

733
00:40:17,302 --> 00:40:22,182
S2: the pieces. Um, and I give you the puzzle with the, uh,

734
00:40:22,182 --> 00:40:25,702
S2: without those pieces that I took away, you could probably

735
00:40:25,982 --> 00:40:28,942
S2: at some level redraw the rest of the puzzle. Right?

736
00:40:28,982 --> 00:40:32,502
S2: You could use your, uh, critical thinking and your intelligence

737
00:40:32,502 --> 00:40:35,382
S2: to see. Oh, okay. I don't have all the pieces here,

738
00:40:35,382 --> 00:40:37,622
S2: but I can see this is clearly a photo of

739
00:40:37,662 --> 00:40:41,182
S2: a waterfall. Right? It's that ability to work with. Not

740
00:40:41,222 --> 00:40:44,182
S2: what we know, but what we don't know. That is

741
00:40:44,182 --> 00:40:48,742
S2: my definition of intelligence. Um, because I think I think

742
00:40:48,782 --> 00:40:51,342
S2: it gets a little confusing because at least in the

743
00:40:51,342 --> 00:40:54,942
S2: early schooling era, they're not really teaching you intelligence. They're

744
00:40:54,942 --> 00:40:57,502
S2: teaching you knowledge. Uh, the teacher will just give you

745
00:40:57,502 --> 00:40:59,832
S2: a fact, and you memorize that fact, then they'll give

746
00:40:59,832 --> 00:41:02,112
S2: you a test and you repeat back the fact. And

747
00:41:02,112 --> 00:41:03,592
S2: that is the metric that a lot of people are

748
00:41:03,592 --> 00:41:07,432
S2: using for large language models. Yeah, they're doing like bullshit, like, oh,

749
00:41:07,472 --> 00:41:11,112
S2: can it do the SATs? Can it do the bar exam? Like,

750
00:41:11,152 --> 00:41:13,032
S2: of course it can do the bar exam. It's a

751
00:41:13,032 --> 00:41:17,472
S2: database of all the answers. It can just regurgitate the answers. Now,

752
00:41:17,512 --> 00:41:21,032
S2: my definition of intelligence would be okay. What happens when

753
00:41:21,032 --> 00:41:22,992
S2: we put things that, like, we ask it to do

754
00:41:22,992 --> 00:41:26,592
S2: things that aren't in its database, which is admittedly very

755
00:41:26,592 --> 00:41:29,632
S2: hard for large language models because, well, they have all

756
00:41:29,632 --> 00:41:32,072
S2: of human knowledge pretty much in there. They have like

757
00:41:32,072 --> 00:41:35,872
S2: every book, every movie, every web page. So it's very

758
00:41:35,872 --> 00:41:39,552
S2: hard to find a task where the AI cannot just

759
00:41:39,552 --> 00:41:42,512
S2: use basic pattern matching and be. It isn't in their

760
00:41:42,512 --> 00:41:45,952
S2: data set. And this is where I say AI is

761
00:41:45,952 --> 00:41:50,992
S2: not in any like any definition of the word intelligent. Um,

762
00:41:50,992 --> 00:41:54,672
S2: and the way I would assess this is by product. Um,

763
00:41:54,672 --> 00:41:57,162
S2: if we ask the AI to do something that it

764
00:41:57,162 --> 00:41:59,362
S2: already has in its database. It's very easily going to

765
00:41:59,362 --> 00:42:01,362
S2: be able to do it because it already knows the answer.

766
00:42:01,682 --> 00:42:05,362
S2: But then we have a lot of problems, like we

767
00:42:05,362 --> 00:42:08,802
S2: as humans have a lot of unanswered questions. We have

768
00:42:08,802 --> 00:42:11,722
S2: mathematical problems that have not yet been solved. We have

769
00:42:11,882 --> 00:42:16,202
S2: incomplete scientific theories. Now, uh, I like to go to

770
00:42:16,202 --> 00:42:20,842
S2: Einstein for this example, because Einstein lived before computers. Uh,

771
00:42:20,842 --> 00:42:23,002
S2: in his day, he could read as many books as

772
00:42:23,002 --> 00:42:26,122
S2: he could read. He couldn't Google things. He couldn't control

773
00:42:26,122 --> 00:42:28,362
S2: F through a book. He would have to read a

774
00:42:28,362 --> 00:42:31,002
S2: book or a scientific paper. And back then I think

775
00:42:31,002 --> 00:42:34,082
S2: they were published in like, actual physical journals. So he

776
00:42:34,082 --> 00:42:37,002
S2: would have had to gone and got a scientific journal

777
00:42:37,202 --> 00:42:39,562
S2: and read through it to learn a little bit about

778
00:42:39,562 --> 00:42:43,922
S2: a thing. And with as little like, I'm not going

779
00:42:43,962 --> 00:42:47,202
S2: to say Einstein wasn't knowledgeable, but with as little knowledge

780
00:42:47,202 --> 00:42:50,362
S2: as he had access to, he was able to create

781
00:42:50,402 --> 00:42:56,052
S2: theories like special and general relativity, which revolutionized physics. Now,

782
00:42:57,092 --> 00:42:59,412
S2: in physics classes today we work with a lot of

783
00:42:59,412 --> 00:43:03,412
S2: his theories. And your average like, not very bright college

784
00:43:03,412 --> 00:43:07,412
S2: student can can work with Einstein's theories because we already

785
00:43:07,412 --> 00:43:10,172
S2: have the complete picture. He's already come up with them.

786
00:43:10,172 --> 00:43:12,212
S2: He's already given us the answers and we can now

787
00:43:12,252 --> 00:43:15,492
S2: work with them. But what made Einstein so amazing is

788
00:43:15,492 --> 00:43:17,692
S2: he did not have the answers. He came up with

789
00:43:17,692 --> 00:43:23,292
S2: them initially. So, um, kind of my point there would be, okay,

790
00:43:23,292 --> 00:43:26,652
S2: so Einstein could come up with this amazing theory with very,

791
00:43:26,652 --> 00:43:30,892
S2: very limited access to information. We have these machines that have,

792
00:43:31,612 --> 00:43:34,572
S2: I don't even know how much trillions and trillions and

793
00:43:34,572 --> 00:43:37,732
S2: trillions and trillions of data points like every book, every

794
00:43:37,732 --> 00:43:43,572
S2: scientific paper, absolutely everything is in their knowledge database. But like,

795
00:43:43,612 --> 00:43:45,772
S2: what have they come up with? If I ask it

796
00:43:45,772 --> 00:43:49,092
S2: for a theory that unifies classical physics and quantum physics,

797
00:43:49,292 --> 00:43:51,772
S2: it just shits out some existing theories that someone else

798
00:43:51,772 --> 00:43:55,742
S2: came up with, and if it had any intelligence at all,

799
00:43:55,902 --> 00:43:58,902
S2: with the sheer amount of knowledge it has access to

800
00:43:59,062 --> 00:44:01,702
S2: and the sheer amount of computing power it has access to,

801
00:44:01,942 --> 00:44:03,822
S2: I would expect it would be able to complete at

802
00:44:03,862 --> 00:44:06,662
S2: least one of those theories. Like the fact that it

803
00:44:06,662 --> 00:44:09,742
S2: just has so much knowledge, yet seems to have done

804
00:44:09,742 --> 00:44:14,502
S2: nothing novel whatsoever. And in my opinion says that not

805
00:44:14,502 --> 00:44:17,822
S2: only does it lack intelligence, it isn't intelligent at all.

806
00:44:18,622 --> 00:44:21,142
S1: Yeah. So. So this is fascinating. This is where I

807
00:44:21,142 --> 00:44:23,862
S1: wanted to get to. So I think a couple of

808
00:44:23,862 --> 00:44:28,502
S1: times you've made this like horrible, horrible error. And this

809
00:44:28,502 --> 00:44:32,222
S1: is why. This is why it's actually a risk to

810
00:44:32,262 --> 00:44:35,342
S1: regular people. So when we were talking about computer science

811
00:44:35,342 --> 00:44:39,102
S1: and programming, you used yourself as an example of being

812
00:44:39,102 --> 00:44:42,502
S1: able to pivot. You have been on the front page

813
00:44:42,502 --> 00:44:49,462
S1: of wired. You. You're an exceptional person. Um, malware analysts

814
00:44:49,822 --> 00:44:54,592
S1: are also exceptional people within computer science and within security.

815
00:44:54,592 --> 00:44:58,472
S1: So very few people in security can do malware analysis.

816
00:44:58,472 --> 00:45:01,752
S1: So you're already exceptional and then exceptional within that group.

817
00:45:02,072 --> 00:45:06,952
S1: And then you just defined intelligence giving the example of Einstein.

818
00:45:06,952 --> 00:45:10,792
S1: Einstein is one of the smartest people, arguably the smartest

819
00:45:10,792 --> 00:45:16,032
S1: person I would argue Newton. Uh, but anyway, um, one

820
00:45:16,032 --> 00:45:19,672
S1: of the smartest people in the entire world that's ever existed.

821
00:45:20,352 --> 00:45:22,792
S1: If you put the bar there, I agree with you.

822
00:45:22,792 --> 00:45:25,552
S1: And I also want to just grant you the overall

823
00:45:25,552 --> 00:45:28,472
S1: point of like, where are the novel discoveries? Here's my

824
00:45:28,472 --> 00:45:31,911
S1: problem with this, Marcus. Like, that is not the bar

825
00:45:31,912 --> 00:45:34,352
S1: that matters. The bar that matters is what's going to

826
00:45:34,352 --> 00:45:38,512
S1: change society. The bar that matters is taking normal people's

827
00:45:38,872 --> 00:45:42,872
S1: level of intelligence and getting above that bar. So you have,

828
00:45:42,912 --> 00:45:46,952
S1: you know, John Smith, you know, working at some job. Again,

829
00:45:46,952 --> 00:45:49,002
S1: I go down the list writing an essay about the

830
00:45:49,002 --> 00:45:52,762
S1: book that you read, reading and answering emails, organizing a conference.

831
00:45:52,962 --> 00:45:55,962
S1: Hundreds of millions of people are being paid a full

832
00:45:55,962 --> 00:46:01,042
S1: living salary to do these jobs. Not very well. That

833
00:46:01,042 --> 00:46:04,442
S1: level of intelligence, which I think you and I can agree,

834
00:46:04,482 --> 00:46:07,362
S1: like it's not that great. And they are not Einstein.

835
00:46:07,362 --> 00:46:10,202
S1: They are not writing malware. They are. The standard here

836
00:46:10,202 --> 00:46:13,082
S1: is extremely low. The amount of critical thinking needed for

837
00:46:13,122 --> 00:46:15,762
S1: that is extremely low. So what I'm talking about is

838
00:46:15,762 --> 00:46:20,082
S1: an AI that can replace that level of intelligence. And

839
00:46:21,122 --> 00:46:24,762
S1: I just think the the definition of inventing net new

840
00:46:24,762 --> 00:46:31,242
S1: things is an unbelievably high like bar for intelligence. To

841
00:46:31,282 --> 00:46:34,522
S1: me it has to be can you be hit with

842
00:46:34,522 --> 00:46:38,322
S1: regular everyday problems and can you solve them using your

843
00:46:38,362 --> 00:46:41,642
S1: knowledge of how the world works? Because another thing, just

844
00:46:41,642 --> 00:46:43,722
S1: one last thing to say here. Every single one of

845
00:46:43,722 --> 00:46:48,492
S1: those mundane tasks that I just gave, They're not actually

846
00:46:48,492 --> 00:46:52,572
S1: just knowledge lookup. They actually require critical thinking for every

847
00:46:52,572 --> 00:46:54,852
S1: single one of them because the email is not the same.

848
00:46:55,172 --> 00:46:57,252
S1: The report that they're going to write is not the same.

849
00:46:57,972 --> 00:47:00,652
S1: Each each one of those, even though it's relatively simple

850
00:47:00,652 --> 00:47:05,252
S1: for a human, it's impossible for pre-human technology or pre

851
00:47:05,292 --> 00:47:09,012
S1: AI technology that wasn't human. So it's in my mind

852
00:47:09,012 --> 00:47:12,012
S1: it's 100% intelligence because it can't be automated.

853
00:47:13,572 --> 00:47:17,172
S2: I think that's an entirely different argument. Well, there's always

854
00:47:17,172 --> 00:47:20,572
S2: 2 or 3 separate arguments going on here. There's first like,

855
00:47:20,612 --> 00:47:23,692
S2: is it enough to replace the average worker and which

856
00:47:23,692 --> 00:47:26,052
S2: you're sort of extrapolating and saying that is a definition

857
00:47:26,052 --> 00:47:28,972
S2: of intelligence. But I would argue the average worker is

858
00:47:28,972 --> 00:47:32,972
S2: not using their intelligence. They're working primarily with knowledge. Like

859
00:47:32,972 --> 00:47:35,652
S2: if we were to take your average like John Smith

860
00:47:35,692 --> 00:47:39,732
S2: office worker, and apply them to a task that was

861
00:47:39,772 --> 00:47:43,852
S2: novel and did require intelligence, they would probably be able

862
00:47:43,852 --> 00:47:46,302
S2: to do that because they have intelligence that is going

863
00:47:46,342 --> 00:47:49,182
S2: to waste in that job, like they're being made to

864
00:47:49,222 --> 00:47:53,662
S2: do essentially busy work, which is primarily knowledge based, not

865
00:47:53,662 --> 00:47:58,862
S2: intelligence based. Now, I think what um, the kind of, uh,

866
00:47:59,102 --> 00:48:04,502
S2: the miscommunication or the misunderstanding here is trying to attribute, uh,

867
00:48:04,902 --> 00:48:08,942
S2: what requires intelligence for a human as intelligence for a

868
00:48:08,942 --> 00:48:11,902
S2: large language model, because these large language models are doing

869
00:48:11,902 --> 00:48:16,022
S2: pattern matching. While, sure, every email isn't the same, it's

870
00:48:16,022 --> 00:48:19,342
S2: not thinking about the differences between emails. It simply just

871
00:48:19,342 --> 00:48:22,902
S2: has a big database of here's all the different conversations

872
00:48:22,902 --> 00:48:26,422
S2: and every conversation that has pretty much ever occurred. Like,

873
00:48:26,622 --> 00:48:29,582
S2: almost every single conversation I've ever had in my life

874
00:48:29,822 --> 00:48:32,302
S2: is a slight variation of a conversation that has already

875
00:48:32,302 --> 00:48:35,182
S2: been had by someone else. Like there is nothing novel

876
00:48:35,222 --> 00:48:37,742
S2: going on. So I would argue that that is not

877
00:48:37,742 --> 00:48:40,662
S2: a sign of, uh, AI intelligence. It is a sign

878
00:48:40,662 --> 00:48:45,192
S2: that you can, uh, you can almost emulate a a

879
00:48:45,352 --> 00:48:49,672
S2: not fully applied intelligence with a very advanced pattern matching algorithm.

880
00:48:49,912 --> 00:48:50,552
S2: And the reason?

881
00:48:50,592 --> 00:48:53,432
S1: Can I jump in? Can I jump in real quick? Yeah.

882
00:48:53,472 --> 00:48:56,112
S1: Sorry to interrupt. Um, I want to give you an

883
00:48:56,112 --> 00:49:01,472
S1: example here. Couples therapy. So a couples therapist studies for

884
00:49:01,472 --> 00:49:03,632
S1: whatever they have, a master's degree or a PhD, whatever

885
00:49:03,632 --> 00:49:06,392
S1: they have. And a couple comes to them and says,

886
00:49:06,392 --> 00:49:08,472
S1: you know, we're we're about to break up our marriage

887
00:49:08,472 --> 00:49:11,872
S1: because of so and so problem. And they, they help

888
00:49:11,872 --> 00:49:14,272
S1: them for 2 or 3 years listening to all their

889
00:49:14,272 --> 00:49:19,312
S1: different problems. And then they're coaching them through all the psychology,

890
00:49:19,312 --> 00:49:23,232
S1: the sociology, all the trauma stuff, whatever, whatever they're talking

891
00:49:23,232 --> 00:49:28,392
S1: to them about cognitive behavior, behavioral therapy, all this stuff, Marcus,

892
00:49:28,392 --> 00:49:32,752
S1: all that stuff is just knowledge. It's just knowledge. And

893
00:49:32,752 --> 00:49:36,312
S1: guess what? That conversation with this couple is kind of

894
00:49:36,352 --> 00:49:39,272
S1: just like the other conversation with the other couple. Like,

895
00:49:39,272 --> 00:49:42,632
S1: there's not really anything net new. Can we really argue

896
00:49:43,152 --> 00:49:47,192
S1: that what that marriage therapist is doing is not isn't

897
00:49:47,472 --> 00:49:50,552
S1: requiring intelligence? Of course it does. Of course it does.

898
00:49:50,592 --> 00:49:53,552
S1: And what I'm arguing is all this knowledge work. It

899
00:49:53,552 --> 00:49:56,272
S1: does to a lesser degree, of course, and to a

900
00:49:56,272 --> 00:49:59,911
S1: lesser degree than my friend who's a cardiologist, because those

901
00:49:59,912 --> 00:50:02,512
S1: are like really difficult things. But when you were trying

902
00:50:02,512 --> 00:50:05,272
S1: to plan a four day trip to Switzerland with a

903
00:50:05,272 --> 00:50:08,272
S1: different type of family, and one has different food needs,

904
00:50:08,392 --> 00:50:11,112
S1: and the weather is also different in Switzerland, it is

905
00:50:11,112 --> 00:50:13,752
S1: a new it is a net new problem each time,

906
00:50:13,752 --> 00:50:17,712
S1: even though the problem specifics look different than previous versions.

907
00:50:19,992 --> 00:50:22,232
S2: I would say it's a net new problem to a

908
00:50:22,232 --> 00:50:27,032
S2: specific human, but not humanity as a whole, because, um, but.

909
00:50:27,032 --> 00:50:30,472
S1: That doesn't matter. Because that's that's what that's what we

910
00:50:30,512 --> 00:50:33,712
S1: solve every day. It doesn't matter what's new to to

911
00:50:33,752 --> 00:50:37,112
S1: humanity overall, day to day, we're being hit with regular

912
00:50:37,112 --> 00:50:40,842
S1: everyday problems. And that that's the work that we have

913
00:50:40,842 --> 00:50:43,602
S1: to do. We we can't leverage all of humanity for that.

914
00:50:43,602 --> 00:50:44,922
S1: We have to solve it ourselves.

915
00:50:45,962 --> 00:50:48,682
S2: Well, the I guess the problem is, are you trying

916
00:50:48,682 --> 00:50:50,922
S2: to argue that AI is intelligent, or are you trying

917
00:50:50,922 --> 00:50:54,522
S2: to argue that it could do the average or like

918
00:50:54,562 --> 00:50:57,602
S2: any of the jobs that you've given? Because my argument.

919
00:50:57,602 --> 00:50:58,162
S1: I'm arguing.

