1
00:00:01,010 --> 00:00:04,399
S1: Welcome to Unsupervised Learning, a security, AI and meaning focused

2
00:00:04,400 --> 00:00:07,250
S1: podcast that looks at how best to thrive as humans

3
00:00:07,250 --> 00:00:11,480
S1: in a post AI world. It combines original ideas, analysis,

4
00:00:11,480 --> 00:00:14,720
S1: and mental models to bring not just the news, but

5
00:00:14,720 --> 00:00:21,610
S1: why it matters and how to respond. All right. Welcome

6
00:00:21,610 --> 00:00:25,840
S1: to unsupervised learning. This is Daniel Meisler. All right. 436

7
00:00:25,930 --> 00:00:29,440
S1: in about a bunch of AI stuff here. So, uh,

8
00:00:29,920 --> 00:00:33,160
S1: big religious day for me on Monday. It was, uh,

9
00:00:33,159 --> 00:00:37,150
S1: WWDC by Apple. And what I said after it was Apple,

10
00:00:37,150 --> 00:00:40,570
S1: just one. I they caught up and passed everyone basically

11
00:00:40,570 --> 00:00:43,570
S1: in one shot. And I think this is because the

12
00:00:43,570 --> 00:00:45,610
S1: thing that they get that most people don't is it

13
00:00:45,610 --> 00:00:48,070
S1: it's all about the platform. It's all about the integrations.

14
00:00:48,070 --> 00:00:51,190
S1: It's all about the ecosystem. And they're essentially the only

15
00:00:51,190 --> 00:00:54,580
S1: one who understands that and is also good at tech.

16
00:00:54,580 --> 00:00:57,760
S1: And they also have this massive platform in the iPhone. Right.

17
00:00:57,760 --> 00:01:00,640
S1: So it's just they're just winning. They're so far ahead

18
00:01:00,640 --> 00:01:03,340
S1: of everyone in terms of long term thinking, like it's

19
00:01:03,340 --> 00:01:06,040
S1: not even close now. This one has since been cleared up,

20
00:01:06,040 --> 00:01:08,200
S1: but at the time it wasn't clear what part of

21
00:01:08,200 --> 00:01:12,250
S1: what they were doing was actually Apple stuff. And what

22
00:01:12,250 --> 00:01:14,800
S1: part was OpenAI. I feel like they could have done

23
00:01:14,800 --> 00:01:17,200
S1: a better job explaining that because I was paying attention,

24
00:01:17,200 --> 00:01:20,320
S1: and I still kind of didn't get it. Looking back,

25
00:01:20,319 --> 00:01:23,410
S1: it was pretty clear somehow I missed it. And I,

26
00:01:23,410 --> 00:01:25,569
S1: I mean, I used to work there, I've worked on

27
00:01:25,569 --> 00:01:28,840
S1: AI quite a bit and I like, I love AI

28
00:01:28,840 --> 00:01:31,000
S1: and I love Apple, and I still didn't get it.

29
00:01:31,000 --> 00:01:32,890
S1: So I feel like they could have been more clear.

30
00:01:32,890 --> 00:01:35,500
S1: So the answer to this question is that most of

31
00:01:35,500 --> 00:01:38,440
S1: what they showed was actually all Apple stuff. It was

32
00:01:38,440 --> 00:01:42,819
S1: not pure OpenAI. And that that's a huge difference, right?

33
00:01:42,819 --> 00:01:46,360
S1: So so essentially what the situation is, is, I would

34
00:01:46,360 --> 00:01:49,360
S1: say probably 90% of what they showed. I don't know,

35
00:01:49,360 --> 00:01:51,550
S1: you can figure it out for yourself. Like if you

36
00:01:51,550 --> 00:01:53,890
S1: did the the math or whatever, it doesn't really matter.

37
00:01:53,890 --> 00:01:57,220
S1: The point is like 90% roughly of everything they talked

38
00:01:57,220 --> 00:02:01,990
S1: about on device, the cloud stuff, the whole new cloud infrastructure,

39
00:02:01,990 --> 00:02:06,250
S1: which is absolutely insane. They basically built a new AI

40
00:02:06,280 --> 00:02:10,630
S1: cloud infrastructure that's completely secure, where, you know, they don't

41
00:02:10,630 --> 00:02:13,540
S1: even have access to the content that's on there. They

42
00:02:13,540 --> 00:02:16,600
S1: don't give that access to any third parties. It's essentially

43
00:02:16,600 --> 00:02:20,650
S1: like a cloud version of the local, very secure enclave.

44
00:02:20,650 --> 00:02:22,960
S1: I'm not sure it's quite that level, but it's pretty

45
00:02:22,960 --> 00:02:26,050
S1: much that level. It is better than anything that exists

46
00:02:26,050 --> 00:02:28,660
S1: in the industry right now in terms of like secure

47
00:02:28,660 --> 00:02:32,560
S1: cloud compute, and it's just extraordinary. And they generated it

48
00:02:32,560 --> 00:02:35,620
S1: just so that they could do these AI features. So

49
00:02:35,620 --> 00:02:39,220
S1: that is essentially their AI story is they're going to

50
00:02:39,220 --> 00:02:42,520
S1: have deep context about you because they're Apple. And this

51
00:02:42,520 --> 00:02:45,459
S1: is essentially life OS is what we're talking about. And

52
00:02:45,460 --> 00:02:49,000
S1: so they have all that information about you and they

53
00:02:49,000 --> 00:02:51,670
S1: have the ability to bring that to AI. And what

54
00:02:51,669 --> 00:02:53,470
S1: they've done is they've made it so that if you

55
00:02:53,470 --> 00:02:55,990
S1: could do it locally, it does it locally, and if

56
00:02:55,990 --> 00:02:57,850
S1: you can't do it locally, it goes up to the

57
00:02:57,850 --> 00:03:02,200
S1: cloud using AI models by Apple. So these are Apple

58
00:03:02,200 --> 00:03:06,040
S1: models running on Apple silicon right. It's not even it's

59
00:03:06,040 --> 00:03:09,519
S1: not even Nvidia stuff. They're using their own chips also

60
00:03:09,520 --> 00:03:12,730
S1: in the cloud and obviously on device. So that's the

61
00:03:12,730 --> 00:03:17,859
S1: whole ecosystem. Now that is for all the context data okay.

62
00:03:17,860 --> 00:03:21,970
S1: That's for basically all of the personal data that you have,

63
00:03:21,970 --> 00:03:24,190
S1: which they don't want to share with other people and

64
00:03:24,190 --> 00:03:26,590
S1: they want to be very, very private. Now in the

65
00:03:26,590 --> 00:03:29,260
S1: case of like looking things up like, oh, get the

66
00:03:29,260 --> 00:03:32,650
S1: current weather and stuff like that. That's where they partnered

67
00:03:32,650 --> 00:03:37,300
S1: with the new Siri, with the new with OpenAI. Right.

68
00:03:37,300 --> 00:03:39,910
S1: So that's that's what they did. So essentially what they

69
00:03:39,910 --> 00:03:44,200
S1: did was they said, okay for live searches, OpenAI is

70
00:03:44,200 --> 00:03:47,500
S1: really good at that. You know, fact checking, stuff like that.

71
00:03:47,500 --> 00:03:50,080
S1: Just just calling the internet. And maybe they're not great

72
00:03:50,080 --> 00:03:52,000
S1: at fact checking, but you know what I mean? So

73
00:03:52,000 --> 00:03:56,800
S1: they're leveraging third party AI for a very specific thing,

74
00:03:56,800 --> 00:03:59,890
S1: but they're not like uploading all this stuff to OpenAI.

75
00:03:59,890 --> 00:04:02,980
S1: It's like the exact opposite of what they're doing. They

76
00:04:02,980 --> 00:04:05,950
S1: built this whole new infrastructure just to keep it all

77
00:04:05,950 --> 00:04:08,680
S1: very secure and very private. So it's very much two

78
00:04:08,680 --> 00:04:11,410
S1: different things. Apple AI, which is like 90% of what

79
00:04:11,410 --> 00:04:13,990
S1: they talked about, which has the on prem and the

80
00:04:13,990 --> 00:04:17,440
S1: cloud component. And then this other third party integration, which

81
00:04:17,440 --> 00:04:20,020
S1: is with ChatGPT. And they even talked about yeah, they're

82
00:04:20,020 --> 00:04:23,080
S1: going to do like they're looking at possibly doing Gemini.

83
00:04:23,080 --> 00:04:25,989
S1: But the point is other third party services for those

84
00:04:25,990 --> 00:04:29,920
S1: kind of integrations. But don't confuse it. That is not

85
00:04:29,950 --> 00:04:34,300
S1: what they're doing with our personal data. Like so Live Bovary.

86
00:04:34,450 --> 00:04:36,670
S1: A whole bunch of people were like, well, how do

87
00:04:36,670 --> 00:04:39,789
S1: I find an android? Like, I can't believe they're sending

88
00:04:39,790 --> 00:04:41,890
S1: all my data to OpenAI. And it's like, that is

89
00:04:41,890 --> 00:04:44,109
S1: not what they said. It's not what they said. It's

90
00:04:44,110 --> 00:04:47,650
S1: like the opposite of what they said. So, um, already

91
00:04:47,650 --> 00:04:52,330
S1: running the demos or the betas for iOS 18, uh,

92
00:04:52,330 --> 00:04:55,870
S1: remarkably stable, but largely because they didn't really change that

93
00:04:55,870 --> 00:04:59,440
S1: much yet, and they definitely haven't added like the AI

94
00:04:59,470 --> 00:05:02,980
S1: services yet. So like, Siri is still the same. That

95
00:05:02,980 --> 00:05:05,950
S1: stuff is coming out like later this year, they said.

96
00:05:05,950 --> 00:05:08,409
S1: So I feel like I didn't really get the AI

97
00:05:08,410 --> 00:05:11,680
S1: features that I was looking for immediately in the beta,

98
00:05:11,680 --> 00:05:14,860
S1: which I was hoping to get. But whatever it happens

99
00:05:14,860 --> 00:05:18,640
S1: when it happens went and trained kickboxing. I was extremely.

100
00:05:18,760 --> 00:05:21,940
S1: Got out it. It was quite sad. I was like

101
00:05:21,940 --> 00:05:25,090
S1: five times worse than a beginner, I think, because it's

102
00:05:25,089 --> 00:05:28,060
S1: hard for me to like, do things slowly and I'm

103
00:05:28,060 --> 00:05:30,760
S1: like imagining things in my brain, but my body won't

104
00:05:30,760 --> 00:05:33,550
S1: do that thing that's in my brain. I'm just very

105
00:05:33,550 --> 00:05:35,410
S1: sort of confused. So I think I'm going to do

106
00:05:35,410 --> 00:05:39,310
S1: some more, like learning the basics of the body coordination,

107
00:05:39,310 --> 00:05:42,190
S1: like on a bag, and then do it with live

108
00:05:42,190 --> 00:05:45,430
S1: people after that. That's my current plan. Plus I'm going

109
00:05:45,430 --> 00:05:48,370
S1: to be doing a Gits as well. And uh, one

110
00:05:48,370 --> 00:05:51,909
S1: of the most important conversations ever on AI safety absolutely.

111
00:05:51,910 --> 00:05:55,660
S1: Got to watch this thing. It's, uh, Leopold Aschenbrenner and

112
00:05:55,660 --> 00:05:59,560
S1: Dwarkesh Patel. I've talked about this million times, but it

113
00:05:59,560 --> 00:06:02,350
S1: is really, really good. Wrote a new fabric pattern called

114
00:06:02,350 --> 00:06:06,370
S1: Capture Thinkers work. So you basically send in, like Hayek

115
00:06:06,370 --> 00:06:11,740
S1: or somebody into, you know, fabric SP capture thinkers work.

