1
00:00:00,000 --> 00:00:02,397
Something about AI that causes
the people in the field I would say to

2
00:00:02,480 --> 00:00:05,837
become both excessively utopian
and excessively apocalyptic. Having said

3
00:00:05,920 --> 00:00:08,997
that, I think what's actually happened
is an enormous amount of technical

4
00:00:09,080 --> 00:00:11,557
progress that built up over time.
And like for for example, we now know the

5
00:00:11,640 --> 00:00:13,997
neural network is the correct
architecture. And I I will tell you like

6
00:00:14,080 --> 00:00:16,757
there was a 60-year run where that was
like a you know, or even 70 years where

7
00:00:16,840 --> 00:00:19,957
that was controversial. And so so
the way I think about what's happening is

8
00:00:20,040 --> 00:00:22,477
basically I think I think about
basically the the period we're in right

9
00:00:22,560 --> 00:00:27,317
now is it's I call it 80-year
overnight success. Right, which is like

10
00:00:27,400 --> 00:00:30,397
it's an overnight success cuz it's like
bam, you know, chat GPT hits and then

11
00:00:30,480 --> 00:00:33,197
and then 01 hits and then you know, open
claw hits and like you know, these are

12
00:00:33,280 --> 00:00:36,597
open these are these are like overnight
like radical overnight transformative

13
00:00:36,680 --> 00:00:40,917
successes, but they're drawing on an
80-year sort of wellspring backlog, you

14
00:00:41,000 --> 00:00:43,197
know, of of of of ideas and thinking.

15
00:00:43,280 --> 00:00:46,157
It's not just that it's all brand new,
it's that it's an unlock of all of these

16
00:00:46,240 --> 00:00:48,957
decades of like very serious hardcore
[music] research. If I were 18, like

17
00:00:49,040 --> 00:00:51,277
this is a hundred this is what I would
be spending all of my time on. This is

18
00:00:51,360 --> 00:00:54,237
like such an incredible
conceptual breakthrough.

19
00:00:54,320 --> 00:00:57,917
Before we get into today's episode,
I just have a small message for listeners.

20
00:00:58,000 --> 00:01:01,837
Thank you. We would not be able to bring
you the AI engineering, science, and

21
00:01:01,920 --> 00:01:04,837
entertainment content that you so
clearly want if you didn't choose to

22
00:01:04,920 --> 00:01:06,997
also click in and tune into our content.

23
00:01:07,080 --> 00:01:10,797
We've been approached by sponsors
on an almost daily basis, but fortunately

24
00:01:10,880 --> 00:01:15,277
enough of you actually subscribe to us
to keep all this sustainable without ads

25
00:01:15,360 --> 00:01:19,797
and we want to keep it that way. But I
just have one favor to ask all of you.

26
00:01:19,880 --> 00:01:23,557
The single most powerful, completely
free thing you can do is to click that

27
00:01:23,640 --> 00:01:27,357
subscribe button. It's the only thing
I'll ever ask of you and it means

28
00:01:27,440 --> 00:01:31,717
absolutely everything to me and my team
that works so hard to bring Latent Space

29
00:01:31,800 --> 00:01:35,837
to you each and every week. If you
do it, I promise you we'll never stop

30
00:01:35,920 --> 00:01:38,477
working to make this show even better.

31
00:01:38,560 --> 00:01:41,440
Now let's get into it.

32
00:01:44,720 --> 00:01:47,397
Everyone, welcome to the Latent Space
podcast. This [music] is Alessio,

33
00:01:47,480 --> 00:01:50,677
founder of Kernel Labs, and I'm joined
by Swyx, editor of Latent Space. Hello,

34
00:01:50,760 --> 00:01:56,157
and we're in a16z with A uh Martin
Casado, welcome. Yes. Yes, A and what,

35
00:01:56,240 --> 00:02:01,357
half of 16? >> [laughter] >> A1. Exactly.
Uh apparently this is the the final few

36
00:02:01,440 --> 00:02:04,677
days in your your current office. You're
moving across the road. Uh we're yeah,

37
00:02:04,760 --> 00:02:07,037
we have a we have we have some we have
some projects underway, but yeah, this

38
00:02:07,120 --> 00:02:09,837
is actually this is the original. We're
in actually the original office. We're

39
00:02:09,920 --> 00:02:12,117
in the we're in the we're in the
we're in the whole thing. Yeah.

40
00:02:12,200 --> 00:02:14,477
>> Great. Thank you.
So I have to come out.

41
00:02:14,560 --> 00:02:18,317
This is a you know, I just wanted
to pick a spicy start. In October 2022, I

42
00:02:18,400 --> 00:02:21,157
just made friends with Rune and

43
00:02:21,240 --> 00:02:26,037
I wanted to give him something to sort
of be spicy about and I said uh you'll

44
00:02:26,120 --> 00:02:29,197
never not be funny that A16Z was
constantly going the future is where the

45
00:02:29,280 --> 00:02:32,557
smart people choose to spend their time
and then going deep into crypto and not

46
00:02:32,640 --> 00:02:37,157
in AI and that was in October 2022
and Rune says there was an internal

47
00:02:37,240 --> 00:02:41,357
meeting in A16Z to reorient around Gen AI.
Obviously you have, but was there a

48
00:02:41,440 --> 00:02:45,117
meeting? What what was that?
I mean, I don't look, I've been doing AI

49
00:02:45,200 --> 00:02:48,197
since the late '80s. So I don't know
like all that. As far as I'm concerned,

50
00:02:48,280 --> 00:02:49,957
this stuff is all Johnny-come-lately.

51
00:02:50,040 --> 00:02:52,277
Yeah, yeah, I mean look, we've been
doing AI our entire existence. I mean,

52
00:02:52,360 --> 00:02:54,437
we've been doing AI machine learning
deep you know, deep learning. We've been

53
00:02:54,520 --> 00:02:57,797
doing this stuff way from
the beginning. But obviously,

54
00:02:57,880 --> 00:03:01,357
AI is just core to computer science.
I I I actually view them as like quite

55
00:03:01,440 --> 00:03:04,197
>> [snorts] >> uh quite continuous. Um you
know, Ben and I both have computer science

56
00:03:04,280 --> 00:03:07,477
degrees. Um you know, we we both Ben
Ben and I actually both are old enough to

57
00:03:07,560 --> 00:03:09,997
remember the actual AI boom in the 1980s.

58
00:03:10,080 --> 00:03:12,837
There was like a there was a big AI boom
at the time. Um and there was a there

59
00:03:12,920 --> 00:03:16,757
was one under names like expert systems
um and with the era of like Lisp and

60
00:03:16,840 --> 00:03:21,077
Lisp machines. Um I I coded in Lisp.
I was coding in Lisp in 1989. Where that

61
00:03:21,160 --> 00:03:23,237
was the the language of the AI future.

62
00:03:23,320 --> 00:03:25,757
Um yeah, so this is something that we're
like completely you're completely

63
00:03:25,840 --> 00:03:28,157
comfortable with. We've been doing
the whole time and are very enthusiastic

64
00:03:28,240 --> 00:03:32,357
about. Is there a strong like this time
is different because uh my closest

65
00:03:32,440 --> 00:03:36,197
analog was 2016-17.
There was there was an AI boom and it

66
00:03:36,280 --> 00:03:39,917
petered out very very quickly. Uh >> Well,
it just it it just didn't sort of invest.

67
00:03:40,000 --> 00:03:42,797
>> Sort of. Sort of. Investment investment
excitement. Although that's really when

68
00:03:42,880 --> 00:03:46,597
the the Nvidia phenomenon really
it was it I would say it was in that

69
00:03:46,680 --> 00:03:49,797
period when it was very clear that it
at the time it the vocabulary was more

70
00:03:49,880 --> 00:03:52,437
machine learning, but it it was very
clear at that time that machine learning

71
00:03:52,520 --> 00:03:53,997
was hitting some sort of takeoff point.

72
00:03:54,080 --> 00:03:56,237
Yeah. Well, and as you guys you guys
have talked about this at length on your

73
00:03:56,320 --> 00:03:59,077
on your thing, but you know, if you
really track what happened, I think the

74
00:03:59,160 --> 00:04:02,757
real story is it was it was the AlexNet
uh basically breakthrough in like 2013

75
00:04:02,840 --> 00:04:06,437
that was the that was the real knee
in the curve. Um and then it was obviously

76
00:04:06,520 --> 00:04:07,997
the transformer breakthrough in '17.

77
00:04:08,080 --> 00:04:10,197
Yeah. Um and then everything that
followed. But but you know, look,

78
00:04:10,280 --> 00:04:12,557
machine learning, you know, there were
you know, look, uh I mean, look, I've

79
00:04:12,640 --> 00:04:15,597
been working, you know, I've been working
with one of my you know, kind of projects

80
00:04:15,680 --> 00:04:19,797
working with Facebook since 2004. Um and
on the board since 2007. And of course,

81
00:04:19,880 --> 00:04:22,357
you know, you know, they they started
using machine learning very early. Um

82
00:04:22,440 --> 00:04:24,957
and you know, I've used it basically,
you know, for like 20 years for, you

83
00:04:25,040 --> 00:04:27,397
know, content, you know,
feed optimization and advertising

84
00:04:27,480 --> 00:04:30,957
optimization and obviously many,
you know, [snorts] financial services, you

85
00:04:31,040 --> 00:04:33,557
know, many many many companies, many
different sectors have been doing this.

86
00:04:33,640 --> 00:04:35,237
And so, it's like one of these things.

87
00:04:35,320 --> 00:04:37,677
It's like it's not a sing it's not
a single thing. Like it's it's like it's

88
00:04:37,760 --> 00:04:41,037
like layers, right? Um and and the
layers arrive at different paces. And

89
00:04:41,120 --> 00:04:44,237
but they kind of build up. Yeah. Uh they
kind of build up over time. And then and

90
00:04:44,320 --> 00:04:47,157
then yeah, and then look, in retrospect,
it was 2017 was kind of the, you know,

91
00:04:47,240 --> 00:04:49,957
the key the key point with the
transformer. And then and then as you

92
00:04:50,040 --> 00:04:52,717
guys know, there was this really weird
like four-year period where it's like

93
00:04:52,800 --> 00:04:55,517
the the transformer existed
and then it was just like >> [snorts]

94
00:04:55,600 --> 00:04:58,877
>> It's cool. Yeah. Well, [laughter] but
but it was just but but between 2020 but

95
00:04:58,960 --> 00:05:02,397
between 2017 and 2021, I mean, that
was the era of which like companies like

96
00:05:02,480 --> 00:05:05,277
Google had internal chatbots, but they
weren't letting anybody use them. Yeah.

97
00:05:05,360 --> 00:05:08,917
Right. And then, you know, and then
OpenAI developed ChatGPT or GPT-2, and

98
00:05:09,000 --> 00:05:11,677
then they told everybody this is way
too dangerous to deploy, right? We're you

99
00:05:11,760 --> 00:05:14,677
know, we can't possibly let normal
people normal people use this thing. And

100
00:05:14,760 --> 00:05:18,317
then do you guys I'm sure remember AI
Dungeon? Um It's a So, the old for there

101
00:05:18,400 --> 00:05:21,997
was like a year where like the only way
for a normal person to use GPT-3 was in

102
00:05:22,080 --> 00:05:24,997
in AI Dungeon. Yeah. And so, you you
we would do this. You'd go in there and

103
00:05:25,080 --> 00:05:27,157
you'd pretend to play Dungeons and
Dragons. In reality, you're just trying

104
00:05:27,240 --> 00:05:31,037
to talk to talk to GPT. And so, there
was this, you know, there was this long,

105
00:05:31,120 --> 00:05:33,557
you know, and the I you know, the big
big companies you know, big companies

106
00:05:33,640 --> 00:05:36,277
are cautious. And you know, the big
companies were cautious. It It By the

107
00:05:36,360 --> 00:05:39,077
way, it took OpenAI you know, they they
they talk about this. It took OpenAI

108
00:05:39,160 --> 00:05:42,237
time to actually adjust, you know,
kind of re- redirect their research path.

109
00:05:42,320 --> 00:05:46,317
>> I think it was at Rosewood, right?
Uh that the dinner they founded OpenAI was

110
00:05:46,400 --> 00:05:49,437
right there. Right. But that that
dinner would have taken place in 20

111
00:05:49,520 --> 00:05:52,637
>> 18 18 The formation
of OpenAI is late as

112
00:05:52,720 --> 00:05:56,717
2018? I I sorry no I'm I'm I'm I'm wrong.
Probably should be 20 They just

113
00:05:56,800 --> 00:06:00,437
celebrated their 10-year anniversary
so it it is 2025. Yeah. That's so 2015.

114
00:06:00,520 --> 00:06:05,557
Yeah, 2015 yeah 2015.
But then Alec Radford did GPT-1 in

115
00:06:05,640 --> 00:06:08,757
>> What? Probably 17 18 yeah.
>> 17 18 so it yeah for the and then and

116
00:06:08,840 --> 00:06:10,517
then they didn't really and then GPT-3

117
00:06:10,600 --> 00:06:16,197
was what? 20 20 20 2020 Because that
became co-pilot immediately yeah. Even

118
00:06:16,280 --> 00:06:18,677
OpenAI which has been you know the leader
of the of this thing in the last

119
00:06:18,760 --> 00:06:21,797
decade you know even they had to adapt
and and and lean into the new thing and

120
00:06:21,880 --> 00:06:26,557
so yeah I I think it's just this process
of basically sort of wave after wave

121
00:06:26,640 --> 00:06:29,157
layer after layer you know building
on itself and then you kind of get these

122
00:06:29,240 --> 00:06:32,357
catalytic moments where where the whole
thing pops. And obviously that's what's

123
00:06:32,440 --> 00:06:35,437
happening now. Is it useful
to think about will there

124
00:06:35,520 --> 00:06:38,277
be an AI winter because there's always
these patterns? Like is it endless

125
00:06:38,360 --> 00:06:42,637
summer? It's something I constantly
think about because do I get do I just

126
00:06:42,720 --> 00:06:46,717
like just get endlessly hyped and just
trust that I will only be early and

127
00:06:46,800 --> 00:06:49,357
never wrong? Or >> [laughter]

128
00:06:49,440 --> 00:06:52,237
>> Or are we will there be a winter?

129
00:06:52,320 --> 00:06:55,317
So there's something about say
the following there's something about AI

130
00:06:55,400 --> 00:06:57,397
that has led to this repeated pattern.

131
00:06:57,480 --> 00:07:00,557
And and and you you guys know this It's
summer winter summer winter summer

132
00:07:00,640 --> 00:07:03,237
winter summer winter
and it goes back 80 years.

133
00:07:03,320 --> 00:07:07,677
80 years. So the original neural network
paper was 1943 right which is which is

134
00:07:07,760 --> 00:07:10,677
amazing that it was it was far back
that long. And then there was you have you

135
00:07:10,760 --> 00:07:13,797
guys ever talked about this on your show
but there was this there was a big there

136
00:07:13,880 --> 00:07:15,277
was an AGI conference at Dartmouth

137
00:07:15,360 --> 00:07:18,877
University in 1955. 55 Yeah. And they

138
00:07:18,960 --> 00:07:22,317
got an NSF grant to for the all the AI
experts at the time to spend the summer

139
00:07:22,400 --> 00:07:25,317
together and they figured if they had
10 weeks together they could get AGI

140
00:07:25,400 --> 00:07:27,037
>> [laughter] >> out the other end.
And they got there by

141
00:07:27,120 --> 00:07:30,197
the way they got the grant they got the 10
weeks and then you know 1955 you know

142
00:07:30,280 --> 00:07:33,197
no no AGI. And like I said
I lived through the 80s

143
00:07:33,280 --> 00:07:35,797
version of this where there was a big
a big boom and a crash. And so so there is

144
00:07:35,880 --> 00:07:38,317
this thing there there is something
about AI that causes the people in the

145
00:07:38,400 --> 00:07:41,517
field I would say to become both
excessively utopian and excessively

146
00:07:41,600 --> 00:07:44,237
apocalyptic Um and it's
we're probably on both sides

147
00:07:44,320 --> 00:07:47,397
of like the the the boom bust cycle
you you kind of see that play out.

148
00:07:47,480 --> 00:07:50,717
Having said that, I think what's
actually happened is like just at in you

149
00:07:50,800 --> 00:07:53,757
know in we now know in retrospect like
an enormous amount of technical progress

150
00:07:53,840 --> 00:07:56,157
that built up over time. And like
for for example, we now know that neural

151
00:07:56,240 --> 00:07:59,117
network is the correct architecture.
And I I will tell you like there was a

152
00:07:59,200 --> 00:08:01,837
60-year run where that was like a you
know or even 70 years where that was

153
00:08:01,920 --> 00:08:04,797
controversial. And and we now know
that that's the case. And so we we now you

154
00:08:04,880 --> 00:08:07,037
know everything we're building on today
is sort of derived from the original

155
00:08:07,120 --> 00:08:11,157
idea 1943. And so so in retrospect we
now know that like these these guys were

156
00:08:11,240 --> 00:08:13,317
right they you know they would get
the timing wrong and they thought you know

157
00:08:13,400 --> 00:08:16,037
capabilities would arrive faster or they
were it could be turned into businesses

158
00:08:16,120 --> 00:08:18,877
sooner or whatever. But like they were
fundamentally the scientists who worked

159
00:08:18,960 --> 00:08:21,437
on this over the course of decades were
fundamentally correct about what they

160
00:08:21,520 --> 00:08:23,877
were doing. >> [snorts] >>
And and and the payoff from from from

161
00:08:23,960 --> 00:08:25,597
all their work is happening now.

162
00:08:25,680 --> 00:08:28,557
And so so the way I think about what's
happening is basically I think I think

163
00:08:28,640 --> 00:08:31,837
about basically the the the period we're
in right now is it's I call it 80-year

164
00:08:31,920 --> 00:08:36,397
overnight success. Right. Which is like
it's an overnight success cuz it's like

165
00:08:36,480 --> 00:08:40,117
bam you know chat GPT hits and then
and then 01 hits and then you know Claude

166
00:08:40,200 --> 00:08:42,597
hits and like you know these are open
these are these are like overnight like

167
00:08:42,680 --> 00:08:46,077
radical overnight transformative
successes but they're drawing on an

168
00:08:46,160 --> 00:08:51,037
80-year sort of wellspring backlog you
know of of of of ideas and thinking.

169
00:08:51,120 --> 00:08:53,997
It's not just that it's all brand new
it's that it's an unlock of all of these

170
00:08:54,080 --> 00:08:58,037
decades of like very serious hardcore
research um and thinking. Look, there

171
00:08:58,120 --> 00:09:01,157
were AI researchers who
spent their entire lives.

172
00:09:01,240 --> 00:09:03,837
They got their PhD. >> [laughter] >>
They worked for they researched for 40

173
00:09:03,920 --> 00:09:06,597
years. They retired. In a lot of cases
they passed away and they never actually

174
00:09:06,680 --> 00:09:10,957
saw it work. Yeah. So sad. It is it is
sad. It is sad. Jeff Hinton was like the

175
00:09:11,040 --> 00:09:14,117
last guy. Yeah [laughter] yeah well
there were guys out there Alan Newell I

176
00:09:14,200 --> 00:09:16,677
mean there was tons of John McCarthy
John McCarthy was like one of the

177
00:09:16,760 --> 00:09:18,637
inventors of the field. He's one
of the guys who organized the Dartmouth

178
00:09:18,720 --> 00:09:21,117
conference and you know he taught
at Stanford for 40 years and passed you

179
00:09:21,200 --> 00:09:24,117
know passed away I don't know
whatever 10 10 years ago or something.

180
00:09:24,200 --> 00:09:27,157
Never never actually got got to see
it happen. But like it is amazing in

181
00:09:27,240 --> 00:09:29,637
retrospect like these guys were
incredibly smart and they worked really

182
00:09:29,720 --> 00:09:33,597
hard and they were correct. So anyway so
then it's like okay you know as they say

183
00:09:33,680 --> 00:09:35,397
history doesn't repeat but it rhymes.

184
00:09:35,480 --> 00:09:37,837
It's like okay does that mean that
there's going to be another like you

185
00:09:37,920 --> 00:09:41,037
know basically boom bust cycle
and I I will tell you like look like

186
00:09:41,120 --> 00:09:43,517
in a sense I guess everything goes
through cycles and you know people get

187
00:09:43,600 --> 00:09:46,477
overly enthusiastic and overly depressed
and there's there's a time there's a

188
00:09:46,560 --> 00:09:50,597
timelessness to that.
Having said that there's just no question.

189
00:09:50,680 --> 00:09:53,317
So The four most the four
most dangerous words in the

190
00:09:53,400 --> 00:09:56,597
>> is different. This time is different.
Do you know the 12 most dangerous words in

191
00:09:56,680 --> 00:09:58,997
investing? >> No.
The four most four most dangerous

192
00:09:59,080 --> 00:10:01,317
words in investing [laughter]
are this time is different.