920
00:50:58,162 --> 00:51:02,002
S2: Both. Okay. So I would agree with you that. Yeah. Um,

921
00:51:02,002 --> 00:51:05,082
S2: a lot of therapy is just pattern matching. Uh, there

922
00:51:05,082 --> 00:51:08,042
S2: are very, uh, I would use like an example like

923
00:51:08,042 --> 00:51:12,362
S2: attachment theory. There's four different types of attachment style. Um,

924
00:51:12,402 --> 00:51:15,722
S2: and everyone fits into one of those four groups. Um,

925
00:51:15,842 --> 00:51:19,082
S2: and there's, uh, some very clearly defined rules of what

926
00:51:19,122 --> 00:51:22,642
S2: tends to cause someone to become a certain attachment style.

927
00:51:23,002 --> 00:51:26,522
S2: But every single situation is going to be slightly different. Um,

928
00:51:26,522 --> 00:51:30,122
S2: but all of those situations map to a single rule

929
00:51:30,242 --> 00:51:33,642
S2: that then maps to your specific attachment style. So essentially

930
00:51:33,642 --> 00:51:35,882
S2: what a therapist is doing there is they're listening to

931
00:51:35,882 --> 00:51:38,612
S2: you and they're trying to pick out like, okay, what

932
00:51:38,612 --> 00:51:41,572
S2: are what are their traumas? Like how did their parents

933
00:51:41,572 --> 00:51:44,732
S2: teach them? And then they're mapping that just to a rule. Now,

934
00:51:44,732 --> 00:51:47,332
S2: as a human, that requires a lot more intelligence than

935
00:51:47,332 --> 00:51:49,732
S2: it would require for a large language model, because the

936
00:51:49,732 --> 00:51:52,812
S2: large language model has way more examples to go on.

937
00:51:52,852 --> 00:51:57,092
S2: So what you're basically taking is um, uh, almost it's

938
00:51:57,092 --> 00:52:00,892
S2: like a, a kind of hand in hand relationship. You can,

939
00:52:00,892 --> 00:52:05,172
S2: to an extent, replace intelligence with knowledge and knowledge with intelligence.

940
00:52:05,372 --> 00:52:09,252
S2: And with the large language model, you have infinitely more knowledge,

941
00:52:09,372 --> 00:52:12,972
S2: which means it needs infinitely less intelligence to replace, uh,

942
00:52:13,212 --> 00:52:16,092
S2: whatever job you want to say it's going to replace.

943
00:52:16,092 --> 00:52:19,532
S2: But my argument is that's not a sign of AI intelligence.

944
00:52:19,532 --> 00:52:22,692
S2: That is a sign of the AI's knowledge. I think

945
00:52:22,732 --> 00:52:26,652
S2: in order to, uh, even demonstrate a small amount of intelligence,

946
00:52:26,972 --> 00:52:29,292
S2: the bar should be a lot higher than for a human.

947
00:52:29,292 --> 00:52:32,372
S2: And that's the reason why I went with the Einstein example. Because, sure,

948
00:52:32,692 --> 00:52:34,812
S2: a lot of people consider Einstein to be the smartest

949
00:52:34,812 --> 00:52:38,262
S2: human on earth. But he was working with a very,

950
00:52:38,262 --> 00:52:41,582
S2: very small amount of information, like a very small amount

951
00:52:41,582 --> 00:52:43,822
S2: of knowledge and a very small amount of data compared

952
00:52:43,822 --> 00:52:47,062
S2: to an AI. So knowledge sort of acts as this

953
00:52:47,102 --> 00:52:51,462
S2: sort of amplifier for intelligence. So if Einstein, with his

954
00:52:51,462 --> 00:52:55,622
S2: amazing intelligence but very limited access to data, could accomplish

955
00:52:55,622 --> 00:53:00,022
S2: such amazing feats, why can something with trillions and trillions

956
00:53:00,022 --> 00:53:04,022
S2: and trillions of data points and books not really accomplish

957
00:53:04,022 --> 00:53:07,422
S2: more than your average Joe, who's not even fully applying

958
00:53:07,422 --> 00:53:10,502
S2: all the intelligence they have. They're just doing data entry

959
00:53:10,622 --> 00:53:14,342
S2: or pattern matching. So my argument there is I'm not

960
00:53:14,342 --> 00:53:17,702
S2: saying that current generation large language models might not be

961
00:53:17,742 --> 00:53:20,662
S2: able to replace certain roles. I'm saying it is not

962
00:53:20,662 --> 00:53:24,062
S2: evidence that they are intelligent. It is intelligent. It's evidence

963
00:53:24,062 --> 00:53:28,022
S2: that their level of knowledge can act in place of intelligence.

964
00:53:28,022 --> 00:53:31,502
S2: And this becomes important because there is a ceiling, there

965
00:53:31,502 --> 00:53:33,952
S2: is a cap where you can't just shove in more

966
00:53:33,952 --> 00:53:35,512
S2: and more knowledge and it just gets more and more

967
00:53:35,512 --> 00:53:38,552
S2: intelligent or sorry, it's not getting intelligent. It's able to

968
00:53:38,592 --> 00:53:41,352
S2: emulate intelligence more and more. At some point you hit

969
00:53:41,352 --> 00:53:44,272
S2: a ceiling and we've already hit that ceiling because we

970
00:53:44,272 --> 00:53:47,752
S2: have these machines that have all of this, like all

971
00:53:47,792 --> 00:53:50,952
S2: of science in their database, and they can't do shit.

972
00:53:50,992 --> 00:53:54,312
S2: Like they cannot complete a single scientific theory.

973
00:53:54,552 --> 00:53:56,512
S1: Yeah. No, no, I love this. First of all, I

974
00:53:56,512 --> 00:53:59,112
S1: just want to say I love this line that you

975
00:53:59,112 --> 00:54:03,272
S1: were on. This is actually not being talked about anywhere. Um, well,

976
00:54:03,472 --> 00:54:05,392
S1: I want to say in all the places that I'm

977
00:54:05,392 --> 00:54:07,672
S1: looking and I'm looking a lot of different places, I

978
00:54:07,712 --> 00:54:11,472
S1: love this point that you just made about how more

979
00:54:11,472 --> 00:54:15,392
S1: and more knowledge is a cheat code to intelligence. It

980
00:54:15,432 --> 00:54:19,752
S1: requires less intelligence. I don't think it matters. And and

981
00:54:19,752 --> 00:54:23,472
S1: here's why. So. So first of all, there's, um, there's

982
00:54:23,472 --> 00:54:25,592
S1: a bunch of post training stuff that gets done. And

983
00:54:25,592 --> 00:54:27,672
S1: I'm not an expert on post training, but there's a

984
00:54:27,672 --> 00:54:31,192
S1: bunch of stuff that you do with these models afterwards,

985
00:54:31,402 --> 00:54:36,122
S1: and I believe that we are making massive inroads on that,

986
00:54:36,602 --> 00:54:42,082
S1: what I call tricks. Um, basically tricking AI as a

987
00:54:42,082 --> 00:54:46,402
S1: system overall to understand how to make these jumps. In

988
00:54:46,402 --> 00:54:49,802
S1: other words, I think it is possible to teach AI

989
00:54:50,442 --> 00:54:55,762
S1: how to go from Newtonian physics to Einsteinian theories, and

990
00:54:55,762 --> 00:54:58,122
S1: I think it's possible to teach it, to generalize, to

991
00:54:58,162 --> 00:55:00,322
S1: be able to do that. And there's already some evidence

992
00:55:00,322 --> 00:55:04,442
S1: of this working. I had a thing in the show recently. Um, basically,

993
00:55:04,442 --> 00:55:10,082
S1: researchers have been working on this one problem, um, with bacteria, bacteriophages,

994
00:55:10,402 --> 00:55:13,322
S1: which are basically viruses that propagate and try to take

995
00:55:13,322 --> 00:55:16,562
S1: over bacteria. And so, um, they've been struggling with this

996
00:55:16,562 --> 00:55:20,042
S1: for like over a decade or maybe two decades. The

997
00:55:20,042 --> 00:55:24,722
S1: absolute pinnacle researchers in this particular thing, they gave it

998
00:55:24,722 --> 00:55:27,762
S1: to this new, uh, Google model, which is specifically designed,

999
00:55:27,922 --> 00:55:30,252
S1: designed to do exactly what you're talking about, Marcus. To

1000
00:55:30,292 --> 00:55:33,252
S1: actually find net new things. And it came back and said,

1001
00:55:33,252 --> 00:55:36,411
S1: here's my hypothesis for why this is happening. They looked

1002
00:55:36,412 --> 00:55:39,572
S1: at it and they said, Holy crap. That is the answer.

1003
00:55:39,652 --> 00:55:42,052
S1: It's a net new answer. They went and tested it.

1004
00:55:42,092 --> 00:55:46,212
S1: It was 100% confirmed. And there's a bunch of more

1005
00:55:46,252 --> 00:55:49,292
S1: examples of this. There's companies that are doing this basically

1006
00:55:49,452 --> 00:55:54,612
S1: harvesting like research that dormant research, raw research and coming

1007
00:55:54,612 --> 00:55:57,332
S1: up with net new hypotheses. So I think that's just

1008
00:55:57,332 --> 00:55:59,812
S1: a matter of like, we just haven't done the work

1009
00:55:59,972 --> 00:56:06,092
S1: of the scaffolding of teaching the AI how to make jumps, right.

1010
00:56:06,132 --> 00:56:08,292
S1: Because like you said, it's been so easy to just

1011
00:56:08,292 --> 00:56:10,692
S1: pull from the knowledge. But I think this is a

1012
00:56:10,692 --> 00:56:12,852
S1: training step that we could do. And this is an

1013
00:56:12,852 --> 00:56:16,172
S1: empirical thing. Like it'll either work or it doesn't. And

1014
00:56:16,172 --> 00:56:20,092
S1: I can send you over the stuff. But the bigger

1015
00:56:20,092 --> 00:56:24,812
S1: point here for me is that, um, this is why

1016
00:56:24,812 --> 00:56:27,132
S1: this matters to me is because of humans. So if

1017
00:56:27,132 --> 00:56:29,142
S1: we go back to the point of like, don't worry

1018
00:56:29,142 --> 00:56:32,382
S1: about I if I could, you know, summarize your thing

1019
00:56:32,422 --> 00:56:35,022
S1: about that saying there's always going to be a place

1020
00:56:35,022 --> 00:56:37,502
S1: for humans. This thing is not a big deal. Um,

1021
00:56:37,502 --> 00:56:41,662
S1: I think you've even called it like, um, auto autocorrect.

1022
00:56:41,702 --> 00:56:45,342
S1: No autocomplete. Right? You call it autocomplete? I'm like, no,

1023
00:56:45,342 --> 00:56:49,662
S1: it is not autocomplete. It's actually doing this work. My

1024
00:56:49,662 --> 00:56:54,902
S1: point to you, Marcus, is my friends cardiology practice is

1025
00:56:54,902 --> 00:57:00,942
S1: nothing but pattern matching. Okay. Most high level work, um,

1026
00:57:00,942 --> 00:57:05,302
S1: is just pattern matching. This, um, couples therapist. They are

1027
00:57:05,302 --> 00:57:07,902
S1: also doing pattern matching. And if you look at the

1028
00:57:07,902 --> 00:57:11,022
S1: average job, they are also doing pattern matching like okay,

1029
00:57:11,062 --> 00:57:14,342
S1: it's we're taking emails. We're writing a report. You can

1030
00:57:14,342 --> 00:57:18,782
S1: reduce all this work that hundreds of millions of people

1031
00:57:18,782 --> 00:57:20,782
S1: are doing. You could reduce that to pattern matching if

1032
00:57:20,782 --> 00:57:26,072
S1: you want. But we're talking about hundreds of millions of jobs, Right?

1033
00:57:26,072 --> 00:57:29,472
S1: So if I could do that work, then it still

1034
00:57:29,472 --> 00:57:31,992
S1: has the impact that I'm concerned about, whatever we want

1035
00:57:31,992 --> 00:57:32,592
S1: to call it.

1036
00:57:34,752 --> 00:57:38,752
S2: I don't disagree, but then the question is like, why

1037
00:57:38,752 --> 00:57:41,712
S2: hasn't it? Are we are we not there yet? Um.

1038
00:57:41,752 --> 00:57:42,112
S2: Are we?

1039
00:57:42,152 --> 00:57:44,912
S1: Because it's just now getting started. It's just now getting started.

1040
00:57:44,952 --> 00:57:47,592
S1: Like the tech is just now getting to the companies.

1041
00:57:47,592 --> 00:57:49,632
S1: These companies don't even know what AI is. They don't

1042
00:57:49,632 --> 00:57:52,272
S1: know what it isn't. It takes time to spin it up.

1043
00:57:52,472 --> 00:57:55,792
S1: Like we see the experiment starting. We see thousands of

1044
00:57:55,792 --> 00:57:58,512
S1: companies trying to adopt it, but they're trying to figure

1045
00:57:58,512 --> 00:58:00,152
S1: out what it is at the same time that they're

1046
00:58:00,152 --> 00:58:02,112
S1: trying to adopt it. So it's just a matter of

1047
00:58:02,152 --> 00:58:04,682
S1: it's just a matter of one, two, three, 4 or

1048
00:58:04,682 --> 00:58:08,312
S1: 5 years in my opinion. And it's already happening. We

1049
00:58:08,312 --> 00:58:10,632
S1: already have evidence that it's happening. It's just a matter

1050
00:58:10,632 --> 00:58:11,072
S1: of time.

1051
00:58:12,472 --> 00:58:16,872
S2: I'm not sure I agree because like companies have, as

1052
00:58:16,872 --> 00:58:18,912
S2: you said, they've been in a rush to roll out AI.

1053
00:58:19,312 --> 00:58:21,832
S2: But then the question is like, if it is so

1054
00:58:21,832 --> 00:58:25,122
S2: close to being able to replace the average knowledge worker.

1055
00:58:25,402 --> 00:58:29,602
S2: Why aren't we seeing any massive replacement? Why aren't we

1056
00:58:29,642 --> 00:58:33,642
S2: seeing any change in the productivity of companies? Because, like,

1057
00:58:33,642 --> 00:58:37,922
S2: we have objective metrics for those things. Um, and one

1058
00:58:37,962 --> 00:58:40,562
S2: sort of tangent I'd want to go on here is

1059
00:58:40,562 --> 00:58:43,322
S2: a lot of industries aren't finite. Like if you think

1060
00:58:43,322 --> 00:58:45,962
S2: of something like, say, farming, there's only so much food

1061
00:58:45,962 --> 00:58:48,162
S2: we need. Uh, if we were able to automate all

1062
00:58:48,202 --> 00:58:50,802
S2: the farming, sure, farmers could go away. Uh, we don't.

1063
00:58:50,882 --> 00:58:54,602
S2: There's no value to. Hey, let's re allocate the farmers

1064
00:58:54,602 --> 00:58:56,762
S2: to figuring out how to make even more food. Like,

1065
00:58:56,762 --> 00:58:59,642
S2: once we've got enough food to feed everyone, we're good. Um,

1066
00:58:59,642 --> 00:59:02,642
S2: but then you have industries which is also overlap with

1067
00:59:02,642 --> 00:59:06,322
S2: the industries that are most invested in AI and are

1068
00:59:06,322 --> 00:59:10,322
S2: most trying to replace workers with AI like tech. Now,

1069
00:59:10,362 --> 00:59:14,082
S2: tech is not a finite industry. There is infinite software

1070
00:59:14,082 --> 00:59:17,162
S2: you can write. There are infinite, uh, feature improvements. There's

1071
00:59:17,162 --> 00:59:21,162
S2: infinite patches now, uh, like, uh, I guess the example

1072
00:59:21,162 --> 00:59:24,332
S2: I give is Microsoft is not a search engine company,

1073
00:59:24,332 --> 00:59:25,852
S2: but they felt the need to come out with a

1074
00:59:25,852 --> 00:59:28,572
S2: search engine. Google is not an operating system company, but

1075
00:59:28,572 --> 00:59:30,852
S2: they felt the need to come up with an operating system.

1076
00:59:31,132 --> 00:59:34,132
S2: And why? Because the more products you make, the more

1077
00:59:34,132 --> 00:59:36,612
S2: market share you can capture, and the more you can

1078
00:59:36,612 --> 00:59:40,972
S2: compete in, uh, in the global market. So from a

1079
00:59:40,972 --> 00:59:44,772
S2: tech company perspective, uh, let's say I can right now

1080
00:59:44,772 --> 00:59:48,452
S2: replace all of my software engineers with AI. Um, the

1081
00:59:48,452 --> 00:59:50,332
S2: illogical thing to do would be to lay off all

1082
00:59:50,332 --> 00:59:53,412
S2: my engineers and do the same thing that I'm already doing,

1083
00:59:53,412 --> 00:59:56,252
S2: but with AI, because all of the companies that are

1084
00:59:56,252 --> 00:59:58,812
S2: smart are just going to do more. They're going to

1085
00:59:58,852 --> 01:00:01,452
S2: just start breaking into all the other industries that all

1086
01:00:01,492 --> 01:00:04,852
S2: their competitors are in, and they're going to dominate every space.

1087
01:00:05,012 --> 01:00:08,692
S2: So the logical path, uh, if I could replace employees,

1088
01:00:08,852 --> 01:00:11,892
S2: would not be laying off your employees. It would be

1089
01:00:11,892 --> 01:00:14,412
S2: to just expand. It's to have your employees do more

1090
01:00:14,412 --> 01:00:16,892
S2: and more things, do more and more products, and capture

1091
01:00:16,892 --> 01:00:20,292
S2: more and more market share. Now, if that was happening,

1092
01:00:20,572 --> 01:00:23,172
S2: we would see an increase in GDP, because GDP is

1093
01:00:23,172 --> 01:00:26,532
S2: the total value of all of the goods and products

1094
01:00:26,532 --> 01:00:30,372
S2: produced within the country. Yet we're not seeing any any

1095
01:00:30,372 --> 01:00:33,132
S2: change in GDP. We're not seeing any change in productivity.

1096
01:00:33,572 --> 01:00:35,612
S2: All of these companies who are claiming that they're going

1097
01:00:35,612 --> 01:00:39,012
S2: to replace all their engineers with AI, we've not seen

1098
01:00:39,012 --> 01:00:44,892
S2: any change in anything. Um, so firstly, I wouldn't especially

1099
01:00:44,892 --> 01:00:47,652
S2: in tech, I would not be worried about employees being

1100
01:00:47,652 --> 01:00:52,372
S2: just replaced, uh, because, um, well, firstly, AI doesn't fully

1101
01:00:52,372 --> 01:00:57,532
S2: replace employee. It accelerates their productivity. So, uh, doing the

1102
01:00:57,532 --> 01:01:00,972
S2: same amount of productivity but with less employees is far

1103
01:01:00,972 --> 01:01:04,972
S2: less desirable than keeping your employees and just doing more. Um,

1104
01:01:04,972 --> 01:01:07,532
S2: but we've not really seen either. We haven't seen employees

1105
01:01:07,532 --> 01:01:10,172
S2: being replaced with AI, and we haven't seen this massive

1106
01:01:10,172 --> 01:01:13,852
S2: explosion in productivity from tech companies. In fact, they're doing

1107
01:01:13,892 --> 01:01:18,222
S2: the opposite. They're laying off their employees to focus on AI.