116
00:06:11,740 --> 00:06:15,700
S1: And I mean, it's extraordinary. In fact, I can show you, okay,

117
00:06:15,700 --> 00:06:19,779
S1: so if we do a capture thinkers work here, let's

118
00:06:19,779 --> 00:06:23,109
S1: do this. Let's first let's blow it up some and

119
00:06:23,110 --> 00:06:26,799
S1: then we'll do who's a philosopher? Hume. We'll send that

120
00:06:26,800 --> 00:06:30,820
S1: to capture thinkers work. So comes up with a single

121
00:06:30,820 --> 00:06:35,919
S1: line sentence up here. Most impactful ideas primary advice or teachings.

122
00:06:35,920 --> 00:06:39,190
S1: The works their main works. So you can go build

123
00:06:39,190 --> 00:06:42,580
S1: a reading list based on this top quotes. Something is

124
00:06:42,580 --> 00:06:46,360
S1: humean if it emphasizes the role of experienced custom and

125
00:06:46,360 --> 00:06:50,800
S1: sentiment over abstract reason and innate ideas, and like advice

126
00:06:50,800 --> 00:06:55,060
S1: of how to like implement humean concepts into your life.

127
00:06:55,060 --> 00:06:57,850
S1: So yeah, love this pattern. It's a great way of

128
00:06:57,850 --> 00:07:00,970
S1: like thinking about someone's work. So for example, if you

129
00:07:00,970 --> 00:07:03,909
S1: do like S or pearl, which it doesn't matter if

130
00:07:03,910 --> 00:07:06,310
S1: you spell it wrong, for example, it looks like maybe

131
00:07:06,310 --> 00:07:08,859
S1: I spelled it wrong because it corrected it to one

132
00:07:08,860 --> 00:07:12,490
S1: R and one L, but yeah. Background where they were born,

133
00:07:12,490 --> 00:07:16,750
S1: the school psychotherapy just really, really powerful. Got a thread

134
00:07:16,750 --> 00:07:19,480
S1: here on how to find a good mentor. I, I

135
00:07:19,480 --> 00:07:21,400
S1: put this out a very long time ago, but I

136
00:07:21,400 --> 00:07:23,410
S1: think it's worth sharing. I'm going to go ahead and

137
00:07:23,410 --> 00:07:27,340
S1: open it up. So it's like yeah, being mentored by

138
00:07:27,340 --> 00:07:29,650
S1: someone ahead of you can really change your life. Don't

139
00:07:29,650 --> 00:07:34,390
S1: overuse flattery. Ask something very specific. Behave like a future

140
00:07:34,390 --> 00:07:39,790
S1: peer like don't. Don't grovel. Basically, show your you've already

141
00:07:39,790 --> 00:07:42,340
S1: done a bunch of work and, you know, don't ask

142
00:07:42,340 --> 00:07:44,770
S1: basic questions that you could have got by reading their

143
00:07:44,770 --> 00:07:48,100
S1: blog or their book. I mean, to some extent you can,

144
00:07:48,100 --> 00:07:51,130
S1: but you don't want to overdo that and ask for

145
00:07:51,130 --> 00:07:53,440
S1: an opinion on something you've created. So it's like, look,

146
00:07:53,440 --> 00:07:57,010
S1: I've done work, I've created something. I'm curious about a

147
00:07:57,010 --> 00:08:00,040
S1: very specific question about what you think about this thing

148
00:08:00,040 --> 00:08:03,100
S1: you don't want to ask open ended things. Kind of

149
00:08:03,100 --> 00:08:05,800
S1: a general rule here is like, don't give them work.

150
00:08:05,800 --> 00:08:10,060
S1: You're you're being respectful of their time by asking a

151
00:08:10,060 --> 00:08:13,300
S1: pretty small thing and just sort of like presenting yourself

152
00:08:13,300 --> 00:08:15,250
S1: as a peer and you're like, hey, look, I'm looking

153
00:08:15,250 --> 00:08:17,530
S1: for feedback on this. I know you've done it a lot,

154
00:08:17,530 --> 00:08:19,510
S1: and that's kind of the vibe. Or you can offer

155
00:08:19,510 --> 00:08:24,430
S1: an improvement or construction a constructive like comment or feedback

156
00:08:24,430 --> 00:08:28,030
S1: on something they've done and most importantly, like produce something

157
00:08:28,030 --> 00:08:30,340
S1: really cool and show it to them. And then they're

158
00:08:30,340 --> 00:08:33,310
S1: likely to just like, treat you like a future peer

159
00:08:33,309 --> 00:08:36,700
S1: and interact with you. And that's the summary here. So

160
00:08:36,700 --> 00:08:40,240
S1: and it got a blog post around it okay. Let's see. Yeah.

161
00:08:40,240 --> 00:08:43,179
S1: Adding source names to stories now which we'll see here

162
00:08:43,179 --> 00:08:46,000
S1: in a second. So basically you'll know what the source

163
00:08:46,000 --> 00:08:48,190
S1: is before you click it. New member piece on the

164
00:08:48,190 --> 00:08:51,130
S1: future of AI okay. Yeah I'm going to go into

165
00:08:51,130 --> 00:08:56,500
S1: this one. This one's pretty major okay. Let's dive into this.

166
00:08:56,500 --> 00:09:00,580
S1: All right. So first thing is like my definition of

167
00:09:00,580 --> 00:09:06,939
S1: AGI and ASI. So my definition of AGI is can

168
00:09:06,940 --> 00:09:10,150
S1: functionally replace an average white collar worker in the US

169
00:09:10,150 --> 00:09:14,560
S1: making 80 K in 2023. And I say 2023 because

170
00:09:14,559 --> 00:09:17,199
S1: it's going to get weird. This definition will get more

171
00:09:17,200 --> 00:09:20,260
S1: weird if you don't lock it into a previous time before.

172
00:09:20,260 --> 00:09:23,620
S1: Like AGI. We're basically saying like, look, a white collar

173
00:09:23,620 --> 00:09:27,309
S1: worker making an average salary, right? Like 80 K in

174
00:09:27,309 --> 00:09:32,440
S1: the US in 2023. And ASI, I'm defining as an

175
00:09:32,440 --> 00:09:36,700
S1: AGI that's smarter and more capable than any human that's

176
00:09:36,700 --> 00:09:40,210
S1: ever lived. And what what this is essentially saying is like,

177
00:09:40,210 --> 00:09:42,820
S1: if you can functionally replace an average white collar worker

178
00:09:42,820 --> 00:09:47,469
S1: in the US, what that is referring to is generality,

179
00:09:47,740 --> 00:09:50,290
S1: because what I'm talking about is the ability for the

180
00:09:50,290 --> 00:09:52,840
S1: boss to say, you know what, you're actually not on

181
00:09:52,870 --> 00:09:56,830
S1: that project anymore. Go get with Julie and Lisa, and

182
00:09:56,830 --> 00:10:00,280
S1: you're being completely retasked to this other project, which is

183
00:10:00,280 --> 00:10:03,100
S1: kind of unrelated to that. And there is some issue

184
00:10:03,100 --> 00:10:06,340
S1: of like specialization, but we're assuming it's still within the

185
00:10:06,340 --> 00:10:10,420
S1: specialization of that person, and it's general enough where you

186
00:10:10,420 --> 00:10:13,510
S1: could be like, look, it's a completely different task. It's

187
00:10:13,510 --> 00:10:16,090
S1: a completely different project. It's got a different time frame,

188
00:10:16,090 --> 00:10:19,559
S1: it's got different goals, it's got different metrics. Hand, what

189
00:10:19,559 --> 00:10:22,650
S1: a human does is basically say, okay, cool, I will

190
00:10:22,650 --> 00:10:25,199
S1: erase what I was thinking before. I will take the

191
00:10:25,200 --> 00:10:27,720
S1: new requirements, I will come up with a new plan

192
00:10:27,720 --> 00:10:29,460
S1: and I will start working on it, which means you

193
00:10:29,460 --> 00:10:31,020
S1: got to break it into tasks. You got to have

194
00:10:31,020 --> 00:10:33,300
S1: a timeline, you got to have a schedule, you got

195
00:10:33,300 --> 00:10:35,250
S1: to have all that. So in other words, it's not

196
00:10:35,250 --> 00:10:39,750
S1: just narrowly executing small tasks, it's also the planning stage.

197
00:10:39,780 --> 00:10:44,040
S1: It's also constant readjustment, constant evolution of like how do

198
00:10:44,040 --> 00:10:46,650
S1: I best do this. And obviously there's different tiers of

199
00:10:46,650 --> 00:10:49,500
S1: quality of people who can do that as well. But

200
00:10:49,500 --> 00:10:54,420
S1: that's why I say average white collar worker making $80,000.

201
00:10:54,420 --> 00:10:58,050
S1: So it's not like exceptional. And it's not like the worst.

202
00:10:58,050 --> 00:11:01,679
S1: So so given that as a definition of AGI, this

203
00:11:01,679 --> 00:11:06,179
S1: general intelligence that can like be retasked as well as

204
00:11:06,179 --> 00:11:10,500
S1: do specific tasks underneath plus answer emails plus be, you know,

205
00:11:10,500 --> 00:11:14,429
S1: communicative and, you know, basically function as a good employee.

206
00:11:14,429 --> 00:11:19,230
S1: That's essentially AGI. So given that what who is able

207
00:11:19,230 --> 00:11:21,719
S1: to do that and basically be a thinker and be

208
00:11:21,720 --> 00:11:25,829
S1: a producer of new ideas, basically smarter and more capable

209
00:11:25,830 --> 00:11:28,170
S1: than any human that's ever lived. And a good way

210
00:11:28,170 --> 00:11:31,800
S1: of thinking about this is like and Aschenbrenner talked about

211
00:11:31,800 --> 00:11:35,370
S1: this as well in the forecast. So while using their

212
00:11:35,370 --> 00:11:40,319
S1: first names, Leopold talked about that in the talk podcast conversation.

213
00:11:40,320 --> 00:11:43,140
S1: It's essentially like, what if you could just make like

214
00:11:43,140 --> 00:11:48,810
S1: 10,000 Einsteins or John von Neumann, right? That would be insane.

215
00:11:48,809 --> 00:11:52,980
S1: So just imagine John von Neumann here, or Einstein or

216
00:11:52,980 --> 00:11:55,410
S1: whoever you want to put as like the smartest person.

217
00:11:55,410 --> 00:11:57,510
S1: And just imagine you can have a whole bunch of

218
00:11:57,510 --> 00:12:00,360
S1: them who could go and do these tasks, like AGI

219
00:12:00,390 --> 00:12:04,439
S1: tasks like, you know, most importantly, come up with new science,

220
00:12:04,440 --> 00:12:06,929
S1: come up with new innovations, come up with new medicines,

221
00:12:06,929 --> 00:12:10,020
S1: but also, yeah, go do all this work, figure out

222
00:12:10,020 --> 00:12:12,809
S1: how how to do it better figure out better ways

223
00:12:12,809 --> 00:12:14,100
S1: to do all the work we have to do and

224
00:12:14,100 --> 00:12:16,709
S1: then go do it. So that's AC. And I'm saying

225
00:12:16,710 --> 00:12:20,670
S1: essentially there's a big divide here between conscious and unconscious.

226
00:12:20,670 --> 00:12:23,189
S1: And it actually doesn't matter all that much. A lot

227
00:12:23,190 --> 00:12:27,000
S1: of people think that consciousness is on this path over here.

228
00:12:27,000 --> 00:12:30,540
S1: So it's like, oh, we're going to get artificial narrow intelligence,

229
00:12:30,540 --> 00:12:32,309
S1: and then we're going to get AGI. And then after

230
00:12:32,309 --> 00:12:35,160
S1: that it becomes conscious and then it becomes super intelligent

231
00:12:35,160 --> 00:12:37,740
S1: or some combination thereof. But I think the better way

232
00:12:37,740 --> 00:12:40,620
S1: to think about this is that we could stay completely

233
00:12:40,620 --> 00:12:43,470
S1: over here and not even get consciousness, but still have

234
00:12:43,470 --> 00:12:48,179
S1: artificial super intelligence, right? So one does not require the other.

235
00:12:48,179 --> 00:12:52,410
S1: And yeah, I've got prediction for a date later on.