193
00:10:01,400 --> 00:10:05,757
So the 12 most dangerous words. And so
like I'll tell you what's different like

194
00:10:05,840 --> 00:10:09,437
Now it's working. Like like there's just
no I mean look there's just no question.

195
00:10:09,520 --> 00:10:12,917
And by the way I'll just give you
guys my take like LLMs like from from

196
00:10:13,000 --> 00:10:15,917
basically the chat GPT moment through to

197
00:10:16,000 --> 00:10:20,277
spring of 25 I think you could still
I think well intentioned well informed

198
00:10:20,360 --> 00:10:24,637
skeptics could still say oh this is just
pattern completion and oh these things

199
00:10:24,720 --> 00:10:27,357
don't really understand what they're
doing and you know the hallucination

200
00:10:27,440 --> 00:10:30,397
rates are way too high and you know
this is going to be great for creative

201
00:10:30,480 --> 00:10:33,757
writing and creating you know
Shakespearean sonnets and you know as as

202
00:10:33,840 --> 00:10:36,637
rap lyrics or whatever like it's going
to be great at all that stuff but we're

203
00:10:36,720 --> 00:10:39,597
not going to be able to harness this to
make this relevant in you know coding or

204
00:10:39,680 --> 00:10:42,517
in medicine or in law or in you know
kind of fields that you know kind of

205
00:10:42,600 --> 00:10:45,517
really really matter.
And I think basically it was the reasoning

206
00:10:45,600 --> 00:10:48,757
breakthrough. It was 01 and then R1
that basically answered that question

207
00:10:48,840 --> 00:10:51,197
basically said oh no we're going
to be able to actually turn this into

208
00:10:51,280 --> 00:10:53,277
something that's going to work
in the real world and then and then it

209
00:10:53,360 --> 00:10:55,557
obviously the coding breakthrough over
the basically the coding breakthrough

210
00:10:55,640 --> 00:10:57,837
that kind of catalyzed over the holiday
break was kind of the third step in

211
00:10:57,920 --> 00:11:00,757
that. Where it's like [snorts] all
right if if you know if Linus Torvalds is

212
00:11:00,840 --> 00:11:04,197
saying that AI coding is now
better than he is like >> [laughter]

213
00:11:04,280 --> 00:11:06,237
>> Like that that's that's
never happened before.

214
00:11:06,320 --> 00:11:08,877
>> That's the benchmark yeah. That's never
happened before and so now we know that

215
00:11:08,960 --> 00:11:11,917
it's it's going to sweep through coding
and and then and then we we know you

216
00:11:12,000 --> 00:11:14,117
know we know that if it's going
to work in coding it's going to work in

217
00:11:14,200 --> 00:11:17,077
everything else right? It's just that
that cuz that's that's like that's like

218
00:11:17,160 --> 00:11:19,357
that's like the hardest in many ways
that's the hardest example and now

219
00:11:19,440 --> 00:11:22,397
everything else is going to be
a derivative of that. And then on top of

220
00:11:22,480 --> 00:11:24,877
that we just got the agent breakthrough
but you know with open claw which is

221
00:11:24,960 --> 00:11:27,837
fantastic which is amazing
and incredibly powerful. And then we just

222
00:11:27,920 --> 00:11:31,157
got the the the auto research you know
know, the self improvement you know

223
00:11:31,240 --> 00:11:33,557
we're now into the self improvement
breakthrough. And so the so the way I

224
00:11:33,640 --> 00:11:36,317
think about it is we've had
four fundamental breakthroughs in

225
00:11:36,400 --> 00:11:39,557
functionality. LLMs reasoning, agents um

226
00:11:39,640 --> 00:11:44,277
and then and and then now RSI. Um and
they're all actually working. Um and so

227
00:11:44,360 --> 00:11:47,357
I'm I'm just as you can I'm
jumping out of my shoes. [laughter]

228
00:11:47,440 --> 00:11:50,677
Like like this is like this is it.
Like this is the culmination of 80 years

229
00:11:50,760 --> 00:11:53,637
worth of worth of work and this is the
time it's becoming real. Yeah. Yeah, I

230
00:11:53,720 --> 00:11:57,037
I'm completely convinced. I think the
anxiety that people feel is like during

231
00:11:57,120 --> 00:12:00,557
the transistor era you had Moore's law
and it's like, all right, we understand

232
00:12:00,640 --> 00:12:03,717
why these things are getting better.
We understand the physics of it. With AI

233
00:12:03,800 --> 00:12:08,117
it's it's so jagged and like the jumps
where like like you said it's like in 3

234
00:12:08,200 --> 00:12:11,917
months you have like this huge jump like
and people are like, well, this can't

235
00:12:12,000 --> 00:12:13,837
keep happening, right?
But then it keeps happening.

236
00:12:13,920 --> 00:12:17,837
>> It'll keep happening. And so like how
do you think about also timelines of like

237
00:12:17,920 --> 00:12:20,877
what's worth building? I think we always
have this question with guests which is

238
00:12:20,960 --> 00:12:24,437
like, you know, should you spend time
building harness for a model versus like

239
00:12:24,520 --> 00:12:27,837
the next model just going to do
it one shot in the latest space.

240
00:12:27,920 --> 00:12:31,117
And how does that inform like how you
think about the shape of the technology?

241
00:12:31,200 --> 00:12:34,237
You know, you talk about how it's
a new computing platform. If you have a

242
00:12:34,320 --> 00:12:37,477
computing platform then like every 6
months it like drastically changes in

243
00:12:37,560 --> 00:12:40,717
what it looks like. It's hard to build
companies on top of it. Yeah, so so a

244
00:12:40,800 --> 00:12:43,757
couple things. So one is like, look,
the the Moore's law was what we now call a

245
00:12:43,840 --> 00:12:47,277
scaling law. Like Moore's law was a
scaling law and for your younger viewers

246
00:12:47,360 --> 00:12:50,277
Moore's law was every chip chip chips
either get twice as powerful or twice as

247
00:12:50,360 --> 00:12:53,117
cheap every every 18 months.
And that and that and that you know,

248
00:12:53,200 --> 00:12:55,597
that it's gotten more complicated in the
last few years, but like that that was

249
00:12:55,680 --> 00:12:59,477
like the 50-year trajectory of of of the
computer industry. And then and then by

250
00:12:59,560 --> 00:13:02,957
the way, that's what took the mainframe
computer from a $25 current dollar thing

251
00:13:03,040 --> 00:13:05,317
into you know, the phone in your pocket
being you know, a million times more

252
00:13:05,400 --> 00:13:08,037
powerful than that. Like that,
you know, for for 500 bucks.

253
00:13:08,120 --> 00:13:10,797
And so that that was a scaling law.
And then and then and then key to any

254
00:13:10,880 --> 00:13:13,997
scaling law including Moore's law
and the AI scaling laws is, you know,

255
00:13:14,080 --> 00:13:17,317
they're not really laws, right? They're
they're they're they're predictions. But

256
00:13:17,400 --> 00:13:20,157
when they work they become self-fulfilling
predictions because they they

257
00:13:20,240 --> 00:13:23,077
they set a benchmark and and then
the entire industry, right? All the smart

258
00:13:23,160 --> 00:13:26,157
people in the industry kind of work to
make sure that that that actually happens.

259
00:13:26,240 --> 00:13:28,677
And so they they kind of motivate
the breakthroughs that are required to to

260
00:13:28,760 --> 00:13:31,637
keep that going. And in and in and
in chips that was a 50-year that was a

261
00:13:31,720 --> 00:13:34,477
50-year run, right? And it it
was amazing and it's still happening in in

262
00:13:34,560 --> 00:13:37,797
some areas of of chips. I think the same
thing is happening with the the core

263
00:13:37,880 --> 00:13:41,557
scaling laws the core scaling laws
in in in AI. You know, they're they're not

264
00:13:41,640 --> 00:13:44,717
really laws, but like they they are
basically they're predictions and then

265
00:13:44,800 --> 00:13:48,157
they're motivating catalysts for the
research work that is required to be and

266
00:13:48,240 --> 00:13:50,757
and and by the way also
the investment dollars

267
00:13:50,840 --> 00:13:53,517
you know, required to basically
keep you know, keep the curves going.

268
00:13:53,600 --> 00:13:55,677
And and look it's going to be
complicated and it's going to be

269
00:13:55,760 --> 00:13:57,917
variable and there you know, there going
to be walls that are going to look like

270
00:13:58,000 --> 00:13:59,837
they're fast approaching and then
they're going to be you know, engineers

271
00:13:59,920 --> 00:14:01,437
are going to get to work and they're
going to figure out a way to punch

272
00:14:01,520 --> 00:14:04,837
through the walls. And obviously that's
you know, that's been happening a lot.

273
00:14:04,920 --> 00:14:06,917
You know, and then look there's going
to be times when it looks like the walls

274
00:14:07,000 --> 00:14:09,197
have you know, the the the laws have
petered out and then they're going to

275
00:14:09,280 --> 00:14:12,277
they're going to pick up again and surge
and then and then and then it it appears

276
00:14:12,360 --> 00:14:15,677
what's happening to the eyes there's now
multiple you know, multiple scaling laws.

277
00:14:15,760 --> 00:14:18,597
Um there's multiple areas of improvement
and and I think you know, I don't know

278
00:14:18,680 --> 00:14:20,637
how many more there already have to be
discovered, but there are probably some

279
00:14:20,720 --> 00:14:23,437
more that we don't know about yet. You
know, like for example, there's probably

280
00:14:23,520 --> 00:14:27,437
some scaling law around um world
models and that we don't fully kind of

281
00:14:27,520 --> 00:14:30,277
acquisition of data at scale in the real
world that we don't fully understand

282
00:14:30,360 --> 00:14:32,997
yet. So that that that one will
probably kick in at some point here.

283
00:14:33,080 --> 00:14:35,717
There's a bunch of really
smart people working on that. Um

284
00:14:35,800 --> 00:14:38,757
And so yeah, I I think the expectation
is that you know, the the scaling laws

285
00:14:38,840 --> 00:14:41,437
generally are going to continue.
Yeah, the the pace of improvement will

286
00:14:41,520 --> 00:14:45,437
continue to move really fast.
Um To your question on like what to build.

287
00:14:45,520 --> 00:14:48,197
So I'm a complete believer the scaling
laws are going to continue. I'm a

288
00:14:48,280 --> 00:14:51,077
complete believer the capabilities
are going to keep getting amazing

289
00:14:51,160 --> 00:14:53,717
you know, leaps and bounds.
The part where I kind of part ways a

290
00:14:53,800 --> 00:14:56,397
little bit with I would what I
would describe as the AI purist

291
00:14:56,480 --> 00:14:59,517
you know, which is which I would
characterize as like the people who are

292
00:14:59,600 --> 00:15:01,757
many ways the smartest people in the
field, but also the people who spend

293
00:15:01,840 --> 00:15:03,797
their entire life like in a lab.

294
00:15:03,880 --> 00:15:06,237
Um [snorts] and have to have I would
say have very little experience in the

295
00:15:06,320 --> 00:15:09,357
outside world. Um the nuance
I would offer is the outside

296
00:15:09,440 --> 00:15:13,077
world of 8 billion people
and institutions and governments and

297
00:15:13,160 --> 00:15:15,317
companies and economic systems and

298
00:15:15,400 --> 00:15:17,957
social systems is really complicated. Um

299
00:15:18,040 --> 00:15:22,597
and and doesn't you know, it it
8 billion people making collective

300
00:15:22,680 --> 00:15:27,077
decisions on planet Earth is not
a simple process of like just like

301
00:15:27,160 --> 00:15:29,357
You see this happening now it's like
a bunch of the AI CEOs have this thing

302
00:15:29,440 --> 00:15:31,597
which is just like they're well there's
just this they just all have this kind

303
00:15:31,680 --> 00:15:33,877
of thing when they talk in public where
they're just like well there's this this

304
00:15:33,960 --> 00:15:36,637
obvious set of things that society needs
to do. Mhm. [snorts] And then they're

305
00:15:36,720 --> 00:15:38,997
like society's not doing
any of those things.

306
00:15:39,080 --> 00:15:41,797
Right? And it's like how can society not
you know what whatever their theory is

307
00:15:41,880 --> 00:15:45,197
how can society not see XYZ? And the
answer is well society is number one

308
00:15:45,280 --> 00:15:48,277
there's no single society it's like 8
billion people. And they like all have a

309
00:15:48,360 --> 00:15:50,477
voice and they all have a vote like
at the end of the day of how they they

310
00:15:50,560 --> 00:15:53,997
react to change and then you know it
just like it's just human reality is

311
00:15:54,080 --> 00:15:55,917
just really complicated and messy.

312
00:15:56,000 --> 00:16:00,557
Um and and and so the specific answer to
your question is like as usual it depends.

313
00:16:00,640 --> 00:16:03,997
Um you know it it depends. Look there's
no question people are going to like

314
00:16:04,080 --> 00:16:06,277
there's no question there're going to be
companies it's already happening. There

315
00:16:06,360 --> 00:16:08,397
are companies that think that they're
building value on top of the models and

316
00:16:08,480 --> 00:16:10,317
then they're just going to get blitzed
by the by the next model. There's no

317
00:16:10,400 --> 00:16:12,917
question that's happening. But I think
there's no question also that just the

318
00:16:13,000 --> 00:16:17,357
process of adaptation of any technology
into the real in into the real messy

319
00:16:17,440 --> 00:16:21,237
world of humanity is is just going to be
messy and complicated. It's it's not

320
00:16:21,320 --> 00:16:23,677
going to be simple and straightforward
it's going to be messy and complicated.

321
00:16:23,760 --> 00:16:27,197
And there are going to be a lot
of companies and a lot of products um and

322
00:16:27,280 --> 00:16:30,477
in in in fact entire industries that are
going to get built that to basically

323
00:16:30,560 --> 00:16:33,567
actually help all of this technology
actually reach real people.

324
00:16:33,650 --> 00:16:35,037
>> [snorts] >> The amount
of capital going into these

325
00:16:35,120 --> 00:16:38,717
companies I mean Dario talked about it
on the Door Cash podcast and Door Cash

326
00:16:38,800 --> 00:16:41,477
was like why don't you just buy 10x more
GPUs and he's like because I'm going to

327
00:16:41,560 --> 00:16:46,357
go bankrupt if the model doesn't exactly
hit the the performance level. How do

328
00:16:46,440 --> 00:16:49,477
you think about that also as a risk
on you know you guys are investors at

329
00:16:49,560 --> 00:16:51,957
Inflection and Inflection AI
and Thinking Machines and World Apps it

330
00:16:52,040 --> 00:16:55,157
seems like we're
leveraging this gaming loss

331
00:16:55,240 --> 00:16:59,197
at a pretty high rate. Like how
comfortable I guess do you feel with the

332
00:16:59,280 --> 00:17:01,237
downside scenario? Like and say like

333
00:17:01,320 --> 00:17:06,437
things peter out you think you can kind
of like restructure these build outs and

334
00:17:06,520 --> 00:17:08,997
you know capital investments? Yeah
so I should start by saying so I lived

335
00:17:09,080 --> 00:17:12,157
through the dot com crash and I
can tell you stories for hours

336
00:17:12,240 --> 00:17:15,917
about the dot com crash and it was
horrible. No it was awful. It was it was

337
00:17:16,000 --> 00:17:19,557
it was apocalyptic. By the way the a lot
of the dot-com crash was actually at the

338
00:17:19,640 --> 00:17:22,397
time it was actually a telecom crash.
It was a bandwidth crash. Like the the

339
00:17:22,480 --> 00:17:24,597
thing that actually crashed that wiped
out all the money with the tele- the

340
00:17:24,680 --> 00:17:26,677
telecom companies. Global Crossing.

341
00:17:26,760 --> 00:17:31,517
Global Global Yes. I'm from Singapore
and they laid so much cable over over

342
00:17:31,600 --> 00:17:34,997
our oceans. Actually, there was a
scaling law in the dot-com era and it

343
00:17:35,080 --> 00:17:38,557
was literally the the US Commerce
Department put out a report in 1996 and

344
00:17:38,640 --> 00:17:41,317
they said internet traffic
was doubling every quarter.

345
00:17:41,400 --> 00:17:44,677
Uh and and actually in 1995 and 1996
internet traffic actually did double

346
00:17:44,760 --> 00:17:46,997
every quarter. And so that became
the scaling law. And so what all these

347
00:17:47,080 --> 00:17:49,557
telecom entrepreneurs did was they went
out and they raised money to build fiber

348
00:17:49,640 --> 00:17:52,997
anticipating that the demand for bandwidth
was going to keep doubling every quarter.

349
00:17:53,080 --> 00:17:55,117
Doubling every quarter though is like,
you know, grains of chess on the

350
00:17:55,200 --> 00:17:58,557
chessboard. Like at some point
the numbers become extremely large, right?

351
00:17:58,640 --> 00:18:01,437
And and and it really and really what
happened was the internet the internet

352
00:18:01,520 --> 00:18:04,277
by the way continuously kept growing
basically since inception and it's, you

353
00:18:04,360 --> 00:18:07,517
know, it's it's continuously grown. It's
never shrunk. And it's grown really fast

354
00:18:07,600 --> 00:18:10,197
compared to anything else, you know,
in in in human history. But it wasn't

355
00:18:10,280 --> 00:18:12,517
doubling every quarter as of 1998-1999.

356
00:18:12,600 --> 00:18:15,077
And so there was this gap in the
expectation of what they thought was the

357
00:18:15,160 --> 00:18:18,477
scaling law versus reality. And that's
actually what caused the dot-com crash,

358
00:18:18,560 --> 00:18:21,797
which was they they wait over companies
like Global Crossing wait overbuilt

359
00:18:21,880 --> 00:18:26,197
fiber. Which is sort of the By the way,
fiber, telecom equipment, you know, so

360
00:18:26,280 --> 00:18:28,997
all that all the networking gear, you
know, and then and then by the way, the

361
00:18:29,080 --> 00:18:31,557
actual physical data centers. Like that
was the beginning of the of the of the

362
00:18:31,640 --> 00:18:34,997
data center build and then and then data
center overbuilt. And so you had the but

363
00:18:35,080 --> 00:18:38,717
it was it was literally I think it
was like $2 trillion got wiped out, right?

364
00:18:38,800 --> 00:18:42,077
It was like >> Jesus It was like
a big It was And by the way,

365
00:18:42,160 --> 00:18:46,277
the other the other subtlety in it was
the internet companies themselves never

366
00:18:46,360 --> 00:18:48,917
really had any debt cuz tech tech
companies generally don't run on debt,

367
00:18:49,000 --> 00:18:50,397
but the telecom companies run on debt.

368
00:18:50,480 --> 00:18:52,997
Physical infrastructure companies run
on debt. And so the companies like Global

369
00:18:53,080 --> 00:18:55,757
Crossing not just raised a lot of equity,
they also raised a lot of debt.

370
00:18:55,840 --> 00:18:57,957
So they're highly levered. And so then
you just do the thing of just like,

371
00:18:58,040 --> 00:19:00,277
okay, you have a highly levered thing
where you're you're just over over

372
00:19:00,360 --> 00:19:03,877
you're overbuilding capacity. Demand
is growing but not as fast as you hoped.

373
00:19:03,960 --> 00:19:06,837
And then boom, bankrupt, right? And then
and then it's and then it's like they

374
00:19:06,920 --> 00:19:09,557
say about the hotel industry, which
is it's always the third owner of a hotel

375
00:19:09,640 --> 00:19:12,557
that makes money. Right?
It has to go bankrupt twice.

376
00:19:12,640 --> 00:19:15,757
Right? You have to wash out all of the
over-optimistic exuberance before it

377
00:19:15,840 --> 00:19:17,967
gets to actually a stable
state and then it makes money.

378
00:19:18,050 --> 00:19:19,357
>> [gasps] >> So,
by the way, all of those data

379
00:19:19,440 --> 00:19:22,557
centers and all of those all the fiber
that they're in use it's all in use

380
00:19:22,640 --> 00:19:26,877
today. But 25 years later. >> [snorts] >>
But it took it actually the elapsed time

381
00:19:26,960 --> 00:19:30,597
was it took 15 years. It took 15 years
from 2000 to 2015 to actually fill the

382
00:19:30,680 --> 00:19:34,957
fill up all the capacity. The cautionary
warning is the the overbuild can happen.

383
00:19:35,040 --> 00:19:37,517
Um and and and and you know you you
get into this thing where basically

384
00:19:37,600 --> 00:19:40,397
everybody everybody who basically has
any sort of institutional capital is

385
00:19:40,480 --> 00:19:43,357
like wow it's just I I don't know how
to invest in these crazy software things.