1108
01:01:18,582 --> 01:01:21,502
S2: They're like, oh, we don't even have enough money to do, uh,

1109
01:01:21,542 --> 01:01:24,702
S2: whatever it was. Like whatever products Google has been, they're

1110
01:01:24,742 --> 01:01:27,502
S2: like binning half of their products and services to go

1111
01:01:27,502 --> 01:01:31,222
S2: and focus on AI. Whereas if I was actually capable

1112
01:01:31,222 --> 01:01:33,382
S2: of what they say it's capable of, they would be

1113
01:01:33,382 --> 01:01:37,862
S2: going in the opposite direction. They'd be making a gaming console,

1114
01:01:37,862 --> 01:01:41,022
S2: they'd be making a desktop operating system better. They would

1115
01:01:41,022 --> 01:01:44,222
S2: be cornering like every single market. And they're not. They're

1116
01:01:44,222 --> 01:01:45,382
S2: actually cutting down.

1117
01:01:46,262 --> 01:01:50,182
S1: Yeah, yeah, yeah. Interesting point, I hear you. Um, I

1118
01:01:50,182 --> 01:01:53,142
S1: think the reason it's not affecting GDP yet is because

1119
01:01:53,222 --> 01:01:56,062
S1: it's it's not rolled out. I mean, this is just

1120
01:01:56,062 --> 01:02:00,422
S1: starting like last year. We weren't even talking about agents yet,

1121
01:02:00,422 --> 01:02:02,342
S1: which is kind of like the the way we try

1122
01:02:02,382 --> 01:02:06,062
S1: to get into, um, all this automation. Um, and now

1123
01:02:06,062 --> 01:02:08,462
S1: agents are just now starting to get serious. My, my

1124
01:02:08,502 --> 01:02:12,662
S1: estimate for this has always been, uh, before 29. So

1125
01:02:12,702 --> 01:02:14,302
S1: I my I think this is going to be like

1126
01:02:14,302 --> 01:02:20,032
S1: a 20, 27 thing when this AGI, um, worker replacement

1127
01:02:20,232 --> 01:02:23,152
S1: technology is good enough to actually start replacing workers and

1128
01:02:23,152 --> 01:02:24,632
S1: it could be way sooner than that. It could be

1129
01:02:24,632 --> 01:02:28,312
S1: this year. It could be whenever, um, that it's not

1130
01:02:28,312 --> 01:02:30,952
S1: until that thing gets rolled out and it starts getting

1131
01:02:30,952 --> 01:02:35,232
S1: implemented by the thousands, by many, many different companies that

1132
01:02:35,232 --> 01:02:38,112
S1: we're actually going to see a GDP bump. So I

1133
01:02:38,152 --> 01:02:41,632
S1: would anticipate that being in like 28, 29, 30 and

1134
01:02:41,632 --> 01:02:45,272
S1: into the 30s, because that's just a slow like massive

1135
01:02:45,272 --> 01:02:50,992
S1: ramp up. Um, and the other thing, um, I can't

1136
01:02:50,992 --> 01:02:52,792
S1: remember your other point. What was the other point that

1137
01:02:52,792 --> 01:02:54,872
S1: you made about, uh, other than the GDP.

1138
01:02:56,552 --> 01:02:59,552
S2: Um, that it would make sense to, uh, not lay

1139
01:02:59,552 --> 01:03:02,112
S2: off employees, but to increase productivity?

1140
01:03:02,392 --> 01:03:05,752
S1: Yeah, yeah. So the difference there, the reason that doesn't

1141
01:03:05,792 --> 01:03:10,392
S1: work is because, um, an employee might cost like, let's

1142
01:03:10,392 --> 01:03:14,722
S1: just call it $200,000 with benefits, but for Depending on

1143
01:03:14,722 --> 01:03:18,402
S1: the level, that could be 3 or $400,000. Well, their

1144
01:03:18,402 --> 01:03:22,562
S1: entire contract with the AI company might be $200,000 if

1145
01:03:22,562 --> 01:03:28,642
S1: they can spin up 40,000 of these employees for $200,000

1146
01:03:28,642 --> 01:03:33,762
S1: instead of the 500 human employees. Then I think they

1147
01:03:33,762 --> 01:03:37,642
S1: would start with getting rid of the the previous ones. Obviously,

1148
01:03:37,642 --> 01:03:39,602
S1: they should do a slow thing, but I think the

1149
01:03:39,602 --> 01:03:41,762
S1: natural thing they're going to do, just based on what

1150
01:03:41,762 --> 01:03:45,082
S1: the CFO says is, yeah, we have these 500 people

1151
01:03:45,442 --> 01:03:49,282
S1: making 400 grand or 300 grand. Yeah, we need to

1152
01:03:49,322 --> 01:03:52,722
S1: phase them out. Keep the top 1%, keep the top 10%,

1153
01:03:52,882 --> 01:03:54,682
S1: and they're going to be the ones, you know, helping

1154
01:03:54,682 --> 01:03:58,842
S1: us move into other areas. But all. Net new actual, um,

1155
01:03:58,842 --> 01:04:01,522
S1: software engineers are going to be this other thing which

1156
01:04:01,522 --> 01:04:04,282
S1: we're already paying for the subscription, which. So it's it's

1157
01:04:04,322 --> 01:04:08,282
S1: like no marginal or. Yeah, the marginal cost is virtually

1158
01:04:08,282 --> 01:04:11,282
S1: zero to add new software engineers.

1159
01:04:12,572 --> 01:04:15,452
S2: But I think that rests on the assumption that the, uh,

1160
01:04:15,532 --> 01:04:19,412
S2: the AI agent can wholly replace an engineer, which is,

1161
01:04:19,412 --> 01:04:21,612
S2: as I said, is something I don't believe will ever,

1162
01:04:21,612 --> 01:04:25,492
S2: ever happen. I think what we will see is maybe

1163
01:04:25,492 --> 01:04:27,772
S2: I'm not even convinced of this yet, but I think

1164
01:04:27,772 --> 01:04:30,732
S2: maybe we will see these AI models get to the

1165
01:04:30,732 --> 01:04:33,732
S2: point where they can actually accelerate the productivity of a

1166
01:04:33,732 --> 01:04:36,132
S2: software engineer. But at the end of the day, there

1167
01:04:36,132 --> 01:04:38,292
S2: is always going to be a human on the end.

1168
01:04:38,412 --> 01:04:41,892
S2: It's just an abstraction. It's like, uh, how when we

1169
01:04:41,892 --> 01:04:44,132
S2: used to write things in machine code, it was super

1170
01:04:44,132 --> 01:04:46,932
S2: slow scrolling ones and zeros on a punch card. And

1171
01:04:46,932 --> 01:04:48,852
S2: then we made assembly language, and that made things a

1172
01:04:48,852 --> 01:04:51,412
S2: bit more productive. And then we made a C, and

1173
01:04:51,412 --> 01:04:54,252
S2: then we made Python. And now you can write whole

1174
01:04:54,252 --> 01:04:57,572
S2: software suites with these like point and click applications. But

1175
01:04:57,612 --> 01:04:59,412
S2: at the end of the day there is always a

1176
01:04:59,412 --> 01:05:02,172
S2: human on the end of that, it is just an

1177
01:05:02,612 --> 01:05:06,052
S2: like an abstraction and an acceleration of a human employee.

1178
01:05:06,292 --> 01:05:09,052
S2: Whereas I think the assumption you're working from here is

1179
01:05:09,052 --> 01:05:11,142
S2: that we get to a point where we can just

1180
01:05:11,142 --> 01:05:14,222
S2: wholly replace the human, which, uh, for the same reason

1181
01:05:14,222 --> 01:05:16,822
S2: I said AI is not intelligent. I don't think we

1182
01:05:16,822 --> 01:05:20,462
S2: can do, because it will always lack that ability to

1183
01:05:20,502 --> 01:05:23,422
S2: fill in the gaps, to come up with answers when

1184
01:05:23,422 --> 01:05:25,982
S2: not all the pieces are there. Like, I don't know

1185
01:05:26,062 --> 01:05:28,942
S2: how much you've done software engineering, but a lot of

1186
01:05:28,942 --> 01:05:31,702
S2: it is like the client doesn't even know what they want. Like,

1187
01:05:31,702 --> 01:05:34,382
S2: how do you prompt ChatGPT to build code when the

1188
01:05:34,382 --> 01:05:37,222
S2: client isn't even sure what it is they want to build? Um,

1189
01:05:37,222 --> 01:05:39,582
S2: and that's always been a large part of software engineering.

1190
01:05:39,622 --> 01:05:43,622
S2: It's not the writing code, which, uh, admittedly it's iffy,

1191
01:05:43,622 --> 01:05:46,542
S2: but you can to an extent do with large language models.

1192
01:05:46,702 --> 01:05:49,742
S2: It's the making the design decisions. It's like, what kind

1193
01:05:49,742 --> 01:05:52,022
S2: of server infrastructure do you want? How do you want

1194
01:05:52,022 --> 01:05:54,262
S2: the applications to talk to each other? And those are

1195
01:05:54,262 --> 01:05:57,462
S2: not decisions that an agent can make. Their decisions that

1196
01:05:57,462 --> 01:05:59,862
S2: need to be made by a human and the human

1197
01:05:59,862 --> 01:06:02,222
S2: making the decisions doesn't even know what decisions they want

1198
01:06:02,262 --> 01:06:02,782
S2: to make.

1199
01:06:03,182 --> 01:06:07,182
S1: Yeah, but so so check this out. Um, I want

1200
01:06:07,222 --> 01:06:10,152
S1: to use your previous points kind of to counter that argument,

1201
01:06:11,032 --> 01:06:15,352
S1: making those server choices and making those customer choices, those

1202
01:06:15,352 --> 01:06:19,752
S1: conversations are exactly the same as the marriage therapists, okay.

1203
01:06:20,192 --> 01:06:25,992
S1: They're they're basing that on good fundamental principles of building

1204
01:06:25,992 --> 01:06:31,152
S1: good applications with servers and applications and network connectivity and

1205
01:06:31,192 --> 01:06:34,152
S1: authentication and security and privacy and all these different things.

1206
01:06:34,152 --> 01:06:36,712
S1: These are fundamental rules. These are like written down and

1207
01:06:36,712 --> 01:06:39,992
S1: you can debate them or whatever. But fundamentally, AI is

1208
01:06:39,992 --> 01:06:41,912
S1: going to be doing the same exact thing that they're

1209
01:06:41,912 --> 01:06:47,392
S1: doing for like writing emails and sending emails and summarizing documents. Fundamentally,

1210
01:06:47,392 --> 01:06:51,552
S1: we're not making a Einsteinian jump here. We're talking about

1211
01:06:51,552 --> 01:06:55,192
S1: teaching an AI. What a fundamental building of an application

1212
01:06:55,192 --> 01:06:58,792
S1: infrastructure looks like, and what good software engineering looks like.

1213
01:06:58,912 --> 01:07:01,752
S1: That is a knowledge base. So the fact that we

1214
01:07:01,792 --> 01:07:03,752
S1: haven't done that yet and we're still stuck in this

1215
01:07:03,752 --> 01:07:06,922
S1: vibe coding land, that's because vibe coding started about 48

1216
01:07:06,922 --> 01:07:11,442
S1: seconds ago in I time, right? It started just now.

1217
01:07:11,762 --> 01:07:15,322
S1: So the vibe coding or the the eyes are not

1218
01:07:15,322 --> 01:07:18,442
S1: good at doing the software engineering piece yet. But it is.

1219
01:07:18,482 --> 01:07:22,482
S1: It is the same. It's not any harder than the cardiologist,

1220
01:07:22,882 --> 01:07:25,322
S1: which is also a knowledge base, or the marriage therapist,

1221
01:07:25,322 --> 01:07:27,602
S1: which is also a knowledge base. It's just a matter

1222
01:07:27,602 --> 01:07:30,642
S1: of time before this stuff becomes more capable. Now in

1223
01:07:30,642 --> 01:07:32,962
S1: terms of like, well, there'll always be a human in

1224
01:07:32,962 --> 01:07:36,282
S1: the loop. I mean, I think the higher level you go,

1225
01:07:36,322 --> 01:07:41,522
S1: the more advanced you go. Like, I think there's going

1226
01:07:41,522 --> 01:07:44,082
S1: to be humans in the loop over agent farms that

1227
01:07:44,082 --> 01:07:46,722
S1: are doing things, and the human's going to be applying

1228
01:07:47,162 --> 01:07:49,722
S1: what I believe to be the most irreplaceable thing, which

1229
01:07:49,722 --> 01:07:54,762
S1: is like taste and judgment and like preference, uh, because

1230
01:07:54,802 --> 01:07:57,642
S1: eyes are not alive. They don't have their own opinions.

1231
01:07:58,082 --> 01:07:59,682
S1: And so I think that's going to be the kind

1232
01:07:59,682 --> 01:08:01,882
S1: of like spiritual shaping these things are going to be

1233
01:08:01,882 --> 01:08:04,892
S1: putting into these agent farms. But in terms of execution,

1234
01:08:04,892 --> 01:08:07,412
S1: of writing the code and making sure it's on a

1235
01:08:07,412 --> 01:08:10,652
S1: good infrastructure that's secure. That, to me is all knowledge

1236
01:08:10,652 --> 01:08:11,332
S1: based stuff.

1237
01:08:12,892 --> 01:08:15,732
S2: I would actually push back and say that I completely

1238
01:08:15,732 --> 01:08:19,972
S2: disagree because unlike with the past examples you've given me,

1239
01:08:20,212 --> 01:08:25,092
S2: tech moves insanely fast. Like we have, uh, there's probably

1240
01:08:25,092 --> 01:08:28,532
S2: been 15 new database, uh, frameworks that have come out

1241
01:08:28,532 --> 01:08:31,892
S2: since I started programming. And back in the day, it

1242
01:08:31,892 --> 01:08:34,732
S2: would be buying everything into an SQL database. And that

1243
01:08:34,732 --> 01:08:37,772
S2: was horrible for the most the majority of applications. And

1244
01:08:37,772 --> 01:08:39,412
S2: then we started coming out with all these new types

1245
01:08:39,412 --> 01:08:42,572
S2: of databases. We came out with NoSQL and all of

1246
01:08:42,572 --> 01:08:47,852
S2: these unstructured, um, like query based. Um, uh, I don't

1247
01:08:47,852 --> 01:08:49,252
S2: even want, I just want to call it like a

1248
01:08:49,252 --> 01:08:52,772
S2: data bin. And then we had like AWS buckets, and

1249
01:08:52,812 --> 01:08:57,171
S2: then we're constantly, uh, inventing new technologies. And the problem

1250
01:08:57,172 --> 01:09:00,372
S2: there is twofold. The first is there is no set

1251
01:09:00,372 --> 01:09:03,782
S2: of rules because we're inventing new technologies at a rate

1252
01:09:03,782 --> 01:09:07,502
S2: in which there it's it's basically it is opinions, it's decisions.

1253
01:09:07,742 --> 01:09:10,142
S2: And there's no hard and fast rule of if this

1254
01:09:10,142 --> 01:09:14,822
S2: then that for database or server design. And also uh,

1255
01:09:15,262 --> 01:09:18,902
S2: unlike with, say, the marriage therapist, almost everything that has

1256
01:09:18,902 --> 01:09:21,702
S2: ever happened to you has happened to someone else. Like,

1257
01:09:22,182 --> 01:09:24,262
S2: like a lot of people would like to believe that

1258
01:09:24,302 --> 01:09:27,502
S2: their experience is, like, unique and no one else has

1259
01:09:27,502 --> 01:09:30,662
S2: lived their life and no one else has their problems.

1260
01:09:30,822 --> 01:09:34,502
S2: But in reality, your average human is like a hundred.

1261
01:09:34,542 --> 01:09:36,422
S2: Other people have been through the same stuff that they

1262
01:09:36,462 --> 01:09:39,822
S2: have been through. Whereas with technology we're trying to build new,

1263
01:09:39,982 --> 01:09:43,062
S2: new applications. We're trying to build stuff that hasn't already

1264
01:09:43,062 --> 01:09:47,142
S2: been built. So if we would say just, uh, confined

1265
01:09:47,142 --> 01:09:50,782
S2: within only the software that already exists, sure. You could

1266
01:09:50,782 --> 01:09:53,182
S2: just use rules, but then what would be the value

1267
01:09:53,182 --> 01:09:55,302
S2: of that? Like what is the value of, hey, we

1268
01:09:55,302 --> 01:09:57,782
S2: can just build software we already have like, oh, I

1269
01:09:57,782 --> 01:10:00,062
S2: can build a HTTP server, why don't I just download

1270
01:10:00,062 --> 01:10:04,272
S2: Apache or nginx. Um, so that is a problem. And

1271
01:10:04,272 --> 01:10:06,992
S2: then the second thing is, uh, in order for the

1272
01:10:06,992 --> 01:10:10,392
S2: large language models to actually make, uh, even emulate making

1273
01:10:10,392 --> 01:10:13,752
S2: those decisions, they need knowledge of those frameworks, which means

1274
01:10:13,752 --> 01:10:16,232
S2: they have to be trained on that data. So someone

1275
01:10:16,232 --> 01:10:19,232
S2: has to go and they have to create massive amounts

1276
01:10:19,232 --> 01:10:21,671
S2: of data to feed into the large language model, because

1277
01:10:21,672 --> 01:10:24,592
S2: it's almost like, uh, I'd call it like a lossy

1278
01:10:24,592 --> 01:10:27,912
S2: compression algorithm. Uh, you can't just put like a really

1279
01:10:27,912 --> 01:10:31,392
S2: good paper on database design into a large language model.

1280
01:10:31,392 --> 01:10:33,032
S2: And now your large language model is really good at

1281
01:10:33,032 --> 01:10:36,552
S2: database design. You need like, thousands and millions of data

1282
01:10:36,552 --> 01:10:39,312
S2: points in order for it to even be mediocre at

1283
01:10:39,352 --> 01:10:41,832
S2: that decision making. So now we're going to have a

1284
01:10:41,832 --> 01:10:44,552
S2: lag where like a new technology will come out. And

1285
01:10:44,552 --> 01:10:47,631
S2: as a human with like actual intelligence, but a lack

1286
01:10:47,632 --> 01:10:49,592
S2: of knowledge, because this technology is new and I'm not

1287
01:10:49,592 --> 01:10:52,392
S2: an expert in it, I can read the documentation and

1288
01:10:52,392 --> 01:10:55,392
S2: I can understand it. I can fill in the missing pieces.

