1
00:00:00,120 --> 00:00:03,560
Speaker 1: I wanted to try to build something useful, but I

2
00:00:03,600 --> 00:00:06,360
didn't think I would build anything particularly great. You've said,

3
00:00:06,440 --> 00:00:10,119
probabilistically it seemed unlikely, but I wouldn't to at least try.

4
00:00:11,039 --> 00:00:13,359
Speaker 2: So you're talking to a room full of people who

5
00:00:13,359 --> 00:00:16,879
are all technical engineers, often some of the most eminent

6
00:00:17,359 --> 00:00:19,280
AI researchers coming up in the game.

7
00:00:20,600 --> 00:00:25,519
Speaker 3: Okay, I I think we should. I think I like

8
00:00:25,559 --> 00:00:27,440
the term engineer better than researcher.

9
00:00:27,800 --> 00:00:32,560
Speaker 1: I suppose that if there's some fundamental algorithmic breakthrough, it's

10
00:00:32,600 --> 00:00:34,439
a research Otherwise it's engineering.

11
00:00:35,039 --> 00:00:37,799
Speaker 2: Maybe let's start way back when you were This is

12
00:00:37,840 --> 00:00:40,920
a room full of eighteen to twenty five year olds.

13
00:00:41,119 --> 00:00:44,880
Excuse younger, because the founder set is younger and younger.

14
00:00:45,359 --> 00:00:47,719
Speaker 4: Can you put yourself back into their shoes?

15
00:00:47,960 --> 00:00:52,840
Speaker 2: When you were eighteen nineteen learning to code, even coming

16
00:00:52,880 --> 00:00:55,759
up with a first idea for ZIP too, What was

17
00:00:55,799 --> 00:00:56,520
that like for you?

18
00:00:56,719 --> 00:00:56,960
Speaker 3: Yeah?

19
00:00:56,960 --> 00:01:00,759
Speaker 1: Back in ninety five, I was basically choice of either

20
00:01:00,880 --> 00:01:04,879
do Grass studies PhD a Stanford and material science actually

21
00:01:04,879 --> 00:01:08,159
working on ultra capacities for potentially use in electric vehicles,

22
00:01:08,239 --> 00:01:10,599
essentially trying to solve the range problem for electric vehicles,

23
00:01:11,319 --> 00:01:14,359
or try to do something in this thing that most

24
00:01:14,359 --> 00:01:16,920
people have never heard of called the Internet. And I

25
00:01:16,959 --> 00:01:19,400
talked to my professor, was Bill Nicks in the material

26
00:01:19,439 --> 00:01:21,760
science problement, and I said, can I defer for a

27
00:01:21,840 --> 00:01:25,680
quarter because this will probably fail and then I'll need

28
00:01:25,719 --> 00:01:27,519
to come back to college. And then he said, this

29
00:01:27,560 --> 00:01:30,359
is probably the last conversation we'll have, and he was right.

30
00:01:30,439 --> 00:01:32,319
But I thought things would most likely failed, not that

31
00:01:32,359 --> 00:01:33,599
they would most likely succeed.

32
00:01:33,719 --> 00:01:34,560
Speaker 3: And then in.

33
00:01:34,560 --> 00:01:38,079
Speaker 1: Ninety five I wrote, basically, I think the first or

34
00:01:38,079 --> 00:01:42,040
close to the first maps, directions, Internet, white pages and

35
00:01:42,120 --> 00:01:45,000
yellow pages on the Internet. I just wrote that personally,

36
00:01:45,200 --> 00:01:46,680
and I didn't even use website, but I just read

37
00:01:46,719 --> 00:01:50,480
the boot directly because I couldn't afford and I could

38
00:01:50,719 --> 00:01:55,400
afford a t one. The original office was on Firman

39
00:01:55,439 --> 00:01:58,799
Avenue in Palelto. There was like an ISP on the

40
00:01:58,799 --> 00:02:02,319
floor below. I rolled a hole through the floor and