386
00:19:43,440 --> 00:19:46,557
But for sure I can put build data
centers and for sure I can buy GPUs and

387
00:19:46,640 --> 00:19:50,597
I can deploy you know the compute
grids and and all these things.

388
00:19:50,680 --> 00:19:53,477
Um and so you know you if you're
a pessimist you could look at this and you

389
00:19:53,560 --> 00:19:56,877
could say wow this is like really set
up to be able to basically replicate you

390
00:19:56,960 --> 00:19:59,277
know what we went through we went
through in 2000. Obviously that would be

391
00:19:59,360 --> 00:20:02,437
bad. The counterargument which
is the one I I agree with which is the

392
00:20:02,520 --> 00:20:05,877
counterargument on the other side is a
couple things. One is the companies that

393
00:20:05,960 --> 00:20:08,437
are investing all the the companies
that are investing the money are like the

394
00:20:08,520 --> 00:20:12,077
bluest chip of companies. So that back
back in the in the like Global Crossing

395
00:20:12,160 --> 00:20:15,317
was like an it was like an entrepreneur
it's like a a new venture. But like the

396
00:20:15,400 --> 00:20:19,157
money that's being deployed now at scale
is Microsoft and you know and Amazon and

397
00:20:19,240 --> 00:20:23,917
Google right and Facebook and Nvidia
and you know these these these and and now

398
00:20:24,000 --> 00:20:25,997
you know by the way Open AI and Anthropic
which are now at like you know

399
00:20:26,080 --> 00:20:28,677
really serious size um you
know as companies with you know

400
00:20:28,760 --> 00:20:32,237
very serious revenue. These are very
large scale companies with like lots

401
00:20:32,320 --> 00:20:36,517
lots of cash lots of debt capacity that
they've they've never used. And so this

402
00:20:36,600 --> 00:20:39,877
is institutional in a way that that
really wasn't at the time. And then the

403
00:20:39,960 --> 00:20:43,397
other is at least for now every dollar
that's being put into anything that

404
00:20:43,480 --> 00:20:47,477
results in a running GPU is being turned
into revenue right away. Like so and you

405
00:20:47,560 --> 00:20:50,357
guys know this like everybody's starved
for capacity. Everybody's starved for

406
00:20:50,440 --> 00:20:53,317
compute capacity and then you know all
the associated things memory and and and

407
00:20:53,400 --> 00:20:57,157
interconnect and everything else.
Um data center space. And so every dollar

408
00:20:57,240 --> 00:20:59,997
right now that's being put in the ground
is turning into revenue. And and it and

409
00:21:00,080 --> 00:21:02,597
in fact I actually think there's
an interesting thing happening which is

410
00:21:02,680 --> 00:21:06,397
because everybody's starved for capacity
the models that we actually have that we

411
00:21:06,480 --> 00:21:09,317
can use today are inferior versions of
what we would have if not for the supply

412
00:21:09,400 --> 00:21:11,877
constraints. That's true.
If if right, suppose a

413
00:21:11,960 --> 00:21:14,837
hypothetical universe in which GPUs
were 10 times cheaper and 10 times more

414
00:21:14,920 --> 00:21:17,917
plentiful, the models would be much
better cuz you would just allocate a lot

415
00:21:18,000 --> 00:21:19,957
more money to training and you'd just
build better models and they would be

416
00:21:20,040 --> 00:21:23,477
better. And so we're we're actually
getting the sandbag version of the

417
00:21:23,560 --> 00:21:26,117
technology. >> Yeah,
no everything we use is quantized

418
00:21:26,200 --> 00:21:30,277
because the the labs have to keep
the the full versions. Right.

419
00:21:30,360 --> 00:21:32,037
>> [laughter] >> We're not
even getting the good stuff.

420
00:21:32,120 --> 00:21:36,477
Yeah. But but getting the good stuff
is is just even if technical progress

421
00:21:36,560 --> 00:21:40,157
stops, once there's like a much bigger
build of like GPU manufacturing capacity

422
00:21:40,240 --> 00:21:42,197
and memory, you [clears throat] know,
all all the things that have to happen

423
00:21:42,280 --> 00:21:45,077
in the course of the next five or 10
years. Once it happens, even the current

424
00:21:45,160 --> 00:21:47,957
technology is going to get going to get
much better. And then as you know, like

425
00:21:48,040 --> 00:21:50,597
there's just like a million ways to use
this stuff. Like there's just like a

426
00:21:50,680 --> 00:21:53,277
million use cases for this. Like it
it you know, this isn't just sending

427
00:21:53,360 --> 00:21:55,997
packets across a a thing, whatever, and
hoping that people find something to do

428
00:21:56,080 --> 00:21:58,717
with it. This is just like, oh, we apply
intelligence in every domain of human

429
00:21:58,800 --> 00:22:02,917
activity and then it works like
incredibly well. Yeah. Um here's what I

430
00:22:03,000 --> 00:22:06,877
know. Here's what I know. Um in the next
three or four years, it's like somewhere

431
00:22:06,960 --> 00:22:08,957
between three or four years out,
basically everything is selling out. So

432
00:22:09,040 --> 00:22:12,397
like the the entire supply chain is is
is is sold out or or selling out. And so

433
00:22:12,480 --> 00:22:15,757
there there's no like it I we're
just going to have like chronic

434
00:22:15,840 --> 00:22:19,077
supply shortage for you know, for years
to come. Um there's going to be a

435
00:22:19,160 --> 00:22:22,117
response from the market that's going
to result in an enormous, you know, it's

436
00:22:22,200 --> 00:22:25,277
happening now, an enormous flood of
investment in a new fab capacity and you

437
00:22:25,360 --> 00:22:28,277
know, every everything else to be able
to do that. At some point the supply

438
00:22:28,360 --> 00:22:31,397
chain constraints will unlock, you know,
at least to some degree. That will be

439
00:22:31,480 --> 00:22:34,437
another accelerant to industry growth
when that happens cuz the products will

440
00:22:34,520 --> 00:22:36,437
get better and everything
will get cheaper.

441
00:22:36,520 --> 00:22:38,717
Um and so so I know that's going
to happen. I know that, you know, the

442
00:22:38,800 --> 00:22:42,197
deployments, you know, the the actual
use cases are like really compelling and

443
00:22:42,280 --> 00:22:44,517
then like I said, you know, with
reasoning and agents and so forth, like

444
00:22:44,600 --> 00:22:47,837
I know they're just going to get like
much much better from here. And so I I I

445
00:22:47,920 --> 00:22:51,037
know the capabilities are like really
real and serious. I also know that the

446
00:22:51,120 --> 00:22:52,637
technical progress is not going to stop.

447
00:22:52,720 --> 00:22:56,237
It it it is it it is accelerating. Like
the breakthroughs are are tremendous. I

448
00:22:56,320 --> 00:22:59,037
mean, even just month over month,
the breakthroughs are really dramatic.

449
00:22:59,120 --> 00:23:02,837
And so, you You I think if you were
a cynic and there there are cynics, you

450
00:23:02,920 --> 00:23:05,877
can look at 2000, you can find echoes,
but I can't even imagine betting on that

451
00:23:05,960 --> 00:23:09,197
this is going to like somehow disappoint
and you know, at least for years to come.

452
00:23:09,280 --> 00:23:11,517
I think it would be essentially
suicidal to make that bet.

453
00:23:11,600 --> 00:23:15,087
Um is that Michael Burry?
Uh oh That's an interesting guy, huh?

454
00:23:15,170 --> 00:23:16,157
>> [laughter] >> We'll pick
on a guy we'll pick let's

455
00:23:16,240 --> 00:23:19,317
pick on one guy. We'll pick cuz he did
he came out with it was it was the He

456
00:23:19,400 --> 00:23:21,757
doesn't mind. It was the Nvidia short,
right? He came out with the Nvidia

457
00:23:21,840 --> 00:23:24,597
short. And then did if you guys probably
talked about this which is the the

458
00:23:24,680 --> 00:23:28,477
analysis now that the current models
are getting better faster at such a rate

459
00:23:28,560 --> 00:23:31,717
that if you are running an Nvidia If
you're running an Nvidia inference chip

460
00:23:31,800 --> 00:23:35,077
today that's 3 years old, you're making
more money on it today than you did 3

461
00:23:35,160 --> 00:23:38,437
years ago. Because the pace of improvement
of the software is is faster

462
00:23:38,520 --> 00:23:41,157
than the the depreciation cycle
of the chip. And then my understanding is

463
00:23:41,240 --> 00:23:44,197
Google is running I don't know if they I
don't know exactly what I this is rumors

464
00:23:44,280 --> 00:23:48,037
that I've heard or maybe it's public,
but I think Google's running very old TPUs

465
00:23:48,120 --> 00:23:51,797
Very profitably. And very profitably
and very profitably. Um and so so it

466
00:23:51,880 --> 00:23:54,277
actually turns out as far as I can tell
that it's actually the opposite of the

467
00:23:54,360 --> 00:23:58,037
Burry thesis is actually he was actually
180° wrong. It's actually the

468
00:23:58,120 --> 00:24:00,357
the old Nvidia chips are
getting more valuable.

469
00:24:00,440 --> 00:24:03,157
Which is something that's like literally
never happened [laughter] before. Like

470
00:24:03,240 --> 00:24:06,517
it's never been the case that you have
an older model chip that becomes more

471
00:24:06,600 --> 00:24:10,477
valuable not less valuable. And then and
again that's an expression of that just

472
00:24:10,560 --> 00:24:14,477
ferocious pace of software progress,
ferocious pace of capability payoff that

473
00:24:14,560 --> 00:24:17,637
you're getting on the other side of this.
And so I just the idea of betting

474
00:24:17,720 --> 00:24:21,897
against that like Yeah, yeah. It's like
an invitation to get your face ripped off.

475
00:24:21,980 --> 00:24:23,877
>> [laughter] >> What
my really hits was like modeling

476
00:24:23,960 --> 00:24:28,357
the lifespan of the H100 and H200s
and and going like you know, usually they

477
00:24:28,440 --> 00:24:31,837
advise like 4 to 7 years and it was you
know, maybe you sort of realistically

478
00:24:31,920 --> 00:24:36,957
cut it down to 2 to 3, but actually it's
going up and not down. And and that's I

479
00:24:37,040 --> 00:24:40,317
mean that's I think that's the dream.
We are finding utilization and I think

480
00:24:40,400 --> 00:24:44,117
utilization solves all problems. Like
you can you can find use use cases for

481
00:24:44,200 --> 00:24:47,797
even like the poor like even memory
we're having a shortage, right? And and

482
00:24:47,880 --> 00:24:51,277
even like the the shittier versions of
of memory that we do have we are finding

483
00:24:51,360 --> 00:24:53,277
use cases for it. So like that's great.

484
00:24:53,360 --> 00:24:57,717
Yeah. How how important is open source AI
and kind of like edge inference in a world

485
00:24:57,800 --> 00:25:01,037
in which you have three years of supply
crunch. Like you think in the like you

486
00:25:01,120 --> 00:25:03,677
know, if you fast forward like five
years, like how do you think about

487
00:25:03,760 --> 00:25:06,957
inference in the data
center versus at the edge?

488
00:25:07,040 --> 00:25:09,077
Well, so just to start, yeah. So,
I think I think open source is very

489
00:25:09,160 --> 00:25:11,277
important for a bunch of reasons.
I think edge edge inference is very

490
00:25:11,360 --> 00:25:13,837
important for a bunch of reasons.
I I think just practically speaking, if

491
00:25:13,920 --> 00:25:16,917
we're just going to have fundamental
construct supply crunches for the next I

492
00:25:17,000 --> 00:25:19,597
mean, you guys know that if you just
project forward demand over the next

493
00:25:19,680 --> 00:25:23,077
three years relative to supply,
one of the dismaying

494
00:25:23,160 --> 00:25:25,517
predictions you can do is what's going
to what's going to happen to the cost of

495
00:25:25,600 --> 00:25:28,197
of inference in the core over the next
three years. And like it may rise

496
00:25:28,280 --> 00:25:31,797
dramatically, right? Like So, so what
and then is is you know, like the the

497
00:25:31,880 --> 00:25:34,357
big model competition is subsidizing
heavily right now, right? And so, so

498
00:25:34,440 --> 00:25:38,077
what's the what will be the average
person's you know, per day per month

499
00:25:38,160 --> 00:25:41,437
token cost, you know, three years from
now to do all the things that they want

500
00:25:41,520 --> 00:25:44,197
to do? And I I don't know what's going
to be like I mean, I have you guys

501
00:25:44,280 --> 00:25:46,597
probably have friends. I have friends
today who are paying a thousand dollars

502
00:25:46,680 --> 00:25:50,477
a day for open cloud for cloud tokens
to run open cloud, right? And so, okay,

503
00:25:50,560 --> 00:25:53,557
$30,000 a month, right? And and by the
way, those friends have like a thousand

504
00:25:53,640 --> 00:25:56,597
more ideas of the things that they want
to claw to do, right? And so, you could

505
00:25:56,680 --> 00:25:59,757
imagine there's like latent demand of up
to I don't know, five or ten thousand

506
00:25:59,840 --> 00:26:04,117
dollars a day of of of tokens for a fully
deployed, you know, personal agent.

507
00:26:04,200 --> 00:26:07,437
And obviously, consumers can't pay that,
right? And so, so but it gives you a

508
00:26:07,520 --> 00:26:10,597
sense of the of the few of the future
scope of demand, right? And so, so even

509
00:26:10,680 --> 00:26:13,317
even if there's a ten x improvement
in price performance, that's still, you

510
00:26:13,400 --> 00:26:15,477
know, goes to a hundred dollars a day,
which is still way beyond what people

511
00:26:15,560 --> 00:26:19,277
can pay. So, there's just going to be like
ferocious demand. By the way, the agent

512
00:26:19,360 --> 00:26:22,237
thing the other interesting thing is I
think the agent thing so, up until now a

513
00:26:22,320 --> 00:26:24,917
lot of the constraints have been GPU
constraints. I think the agent thing now

514
00:26:25,000 --> 00:26:28,837
also translates into CPU constraints,
CPU and memory, yes. CPU and memory,

515
00:26:28,920 --> 00:26:31,477
right? And so, like the entire
chip ecosystem is just going to get

516
00:26:31,560 --> 00:26:33,997
>> With with our network constraints,
that will be the killer. It's all

517
00:26:34,080 --> 00:26:37,157
bottlenecked potentially for years.
And so, so I I think that Brad and I think

518
00:26:37,240 --> 00:26:39,517
it's actually possible, I mean,
generally inference costs are going to

519
00:26:39,600 --> 00:26:42,037
keep coming down, but I think the let's
put it this way, the rate of decline I

520
00:26:42,120 --> 00:26:45,197
think may level out here for a bit
because of these supply constraints. And

521
00:26:45,280 --> 00:26:48,037
then at some point maybe the labs stop
subsidizing so much and that that that

522
00:26:48,120 --> 00:26:51,197
again will be a be an issue. And so,
there's just going to be so much more

523
00:26:51,280 --> 00:26:54,597
demand for inference than than can
be satisfied you know kind of with a

524
00:26:54,680 --> 00:26:57,397
centralized model and then and then you
know you guys know this but like all the

525
00:26:57,480 --> 00:26:59,797
just the dramatic I mean just
the dramatic innovations that happened in

526
00:26:59,880 --> 00:27:03,317
the Apple Silicon to be able to do
inferences it's quite amazing. The level

527
00:27:03,400 --> 00:27:05,997
of effort being put like the open source
guys are putting incredible effort into

528
00:27:06,080 --> 00:27:08,277
getting you know this recurring
pattern where the big

529
00:27:08,360 --> 00:27:11,357
model will never run on a PC and then
six months later up it runs on a PC.

530
00:27:11,440 --> 00:27:15,197
Right. It's like amazing. And there's
very smart people working on that. So

531
00:27:15,280 --> 00:27:18,117
there's all that and then look there's
also you know there's also like other

532
00:27:18,200 --> 00:27:20,917
there's other motivators there's other
motivators which is just like okay how

533
00:27:21,000 --> 00:27:24,157
much trust are the big centralized
model providers you know

534
00:27:24,240 --> 00:27:27,237
how much trust are they building in the
market versus you know how much are you

535
00:27:27,320 --> 00:27:30,317
know at least for in certain cases with
some people for certain use cases people

536
00:27:30,400 --> 00:27:33,157
being like well I'm not willing to just
like turn everything over. So there

537
00:27:33,240 --> 00:27:36,837
there there's all the trust issues.
By the way there's also just like straight

538
00:27:36,920 --> 00:27:40,237
up price optimization. There's many uses
of AI where you don't need Einstein in

539
00:27:40,320 --> 00:27:43,997
the cloud you just need like
a smart local model. There's also

540
00:27:44,080 --> 00:27:47,317
performance issues where you want to you
know you want you know you're going to

541
00:27:47,400 --> 00:27:48,917
want your doorknob
to have an AI model in it.

542
00:27:49,000 --> 00:27:50,877
>> Right. You know to be
able to you know do you

543
00:27:50,960 --> 00:27:52,837
know to be able to do access control.

544
00:27:52,920 --> 00:27:56,157
Obviously like everything with a chip
is going to have an AI model in it and a

545
00:27:56,240 --> 00:27:57,797
lot of those are going to be local.

546
00:27:57,880 --> 00:28:00,397
And so yeah no like I think I think
you're going to have and then you're

547
00:28:00,480 --> 00:28:02,717
going to by the way also wearable
devices you know you you don't want to

548
00:28:02,800 --> 00:28:05,757
do a complete round trip you want you
know whatever your smart devices are you

549
00:28:05,840 --> 00:28:07,877
want it to be like super low latency.

550
00:28:07,960 --> 00:28:11,437
Yeah. The question do we care who makes
it? One of the biggest news this week

551
00:28:11,520 --> 00:28:13,237
was the collapse of AI2 the Allen

552
00:28:13,320 --> 00:28:18,517
Institute one of the actual American
open source model labs. Yeah. And

553
00:28:18,600 --> 00:28:21,757
I'm not that optimistic in the on
American open source. Like you you guys

554
00:28:21,840 --> 00:28:25,997
invested in Mistral and Mistral's doing
extremely well outside of China. That's

555
00:28:26,080 --> 00:28:28,397
about it. Yeah we'll see. We'll see.

556
00:28:28,480 --> 00:28:32,557
Look I number one I do think we care I do
think we I do think we care who makes it.

557
00:28:32,640 --> 00:28:34,997
I would say this the the previous
presidential administration

558
00:28:35,080 --> 00:28:38,557
wanted to kill it in the US.
They wanted to drown it in the bathtub.

559
00:28:38,640 --> 00:28:39,597
And [snorts] so they wanted to kill it.

560
00:28:39,680 --> 00:28:42,117
So at least we have a government now
that actually like actually wants it

561
00:28:42,200 --> 00:28:44,237
wants it to happen.
>> And you're on the council? And and I

562
00:28:44,320 --> 00:28:47,317
yeah and the new and the PCAST yeah.
So the you know this admin a for whatever

563
00:28:47,400 --> 00:28:49,277
other political issues people have,
which are many, you know, this

564
00:28:49,360 --> 00:28:51,717
administration has I think a very
enlightened view and in particular an

565
00:28:51,800 --> 00:28:55,117
enlightened view on AI and in particular
on open source AI. Uh and so, they're

566
00:28:55,200 --> 00:28:58,877
very supportive. Um my read is the Chi-
the Chinese have a very the various

567
00:28:58,960 --> 00:29:01,637
Chinese companies have a very specific
reason to do open source, which is they

568
00:29:01,720 --> 00:29:04,997
they they don't fundamentally they don't
think they can sell commercial uh AI

569
00:29:05,080 --> 00:29:08,837
outside of China right now and or at least
specifically not not in the US for

570
00:29:08,920 --> 00:29:11,717
a combination of reasons. And so, they
they kind of view I think open source AI

571
00:29:11,800 --> 00:29:14,957
is a bit of a loss leader against
basically domestic um you know, paid

572
00:29:15,040 --> 00:29:18,757
paid services and then kind of you know,
kind of ancillary products. You know,

573
00:29:18,840 --> 00:29:20,117
they're they're very excited about it.

574
00:29:20,200 --> 00:29:22,557
By the way, I think it's great. I think
it's great that they're doing it. Um you

575
00:29:22,640 --> 00:29:25,877
know, I think Deep Seek was like a gift
to the world. Um I think the great thing

576
00:29:25,960 --> 00:29:29,117
about open source open source the the
the impact of open source is felt two

577
00:29:29,200 --> 00:29:32,277
ways. One is you you get the software
for free. But the other is you get to

578
00:29:32,360 --> 00:29:35,277
learn how it works. Right.
And so, like the paper Uh the

579
00:29:35,360 --> 00:29:39,637
paper the paper and and the code. Right,
and the code. And so, like for example,

580
00:29:39,720 --> 00:29:42,677
I thought this was amazing. So, OpenAI
comes out with O1 and it's an amazing

581
00:29:42,760 --> 00:29:45,197
technical breakthrough and it's just
like absolutely fantastic. But of

582
00:29:45,280 --> 00:29:47,517
course, they don't explain how it works
in detail and then of course, they hide

583
00:29:47,600 --> 00:29:49,237
the they hide the reasoning traces.