1289
01:10:55,392 --> 01:10:57,592
S2: I can make decisions about how that might pertain to

1290
01:10:57,592 --> 01:11:01,072
S2: my software. but the large language model. Now, it doesn't

1291
01:11:01,072 --> 01:11:03,472
S2: have any intelligence to begin with, and it doesn't have

1292
01:11:03,472 --> 01:11:05,152
S2: the knowledge because it's not being put out there on

1293
01:11:05,152 --> 01:11:07,512
S2: the internet yet for it to be trained on. So

1294
01:11:07,512 --> 01:11:10,552
S2: now we just have this massive lag where an actual

1295
01:11:10,552 --> 01:11:13,512
S2: human developer is going to be better than the AI

1296
01:11:13,552 --> 01:11:15,552
S2: because they're going to know more and they're going to

1297
01:11:15,552 --> 01:11:18,312
S2: be able to work with newer technologies better. So I

1298
01:11:18,312 --> 01:11:21,472
S2: would really push back on that, especially with software engineering,

1299
01:11:21,672 --> 01:11:24,152
S2: because a lot of fields are very static. It's like

1300
01:11:24,152 --> 01:11:29,032
S2: things haven't changed in like decades or millennia. Whereas technology

1301
01:11:29,072 --> 01:11:31,112
S2: like in the time we've been recording this video, there's

1302
01:11:31,112 --> 01:11:33,192
S2: probably like 15 new technologies. I'm going to have to

1303
01:11:33,192 --> 01:11:35,312
S2: go and learn. So I would.

1304
01:11:35,352 --> 01:11:39,192
S1: Disagree. I hear you there. I hear you there. Um,

1305
01:11:39,192 --> 01:11:43,952
S1: let me respond to that. Um, so essentially, who do

1306
01:11:43,952 --> 01:11:45,872
S1: you think is going to be better at learning a

1307
01:11:45,872 --> 01:11:49,792
S1: new tech? Uh, so so I think you're incorrect about

1308
01:11:50,352 --> 01:11:52,751
S1: you need to go and get, um, you know, thousands

1309
01:11:52,752 --> 01:11:56,112
S1: or millions of examples, uh, what people are currently doing, uh,

1310
01:11:56,232 --> 01:11:58,522
S1: in state of the art of building this stuff now,

1311
01:11:58,802 --> 01:12:00,482
S1: and I'm pretty sure you and I could just test

1312
01:12:00,482 --> 01:12:04,002
S1: this offline afterwards. Um, we could potentially make, like, a

1313
01:12:04,002 --> 01:12:07,842
S1: fake language, a fake new programming language, also using AI,

1314
01:12:08,162 --> 01:12:11,762
S1: and then say, rewrite this working application in this new

1315
01:12:11,762 --> 01:12:15,602
S1: application in this new, um, programming language, and you give

1316
01:12:15,602 --> 01:12:18,282
S1: it a full programming language spec and it could just

1317
01:12:18,282 --> 01:12:22,522
S1: port automatically over. So I think you're making my point

1318
01:12:22,522 --> 01:12:26,041
S1: for me. So if you think about the concept of

1319
01:12:26,082 --> 01:12:31,802
S1: creating a. Net new application, um, going back to human psychology, um,

1320
01:12:31,922 --> 01:12:36,082
S1: and going back to philosophical principles, think about this. There

1321
01:12:36,122 --> 01:12:38,322
S1: are only so many net new things that you can make.

1322
01:12:38,362 --> 01:12:40,082
S1: And they're going to be oriented. They're going to have

1323
01:12:40,082 --> 01:12:43,842
S1: the shape, the pothole shape of human problems. So like

1324
01:12:43,882 --> 01:12:46,002
S1: you're not going to make something up that is net

1325
01:12:46,042 --> 01:12:48,602
S1: new to a computer that it that the AI has

1326
01:12:48,602 --> 01:12:50,802
S1: never heard of before. You're going to make things like,

1327
01:12:50,842 --> 01:12:52,322
S1: oh it's going to be a game. Oh it's going

1328
01:12:52,322 --> 01:12:54,282
S1: to be an application where I submit this and get

1329
01:12:54,282 --> 01:12:57,171
S1: this back. And when you show it a new programming language,

1330
01:12:57,172 --> 01:12:59,572
S1: it's not going to be like, Holy crap, I've never

1331
01:12:59,572 --> 01:13:01,532
S1: seen that before. You've blown my mind. It's going to

1332
01:13:01,532 --> 01:13:04,812
S1: be like, are you kidding me? I know all programming languages,

1333
01:13:04,812 --> 01:13:07,412
S1: and I see what you've done here with the spec. Okay, cool. Yeah,

1334
01:13:07,412 --> 01:13:11,252
S1: I could use that. So if there are 25 new

1335
01:13:11,252 --> 01:13:15,572
S1: programming paradigms and programming languages and programming specs that come

1336
01:13:15,572 --> 01:13:18,412
S1: out while we're doing this, the advantage of these new

1337
01:13:18,572 --> 01:13:23,291
S1: software engineers, uh, things, they're they're parsing those all the time.

1338
01:13:23,292 --> 01:13:26,012
S1: That's just part of the agent infrastructure to constantly be

1339
01:13:26,012 --> 01:13:29,492
S1: re ingesting these new languages. And here's the crazy part.

1340
01:13:29,532 --> 01:13:32,652
S1: If it hadn't done that, then when you say just

1341
01:13:32,692 --> 01:13:35,211
S1: I want you to program in, uh, you know, booga

1342
01:13:35,212 --> 01:13:37,252
S1: booga or whatever it is, it's like, I don't know

1343
01:13:37,252 --> 01:13:39,532
S1: what that is. Hold on. Give me a second, okay?

1344
01:13:39,532 --> 01:13:42,652
S1: I just consumed everything anyone's ever said about it because

1345
01:13:42,652 --> 01:13:44,972
S1: I live crawled the internet about it. I read all

1346
01:13:44,972 --> 01:13:48,171
S1: the forums, I read the entire programming spec. I brought

1347
01:13:48,172 --> 01:13:50,252
S1: that in. That's now part of my knowledge. Would you

1348
01:13:50,252 --> 01:13:52,052
S1: like me to rewrite this in? Booga booga.

1349
01:13:53,022 --> 01:13:56,302
S2: I don't, uh, so I don't agree for multiple reasons.

1350
01:13:56,302 --> 01:14:01,982
S2: The first is that I think we're talking more about, um, uh,

1351
01:14:02,382 --> 01:14:05,022
S2: I don't know what actually is the correct term for it,

1352
01:14:05,022 --> 01:14:07,662
S2: but you have two functionalities to a large language model.

1353
01:14:07,662 --> 01:14:10,982
S2: You have the actual training database where they've ingested the data,

1354
01:14:10,982 --> 01:14:14,622
S2: they've run it through their training system, they've done their, uh,

1355
01:14:14,622 --> 01:14:17,902
S2: human like reinforcement learning on the data. And that's how

1356
01:14:17,902 --> 01:14:20,822
S2: you get the AI being able to do something. Now,

1357
01:14:20,822 --> 01:14:22,622
S2: on top of that, it can search the internet, it

1358
01:14:22,622 --> 01:14:25,182
S2: can grab some new data. And then based on its

1359
01:14:25,182 --> 01:14:30,262
S2: existing knowledge base, it can to an extent parse that data.

1360
01:14:30,262 --> 01:14:32,542
S2: But I would argue that firstly, like even something as

1361
01:14:32,542 --> 01:14:35,501
S2: simple as a new programming language would not be something

1362
01:14:35,502 --> 01:14:38,062
S2: that it could just go out, fetch the spec, and

1363
01:14:38,062 --> 01:14:41,142
S2: then using its existing database, be able to write in

1364
01:14:41,182 --> 01:14:43,142
S2: that language. I think it would have to be not

1365
01:14:43,142 --> 01:14:45,822
S2: just retrained, but also it would have to go through

1366
01:14:45,822 --> 01:14:49,582
S2: the human, uh, reinforcement learning process. And then my second

1367
01:14:49,582 --> 01:14:52,552
S2: point is I've actually done this, um, not even with

1368
01:14:52,552 --> 01:14:55,352
S2: a new programming language, but, uh, there's a niche programming

1369
01:14:55,352 --> 01:14:58,152
S2: language that I have to use quite a lot. And

1370
01:14:58,152 --> 01:15:01,392
S2: for the life of me, I cannot get any large

1371
01:15:01,392 --> 01:15:04,432
S2: language model to write functional code in that language, because

1372
01:15:04,432 --> 01:15:07,352
S2: there's just not enough data points about it. I can

1373
01:15:07,352 --> 01:15:09,352
S2: go on to a forum, I can grab a script

1374
01:15:09,352 --> 01:15:12,352
S2: and it'll work. But whenever I ask the large language

1375
01:15:12,352 --> 01:15:15,392
S2: model to make me something in that language, it just

1376
01:15:15,392 --> 01:15:17,631
S2: falls flat on its face like it cannot do it.

1377
01:15:17,912 --> 01:15:20,912
S2: So like based on my experience and we're talking something

1378
01:15:20,912 --> 01:15:24,022
S2: as simple as a programming language, which can be 1

1379
01:15:24,022 --> 01:15:27,112
S2: to 1 mapped to an existing programming language, when we

1380
01:15:27,112 --> 01:15:30,352
S2: talk about like brand new database frameworks, it's not like, oh,

1381
01:15:30,352 --> 01:15:33,672
S2: this is just SQL, but the words are different. Like,

1382
01:15:33,672 --> 01:15:37,072
S2: this is an entirely new framework with entirely new, uh,

1383
01:15:37,072 --> 01:15:40,232
S2: design decisions and things to comprehend. Um, and if I

1384
01:15:40,272 --> 01:15:43,312
S2: can't even get these large language models to write me

1385
01:15:43,312 --> 01:15:45,512
S2: a language where it literally is a 1 to 1

1386
01:15:45,512 --> 01:15:48,712
S2: map of existing languages, I can't imagine it being able

1387
01:15:48,712 --> 01:15:51,722
S2: to just on the fly, consume, uh, say like a

1388
01:15:51,722 --> 01:15:54,802
S2: documentation or a press release about a new technology and

1389
01:15:54,802 --> 01:15:57,282
S2: then be able to immediately work with that technology.

1390
01:15:58,122 --> 01:16:00,962
S1: Yeah. I mean, I would agree it's not possible. Now, um,

1391
01:16:01,202 --> 01:16:03,722
S1: I've actually tried this as well. I mean, there's also

1392
01:16:03,722 --> 01:16:06,642
S1: just languages that a given model is better with, and

1393
01:16:06,642 --> 01:16:09,722
S1: everyone kind of knows that. So it's not as simple

1394
01:16:09,722 --> 01:16:12,282
S1: as just go grab the spec for the language. And

1395
01:16:12,282 --> 01:16:14,722
S1: now it's suddenly better because of the previous point that

1396
01:16:14,722 --> 01:16:17,881
S1: you made. But this is the type of thing that

1397
01:16:17,882 --> 01:16:22,002
S1: billions of dollars are being spent on, like, uh, cursor and,

1398
01:16:22,042 --> 01:16:25,001
S1: you know, um, well, I guess all the pinnacle models

1399
01:16:25,002 --> 01:16:28,282
S1: are obviously working on this, but the kind of the

1400
01:16:28,282 --> 01:16:31,362
S1: biggest challenge that everyone's trying to solve right now is

1401
01:16:31,362 --> 01:16:34,762
S1: how to get that context into the current working model.

1402
01:16:35,082 --> 01:16:39,082
S1: And that is accelerating at like a crazy amount. I

1403
01:16:39,082 --> 01:16:42,002
S1: don't think it requires retraining. Um, I think fine tuning

1404
01:16:42,002 --> 01:16:44,122
S1: is largely kind of blown up and just not a

1405
01:16:44,122 --> 01:16:50,171
S1: great thing. Um, I expect that to be very possible, uh,

1406
01:16:50,292 --> 01:16:53,732
S1: very soon. And I'm not saying, like, um, an entire

1407
01:16:53,732 --> 01:16:57,532
S1: new database structure, entirely new programming language. Um, but it

1408
01:16:57,532 --> 01:17:00,572
S1: takes time for a human to learn a programming language

1409
01:17:00,572 --> 01:17:02,212
S1: as well. It's not like the human just go reads

1410
01:17:02,212 --> 01:17:04,092
S1: the spec and now they can suddenly do it. That's

1411
01:17:04,092 --> 01:17:06,251
S1: a thing that also takes a human years to learn.

1412
01:17:06,492 --> 01:17:09,892
S1: I think AIS are going to be way faster, um,

1413
01:17:09,932 --> 01:17:12,652
S1: at doing that. And I think that'll speed up, um,

1414
01:17:12,652 --> 01:17:15,692
S1: pretty dramatically over the next. I don't know who knows

1415
01:17:15,692 --> 01:17:17,092
S1: how fast 1 to 3 years.

1416
01:17:19,692 --> 01:17:22,012
S2: Yeah. So that's kind of where I think we our

1417
01:17:22,012 --> 01:17:25,732
S2: core disagreement is, which is basically, I think in order

1418
01:17:25,732 --> 01:17:29,572
S2: to be able to do that, you need actual functional intelligence,

1419
01:17:29,812 --> 01:17:32,211
S2: which I do not believe a machine can ever have.

1420
01:17:32,452 --> 01:17:35,692
S2: So I think we can only keep building these models

1421
01:17:36,012 --> 01:17:40,492
S2: that they emulate intelligence on specific tasks. But I actually

1422
01:17:40,492 --> 01:17:42,812
S2: think it's going to take longer to build those models

1423
01:17:42,812 --> 01:17:45,292
S2: for these new technologies than it is for a bunch

1424
01:17:45,292 --> 01:17:48,342
S2: of humans to just learn about the tech. So I

1425
01:17:48,342 --> 01:17:50,062
S2: think we're actually just going to we're going to hit

1426
01:17:50,062 --> 01:17:53,582
S2: a wall where a lot of people are thinking that, um,

1427
01:17:53,622 --> 01:17:55,342
S2: these models are just get more and more advanced, more

1428
01:17:55,342 --> 01:17:57,742
S2: and more intelligent, and at some point they will reach

1429
01:17:57,742 --> 01:18:00,942
S2: human level intelligence. Whereas my personal belief is that we

1430
01:18:00,942 --> 01:18:03,542
S2: have already capped out and we're now at the stage

1431
01:18:03,542 --> 01:18:05,542
S2: of we will just build some models that can do

1432
01:18:05,542 --> 01:18:09,622
S2: this specific task. So now you're, uh, essentially now your

1433
01:18:09,622 --> 01:18:12,902
S2: question is, can I learn a new framework or a

1434
01:18:12,902 --> 01:18:16,461
S2: new language or a new technology, then understand it enough

1435
01:18:16,462 --> 01:18:19,902
S2: to build a submodel, build the Submodel, make sure it

1436
01:18:19,902 --> 01:18:23,542
S2: actually works functionally faster than someone could, just learn it

1437
01:18:23,542 --> 01:18:26,542
S2: and write code in it. And I, I don't see

1438
01:18:26,542 --> 01:18:30,142
S2: any way in which without getting to a model that

1439
01:18:30,142 --> 01:18:34,462
S2: has actual human level intelligence, not knowledge, actual intelligence, we

1440
01:18:34,462 --> 01:18:36,421
S2: would ever get to the point where it can learn

1441
01:18:36,422 --> 01:18:38,102
S2: that quickly without retraining.

1442
01:18:39,462 --> 01:18:41,742
S1: Well, keep in mind it only has to do it once, right?

1443
01:18:41,742 --> 01:18:44,152
S1: Because then it just becomes like an MCP that somebody

1444
01:18:44,152 --> 01:18:46,631
S1: can call so the whole world can use it, right?

1445
01:18:46,672 --> 01:18:49,872
S1: So I think that's just the scale of the the

1446
01:18:49,872 --> 01:18:52,791
S1: scale of the benefit of doing it this way, um,

1447
01:18:52,832 --> 01:18:55,392
S1: is extraordinary. The other thing is like, this is not

1448
01:18:55,392 --> 01:18:58,312
S1: even talk about like actual ASI, which I consider to

1449
01:18:58,312 --> 01:19:01,912
S1: be like true general intelligence, which I'm still a bit

1450
01:19:01,952 --> 01:19:04,472
S1: agnostic on. I would say I'm like 90% that it

1451
01:19:04,472 --> 01:19:07,072
S1: gets there eventually, but I have no idea when. But

1452
01:19:07,072 --> 01:19:11,192
S1: more importantly, like, I tend to like, listen, when somebody like, uh,

1453
01:19:11,512 --> 01:19:14,992
S1: Karpathy or, uh, Ilya or people like this, people who

1454
01:19:14,992 --> 01:19:17,671
S1: are like knee deep into this and are the true

1455
01:19:17,672 --> 01:19:21,192
S1: experts on what the frontier looks like. Um, plus a

1456
01:19:21,192 --> 01:19:22,992
S1: lot of the people that I follow, actually, at the

1457
01:19:22,992 --> 01:19:26,232
S1: labs who are not like the marketing people, not the CEO,

1458
01:19:26,512 --> 01:19:30,192
S1: but the actual researchers talking about how fast they're making

1459
01:19:30,192 --> 01:19:34,392
S1: these jumps. Um, I think this whole thing of, like,

1460
01:19:34,432 --> 01:19:37,592
S1: learn a new framework for programming is going to be

1461
01:19:37,592 --> 01:19:43,642
S1: relatively small in the difficulty, um, category. Um, but I

1462
01:19:43,682 --> 01:19:46,322
S1: do hear your main point. I would say that this

1463
01:19:46,322 --> 01:19:48,562
S1: whole thing we've been talking about here is not even

1464
01:19:48,562 --> 01:19:51,802
S1: my main point. My main point is that what I

1465
01:19:51,842 --> 01:19:55,762
S1: care about is the impact of humans, and that if,

1466
01:19:56,362 --> 01:20:01,402
S1: you know, cardiologists and marriage therapists and really advanced, like,

1467
01:20:01,442 --> 01:20:05,642
S1: highly trained people are largely doing rote knowledge and having

1468
01:20:05,682 --> 01:20:10,842
S1: an an interaction about rote knowledge that could be defined,

1469
01:20:10,882 --> 01:20:14,882
S1: that could be applied to software engineering. Okay, what if

1470
01:20:14,882 --> 01:20:18,322
S1: we just didn't start using brand new tech all the

1471
01:20:18,322 --> 01:20:21,082
S1: time to build new tech? So let's say, for example,

1472
01:20:21,082 --> 01:20:24,882
S1: we actually locked in on TypeScript and a certain back

1473
01:20:24,922 --> 01:20:27,642
S1: end technology and a certain, you know, database and a

1474
01:20:27,642 --> 01:20:30,881
S1: certain server structure. And then the AI got really, really

1475
01:20:30,882 --> 01:20:34,042
S1: good at that. And then the whole goal became, let's

1476
01:20:34,042 --> 01:20:38,162
S1: maximize GDP for the planet or whatever it is that

1477
01:20:38,802 --> 01:20:43,732
S1: that would potentially make a software engineer extremely an AI

1478
01:20:43,772 --> 01:20:47,492
S1: based software engineer way more productive and way lower marginal

1479
01:20:47,492 --> 01:20:50,372
S1: cost than hiring a human right. There's no rule that

1480
01:20:50,372 --> 01:20:53,092
S1: says we must adopt new technologies all the time as

1481
01:20:53,092 --> 01:20:55,492
S1: they come out the following day, even though that is

1482
01:20:55,532 --> 01:20:58,972
S1: like the way humans do it. But I just don't

1483
01:20:58,972 --> 01:21:02,892
S1: think we can look at the top tier of someone

1484
01:21:02,892 --> 01:21:06,972
S1: like a Marcus, someone like myself, someone like a malware engineer,

1485
01:21:07,212 --> 01:21:10,652
S1: someone like, you know, a cardiologist or whatever, and say,

1486
01:21:10,692 --> 01:21:14,092
S1: you know, there's parts of their job that they're going

1487
01:21:14,092 --> 01:21:16,372
S1: to be able to pivot because they're exceptional. There's parts

1488
01:21:16,372 --> 01:21:18,412
S1: of their job and I might not be able to do.