236
00:12:52,410 --> 00:12:55,980
S1: We are now inside the post itself. Yeah. So how

237
00:12:55,980 --> 00:12:58,230
S1: I see the paths of development and the thing that

238
00:12:58,230 --> 00:13:00,690
S1: got me talking about this, or thinking about this a

239
00:13:00,690 --> 00:13:05,069
S1: lot more, is just this whole Leopold and Dwarkesh thing. Yeah.

240
00:13:05,070 --> 00:13:09,750
S1: So I talked about that path and that's the conversation. Yeah. Okay.

241
00:13:09,750 --> 00:13:13,800
S1: So this is basically my argument of like how we're

242
00:13:13,800 --> 00:13:17,820
S1: going to get to AGI and ASI. So I think

243
00:13:17,820 --> 00:13:21,179
S1: this is a hypothesis. Right. It's it's an idea. It's

244
00:13:21,179 --> 00:13:24,300
S1: a thought only so many components to understand the world.

245
00:13:24,300 --> 00:13:28,230
S1: Some of these are easily attainable and some are like impossible.

246
00:13:28,230 --> 00:13:30,809
S1: So an impossible one is like you can't calculate all

247
00:13:30,809 --> 00:13:35,069
S1: the combinations of things like numbers get crazy very quickly

248
00:13:35,070 --> 00:13:39,150
S1: with combinations. So unless you can brute force everything like

249
00:13:39,150 --> 00:13:42,780
S1: like God or something, it's really hard. But I think

250
00:13:42,780 --> 00:13:48,420
S1: there are really realistic and approachable abstractions of that complexity.

251
00:13:48,420 --> 00:13:51,480
S1: So my intuition is that there are likely abstractions of

252
00:13:51,480 --> 00:13:55,170
S1: those that we haven't yet attained. I think that's pretty obvious.

253
00:13:55,170 --> 00:13:58,109
S1: And then an analogy I use here is like fluid

254
00:13:58,110 --> 00:14:02,910
S1: dynamics representing all the molecules in a liquid, or Newtonian

255
00:14:02,910 --> 00:14:06,180
S1: physics being an abstraction for like quantum or whatever the

256
00:14:06,179 --> 00:14:09,329
S1: bottom layer is. So my friend's argument is I was

257
00:14:09,330 --> 00:14:13,079
S1: walking with him recently and like, I think it's really smart.

258
00:14:13,080 --> 00:14:16,170
S1: He basically said there's limitations to what we can do

259
00:14:16,170 --> 00:14:20,070
S1: even with superintelligence, because of the one way nature of calculation.

260
00:14:20,070 --> 00:14:23,460
S1: So basically P and P, you could try lots of things,

261
00:14:23,460 --> 00:14:25,500
S1: but it's a one way function. You can see if

262
00:14:25,500 --> 00:14:28,800
S1: it works, but you can't like immediately see all the options.

263
00:14:28,800 --> 00:14:32,040
S1: And that's kind of like P and P. So he

264
00:14:32,040 --> 00:14:34,590
S1: came up with a great definition of intelligence, which I've

265
00:14:34,590 --> 00:14:37,620
S1: since seen like other people talking about it as well,

266
00:14:37,620 --> 00:14:40,950
S1: which is like the ability to quickly reduce a search

267
00:14:40,950 --> 00:14:44,130
S1: space for a problem, like like pruning trees. So there's

268
00:14:44,130 --> 00:14:47,729
S1: like 100 trillion options. What can we do to get

269
00:14:47,730 --> 00:14:50,730
S1: that down to like 10 or 100 or 1000 or

270
00:14:50,730 --> 00:14:54,090
S1: whatever as fast as possible. And the faster you could

271
00:14:54,090 --> 00:14:55,590
S1: do that, the more you could do that, the more

272
00:14:55,590 --> 00:14:58,620
S1: intelligent you are. And actually reminds me a lot of

273
00:14:58,620 --> 00:15:01,080
S1: quantum computing for the first time. I've never seen like

274
00:15:01,080 --> 00:15:05,400
S1: an intersection there, but imagine that you could test multiple

275
00:15:05,400 --> 00:15:07,710
S1: options all at once, which is kind of the whole

276
00:15:07,710 --> 00:15:11,160
S1: thing with quantum anyway, so here are the components I

277
00:15:11,160 --> 00:15:15,000
S1: see leading to superintelligence. One is a super complex model

278
00:15:15,000 --> 00:15:18,030
S1: of the world, like how the cell works, how medicines interact.

279
00:15:18,170 --> 00:15:23,330
S1: With cells, you know, molecular interaction with different body parts.

280
00:15:23,330 --> 00:15:27,560
S1: So it's essentially like testing drugs, but doing it with

281
00:15:27,560 --> 00:15:31,250
S1: models instead of like having to actually try it in labs,

282
00:15:31,250 --> 00:15:35,390
S1: in people, in mice. So think of that at like

283
00:15:35,390 --> 00:15:39,740
S1: all levels of physics. So like quantum atomic molecules, cells

284
00:15:39,740 --> 00:15:43,850
S1: in bigger and bigger, eventually you're at like psychology, sociology, whatever.

285
00:15:43,850 --> 00:15:48,170
S1: And in some cases it's just feeding our current understanding

286
00:15:48,170 --> 00:15:51,530
S1: as text. But even better, if you like, have actual

287
00:15:51,530 --> 00:15:55,310
S1: recordings of these things happening. Similar to how Tesla taught

288
00:15:55,310 --> 00:15:59,570
S1: its new full Self-Driving. It's just like, yeah, this is here.

289
00:15:59,570 --> 00:16:02,930
S1: Here are the actual interactions. Learn from that. So the

290
00:16:02,930 --> 00:16:06,440
S1: use case I kept using for this was solving human aging.

291
00:16:06,440 --> 00:16:09,950
S1: So let's say it's like the telomeres or whatever the

292
00:16:09,950 --> 00:16:11,930
S1: length of the telomeres. And there's a whole bunch of

293
00:16:11,930 --> 00:16:14,990
S1: different possible mechanisms of how aging could be happening, and

294
00:16:14,990 --> 00:16:17,240
S1: therefore how we might be able to turn it off

295
00:16:17,240 --> 00:16:19,940
S1: or reverse it or whatever. But the most important thing

296
00:16:19,940 --> 00:16:23,360
S1: is like have a deep understanding of how everything works,

297
00:16:23,360 --> 00:16:25,550
S1: so you know where you could actually start to go

298
00:16:25,550 --> 00:16:28,130
S1: and look. So the next big thing for this, in

299
00:16:28,130 --> 00:16:31,940
S1: my opinion, is the patterns that analogies and metaphors. So

300
00:16:31,940 --> 00:16:35,239
S1: links between things. And we're already seeing like models get

301
00:16:35,240 --> 00:16:37,280
S1: really good at that. And I think the next one

302
00:16:37,280 --> 00:16:40,640
S1: is the size of the working memory. Right. So not

303
00:16:40,640 --> 00:16:43,490
S1: exactly sure what this means, but it's like it's it's

304
00:16:43,490 --> 00:16:47,510
S1: like your Ram or your functional memory. It's like, how

305
00:16:47,510 --> 00:16:49,220
S1: much of this map can you fit in your brain

306
00:16:49,220 --> 00:16:51,830
S1: at once without having to slide it over in parts

307
00:16:51,830 --> 00:16:53,420
S1: of the map, fall off the table and you can't

308
00:16:53,420 --> 00:16:56,390
S1: see them. So maybe that's a combination of multiple tiers.

309
00:16:56,390 --> 00:16:58,610
S1: So you have like you have like Ram. Then you

310
00:16:58,610 --> 00:17:02,180
S1: have like L1, L2 inside of the CPU, which is

311
00:17:02,180 --> 00:17:04,880
S1: really the brain, not really memory, but kind of all

312
00:17:04,880 --> 00:17:07,669
S1: the same. Right. And then you've got disk which is

313
00:17:07,670 --> 00:17:10,610
S1: much slower, which has to go into memory to get

314
00:17:10,609 --> 00:17:14,270
S1: loaded up. And when there's too much on disk and

315
00:17:14,270 --> 00:17:17,240
S1: not enough Ram, like I said, you got to swap, right.

316
00:17:17,240 --> 00:17:20,360
S1: So these are the types of things that I think

317
00:17:20,359 --> 00:17:24,139
S1: the overall size of that memory that can do the

318
00:17:24,140 --> 00:17:28,490
S1: patterns based on the world model is essentially what gets

319
00:17:28,490 --> 00:17:31,970
S1: us there. That that's my theory. And essentially this theory

320
00:17:31,970 --> 00:17:35,869
S1: is that with scale inside of neural networks, those three

321
00:17:35,869 --> 00:17:39,350
S1: things just keep getting better and better, plus some post-processing,

322
00:17:39,350 --> 00:17:41,840
S1: which we'll talk about, and a whole bunch of hacks

323
00:17:41,840 --> 00:17:43,699
S1: and tricks that are basically just going to come with

324
00:17:43,700 --> 00:17:47,090
S1: the industry. Because my friend pointed out that just the

325
00:17:47,090 --> 00:17:49,940
S1: size of the model enough alone is not going to

326
00:17:49,940 --> 00:17:53,390
S1: be enough, and that post training is super critical. And

327
00:17:53,390 --> 00:17:56,240
S1: once I got what he was saying, like, we definitely

328
00:17:56,240 --> 00:17:58,310
S1: agree on that now, it was just something that I

329
00:17:58,310 --> 00:18:01,369
S1: wasn't aware of and that got us kind of in line.

330
00:18:01,369 --> 00:18:03,980
S1: So this is my way of capturing this. So my

331
00:18:03,980 --> 00:18:07,220
S1: hypothesis is that there are very few fundamental components that

332
00:18:07,220 --> 00:18:11,119
S1: allow AI to scale up from narrow AI to AGI

333
00:18:11,150 --> 00:18:15,020
S1: to ASI. So it's a world model, sufficient training examples

334
00:18:15,020 --> 00:18:18,890
S1: to allow deep enough ability to find patterns and similarities

335
00:18:18,890 --> 00:18:23,240
S1: between phenomenon within that world model, sufficient working memory, and

336
00:18:23,240 --> 00:18:27,560
S1: then ultimately the ability to like, model the scientific method

337
00:18:27,560 --> 00:18:32,720
S1: and test things. Ideally not having to actually do that test,

338
00:18:32,720 --> 00:18:36,770
S1: but being able to test things in an abstraction. Right.

339
00:18:36,770 --> 00:18:40,370
S1: And I'm not sure how smart or how good that is,

340
00:18:40,369 --> 00:18:44,240
S1: because if you're not testing, you're not really testing. I

341
00:18:44,240 --> 00:18:47,090
S1: think this might be possible. And yeah, like I said,

342
00:18:47,090 --> 00:18:51,139
S1: I'm least confident that number four is required or that

343
00:18:51,140 --> 00:18:54,500
S1: it's possible. However, I would say that these three are

344
00:18:54,500 --> 00:18:58,399
S1: amazing and possibly enough. And this one, if it were possible,

345
00:18:58,400 --> 00:19:01,010
S1: to whatever degree, it would massively enhance the other three.