584
00:29:49,320 --> 00:29:51,917
Right. [snorts] And and then and then
everybody's like, okay, this is great,

585
00:29:52,000 --> 00:29:53,997
but like who's going to be able
to replicate this? How are other people

586
00:29:54,080 --> 00:29:56,277
going to be able to do this? You know,
is there secret sauce in there? And then

587
00:29:56,360 --> 00:29:59,437
R1 comes out and it's just like, there's
the code and there's the paper. And now

588
00:29:59,520 --> 00:30:02,237
the whole world knows how to do it.
And then, you know, 3 months later every

589
00:30:02,320 --> 00:30:03,837
other AI model is is adding reasoning.

590
00:30:03,920 --> 00:30:07,197
And so, so you get this kind of double
like even if the Chinese models

591
00:30:07,280 --> 00:30:10,037
themselves are not the models that get
used, the education that's taken place

592
00:30:10,120 --> 00:30:13,157
to the rest of the world the information
diffusion, you know, is incredibly

593
00:30:13,240 --> 00:30:16,037
powerful. So, that happens. And then I
don't know. We'll we'll see. You know,

594
00:30:16,120 --> 00:30:19,037
there are a bunch of American,
you know, open source you know, AI model

595
00:30:19,120 --> 00:30:21,837
companies. I mean, look, there's
going to be tremendous, you know, there

596
00:30:21,920 --> 00:30:24,437
already is. There's you know, there's
going to be treme- there's tremendous

597
00:30:24,520 --> 00:30:28,117
competition among the primary
model companies. You know, there's

598
00:30:28,200 --> 00:30:32,397
depending on how you count, there's like
four or five, you know, big co- model

599
00:30:32,480 --> 00:30:36,317
companies now that are, you know, kind
of neck and neck in different ways. Um

600
00:30:36,400 --> 00:30:39,117
you know, and and and um you know, and
then obviously both both both X and then

601
00:30:39,200 --> 00:30:41,917
Meta where I'm involved you know, both
have huge, you know, huge attempts to,

602
00:30:42,000 --> 00:30:44,637
you know, kind of to kind of kind
of leapfrog on the way. And then you've

603
00:30:44,720 --> 00:30:46,877
got, you know, a whole fleet of startups,
new companies, including a

604
00:30:46,960 --> 00:30:49,077
whole bunch that were backed in that are,
you know, trying to kind of come up

605
00:30:49,160 --> 00:30:51,677
with different approaches. And then
you've got whatever it is, I don't know.

606
00:30:51,760 --> 00:30:55,157
How How many How many like main line
foundation model companies are there in

607
00:30:55,240 --> 00:30:59,477
China at this point? It's probably six >>
tigers, is what they call it. Uh Quen is

608
00:30:59,560 --> 00:31:02,277
in in questionable because
of there's changing leadership.

609
00:31:02,360 --> 00:31:04,997
>> Right. Yeah. But that doesn't
include that that includes like Moonshot.

610
00:31:05,080 --> 00:31:08,157
>> Yes. Kimi, DeepSeek, uh

611
00:31:08,240 --> 00:31:13,997
ZAI, um Quen, 01 is in there. Right. And
then um and ByteDance and and then you you

612
00:31:14,080 --> 00:31:16,957
see that ByteDance would be like
the next tier. They weren't as prominent.

613
00:31:17,040 --> 00:31:19,477
>> They weren't but now they're Yeah,
but they're you know, at least you know,

614
00:31:19,560 --> 00:31:21,797
SeaDance is very inspiring
and presumably they have more stuff coming

615
00:31:21,880 --> 00:31:24,957
and Tencent probably has more stuff
coming and and so forth. And so So So

616
00:31:25,040 --> 00:31:27,477
like look, here here would be a thing
you can anticipate, which is there are

617
00:31:27,560 --> 00:31:30,397
not these markets There are not going
to be Between the US and China right now,

618
00:31:30,480 --> 00:31:33,157
there's like a dozen primary foundation
model companies that are like at scale

619
00:31:33,240 --> 00:31:34,957
at some level of like critical mass.

620
00:31:35,040 --> 00:31:38,117
It's not going to be a dozen in 3 years,
right? Like it just because these

621
00:31:38,200 --> 00:31:40,557
industries don't bear a dozen. It's It's
going to be three or you know, there's

622
00:31:40,640 --> 00:31:43,117
going to be three or four big winners
or maybe one or two big winners. And so

623
00:31:43,200 --> 00:31:45,197
there's going to be like a whole bunch
of those guys that are going to have to

624
00:31:45,280 --> 00:31:48,317
figure out alternate strategies. Um and I
think like open source is one of those

625
00:31:48,400 --> 00:31:51,797
strategies. And so I I think you
could see like a whole I I I think the

626
00:31:51,880 --> 00:31:54,197
question's like who's going to do open
source? I think that could change really

627
00:31:54,280 --> 00:31:57,357
fast. I I think that that that's a very
dynamic thing. I think it's very hard to

628
00:31:57,440 --> 00:32:00,677
predict what happens. And And I think
it's very important. Nvidia's doing a

629
00:32:00,760 --> 00:32:02,837
lot. You mentioned Well, I was going
to say Well, exactly. And then you got

630
00:32:02,920 --> 00:32:06,037
Nvidia and then and then, you know, just
again industry for There's an old thing

631
00:32:06,120 --> 00:32:08,557
in business strategy which is called
a commoditize the commoditize the

632
00:32:08,640 --> 00:32:11,157
complement. And so right, it's like if
your Jensen is just kind of obvious. Of

633
00:32:11,240 --> 00:32:13,997
course you want to commoditize
the software. And he's And to his enormous

634
00:32:14,080 --> 00:32:16,837
credit, he's putting enormous resources
behind that. And so maybe maybe it's

635
00:32:16,920 --> 00:32:19,357
literally Nvidia.
And I think that would be great.

636
00:32:19,440 --> 00:32:23,977
Yeah. Uh narrative violation two
European projects. Uh innovating.

637
00:32:24,060 --> 00:32:27,717
>> [laughter] >> Bam.
I I'm mostly in my uh Europe uh

638
00:32:27,800 --> 00:32:29,437
conference soon and I got both of them.

639
00:32:29,520 --> 00:32:33,157
They got us. They got us, Mark. They got
us all. Wait a minute. Where was Peter

640
00:32:33,240 --> 00:32:35,277
So where was Steinberg
when he did Open Indiana?

641
00:32:35,360 --> 00:32:36,637
>> Yeah, yeah, yeah.
He was in what? He was in Vienna.

642
00:32:36,720 --> 00:32:38,797
>> Oh, he was in Vienna.
And then where is he now?

643
00:32:38,880 --> 00:32:40,357
Uh he's moving to SF. Okay. Okay.

644
00:32:40,440 --> 00:32:41,797
[laughter] All right. Okay. There we go.

645
00:32:41,880 --> 00:32:44,397
And then yeah, the Pi guy right? The Pi
guys are European. Yeah, they're they're

646
00:32:44,480 --> 00:32:46,077
buddies in Austria. Mario's also there.

647
00:32:46,160 --> 00:32:49,437
Yeah. Right. And are they Yeah, they
haven't announced yet any sort of change

648
00:32:49,520 --> 00:32:52,717
change or have they? No, they're they
have a company there. Okay, good. Okay,

649
00:32:52,800 --> 00:32:56,237
good. Good, good. Um Yeah, good.
Anyways, yeah, I think Pi

650
00:32:56,320 --> 00:32:59,677
and Open Claw are very important
software things and and I just wanted

651
00:32:59,760 --> 00:33:02,837
you to just go off on what you think.

652
00:33:02,920 --> 00:33:06,357
Yeah. So, I think in the combination of
the two of them, I think it's one of the

653
00:33:06,440 --> 00:33:10,357
10 most important software Open Claw got
all the attention, but talk about Pi. Pi

654
00:33:10,440 --> 00:33:13,317
Pi's kind of the end Yeah, Pi's Pi's
kind of the architectural breakthrough.

655
00:33:13,400 --> 00:33:16,357
For those of us who are older, there was
this whole thing that was very important

656
00:33:16,440 --> 00:33:20,237
in the world of software basically from
like 1970 to I don't know. It still is

657
00:33:20,320 --> 00:33:22,917
very important, but like 19 from 1970
through to like basically the creation

658
00:33:23,000 --> 00:33:25,557
of Linux, which is basically this this
thing we used to call like the Unix

659
00:33:25,640 --> 00:33:28,997
mindset. Like so so cuz there were
all these different, you know, theories

660
00:33:29,080 --> 00:33:31,717
there all these different operating
systems and mainframes and and then, you

661
00:33:31,800 --> 00:33:34,317
know, all these Windows and Mac and all
these things. And then there was this

662
00:33:34,400 --> 00:33:37,237
but kind of behind it all was this idea
of kind of the Unix mindset. And the

663
00:33:37,320 --> 00:33:40,677
Unix mindset was this thing where
basically you don't have these like like

664
00:33:40,760 --> 00:33:43,837
in the old days like like the operating
system that like made the computer

665
00:33:43,920 --> 00:33:47,197
actually really work like in the 1960s
was this thing called OS 360, which was

666
00:33:47,280 --> 00:33:49,717
this big operating system that IBM
developed that was supposed to basically

667
00:33:49,800 --> 00:33:52,757
run everything. And it was this like
giant monolithic architecture in the

668
00:33:52,840 --> 00:33:56,317
sky. It was like a you know, it was like
a giant castle um of software. And and

669
00:33:56,400 --> 00:33:58,637
by the way, it worked really well
and they were very successful with it, but

670
00:33:58,720 --> 00:34:00,877
like it was this huge castle in the sky,
but it was this thing it was almost

671
00:34:00,960 --> 00:34:03,797
unapproachable, which is like you had
to be kind of inside IBM or very close to

672
00:34:03,880 --> 00:34:06,597
IBM and you had to really understand
every aspect of how the system worked.

673
00:34:06,680 --> 00:34:09,517
And then the the Unix guys originally
out of AT&T and then out of out of

674
00:34:09,600 --> 00:34:12,437
Berkeley um you know, came out and they
said, "No, let's have a completely

675
00:34:12,520 --> 00:34:14,357
different architecture." And the way
architecture is going to work is we're

676
00:34:14,440 --> 00:34:17,677
going to have we're going to have
a a prompt and and a shell. And then and

677
00:34:17,760 --> 00:34:19,557
then we're going to all the functionality
is going to be in the form

678
00:34:19,640 --> 00:34:21,717
of these discrete modules. And then
you're going to be able to chain the

679
00:34:21,800 --> 00:34:24,997
modules together. And so like the the it's
almost like the operating operating

680
00:34:25,080 --> 00:34:27,917
system itself is going to be
a programming language. Um and then that

681
00:34:28,000 --> 00:34:31,077
led led to the the the sort of centrality
of the shell um and then that

682
00:34:31,160 --> 00:34:34,157
led to the sort of you know, basically
chaining together Unix tools and then

683
00:34:34,240 --> 00:34:37,237
that led to the emergence of these these
scripting languages like Pearl where you

684
00:34:37,320 --> 00:34:39,317
could basically kind
of very easily do this.

685
00:34:39,400 --> 00:34:41,437
And then the shells got more
sophisticated. And then and then and

686
00:34:41,520 --> 00:34:44,517
then look like you know that that number
one that worked. And that that was the

687
00:34:44,600 --> 00:34:47,757
world I grew up in like I was I was a
Unix guy you know sort of from call it

688
00:34:47,840 --> 00:34:51,677
1988 to you know kind of all the way
through my work. And it worked really

689
00:34:51,760 --> 00:34:55,357
well. It it's in the background
you know normal people don't need it

690
00:34:55,440 --> 00:34:58,037
didn't need it necessarily know about it.
But like if you were doing like

691
00:34:58,120 --> 00:35:00,437
system architecture
application development

692
00:35:00,520 --> 00:35:03,517
you knew all about it. And then you know
it's been in the background ever since

693
00:35:03,600 --> 00:35:06,517
and you know look your Mac still has
a Unix shell you know kind of in there and

694
00:35:06,600 --> 00:35:09,517
your iPhone still has a Unix shell kind
of buried in there somewhere. So they're

695
00:35:09,600 --> 00:35:11,957
kind of in there and then you know the
Windows shell is kind of a you know sort

696
00:35:12,040 --> 00:35:15,277
of a weird derivative of that but you
know but look the the internet runs on

697
00:35:15,360 --> 00:35:19,037
Unix and then smartphones actually both
iOS and Android are Unix derivatives and

698
00:35:19,120 --> 00:35:22,277
so you know kind of Unix did end up
winning but but anyway and then we just

699
00:35:22,360 --> 00:35:24,957
started taking that for granted and then
and then so so basically the way I think

700
00:35:25,040 --> 00:35:28,117
about what happened with Pi and then
with Open Claw is basically what those

701
00:35:28,200 --> 00:35:30,837
guys figured out is I would say
the the great breakthroughs are obvious in

702
00:35:30,920 --> 00:35:33,037
retrospect right which is the best kind.

703
00:35:33,120 --> 00:35:35,917
The best kind. They weren't obvious
at the time or somebody else would have

704
00:35:36,000 --> 00:35:38,757
done them already.
And so there is a like a real conceptual

705
00:35:38,840 --> 00:35:41,397
leap but then you look at it sort
of backwards looking and you're just like

706
00:35:41,480 --> 00:35:43,957
oh of course. Like to me
those are always the best

707
00:35:44,040 --> 00:35:46,117
breakthroughs. Well actually language
models themselves are like that it's

708
00:35:46,200 --> 00:35:49,317
just like oh next token completion oh
of course. Yeah what other objective

709
00:35:49,400 --> 00:35:52,357
mattered? Yeah exactly [laughter] but
but like right but she's even saying it

710
00:35:52,440 --> 00:35:54,837
wasn't obvious until somebody actually
did it right. And so the conceptual

711
00:35:54,920 --> 00:35:57,757
breakthrough is real and deep and
powerful and and very important. And so

712
00:35:57,840 --> 00:36:00,957
the way I think about Pi and Open
Claw is it's basically marrying the the

713
00:36:01,040 --> 00:36:05,437
language model mindset to the to the Unix
basically shell prompt mindset. And

714
00:36:05,520 --> 00:36:09,677
so it's it's basically this idea
that what what so what is an agent? Right

715
00:36:09,760 --> 00:36:11,917
and as you know like many smart people
have been trying to figure out what an

716
00:36:12,000 --> 00:36:14,677
agent is for for for decades and they've
had many architectures to build agents

717
00:36:14,760 --> 00:36:17,077
and the whole thing and it turns out
what is an agent. So it turns out what

718
00:36:17,160 --> 00:36:20,557
we now know is an agent is the following.
It's it's it's language model

719
00:36:20,640 --> 00:36:24,237
and then above that it's a bash it's
a bash shell. So it's it's a Unix shell.

720
00:36:24,320 --> 00:36:27,637
And then it's in then the agent has
access has access to to the shell in you

721
00:36:27,720 --> 00:36:30,957
know hopeful hopefully in a sandbox
maybe in maybe in a sandbox.

722
00:36:31,040 --> 00:36:34,957
So it's it's the model it's the shell
and then it's a it's a file system

723
00:36:35,040 --> 00:36:36,917
and then the state is stored in files.

724
00:36:37,000 --> 00:36:39,157
And then you know there's the markdown
format for the you know for for the

725
00:36:39,240 --> 00:36:41,717
files themselves and then and then
there's basically what in Unix is called

726
00:36:41,800 --> 00:36:44,357
a cron job. There's a loop and then
there's a heartbeat there's a heartbeat

727
00:36:44,440 --> 00:36:47,077
and and the thing basically
wake wakes up wakes up.

728
00:36:47,160 --> 00:36:51,597
So it's basically LLM plus shell plus
file system plus markdown plus cron and

729
00:36:51,680 --> 00:36:54,797
it turns out that's an agent and every
part of that other than the model is

730
00:36:54,880 --> 00:36:57,237
something that we already completely
know and understand and in fact it turns

731
00:36:57,320 --> 00:37:00,877
out that like the latent power
of the Unix shell is like extraordinary.

732
00:37:00,960 --> 00:37:04,757
Because basically like all like there's
just like there's just enormous latent

733
00:37:04,840 --> 00:37:07,077
power in the shell there's enormous
numbers of Unix commands there's

734
00:37:07,160 --> 00:37:09,437
enormous number of command line
interfaces into all kinds of things

735
00:37:09,520 --> 00:37:12,397
already in the you know your entire I
mean your entire just to start with your

736
00:37:12,480 --> 00:37:15,757
computer runs on a shell. If you're
running a Mac or a phone your computer

737
00:37:15,840 --> 00:37:17,397
your computer is running on a shell.

738
00:37:17,480 --> 00:37:20,197
Already and so like the full power
of your computer is available at the

739
00:37:20,280 --> 00:37:23,517
command line level and then it turns
out it's really easy to expose other

740
00:37:23,600 --> 00:37:24,957
functions as a command line interface.

741
00:37:25,040 --> 00:37:28,037
And so like this this whole idea where
we need like MCP and these like product

742
00:37:28,120 --> 00:37:30,797
fancy protocols whatever it's like no
we don't we just need like a command

743
00:37:30,880 --> 00:37:33,477
command line thing. So that's the
architecture and then it turns out what

744
00:37:33,560 --> 00:37:36,917
is your agent? Your agent is a bunch
of files stored in a file system and then

745
00:37:37,000 --> 00:37:38,957
there's the thing that just like
completely blew my mind when I wrapped

746
00:37:39,040 --> 00:37:41,477
my head around it as a result
of this which is like okay.

747
00:37:41,560 --> 00:37:44,037
This means your agent is now actually
independent of the model that it's

748
00:37:44,120 --> 00:37:47,477
running on because you can actually
swap out a different LLM underneath your

749
00:37:47,560 --> 00:37:50,397
agent and your your agent will change
personality somewhat cuz the model is

750
00:37:50,480 --> 00:37:53,717
different but all of the state stored
in the files will be retained. Different

751
00:37:53,800 --> 00:37:55,397
instruction set but you just compile it.

752
00:37:55,480 --> 00:37:57,477
>> Right exactly and then it's all right
it's like right swapping out a chip and

753
00:37:57,560 --> 00:38:00,157
recompiling but it's it's still
it's still your agent with all of its

754
00:38:00,240 --> 00:38:03,437
memories and with all of its
capabilities and then by the way you can

755
00:38:03,520 --> 00:38:06,797
also swap out the shell so you can move
it to a different execution environment

756
00:38:06,880 --> 00:38:09,557
that is also is also a bash shell. By
the way you can also switch out the file

757
00:38:09,640 --> 00:38:13,477
system. Right and you can and you can and
you can swap out the the the heartbeat for

758
00:38:13,560 --> 00:38:17,477
the the cron framework the the loop
the the agent framework itself and so your

759
00:38:17,560 --> 00:38:21,557
agent basically is basically at the end
of the day it's just it's just its files

760
00:38:21,640 --> 00:38:23,877
Um, then and then there is of course the
purple cloth. Yeah, it's it's basically

761
00:38:23,960 --> 00:38:26,077
it's just the files. >> [laughter]
>> Um, and then by the way, as a

762
00:38:26,160 --> 00:38:28,997
consequence of that, the agent its
and then the agent itself, it turns out a

763
00:38:29,080 --> 00:38:32,477
couple important things. So, one is it
it's it it can migrate itself. Right?

764
00:38:32,560 --> 00:38:34,957
And so, you you you you can instruct
your agent, "Migrate yourself to a

765
00:38:35,040 --> 00:38:36,557
different uh runtime environment.

766
00:38:36,640 --> 00:38:39,037
Migrate yourself to a different file
system. Migrate yourself to a different

767
00:38:39,120 --> 00:38:41,117
you know, like we swap out the language
model." Your agent will do all that

768
00:38:41,200 --> 00:38:44,117
stuff for you. And then there's
the final thing, which is just amazing,

769
00:38:44,200 --> 00:38:47,197
which is the agent is the agent actually
has full introspection. It actually it

770
00:38:47,280 --> 00:38:50,237
actually knows about its own files,
and it could rewrite its own files.