1489
01:21:18,572 --> 01:21:22,092
S1: I'm worried about the other 99% who are not doing

1490
01:21:22,092 --> 01:21:26,612
S1: anything extraordinarily exceptional. All those examples that I gave of,

1491
01:21:26,652 --> 01:21:29,452
S1: like regular tasks, that's what most people are being paid

1492
01:21:29,452 --> 01:21:33,171
S1: money to do. And that's what I can already do.

1493
01:21:33,212 --> 01:21:37,171
S1: It just hasn't been like systematized and like brought into

1494
01:21:37,172 --> 01:21:40,132
S1: companies where it's actually running as an engine, where they

1495
01:21:40,132 --> 01:21:44,492
S1: can start laying everyone off and hiring more of those things.

1496
01:21:44,812 --> 01:21:47,372
S1: So if you could replace doctors like this, if you

1497
01:21:47,372 --> 01:21:52,772
S1: could replace, you know, um, marriage counselors, then it's going

1498
01:21:52,812 --> 01:21:54,692
S1: to head knowledge work. It's going to hit it really bad.

1499
01:21:56,492 --> 01:22:00,331
S2: Um, I mean, I guess it's a big if, but

1500
01:22:00,332 --> 01:22:04,292
S2: if it could, then. Sure. Yeah. Um, but then there's

1501
01:22:04,292 --> 01:22:06,852
S2: there's two, two points I would make. The first is,

1502
01:22:06,852 --> 01:22:09,211
S2: does it all happen at once? Does it happen so

1503
01:22:09,212 --> 01:22:12,452
S2: fast that all of these people are just out of work? Um,

1504
01:22:12,772 --> 01:22:15,852
S2: and two, is that, like, do we not find a

1505
01:22:15,852 --> 01:22:19,212
S2: way to deal with that? Um, I would argue that

1506
01:22:19,212 --> 01:22:21,932
S2: there's always going to be critical thinking work in any

1507
01:22:21,932 --> 01:22:25,812
S2: role that cannot be automated with large language models. So

1508
01:22:25,812 --> 01:22:28,132
S2: there is always going to be a place for someone

1509
01:22:28,252 --> 01:22:32,372
S2: who has the knowledge of that domain and the critical thinking. Um,

1510
01:22:32,372 --> 01:22:35,012
S2: I see this a lot in security, actually. Uh, we,

1511
01:22:35,012 --> 01:22:37,822
S2: we try and secure medical systems. It's like, I don't

1512
01:22:37,822 --> 01:22:40,422
S2: know shit about medical systems. I can be like, oh,

1513
01:22:40,462 --> 01:22:42,222
S2: let's fire a wall it off from the internet. And

1514
01:22:42,222 --> 01:22:44,142
S2: it's like, oh no, that person is dead now. It

1515
01:22:44,142 --> 01:22:47,102
S2: needed to talk to the cloud to set their heartbeat

1516
01:22:47,102 --> 01:22:50,142
S2: pace or some shit. Um, so then we have medical, uh,

1517
01:22:50,142 --> 01:22:53,182
S2: professionals come in and they understand these technologies, they understand

1518
01:22:53,182 --> 01:22:55,982
S2: the patient and they understand the needs of the patient,

1519
01:22:55,982 --> 01:22:58,342
S2: and they, uh, they communicate them to us, and then

1520
01:22:58,342 --> 01:23:01,182
S2: we do the security. And I think there's always going

1521
01:23:01,182 --> 01:23:04,142
S2: to be an equivalent of that for anything you can

1522
01:23:04,142 --> 01:23:06,062
S2: do with AI. There is always going to be a

1523
01:23:06,062 --> 01:23:09,022
S2: need for someone who has the same knowledge as the AI,

1524
01:23:09,422 --> 01:23:13,222
S2: but also the, uh, the critical thinking, the intelligence to

1525
01:23:13,262 --> 01:23:17,142
S2: actually work out the problems. So it might be we just, uh,

1526
01:23:17,382 --> 01:23:22,222
S2: everyone becomes project managers, like, instead of having, uh, humans doing, like,

1527
01:23:22,262 --> 01:23:28,262
S2: data entry and busywork, everyone becomes a product, uh, project manager.

1528
01:23:28,262 --> 01:23:31,182
S2: And the AI is now the low level employee. Now, personally,

1529
01:23:31,182 --> 01:23:33,822
S2: I would love if that was possible. If we could

1530
01:23:34,072 --> 01:23:38,272
S2: take the entire working class and move everyone up a

1531
01:23:38,272 --> 01:23:41,791
S2: level and have all those people make, uh, good salaries

1532
01:23:41,832 --> 01:23:44,792
S2: like project manager salaries, and then have the AI do

1533
01:23:44,792 --> 01:23:46,711
S2: all the grunt work that no one wants to do.

1534
01:23:46,992 --> 01:23:49,112
S2: I think that would be amazing, but I also don't

1535
01:23:49,112 --> 01:23:51,792
S2: think it's realistic. I don't think we're actually going to

1536
01:23:51,792 --> 01:23:54,671
S2: get to the point where it can even do enough

1537
01:23:54,672 --> 01:23:56,072
S2: to get us to that level.

1538
01:23:57,072 --> 01:23:59,312
S1: Don't you think the AIS are going to be okay

1539
01:23:59,312 --> 01:24:01,712
S1: if the AI could do cardiologist work, don't you think

1540
01:24:01,752 --> 01:24:07,792
S1: it could be a better project manager? Nope. Isn't project management?

1541
01:24:07,792 --> 01:24:10,072
S1: Isn't all the things that you just described? That is

1542
01:24:10,072 --> 01:24:14,392
S1: knowledge that that thing of like, well, we didn't know

1543
01:24:14,392 --> 01:24:16,912
S1: how to secure it because it's a medical thing. That

1544
01:24:16,912 --> 01:24:20,312
S1: medical thing is knowledge that is captured somewhere. Like, this

1545
01:24:20,312 --> 01:24:23,232
S1: is all just a matter of orchestration. I mean, you

1546
01:24:23,232 --> 01:24:25,152
S1: could feed all this to a project manager and it

1547
01:24:25,152 --> 01:24:28,072
S1: would know, hey, Marcus, don't, like, set up that firewall

1548
01:24:28,072 --> 01:24:32,482
S1: rule because so-and-so is going to die like this. This is, um.

1549
01:24:32,522 --> 01:24:34,482
S1: This is not net new stuff. This goes back to

1550
01:24:34,482 --> 01:24:37,322
S1: your previous point of like. Well, now you're just calling

1551
01:24:37,322 --> 01:24:39,442
S1: up knowledge. Yeah, that's that's the whole game. The whole

1552
01:24:39,442 --> 01:24:40,762
S1: game is calling up knowledge.

1553
01:24:41,482 --> 01:24:44,122
S2: Well I meant that example, not as an example of

1554
01:24:44,122 --> 01:24:46,961
S2: something I couldn't do, but as an example of when

1555
01:24:46,962 --> 01:24:49,482
S2: we need to bring in people with one skill set

1556
01:24:49,522 --> 01:24:52,522
S2: to another. Um, and the same would be true for

1557
01:24:52,682 --> 01:24:55,722
S2: the skill set essentially just being intelligence, like critical thinking,

1558
01:24:55,962 --> 01:24:59,122
S2: the the parts that are missing from AI, there would

1559
01:24:59,122 --> 01:25:03,082
S2: exist that in every field. And you can't take someone

1560
01:25:03,082 --> 01:25:05,602
S2: who has, uh, like what I was trying to get

1561
01:25:05,602 --> 01:25:07,602
S2: at is you can't take someone who has the intelligence

1562
01:25:07,602 --> 01:25:10,082
S2: and the knowledge, sorry, the intelligence and the critical thinking,

1563
01:25:10,082 --> 01:25:12,762
S2: but not the knowledge, and then give them the knowledge

1564
01:25:12,762 --> 01:25:15,522
S2: of the AI and make it whole. Uh, because that

1565
01:25:15,522 --> 01:25:17,842
S2: person now has no ability to fact check the AI.

1566
01:25:18,242 --> 01:25:21,802
S2: They have no ability to work with that information because

1567
01:25:21,802 --> 01:25:23,282
S2: they don't know if it's true or not. They just

1568
01:25:23,282 --> 01:25:25,762
S2: know what this model is telling them. Which is why

1569
01:25:25,762 --> 01:25:29,802
S2: I say, like with the medical professionals, I don't know

1570
01:25:29,802 --> 01:25:33,852
S2: about how medical devices work. I could ask ChatGPT. It's

1571
01:25:33,852 --> 01:25:35,452
S2: going to give me a horrible answer and I'm going

1572
01:25:35,492 --> 01:25:38,612
S2: to end up killing someone. So we need someone who understands,

1573
01:25:38,612 --> 01:25:40,812
S2: who has both the knowledge in the medical field and

1574
01:25:40,812 --> 01:25:44,972
S2: the critical thinking and intelligence. So my argument is that, uh,

1575
01:25:45,092 --> 01:25:47,692
S2: these large language models will always be able to replace

1576
01:25:47,692 --> 01:25:50,331
S2: the knowledge. They will always be able to, to an extent,

1577
01:25:50,332 --> 01:25:53,692
S2: substitute knowledge, but they can never do the critical thinking.

1578
01:25:53,692 --> 01:25:56,532
S2: So you always need a human for that. And it

1579
01:25:56,532 --> 01:25:58,292
S2: can't just be any human. It's going to have to

1580
01:25:58,292 --> 01:26:01,532
S2: be someone who also has some of the knowledge to

1581
01:26:01,732 --> 01:26:03,412
S2: actually work with the model.

1582
01:26:04,372 --> 01:26:07,412
S1: Yeah. So here's what I think the fundamental issue is

1583
01:26:07,412 --> 01:26:11,172
S1: that you gave an example of like when AI is

1584
01:26:11,172 --> 01:26:14,892
S1: using its knowledge and its vast knowledge, it's not actually

1585
01:26:14,892 --> 01:26:18,171
S1: being intelligent. Well guess what? When you bring over Sarah

1586
01:26:18,172 --> 01:26:22,092
S1: Meier to be the medical expert to help us not

1587
01:26:22,092 --> 01:26:25,692
S1: do the wrong firewall rule, she is not being Einstein right?

1588
01:26:25,692 --> 01:26:28,902
S1: Then she's just calling on her knowledge. She is also

1589
01:26:28,902 --> 01:26:34,142
S1: not invoking her creativity and her massive critical thinking. She's

1590
01:26:34,142 --> 01:26:38,342
S1: calling a database inside her own brain. That database can

1591
01:26:38,382 --> 01:26:42,022
S1: be in the context that is given when this AI

1592
01:26:42,062 --> 01:26:44,662
S1: tries to write the firewall rule. When the AI goes

1593
01:26:44,662 --> 01:26:47,622
S1: to consider writing the firewall rule, it will pull from

1594
01:26:47,622 --> 01:26:51,102
S1: its knowledge, which includes Sarah Meyer's knowledge, which is you

1595
01:26:51,102 --> 01:26:54,742
S1: don't put firewall rules for egress traffic inside of medical,

1596
01:26:55,102 --> 01:26:58,622
S1: you know, um, high criticality systems because you might block

1597
01:26:58,662 --> 01:27:02,542
S1: the heartbeat monitor that that was not critical thinking that

1598
01:27:02,542 --> 01:27:06,462
S1: Sarah did she she called on her knowledge. So the

1599
01:27:06,462 --> 01:27:10,421
S1: same exact thing that how how is that how is that?

1600
01:27:11,262 --> 01:27:14,742
S2: Because essentially, when you're trying to bridge these two different fields,

1601
01:27:14,742 --> 01:27:17,501
S2: like we're trying to secure a medical system that's taking

1602
01:27:17,502 --> 01:27:20,902
S2: security and medical systems, there is a gap there. There

1603
01:27:20,902 --> 01:27:23,302
S2: is a gap where let's say we have Sarah, who

1604
01:27:23,302 --> 01:27:25,942
S2: knows like a lot about medical systems. And then there's

1605
01:27:25,942 --> 01:27:28,912
S2: me who knows a lot about cybersecurity. Well, there needs

1606
01:27:28,912 --> 01:27:30,631
S2: to be communication. There needs to be thought, there needs

1607
01:27:30,632 --> 01:27:33,272
S2: to be problem solving. And that is where the intelligence

1608
01:27:33,272 --> 01:27:35,592
S2: comes in. If we just sit there with our collective

1609
01:27:35,592 --> 01:27:38,152
S2: knowledge and we sit in a room together and we're like,

1610
01:27:38,152 --> 01:27:41,872
S2: I'm knowledgeable and you're knowledgeable, nothing's going to happen. We

1611
01:27:41,872 --> 01:27:43,872
S2: have to discuss and we have to reason and we

1612
01:27:43,872 --> 01:27:46,552
S2: have to think. And that is the part that large

1613
01:27:46,552 --> 01:27:49,872
S2: language models cannot do. They cannot reason. They cannot think.

1614
01:27:50,152 --> 01:27:53,512
S2: They just have the knowledge. Now you need the intelligence

1615
01:27:53,512 --> 01:27:56,872
S2: to apply the knowledge. And that is where we come in.

1616
01:27:57,192 --> 01:28:00,232
S1: So so yeah, I love this example by the way.

1617
01:28:00,232 --> 01:28:03,592
S1: This is great. So imagine you're in the room with Sarah.

1618
01:28:04,272 --> 01:28:09,112
S1: And again we're going back to the fact that intelligence

1619
01:28:09,272 --> 01:28:12,192
S1: is needed when you don't have the knowledge. Right. So

1620
01:28:12,192 --> 01:28:14,832
S1: so what's this? You're sitting in the, in the room

1621
01:28:14,832 --> 01:28:17,512
S1: with all the knowledge about cybersecurity. She's sitting in the

1622
01:28:17,512 --> 01:28:21,032
S1: room with all the knowledge about medical systems, and you're

1623
01:28:21,032 --> 01:28:25,282
S1: having a conversation. It turns out you're just exchanging knowledge,

1624
01:28:26,042 --> 01:28:31,041
S1: because the fact of what should be done is already known. Right?

1625
01:28:31,282 --> 01:28:36,242
S2: I disagree. Yeah. Like you're making a situation where, like,

1626
01:28:36,282 --> 01:28:38,961
S2: whatever it is we're trying to solve, there is already

1627
01:28:38,962 --> 01:28:42,162
S2: a known solution. Um, whereas if that was the case,

1628
01:28:42,162 --> 01:28:44,482
S2: then obviously security would be solved, right? Like, if there

1629
01:28:44,482 --> 01:28:48,842
S2: was a known solution to every problem, we wouldn't have jobs. Um,

1630
01:28:48,842 --> 01:28:52,122
S2: I actually don't know if you work in cyber security still. Um, but.

1631
01:28:52,722 --> 01:28:55,162
S1: So so check this out. That, that that kind of

1632
01:28:55,202 --> 01:28:58,482
S1: is the case, though, isn't it? Marcus doesn't. Has anyone

1633
01:28:58,482 --> 01:29:00,282
S1: come up with an answer, like, have you coming up

1634
01:29:00,282 --> 01:29:03,642
S1: with an answer for a customer where? Um, it hadn't

1635
01:29:03,642 --> 01:29:06,682
S1: been talked about, like basically doing that solution hadn't been

1636
01:29:06,682 --> 01:29:09,602
S1: talked about a million times throughout the history of cybersecurity.

1637
01:29:11,282 --> 01:29:14,642
S2: Yes, I personally have, um, but a lot of the

1638
01:29:14,642 --> 01:29:17,922
S2: time it isn't just like, it's kind of hard to

1639
01:29:17,922 --> 01:29:21,202
S2: explain in a way that the non-security viewers will will

1640
01:29:21,322 --> 01:29:26,172
S2: easily grasp. but a lot of security is known solutions.

1641
01:29:26,732 --> 01:29:30,932
S2: They're just not good solutions. Like, let's say, password theft.

1642
01:29:30,932 --> 01:29:33,532
S2: Your password gets stolen. What's the everyone's solution to that

1643
01:29:33,532 --> 01:29:37,492
S2: two factor authentication? You send an SMS to their phone. Um,

1644
01:29:37,932 --> 01:29:41,892
S2: why doesn't everything have two factor authentication? Well, because there's

1645
01:29:41,892 --> 01:29:45,532
S2: problems with that. Okay. Let's say we choose phone based

1646
01:29:45,532 --> 01:29:48,332
S2: two factor authentication. What if I lose my phone?

1647
01:29:49,052 --> 01:29:49,532
S1: Sure.

1648
01:29:49,572 --> 01:29:51,252
S2: Now I have to go down to it and I

1649
01:29:51,252 --> 01:29:53,852
S2: have to get my password reset. And I'm losing productivity,

1650
01:29:54,132 --> 01:29:57,052
S2: and it has to somehow verify my identity and know

1651
01:29:57,052 --> 01:29:59,452
S2: that I'm not a scammer pretending to be me. So

1652
01:29:59,452 --> 01:30:02,292
S2: there's a huge productivity loss there. And then what happens

1653
01:30:02,292 --> 01:30:04,132
S2: when I pick up my phone to to get my

1654
01:30:04,172 --> 01:30:07,732
S2: toufar code? And I see there's a notification from my girlfriend.