346
00:19:01,010 --> 00:19:03,859
S1: And I think these are the main components. I mean,

347
00:19:03,859 --> 00:19:07,520
S1: I've spent relatively little time thinking about this. I could

348
00:19:07,520 --> 00:19:09,800
S1: revise it in the future. The other thing to say

349
00:19:09,800 --> 00:19:13,580
S1: about this is that number four, which is the testing piece,

350
00:19:13,580 --> 00:19:18,919
S1: might just be implied somehow if number one is sufficiently advanced,

351
00:19:18,920 --> 00:19:22,460
S1: if the world model is good enough, maybe we start

352
00:19:22,460 --> 00:19:25,460
S1: getting to this, I don't know. And maybe it requires

353
00:19:25,460 --> 00:19:29,180
S1: the patterns and similarities. And obviously I think it a

354
00:19:29,180 --> 00:19:31,909
S1: lot of this requires the memory. So I guess what

355
00:19:31,910 --> 00:19:34,639
S1: I'm saying is the whole thing might be two big things,

356
00:19:34,640 --> 00:19:37,280
S1: the complexity of the world model and the size of

357
00:19:37,280 --> 00:19:40,040
S1: the active memory that it's able to use. And maybe

358
00:19:40,040 --> 00:19:43,280
S1: the patterns and the testing stuff might come along with

359
00:19:43,280 --> 00:19:47,420
S1: those two. And I guess the most controversial thing is that,

360
00:19:47,420 --> 00:19:51,290
S1: and this kind of undergirds everything I'm saying is the

361
00:19:51,290 --> 00:19:55,640
S1: hypothesis that the universe has like an actual complexity that

362
00:19:55,640 --> 00:19:59,930
S1: is approachable or finite, and that models only need to

363
00:19:59,930 --> 00:20:05,780
S1: cross a certain threshold of depth of understanding or quality

364
00:20:05,780 --> 00:20:10,700
S1: of abstractions or depth of abstractions or height of abstractions

365
00:20:10,700 --> 00:20:15,800
S1: to become very similar to like a Laplacian demon in

366
00:20:15,800 --> 00:20:19,190
S1: the sense of like full knowledge. So I don't think

367
00:20:19,190 --> 00:20:22,250
S1: they can have full knowledge. I think that's God level

368
00:20:22,250 --> 00:20:26,570
S1: stuff that's supernatural. At least according to our current understanding

369
00:20:26,570 --> 00:20:30,139
S1: and my belief. But that if you hit a certain

370
00:20:30,140 --> 00:20:35,210
S1: threshold or depth of abstraction quality, it's functionally very similar.

371
00:20:35,210 --> 00:20:40,280
S1: That is what I'm basing my intuition, my hypothesis of

372
00:20:40,280 --> 00:20:44,869
S1: AGI and ASI on. I got another concept called human 3.0.

373
00:20:44,869 --> 00:20:48,439
S1: So it's essentially about how many people believe that they

374
00:20:48,440 --> 00:20:50,750
S1: have ideas that are useful to the world. How many

375
00:20:50,750 --> 00:20:53,659
S1: people believe they could write a short book, percentage of

376
00:20:53,660 --> 00:20:56,990
S1: a person's creativity that they're actively using in their life,

377
00:20:56,990 --> 00:21:00,800
S1: percentage of a person's total capabilities that they actually broadcast

378
00:21:00,800 --> 00:21:04,580
S1: the world as some sort of value, and the degree

379
00:21:04,580 --> 00:21:08,090
S1: to which someone lives as their full spectrum self. So

380
00:21:08,090 --> 00:21:13,250
S1: in human 2.0, we're essentially living and working for capitalism.

381
00:21:13,250 --> 00:21:16,160
S1: Capitalism is like the most important thing. So we're kind

382
00:21:16,160 --> 00:21:20,330
S1: of like 10%. We're putting like 10% of our value

383
00:21:20,330 --> 00:21:23,360
S1: in what we can offer to the world, I would

384
00:21:23,359 --> 00:21:25,940
S1: argue in our CV because it's like, oh, I'm really

385
00:21:25,940 --> 00:21:29,210
S1: good with Microsoft Excel, and I can send emails or whatever.

386
00:21:29,210 --> 00:21:31,790
S1: And it's like, is that really you? I don't think so.

387
00:21:31,790 --> 00:21:35,359
S1: The other aspect is that human 2.0, I would say

388
00:21:35,359 --> 00:21:39,950
S1: current model is essentially most people believing that they're not special,

389
00:21:39,950 --> 00:21:44,600
S1: most people believing that only special people write books, write blogs,

390
00:21:44,600 --> 00:21:49,760
S1: give presentations, do things like that, and regular people don't.

391
00:21:49,760 --> 00:21:54,440
S1: So two main things bifurcated lives broken into personal and professional.

392
00:21:54,440 --> 00:21:58,340
S1: This is 2.0, and the percentage of people who think

393
00:21:58,340 --> 00:22:00,320
S1: that they have something to offer the world, which I

394
00:22:00,320 --> 00:22:04,490
S1: would argue is extremely low and human 3.0 is the

395
00:22:04,490 --> 00:22:08,090
S1: transition to a world where people understand that everyone has

396
00:22:08,090 --> 00:22:10,280
S1: something to offer to the world, and they live as

397
00:22:10,280 --> 00:22:14,720
S1: full spectrum people. So it's like living for other humans,

398
00:22:14,720 --> 00:22:19,910
S1: living as humans, for humans, as opposed to living as

399
00:22:19,910 --> 00:22:24,919
S1: humans who like it's like living to work instead of

400
00:22:24,920 --> 00:22:28,190
S1: working to live and wouldn't even be working to live.

401
00:22:28,220 --> 00:22:30,649
S1: Human 3.0 would even be more advanced than that. It

402
00:22:30,650 --> 00:22:32,989
S1: would be. So it means like your public profile is

403
00:22:32,990 --> 00:22:36,230
S1: everything about you, right? How caring you are, how smart

404
00:22:36,230 --> 00:22:39,680
S1: you are, how funny you are, your favorite things in life,

405
00:22:39,680 --> 00:22:42,800
S1: your best ideas, the projects you like to work on,

406
00:22:42,800 --> 00:22:45,560
S1: the problems you think are important and meaningful in the world,

407
00:22:45,560 --> 00:22:49,760
S1: plus your technical skills, of course. And they know that

408
00:22:49,760 --> 00:22:53,450
S1: they're valuable because they are all these things, not just

409
00:22:53,450 --> 00:22:56,630
S1: what they can do for a corporation. And so my

410
00:22:56,630 --> 00:22:59,450
S1: thoughts on this are like, because you have to start wondering,

411
00:22:59,450 --> 00:23:03,080
S1: when you listen to the Leopold and Aakash conversation, it's like,

412
00:23:03,080 --> 00:23:06,800
S1: is human 3.0 in danger? Are we in danger of

413
00:23:06,800 --> 00:23:11,270
S1: losing a world where humans exist and do what they

414
00:23:11,270 --> 00:23:14,240
S1: do for other humans? Because ASI is going to take

415
00:23:14,240 --> 00:23:18,320
S1: over and whatever turn us into whatever? Maybe, hopefully. Well,

416
00:23:18,320 --> 00:23:22,190
S1: I'm pushing for a possible AGI or a that allows

417
00:23:22,190 --> 00:23:25,790
S1: us to do human 3.0 either. It would be kind

418
00:23:25,790 --> 00:23:28,160
S1: of scary if it just took over and implemented it.

419
00:23:28,160 --> 00:23:30,740
S1: That would be scary. Hopefully it would arrive at something

420
00:23:30,740 --> 00:23:34,580
S1: like human 3.0 and keep us alive and thriving. Or

421
00:23:34,580 --> 00:23:37,070
S1: it doesn't take over. It's still used as a tool,

422
00:23:37,070 --> 00:23:40,640
S1: but it's in the hands of someone somewhat benign like

423
00:23:40,640 --> 00:23:45,919
S1: the US, hopefully remains benign. And at that point we

424
00:23:45,950 --> 00:23:50,300
S1: we basically give it the criteria of help us maintain

425
00:23:50,300 --> 00:23:53,240
S1: a society in which we move towards and maintain something

426
00:23:53,240 --> 00:23:56,390
S1: like human 3.0. Another thing I'm really worried about is

427
00:23:56,390 --> 00:24:00,290
S1: like how immigration will affect UBI. So if the government

428
00:24:00,290 --> 00:24:03,950
S1: is basically the jobs are gone, know very few people

429
00:24:03,950 --> 00:24:07,040
S1: can do something that I can't. And that's like, whatever, 10%

430
00:24:07,040 --> 00:24:09,649
S1: of the world or 1% of the world, or 25%

431
00:24:09,650 --> 00:24:13,609
S1: of the world, whatever. Different times, different numbers, different ways

432
00:24:13,609 --> 00:24:17,090
S1: of measuring. But let's say it's 10% and the other 90%

433
00:24:17,090 --> 00:24:21,050
S1: just aren't employable because, like Harari said, you know, they're

434
00:24:21,050 --> 00:24:23,450
S1: a useless class. They have nothing that they can offer

435
00:24:23,450 --> 00:24:26,359
S1: the economy. And this is human 2.0. Still, this is

436
00:24:26,359 --> 00:24:30,440
S1: not crossed over into this hopeful land of human 3.0.

437
00:24:30,440 --> 00:24:33,560
S1: In the meantime, I mean, because some people, like Leopold

438
00:24:33,560 --> 00:24:36,949
S1: are worried about ASI happening in like a number of

439
00:24:36,950 --> 00:24:41,900
S1: years and just like completely redoing society and the government's

440
00:24:41,900 --> 00:24:43,699
S1: obviously going to have to be involved there. They're going

441
00:24:43,700 --> 00:24:47,420
S1: to have to pay people to basically not riot, not

442
00:24:47,420 --> 00:24:53,480
S1: kill themselves and, you know, provide education, provide gaming, provide entertainment,

443
00:24:53,480 --> 00:24:57,590
S1: provide something some sort of basis for meaning that's going

444
00:24:57,590 --> 00:25:00,649
S1: to have to be included. But what does that look like?

445
00:25:00,650 --> 00:25:04,310
S1: I mean, there's like a welfare state and there's like

446
00:25:04,310 --> 00:25:07,130
S1: the the government's giving out too much money, and then

447
00:25:07,130 --> 00:25:09,919
S1: there's a whole nother level when it's actually UBI, because

448
00:25:09,920 --> 00:25:13,010
S1: no one's even able to work. Right. It's not that

449
00:25:13,010 --> 00:25:16,550
S1: they can't find a job when when people are actually hiring,

450
00:25:16,550 --> 00:25:19,160
S1: it's like, no. Nobody's hiring. All the work's being done

451
00:25:19,160 --> 00:25:23,900
S1: by AI or whatever, 90%, and AI is producing trillions

452
00:25:23,900 --> 00:25:27,470
S1: of dollars in value. And that raises another question. Okay,

453
00:25:27,470 --> 00:25:30,110
S1: you're making all this new stuff. You have all these

454
00:25:30,109 --> 00:25:34,400
S1: new creators, you have all this new AI generated productivity

455
00:25:34,520 --> 00:25:39,419
S1: and products and services. For who? Who's buying the stuff?

456
00:25:39,420 --> 00:25:42,300
S1: Who's buying the stuff? If the top 10% are the

457
00:25:42,300 --> 00:25:45,689
S1: only ones who can make things and sell things. Who's

458
00:25:45,690 --> 00:25:49,410
S1: buying the stuff? And the natural answer looks like, which

459
00:25:49,410 --> 00:25:54,510
S1: is quite gross. It's like, I guess the government taxes

460
00:25:54,510 --> 00:25:57,840
S1: all the income that's being made. Again, where's the income

461
00:25:57,840 --> 00:26:00,869
S1: coming from? It's almost like there's an exchange between the

462
00:26:00,869 --> 00:26:04,590
S1: government and the producers, and the producers are making all

463
00:26:04,590 --> 00:26:07,050
S1: this awesome stuff. The government is using all the awesome

464
00:26:07,050 --> 00:26:10,139
S1: stuff to generate all these awesome things. But then the

465
00:26:10,140 --> 00:26:12,959
S1: government has to take this money, which is very strange

466
00:26:12,960 --> 00:26:15,390
S1: word for it now, because it's not being generated by

467
00:26:15,390 --> 00:26:18,690
S1: the people and being taxed by the government. In fact,

468
00:26:18,690 --> 00:26:22,260
S1: it's like it's just a money idea. The money idea

469
00:26:22,260 --> 00:26:26,790
S1: goes to the government. The government then gives the population money,

470
00:26:26,790 --> 00:26:29,580
S1: the ability to purchase these goods and services, which are

471
00:26:29,580 --> 00:26:32,580
S1: being generated by the small percentage of the population to

472
00:26:32,580 --> 00:26:36,240
S1: the tune of whatever trillions of trillions upon trillions of dollars.