771
00:38:50,320 --> 00:38:52,837
Right? Which which by the way,
is basically no widely deployed software

772
00:38:52,920 --> 00:38:55,637
system in history where the the the
thing that you're using actually has

773
00:38:55,720 --> 00:38:58,597
full introspective knowledge of how
it itself works and is able to modify

774
00:38:58,680 --> 00:39:02,277
itself like that. That I mean, there've
been toy systems that have had that, but

775
00:39:02,360 --> 00:39:04,237
there there's never been a widely
deployed system that has that

776
00:39:04,320 --> 00:39:07,997
capability. And then that leads you to
the capability that just like completely

777
00:39:08,080 --> 00:39:10,437
blew my mind when I wrapped my head
around it, which is you can tell the

778
00:39:10,520 --> 00:39:15,277
agent to add new functions and features
to itself, and it can do that.

779
00:39:15,360 --> 00:39:17,917
Right? Ex- extend yourself. Like,
extend yourself. Give yourself a new

780
00:39:18,000 --> 00:39:20,917
capability. Right? And so, and so,
literally it's just like you run into

781
00:39:21,000 --> 00:39:23,397
somebody at a party, and they're like,
"Oh, I have my open claw do whatever,

782
00:39:23,480 --> 00:39:25,597
connect to my eight sleep bed, and it
gives me better advice in sleep." And

783
00:39:25,680 --> 00:39:28,077
you go home at night, and you tell your
claw, or if they're at the party, by the

784
00:39:28,160 --> 00:39:30,437
way, you tell your claw, "Oh, add this
capability to yourself." And your claw

785
00:39:30,520 --> 00:39:32,997
will say, "Oh, okay, no problem." And
it'll go out on the internet, and it'll

786
00:39:33,080 --> 00:39:35,637
figure out whatever it needs, and then
it'll go out to claw code or whatever,

787
00:39:35,720 --> 00:39:37,797
it'll write whatever it needs, and then
the next thing you know, it has this new

788
00:39:37,880 --> 00:39:41,237
capability. And so, you don't even have
to like you can have it upgrade itself

789
00:39:41,320 --> 00:39:44,437
without even having to without having
to do anything other than tell it that you

790
00:39:44,520 --> 00:39:47,677
want it to do that. And so, anyway,
so the the combination of all this is just

791
00:39:47,760 --> 00:39:51,757
I mean, this is just like a massive
incredible I mean, it's just

792
00:39:51,840 --> 00:39:54,557
incredible. Like, if I if I were if I
were 18, like this is a hundred this is

793
00:39:54,640 --> 00:39:56,797
what I would be spending all of my time
on. This is like such an incredible

794
00:39:56,880 --> 00:39:59,757
conceptual breakthrough. And again, pe-
people are going to look at it, and they

795
00:39:59,840 --> 00:40:01,277
already get this response. People are
going to look at it, and they're going

796
00:40:01,360 --> 00:40:03,357
to say, "Oh, well, where's
the breakthrough?" Cuz these all of these

797
00:40:03,440 --> 00:40:05,117
components were already known before.

798
00:40:05,200 --> 00:40:07,877
But but this is the key. The key
to the breakthrough was by using all these

799
00:40:07,960 --> 00:40:10,877
components that were known before,
you get all of the underlying capability

800
00:40:10,960 --> 00:40:13,597
that's buried in there. And so, all
and so, for example, computer use all of a

801
00:40:13,680 --> 00:40:16,237
sudden just kind of falls trivially
trivial. Of course, it's going to be

802
00:40:16,320 --> 00:40:19,397
able to use a computer. It [snorts]
has full access to the shell. Right? And

803
00:40:19,480 --> 00:40:21,677
then And then you just you you give it
access to a browser and then you've got

804
00:40:21,760 --> 00:40:24,357
the computer and the browser and and off
and away it goes. And then then you've

805
00:40:24,440 --> 00:40:27,397
got all the abilities of the browser,
also. Um and so and so the capability

806
00:40:27,480 --> 00:40:31,037
unlock here is profound. My friends who
were, you know, deepest into this are

807
00:40:31,120 --> 00:40:34,637
having their claw do like like literally
like a thousand things in their lives.

808
00:40:34,720 --> 00:40:36,717
They have new ideas every day.
They're just like constantly throwing new

809
00:40:36,800 --> 00:40:40,277
challenges at the thing. And by the way,
it's early and you know, these are you

810
00:40:40,360 --> 00:40:42,997
know, these are prototypes and there
is you know, as you guys know, there's

811
00:40:43,080 --> 00:40:46,157
security issues. And and so you know,
there's a bunch of stuff to be ironed

812
00:40:46,240 --> 00:40:51,637
out, but the the unlock of capability is
just incredible. And I I have absolutely

813
00:40:51,720 --> 00:40:54,397
no doubt that everybody in the world
is going to is going to have at least, you

814
00:40:54,480 --> 00:40:57,117
know, an agent like this, if not an entire
family of agents, and we're going

815
00:40:57,200 --> 00:40:59,957
to be living in a world where I
think it's almost inevitable now that

816
00:41:00,040 --> 00:41:02,637
this is the way people are going
to use computers. I was going to say for

817
00:41:02,720 --> 00:41:05,837
someone who's deeply familiar with
social networks, the next step is your

818
00:41:05,920 --> 00:41:09,477
claw talking to my claw.
Posting on claw Facebook, posting their

819
00:41:09,560 --> 00:41:13,877
jobs on claw LinkedIn, and close posting
their tweets on claw XAI or whatever,

820
00:41:13,960 --> 00:41:17,477
you know. >> [laughter] >> Um I
do think that that is how uh you

821
00:41:17,560 --> 00:41:21,437
know, we we get into some danger there
in terms of like alignment and whether

822
00:41:21,520 --> 00:41:24,717
or not we want these things to to to run.
You guys know about Rent-A-

823
00:41:24,800 --> 00:41:26,917
Rent-A-Human.com? >> Yeah, Rent-A- Yeah.

824
00:41:27,000 --> 00:41:29,757
I mean, it's Fiverr,
it's Task Rabbit, it's >> course.

825
00:41:29,840 --> 00:41:32,997
Mechanical Mechanical Turk.
>> Yeah, but flipped. Right? The agent

826
00:41:33,080 --> 00:41:35,497
hiring the people. Yeah.
Which of course is going to happen.

827
00:41:35,580 --> 00:41:37,117
>> [laughter] >> Right?
It's obviously going to happen.

828
00:41:37,200 --> 00:41:40,357
I'm curious if you have any thoughts on
the engineering side. So, when you built

829
00:41:40,440 --> 00:41:44,037
the browser, the internet,
you know, just a bunch of

830
00:41:44,120 --> 00:41:46,157
mostly plain text file plus some images.

831
00:41:46,240 --> 00:41:50,517
And today the every website and app
is like so complex and like somehow, you

832
00:41:50,600 --> 00:41:53,637
know, the browser kept
evolving to fit that in.

833
00:41:53,720 --> 00:41:57,877
Are there any design choices that were
made like early in the browser and kind

834
00:41:57,960 --> 00:42:02,477
of like the internet and the protocols
that you're seeing agents similar today

835
00:42:02,560 --> 00:42:05,437
say, "Hey, this thing is just not
going to work for like this type of new

836
00:42:05,520 --> 00:42:08,917
compute and we should just rip
it out right now." There were a

837
00:42:09,000 --> 00:42:09,917
whole bunch, but I'll give you a couple.

838
00:42:10,000 --> 00:42:13,117
So, one is um and we did
you know, to be clear, like

839
00:42:13,200 --> 00:42:15,277
this this was not you know this was
totally different we didn't have the

840
00:42:15,360 --> 00:42:18,157
capabilities we have today but we didn't
have the language models underneath this

841
00:42:18,240 --> 00:42:22,917
but we did have this idea that human
readability actually mattered a great deal

842
00:42:23,000 --> 00:42:25,877
and so specifically in those days it was
it was not so much English language but

843
00:42:25,960 --> 00:42:29,037
it was there was a design decision to be
made between binary protocols and text

844
00:42:29,120 --> 00:42:32,197
protocols. And basically every every every

845
00:42:32,280 --> 00:42:36,237
basically old school systems architect
that had grown up between like the 1960s

846
00:42:36,320 --> 00:42:39,317
and 1990s basically said you know
the internet is what do you know about the

847
00:42:39,400 --> 00:42:43,197
internet it's starved for bandwidth.
You just you have these very narrow straws

848
00:42:43,280 --> 00:42:45,717
you know like people when we did
the work on Mosaic like people who had the

849
00:42:45,800 --> 00:42:48,837
internet at home had a 14 kilobit modem
right so you're you're trying to like

850
00:42:48,920 --> 00:42:52,197
hyper optimize every bit of data that
that travels over the network. And so

851
00:42:52,280 --> 00:42:54,837
obviously if you're going to design
a protocol like HTTP you're going to want

852
00:42:54,920 --> 00:42:57,837
it to be a you know highly compressed
binary protocol for maximum efficiency

853
00:42:57,920 --> 00:43:00,077
and you're going to want to have it be
like a single connection that persists

854
00:43:00,160 --> 00:43:02,037
and you're you're the last thing you're
going to want to do is like bring up and

855
00:43:02,120 --> 00:43:04,437
tear down new connections.
And you definitely you're not going to

856
00:43:04,520 --> 00:43:07,317
not going to want a text protocol and so
of course we said no we actually want to

857
00:43:07,400 --> 00:43:11,597
go completely the other direction and it's
obviously we only want text protocols.

858
00:43:11,680 --> 00:43:15,157
By the way same thing in HTML itself we
want HTML to be relatively verbose you

859
00:43:15,240 --> 00:43:18,477
know we want the tags to actually be
like human readable. We want to use the

860
00:43:18,560 --> 00:43:19,837
>> The most inefficient things possible.

861
00:43:19,920 --> 00:43:22,757
>> Yeah we want to do the do the do the
inefficient things. You're the original

862
00:43:22,840 --> 00:43:26,197
token maxer. Yeah exactly yeah
yeah yeah basically it's just like

863
00:43:26,280 --> 00:43:28,597
Well yeah well actually this was this
was actually the the conscious thing

864
00:43:28,680 --> 00:43:32,037
which basically says just like assume
assume a future of infinite bandwidth

865
00:43:32,120 --> 00:43:35,677
build for that and then basically what
it was is it was a bet that it was a bet

866
00:43:35,760 --> 00:43:39,037
that if the system if the if the latent
capabilities of the system were powerful

867
00:43:39,120 --> 00:43:41,637
enough and that was obvious enough
to people that would create the demand for

868
00:43:41,720 --> 00:43:44,157
the bandwidth that would cause
the supply of bandwidth to get built that

869
00:43:44,240 --> 00:43:46,877
would actually make the whole thing work.
And then specifically what we

870
00:43:46,960 --> 00:43:49,117
wanted was we wanted everything
to be human readable because we at the

871
00:43:49,200 --> 00:43:51,597
engineering level we wanted people to be
able to read the protocol coming over

872
00:43:51,680 --> 00:43:53,157
the wire and be able to understand it.

873
00:43:53,240 --> 00:43:55,877
With their with their bare eyes
without having to like disassemble it or

874
00:43:55,960 --> 00:43:59,037
whatever right you have to convert it
out of binary right and so the the all

875
00:43:59,120 --> 00:44:02,797
the you know HTTP and everything else
were were it was always text protocols

876
00:44:02,880 --> 00:44:05,797
and the same thing with HTML and in many
ways some people say that the key break

877
00:44:05,880 --> 00:44:08,997
through in the browser was the view
source option, which is every webpage

878
00:44:09,080 --> 00:44:11,357
you go to you could view source, which
means you could see how it worked, which

879
00:44:11,440 --> 00:44:15,557
means you could teach yourself how
to build right new to build new webpages.

880
00:44:15,640 --> 00:44:19,517
There was that. So, human readability
and again human readability in those

881
00:44:19,600 --> 00:44:22,357
days still meant technical, you know,
specs. You know, now it means English

882
00:44:22,440 --> 00:44:25,677
language, but there's an incredible
latent power in giving everybody who

883
00:44:25,760 --> 00:44:28,477
uses the system the option to be able
to drop down and actually understand and

884
00:44:28,560 --> 00:44:31,037
see how it's working. And that worked
really well for the web and I think it's

885
00:44:31,120 --> 00:44:35,277
working really well for AI.
That was one. What was the other um A

886
00:44:35,360 --> 00:44:38,557
big part of the idea of web servers
was actually surface the underlying latent

887
00:44:38,640 --> 00:44:41,717
capability of the operating system
and to be able to surface the also the

888
00:44:41,800 --> 00:44:44,517
underlying latent capability of the
database. Because basically what was a

889
00:44:44,600 --> 00:44:47,117
web server? What what what what is a web
server fundamentally architecturally?

890
00:44:47,200 --> 00:44:49,997
It's it's it's the operating system. So,
it's it's the operating system's ability

891
00:44:50,080 --> 00:44:53,157
to, you know, it's running on top of an
OS. So, it's the OS's ability to manage

892
00:44:53,240 --> 00:44:56,557
the file system and do everything else
that you want to do, process everything.

893
00:44:56,640 --> 00:44:58,997
And then of course a lot of really, you
know, a lot a lot of websites are front

894
00:44:59,080 --> 00:45:01,957
ends to databases. And so, you wanted
to you wanted to unleash the underlying

895
00:45:02,040 --> 00:45:05,077
latent power of whether it was an Oracle
database or some other, you know, some

896
00:45:05,160 --> 00:45:08,357
other postgress or whatever whatever
it was. And so, a lot of the function of

897
00:45:08,440 --> 00:45:12,317
the web server was to just bridge from
that internet connection coming in to be

898
00:45:12,400 --> 00:45:15,117
able to unlock the underlying
power of the OS and the database.

899
00:45:15,200 --> 00:45:17,557
And again, people looked at it at the
time and they were like, well, is this

900
00:45:17,640 --> 00:45:19,997
really does this really matter?
Like is this important because we've had

901
00:45:20,080 --> 00:45:22,997
databases forever and we've always had,
you know, user interfaces for databases

902
00:45:23,080 --> 00:45:25,317
and this is just another
user interface for a database.

903
00:45:25,400 --> 00:45:27,957
It's like, okay, yeah, fair enough.
But on the other side of that is just like

904
00:45:28,040 --> 00:45:31,077
this is now a much better interface
to databases and one that 8 billion people

905
00:45:31,160 --> 00:45:34,677
are going to use and is going to be like
far easier to use and far more flexible

906
00:45:34,760 --> 00:45:37,797
and and and and you're not just going
to have old databases. Now you have a

907
00:45:37,880 --> 00:45:40,477
system where people can actually
understand why they want to build, you

908
00:45:40,560 --> 00:45:43,397
know, a million times more database apps
than they have in the past. And then the

909
00:45:43,480 --> 00:45:46,077
number of databases in the world exploded.
And so, again this goes to

910
00:45:46,160 --> 00:45:49,077
this thing of like building building
in layers. Some of the smartest people in

911
00:45:49,160 --> 00:45:51,957
industry look at any new challenge and
they're like, okay, I I need to build a

912
00:45:52,040 --> 00:45:53,997
new kind of application. So, the first
thing I need to do is build a new

913
00:45:54,080 --> 00:45:56,397
programming language. Right?
And then the next thing I need to

914
00:45:56,480 --> 00:45:57,497
do is build a new operating system.

915
00:45:57,580 --> 00:45:59,277
[laughter] Right? And then the next
thing I need to do is I need to build a

916
00:45:59,360 --> 00:46:02,717
new chip. Right? And they they kind
of want to reinvent everything. And I've

917
00:46:02,800 --> 00:46:06,157
I've always had maybe it's just I don't
know pragmatic mentality or something or

918
00:46:06,240 --> 00:46:08,997
maybe an engineering over science
mentality, but it's more like no, you

919
00:46:09,080 --> 00:46:12,437
have just like all of this latent power
in the existing systems and you don't

920
00:46:12,520 --> 00:46:15,317
want to be held back by their constraints,
but what you want to do is

921
00:46:15,400 --> 00:46:18,397
you want to kind of liberate that power
and open it up. So I I think I think

922
00:46:18,480 --> 00:46:21,037
that and I think the web did that for
those reasons and I think it's the same

923
00:46:21,120 --> 00:46:24,517
thing that's happening.
Programming languages is another good

924
00:46:24,600 --> 00:46:28,077
thing. We have Brett Taylor on the
podcast and we were talking about rust

925
00:46:28,160 --> 00:46:31,597
and you know, rust is memory safe
by default. So why are we teaching the

926
00:46:31,680 --> 00:46:33,277
model to not write memory unsafe code?

927
00:46:33,360 --> 00:46:36,797
Just use rust and then you get it for
free. How much do you think there's like

928
00:46:36,880 --> 00:46:40,197
time to be spent like recreating some
of these things instead of taking them for

929
00:46:40,280 --> 00:46:43,077
granted? I'll be like,
oh okay, python is kind of slow.

930
00:46:43,160 --> 00:46:47,597
You know, it's like yeah. As as imperfect
as they are, they are the lingua franca.

931
00:46:47,680 --> 00:46:50,277
I mean, I think this is going to change
a lot cuz I don't think the models care

932
00:46:50,360 --> 00:46:52,477
what language they program in and I
think they're going to be good at

933
00:46:52,560 --> 00:46:54,517
programming in every language and I
think they're going to be good at

934
00:46:54,600 --> 00:46:57,957
translating from any language to any
other language. Like okay, so this gets

935
00:46:58,040 --> 00:47:01,637
into the coding side of things. I think
we're going through a really fundamental

936
00:47:01,720 --> 00:47:04,757
change. And then look, I I grew up you
know, I grew up hand code. You know, I I

937
00:47:04,840 --> 00:47:07,637
grew up hand coding. Everything I did
was actually everything I did actually

938
00:47:07,720 --> 00:47:11,597
was written in C. I wasn't even I
wasn't even using C++. So I or like

939
00:47:11,680 --> 00:47:15,317
Java or any of this stuff, right? And so
everything everything I ever did I was

940
00:47:15,400 --> 00:47:18,277
like managing my own
memory at the level of C.

941
00:47:18,360 --> 00:47:20,837
And then I you know, I still from
the generation that you know, I knew

942
00:47:20,920 --> 00:47:23,237
assembly language and you
know, I I you know,

943
00:47:23,320 --> 00:47:26,157
so I could drop down and do
things right on the ship.

944
00:47:26,240 --> 00:47:29,437
And so we've just we've all all of us
we've always lived in a world in which

945
00:47:29,520 --> 00:47:32,757
software is like this precious thing
that like you have to think about very

946
00:47:32,840 --> 00:47:35,397
carefully and it's like really hard
to generate good software and there's only

947
00:47:35,480 --> 00:47:37,957
a small number of people who can do
it and like you have to be very like

948
00:47:38,040 --> 00:47:40,397
jealous in terms of thinking about like
how do you allocate? Like what are your

949
00:47:40,480 --> 00:47:43,397
engineers working on and how many good
engineers do you actually have and how

950
00:47:43,480 --> 00:47:46,197
much software can they write and how can
how much software can human beings you

951
00:47:46,280 --> 00:47:49,237
know, kind of maintain? And I think
like all those assumptions are being shot

952
00:47:49,320 --> 00:47:51,437
right out the window right now.
Like I think they're

953
00:47:51,520 --> 00:47:54,557
I think those days are just over and I
think the new world is like actually

954
00:47:54,640 --> 00:47:57,237
high quality software is just
like infinitely available.

955
00:47:57,320 --> 00:48:00,717
And [snorts] if you need new software
to do XYZ, like you're just going to wave

956
00:48:00,800 --> 00:48:01,677
your hand and you're going to get it.

957
00:48:01,760 --> 00:48:04,357
And then if it's if you don't like the
language it's written in, you just tell

958
00:48:04,440 --> 00:48:07,117
the thing, all right, I want
the right now I want the rest version.

959
00:48:07,200 --> 00:48:09,917
Um or you know, secure you know, secure
we're about to by the way, we're about

960
00:48:10,000 --> 00:48:12,437
to go through computer security about
to go through the most dramatic change

961
00:48:12,520 --> 00:48:15,237
ever, which is number one, like every
single latent security bug is about to

962
00:48:15,320 --> 00:48:17,317
be exposed. >> Right.
So it's we're going to have like

963
00:48:17,400 --> 00:48:20,277
the we're we're set up here for like
the computer security apocalypse for a

964
00:48:20,360 --> 00:48:23,317
while. But but but on the other side
of it, now we have coding agents that can

965
00:48:23,400 --> 00:48:25,917
go in and actually fix all the security
bugs. And so how how are you going to

966
00:48:26,000 --> 00:48:28,237
secure software in the future? You're
going to tell the tell the bot to secure

967
00:48:28,320 --> 00:48:30,597
it and it's going to go
through and and fix it all.

968
00:48:30,680 --> 00:48:33,637
And so so this thing that was
this incredibly scarce resource of

969
00:48:33,720 --> 00:48:36,917
high-quality software is just going
to become a completely fungible thing that

970
00:48:37,000 --> 00:48:38,677
you're just going to have
as much as you want.