1655
01:30:07,972 --> 01:30:09,972
S2: So I start talking to my girlfriend and we've just

1656
01:30:09,972 --> 01:30:12,732
S2: now lost a bunch of productivity. And then you amplify

1657
01:30:12,732 --> 01:30:17,572
S2: that over an entire office space and to whatever the

1658
01:30:17,572 --> 01:30:21,422
S2: cost of password theft was. Um, being hacked was you

1659
01:30:21,462 --> 01:30:24,982
S2: probably lost more in productivity loss. So while we have

1660
01:30:24,982 --> 01:30:29,502
S2: existing solutions, we actually don't have good solutions. Everything is

1661
01:30:29,502 --> 01:30:33,381
S2: just trying to get the best possible, uh, the best

1662
01:30:33,382 --> 01:30:37,862
S2: possible idea for our specific scenario, which is really what

1663
01:30:37,902 --> 01:30:40,822
S2: cybersecurity is about. And that is the critical thinking. That's

1664
01:30:40,822 --> 01:30:43,662
S2: the intelligence. And I can just go, are your passwords

1665
01:30:43,662 --> 01:30:47,622
S2: are being stolen? Use Toofar. Okay. What? Toofar. Yeah. How

1666
01:30:47,622 --> 01:30:49,062
S2: do I handle password reset?

1667
01:30:49,502 --> 01:30:52,342
S1: But that's not what an AI is doing. Okay. So

1668
01:30:52,382 --> 01:30:56,022
S1: modern AIS like because I do this with risk assessment

1669
01:30:56,182 --> 01:30:58,381
S1: all the time, I do. Yeah. To answer your question,

1670
01:30:58,382 --> 01:31:01,702
S1: I'm doing this all the time for actual customers when

1671
01:31:01,702 --> 01:31:04,702
S1: you give the thing proper context. So if you give

1672
01:31:04,702 --> 01:31:06,862
S1: it the thing that look toofar doesn't work in this

1673
01:31:06,862 --> 01:31:09,742
S1: situation because of this, um, you don't even have to

1674
01:31:09,742 --> 01:31:12,342
S1: give it that. You could just say, this situation exists.

1675
01:31:12,382 --> 01:31:15,622
S1: This situation exists. We have this complexity over here. We

1676
01:31:15,622 --> 01:31:17,822
S1: have this complexity over here. What do you think the

1677
01:31:17,822 --> 01:31:21,032
S1: security control should be. It will say things like, well,

1678
01:31:21,592 --> 01:31:24,392
S1: based on your current state of your cloud infrastructure and

1679
01:31:24,392 --> 01:31:26,232
S1: based on the fact that you do business in France

1680
01:31:26,232 --> 01:31:28,192
S1: and based on the fact that tufa won't work for

1681
01:31:28,192 --> 01:31:32,711
S1: this and you can't actually do SMS because, uh, whatever,

1682
01:31:32,712 --> 01:31:36,312
S1: it's it's too easy to do, um, spoofing of, you know,

1683
01:31:36,592 --> 01:31:38,792
S1: cell phones. So we're going to rule that one out,

1684
01:31:38,792 --> 01:31:42,472
S1: especially in this scenario because of this jurisdiction whatever. And

1685
01:31:42,472 --> 01:31:44,752
S1: you start giving it the more context you give it,

1686
01:31:44,752 --> 01:31:48,192
S1: the more it's going to navigate all those special situations

1687
01:31:48,192 --> 01:31:50,631
S1: just the way that that you would or that I would.

1688
01:31:51,072 --> 01:31:52,792
S1: My argument to you is that when you look at

1689
01:31:52,792 --> 01:31:56,112
S1: an average pen test report and again, we can't use

1690
01:31:56,152 --> 01:31:59,232
S1: me or you or some other person who's been doing

1691
01:31:59,232 --> 01:32:01,952
S1: this forever as the example, because, yeah, maybe we can

1692
01:32:01,952 --> 01:32:04,192
S1: come up with a novel thing that's never been thought

1693
01:32:04,192 --> 01:32:06,712
S1: of before. But if you go take all the pen

1694
01:32:06,712 --> 01:32:09,712
S1: test reports produced last week and we go step by

1695
01:32:09,712 --> 01:32:12,631
S1: step through them, what kind of novelty is in there?

1696
01:32:12,632 --> 01:32:15,712
S1: What kind of new solutions are being proposed? Aren't they

1697
01:32:15,712 --> 01:32:19,482
S1: mostly saying, um, this thing was wide open. You've got

1698
01:32:19,482 --> 01:32:22,002
S1: a config problem here. Uh, that config should not be

1699
01:32:22,002 --> 01:32:23,881
S1: in that way. Oh, by the way, you should review

1700
01:32:23,922 --> 01:32:26,602
S1: your changes before you actually publish them. Oh, by the way,

1701
01:32:26,602 --> 01:32:29,842
S1: you have, you know, your credentials are open. Uh, like,

1702
01:32:29,882 --> 01:32:32,442
S1: it's going to be very common stuff. Like you've got

1703
01:32:32,482 --> 01:32:34,362
S1: to patch this stuff, you've got to patch.

1704
01:32:35,602 --> 01:32:37,042
S2: But who's giving it the context?

1705
01:32:38,922 --> 01:32:41,362
S1: The point is people are being paid to do this work.

1706
01:32:41,402 --> 01:32:43,282
S1: They don't have to give any context. They can just

1707
01:32:43,282 --> 01:32:45,522
S1: give the list of vulnerabilities. They can give the list

1708
01:32:45,522 --> 01:32:47,402
S1: of vulnerabilities. They can say, I was able to get

1709
01:32:47,402 --> 01:32:49,842
S1: in through this. Um, you have to patch that system

1710
01:32:49,842 --> 01:32:52,442
S1: because it's critical. And the people read it and they're like,

1711
01:32:52,442 --> 01:32:54,882
S1: oh yeah, yeah, it's a critical system. This is what

1712
01:32:54,882 --> 01:32:57,921
S1: my team has been telling me for 14 years. Everything

1713
01:32:57,922 --> 01:32:59,881
S1: in this report, my team has already told me a

1714
01:32:59,882 --> 01:33:02,762
S1: million times, I just needed to get this report because

1715
01:33:02,762 --> 01:33:05,802
S1: now I can justify the security spend. Like how much

1716
01:33:05,802 --> 01:33:11,522
S1: of novel, quote unquote novel things that very highly paid

1717
01:33:11,522 --> 01:33:15,922
S1: people are being paid for is actually exactly the thing

1718
01:33:15,922 --> 01:33:19,682
S1: that you described before, which is it's very much known.

1719
01:33:19,722 --> 01:33:22,962
S1: Things just applied in a specific way for a specific customer.

1720
01:33:23,242 --> 01:33:26,522
S1: It's not always the same. It's non-deterministic because it's a

1721
01:33:26,522 --> 01:33:30,842
S1: particular tester doing it. But the output looks remarkably similar,

1722
01:33:31,362 --> 01:33:34,002
S1: almost identical to the output of another pentester doing the

1723
01:33:34,002 --> 01:33:34,802
S1: same exact thing.

1724
01:33:36,162 --> 01:33:39,362
S2: Yeah, but now we're switching from the reports to the

1725
01:33:39,362 --> 01:33:42,802
S2: decision making. Because, Ali, you said about the context, like, okay,

1726
01:33:42,842 --> 01:33:45,042
S2: the company is operating in this jurisdiction, blah blah, blah, blah,

1727
01:33:45,042 --> 01:33:48,002
S2: blah blah blah, but who is giving it that context? Like,

1728
01:33:48,042 --> 01:33:51,162
S2: surely there has to be a person who is gathering

1729
01:33:51,162 --> 01:33:53,482
S2: and giving the model that context.

1730
01:33:53,682 --> 01:33:58,082
S1: Sure, sure. The agent, an agent is crawling all the docs.

1731
01:33:58,442 --> 01:34:01,362
S1: It's having a chat, interviews with people. It's doing voice

1732
01:34:01,362 --> 01:34:04,562
S1: calls and interviewing people. Like these are this is all

1733
01:34:04,562 --> 01:34:09,002
S1: just like common common stakes for like something an AI

1734
01:34:09,042 --> 01:34:11,122
S1: could do. It could do an interview, it could talk

1735
01:34:11,122 --> 01:34:14,412
S1: to you and extract information. It could read an entire wiki.

1736
01:34:14,452 --> 01:34:17,251
S1: It could read every doctrine on, uh, you know, Google

1737
01:34:17,252 --> 01:34:20,492
S1: Docs and Confluence. Like, it can go and gather that context.

1738
01:34:21,292 --> 01:34:23,092
S2: You see, I don't think it can. Like, I don't

1739
01:34:23,092 --> 01:34:26,892
S2: think it has the capability or the intelligence to gather

1740
01:34:26,892 --> 01:34:30,172
S2: that kind of context. I think maybe there is an

1741
01:34:30,172 --> 01:34:34,012
S2: argument that in the future it could, uh, but given

1742
01:34:34,012 --> 01:34:37,852
S2: any current generation model I have used, we're not even

1743
01:34:37,852 --> 01:34:39,852
S2: close to being able to gather the context. We can't

1744
01:34:39,852 --> 01:34:43,052
S2: even answer the questions. Like I can, I've had multiple

1745
01:34:43,052 --> 01:34:45,532
S2: times where I've asked the AI stuff, and because it's

1746
01:34:45,532 --> 01:34:48,212
S2: been trained on all of the past knowledge, which, as

1747
01:34:48,212 --> 01:34:50,892
S2: I mentioned, is constantly changing, it'll be like, oh, we

1748
01:34:50,892 --> 01:34:54,131
S2: should do password rotation. And then like we've disproven password

1749
01:34:54,132 --> 01:34:58,131
S2: rotation like 20 years ago. Like, why are you suggesting this? Um,

1750
01:34:58,492 --> 01:35:02,412
S2: and so we're right now, we're still running into the

1751
01:35:02,412 --> 01:35:05,612
S2: problem of, oh, the best practices have changed in the

1752
01:35:05,612 --> 01:35:08,251
S2: last five years. Maybe it's even the last one year

1753
01:35:08,492 --> 01:35:11,662
S2: and the model hasn't even updated to to account for that.

1754
01:35:11,662 --> 01:35:14,542
S2: And now you're saying we could have the model that

1755
01:35:14,542 --> 01:35:17,462
S2: can barely even keep up with the current state of

1756
01:35:17,462 --> 01:35:21,302
S2: the world? Now, actually, like interviewing people and like profiling

1757
01:35:21,302 --> 01:35:25,182
S2: my company and like the local laws and the documentations.

1758
01:35:25,182 --> 01:35:28,982
S2: And I think it may be in ten, 20 years

1759
01:35:28,982 --> 01:35:32,462
S2: is something that I could do. Uh, but there is

1760
01:35:32,462 --> 01:35:36,142
S2: no current generation model I've seen that even comes close

1761
01:35:36,422 --> 01:35:37,662
S2: to being able to do that.

1762
01:35:37,702 --> 01:35:40,422
S1: But but it's not a model, though, Marcus. It's like.

1763
01:35:40,422 --> 01:35:45,222
S1: It's not like it's not like, um, O3 or some

1764
01:35:45,222 --> 01:35:48,942
S1: model has has capabilities like that. It's more like we

1765
01:35:48,982 --> 01:35:52,542
S1: hired this, I don't know. Securus. This Securus company comes

1766
01:35:52,542 --> 01:35:54,782
S1: out and it's like our model Securus is the smartest

1767
01:35:54,782 --> 01:35:58,302
S1: model ever, and it will do a security assessment for

1768
01:35:58,302 --> 01:36:01,422
S1: you when you give the thing to Securus. It's actually

1769
01:36:01,422 --> 01:36:03,862
S1: smoke and mirrors. It's actually a whole bunch of agents

1770
01:36:03,862 --> 01:36:07,382
S1: sitting behind it. So it's really this really smart orchestrator

1771
01:36:07,382 --> 01:36:11,032
S1: in the front. It has a million different gatherers that

1772
01:36:11,032 --> 01:36:13,872
S1: go and answer questions that haven't been talked about. So

1773
01:36:13,872 --> 01:36:17,352
S1: you handed us this thing that's like, hey, we're building this, uh,

1774
01:36:17,712 --> 01:36:21,672
S1: you know, whatever it is, um, this special new application.

1775
01:36:21,832 --> 01:36:24,312
S1: And the top level agent is like. I have no

1776
01:36:24,312 --> 01:36:27,032
S1: idea what that is. Let me go crawl all the documentation.

1777
01:36:27,032 --> 01:36:31,432
S1: So it's all smoke and mirrors. Orchestration of multiple things

1778
01:36:31,432 --> 01:36:34,352
S1: going together. It's like, hey, what about this application? Well,

1779
01:36:34,392 --> 01:36:37,752
S1: it's being developed by so-and-so team. Okay, who's on that team? Okay,

1780
01:36:37,752 --> 01:36:39,672
S1: here's the list of people. Okay. Set up meetings and

1781
01:36:39,672 --> 01:36:42,352
S1: go talk to them or send them this form. Have

1782
01:36:42,352 --> 01:36:43,832
S1: them fill it out and bring it back to us,

1783
01:36:43,832 --> 01:36:48,072
S1: and we'll parse that. So it's nothing about those individual

1784
01:36:48,072 --> 01:36:52,152
S1: steps are difficult. What's difficult is just combining all those

1785
01:36:52,152 --> 01:36:55,392
S1: steps up into this overall product, so that it looks

1786
01:36:55,392 --> 01:36:57,832
S1: like with a customer, when you're talking to it, it's

1787
01:36:57,832 --> 01:37:00,712
S1: figuring it all out when behind the scenes, it's like

1788
01:37:00,712 --> 01:37:03,912
S1: you were saying before, it's a whole bunch of subtasks

1789
01:37:03,912 --> 01:37:07,312
S1: that are being doled out to this thing. And ultimately,

1790
01:37:08,202 --> 01:37:11,842
S1: that's just orchestration. That's just like all these little pieces

1791
01:37:11,842 --> 01:37:15,762
S1: being combined to produce the illusion of an overall intelligence.

1792
01:37:16,322 --> 01:37:20,082
S1: And my argument to you is that that capability is

1793
01:37:20,082 --> 01:37:23,402
S1: going to be functionally the same as the capability of

1794
01:37:23,402 --> 01:37:25,282
S1: you hired someone else to do it. You hire an

1795
01:37:25,282 --> 01:37:27,282
S1: actual person to do it. Because at the end of

1796
01:37:27,282 --> 01:37:28,722
S1: the day, it's going to be like the Turing test.

1797
01:37:28,722 --> 01:37:30,722
S1: You're going to look at two different pen test reports,

1798
01:37:31,402 --> 01:37:33,881
S1: and one was just you or me doing this assessment

1799
01:37:33,882 --> 01:37:36,362
S1: for three weeks where we had to do the interviews,

1800
01:37:36,362 --> 01:37:40,282
S1: and this other one launched off 780 agents and came back.

1801
01:37:40,282 --> 01:37:42,922
S1: And because the cost of inference is so short, you know,

1802
01:37:42,962 --> 01:37:45,802
S1: the whole thing cost them $13, whereas they would have

1803
01:37:45,802 --> 01:37:47,442
S1: paid you or me 90 grand to do it.

1804
01:37:49,282 --> 01:37:51,602
S2: I think that only works if you can build the

1805
01:37:51,602 --> 01:37:54,762
S2: system with the system, because if you're making all of

1806
01:37:54,762 --> 01:37:56,921
S2: these sub models to handle all of these, like niche

1807
01:37:56,962 --> 01:38:02,122
S2: edge cases and these like specific domains of expertise, you

1808
01:38:02,122 --> 01:38:05,242
S2: need people in those domains of expertise to build those systems.

1809
01:38:05,532 --> 01:38:07,732
S2: So now essentially all you have is rather than a

1810
01:38:07,732 --> 01:38:11,052
S2: bunch of pen test companies, you have the pen test company,

1811
01:38:11,212 --> 01:38:14,052
S2: which is every pen tester now working on building some

1812
01:38:14,052 --> 01:38:17,412
S2: model that can, uh, roughly emulate what a human pen

1813
01:38:17,412 --> 01:38:20,372
S2: tester would do. Um, but that I would not see

1814
01:38:20,412 --> 01:38:23,212
S2: as a job replacement. I would see as a job shift. Uh,

1815
01:38:23,212 --> 01:38:27,131
S2: you would essentially just end up with this horrible late

1816
01:38:27,132 --> 01:38:32,532
S2: stage capitalist bullshit where everyone works for a single AI company. Um,

1817
01:38:32,532 --> 01:38:34,252
S2: and no one has been replaced. We still need the

1818
01:38:34,252 --> 01:38:36,932
S2: same number of people just there, building models to automate

1819
01:38:36,932 --> 01:38:40,972
S2: things instead of doing the things. Um, but I don't

1820
01:38:40,972 --> 01:38:45,132
S2: see it as such a productivity boost that we would

1821
01:38:45,172 --> 01:38:48,172
S2: actually see like massive amounts of employees being lost. We

1822
01:38:48,172 --> 01:38:51,612
S2: would just see them going over to these, uh, these AGI,

1823
01:38:51,652 --> 01:38:54,052
S2: I guess we would call it companies to make these

1824
01:38:54,052 --> 01:38:57,372
S2: fake Agis by like, cobbling together a bunch of, uh,

1825
01:38:57,492 --> 01:39:00,972
S2: agents and sub models. But the amount of humans that

1826
01:39:00,972 --> 01:39:04,022
S2: you would need to build and maintain such a system

1827
01:39:04,022 --> 01:39:07,622
S2: would be like, phenomenal. It would be, uh, it would

1828
01:39:07,622 --> 01:39:10,182
S2: be like the biggest company on earth. So I wouldn't

1829
01:39:10,182 --> 01:39:13,102
S2: see that as people are going to be losing their jobs.

1830
01:39:13,102 --> 01:39:14,982
S2: I would see it as problematic because now you just

1831
01:39:14,982 --> 01:39:18,942
S2: have a monopoly that just it does everything. Um, and

1832
01:39:18,942 --> 01:39:22,062
S2: that's going to be horrible for working conditions. Um, but

1833
01:39:22,062 --> 01:39:24,062
S2: I wouldn't see that as a risk of, okay, now

1834
01:39:24,062 --> 01:39:26,902
S2: pen testers are obsolete. There's no more pen testers. Uh, okay.

1835
01:39:26,942 --> 01:39:29,902
S2: Medical professionals obsolete. There's no more medical professionals. It would

1836
01:39:29,902 --> 01:39:33,102
S2: just be. Everyone would just be like a medical professional

1837
01:39:33,102 --> 01:39:34,862
S2: working for an AI company.

1838
01:39:35,982 --> 01:39:39,822
S1: Yeah, but aren't these all just modules, though? So you

1839
01:39:39,822 --> 01:39:42,142
S1: have a module for going to collect more information when

1840
01:39:42,142 --> 01:39:45,422
S1: you don't know something, you have a module for summarizing

1841
01:39:45,422 --> 01:39:48,142
S1: that information and bringing it into the context. You have

1842
01:39:48,142 --> 01:39:53,621
S1: a module for um, you know. Net learning some new thing, right?