473
00:26:36,240 --> 00:26:38,550
S1: And then they give that money out to the people

474
00:26:38,550 --> 00:26:41,100
S1: and the people that can then use that to buy

475
00:26:41,100 --> 00:26:43,800
S1: the services that are being made by this 10%. And

476
00:26:43,800 --> 00:26:47,340
S1: that feels very not quite Ponzi, but like shell game,

477
00:26:47,340 --> 00:26:50,310
S1: it's like, oh, it's fake money. We're moving it around.

478
00:26:50,310 --> 00:26:53,520
S1: It just feels very strange when when the, the human

479
00:26:53,520 --> 00:26:59,070
S1: population is not producing not participating in that creation. Now

480
00:26:59,070 --> 00:27:02,880
S1: in human 3.0, I think they will be I'm not

481
00:27:02,880 --> 00:27:05,340
S1: sure exactly what that economy will look like. I'm sure

482
00:27:05,340 --> 00:27:08,490
S1: much smarter people can help me think about that. It's

483
00:27:08,490 --> 00:27:12,449
S1: essentially more of a human to human exchange barter system.

484
00:27:12,450 --> 00:27:15,840
S1: Obviously a currency like a maybe like a doctor, a

485
00:27:15,960 --> 00:27:19,139
S1: coffee type of currency or something. But you can also

486
00:27:19,140 --> 00:27:22,080
S1: buy regular goods and services, but you could also pay

487
00:27:22,080 --> 00:27:25,560
S1: people for making you happy or whatever. Like a lot

488
00:27:25,560 --> 00:27:27,570
S1: of people will be creating art and doing lots of

489
00:27:27,570 --> 00:27:30,570
S1: stuff like that. So I'm hoping there's a large economy

490
00:27:30,570 --> 00:27:33,450
S1: for sharing on this full spectrum that we talked about

491
00:27:33,450 --> 00:27:37,020
S1: with human 3.0, and that will involve money and stuff

492
00:27:37,020 --> 00:27:40,109
S1: like that. But what I'm talking about here is really

493
00:27:40,109 --> 00:27:43,440
S1: human 2.0 stuff. This is like this year, next year,

494
00:27:43,440 --> 00:27:45,930
S1: the next five years, the next ten years, you know,

495
00:27:45,930 --> 00:27:47,730
S1: what do we do in the meantime with all these

496
00:27:47,730 --> 00:27:50,610
S1: people who are going to be needing money? I mean,

497
00:27:50,609 --> 00:27:54,600
S1: what happens with an economy with millions of immigrants who

498
00:27:54,600 --> 00:28:00,150
S1: are undocumented, who have families, they need to feed them,

499
00:28:00,150 --> 00:28:02,490
S1: they need to house them. And what if the government

500
00:28:02,490 --> 00:28:05,160
S1: is like, well, what if the jobs go away? Well,

501
00:28:05,160 --> 00:28:07,800
S1: maybe they leave. Maybe they don't leave, though, and they're

502
00:28:07,800 --> 00:28:10,740
S1: very angry because they can no longer get work, and

503
00:28:10,740 --> 00:28:14,580
S1: they're also not being paid like the local citizens. I

504
00:28:14,580 --> 00:28:17,639
S1: don't know what the solution here is, but it's I'm

505
00:28:17,640 --> 00:28:19,530
S1: saying we need to think about it. So my thoughts

506
00:28:19,530 --> 00:28:24,180
S1: on conscious AI are, I think, fairly simple actually. I

507
00:28:24,180 --> 00:28:27,330
S1: think the way humans got consciousness is that it was

508
00:28:27,330 --> 00:28:31,350
S1: adaptive for helping us accelerate evolution. So basically winning and

509
00:28:31,350 --> 00:28:35,159
S1: losing is what powers, evolution, blame and praise are smaller

510
00:28:35,160 --> 00:28:38,730
S1: versions of winning and losing plants and insects win and

511
00:28:38,730 --> 00:28:43,140
S1: lose as well, evolutionarily, but it seems like only humans

512
00:28:43,140 --> 00:28:45,750
S1: and a few other species. And maybe it's a lot,

513
00:28:45,750 --> 00:28:50,190
S1: but it's a subset. Very small subset, have subjective experiences,

514
00:28:50,190 --> 00:28:53,130
S1: and it's not clear that anyone else other than humans

515
00:28:53,130 --> 00:28:56,340
S1: has like blame and praise. Well, maybe dogs do. I

516
00:28:56,340 --> 00:29:00,390
S1: feel like they understand, blame and praise. Good boy, bad girl, right?

517
00:29:00,390 --> 00:29:03,240
S1: I feel like that works based on their facial expressions.

518
00:29:03,240 --> 00:29:06,360
S1: Who knows if that's right. But anyway, bottom line is

519
00:29:06,360 --> 00:29:11,340
S1: blame and praise might empower evolution if we experience winning

520
00:29:11,340 --> 00:29:15,390
S1: and losing, being blamed and praised and feeling that at

521
00:29:15,390 --> 00:29:20,340
S1: a visceral level versus having no inter like immediate repercussions

522
00:29:20,340 --> 00:29:23,190
S1: of not doing well. So maybe evolution still gets you

523
00:29:23,190 --> 00:29:27,780
S1: there if you don't have blame and praise and subjective experience.

524
00:29:27,780 --> 00:29:31,710
S1: But I think this whole kind of container is how

525
00:29:31,710 --> 00:29:34,830
S1: we ended up being conscious, because it enables this, which

526
00:29:34,830 --> 00:29:38,670
S1: makes evolution better. That that's my guess. If that's right,

527
00:29:38,670 --> 00:29:42,810
S1: then consciousness might have simply emerged from evolution as our

528
00:29:42,810 --> 00:29:46,440
S1: brains got bigger and we became better at getting better,

529
00:29:46,440 --> 00:29:53,040
S1: which means it could emerge naturally through like self similar loops,

530
00:29:53,040 --> 00:29:56,400
S1: definitely some RL related stuff. Or we could kind of

531
00:29:56,400 --> 00:29:58,710
S1: just like hack the system and like add it on

532
00:29:58,710 --> 00:30:02,010
S1: and be like, hey, develop this so that you can

533
00:30:02,010 --> 00:30:04,680
S1: have this advantage or whatever. We can try to shortcut

534
00:30:04,680 --> 00:30:07,620
S1: what evolution did over millions of years. Yeah. And then

535
00:30:07,620 --> 00:30:10,860
S1: one question is like, well, if we hacked it that way,

536
00:30:10,860 --> 00:30:13,650
S1: would it be real consciousness? And me as a materialist,

537
00:30:13,650 --> 00:30:17,070
S1: I'm like, well, what is real consciousness like? I think

538
00:30:17,070 --> 00:30:20,700
S1: at some level it's all an illusion anyway. But practically,

539
00:30:20,700 --> 00:30:25,350
S1: if you feel yourself feeling and it hurts, that's real, right?

540
00:30:25,350 --> 00:30:29,790
S1: It doesn't matter how mechanistic the substrate below actually is,

541
00:30:29,790 --> 00:30:32,940
S1: because at some point you can pick along the way

542
00:30:32,940 --> 00:30:36,000
S1: and just be like, this is mechanical, this is mechanical.

543
00:30:36,000 --> 00:30:40,340
S1: This is. Mechanical. Therefore it's all mechanical. Doesn't matter. Friendship

544
00:30:40,340 --> 00:30:43,640
S1: still is awesome. Love is awesome. Ice cream is awesome.

545
00:30:43,640 --> 00:30:45,739
S1: Doesn't make it any less awesome than it's made out

546
00:30:45,740 --> 00:30:50,060
S1: of atoms, right? So my AC prediction I'm ferm on

547
00:30:50,060 --> 00:30:54,770
S1: my AGI prediction of 2025 and 2028. Somewhere in there,

548
00:30:54,770 --> 00:30:59,480
S1: maybe 2026 2027 is like ideal. The sweet spot for me.

549
00:30:59,480 --> 00:31:03,350
S1: I'm going to give soft prediction, soft prediction of a

550
00:31:03,380 --> 00:31:07,940
S1: C 2027 to 31. So what is that. That's three

551
00:31:07,940 --> 00:31:12,440
S1: four plus two. So that's roughly 2029. Don't feel great

552
00:31:12,440 --> 00:31:16,010
S1: about it. That's like a 30 to 40% prediction like

553
00:31:16,010 --> 00:31:21,470
S1: confidence level. This one I'm saying 90% by 2028. End

554
00:31:21,470 --> 00:31:25,760
S1: of 2028 I'm saying like 60% by end of 2025.

555
00:31:25,760 --> 00:31:27,740
S1: So I feel a lot more confident about that one.

556
00:31:27,740 --> 00:31:30,860
S1: But who knows. Predictions are hard, especially when they're about

557
00:31:30,860 --> 00:31:35,450
S1: the future. Niels Bohr said that. All right, unifying thoughts. Leopold,

558
00:31:35,450 --> 00:31:39,440
S1: you want to watch him on AI safety? Completely. Absolutely

559
00:31:39,440 --> 00:31:41,990
S1: impressive guy. And you want to go read his essays

560
00:31:41,990 --> 00:31:44,870
S1: that are released on this as well. Talk to Aakash

561
00:31:44,870 --> 00:31:48,740
S1: Patel is becoming one of my favorite podcasters ever right now.

562
00:31:48,740 --> 00:31:53,719
S1: He is basically an autodidact. He's a crazy reader. When

563
00:31:53,720 --> 00:31:57,830
S1: Dwarkesh talks to Cowan or Cowan, that's how he pronounces

564
00:31:57,830 --> 00:32:02,270
S1: it Tyler Cowan. When they talk, they jump around, they're like,

565
00:32:02,270 --> 00:32:04,370
S1: oh yeah, you know, I was I was knitting this

566
00:32:04,370 --> 00:32:06,830
S1: thing the other day and I was working. I was,

567
00:32:06,830 --> 00:32:10,190
S1: you know, making wood figurines on my lathe in the

568
00:32:10,190 --> 00:32:13,820
S1: back backyard. And I was thinking about, you know, whatever,

569
00:32:13,820 --> 00:32:16,280
S1: Roman history. And then I was thinking about, you know,

570
00:32:16,280 --> 00:32:19,700
S1: Adams economics. And that reminds me of this one poem

571
00:32:19,700 --> 00:32:22,910
S1: by Rilke or whatever. And they just jump from here

572
00:32:22,910 --> 00:32:25,460
S1: to there and they're like, oh, that reminds me of this. Hey,

573
00:32:25,460 --> 00:32:27,860
S1: what about World War Two? And this thing happened and

574
00:32:27,860 --> 00:32:30,230
S1: they're just jumping from place to place and like, I

575
00:32:30,230 --> 00:32:32,180
S1: feel like I'm pretty good at that. I feel like

576
00:32:32,180 --> 00:32:36,469
S1: that's one of my superpowers. But Cowan, seeing Cowan do this,

577
00:32:36,470 --> 00:32:40,370
S1: it is insane. And Dwarkesh is like a mini Cowan.