971
00:48:38,760 --> 00:48:41,077
>> Right. Uh and and that has
like, you know, that has like

972
00:48:41,160 --> 00:48:43,037
tons and tons of consequences.

973
00:48:43,120 --> 00:48:45,157
In some sense, the answer to the
question you posed I I think is just

974
00:48:45,240 --> 00:48:47,157
somewhat, I don't know, simple
or something or straightforward, which is

975
00:48:47,240 --> 00:48:49,397
just if you want all your software in
rest, you just tell the bot you want all

976
00:48:49,480 --> 00:48:53,317
your software in rest. Like
the things that used to be like hard or

977
00:48:53,400 --> 00:48:56,277
even like seem like an insurmountable
mountain to get to get through all of a

978
00:48:56,360 --> 00:48:59,197
sudden I think become very easy.
Uh I think Brett had a theory that there

979
00:48:59,280 --> 00:49:03,597
would be a more optimal language
for LLMs. And so the contention is

980
00:49:03,680 --> 00:49:05,317
there isn't. Like just don't bother.

981
00:49:05,400 --> 00:49:09,757
Just whatever humans already use, LLMs
are perfectly capable porting. I think

982
00:49:09,840 --> 00:49:12,117
we're pretty close to being I don't know
if this would work today. I think we're

983
00:49:12,200 --> 00:49:16,477
pretty close to being able to ask the AI
what what its optimal language be and let

984
00:49:16,560 --> 00:49:17,927
it right and let it design it.

985
00:49:18,010 --> 00:49:21,237
>> [laughter] >> True.
Okay, here's the question.

986
00:49:21,320 --> 00:49:24,037
Are you going to even going to have
programming languages in the future? Or

987
00:49:24,120 --> 00:49:26,717
the AI or the AI is just
going to be emitting binaries.

988
00:49:26,800 --> 00:49:29,197
Let's assume for a moment that humans
aren't coding anymore. Let's assume it's

989
00:49:29,280 --> 00:49:32,037
all bots. The what levels
of intermediate abstraction

990
00:49:32,120 --> 00:49:35,637
do the bots even need? Yeah. Or are they
just coding binary directly? Did you see

991
00:49:35,720 --> 00:49:37,597
there's actually an exper- there's
somebody just did this thing where they

992
00:49:37,680 --> 00:49:40,397
have a they have a language model now
that actually emits model weights for a

993
00:49:40,480 --> 00:49:44,757
new language model. >> [laughter]
>> Right. And so will the bots be

994
00:49:44,840 --> 00:49:47,597
>> Just predict the weights. Well, yeah,
will the bots literally be emitting not

995
00:49:47,680 --> 00:49:50,397
just coding binaries, but will they will
will they actually be emitting weights

996
00:49:50,480 --> 00:49:55,317
for new for new models.
Direct, directly. And conceptually

997
00:49:55,400 --> 00:49:59,117
there's no reason why they
can't do both of those things.

998
00:49:59,200 --> 00:50:01,837
Just architecturally both of those
things seem completely possible.

999
00:50:01,920 --> 00:50:05,157
>> Very inefficient. You're basically
a simulation of a simulation in a

1000
00:50:05,240 --> 00:50:06,597
simulation inside of the weights. Yeah.

1001
00:50:06,680 --> 00:50:08,637
Yeah, very inefficient.
But like look, LLMs

1002
00:50:08,720 --> 00:50:10,237
are already like incredibly inefficient.

1003
00:50:10,320 --> 00:50:14,157
Ask I'm saying the same thing. Ask Claude
2 + 2 = 4, right? It's just like you know,

1004
00:50:14,240 --> 00:50:16,997
it's like you know, it's it's it's like
whatever billions and billions of times

1005
00:50:17,080 --> 00:50:19,997
more inefficient than using
your pocket calculator.

1006
00:50:20,080 --> 00:50:23,437
But but but yeah, the the payoff is so
great of the general capability. So

1007
00:50:23,520 --> 00:50:26,957
anyway, like I I kind of think in 10
years like I'm not sure yeah, like I'm

1008
00:50:27,040 --> 00:50:30,717
not sure there will even be a salient
concept of a programming language

1009
00:50:30,800 --> 00:50:31,957
in the way that we understand it today.

1010
00:50:32,040 --> 00:50:35,157
And in fact, what we may be doing more
and more is a form of interpretability,

1011
00:50:35,240 --> 00:50:38,957
which is we're trying to understand why
the bots have decided to structure code

1012
00:50:39,040 --> 00:50:41,477
in the way that they have. I mean,
if you play it through, you don't need

1013
00:50:41,560 --> 00:50:45,277
browsers then. Like that's the death
of the browser. Well, so I I would take it

1014
00:50:45,360 --> 00:50:48,437
a step further, which is you
may not need user interfaces.

1015
00:50:48,520 --> 00:50:51,637
So who is going to use
software in the future?

1016
00:50:51,720 --> 00:50:54,237
Other bots. Other bots. Yeah. And so

1017
00:50:54,320 --> 00:50:59,457
>> need to I don't know pipe
information in Do we? and out. Really?

1018
00:50:59,540 --> 00:51:01,197
>> [laughter] >> Well,
what are you going to do then?

1019
00:51:01,280 --> 00:51:02,997
>> Are you sure? You're just
going to log off and touch grass?

1020
00:51:03,080 --> 00:51:04,037
>> Whatever you want, exactly.

1021
00:51:04,120 --> 00:51:06,037
>> [laughter] >> Isn't that better?

1022
00:51:06,120 --> 00:51:07,717
I want software to do stuff for me.

1023
00:51:07,800 --> 00:51:10,037
Isn't that but isn't that better?

1024
00:51:10,120 --> 00:51:12,357
You know, I don't know like like you
know, you know, you know, there are

1025
00:51:12,440 --> 00:51:15,437
arguments here, you know. >> [laughter]
>> It was not that long ago that 99% of

1026
00:51:15,520 --> 00:51:18,197
humanity was behind a plow. Right.
Right. And what are people going

1027
00:51:18,280 --> 00:51:20,877
to do if they're not plowing fields all
day to to grow food, right? And it just

1028
00:51:20,960 --> 00:51:22,997
turns out there's like much better ways
for people to spend time than plowing

1029
00:51:23,080 --> 00:51:25,657
fields. >> Yeah,
do a scroll in. Uh exactly.

1030
00:51:25,740 --> 00:51:27,717
>> [laughter] >> Exactly,
or you know, talking to their

1031
00:51:27,800 --> 00:51:30,837
friends. And look, and I'm not
an absolutist and I'm not a utopian and I

1032
00:51:30,920 --> 00:51:32,957
and to be clear like I have I have
an 11-year-old and he's learning how to

1033
00:51:33,040 --> 00:51:35,397
code and like I'm you know, I I think
it's still like a really good idea to

1034
00:51:35,480 --> 00:51:38,797
learn how to code and so forth. But I
just if you project forward, you just

1035
00:51:38,880 --> 00:51:41,077
have to think forward to a world
in which it's just like okay, I'm just

1036
00:51:41,160 --> 00:51:43,957
going to tell the thing what I need
and it's going to do it. And then and then

1037
00:51:44,040 --> 00:51:47,357
it's going to do it in whatever way
is most optimal for it to do it. Yeah. I

1038
00:51:47,440 --> 00:51:51,037
tell it to do it non-optimally. Like if
I tell it to do it in Java or in Rust or

1039
00:51:51,120 --> 00:51:53,957
whatever, it'll do it, I'm sure. But like
if I'm just going to tell it to do

1040
00:51:54,040 --> 00:51:56,277
it, it's going to do it in whatever way
is like the optimal way to do it. And

1041
00:51:56,360 --> 00:51:58,757
then I And then if I need to understand
how it works, I'm going to ask it to

1042
00:51:58,840 --> 00:52:01,717
explain to me how it works. Right?
And so it's going to be doing its

1043
00:52:01,800 --> 00:52:04,197
own interpret- it's going to be
the engine of interpretability to explain

1044
00:52:04,280 --> 00:52:07,557
itself. And I I just am not convinced
that that I'm not I'm not convinced that

1045
00:52:07,640 --> 00:52:11,637
in that world you have these historical
The goals of the abstractions will be

1046
00:52:11,720 --> 00:52:12,957
whatever the boss needed with the user.

1047
00:52:13,040 --> 00:52:16,557
Yeah. Yeah, that Well, I I'm curious
though, like if that's true, then

1048
00:52:16,640 --> 00:52:20,437
shouldn't the models providers
be building some internal language

1049
00:52:20,520 --> 00:52:24,717
representation that they can
do extreme kind of like RL

1050
00:52:24,800 --> 00:52:27,757
uh and reward modeling around. Because
it's like today they're kind of like

1051
00:52:27,840 --> 00:52:31,837
tied to like TypeScript and Python
because the users need to write in that

1052
00:52:31,920 --> 00:52:35,677
language. Versus they can have their own
thing internally and like they don't

1053
00:52:35,760 --> 00:52:38,797
need to teach it to anybody. They just
need to teach their model. And I think

1054
00:52:38,880 --> 00:52:41,397
that's how you get maybe the version
between the models. Like going back to

1055
00:52:41,480 --> 00:52:45,357
like the Pi open cloud thing, it's like,
"Oh, I built all the software using the

1056
00:52:45,440 --> 00:52:48,677
open AI model, and now switch to the
Anthropic model, but the Anthropic model

1057
00:52:48,760 --> 00:52:52,597
doesn't understand the thing." So I
would It feels like there still needs to

1058
00:52:52,680 --> 00:52:57,317
be some abstraction. But maybe not.
Maybe that's the lock-in

1059
00:52:57,400 --> 00:53:00,237
that the model providers want to have.
I don't I'm not even sure that's lock-in

1060
00:53:00,320 --> 00:53:02,557
though, because why can't the second
model just learn what the first model

1061
00:53:02,640 --> 00:53:05,877
has done? Like Exactly. >> Okay,
so Okay, give me an example. So as

1062
00:53:05,960 --> 00:53:08,317
you know, models can now reverse
engineer software binaries. This whole

1063
00:53:08,400 --> 00:53:10,597
thing now where people are reverse
engineering like Nintendo Nintendo game

1064
00:53:10,680 --> 00:53:13,997
binaries. Yeah. So you have like You've
You've seen a bunch of reports like this

1065
00:53:14,080 --> 00:53:16,917
where somebody has like a favorite game
from the 1980s. And the source code is

1066
00:53:17,000 --> 00:53:19,397
like long dead, but they have like
a binary burned into a chip or something,

1067
00:53:19,480 --> 00:53:21,677
and now they're reverse engineering
to get a version that runs on their Mac,

1068
00:53:21,760 --> 00:53:24,037
right? And so if you reverse it if
This is what I kind of say. If you're

1069
00:53:24,120 --> 00:53:28,357
reversing like x86 binaries,
then why can't you reverse engineer

1070
00:53:28,440 --> 00:53:32,037
whatever they create. Yeah. And because
we're on a Unix-based system, it has to

1071
00:53:32,120 --> 00:53:36,077
be reversible because it needs to run
on the target. Yeah. Yeah, yeah. Yeah,

1072
00:53:36,160 --> 00:53:39,277
yeah, basically. And so I just I just
think it's this thing where it's just like

1073
00:53:39,360 --> 00:53:41,477
And by the way, everything we're
describing is something that human

1074
00:53:41,560 --> 00:53:44,317
beings in theory could have
done before, but just with

1075
00:53:44,400 --> 00:53:47,917
>> Right, yeah. but with enormous
but it was just always like cost and

1076
00:53:48,000 --> 00:53:51,477
labor prohibitive. Reverse engineer I
learned how to reverse engineer was like

1077
00:53:51,560 --> 00:53:54,357
human beings can reverse engineer
binaries. It's just for any complex

1078
00:53:54,440 --> 00:53:57,717
binary you need like a thousand years
to do it. But now with the model you

1079
00:53:57,800 --> 00:54:00,957
don't. And so all of a sudden you get
you get these things or another way to

1080
00:54:01,040 --> 00:54:04,557
think about it is so much of human built
systems are to compensate for the human

1081
00:54:04,640 --> 00:54:07,077
limitations. Right?
And if you don't have the human

1082
00:54:07,160 --> 00:54:09,917
limitations anymore then
all of a sudden you have

1083
00:54:10,000 --> 00:54:13,037
won't have abstractions but you'll
have a different kind of abstraction.

1084
00:54:13,120 --> 00:54:16,557
I have two topics to bring us to a close
and you can pick whichever ones. I just

1085
00:54:16,640 --> 00:54:20,197
talking about protocols.
Was it you or someone else I forget my

1086
00:54:20,280 --> 00:54:22,717
internet history who said that like
the biggest mistake that we didn't figure

1087
00:54:22,800 --> 00:54:26,717
out in the early days was payments.
Was that you? Yes, it was a 402

1088
00:54:26,800 --> 00:54:29,477
402 payment required.
We have a chance now.

1089
00:54:29,560 --> 00:54:31,837
I don't think we're going to figure it
out. I don't know. Like what's your take?

1090
00:54:31,920 --> 00:54:34,237
>> we will. Yeah. No, now I think it's
going to happen for sure. Yeah. Yeah,

1091
00:54:34,320 --> 00:54:36,277
and there's two reasons it's going to
happen for sure. One is we actually have

1092
00:54:36,360 --> 00:54:39,397
internet native money now in the form of
crypto stable coins and crypto. And this

1093
00:54:39,480 --> 00:54:42,877
is I think this is the grand unification
basically of AI and crypto is what's

1094
00:54:42,960 --> 00:54:45,717
about to happen now.
I think AI is the crypto killer app. I

1095
00:54:45,800 --> 00:54:48,477
think it's where where this is really
going to come out. And then the other is

1096
00:54:48,560 --> 00:54:51,477
it's just it I mean it's just I think
it's now obvious. It's like obviously AI

1097
00:54:51,560 --> 00:54:53,037
agents are going to need money.

1098
00:54:53,120 --> 00:54:55,397
And it's already happening, right?
If you've got a if you've got a claw and

1099
00:54:55,480 --> 00:54:58,557
you want it to buy things for you,
you have to give it money in some form. I

1100
00:54:58,640 --> 00:55:01,837
would say the adoption is probably
like 0.1% if if that, but yeah.

1101
00:55:01,920 --> 00:55:04,917
>> Oh, today? Yeah, yeah, yeah,
but think forward. Like where is it going?

1102
00:55:05,000 --> 00:55:07,877
Forward thinking. The ultimate principle
of everything and and everything that I

1103
00:55:07,960 --> 00:55:10,757
think I we do is it's the William Gibson
quote which is the future is already

1104
00:55:10,840 --> 00:55:15,117
here it just isn't distributed. It isn't
distributed yet. My friends who are the

1105
00:55:15,200 --> 00:55:19,557
most aggressive users of open claw just
like have given their claws bank accounts

1106
00:55:19,640 --> 00:55:23,237
and credit cards. And and and and not
only have they done it, it's obvious

1107
00:55:23,320 --> 00:55:25,677
that they needed to do it because it's
obvious that they needed to be able to

1108
00:55:25,760 --> 00:55:27,117
spend money on their behalf. Yeah, yeah.

1109
00:55:27,200 --> 00:55:29,637
It's just completely obvious. And so
and again like so the number of people who

1110
00:55:29,720 --> 00:55:31,997
have done that today to your point
is like I don't know probably 5,000 or

1111
00:55:32,080 --> 00:55:35,397
something. But it'll grow. That's how
these things start. Actually, I mean

1112
00:55:35,480 --> 00:55:38,157
since you keep mentioning And by the way
open claw by the way if you don't give

1113
00:55:38,240 --> 00:55:40,457
it a bank account, it's just going
to break into your your bank account.

1114
00:55:40,540 --> 00:55:41,917
>> [laughter] >> It's going
to break It's going to break

1115
00:55:42,000 --> 00:55:44,637
It's going to break into your bank
account anyway and take your money. So,

1116
00:55:44,720 --> 00:55:47,557
you you might as you might as well do it.
You might as well do it. Uh yeah. By

1117
00:55:47,640 --> 00:55:51,117
the way, I really love I got to tell you
I really love the phenomenon. I love the

1118
00:55:51,200 --> 00:55:54,637
YOLO. Um I'm not doing it myself, to be
clear, but but I love the people that

1119
00:55:54,720 --> 00:55:58,317
are just like you What is it?
Skip Skip permissions?

1120
00:55:58,400 --> 00:55:59,557
By the way, it's a Facebook thing.

1121
00:55:59,640 --> 00:56:02,677
>> Okay. >> [laughter] >> Cuz
cuz we In Facebook, they they have

1122
00:56:02,760 --> 00:56:06,917
this culture to name the thing dangerous
so that you are aware when you enable

1123
00:56:07,000 --> 00:56:09,237
the flag that you're opting into
a dangerous thing. Okay, good. They

1124
00:56:09,320 --> 00:56:13,997
brought it into OpenClaw AI. And of course
that makes it enticing. Sam runs Codex

1125
00:56:14,080 --> 00:56:15,837
with skip permissions on on his laptop.

1126
00:56:15,920 --> 00:56:19,517
Yes, 100%. And so, I I I think the way
to actually see the future is to find

1127
00:56:19,600 --> 00:56:21,067
the people who are doing that.

1128
00:56:21,150 --> 00:56:22,837
>> [laughter] >> There's
a man who you know, they're you

1129
00:56:22,920 --> 00:56:25,917
know Log everything, you know,
just watch it watch the logs.

1130
00:56:26,000 --> 00:56:28,957
>> But like, let's actually find out
what the thing can do. And the way to find

1131
00:56:29,040 --> 00:56:32,277
out what the thing can do is just like
Yeah, let it try everything. Let it

1132
00:56:32,360 --> 00:56:34,557
unlock everything. By the way, that's
how you're going to find all the good

1133
00:56:34,640 --> 00:56:36,877
stuff it can do. By the way, that's also
how you're going to find all the flaws.

1134
00:56:36,960 --> 00:56:39,637
Yeah. I think the people who turn
that on for bots are like they're like

1135
00:56:39,720 --> 00:56:42,637
murders to the progress of human
civilization. Like I feel very bad for

1136
00:56:42,720 --> 00:56:44,597
their descendants that their bank
accounts are going to get looted by

1137
00:56:44,680 --> 00:56:47,717
their bots in the first like 20 minutes,
but I think the contribution that

1138
00:56:47,800 --> 00:56:49,677
they're making to the future
of our species is amazing.

1139
00:56:49,760 --> 00:56:51,037
>> It's like gentleman science, you know.

1140
00:56:51,120 --> 00:56:54,917
Yes, it's Yes, yes. It's a Ben Franklin
out with a trying to trying trying to

1141
00:56:55,000 --> 00:56:57,837
get lightning strike his [laughter] his
his balloon and see seeing if he gets

1142
00:56:57,920 --> 00:57:01,017
electrocuted. Yeah. It It's
a Jonas Salk with the polio vaccine.

1143
00:57:01,100 --> 00:57:04,117
>> [laughter] >> Injecting it Yes.
So, yes, I I I I think

1144
00:57:04,200 --> 00:57:07,637
we should have like a glory we should
have like flags and like we should have

1145
00:57:07,720 --> 00:57:10,837
like monuments to [laughter] the people
who just let OpenClaw run their lives.

1146
00:57:10,920 --> 00:57:14,197
More anecdotes. What what are
the craziest or interesting things that

1147
00:57:14,280 --> 00:57:18,277
people listening to this should go go
home and do? I mean, this is this is the

1148
00:57:18,360 --> 00:57:21,117
this is the the extreme thing is just
like the straight YOLO. Like just yeah,

1149
00:57:21,200 --> 00:57:23,077
turn turn your life >> But that's
a general capability. Is

1150
00:57:23,160 --> 00:57:26,917
there like a specific story that was
like wow and and everyone in the group

1151
00:57:27,000 --> 00:57:29,797
chat just lit up? I mean, like you know,
so there's tons of there's already tons

1152
00:57:29,880 --> 00:57:32,677
of health. You know, there's you know,
that the health dashboard stuff is just

1153
00:57:32,760 --> 00:57:34,437
is just absolutely absolutely amazing.

1154
00:57:34,520 --> 00:57:37,277
>> Yeah. The number of stories on I'm
trying I just don't want to violate

1155
00:57:37,360 --> 00:57:40,477
people's, you know,
obviously personal anonymities, but

1156
00:57:40,560 --> 00:57:42,837
you know, one of the things open claws
are really good at is hacking into all

1157
00:57:42,920 --> 00:57:45,877
the stuff in your LAN.
It's really good at it. So, you know,

1158
00:57:45,960 --> 00:57:48,797
internet of things aka internet of

1159
00:57:48,880 --> 00:57:50,917
Like super insecure, but great.