1843
01:39:54,222 --> 01:39:56,702
S1: I mean, let's go back to the cardiologist. Think of

1844
01:39:56,702 --> 01:40:00,662
S1: how much schooling that cardiologist had to have, right. And

1845
01:40:00,662 --> 01:40:04,671
S1: how different the different scenarios are. I would argue it's

1846
01:40:04,672 --> 01:40:10,352
S1: not that much more complex than, um, or less complex

1847
01:40:10,352 --> 01:40:12,792
S1: than pen testing. So I think what you have a

1848
01:40:12,792 --> 01:40:15,872
S1: module for understanding pen testing. Really well, a lot of

1849
01:40:15,872 --> 01:40:19,512
S1: these concepts are general enough. They are abstracted enough that

1850
01:40:19,512 --> 01:40:24,072
S1: when an AI understands them deeply enough, it is then

1851
01:40:24,232 --> 01:40:27,752
S1: using your analogy here or your concept, which I really like,

1852
01:40:28,072 --> 01:40:31,632
S1: it becomes knowledge. It becomes knowledge as opposed to intelligence.

1853
01:40:32,032 --> 01:40:36,831
S1: I really like this distinction. So you when it understands.

1854
01:40:36,832 --> 01:40:39,552
S1: And by the way, I would argue that we've already

1855
01:40:39,712 --> 01:40:43,152
S1: reached this with security. Um, it's already really, really good

1856
01:40:43,152 --> 01:40:47,112
S1: at cybersecurity. We have. Let me give you an example here. Okay.

1857
01:40:47,112 --> 01:40:50,832
S1: Let me give you an example of an extremely dynamic situation. Um,

1858
01:40:51,552 --> 01:40:57,512
S1: right now on hacker one, a completely automated, completely automated

1859
01:40:58,032 --> 01:41:01,802
S1: AI system has the As the number one spot. That

1860
01:41:01,802 --> 01:41:06,282
S1: means it's dealing with all sorts of random situations it's

1861
01:41:06,282 --> 01:41:10,041
S1: never seen before. It's doing its own research. It's finding

1862
01:41:10,042 --> 01:41:13,242
S1: new information on new technologies that's never seen before. It's

1863
01:41:13,242 --> 01:41:16,882
S1: launching all the attacks. It's actually showing proof and evidence

1864
01:41:17,482 --> 01:41:20,362
S1: that it can exploit it. It's actually submitting the reports

1865
01:41:20,762 --> 01:41:25,642
S1: and getting back the points with humans hand off. So

1866
01:41:25,682 --> 01:41:29,722
S1: I mean, if that's not the level of complexity that

1867
01:41:29,722 --> 01:41:33,042
S1: could replace a knowledge worker and let alone do a

1868
01:41:33,042 --> 01:41:37,282
S1: pen test, I mean, we already have all the evidence here,

1869
01:41:37,322 --> 01:41:41,082
S1: like the models are already good at security. And if

1870
01:41:41,082 --> 01:41:44,122
S1: you could just gather more knowledge to put into the context,

1871
01:41:44,522 --> 01:41:45,442
S1: that's all you need.

1872
01:41:47,482 --> 01:41:49,322
S2: But I think you have to ask yourself, like how

1873
01:41:49,322 --> 01:41:53,122
S2: many people were working on that model? Like, is this

1874
01:41:53,122 --> 01:41:55,722
S2: now a model that is just on autopilot forever that

1875
01:41:55,722 --> 01:41:59,172
S2: requires no human interaction whatsoever. And it just sits. And

1876
01:41:59,172 --> 01:42:02,412
S2: it does pen tests. Or does someone keep having to

1877
01:42:02,452 --> 01:42:06,332
S2: update systems behind the scenes because, uh, it kind of

1878
01:42:06,372 --> 01:42:07,732
S2: reminds me of that. I don't remember the name of

1879
01:42:07,732 --> 01:42:10,732
S2: the specific company, but there was this one company that was, uh,

1880
01:42:10,732 --> 01:42:13,172
S2: they were selling Vibe coding, and they went bankrupt after

1881
01:42:13,172 --> 01:42:16,291
S2: it was found that there was no I it was

1882
01:42:16,292 --> 01:42:18,412
S2: just a bunch of people in a, in an impoverished

1883
01:42:18,412 --> 01:42:21,132
S2: country writing the code for them. And essentially, this is

1884
01:42:21,132 --> 01:42:23,572
S2: all you're doing is an abstraction of this is you're

1885
01:42:23,572 --> 01:42:26,252
S2: having you don't know how many people are behind the

1886
01:42:26,252 --> 01:42:30,131
S2: scenes working the AI. Um, and I think that's kind

1887
01:42:30,132 --> 01:42:32,892
S2: of where the metrics get to get a bit skewed,

1888
01:42:32,892 --> 01:42:35,772
S2: because when we say someone was replaced with AI, we

1889
01:42:35,812 --> 01:42:38,852
S2: don't count how many jobs were gained by the creation

1890
01:42:38,852 --> 01:42:42,092
S2: of that AI. We're just like, oh, Jeff has been

1891
01:42:42,092 --> 01:42:45,852
S2: replaced with Pentester AI. Um, that means one person has

1892
01:42:45,852 --> 01:42:48,171
S2: now lost their job to AI, and what you don't

1893
01:42:48,172 --> 01:42:50,892
S2: see is on the back end. There's 500 people making this.

1894
01:42:50,892 --> 01:42:53,932
S2: I actually do anything at all. Um, so I think

1895
01:42:53,932 --> 01:42:57,222
S2: that's kind of the that's my issue with this argument

1896
01:42:57,222 --> 01:43:02,541
S2: of like the job replacement, that people being unemployed because, um,

1897
01:43:02,582 --> 01:43:05,542
S2: at least from what I'm seeing, the amount of work

1898
01:43:05,542 --> 01:43:08,782
S2: that goes into actually maintaining and running these systems is

1899
01:43:08,782 --> 01:43:12,262
S2: equivalent to doing it manually still. So I don't see

1900
01:43:12,702 --> 01:43:14,702
S2: any world in which we're going to see something as

1901
01:43:14,702 --> 01:43:18,262
S2: crazy as like 90% job loss. Um, and then that

1902
01:43:18,262 --> 01:43:22,782
S2: also assumes that everything is capped, right? That maybe pen

1903
01:43:22,782 --> 01:43:24,902
S2: testing is capped, like maybe there is only so much

1904
01:43:24,902 --> 01:43:28,102
S2: security we can do. But then there's other industries where

1905
01:43:28,102 --> 01:43:30,622
S2: it's like, oh, we gained some productivity. We can do

1906
01:43:30,622 --> 01:43:33,662
S2: more now. Um, so I would think that we would

1907
01:43:33,662 --> 01:43:37,742
S2: just see, uh, assuming we can get an AI model

1908
01:43:37,982 --> 01:43:39,542
S2: that just for the sake of argument, let's say it

1909
01:43:39,542 --> 01:43:42,262
S2: doesn't require any people. Like somehow we built an AI

1910
01:43:42,302 --> 01:43:45,822
S2: that just pilots itself, doesn't need any programmers, doesn't need

1911
01:43:45,822 --> 01:43:48,302
S2: any tweaks. It just works on its own. Would we

1912
01:43:48,342 --> 01:43:51,102
S2: not just take those systems and then focus into doing

1913
01:43:51,102 --> 01:43:55,702
S2: something more like, um, maybe we go in mined materials

1914
01:43:55,702 --> 01:43:59,262
S2: from asteroids. Maybe, uh, we start sending probes to, uh,

1915
01:43:59,262 --> 01:44:01,662
S2: to other planets within the galaxy. Like, there's always more

1916
01:44:01,662 --> 01:44:04,542
S2: work to do. Um, and the idea of, like, job

1917
01:44:04,542 --> 01:44:07,502
S2: replacement and job loss kind of hinges on this idea

1918
01:44:07,662 --> 01:44:11,662
S2: that there is finite work, there's finite productivity, um, and

1919
01:44:11,662 --> 01:44:14,462
S2: that a, the productivity isn't just being transferred to the

1920
01:44:14,462 --> 01:44:16,862
S2: AI side, and those people are just doing the same

1921
01:44:16,862 --> 01:44:19,142
S2: thing they were already doing. But for an AI company

1922
01:44:19,502 --> 01:44:21,782
S2: and that like there is just a ceiling and we

1923
01:44:21,862 --> 01:44:23,902
S2: hit this ceiling, it's like, okay, we don't need people anymore.

1924
01:44:23,942 --> 01:44:26,622
S2: Like we've done all the things we need to do. Um,

1925
01:44:26,622 --> 01:44:28,421
S2: and I just I don't see either.

1926
01:44:29,422 --> 01:44:32,942
S1: Yeah. So, so, so I think let's say it costs.

1927
01:44:32,982 --> 01:44:35,382
S1: Let's say it took, um, I actually don't know how

1928
01:44:35,382 --> 01:44:37,422
S1: big this company is that actually did the thing that

1929
01:44:37,422 --> 01:44:41,022
S1: I'm describing. I don't know how much, you know, human

1930
01:44:41,022 --> 01:44:43,902
S1: effort went into building the agents, but the thing running is,

1931
01:44:43,942 --> 01:44:47,222
S1: is not involving humans. It's actually the agents running that

1932
01:44:47,222 --> 01:44:49,342
S1: are doing the thing. Now, you have a good point

1933
01:44:49,342 --> 01:44:51,942
S1: of like, well, those agents start to rot, and we

1934
01:44:51,942 --> 01:44:54,512
S1: have to maintain the agent. Sure, but let's say that's

1935
01:44:54,512 --> 01:44:57,752
S1: only ten people or 100 people. The point is that

1936
01:44:57,752 --> 01:45:02,472
S1: is now replaceable infrastructure that the entire planet can use.

1937
01:45:02,912 --> 01:45:05,711
S1: So the planet can go and now run this, this

1938
01:45:05,712 --> 01:45:09,832
S1: company's application and pointed at targets, and it could find

1939
01:45:09,832 --> 01:45:15,632
S1: vulnerabilities and submit reports. Now that, um, the scaling of

1940
01:45:15,672 --> 01:45:19,232
S1: that infrastructure there to run that I don't think compares

1941
01:45:19,232 --> 01:45:23,992
S1: at all to, let's say, um, a million tests, let's

1942
01:45:23,992 --> 01:45:26,552
S1: say a million pen tests, a million bug bounties or

1943
01:45:26,552 --> 01:45:29,232
S1: pen tests or whatever. You're going to stick this thing

1944
01:45:29,232 --> 01:45:33,752
S1: on scaling up to how many humans are going to

1945
01:45:33,792 --> 01:45:36,552
S1: run that thing and how much they're going to cost,

1946
01:45:37,072 --> 01:45:41,512
S1: versus a million instances of this AI, especially given the

1947
01:45:41,552 --> 01:45:43,832
S1: I mean, look at the inference costs of AI and

1948
01:45:43,832 --> 01:45:45,831
S1: how they come down over like the last two years.

1949
01:45:45,872 --> 01:45:50,402
S1: I think we're paying like 0.01% or something of the

1950
01:45:50,402 --> 01:45:53,121
S1: inference cost of what we were paying to two years ago.

1951
01:45:53,162 --> 01:45:57,082
S1: It's something extraordinarily like that. So there's no reason to

1952
01:45:57,122 --> 01:45:59,362
S1: believe that that's not going to keep falling. Who knows

1953
01:45:59,362 --> 01:46:02,282
S1: what the bottom level is? But the marginal cost for

1954
01:46:02,282 --> 01:46:06,162
S1: doing a pen test using this automated system is going

1955
01:46:06,162 --> 01:46:09,602
S1: to be way lower than adding yet another tester. Not

1956
01:46:09,602 --> 01:46:12,762
S1: only that, but the testers, they're humans, they move on.

1957
01:46:12,762 --> 01:46:15,642
S1: They stop being testers and move on to they become

1958
01:46:15,642 --> 01:46:19,402
S1: managers or whatever. And so I think the scalability there's

1959
01:46:19,442 --> 01:46:20,682
S1: like no comparison.

1960
01:46:24,322 --> 01:46:28,042
S2: Um, I could like, given your assumptions, I could maybe

1961
01:46:28,082 --> 01:46:31,722
S2: see that, um, I don't know about, uh, obviously the

1962
01:46:31,722 --> 01:46:33,922
S2: first thing I would steer away from is the whole

1963
01:46:33,922 --> 01:46:37,002
S2: prior performance, future results thing. Like, if the costs are

1964
01:46:37,002 --> 01:46:38,842
S2: falling that we can't just expect, they're going to keep

1965
01:46:38,842 --> 01:46:40,202
S2: falling forever. Sure, sure.

1966
01:46:40,322 --> 01:46:42,881
S1: Um, that's why I said I know where the bottom is.

1967
01:46:43,602 --> 01:46:45,682
S2: In fact, we may actually see the opposite happen as

1968
01:46:45,682 --> 01:46:49,652
S2: we start to, uh, as, like, resources start to get constrained,

1969
01:46:49,652 --> 01:46:51,652
S2: we might actually see the costs start going back up

1970
01:46:51,652 --> 01:46:54,812
S2: because that's true. Silicon is pretty hard to get. Um,

1971
01:46:54,812 --> 01:46:58,131
S2: energy is um, we're kind of hitting our cap on

1972
01:46:58,132 --> 01:46:59,852
S2: energy right now. So we're having to build out more

1973
01:46:59,852 --> 01:47:02,732
S2: power plants, and those will end up getting factored into

1974
01:47:02,732 --> 01:47:06,612
S2: the costs of running these models. Um, but with something

1975
01:47:06,612 --> 01:47:11,092
S2: like that, say Pentesting, I could see in that specific example. Yeah,

1976
01:47:11,092 --> 01:47:14,492
S2: I think we could probably automate it to make it cheaper. Um,

1977
01:47:14,652 --> 01:47:18,211
S2: that's not because I think there's something fundamentally wrong with, uh,

1978
01:47:18,692 --> 01:47:22,211
S2: with people or the AI is fundamentally good. It's because

1979
01:47:22,212 --> 01:47:24,932
S2: the industry is kind of a scam. We go and

1980
01:47:24,932 --> 01:47:28,852
S2: we bill like $100,000 for some dude to, like, spend

1981
01:47:28,852 --> 01:47:33,372
S2: 15 minutes running Nessus or something. Um, so I think

1982
01:47:33,652 --> 01:47:36,852
S2: in a lot of those spaces we could see, uh,

1983
01:47:37,692 --> 01:47:41,092
S2: I guess full AI automation, but a lot of that

1984
01:47:41,092 --> 01:47:44,532
S2: stuff is really just it was always automatable like, you

1985
01:47:44,532 --> 01:47:46,742
S2: could have scripted a lot of that. It's just that

1986
01:47:46,742 --> 01:47:50,381
S2: no one was doing it. The real work, which is

1987
01:47:50,382 --> 01:47:54,542
S2: not being AI automated, is the stuff that is actually dynamic. Um,

1988
01:47:54,582 --> 01:47:56,662
S2: because like one of the things with bug bounties is

1989
01:47:56,662 --> 01:48:00,862
S2: it is very, very, uh replicatable. Like, you can reproduce it. Um,

1990
01:48:00,902 --> 01:48:04,662
S2: there's only a very few classes of vulnerabilities. And if

1991
01:48:04,662 --> 01:48:07,662
S2: you go for something like low hanging fruit, like cross-site scripting,

1992
01:48:07,902 --> 01:48:09,662
S2: you could write a Python script to go and find

1993
01:48:09,662 --> 01:48:11,982
S2: a million of those. And a lot of the older

1994
01:48:11,982 --> 01:48:15,102
S2: pen testers did, sorry, bug bounty testers did do that.

1995
01:48:15,342 --> 01:48:18,542
S2: They were also automated, just not with AI. So that

1996
01:48:18,542 --> 01:48:21,342
S2: is a, uh, like that is something that has always

1997
01:48:21,342 --> 01:48:24,222
S2: been automatable. Like, I don't think AI has really made

1998
01:48:24,222 --> 01:48:27,902
S2: any amazing breakthroughs there. Um, it might have maybe increased

1999
01:48:27,902 --> 01:48:31,662
S2: the scope of vulnerabilities that are automatable to a, I

2000
01:48:31,662 --> 01:48:33,982
S2: would say a very small degree, but I don't think

2001
01:48:33,982 --> 01:48:36,142
S2: we're going to start seeing like large language models who

2002
01:48:36,142 --> 01:48:40,182
S2: are like writing novel exploits. Um, I think we're just

2003
01:48:40,182 --> 01:48:43,581
S2: going to see stuff that like, um, was already automated,

2004
01:48:43,582 --> 01:48:49,032
S2: like fuzzing, Like fuzzing. I guess you could call automation like, it's, uh,

2005
01:48:49,072 --> 01:48:51,392
S2: you basically just the equivalent of shaking a tree and

2006
01:48:51,392 --> 01:48:54,592
S2: seeing what falls out. Uh, you could put slap AI

2007
01:48:54,592 --> 01:48:56,592
S2: on that and say, oh, look, we have AI, automated

2008
01:48:56,592 --> 01:48:59,912
S2: pen testing, but a lot of those things were already

2009
01:48:59,912 --> 01:49:06,472
S2: automated or automatable. Um, so I think I'd caution like the, uh,

2010
01:49:06,472 --> 01:49:09,671
S2: going from, hey, look, this thing is doing very well

2011
01:49:09,792 --> 01:49:15,552
S2: in an automated sense to cybersecurity is already like, automatable. Um,

2012
01:49:15,592 --> 01:49:19,631
S2: because I, I personally stay away from pen testing. Um,

2013
01:49:19,632 --> 01:49:21,832
S2: but I would say that's a very, very small portion

2014
01:49:21,832 --> 01:49:25,592
S2: of cybersecurity. And most of the parts of cybersecurity, uh,

2015
01:49:25,832 --> 01:49:29,712
S2: that like really do need doing isn't automatable. Like, it's

2016
01:49:29,712 --> 01:49:32,152
S2: not stuff we could just set, uh, pen test bot

2017
01:49:32,192 --> 01:49:35,952
S2: on and it'll it'll fix the system. It's human problems. It's, uh,

2018
01:49:35,992 --> 01:49:40,112
S2: psychological problems. Um, it's like actual complex problems that require

2019
01:49:40,112 --> 01:49:43,242
S2: critical thinking versus. Let's scan this code and look for

2020
01:49:43,242 --> 01:49:48,682
S2: XSS or integer overflows. Um, I think the what seems

2021
01:49:48,682 --> 01:49:51,921
S2: so significant there is that we just weren't doing it.