41
00:02:02,519 --> 00:02:05,959
just ran a land cable directly to the ISP, and

42
00:02:06,200 --> 00:02:08,960
my brother joined me and another co founder of Greg Curry,

43
00:02:08,960 --> 00:02:11,319
who passed away. And at the time we couldn't even

44
00:02:11,360 --> 00:02:13,680
afford a place to stay, so we just The office

45
00:02:13,719 --> 00:02:15,159
was five hunred bucks a months, so we just left

46
00:02:15,159 --> 00:02:17,520
in the office and then shouted the YMCA on page

47
00:02:17,560 --> 00:02:20,639
mill and al communo and yeah, and I guess we

48
00:02:20,719 --> 00:02:23,360
ended up doing a little bit of a useful company

49
00:02:23,599 --> 00:02:26,159
was up to in the beginning, and I did build

50
00:02:26,199 --> 00:02:31,319
a lot of really really good software technology, but were

51
00:02:31,360 --> 00:02:35,039
somewhat captured by the legacy media companies and that night

52
00:02:35,120 --> 00:02:39,080
rout or New York Times, the host whatnot were investors

53
00:02:39,199 --> 00:02:43,120
and customers and also on the board. So I kept

54
00:02:43,120 --> 00:02:45,240
wanting to use software in ways that made no sense.

55
00:02:46,000 --> 00:02:48,280
So I wanted to go direct to consumers anyway. As

56
00:02:48,319 --> 00:02:50,319
long story dwelling too much on z up too, But

57
00:02:51,120 --> 00:02:52,639
I really just wanted to do something useful on the

58
00:02:52,639 --> 00:02:56,000
Internet because I had two choices, do I PhD and

59
00:02:56,159 --> 00:03:00,000
watch people build the internet or help build the internet

60
00:03:00,039 --> 00:03:02,159
in some small way. And I was like, I guess

61
00:03:02,159 --> 00:03:03,879
I can always try and fail and then go back

62
00:03:03,919 --> 00:03:06,000
to grass Studies and away.

63
00:03:06,000 --> 00:03:07,840
Speaker 3: That ended up being like reasonably successful.

64
00:03:08,439 --> 00:03:10,199
Speaker 1: It sold for like three hundred million dollars, which is

65
00:03:10,199 --> 00:03:12,439
a lot at the time these days, that's I think

66
00:03:12,439 --> 00:03:14,599
the minimum impulse, but for an AI startup is like

67
00:03:14,599 --> 00:03:15,560
a billion dollars.

68
00:03:15,759 --> 00:03:16,159
Speaker 3: It's like.

69
00:03:17,960 --> 00:03:21,479
Speaker 1: There's so many freaking unicorns. It's like heard of unicorns

70
00:03:21,520 --> 00:03:23,400
at this point is a billion dollar situation.

71
00:03:24,199 --> 00:03:27,080
Speaker 4: There's been inflation since, so quite a bit more money.

72
00:03:27,960 --> 00:03:32,000
Speaker 1: Yeah, like could probably buy a burger for a nickel.

73
00:03:32,120 --> 00:03:34,400
Not quite, but yeah, there has been a lot of inflation.

74
00:03:35,039 --> 00:03:37,879
But at the high level in AI is pretty intense.

75
00:03:38,439 --> 00:03:41,879
As you've seen companies that are I don't know, less

76
00:03:41,879 --> 00:03:45,039
than a year old getting sometimes a billion dollar, multi

77
00:03:45,080 --> 00:03:48,800
billion dollar valuations, which I guess could pan out and

78
00:03:49,039 --> 00:03:51,759
probably will pan out in some cases, but it is

79
00:03:51,800 --> 00:03:53,639
I wordering to see some of these valuations.

80
00:03:54,520 --> 00:03:59,319
Speaker 4: Yeah, what do you think, I'm pretty bullish. I'm pretty bullish, honestly, so.