1160
00:57:51,000 --> 00:57:53,517
Discoverable. Yeah, it's a discoverable.

1161
00:57:53,600 --> 00:57:56,637
Open claw is happy to scan your network,
identify all the things and then my my

1162
00:57:56,720 --> 00:57:58,997
my friends who are most aggressive
at this are having open claw take over

1163
00:57:59,080 --> 00:58:01,837
everything in their house. Yeah.
It takes over their security cameras, it

1164
00:58:01,920 --> 00:58:04,237
takes over their their, you know,
their whatever their their access control

1165
00:58:04,320 --> 00:58:07,517
systems, it takes over their webcams.
I have a friend whose claw watches him

1166
00:58:07,600 --> 00:58:11,037
sleep. Put a webcam in your bedroom, put
the put the claw put the claw on a loop.

1167
00:58:11,120 --> 00:58:13,837
I have it wake up frequently and have
it watch it just tell it watch me sleep.

1168
00:58:13,920 --> 00:58:17,237
And and I've I've seen the transcripts
and it's literally like Joe's asleep.

1169
00:58:17,320 --> 00:58:19,837
This is good. This is good that Joe's
asleep cuz, you know, I have I have his

1170
00:58:19,920 --> 00:58:22,157
health data and I know that he hasn't
been getting enough sleep and so it's

1171
00:58:22,240 --> 00:58:24,477
really good that he's getting sleep.
I really hope he gets his full whatever,

1172
00:58:24,560 --> 00:58:27,117
you know, 5 hours of REM sleep.

1173
00:58:27,200 --> 00:58:30,437
Uh Joe's moving. Joe's moving.

1174
00:58:30,520 --> 00:58:33,157
Uh Joe might be waking waking up. This
is a real pri If Joe wakes up now, he's

1175
00:58:33,240 --> 00:58:34,717
going to ruin his sleep cycle.

1176
00:58:34,800 --> 00:58:37,957
Oh, okay, it's okay. Joe just rolled over.
Okay, he's gone back to bed. Okay,

1177
00:58:38,040 --> 00:58:40,117
good. All right. Okay,
I can relax. This is fine.

1178
00:58:40,200 --> 00:58:41,887
>> He's monitoring the situation.

1179
00:58:41,970 --> 00:58:44,237
>> [laughter] >> Monitoring
the situation. And and being

1180
00:58:44,320 --> 00:58:47,437
a bot like, you know, it's just
like very focused. [laughter]

1181
00:58:47,520 --> 00:58:50,637
It's just like this is like
his reason for existence

1182
00:58:50,720 --> 00:58:53,717
is to watch Joe sleep. And then and then
I was telling my friend who did this is

1183
00:58:53,800 --> 00:58:55,597
like, you know, on the one hand it's
like, all right, this is weird and

1184
00:58:55,680 --> 00:58:58,557
creepy. And I need to I need to I maybe
this is taking over my life. And then

1185
00:58:58,640 --> 00:58:59,517
the other thing is like, you know, what?

1186
00:58:59,600 --> 00:59:02,037
If I had a heart attack in the middle
of the night, literally would like freak

1187
00:59:02,120 --> 00:59:05,077
out and call 911. Like there's no
question this thing would figure out how

1188
00:59:05,160 --> 00:59:08,517
to like alert medical authorities and like
probably summon SWAT teams and like

1189
00:59:08,600 --> 00:59:10,997
do whatever would be required
[laughter] to save my life, right?

1190
00:59:11,080 --> 00:59:15,717
And so it's like, you know, like yeah,
like that's happening. What else? Um

1191
00:59:15,800 --> 00:59:18,717
They give us this company
Unitree that makes the

1192
00:59:18,800 --> 00:59:22,437
robot dogs. And I actually
have one at home which is

1193
00:59:22,520 --> 00:59:25,437
is it's actually really
the Chinese companies.

1194
00:59:25,520 --> 00:59:27,957
The Chinese companies are
so aggressive at adopting

1195
00:59:28,040 --> 00:59:32,517
uh new technology, but they don't always
like take the time to really package it.

1196
00:59:32,600 --> 00:59:35,517
>> Package it and maybe think it all the
way through. And so so the at least the

1197
00:59:35,600 --> 00:59:40,157
Unitree dog I have so it has a old
non-LLM just control system. Which by

1198
00:59:40,240 --> 00:59:42,997
the way is not very good. Markets well,
but it in practice is it's not that

1199
00:59:43,080 --> 00:59:44,997
good. It has trouble with stairs and so
forth. And so it's not quite what it

1200
00:59:45,080 --> 00:59:47,517
should be. But then the language model
thing comes out and the voice. So they

1201
00:59:47,600 --> 00:59:51,077
they add so they add LLM capability
and then they they add a voice mode to it.

1202
00:59:51,160 --> 00:59:55,257
Um but but that LLM capability is not
at all connected to the control system.

1203
00:59:55,340 --> 00:59:58,357
>> [laughter] >> So so you've
got this schizophrenic dog

1204
00:59:58,440 --> 01:00:00,637
that like is a complete idiot when it
comes to climbing the stairs, but it

1205
01:00:00,720 --> 01:00:02,457
will happily teach you quantum mechanics.

1206
01:00:02,540 --> 01:00:05,677
>> [laughter] >> Right.
And like a plum English accent, right?

1207
01:00:05,760 --> 01:00:06,837
It's just like absolutely amazing.

1208
01:00:06,920 --> 01:00:10,117
>> Jagged intelligence. Yeah.
Talk about jagged. And then now

1209
01:00:10,200 --> 01:00:12,237
obviously what's going to happen in the
future is is they're going to connect

1210
01:00:12,320 --> 01:00:13,917
together, but >> They'll do it.
but right now it's it's

1211
01:00:14,000 --> 01:00:15,077
and so right now it's not that useful.

1212
01:00:15,160 --> 01:00:18,037
And so I I have a friend who has one of
these who had his claw basically hack in

1213
01:00:18,120 --> 01:00:20,117
and rewrite the code rewrite
the write new firmware.

1214
01:00:20,200 --> 01:00:22,877
>> Yeah. Rewrite new firmware
for the for the Unitree robot.

1215
01:00:22,960 --> 01:00:27,037
And now it's now it's an actual dog
for his kids. You can do that before and

1216
01:00:27,120 --> 01:00:29,677
after like the motion? Yeah. It's He
said it's completely different. He said

1217
01:00:29,760 --> 01:00:32,077
it's a complete transformation.
And whenever there's an issue in the thing

1218
01:00:32,160 --> 01:00:35,157
now the claw just like rewrites the code.
You know you know you go you does

1219
01:00:35,240 --> 01:00:37,957
does the code and so it's a good kind
of goes to your thing here. So so like all

1220
01:00:38,040 --> 01:00:40,917
of a sudden this is why like we're going
to think about AI code. AI coding is not

1221
01:00:41,000 --> 01:00:43,477
just like writing new apps. It's also
going in and rewriting all the old stuff

1222
01:00:43,560 --> 01:00:46,437
that should have worked that never worked.
And so like I I think I think

1223
01:00:46,520 --> 01:00:49,117
basically I think the internet
the internet of is basically over. Like

1224
01:00:49,200 --> 01:00:51,437
I I think everything there's
a potentially here where like all these

1225
01:00:51,520 --> 01:00:53,837
devices in your house that have been
like basically marginal or you know

1226
01:00:53,920 --> 01:00:56,717
basically dumb, you know,
like all of a sudden they might all get

1227
01:00:56,800 --> 01:01:01,397
really smart. Now Smart home. you
have to decide if yes, there are horror

1228
01:01:01,480 --> 01:01:05,437
movies in which this is of which this
is the premise. And so you have to decide

1229
01:01:05,520 --> 01:01:08,117
if you want this. >> Yeah.
But but but this is the first time

1230
01:01:08,200 --> 01:01:11,437
I can say with confidence I now know
how you could actually have a smart home

1231
01:01:11,520 --> 01:01:14,037
Yeah. with 30 different kinds of things
with chips and internet access where it

1232
01:01:14,120 --> 01:01:16,477
actually all makes sense and all works
together and it's all coherent and the

1233
01:01:16,560 --> 01:01:19,677
and the whole thing. And to have that
unlock without a human being having to

1234
01:01:19,760 --> 01:01:24,437
go do do of that work like Yeah.
So, it's fine. I'm waiting for a sorry,

1235
01:01:24,520 --> 01:01:27,607
Mark. Uh I can't let you open that
fridge door. You know, like Exactly.

1236
01:01:27,690 --> 01:01:31,117
[laughter] Exactly. Yes. YES. >> BECAUSE
YOU'RE NOT SUPPOSED TO EAT RIGHT NOW.

1237
01:01:31,200 --> 01:01:34,157
>> all of you. Yes. I have every shred of
health information. You know, and I know

1238
01:01:34,240 --> 01:01:35,957
you think you're doing, you know,
da da da. And I know you think you can do

1239
01:01:36,040 --> 01:01:38,037
this, but you know, this is a real Are
you really, [laughter] you know, are you

1240
01:01:38,120 --> 01:01:40,717
really sure? And you know,
you told, you know, you

1241
01:01:40,800 --> 01:01:42,997
told me last night you really don't want
me to let you do this. And so, you know,

1242
01:01:43,080 --> 01:01:44,637
I'm sorry, but the fridge door is locked.

1243
01:01:44,720 --> 01:01:46,557
>> [laughter] >> Um yes.
>> Open the fridge doors.

1244
01:01:46,640 --> 01:01:49,117
>> Exactly. And by the way, I know you're
supposed to be studying for a test. So,

1245
01:01:49,200 --> 01:01:52,277
why don't we why don't you go when you
can pass the test, um I will open the

1246
01:01:52,360 --> 01:01:54,997
fridge door for you.
Final photo call and then and then we

1247
01:01:55,080 --> 01:01:57,037
can wrap up. Uh proof of human. Yes.

1248
01:01:57,120 --> 01:01:59,597
>> right? That's the last
piece that we got to figure out.

1249
01:01:59,680 --> 01:02:04,477
>> Yeah. So, I would say there's there's
two massive, I would say, um uh sort of

1250
01:02:04,560 --> 01:02:08,277
asymmetries in the world right now where
we've known these asymmetries exist and

1251
01:02:08,360 --> 01:02:11,157
we we societally have been unwilling
to grapple with them. And I think they're

1252
01:02:11,240 --> 01:02:14,037
both tipping right now. And And they're
they're they're they're the same thing.

1253
01:02:14,120 --> 01:02:16,197
It's virtual world version. It's
physical world version. So, the virtual

1254
01:02:16,280 --> 01:02:18,237
world version is is the bot problem.

1255
01:02:18,320 --> 01:02:20,997
Where just like, you know, the internet
internet is just like a washing bots.

1256
01:02:21,080 --> 01:02:23,757
The internet's a washing fake people.
It has been forever. Um by the way, a lot

1257
01:02:23,840 --> 01:02:26,277
of that has to do with lack of money,
you know, and so this you know, this is

1258
01:02:26,360 --> 01:02:28,597
the this is the this is the My spicy
take was these two are the same thing

1259
01:02:28,680 --> 01:02:31,997
and corporations are people,
too. You know, >> [laughter]

1260
01:02:32,080 --> 01:02:33,237
>> Interesting. Yeah. Yeah. Yeah. Okay.

1261
01:02:33,320 --> 01:02:34,837
>> So, a bank account is proof of human.

1262
01:02:34,920 --> 01:02:37,557
Yeah. Okay. >> [laughter] >> Yeah,
until you until you give the bots

1263
01:02:37,640 --> 01:02:39,397
bank accounts. Yeah, exactly. So, okay.

1264
01:02:39,480 --> 01:02:42,797
Yeah, so there's that. But yeah, look
Look, the bot I mean, every social media

1265
01:02:42,880 --> 01:02:45,597
user knows this. The bot The bot problem
is a big problem. You know, the bot The

1266
01:02:45,680 --> 01:02:47,837
bot problem has been a big problem
forever. It's It's a huge problem. And

1267
01:02:47,920 --> 01:02:51,917
it's never really been confronted
directly. Like, at any point. By the way,

1268
01:02:52,000 --> 01:02:55,277
the physical world version of this is
the drone the drone problem. Um right?

1269
01:02:55,360 --> 01:02:57,997
And so, we we've known for, you know,
we've known for 20 years now that the

1270
01:02:58,080 --> 01:03:01,837
asymmetric threat both in military
in actual military conflict, but also in

1271
01:03:01,920 --> 01:03:05,037
just like security, like like, you know,
security on the home front, the big

1272
01:03:05,120 --> 01:03:06,477
threat is is the cheap attack drone.

1273
01:03:06,560 --> 01:03:09,597
Right? The the the cheap the cheap
suicide, you know, drone with a bomb.

1274
01:03:09,680 --> 01:03:12,157
And we've known that forever.
And by the way, like, you know, it's very

1275
01:03:12,240 --> 01:03:15,197
disconcerting how like every, you know,
every office complex in the in the cut,

1276
01:03:15,280 --> 01:03:18,917
know, in the world is like unprotected
from drone attacks. Every every stadium,

1277
01:03:19,000 --> 01:03:22,477
every school, every prison
like is like >> Sure.

1278
01:03:22,560 --> 01:03:25,117
Okay, we've known that. We've never done
anything about it. What are we going to

1279
01:03:25,200 --> 01:03:27,277
do about it? Well, one possibility
is just leave leave them unprotected

1280
01:03:27,360 --> 01:03:30,157
forever and live in a world of like
asymmetric terrorism forever. The other

1281
01:03:30,240 --> 01:03:32,757
is take the problem seriously
and figure out the set of techniques and

1282
01:03:32,840 --> 01:03:35,477
technologies required to to be able to
deal with that, whether those are lasers

1283
01:03:35,560 --> 01:03:39,157
or jammers or early warning systems or,
you know, don't Personal force fields.

1284
01:03:39,240 --> 01:03:41,517
>> Kinetic personal For Dune,
>> [laughter] >> uh, for

1285
01:03:41,600 --> 01:03:45,317
For personal force fields. Exactly.
And in both cases, the these are these are

1286
01:03:45,400 --> 01:03:47,917
economic asymmetries. These are economic
asymmetries, right? Cuz it's really

1287
01:03:48,000 --> 01:03:51,037
cheap to field a bot, but it's very hard
to tell if something a bot. It's very

1288
01:03:51,120 --> 01:03:53,797
cheap to field a drone. It's very hard
It's very expensive to defend against a

1289
01:03:53,880 --> 01:03:56,557
drone. But you see what I'm saying
is it's it's it's the it's the virtual

1290
01:03:56,640 --> 01:03:59,077
version of the problem and it's the
physical version of the problem. Uh, the

1291
01:03:59,160 --> 01:04:01,597
virtual version of the problem what we
what we need quite literally is proof of

1292
01:04:01,680 --> 01:04:03,557
human. The reason is because you're
you're not you're not going to have

1293
01:04:03,640 --> 01:04:06,837
proof of bot. The the the especially now
that the bots are too good. The the the

1294
01:04:06,920 --> 01:04:09,197
the bots can pass the Turing test.
And if the bots can pass the Turing test,

1295
01:04:09,280 --> 01:04:11,197
then you can't you can't screen for bot.

1296
01:04:11,280 --> 01:04:14,637
You can't have proof of not a bot. But
what you can have is you can have proof

1297
01:04:14,720 --> 01:04:17,157
of human. You can have, you know,
cryptographically validated this is

1298
01:04:17,240 --> 01:04:19,757
definitely a person. And this is And
then you can have cryptographically

1299
01:04:19,840 --> 01:04:22,837
validated this is definitely like
something that a person said. This video

1300
01:04:22,920 --> 01:04:27,397
is real. Right? Just to double click
on on, uh, do you think Alex Blania with

1301
01:04:27,480 --> 01:04:30,677
World, do you think he's got it or is
there alternative? Oh, so I mean there's

1302
01:04:30,760 --> 01:04:32,997
going to be I think there'll be I mean
many people will try. We're one of the

1303
01:04:33,080 --> 01:04:35,797
key, you know, participants in in the
World in the World project and I didn't

1304
01:04:35,880 --> 01:04:38,157
know that. I didn't know that. Yeah,
so we're we're partisans. But yeah, I I

1305
01:04:38,240 --> 01:04:40,237
think So we think World
is exactly correct.

1306
01:04:40,320 --> 01:04:42,237
>> Okay. And the reason
is it has it has to

1307
01:04:42,320 --> 01:04:45,797
be it it has to be proof of human. It
it has because you can't do proof of not

1308
01:04:45,880 --> 01:04:48,837
bot. You have to do proof of human. To do
proof of human, you you need you need

1309
01:04:48,920 --> 01:04:51,957
biological validation. You you need
it to start with this was actually a

1310
01:04:52,040 --> 01:04:55,117
person. Right? Cuz otherwise you have
bots signing up as fake people. All

1311
01:04:55,200 --> 01:04:57,557
right? So you you have to have like
something You have to have a bio

1312
01:04:57,640 --> 01:05:00,357
biometric. And then you have to have
cryptographic validation and then the

1313
01:05:00,440 --> 01:05:02,517
ability to do to do to do the lookup.

1314
01:05:02,600 --> 01:05:04,557
And then by the way, the other thing
you need which is that you you also need

1315
01:05:04,640 --> 01:05:07,957
selective disclosure. Um, so you need
to be able to do proof of human without

1316
01:05:08,040 --> 01:05:10,237
revealing all the underlying information.
By the way, another thing

1317
01:05:10,320 --> 01:05:12,557
you're going to need you're going
to need proof of age, right? Cuz there's

1318
01:05:12,640 --> 01:05:15,117
all these laws in all these different
countries now around you need to be 13

1319
01:05:15,200 --> 01:05:17,477
or 16 or 18 or whatever to do different
things. And so you're going to you're

1320
01:05:17,560 --> 01:05:20,557
going to need a you know, sort of validate
a proof of age. Um you know, to

1321
01:05:20,640 --> 01:05:22,957
be able to legally operate, right?
And so that that's coming and then you're

1322
01:05:23,040 --> 01:05:25,997
going to want like proof of credit score
and you know, proof of like, you know,

1323
01:05:26,080 --> 01:05:28,277
100 other >> That's a tricky one.
It is a tricky one,

1324
01:05:28,360 --> 01:05:30,797
but you're going to you're going to need
there's no reason like if somebody's

1325
01:05:30,880 --> 01:05:33,237
checking on your credit somebody
shouldn't I'll put this I'll give you an

1326
01:05:33,320 --> 01:05:35,677
example. Somebody shouldn't need to know
your name in order to be able to find

1327
01:05:35,760 --> 01:05:36,797
out whether you're credit worthy, right?

1328
01:05:36,880 --> 01:05:39,517
I see. Independently verifiable
pieces of information.

1329
01:05:39,600 --> 01:05:42,197
>> Pieces of information. So like so like
disclosed and this is the answer to the

1330
01:05:42,280 --> 01:05:45,317
privacy problem writ large, which is I
I only need to prove I need to prove at

1331
01:05:45,400 --> 01:05:47,237
that moment. >> So like
you're going to need that and I

1332
01:05:47,320 --> 01:05:50,277
think that their architecture makes sense.
So that needs to get solved. I

1333
01:05:50,360 --> 01:05:54,517
think language models have tipped
the bots are now too good. Uh and and so

1334
01:05:54,600 --> 01:05:56,917
they're undetectable and so as a
consequence you we now need to go and

1335
01:05:57,000 --> 01:05:59,117
confront that problem directly. And then
and like I said and then the other

1336
01:05:59,200 --> 01:06:01,317
problem is we we need to go
actually confront the drone problem.

1337
01:06:01,400 --> 01:06:04,077
>> Business. The Ukraine
conflict has really unlocked

1338
01:06:04,160 --> 01:06:07,557
a lot of thinking on that. And now
the um and now the the the the the Iran

1339
01:06:07,640 --> 01:06:09,957
situation is also unlocking that.
And so I think there's going to be

1340
01:06:10,040 --> 01:06:12,357
>> just like this incredible explosion
of of both drone and counter drone. Our

1341
01:06:12,440 --> 01:06:13,847
drones are better than their guns.

1342
01:06:13,930 --> 01:06:16,357
>> [laughter] >> It's
also keep it that way.