2022
01:49:51,962 --> 01:49:54,362
S2: Like we have had the capability to automate a lot

2023
01:49:54,362 --> 01:49:56,961
S2: of this for a lot of time, but we've just

2024
01:49:56,962 --> 01:49:58,002
S2: decided not to.

2025
01:49:58,122 --> 01:50:01,642
S1: Um, doesn't that doesn't that apply to hundreds of millions

2026
01:50:01,642 --> 01:50:02,242
S1: of jobs?

2027
01:50:04,162 --> 01:50:08,242
S2: Um, I don't know. That's the thing is, the issue

2028
01:50:08,242 --> 01:50:12,242
S2: is kind of that the people aren't doing the jobs. Um,

2029
01:50:12,442 --> 01:50:16,682
S2: like one thing I've noticed, uh, with vulnerability research is

2030
01:50:16,682 --> 01:50:20,122
S2: a lot of times what will happen is a company will, uh,

2031
01:50:20,122 --> 01:50:22,242
S2: they'll fix a vulnerability in their code base, like a

2032
01:50:22,242 --> 01:50:25,042
S2: very specific type of vulnerability, and they won't go and

2033
01:50:25,042 --> 01:50:27,322
S2: look for that same vulnerability in other parts of their

2034
01:50:27,322 --> 01:50:30,562
S2: own code base. 100% agree they will. Yeah. So they

2035
01:50:30,562 --> 01:50:35,522
S2: will fix one very specific instance of vulnerability. And could

2036
01:50:35,522 --> 01:50:37,602
S2: we make a large language model that goes and finds

2037
01:50:37,602 --> 01:50:40,242
S2: the rest? Sure. But they could they could have and

2038
01:50:40,242 --> 01:50:42,892
S2: they should have done that. And essentially all we're doing

2039
01:50:42,892 --> 01:50:46,092
S2: is we're making this huge, expensive model to pick up

2040
01:50:46,092 --> 01:50:48,772
S2: the slack. That was just we just decided we could

2041
01:50:48,812 --> 01:50:50,372
S2: have done it. We just decided not to.

2042
01:50:50,772 --> 01:50:53,172
S1: So. So yeah, this that's a perfect example of what

2043
01:50:53,172 --> 01:50:55,412
S1: I'm saying. So check this out. That thing I told

2044
01:50:55,412 --> 01:51:00,732
S1: you about that bacteriophages like for researchers, the best in

2045
01:51:00,732 --> 01:51:03,652
S1: the entire world. It turns out that they had a

2046
01:51:03,692 --> 01:51:06,852
S1: bias in their brain. They thought that, um, I'm going

2047
01:51:06,892 --> 01:51:08,252
S1: to mess this up because I don't know anything about

2048
01:51:08,252 --> 01:51:10,372
S1: the science, but there's something about a head and a

2049
01:51:10,372 --> 01:51:13,932
S1: tail of a bacteriophage. And like they combine in a

2050
01:51:13,932 --> 01:51:16,132
S1: certain way to be able to move and go spread.

2051
01:51:16,572 --> 01:51:18,412
S1: And so they were making this thing in their mind

2052
01:51:18,452 --> 01:51:20,292
S1: of like, well, it can't be this and it can't

2053
01:51:20,292 --> 01:51:22,612
S1: be this because heads can only go on tails the

2054
01:51:22,612 --> 01:51:25,492
S1: following way. So the I was like, can't you just

2055
01:51:25,492 --> 01:51:29,532
S1: do this instead? That's probably why it's happening. When they

2056
01:51:29,532 --> 01:51:31,292
S1: both looked at it, they were like, oh my God,

2057
01:51:31,292 --> 01:51:35,092
S1: that is so simple. So a mental human human bias

2058
01:51:35,732 --> 01:51:38,572
S1: caused them to miss this, which would have propelled science

2059
01:51:38,702 --> 01:51:41,742
S1: forward for so long? So it's exactly the same thing.

2060
01:51:41,982 --> 01:51:46,702
S1: So basically, did I make a discovery there? It actually

2061
01:51:46,702 --> 01:51:51,102
S1: just found a human error. But but at scale. So

2062
01:51:51,102 --> 01:51:53,102
S1: so now everyone in the in the world can use

2063
01:51:53,102 --> 01:51:56,022
S1: this model. And they can now do the same thing

2064
01:51:56,062 --> 01:52:01,461
S1: to all this hidden research that's sitting inside this, this, um, this, um,

2065
01:52:01,462 --> 01:52:04,421
S1: archive of, like, raw data that's just sitting there. This

2066
01:52:04,422 --> 01:52:07,982
S1: raw data, in my opinion, contains, hey, you know, if

2067
01:52:07,982 --> 01:52:10,622
S1: you put this molecule to to this thing in this

2068
01:52:10,622 --> 01:52:12,462
S1: part of the cell, it will actually just do this

2069
01:52:12,502 --> 01:52:15,782
S1: and like, it'll change aging completely. And anybody who sees

2070
01:52:15,782 --> 01:52:18,942
S1: that when the AI says it, they'll be like, that

2071
01:52:18,942 --> 01:52:23,342
S1: was obvious. That was completely obvious. So somebody like yourself

2072
01:52:23,342 --> 01:52:27,381
S1: could say, well, that wasn't actually intelligence. That was actually

2073
01:52:27,382 --> 01:52:30,382
S1: just noticing something and applying a rule. And what I'm

2074
01:52:30,382 --> 01:52:34,342
S1: saying is that that mentality is missing the point of

2075
01:52:34,342 --> 01:52:38,352
S1: the benefit of the AI, because whether it's pentesting or

2076
01:52:38,392 --> 01:52:43,072
S1: finding new drugs, or finding new interactions or helping couples

2077
01:52:43,392 --> 01:52:47,272
S1: or solving, you know, helping people get healthier heart, it

2078
01:52:47,272 --> 01:52:51,312
S1: doesn't matter if it's basic, if it should be basic

2079
01:52:51,312 --> 01:52:53,952
S1: or simple for a human. I agree it should be,

2080
01:52:53,952 --> 01:52:58,552
S1: but it's not so. So actually doing these complex things

2081
01:52:58,552 --> 01:53:03,711
S1: of pen testing and marriage counseling and cardiac cardiology, doing

2082
01:53:03,712 --> 01:53:06,272
S1: those things at scale as good or better than most

2083
01:53:06,272 --> 01:53:09,112
S1: people do them, most humans do them is still a

2084
01:53:09,112 --> 01:53:10,552
S1: massive boon to society.

2085
01:53:12,952 --> 01:53:16,711
S2: I, I agree with, uh, like what you just said

2086
01:53:16,712 --> 01:53:19,671
S2: in a vacuum. The issue I have, which I think

2087
01:53:19,672 --> 01:53:23,752
S2: we we discussed previously on text, is that in order

2088
01:53:23,752 --> 01:53:26,952
S2: for these systems to exist and sustain that, human knowledge

2089
01:53:26,952 --> 01:53:29,631
S2: has to remain right. Like, we can't set and forget

2090
01:53:29,632 --> 01:53:32,392
S2: this AI, and then suddenly we don't need to learn

2091
01:53:32,392 --> 01:53:36,722
S2: this entire subject. Right. Um, and one of the really

2092
01:53:36,722 --> 01:53:39,522
S2: big problems, which I've actually blogged about a lot, is

2093
01:53:39,522 --> 01:53:43,961
S2: the amount of data these models need in order to even, like,

2094
01:53:44,322 --> 01:53:48,042
S2: understand at a third grade level, a area of expertise

2095
01:53:48,042 --> 01:53:51,562
S2: is phenomenal. And what happens is once we start relying

2096
01:53:51,562 --> 01:53:54,602
S2: on the systems, we lose our own knowledge because we're

2097
01:53:54,602 --> 01:53:57,522
S2: now just deferring it to this system. But the system

2098
01:53:57,522 --> 01:54:01,362
S2: cannot sustain without this constant, uh, incoming flow of knowledge.

2099
01:54:01,402 --> 01:54:05,122
S2: And my big worry is that the more we automate

2100
01:54:05,122 --> 01:54:08,042
S2: with AI, the more of the, the more skill loss

2101
01:54:08,042 --> 01:54:10,802
S2: we get on our end, which then down the line

2102
01:54:10,802 --> 01:54:13,282
S2: impacts our ability to feed it back into the AI

2103
01:54:13,642 --> 01:54:17,482
S2: and keep these AIS, uh, like in tip top shape. Um,

2104
01:54:17,482 --> 01:54:21,962
S2: which is where we would need like an actual genuine AGI.

2105
01:54:22,282 --> 01:54:24,722
S2: We would need an AI that at some point it

2106
01:54:24,722 --> 01:54:28,442
S2: can learn for itself. It can just, um, it can

2107
01:54:28,442 --> 01:54:31,402
S2: teach itself new skills, which I don't think is possible.

2108
01:54:31,602 --> 01:54:33,642
S2: So then we run the risk of. sure in the

2109
01:54:33,642 --> 01:54:37,242
S2: short term. Can we, like, automate these bug bounties? Sure.

2110
01:54:37,282 --> 01:54:41,202
S2: Can we, uh, like, find the solution to this bacteria thing? Sure.

2111
01:54:41,562 --> 01:54:43,602
S2: But in like five, ten years, when we now no

2112
01:54:43,602 --> 01:54:47,322
S2: longer have any pen testers and, uh, no longer any

2113
01:54:47,322 --> 01:54:51,002
S2: biologists or whatever the word is for people who do bacteria,

2114
01:54:51,602 --> 01:54:55,442
S2: then what? What happens now when, like a new problem arises,

2115
01:54:55,482 --> 01:54:57,402
S2: our AI needs to be trained on it. It needs

2116
01:54:57,402 --> 01:55:00,482
S2: a model built to address that problem. And suddenly, like,

2117
01:55:00,482 --> 01:55:03,362
S2: no one can do the thing anymore. It's actually, uh. Yeah,

2118
01:55:03,802 --> 01:55:05,442
S2: it reminds me of. I don't remember the name of

2119
01:55:05,442 --> 01:55:07,722
S2: the movie, but there's a movie where they're, uh. I

2120
01:55:07,722 --> 01:55:10,962
S2: believe the aliens, they become so advanced that they forget

2121
01:55:10,962 --> 01:55:12,642
S2: how to do all the basics, and they have to

2122
01:55:12,642 --> 01:55:15,842
S2: contact humanity and be like, yo, we actually kind of

2123
01:55:15,882 --> 01:55:18,122
S2: forgot how to do, like, this basic shit. Can you, uh,

2124
01:55:18,122 --> 01:55:21,002
S2: can you help us? Can you, like, bail us out, please?

2125
01:55:21,962 --> 01:55:27,962
S1: Yeah, no, I agree. Yeah. It's an interesting variation of, uh, thing.

2126
01:55:27,962 --> 01:55:29,642
S1: I don't know if you saw the MIT paper that

2127
01:55:29,642 --> 01:55:32,692
S1: basically said people who are using AI and kind of

2128
01:55:32,732 --> 01:55:37,092
S1: relying on it, their brains actually changed. Like, you could

2129
01:55:37,092 --> 01:55:39,812
S1: physically see in their brains that they were less smart

2130
01:55:40,052 --> 01:55:43,852
S1: as a result of like. Using it as a crutch. Yeah, yeah,

2131
01:55:43,852 --> 01:55:46,052
S1: it was crazy. The sample size was really small. It

2132
01:55:46,052 --> 01:55:50,852
S1: was only 54 people, but it was the first example of, um.

2133
01:55:51,212 --> 01:55:55,932
S1: Like tangible difference of, like somebody who's thinking for themselves versus, uh,

2134
01:55:56,092 --> 01:56:01,292
S1: versus using AI. So I do take your point there. Um, well,

2135
01:56:01,292 --> 01:56:03,732
S1: I think we've been going for a while. Uh, any

2136
01:56:03,732 --> 01:56:07,852
S1: any final thoughts? I got to go get someone from Bart, but, um,

2137
01:56:07,852 --> 01:56:10,572
S1: any final thoughts? We could also do a second version, too. So.

2138
01:56:11,452 --> 01:56:13,732
S2: Yeah, I'm down to do a follow up. Yeah. No, my.

2139
01:56:13,732 --> 01:56:15,732
S2: I did not actually see that paper, but that that

2140
01:56:15,732 --> 01:56:19,132
S2: is terrifying. Like, if people are actually losing, like, critical

2141
01:56:19,132 --> 01:56:22,332
S2: thinking ability, that's even worse than just, like, knowledge loss.

2142
01:56:22,492 --> 01:56:24,692
S2: That's like full on. We're just going to end up

2143
01:56:24,692 --> 01:56:29,052
S2: with Idiocracy. So that is that is very concerning. Um,

2144
01:56:29,142 --> 01:56:31,662
S2: but yeah, like my, my primary concerns have been purely

2145
01:56:31,702 --> 01:56:34,502
S2: like the economic model of it. If you've got these

2146
01:56:34,542 --> 01:56:37,742
S2: eyes just ingesting all this data, they're essentially stealing it. Like,

2147
01:56:37,742 --> 01:56:40,102
S2: people don't like me using that term, but they are

2148
01:56:40,102 --> 01:56:43,662
S2: just taking copyrighted content. And the people who make that content,

2149
01:56:43,662 --> 01:56:46,782
S2: they make money from posting it from ad revenue. Um,

2150
01:56:46,782 --> 01:56:48,662
S2: so my belief has always been that they would just,

2151
01:56:48,942 --> 01:56:51,622
S2: just collapse under their own weight, they would cannibalize their

2152
01:56:51,622 --> 01:56:55,182
S2: source of information, and they would fail that way. But

2153
01:56:55,182 --> 01:56:58,942
S2: if they're also making humans dumber, that is an even

2154
01:56:58,942 --> 01:57:02,142
S2: more pressing problem, because then we're just I mean, we're

2155
01:57:02,142 --> 01:57:04,582
S2: just doomed at that point. So I'm I'm hoping that

2156
01:57:04,582 --> 01:57:07,222
S2: turns out to just be like, the sample size was

2157
01:57:07,222 --> 01:57:10,102
S2: too small, and maybe there was like some sampling bias

2158
01:57:10,102 --> 01:57:14,062
S2: where there's like a, a correlation between unintelligent people and

2159
01:57:14,062 --> 01:57:17,902
S2: just deferring their critical thinking to AI. Uh, but if

2160
01:57:17,902 --> 01:57:20,382
S2: that is the case that like, it's actually making intelligent

2161
01:57:20,382 --> 01:57:24,262
S2: people dumber. Um, um, I think we're going to have

2162
01:57:24,262 --> 01:57:26,702
S2: to unite to, to do something about that.

2163
01:57:27,072 --> 01:57:29,992
S1: Yeah. Yeah. I unfortunately, I don't think it's going to

2164
01:57:29,992 --> 01:57:32,552
S1: be a sample size issue. I think that's going to

2165
01:57:32,552 --> 01:57:35,992
S1: be reproducible. Uh, my explanation for it, though, is that

2166
01:57:35,992 --> 01:57:39,632
S1: there are some people who will do this. Arguably even

2167
01:57:39,632 --> 01:57:42,672
S1: most people will do this. They'll just like rely, rely, rely.

2168
01:57:42,672 --> 01:57:45,152
S1: And pretty soon they won't be able to think for themselves.

2169
01:57:45,152 --> 01:57:47,992
S1: And that's that's extremely troubling. But there is also some

2170
01:57:47,992 --> 01:57:51,792
S1: other group, um, which I hopefully will include myself in,

2171
01:57:52,032 --> 01:57:54,792
S1: who is going to like, train my eye to constantly

2172
01:57:54,792 --> 01:57:57,192
S1: badger me in, like sort of a Socratic type of

2173
01:57:57,232 --> 01:58:00,672
S1: way to make sure that never happens. Right? Because because

2174
01:58:00,672 --> 01:58:03,752
S1: I'm going to be actively like defending against this. Um,

2175
01:58:03,752 --> 01:58:06,312
S1: so I think people like that, they're going to accelerate,

2176
01:58:06,312 --> 01:58:08,712
S1: they're going to get smarter, even better at critical thinking.

2177
01:58:09,072 --> 01:58:12,512
S1: But that doesn't speak to the issue of like the other, um,

2178
01:58:12,752 --> 01:58:13,712
S1: larger percent.

2179
01:58:14,552 --> 01:58:17,432
S2: Yeah, I think that's going to be a very, very

2180
01:58:17,432 --> 01:58:19,592
S2: large percentage because it's the same. It's the same thing

2181
01:58:19,592 --> 01:58:22,792
S2: that makes social media work and. Yeah, and makes, uh,

2182
01:58:22,792 --> 01:58:26,962
S2: llms so attractive. It's the the instant gratification that I

2183
01:58:26,962 --> 01:58:28,882
S2: don't have to put in too much work to get

2184
01:58:28,882 --> 01:58:31,682
S2: the results that I want. Um, so I think that

2185
01:58:31,722 --> 01:58:34,602
S2: it's almost it's basically dopamine addiction to an extent. It's like, yeah,

2186
01:58:34,642 --> 01:58:36,482
S2: people just love the idea of I don't have to

2187
01:58:36,482 --> 01:58:38,322
S2: go and read a hundred books on programming. I can

2188
01:58:38,322 --> 01:58:41,122
S2: just make an app. So I think, uh, what's going

2189
01:58:41,122 --> 01:58:43,122
S2: to happen is sure there will be people like you

2190
01:58:43,162 --> 01:58:45,922
S2: who who do make their AI like, keep them in,

2191
01:58:45,962 --> 01:58:49,722
S2: like just challenge them. But I think for the overwhelming

2192
01:58:49,722 --> 01:58:53,122
S2: majority of the population that like instant gratification is just

2193
01:58:53,122 --> 01:58:54,762
S2: going to take them down that rabbit hole of I'm

2194
01:58:54,762 --> 01:58:56,842
S2: just going to I'm not even going to think about

2195
01:58:56,842 --> 01:58:58,402
S2: the question I was just asked. I'm just going to

2196
01:58:58,402 --> 01:59:02,002
S2: type it into ChatGPT. Yeah. Um, which now is giving

2197
01:59:02,002 --> 01:59:03,242
S2: me a lot of anxiety.

2198
01:59:04,362 --> 01:59:07,642
S1: Yeah. Well, Marcus, this has been super fun. Uh, I

2199
01:59:07,642 --> 01:59:10,842
S1: think the conversation was great. And, um. Yeah, we should

2200
01:59:10,842 --> 01:59:11,842
S1: talk about doing a follow up.

2201
01:59:11,842 --> 01:59:14,882
S2: Maybe just challenge them, but I think for them.

2202
01:59:15,122 --> 01:59:16,362
S1: And, uh, talk to you soon.

2203
01:59:17,362 --> 01:59:18,882
S2: Awesome. Thanks so much for having me on.

2204
01:59:19,202 --> 01:59:19,962
S1: All right. See you.