81
00:03:59,520 --> 00:04:01,039
Speaker 2: I think the people in this room are going to

82
00:04:01,039 --> 00:04:04,199
create a lot of the value that a billion people

83
00:04:04,240 --> 00:04:06,960
in the world should be using this stuff, and we're

84
00:04:07,000 --> 00:04:10,039
not even we're scratching the surface of it. I love

85
00:04:10,080 --> 00:04:13,400
the Internet story in that, even back then, you are

86
00:04:13,439 --> 00:04:15,919
a lot like the people in this room back then,

87
00:04:16,399 --> 00:04:20,360
in that you know, the heads of all the CEOs

88
00:04:20,399 --> 00:04:23,160
of all the legacy media companies look to you as

89
00:04:23,160 --> 00:04:26,000
the person who understood the Internet, and a lot of

90
00:04:26,040 --> 00:04:28,879
the world, the corporate world, like the world out large,

91
00:04:28,920 --> 00:04:31,680
that does not understand what's happening with AI, they're going

92
00:04:31,759 --> 00:04:34,360
to look to the people in this room for exactly that.

93
00:04:34,439 --> 00:04:36,319
Speaker 4: It sounds like, what are some of the tangible lessons.

94
00:04:36,319 --> 00:04:38,319
Speaker 2: It sounds like one of them is don't give up

95
00:04:38,360 --> 00:04:41,639
board control or be careful about have a really good lawyer.

96
00:04:42,199 --> 00:04:44,480
Speaker 1: Well, I guess, well, the first my first thought of

97
00:04:44,720 --> 00:04:48,079
the big really the mistake was having too much shareholder

98
00:04:48,160 --> 00:04:53,399
and board control from legacy media companies who then necessarily

99
00:04:53,439 --> 00:04:57,600
see things through the lens of legacy media, and that

100
00:04:57,800 --> 00:05:00,319
they'll make you do things that seems sense full to

101
00:05:00,360 --> 00:05:03,319
them but aren't really don't make sense within your technology.

102
00:05:03,800 --> 00:05:06,040
I know I should point out that I that I

103
00:05:06,399 --> 00:05:09,439
didn't actually first intend to start a company. I try

104
00:05:09,480 --> 00:05:12,160
to get a job at Netscape. I SIMI resume to

105
00:05:12,279 --> 00:05:15,360
Netscape and Mark Kandrison knows about this, but I don't

106
00:05:15,360 --> 00:05:17,959
think you gave us a myristhume. And then nobody responded.

107
00:05:18,399 --> 00:05:19,959
And then I tried hanging out in the lobby of

108
00:05:20,040 --> 00:05:22,600
Netscape see if I could like bump into someone, but

109
00:05:22,639 --> 00:05:24,639
I was like too shy to talk to anyone. So

110
00:05:24,680 --> 00:05:26,879
I'm like, man, this is ridiculous. So I'll just write

111
00:05:26,920 --> 00:05:28,639
sof for myself and see how it goes. So it

112
00:05:28,680 --> 00:05:30,040
wasn't actually from the stand point, if I want to

113
00:05:30,040 --> 00:05:31,439
start a company, I just want to be part of

114
00:05:31,439 --> 00:05:35,360
building the Internet in some way. And since I couldn't

115
00:05:35,360 --> 00:05:37,319
get a job at an area company, I had to

116
00:05:37,319 --> 00:05:42,560
start a in area company anyway. The AI will so

117
00:05:42,720 --> 00:05:47,680
profoundly change the future, it's difficult to fathom how much.

118
00:05:48,040 --> 00:05:54,399
But at the economy is, assuming we don't things don't

119
00:05:54,399 --> 00:05:58,519
go awry, and like AI doesn't kill us all in itself,

120
00:05:59,000 --> 00:06:05,560
then you'll see ultimately an economy that is not ten

121
00:06:05,600 --> 00:06:08,120
times more than the current economy. Ultimately, if we become

122
00:06:08,199 --> 00:06:14,240
say or whatever, our future machine descendants, or but mostly

123
00:06:14,319 --> 00:06:19,680
machine descendants become like a ship scale to civilization or beyond,

124
00:06:20,199 --> 00:06:25,920
we're talking about an economy that is thousands of times,

125
00:06:26,199 --> 00:06:30,839
maybe millions of times bigger than the economy today. Yeah,

126
00:06:31,040 --> 00:06:32,319
I did feel a bit like when I was in

127
00:06:32,399 --> 00:06:34,399
DC taking a lot of flag from getting rid of

128
00:06:34,519 --> 00:06:37,800
waste and fraud, which is an interesting side quest as

129
00:06:37,839 --> 00:06:38,519
side quests go.