1343
01:06:16,440 --> 01:06:20,477
And counter and counter drones. I think
we can sneak in one more question. Um

1344
01:06:20,560 --> 01:06:23,397
I'm trying to tie together a lot
of things that you said over the years. So

1345
01:06:23,480 --> 01:06:26,957
at the Milken Institute debate with Thiel,
which is amazing, um you talked

1346
01:06:27,040 --> 01:06:31,117
about the lag between a new technology
and kind of like the GDP um impact of

1347
01:06:31,200 --> 01:06:33,997
it. The other idea you talked about is

1348
01:06:34,080 --> 01:06:38,277
bourgeois capitalism and how you know,
this kind of managerial class was needed

1349
01:06:38,360 --> 01:06:41,637
because of this complexity and I think
if you bring AI into the fold you have

1350
01:06:41,720 --> 01:06:44,357
like much higher leverage of people. So

1351
01:06:44,440 --> 01:06:48,877
like if you have, you know, the Musk
industries um and you give Elon a GI you

1352
01:06:48,960 --> 01:06:50,957
can run a lot more things uh at once.

1353
01:06:51,040 --> 01:06:54,277
That's right. And then you have the social
contract and I know you retweeted

1354
01:06:54,360 --> 01:06:57,397
a clip of Sam Altman saying um we're
rethinking the whole thing and you're

1355
01:06:57,480 --> 01:07:00,437
like absolutely not. >> Yes.
Under Uh and I was I was at that

1356
01:07:00,520 --> 01:07:03,637
event with Sam last night uh and he
actually said in the last couple weeks

1357
01:07:03,720 --> 01:07:07,077
he felt like now people are taking that
seriously. So I'm just curious like how

1358
01:07:07,160 --> 01:07:09,917
you're seeing the structure
of organization changing,

1359
01:07:10,000 --> 01:07:13,197
especially when you invest in early
stage companies, and um yeah, just like

1360
01:07:13,280 --> 01:07:16,797
how the impact of work structure
and all of that is playing out.

1361
01:07:16,880 --> 01:07:19,317
>> Yeah. So, there's a whole bunch
of there's a whole bunch of time

1362
01:07:19,400 --> 01:07:20,997
we [laughter] could spend I don't know
by the way we'd be happy to spend more

1363
01:07:21,080 --> 01:07:23,997
time, but we could we could spend more
time on all that. So, just it for people

1364
01:07:24,080 --> 01:07:26,757
who haven't followed this. So, that
this this this term managerial comes from

1365
01:07:26,840 --> 01:07:29,957
this thinker in the 20th century James
Burnham who um is one of the great kind

1366
01:07:30,040 --> 01:07:32,917
of 20th century political thinkers uh
societal thinkers. And he sort of said

1367
01:07:33,000 --> 01:07:35,797
as and he was writing
in like 1940s, 1950s.

1368
01:07:35,880 --> 01:07:38,597
Um and he said kind of that the whole
history of capitalism until that point

1369
01:07:38,680 --> 01:07:41,397
had been in two phases. Number one had
been what he called bourgeois capitalism

1370
01:07:41,480 --> 01:07:45,077
which was think about it as like name
on the door. Like Ford Motor Company cuz

1371
01:07:45,160 --> 01:07:48,197
Henry Ford runs the company. Um
and Henry it's like a dictate dictatorial

1372
01:07:48,280 --> 01:07:50,797
model and Henry Ford just like
tells everybody what to do.

1373
01:07:50,880 --> 01:07:53,677
And he said the problem with bourgeois
capitalism is it doesn't scale cuz Henry

1374
01:07:53,760 --> 01:07:56,277
Ford can only tell so many people to do
so many things and then he runs out of

1375
01:07:56,360 --> 01:07:59,997
time in the day. And so, um he said
the second phase of capitalism was what he

1376
01:08:00,080 --> 01:08:02,957
called managerial capitalism which was
the creation of a professional class of

1377
01:08:03,040 --> 01:08:06,477
managers. Um that are
trained not to be like car

1378
01:08:06,560 --> 01:08:09,877
experts or to be whatever experts in any
particular field, but are trained to be

1379
01:08:09,960 --> 01:08:12,917
experts in management. And then that led
to, you know, the importance of like

1380
01:08:13,000 --> 01:08:15,517
Harvard Business, you know, business
schools and management consulting firms

1381
01:08:15,600 --> 01:08:19,157
and all these things. And then you look
at every big company today and like most

1382
01:08:19,240 --> 01:08:22,397
of the executives at most of the Fortune
500 companies are not domain experts in

1383
01:08:22,480 --> 01:08:24,917
whatever the company does and they're
certainly not the founders of those

1384
01:08:25,000 --> 01:08:27,997
companies, but they're professional
managers. And in fact, in the course of

1385
01:08:28,080 --> 01:08:30,837
their careers they'll probably manage
many different kinds of businesses.

1386
01:08:30,920 --> 01:08:33,077
They'll rotate around and they might
work in health care for a while and then

1387
01:08:33,160 --> 01:08:35,357
work in financial services and then go
work in something else, you know, come

1388
01:08:35,440 --> 01:08:37,837
work in tech. And what
Burnham said is he said that

1389
01:08:37,920 --> 01:08:41,077
transition is absolutely required
because the the the problem with

1390
01:08:41,160 --> 01:08:44,797
bourgeois capitalism is is it doesn't
scale. Henry Ford doesn't scale. And so,

1391
01:08:44,880 --> 01:08:47,517
if you're going to run capitalist
enterprises that are going to have

1392
01:08:47,600 --> 01:08:50,637
millions to billions of customers,
um you're going to need to you they're

1393
01:08:50,720 --> 01:08:53,037
going to be operating a level of scale
and complexity that's going to require

1394
01:08:53,120 --> 01:08:54,757
this professional management class.

1395
01:08:54,840 --> 01:08:56,757
And he said, "Look, that the
professional management class has its

1396
01:08:56,840 --> 01:08:59,557
downsides. Like they're not necessarily
experts at doing the thing. They're not

1397
01:08:59,640 --> 01:09:02,757
as inventive, you know, they're not going
to create the next breakthrough thing."

1398
01:09:02,840 --> 01:09:04,757
But he's like whether you think that's
good or bad or whatever it's what's

1399
01:09:04,840 --> 01:09:07,677
going to be required. And basically
that's what happened, right? And so he

1400
01:09:07,760 --> 01:09:09,837
wrote that book originally like 1940.

1401
01:09:09,920 --> 01:09:12,877
You know, over the course of the next
50 years basically managerialism Well, I

1402
01:09:12,960 --> 01:09:15,917
mean, today up till today managerial
managerialism basically took over

1403
01:09:16,000 --> 01:09:18,357
everything and you know, what I'm
describing is basically how all big

1404
01:09:18,440 --> 01:09:21,597
companies run and how all governments
run and how all large-scale nonprofits

1405
01:09:21,680 --> 01:09:24,117
run and kind of everything
you know, everything runs.

1406
01:09:24,200 --> 01:09:28,357
Basically what what what venture capital
does is we basically are a rump sort of

1407
01:09:28,440 --> 01:09:32,157
protest movement to that to try to find
the next Henry Ford or or just to say

1408
01:09:32,240 --> 01:09:35,197
Elon Musk or or the next or the next
Elon Musk or the next Steve Jobs or the

1409
01:09:35,280 --> 01:09:37,877
next Bill Gates or the next Mark
Zuckerberg. And so we we we we start

1410
01:09:37,960 --> 01:09:39,917
these companies in the old model, right?

1411
01:09:40,000 --> 01:09:44,157
We we we start them out as as as as like
in the Henry Ford model. And so we start

1412
01:09:44,240 --> 01:09:47,037
them out with a founder or a or a or
a founder with with colleagues, but you

1413
01:09:47,120 --> 01:09:48,957
know, there's there's a founder CEO.

1414
01:09:49,040 --> 01:09:52,677
Um and then we basically bet that we
basically bet that the startup is going

1415
01:09:52,760 --> 01:09:56,117
to be able to do things specifically
innovate in ways that the big incumbents

1416
01:09:56,200 --> 01:09:58,557
in that industry are not going to be
able to do. And so it's a bet that by

1417
01:09:58,640 --> 01:10:02,157
basically by relighting the sort
of name you know, kind of thing

1418
01:10:02,240 --> 01:10:07,197
this new innovative thing with like
a king monarchical political structure. Um

1419
01:10:07,280 --> 01:10:08,997
that they're going to be able
to innovate in a way that the incumbent is

1420
01:10:09,080 --> 01:10:11,757
not going to be able to because
the incumbent is is being run by managers,

1421
01:10:11,840 --> 01:10:14,597
right? And and and and by the way
and of course venture being what it is

1422
01:10:14,680 --> 01:10:17,077
sometimes that works sometimes it doesn't,
but but we're constantly doing

1423
01:10:17,160 --> 01:10:19,917
that. But I've always viewed it
my entire life as like we're like raging

1424
01:10:20,000 --> 01:10:21,837
against the dying of the light. Like
we're we're we're [clears throat] we're

1425
01:10:21,920 --> 01:10:25,277
sort of constantly trying to fight off
managerialism just basically swamping

1426
01:10:25,360 --> 01:10:29,317
everything and everything getting
basically boring and gray and dumb and

1427
01:10:29,400 --> 01:10:32,557
old. Right. And it's we're trying
to keep some level of energy and vitality

1428
01:10:32,640 --> 01:10:35,717
in the system.
AI is the thing that would lead you to

1429
01:10:35,800 --> 01:10:38,237
think wow, maybe there's a third model.

1430
01:10:38,320 --> 01:10:40,797
Right. And and maybe and way to think
about it would be maybe it's a

1431
01:10:40,880 --> 01:10:43,837
combination of the two. Maybe the new
Henry Ford or the new Elon or the new

1432
01:10:43,920 --> 01:10:47,317
Steve Jobs plus AI is the best of both.

1433
01:10:47,400 --> 01:10:50,597
Right? Cuz it's it's it's sort of the
spark of genius of the name on the door

1434
01:10:50,680 --> 01:10:54,557
model, the Henry Ford model, but then
it's give that person AI superpowers to

1435
01:10:54,640 --> 01:10:58,237
do all the managerial stuff and let
the boss do all the managerial stuff. That

1436
01:10:58,320 --> 01:11:00,197
may be the actual secret formula.

1437
01:11:00,280 --> 01:11:03,077
And we've never even known that we
wanted this because we never even

1438
01:11:03,160 --> 01:11:06,277
thought it was a possibility. But yeah,
I mean, you know this. The What is the

1439
01:11:06,360 --> 01:11:08,397
thing that these bots are really good at?
They're really good at doing

1440
01:11:08,480 --> 01:11:12,037
paperwork. Like, they're really good
at filling out forms. Like, they're really

1441
01:11:12,120 --> 01:11:14,717
good at writing reports. They're really
good at reading them. They're really

1442
01:11:14,800 --> 01:11:16,397
good at doing all the managerial work.

1443
01:11:16,480 --> 01:11:20,117
Like, they're amazing at it. And so,
yeah. So, I I think I think the I 100% I

1444
01:11:20,200 --> 01:11:24,037
think the answer the answer very well
might be to get the best best best of

1445
01:11:24,120 --> 01:11:26,997
both worlds by doing this. And then
the challenge is going to be twofold. It's

1446
01:11:27,080 --> 01:11:29,517
going to be for the innovators to really
figure out how to leverage AI to

1447
01:11:29,600 --> 01:11:33,277
actually do this. Right? And then And then
the the other challenge is going to

1448
01:11:33,360 --> 01:11:36,517
be for the for the incumbents that are
managerial to figure out like, okay,

1449
01:11:36,600 --> 01:11:38,557
what does that mean? Cuz now they're
going to they're they're going to be

1450
01:11:38,640 --> 01:11:41,397
facing a different kind of insurgent
competitor that has a different set of

1451
01:11:41,480 --> 01:11:45,437
capabilities than they're used to. And so,
this really, I think, is going to force

1452
01:11:45,520 --> 01:11:48,637
a lot of big companies to kind of figure
out innovation. Either I say figure out

1453
01:11:48,720 --> 01:11:51,877
innovation or die trying.
Do you feel like that structure

1454
01:11:51,960 --> 01:11:56,237
accelerates the impact on the actual
GDP and economy? If you could space out,

1455
01:11:56,320 --> 01:11:58,317
it's like the growth is like so fast.

1456
01:11:58,400 --> 01:12:01,277
And like, instead of having these
companies kind of like peter out in

1457
01:12:01,360 --> 01:12:04,397
growth and impact, they can kind
of like keep going, if not accelerating.

1458
01:12:04,480 --> 01:12:08,237
>> Yeah. That's for sure the hope. The
The challenge and and you know, and and

1459
01:12:08,320 --> 01:12:11,077
look, the AI utopian view is a Of course.
And and and that's going to be

1460
01:12:11,160 --> 01:12:13,717
the future of the economy, and it's
going to grow 10x and 100x and 1,000x,

1461
01:12:13,800 --> 01:12:16,837
and we're entering this regime of like
much higher economic growth forever, and

1462
01:12:16,920 --> 01:12:19,717
consumer cornucopia of everything, and
it's going to be great. And I And I hope

1463
01:12:19,800 --> 01:12:22,557
that's true. I hope that's That's like
the you know, that's the current kind of

1464
01:12:22,640 --> 01:12:25,637
utopian vision. I hope that's true.
The problem is goes back again. The real

1465
01:12:25,720 --> 01:12:29,077
world is really messy. Um, and I'll give
you an example of how the real world is

1466
01:12:29,160 --> 01:12:32,677
really messy. It requires 900 hours
of professional certification training to

1467
01:12:32,760 --> 01:12:35,157
become a hairdresser
in the state of California.

1468
01:12:35,240 --> 01:12:40,037
Um, so, it's like 35% of the economy,
something like that, you have to get

1469
01:12:40,120 --> 01:12:42,877
some sort of professional
certification to do the job.

1470
01:12:42,960 --> 01:12:45,797
Which [snorts] is to say that the the
professions are all cartels, right? And

1471
01:12:45,880 --> 01:12:48,517
so, you have to get licensed as a doctor.
You have to get licensed as a

1472
01:12:48,600 --> 01:12:53,237
lawyer. You have to get licensed
as a You have to get into a union.

1473
01:12:53,320 --> 01:12:56,357
By the way, to to work for the government,
you need to be you have both

1474
01:12:56,440 --> 01:12:59,397
civil service protections and you
have public sector unions. You have two

1475
01:12:59,480 --> 01:13:02,157
layers of insulation against
ever getting fired for anything

1476
01:13:02,240 --> 01:13:03,227
or anything anything ever changing.

1477
01:13:03,310 --> 01:13:06,477
[laughter] I'll give you another example.
The the dock work the dock workers went on

1478
01:13:06,560 --> 01:13:10,637
strike a couple years ago cuz they're
robotics if you go look at a modern dock

1479
01:13:10,720 --> 01:13:13,997
like in Asia it's all robots. If you
go to American dock it's like all still

1480
01:13:14,080 --> 01:13:16,677
guys dragging dragging stuff by hand.

1481
01:13:16,760 --> 01:13:19,957
The dock workers went on a strike.
It turns out there are 25,000 dock workers

1482
01:13:20,040 --> 01:13:22,077
working on docks in America.
Turns out they have

1483
01:13:22,160 --> 01:13:25,437
incredible political power cuz it's
a it's one of these unified blocks of

1484
01:13:25,520 --> 01:13:28,477
things. They won their strike and so
they got commitments from the dock

1485
01:13:28,560 --> 01:13:30,437
owners to not implement more automation.

1486
01:13:30,520 --> 01:13:31,437
We learned a couple of things in that.

1487
01:13:31,520 --> 01:13:34,197
So number one we learned that even
a union as small as 25,000 people still

1488
01:13:34,280 --> 01:13:37,517
has like tremendous political stroke. We
also learned that they it actually turns

1489
01:13:37,600 --> 01:13:40,797
out the dock workers union has 50,000
people in it cuz there's 20 they have

1490
01:13:40,880 --> 01:13:43,437
25,000 people working at docks they
have 25,000 people drawing full paycheck

1491
01:13:43,520 --> 01:13:48,117
sitting at home from prior union
agreements from prior union agreements.

1492
01:13:48,200 --> 01:13:49,317
I'll give you another great example.

1493
01:13:49,400 --> 01:13:52,877
There are government agencies there are
federal government agencies where the

1494
01:13:52,960 --> 01:13:56,197
employees right of have civil service
protections and they're in public sector

1495
01:13:56,280 --> 01:13:59,437
unions. There are entire federal
government agencies that struck new

1496
01:13:59,520 --> 01:14:03,117
collective bargaining agreements during
COVID. Where not only are they have

1497
01:14:03,200 --> 01:14:05,957
their jobs guaranteed in perpetuity
but they only have to report to work in an

1498
01:14:06,040 --> 01:14:09,357
office one day per month.
And so there are entire office buildings

1499
01:14:09,440 --> 01:14:13,357
in Washington DC that are empty 29 out
of 30 days of the year that are still

1500
01:14:13,440 --> 01:14:15,797
operating and are still
we're all still paying for it.

1501
01:14:15,880 --> 01:14:17,677
And so and then what they do it turns
out what the employees do is they're

1502
01:14:17,760 --> 01:14:20,437
very they're very smart in the in in
this way and so they figure it out. They

1503
01:14:20,520 --> 01:14:23,477
come in on the last day of a month
and the first day of the next month.

1504
01:14:23,560 --> 01:14:27,597
And so they're in the they're
in the office two days per 60 days.

1505
01:14:27,680 --> 01:14:30,347
Which means these buildings
are empty for 58 days at a time.

1506
01:14:30,430 --> 01:14:32,557
>> [laughter] >> And you
see you see where I'm heading

1507
01:14:32,640 --> 01:14:35,637
with this? Like this is like locked in.

1508
01:14:35,720 --> 01:14:39,477
Right? This is like locked in in a way
that has nothing to do with like people

1509
01:14:39,560 --> 01:14:41,957
say capitalist it's like
anti-capitalistic. It's like it's it's

1510
01:14:42,040 --> 01:14:43,917
basically it's restrictions on trade.

1511
01:14:44,000 --> 01:14:45,437
It's restrictions on the ability to like

1512
01:14:45,520 --> 01:14:48,357
change the workforce. And so, so much of

1513
01:14:48,440 --> 01:14:52,477
our economy is is you know the I'm
describing the entire healthcare

1514
01:14:52,560 --> 01:14:56,677
system. I'm describing the entire legal
profession. I'm describing the entire

1515
01:14:56,760 --> 01:15:01,157
housing industry. I'm describing
the entire education system. Right, K-12

1516
01:15:01,240 --> 01:15:03,957
schools in the United States they're
a literal government monopoly.

1517
01:15:04,040 --> 01:15:06,157
How are we going to apply AI in education?
The answer is we're not

1518
01:15:06,240 --> 01:15:08,957
because it's a literal government
monopoly. It is never going to change.

1519
01:15:09,040 --> 01:15:12,237
The end and there is nothing to do.
By the way, you can create an entirely new

1520
01:15:12,320 --> 01:15:14,597
school system. Like that's the one
thing you can do is you can do what Alpha

1521
01:15:14,680 --> 01:15:16,997
School is doing. You can create
an entirely new school system. Other than

1522
01:15:17,080 --> 01:15:19,277
that, you're not going to go and change
what's happening in the American

1523
01:15:19,360 --> 01:15:21,477
classroom like K-12. There's no chance.

1524
01:15:21,560 --> 01:15:22,997
The teachers are 100% opposed to it.

1525
01:15:23,080 --> 01:15:26,237
It's 100% not going to happen. So So you
see what I'm saying is like there's this

1526
01:15:26,320 --> 01:15:30,157
like massive slippage that's going to take
place. Both the AI utopians and the

1527
01:15:30,240 --> 01:15:31,797
AI doomers are far too optimistic.

1528
01:15:31,880 --> 01:15:35,277
Right. >> [laughter] >> You see
what I'm saying because because

1529
01:15:35,360 --> 01:15:38,157
they believe that because the technology
makes something possible that 8 billion

1530
01:15:38,240 --> 01:15:40,157
people all of a sudden are going
to change how they behave and it's just

1531
01:15:40,240 --> 01:15:44,637
like, nope. So much of how the existing
economy works is just it's just like

1532
01:15:44,720 --> 01:15:48,477
wire done. And so, we're going to be
lucky as a society we're going to be

1533
01:15:48,560 --> 01:15:50,477
lucky if AI adoption happens quickly.

1534
01:15:50,560 --> 01:15:53,117
Right, because if it doesn't what
we're just going to have is stagnation.

1535
01:15:53,200 --> 01:15:55,357
As a matter of fact, I know you got
to run. Yeah, we all know if or still

1536
01:15:55,440 --> 01:15:59,037
welcome but uh it was such a pleasure
talking to you. Uh we're truly living in

1537
01:15:59,120 --> 01:16:02,677
an age of science fiction coming
to real life. Yes. Yes. Could not be more

1538
01:16:02,760 --> 01:16:05,797
excited. Really thank you Mark
for you guys. Awesome. Thank you.

1539
01:16:05,880 --> 01:16:08,560
Good. Thank you.

1540
01:16:15,230 --> 01:16:17,250
>> [music]