578
00:32:40,370 --> 00:32:43,610
S1: He's like coming up so fast. It's the reason that

579
00:32:43,610 --> 00:32:48,110
S1: his interviews are so good. He's, like, fiercely curious and

580
00:32:48,110 --> 00:32:52,250
S1: an autodidact, reading, like, whatever, 12 books a second, whatever

581
00:32:52,250 --> 00:32:57,350
S1: his current metric is, it's insane. Anyway. Follow them. It's really,

582
00:32:57,350 --> 00:33:02,090
S1: really good stuff. And human 3.0. My best optimism here

583
00:33:02,090 --> 00:33:05,210
S1: is we get AGI, but we don't get AC for

584
00:33:05,210 --> 00:33:07,550
S1: a while, which will give us a chance to hopefully

585
00:33:07,550 --> 00:33:11,210
S1: move towards 3.0 or we get AC, but it's like

586
00:33:11,210 --> 00:33:14,690
S1: controlled by the US and or benign US actors and

587
00:33:14,690 --> 00:33:20,090
S1: we move towards human 3.0 or AC comes online, it's

588
00:33:20,090 --> 00:33:23,330
S1: benign and it kind of takes over and it builds

589
00:33:23,330 --> 00:33:26,630
S1: human 3.0 because that's the best future for humanity or

590
00:33:26,630 --> 00:33:29,300
S1: something that looks similar. Right. And this is what I

591
00:33:29,300 --> 00:33:32,390
S1: was saying before. I don't actually care how unlikely these

592
00:33:32,390 --> 00:33:37,010
S1: scenarios are, because it's kind of the only good path

593
00:33:37,010 --> 00:33:39,620
S1: that I could see happening from all this AI stuff

594
00:33:39,620 --> 00:33:43,980
S1: is we eventually get to human 3.0. A a human

595
00:33:43,980 --> 00:33:47,940
S1: world for humans, and I would say future versions of

596
00:33:47,940 --> 00:33:51,540
S1: humans which might be hybrids with AI or whatever. But

597
00:33:51,540 --> 00:33:54,030
S1: you got to maintain the humanity. We got to be

598
00:33:54,030 --> 00:33:57,180
S1: doing what we're doing for humans, not for the sake

599
00:33:57,180 --> 00:34:00,870
S1: of tech. So I'm optimistic about that. I am trying

600
00:34:00,870 --> 00:34:02,880
S1: to help build that. That's my whole purpose in like

601
00:34:02,880 --> 00:34:06,450
S1: doing all of this so that that's what I'm shooting for.

602
00:34:06,450 --> 00:34:09,299
S1: And I don't care if it's a 20% chance of

603
00:34:09,300 --> 00:34:11,339
S1: us getting there. First of all, nobody can know what

604
00:34:11,340 --> 00:34:14,190
S1: this number is, so it doesn't really matter. But second

605
00:34:14,190 --> 00:34:17,040
S1: of all, if it's a 10% chance, guess what? I'm

606
00:34:17,040 --> 00:34:19,260
S1: shooting for the 10%. I'm trying to make it 11.

607
00:34:19,260 --> 00:34:21,780
S1: Try to make it a 50, whatever. Try to make

608
00:34:21,780 --> 00:34:24,210
S1: it a 90. If it's a 60% chance and it's

609
00:34:24,210 --> 00:34:26,610
S1: a 40% chance, we're all going to die. Whatever. I'm

610
00:34:26,610 --> 00:34:29,310
S1: not going to sit around thinking about the 40%. I mean,

611
00:34:29,310 --> 00:34:33,270
S1: that's my vibe. I am shooting for human 3.0. And

612
00:34:33,270 --> 00:34:35,730
S1: if someone's like, yeah, well, you don't know, China can

613
00:34:35,730 --> 00:34:37,770
S1: take it over and blah blah blah asi and it

614
00:34:37,770 --> 00:34:41,069
S1: can kill us all and whatever paper clips and I

615
00:34:41,070 --> 00:34:45,270
S1: just don't see the value of going hard down the

616
00:34:45,270 --> 00:34:50,250
S1: path of like, the doomers because let's say they're right.

617
00:34:50,250 --> 00:34:53,460
S1: Their answer is turn it off, don't build it. That's

618
00:34:53,460 --> 00:34:56,850
S1: not going to happen. Next answer. They're like, I don't

619
00:34:56,850 --> 00:34:59,160
S1: have a next answer. We should stop talking about this.

620
00:34:59,160 --> 00:35:01,680
S1: We should stop building it. You have exited yourself from

621
00:35:01,680 --> 00:35:04,170
S1: the adult table. You are no longer in the conversation

622
00:35:04,170 --> 00:35:08,430
S1: because of Moloch. We are building this. It is happening.

623
00:35:08,430 --> 00:35:12,030
S1: We are running full speed with scissors in both hands.

624
00:35:12,030 --> 00:35:15,480
S1: And there's nothing we can do to stop that. My

625
00:35:15,480 --> 00:35:17,969
S1: argument is, the only thing we could do is try

626
00:35:17,969 --> 00:35:20,370
S1: to aim it in a good direction. Try to point

627
00:35:20,370 --> 00:35:23,520
S1: towards human 3.0 and that's what I'm trying to do.

628
00:35:23,520 --> 00:35:28,410
S1: All right, huge diversion, but worth talking about. That was

629
00:35:28,410 --> 00:35:31,740
S1: the AI piece collection of thoughts and predictions about AI.

630
00:35:31,770 --> 00:35:35,640
S1: June of 2024. All right. Security got a whole bunch

631
00:35:35,640 --> 00:35:40,620
S1: of misinformation sneaking into like, little man stuffed news. So

632
00:35:40,620 --> 00:35:44,070
S1: this is like manosphere, like podcasts or whatever, or news

633
00:35:44,070 --> 00:35:50,160
S1: sites or whatever. They're actually republishing Russian propaganda, RT network content.

634
00:35:50,160 --> 00:35:52,890
S1: So it's like, yeah, it can seep in and go

635
00:35:52,890 --> 00:35:55,080
S1: lots of different places. This is why you actually need

636
00:35:55,110 --> 00:35:58,950
S1: AI to be watching everything. And basically when a new

637
00:35:58,950 --> 00:36:04,470
S1: information or disinformation or misinformation propaganda campaign comes out, when

638
00:36:04,469 --> 00:36:08,310
S1: it comes out, we should categorize, okay, what the core

639
00:36:08,310 --> 00:36:12,000
S1: concept is, the core idea. It's trying to spread with AI,

640
00:36:12,000 --> 00:36:15,299
S1: and then you monitor all the different stories coming out.

641
00:36:15,300 --> 00:36:19,950
S1: It might be like Saskatchewan news from South Parish, you know,

642
00:36:19,950 --> 00:36:24,450
S1: Wisconsin or whatever. And they're spreading the same story, because

643
00:36:24,450 --> 00:36:26,850
S1: what the attackers will do is they'll figure out, okay,

644
00:36:26,850 --> 00:36:29,400
S1: where can we get it in? Where can we push

645
00:36:29,400 --> 00:36:31,920
S1: this narrative? And maybe, maybe the front door is closed,

646
00:36:31,920 --> 00:36:35,129
S1: but maybe there's a thousand different backdoors. That was that one.

647
00:36:35,130 --> 00:36:38,010
S1: TikTok had an issue with DMs. There was a zero

648
00:36:38,010 --> 00:36:40,350
S1: day in DMs where you click it and open it.

649
00:36:40,350 --> 00:36:44,460
S1: It's pretty bad. The current tactic for what you basically

650
00:36:44,460 --> 00:36:49,200
S1: get compromise for compromise of your TikTok if you did that. And, uh,

651
00:36:49,200 --> 00:36:53,730
S1: new tactic TikTok is using to avoid the lawsuit is

652
00:36:53,730 --> 00:36:56,100
S1: basically to say they're going to build out a separate

653
00:36:56,100 --> 00:36:58,950
S1: algorithm and separate data just for us. So we'll see

654
00:36:58,950 --> 00:37:01,830
S1: if that works. New paper claims that GPT four can

655
00:37:01,830 --> 00:37:06,180
S1: autonomously hack zero day vulnerabilities. 53% success rate going to

656
00:37:06,180 --> 00:37:07,500
S1: be a lot of these papers. A lot of these

657
00:37:07,500 --> 00:37:11,009
S1: papers are not very good. A lot of them are decent.

658
00:37:11,010 --> 00:37:13,440
S1: I will keep reading and posting these when the claims

659
00:37:13,440 --> 00:37:17,040
S1: are interesting and they're from decent outfits, but you want

660
00:37:17,040 --> 00:37:20,549
S1: to watch the quality of the paper very closely. Look,

661
00:37:20,580 --> 00:37:24,390
S1: especially for how much of their testing methodology that they publish.

662
00:37:24,390 --> 00:37:27,000
S1: Did they show exactly what they tried against these different things?

663
00:37:27,000 --> 00:37:30,419
S1: Did they show all the different examples? Did they, you know,

664
00:37:30,420 --> 00:37:33,750
S1: give you all the stuff to do the replication? If not,

665
00:37:33,750 --> 00:37:36,000
S1: then they're kind of like hiding the ball and you

666
00:37:36,000 --> 00:37:39,270
S1: can't really trust their score. And we got to analyze

667
00:37:39,270 --> 00:37:42,149
S1: paper fabric pattern that you can use for this. And

668
00:37:42,150 --> 00:37:44,190
S1: when I ran it against this one it got a

669
00:37:44,190 --> 00:37:48,900
S1: six on rigor benchmark and Ablations look good but some

670
00:37:48,900 --> 00:37:52,890
S1: details missing and stats lacking. That's why that's why we

671
00:37:52,890 --> 00:37:56,910
S1: have fabric. Okay Chinese drone photographer got snagged by the

672
00:37:56,910 --> 00:38:02,850
S1: US Espionage Act taking pics of military shipyards. So there's

673
00:38:02,850 --> 00:38:05,460
S1: another thing that's being talked about right now with China

674
00:38:05,460 --> 00:38:08,280
S1: buying up a whole bunch of land right next to

675
00:38:08,280 --> 00:38:12,270
S1: military bases, they're buying like massive amounts of farmland, and

676
00:38:12,270 --> 00:38:14,279
S1: they start setting up all this gear. It's like, oh,

677
00:38:14,280 --> 00:38:16,770
S1: don't pay any attention to us. We're just building some

678
00:38:16,770 --> 00:38:20,490
S1: stuff over here. So I hope Biden or Trump or

679
00:38:20,489 --> 00:38:23,370
S1: whoever just like, finds a list of these. And just

680
00:38:23,370 --> 00:38:27,239
S1: like search and seizure, good bye. Do not pass go.

681
00:38:27,239 --> 00:38:30,930
S1: And maybe some of those are benign, right. But I

682
00:38:30,930 --> 00:38:36,630
S1: have to basically be like, well, based on the, you know, 9566,

683
00:38:36,719 --> 00:38:40,650
S1: you know, fucking billion times that you've actually tried to

684
00:38:40,650 --> 00:38:44,880
S1: steal from us and. And actually hacked us, then you

685
00:38:44,880 --> 00:38:47,279
S1: know you're going to lose some for the home team

686
00:38:47,280 --> 00:38:50,490
S1: when you're actually doing something benign and actually just setting

687
00:38:50,489 --> 00:38:53,459
S1: up a farm next to a military base on accident.

688
00:38:53,460 --> 00:38:56,250
S1: And they didn't actually mean anything by that. Too bad

689
00:38:56,250 --> 00:38:59,790
S1: you ruined your reputation. Therefore get out. Okay. OpenAI revealed

690
00:38:59,790 --> 00:39:02,069
S1: that its AI tech was used by Russia and China

691
00:39:02,070 --> 00:39:05,610
S1: for sneaky influence ops. Yeah, they've said this a few

692
00:39:05,610 --> 00:39:08,879
S1: times and I like it. I like the transparency. They're

693
00:39:08,880 --> 00:39:12,360
S1: basically like, look, this keeps happening. We keep catching it

694
00:39:12,360 --> 00:39:14,520
S1: and we're going to try to do better. All right.

695
00:39:14,520 --> 00:39:19,259
S1: WWDC keynote got an analysis here. This is a fabric

696
00:39:19,260 --> 00:39:22,920
S1: analysis from a TechCrunch article. As I talked about before,

697
00:39:22,920 --> 00:39:25,710
S1: Apple basically killed the AI thing. They also did a

698
00:39:25,710 --> 00:39:27,450
S1: bunch of really cool stuff. One of the things is

699
00:39:27,450 --> 00:39:29,700
S1: you bring your phone next to someone else, you can

700
00:39:29,700 --> 00:39:32,910
S1: transfer them cash. I've been waiting that for that forever.