130
00:06:39,199 --> 00:06:41,240
Speaker 4: But I've got to get back to the main quest.

131
00:06:41,800 --> 00:06:43,560
Speaker 1: Yeah, i gotta get back to the main quest here,

132
00:06:43,959 --> 00:06:46,199
So back to the main quest. But I did feel

133
00:06:46,199 --> 00:06:49,240
a little bit like this. It's like fixing the government.

134
00:06:49,279 --> 00:06:52,120
It's kind of say, the beach is dirty and there's

135
00:06:52,120 --> 00:06:54,839
some needles and vcs and like trash, and you want

136
00:06:54,839 --> 00:06:57,439
to clean up the beach. But then there's also this

137
00:06:57,560 --> 00:07:01,639
like thousand foot wool of water, which is tsunami of AI.

138
00:07:02,199 --> 00:07:04,560
And how much does cleaning the beach really matter if

139
00:07:04,600 --> 00:07:06,800
you got a thousand foot tsunami about to hit.

140
00:07:07,839 --> 00:07:08,399
Speaker 3: Not that much.

141
00:07:09,759 --> 00:07:12,000
Speaker 2: Oh well, we're glad you're back on the main quest.

142
00:07:12,040 --> 00:07:13,000
It's very important.

143
00:07:13,240 --> 00:07:15,720
Speaker 1: Yeah, back to the may quest. Building technology, which is

144
00:07:15,759 --> 00:07:18,199
what I like doing. It's just so much noise like this.

145
00:07:18,360 --> 00:07:21,000
The single noise ratio in politics is terrible.

146
00:07:21,240 --> 00:07:23,040
Speaker 4: I live in San Francisco, so you don't need to

147
00:07:23,079 --> 00:07:23,920
tell me twice.

148
00:07:24,319 --> 00:07:26,639
Speaker 3: Yeah, DC's I guess it's all politics in DC.

149
00:07:26,800 --> 00:07:32,160
Speaker 1: But the if you're trying to build a rocket or cars,

150
00:07:32,279 --> 00:07:35,439
or you're trying to have software that compiles and runs reliably,

151
00:07:35,920 --> 00:07:38,800
then you have to be maximally true seeking or your

152
00:07:38,839 --> 00:07:42,439
software or your hardware won't work like you can't full

153
00:07:42,879 --> 00:07:46,079
like math and physics are rigorous judges. So I'm used

154
00:07:46,120 --> 00:07:49,199
to being in a maximally true seeking environment and that's

155
00:07:49,199 --> 00:07:52,319
definitely not politics. So I'm good, glad to be back

156
00:07:52,319 --> 00:07:53,319
in technology.

157
00:07:54,079 --> 00:07:56,360
Speaker 2: I guess I'm curious going back to the zip to moment,

158
00:07:56,680 --> 00:07:58,959
you had hundreds of millions of dollars or you had

159
00:07:58,959 --> 00:08:01,279
to exit a worth millions of dollars.

160
00:08:01,399 --> 00:08:02,839
Speaker 3: I got twenty million dollars, right.

161
00:08:03,480 --> 00:08:06,399
Speaker 2: Okay, so you solved the money problem at least, and

162
00:08:06,480 --> 00:08:09,319
you basically took it and you rolled. You kept rolling

163
00:08:09,399 --> 00:08:13,600
with X dot Com, which in PayPal and Confinity.

164
00:08:13,680 --> 00:08:15,600
Speaker 3: Yes, I kept the chips on the table.

165
00:08:16,040 --> 00:08:18,600
Speaker 4: Yeah, so not everyone does that.

166
00:08:18,800 --> 00:08:20,399
Speaker 2: A lot of the people in this room will have

167
00:08:20,439 --> 00:08:23,240
to make that decision. Actually, what drove you to jump

168
00:08:23,279 --> 00:08:24,040
back into the ring?

169
00:08:25,439 --> 00:08:25,720
Speaker 3: O Thay?

170
00:08:25,920 --> 00:08:28,519
Speaker 1: I felt for Up too. We'd both like incredible technology

171
00:08:28,519 --> 00:08:31,680
but never really got used. I think, at least from

172
00:08:31,720 --> 00:08:34,759
my perspective, we had better technology than say Yahoo or

173
00:08:34,759 --> 00:08:38,320
anyone else, but it was constrained by our customers. And

174
00:08:38,559 --> 00:08:41,120
so I wanted to do something that where okay, we

175
00:08:41,120 --> 00:08:43,720
wouldn't be constrained by our customers, go direct consumer. And

176
00:08:43,759 --> 00:08:47,240
that's what ended up being like X dot Com PayPal,

177
00:08:47,519 --> 00:08:52,360
essentially X dot Com merging with Confinity, which together created PayPal,

178
00:08:52,559 --> 00:08:56,159
and then that actually the sort of PayPal Diespora has it.

179
00:08:56,159 --> 00:08:59,639
It might have created more companies then so more companies

180
00:08:59,679 --> 00:09:03,039
than probaly any anything in the twenty first century, so

181
00:09:03,039 --> 00:09:06,159
many talented people were at the combination of Confinity X

182
00:09:06,200 --> 00:09:08,879
dot com. So I just wanted to, like I felt,

183
00:09:08,879 --> 00:09:11,440
we got our wings clift somewhat with ZIP two, and

184
00:09:11,440 --> 00:09:13,159
it's okay, what if our wings are clipped and we

185
00:09:13,279 --> 00:09:16,440
go direct to consumer and that's what PayPal ended up being.

186
00:09:17,200 --> 00:09:20,759
But yeah, with I got that twenty million dollar check

187
00:09:20,960 --> 00:09:24,039
for my share of ZIP two. At the time, I

188
00:09:24,080 --> 00:09:27,720
was living in a house with four housemates, and I

189
00:09:27,799 --> 00:09:30,840
had a ten grand in the bank, and then this

190
00:09:31,000 --> 00:09:34,000
check arrived in the mail, of all places in the mail,

191
00:09:34,120 --> 00:09:37,000
and then now then my bank bounce went from ten

192
00:09:37,399 --> 00:09:40,480
to twenty million and ten thousand. Okay, seluck to pay

193
00:09:40,559 --> 00:09:42,720
taxes on that at all. But then I ended up

194
00:09:42,720 --> 00:09:45,840
putting almost all of that into x dot com and

195
00:09:45,919 --> 00:09:48,360
as you said, just keeping almost all the chips on

196
00:09:48,399 --> 00:09:48,840
the table.

197
00:09:49,600 --> 00:09:53,240
Speaker 3: And yeah, and then after PayPal was like, I was

198
00:09:53,279 --> 00:09:53,759
curious as to.

199
00:09:53,799 --> 00:09:56,840
Speaker 1: Why we had not sent anyone to Mars, and I

200
00:09:56,919 --> 00:09:59,080
went on the NASA website to find out when we're

201
00:09:59,120 --> 00:10:02,000
sending people tomorrow, and there was no date. I thought,

202
00:10:02,000 --> 00:10:04,919
maybe it's just hard to find on the website, but

203
00:10:05,000 --> 00:10:07,600
in fact, there was no real plan to send people

204
00:10:07,600 --> 00:10:11,159
to Mars. So then I've come to this is such

205
00:10:11,200 --> 00:10:12,559
a long story, so I don't want to take up

206
00:10:12,559 --> 00:10:14,320
too much time here, but I.

207
00:10:14,279 --> 00:10:16,039
Speaker 4: Think we're all listening with rapt attention.

208
00:10:16,639 --> 00:10:17,159
Speaker 3: So I was.