701
00:39:32,910 --> 00:39:35,460
S1: Problem is, a lot of people won't have that feature

702
00:39:35,460 --> 00:39:38,640
S1: turned on, or they'll have an Android and you won't

703
00:39:38,640 --> 00:39:40,770
S1: be able to do it like so. A lot of

704
00:39:40,770 --> 00:39:44,100
S1: people are switching to like Zelle or Venmo. I don't

705
00:39:44,100 --> 00:39:48,240
S1: like Venmo, and Zelle is slow. Whatever. I wish more

706
00:39:48,239 --> 00:39:51,660
S1: people had the ability to do these swiping things and

707
00:39:51,660 --> 00:39:55,319
S1: Apple things, but kind of rude to require that everyone

708
00:39:55,320 --> 00:39:58,350
S1: has an iPhone, right? I mean, that's just not realistic

709
00:39:58,350 --> 00:40:00,840
S1: and also not cool. All right. All the different startups

710
00:40:00,840 --> 00:40:06,060
S1: that Apple killed today always fun. After WWDC or an

711
00:40:06,060 --> 00:40:11,280
S1: OpenAI event, US needs 225,000 more cyber security workers. I

712
00:40:11,280 --> 00:40:14,879
S1: never trust numbers like this because it's hard to know. Like,

713
00:40:14,880 --> 00:40:18,000
S1: what are they asking? Are they asking it right? What

714
00:40:18,000 --> 00:40:20,580
S1: are the questions that they accept? Like, I just feel

715
00:40:20,580 --> 00:40:22,860
S1: like the numbers are bad. I feel like that number

716
00:40:22,860 --> 00:40:25,110
S1: is going to go down because of AI, and I

717
00:40:25,110 --> 00:40:27,150
S1: feel like it's going to go up because of AI.

718
00:40:27,180 --> 00:40:29,219
S1: So it's going to go down because of AI, because

719
00:40:29,219 --> 00:40:31,710
S1: the jobs are going to get easier to do without people,

720
00:40:31,710 --> 00:40:35,250
S1: but so much more stuff is going to get made

721
00:40:35,250 --> 00:40:38,730
S1: that we're just going to be injecting insecurity into the world,

722
00:40:38,730 --> 00:40:41,400
S1: especially when it's like getting made with AI. So it's

723
00:40:41,400 --> 00:40:43,830
S1: like AI going in, building a bunch of stuff and

724
00:40:43,830 --> 00:40:45,839
S1: have that stuff is going to be really nasty in

725
00:40:45,840 --> 00:40:48,989
S1: terms of security vulnerabilities. So it's like still going to

726
00:40:48,989 --> 00:40:53,280
S1: need people for a while, like assuming all this AI

727
00:40:53,310 --> 00:40:56,549
S1: replacement stuff goes really, really fast. We're still talking about

728
00:40:56,550 --> 00:41:01,109
S1: years and years. I mean, super fast for like an 80%

729
00:41:01,110 --> 00:41:05,160
S1: replacement of all jobs, super fast. And these are crazy

730
00:41:05,160 --> 00:41:08,759
S1: numbers I'm just throwing out. But but like Nuclear fast

731
00:41:08,760 --> 00:41:12,330
S1: is like ten years now as far as starting and

732
00:41:12,330 --> 00:41:15,480
S1: actually having major impact, I think it's already starting and

733
00:41:15,480 --> 00:41:18,180
S1: it's already like a year or two from now we're

734
00:41:18,180 --> 00:41:20,190
S1: going to start feeling it. So I think that's going

735
00:41:20,190 --> 00:41:23,489
S1: to be really fast. There's going to be a regulatory response.

736
00:41:23,489 --> 00:41:27,180
S1: It's slow to make any change in any organization, and

737
00:41:27,180 --> 00:41:29,940
S1: the bigger and more bureaucratic it is, the slower it

738
00:41:29,940 --> 00:41:33,330
S1: will be. So just because something could happen in a

739
00:41:33,330 --> 00:41:36,360
S1: year or six months doesn't mean it will. And it's

740
00:41:36,360 --> 00:41:39,120
S1: more likely to happen in 2 or 3 or 4

741
00:41:39,120 --> 00:41:42,420
S1: or 10 years. Got to keep that in mind. It's

742
00:41:42,420 --> 00:41:44,790
S1: going to go very fast. It's going to go faster

743
00:41:44,790 --> 00:41:47,190
S1: than you think and slower than you think both at

744
00:41:47,190 --> 00:41:50,520
S1: the same time. Apple just hit the three. $3 trillion

745
00:41:50,520 --> 00:41:53,790
S1: mark again. But Nvidia jumped ahead. And now Nvidia is

746
00:41:53,790 --> 00:41:58,140
S1: even more valuable than Apple. And Microsoft is still number

747
00:41:58,140 --> 00:42:02,069
S1: one in. Saint Satya is killing it over there. Cartwheel

748
00:42:02,100 --> 00:42:07,170
S1: tool olama CLI wired JS for us for UI design.

749
00:42:07,170 --> 00:42:10,740
S1: Shade map project that turns every mountain building and tree

750
00:42:10,739 --> 00:42:14,100
S1: shadow for any date and time. How to think like

751
00:42:14,100 --> 00:42:18,420
S1: a computer scientist. Pretty cool. Uh, interactive edition here. Starship

752
00:42:18,420 --> 00:42:21,750
S1: just landed its first one. Like all the other ones

753
00:42:21,750 --> 00:42:24,540
S1: either blew up on the way out or, uh, blew

754
00:42:24,540 --> 00:42:27,360
S1: up on the way in, but he finally landed one

755
00:42:27,360 --> 00:42:32,670
S1: very impressive, I think 5000 tons. So that's what, £20,000?

756
00:42:32,670 --> 00:42:38,340
S1: Five times two, £2,000. 2000 times 5000. Holy crap, that's

757
00:42:38,340 --> 00:42:42,150
S1: a lot of pounds. World's largest solar farm just went online.

758
00:42:42,150 --> 00:42:46,739
S1: Three point gigawatt Marty just went live in China. Okay,

759
00:42:46,739 --> 00:42:49,500
S1: here's my question. Why can't we build like ten of

760
00:42:49,500 --> 00:42:52,469
S1: these in the US? Okay, we know we need all

761
00:42:52,469 --> 00:42:56,250
S1: this power for AI. So I say we go Biden.

762
00:42:56,250 --> 00:42:58,860
S1: Trump does this. You basically go to each state and

763
00:42:58,860 --> 00:43:03,060
S1: be like, look, I need you to build 3.5GW in

764
00:43:03,060 --> 00:43:06,810
S1: your state within five years. And they're like, yeah, that's impossible.

765
00:43:06,810 --> 00:43:09,900
S1: We'd have to hire all these people. And the government's like, whatever.

766
00:43:09,900 --> 00:43:12,479
S1: I mean, we have money printers, so we're just going

767
00:43:12,480 --> 00:43:15,030
S1: to print all this money, we're going to inject it

768
00:43:15,030 --> 00:43:18,960
S1: into the economy. Most importantly, giving it to these training

769
00:43:18,960 --> 00:43:21,780
S1: programs will, which will train the people how to build

770
00:43:21,780 --> 00:43:24,660
S1: the wind farms and the solar farms, whatever the best

771
00:43:24,660 --> 00:43:28,469
S1: renewable energy is for that particular state, like Nevada should

772
00:43:28,469 --> 00:43:32,400
S1: build like 100 of these things for for sun. I mean,

773
00:43:32,400 --> 00:43:35,040
S1: I fly over it all the time. There's nothing there.

774
00:43:35,040 --> 00:43:37,680
S1: In fact, you fly over the entire United States. There's

775
00:43:37,680 --> 00:43:41,069
S1: nothing there. It's all empty. And meanwhile the sun just

776
00:43:41,070 --> 00:43:44,600
S1: hits it every. Every day, like nine times out of ten.

777
00:43:44,600 --> 00:43:47,210
S1: For every day of the year, the sun comes up.

778
00:43:47,210 --> 00:43:49,580
S1: In fact, I think it might be higher than 90%

779
00:43:49,580 --> 00:43:52,250
S1: when the sun comes up. In fact, it might be 100%.

780
00:43:52,250 --> 00:43:54,980
S1: The sun always comes up and it is hitting us

781
00:43:54,980 --> 00:43:59,000
S1: with like so much energy. And we just it soaks in,

782
00:43:59,090 --> 00:44:03,319
S1: you know, goes off into the atmosphere. Capture it. Okay, look,

783
00:44:03,320 --> 00:44:05,540
S1: we've got people who don't have jobs and don't have

784
00:44:05,540 --> 00:44:07,940
S1: meaning in their lives. We know that AI is going

785
00:44:07,940 --> 00:44:10,790
S1: to require all this power. So millions of people have

786
00:44:10,790 --> 00:44:12,710
S1: no meaning in their lives because they don't have work.

787
00:44:12,710 --> 00:44:16,220
S1: We talk about the loss of manufacturing jobs. Go build

788
00:44:16,219 --> 00:44:19,370
S1: all these solar panels, go build all these solar plants,

789
00:44:19,370 --> 00:44:22,130
S1: have the people working there. Eventually they'll be replaced by

790
00:44:22,130 --> 00:44:24,890
S1: robots in AI. That's fine. That'll take time like we

791
00:44:24,890 --> 00:44:28,640
S1: talked about before. In the meantime, let's go build 100

792
00:44:28,640 --> 00:44:34,700
S1: of these, whatever, 3.5GW facilities or ten gigawatt facilities, whatever.

793
00:44:34,700 --> 00:44:39,469
S1: Giant solar farms, giant wind farms. And we put millions

794
00:44:39,469 --> 00:44:42,500
S1: of people to work, which gives them meaning, and they're

795
00:44:42,500 --> 00:44:45,290
S1: building towards a renewable future. So, like, even if I

796
00:44:45,290 --> 00:44:48,169
S1: didn't happen whatever, we got renewable energy. Now we can

797
00:44:48,170 --> 00:44:50,210
S1: export that energy. We could sell it to other places.

798
00:44:50,210 --> 00:44:53,120
S1: We just have it for whatever purpose. And you have

799
00:44:53,120 --> 00:44:54,950
S1: to build the pipelines as well, right? You got to

800
00:44:54,950 --> 00:44:57,080
S1: get that energy to where it needs to go. So

801
00:44:57,080 --> 00:45:01,760
S1: that's like pipes, cables, like, you know, wires, all that stuff,

802
00:45:01,760 --> 00:45:05,450
S1: all the infrastructure for moving energy. Where's the downside here?

803
00:45:05,450 --> 00:45:09,620
S1: It's win win jobs and and tons of energy and

804
00:45:09,620 --> 00:45:12,860
S1: no reliance on external energy. And we get to move

805
00:45:12,860 --> 00:45:16,400
S1: on AI faster. So really bothers me when I see

806
00:45:16,400 --> 00:45:20,480
S1: China understanding this and building so quickly. And we're like, man,

807
00:45:20,510 --> 00:45:23,779
S1: let's fight with each other instead. Oh, in nuclear, nuclear

808
00:45:23,780 --> 00:45:27,020
S1: is the other one. Okay, so USA's although I prefer solar,

809
00:45:27,020 --> 00:45:30,200
S1: but someone knows better than me. USA's solar panel manufacturing

810
00:45:30,200 --> 00:45:34,040
S1: capacity jumped by 71%. I like it, I like it.

811
00:45:34,040 --> 00:45:37,250
S1: Florida and Texas are leading. Let's go. What about Nevada?

812
00:45:37,250 --> 00:45:39,979
S1: Where are you at? Florida. Forget about it. It rains

813
00:45:39,980 --> 00:45:41,750
S1: all the time and it's sinking. It's not going to

814
00:45:41,750 --> 00:45:44,960
S1: work out. I guess Florida's energy generation would have to

815
00:45:44,960 --> 00:45:48,080
S1: be based on, like being flooded or wins. No, you

816
00:45:48,080 --> 00:45:50,899
S1: can't do can't do wind farms and hurricanes. That would

817
00:45:50,900 --> 00:45:53,930
S1: be a bit rough. All right. Humans. David Brooks argues

818
00:45:53,930 --> 00:45:56,540
S1: that progressive energy has shifted from the working class to

819
00:45:56,540 --> 00:46:00,770
S1: elite universities, making them more left leaning than ever. Points

820
00:46:00,770 --> 00:46:04,640
S1: to protests and progressive opinions. Yeah, it's all bad. Highly

821
00:46:04,640 --> 00:46:08,210
S1: recommend you read David Brooks stuff The Road to Character.

822
00:46:08,210 --> 00:46:13,430
S1: My favorite one is the Second Mountain. Absolutely unbelievable book. Unbelievable.

823
00:46:13,430 --> 00:46:15,560
S1: And I love his stuff. The New York Times, he's

824
00:46:15,560 --> 00:46:19,969
S1: like a centrist, uh, kind of conservative, but like, yeah, logical.

825
00:46:19,969 --> 00:46:22,130
S1: I like him. Okay. This one turned out to be

826
00:46:22,130 --> 00:46:25,220
S1: not great. A couple people. I looked at it again.

827
00:46:25,219 --> 00:46:28,130
S1: People don't like me linking to Substack. I'm going to

828
00:46:28,130 --> 00:46:31,190
S1: link to Substack. The question isn't, is it Substack or

829
00:46:31,190 --> 00:46:34,790
S1: is it whatever, New York Times, because all these different

830
00:46:34,790 --> 00:46:37,640
S1: sources can be bad. What I care about is like

831
00:46:37,640 --> 00:46:41,480
S1: factually being wrong or like making claims and not backing

832
00:46:41,480 --> 00:46:44,240
S1: them up in some sort of in any sort of way,

833
00:46:44,239 --> 00:46:48,380
S1: or just being ideologue and just like being hateful or

834
00:46:48,380 --> 00:46:51,530
S1: rambling or whatever, like, well, I'm rambling, so not quite

835
00:46:51,530 --> 00:46:54,560
S1: the same. But the point is, the best people in

836
00:46:54,560 --> 00:46:58,489
S1: the world at thinking and being rigorous and having good

837
00:46:58,489 --> 00:47:03,379
S1: ideas are going to Substack, okay, and beehive and all

838
00:47:03,380 --> 00:47:07,580
S1: these other platforms and YouTube and TikTok. Even so, it's

839
00:47:07,580 --> 00:47:10,069
S1: not the source that I'm linking to. That's the problem.

840
00:47:10,070 --> 00:47:13,219
S1: The quality of the content is the problem because before

841
00:47:13,219 --> 00:47:16,730
S1: too long, CNBC and like Fox, first of all, Fox News,

842
00:47:16,730 --> 00:47:18,920
S1: I don't link to that usually. I'm sure they have

843
00:47:18,920 --> 00:47:22,040
S1: good news sometimes, but you get the point. It's not

844
00:47:22,040 --> 00:47:26,120
S1: about the type of source, okay? Blogs are just as

845
00:47:26,120 --> 00:47:30,980
S1: authoritative as YouTube as a lot of the media networks

846
00:47:30,980 --> 00:47:33,470
S1: at this point. It's more about the individual and the

847
00:47:33,469 --> 00:47:37,339
S1: actual article. I am making these links available so that

848
00:47:37,340 --> 00:47:40,310
S1: someone can my link numbers or my click numbers are

849
00:47:40,310 --> 00:47:44,090
S1: probably going to go down because they will be like, well,

850
00:47:44,140 --> 00:47:46,239
S1: I don't want to read from this guy, or I

851
00:47:46,239 --> 00:47:48,640
S1: don't want to read from Hacker News or Financial Times.

852
00:47:48,640 --> 00:47:51,339
S1: I hate that place or whatever. Hopefully you don't need

853
00:47:51,340 --> 00:47:53,980
S1: it anyway. Hopefully this one sentence is good enough. That's

854
00:47:53,980 --> 00:47:55,930
S1: the whole point of doing the summary. All right. New

855
00:47:55,930 --> 00:47:59,469
S1: York Times deep dive into devastation of Ukraine. Oh, this

856
00:47:59,469 --> 00:48:03,399
S1: story was absolutely nuts. And yeah, New York Times good stuff.

857
00:48:03,400 --> 00:48:06,100
S1: They're staying around at a company more than two years

858
00:48:06,100 --> 00:48:10,690
S1: could massively hurt your earning potential. U.S. job openings hit

859
00:48:10,719 --> 00:48:15,460
S1: a three year low. Layoffs are surprisingly low, it says CNN.

860
00:48:15,460 --> 00:48:19,900
S1: See above for commentary on sources. Viagra improves blood flow

861
00:48:19,900 --> 00:48:24,910
S1: and could help to prevent dementia. University of Oxford 49%

862
00:48:24,910 --> 00:48:28,540
S1: of independents think Trump should drop out because of his

863
00:48:28,540 --> 00:48:31,870
S1: guilty verdict. I'm making a prediction. I think this effect

864
00:48:31,870 --> 00:48:34,569
S1: wears off and Trump will be pulling even with Biden

865
00:48:34,570 --> 00:48:36,460
S1: within a month. You know what I'm going to do?

866
00:48:36,460 --> 00:48:42,169
S1: Let's go look at 538 538 polls. What? Wow. Look

867
00:48:42,170 --> 00:48:46,910
S1: at the impact. Look at the impact of the guilty verdict. Yeah,

868
00:48:46,910 --> 00:48:51,020
S1: it happened right in here somewhere. Oh, look. No impact.

869
00:48:51,020 --> 00:48:53,930
S1: And maybe I'm reading this wrong. Maybe this doesn't account

870
00:48:53,930 --> 00:48:57,840
S1: for independence. I mean, this is this is a June

871
00:48:57,840 --> 00:49:02,870
S1: 12th 10th through 12th result. Echelon insights. I don't know

872
00:49:02,870 --> 00:49:07,219
S1: how big this thing is. Okay, YouGov I know that's respectable.

873
00:49:07,219 --> 00:49:12,050
S1: Biden 40, Trump 42. Okay. We got two different echelon

874
00:49:12,050 --> 00:49:16,310
S1: ones daily cause what do we got? 4545 tide. So

875
00:49:16,310 --> 00:49:20,629
S1: so I read this article and it's like, yeah, 49%

876
00:49:20,630 --> 00:49:23,689
S1: of independents think Trump should drop out after his guilty verdict.

877
00:49:23,690 --> 00:49:27,859
S1: Are these polls okay? That's one argument. Independents don't respond

878
00:49:27,860 --> 00:49:30,950
S1: to polls. And supposedly these independents are going to switch

879
00:49:30,950 --> 00:49:33,440
S1: the thing and they're going to be mad at Trump,

880
00:49:33,440 --> 00:49:36,109
S1: and it's going to go for Biden. I don't buy it.

881
00:49:36,320 --> 00:49:39,080
S1: If anything, in my opinion, this is just my opinion.

882
00:49:39,080 --> 00:49:42,980
S1: This is politics. Like, who knows anything, right? In my opinion,

883
00:49:42,980 --> 00:49:46,880
S1: these numbers are soft for Trump. I think Trump is

884
00:49:46,880 --> 00:49:49,760
S1: probably winning by another 5 or 10 points right now.

885
00:49:49,760 --> 00:49:54,109
S1: That's my guess. And I don't think him being convicted

886
00:49:54,110 --> 00:49:59,030
S1: is going to do anything for his numbers if if not,

887
00:49:59,030 --> 00:50:02,300
S1: make it better for him. Just ridiculous. This is why

888
00:50:02,300 --> 00:50:04,550
S1: I think about I. I literally do not look at

889
00:50:04,550 --> 00:50:08,900
S1: news anymore other other than like I brings me political

890
00:50:08,900 --> 00:50:11,390
S1: news and analyzes for me and then I have to

891
00:50:11,390 --> 00:50:13,580
S1: go like read the story, I have to do my

892
00:50:13,580 --> 00:50:16,370
S1: analysis or whatever. But like I do not pay attention

893
00:50:16,370 --> 00:50:19,009
S1: to this stuff. I'm trying to build human 3.0 I

894
00:50:19,010 --> 00:50:22,460
S1: don't like. I'm just so disillusioned with everything. I'm just

895
00:50:22,460 --> 00:50:25,760
S1: like checked out. All right, X is doing adult content.

896
00:50:25,760 --> 00:50:29,390
S1: Good for them. Ernest Hemingway wrote a deeply moving letter

897
00:50:29,390 --> 00:50:32,839
S1: to friends who lost their son, saying, those who truly

898
00:50:32,840 --> 00:50:40,120
S1: live never really die. I love Maria Popova. The Marginalien. Yeah,

899
00:50:40,120 --> 00:50:42,040
S1: I like the old name that she had for that

900
00:50:42,040 --> 00:50:45,790
S1: site before. I can't remember what it was. The yellow site. Anyway,

901
00:50:45,790 --> 00:50:48,250
S1: one of the coolest life wisdom reads I've come across

902
00:50:48,250 --> 00:50:50,049
S1: in a while. This is a really cool one. I'll

903
00:50:50,050 --> 00:50:52,690
S1: put this one up. You can never expect honesty from

904
00:50:52,690 --> 00:50:55,600
S1: people who even lie to themselves. You'll be ten times

905
00:50:55,600 --> 00:50:57,970
S1: happier if you forgive your parents and stop blaming them

906
00:50:57,969 --> 00:51:01,630
S1: for your misery for yourself. From society's advice, most of

907
00:51:01,630 --> 00:51:03,880
S1: them have no idea of what they're doing. If you

908
00:51:03,880 --> 00:51:07,030
S1: continue to wait for the right time, you'll waste your

909
00:51:07,030 --> 00:51:09,910
S1: entire life and nothing will ever happen. It's a massive

910
00:51:09,910 --> 00:51:13,270
S1: list of these. They're quite, quite good. You'll lose 99%

911
00:51:13,270 --> 00:51:16,150
S1: of your close friends when you start to improve your life.

912
00:51:16,150 --> 00:51:19,390
S1: That's bleak. I wouldn't go with 99%, but I get

913
00:51:19,390 --> 00:51:23,290
S1: the point. You don't need 100 self-help books. All you

914
00:51:23,290 --> 00:51:26,650
S1: need is action and self-discipline. It's a great list. It's

915
00:51:26,650 --> 00:51:29,080
S1: a great list. You should check it out. Ideas and

916
00:51:29,080 --> 00:51:33,310
S1: analysis I think I figured out the simplest possible answer

917
00:51:33,310 --> 00:51:37,030
S1: for why Trump is still doing so well. Oh my god,

918
00:51:37,030 --> 00:51:40,600
S1: I didn't include the link. Oh man, I didn't include

919
00:51:40,600 --> 00:51:43,540
S1: the link in here. It's like, oh really? You did great.

920
00:51:43,540 --> 00:51:46,270
S1: Thanks for not sharing it. And recommendation for the week

921
00:51:46,270 --> 00:51:50,020
S1: is Tyler Cowen and Aakash Patel. As I talked about

922
00:51:50,020 --> 00:51:52,480
S1: in the aphorism of the week live as if you

923
00:51:52,480 --> 00:51:55,180
S1: were to die tomorrow. Learn as if you were to

924
00:51:55,180 --> 00:51:58,330
S1: live forever. Live as if you were to die tomorrow.

925
00:51:58,330 --> 00:52:01,930
S1: Learn as if you were to live forever. Mahatma Gandhi.

926
00:52:03,150 --> 00:52:06,239
S1: Unsupervised Learning is produced and edited by Daniel Meisler on

927
00:52:06,239 --> 00:52:10,860
S1: a Neumann U87 AI microphone using Hindenburg. Intro and outro

928
00:52:10,860 --> 00:52:14,160
S1: music is by zombie with the why and to get

929
00:52:14,160 --> 00:52:16,259
S1: the text and links from this episode, sign up for

930
00:52:16,260 --> 00:52:21,910
S1: the newsletter version of the show at Daniel meisler.com/newsletter. We'll

931
00:52:21,910 --> 00:52:22,719
S1: see you next time.