209
00:10:17,159 --> 00:10:19,679
Speaker 1: Actually I was on the long end of Expressway with

210
00:10:19,720 --> 00:10:22,360
my friend day O Ressi. We were like housemates in college,

211
00:10:22,840 --> 00:10:24,679
and Daa was asking me what I'm what we're gonna do?

212
00:10:24,799 --> 00:10:26,840
What am I gonna do after a PayPal and I

213
00:10:26,879 --> 00:10:28,919
was like, I don't know. I guess maybe I'd like

214
00:10:28,919 --> 00:10:31,360
to do something philanthropic in space, because I didn't think

215
00:10:31,360 --> 00:10:33,840
I could actually do anything commotional in space because that

216
00:10:33,840 --> 00:10:37,559
seemed like the purview of nations. But I'm curious as

217
00:10:37,600 --> 00:10:40,000
to when we're gonna send people to Mars. And that's

218
00:10:40,000 --> 00:10:41,519
when I was like, oh, it's not on the website,

219
00:10:41,559 --> 00:10:45,000
and I started digging. There's nothing on the NAST website.

220
00:10:45,039 --> 00:10:48,519
So then I started digging in and I'm definitely summarizing.

221
00:10:48,000 --> 00:10:48,519
Speaker 3: A lot here.

222
00:10:48,600 --> 00:10:52,279
Speaker 1: But my first idea was to do a philanthropic mission

223
00:10:52,320 --> 00:10:56,480
to Mars, Cooled Life to Mars, where would send a

224
00:10:56,519 --> 00:11:00,919
small greenhouse with CS and dehydrated nutrient gael Land out

225
00:11:00,919 --> 00:11:03,600
our mosque and grow hydrate the jel and then you

226
00:11:03,720 --> 00:11:06,720
have this great sort of money shot of green plants

227
00:11:06,759 --> 00:11:09,399
on a red background. For the longest time, I by

228
00:11:09,440 --> 00:11:11,039
the way, I didn't realize money shot. I think is

229
00:11:11,039 --> 00:11:13,600
a porn reference. But anyway, the point is that would

230
00:11:13,639 --> 00:11:15,960
be the great shot of green plants on a red background,

231
00:11:16,080 --> 00:11:18,919
and to try to inspire a NASA and the public

232
00:11:19,480 --> 00:11:22,399
to Santa astronauts to Mars. As I learned more, I

233
00:11:22,480 --> 00:11:25,240
came to realize along the way. By the way, I

234
00:11:25,279 --> 00:11:27,799
went to Russia in two thousand and one and two

235
00:11:27,840 --> 00:11:32,200
thousand and two to buy ICBMs, which is that's an

236
00:11:32,240 --> 00:11:34,559
adventure where you go and meet with Russian high command

237
00:11:34,559 --> 00:11:36,240
and say I'd like to buy some ICBMs.

238
00:11:37,240 --> 00:11:40,320
Speaker 4: This was to get to space, yeah, as a rocket, not.

239
00:11:40,240 --> 00:11:42,799
Speaker 1: To nuke anyone, but they had to as a result

240
00:11:42,879 --> 00:11:45,639
of armstroduction talks. They had to actually destroy a bunch

241
00:11:45,679 --> 00:11:49,039
of their big nuclear missiles. So I was like, how

242
00:11:49,039 --> 00:11:51,279
about if we take two of those minus the luke

243
00:11:51,759 --> 00:11:54,120
added an additional opper stage for Mars.

244
00:11:54,600 --> 00:11:56,080
Speaker 3: But it was trippy.

245
00:11:55,960 --> 00:11:58,840
Speaker 1: Being in Moscow in one of two thousand and one

246
00:11:58,919 --> 00:12:04,519
negotiating with the Russian military to buy icvms, like that's crazy,

247
00:12:07,200 --> 00:12:10,360
but the kept also raising the price on me, so

248
00:12:10,399 --> 00:12:12,559
that so literally it's like the opposite of what a

249
00:12:12,600 --> 00:12:15,519
negotiation to do. So I was like, man, these things

250
00:12:15,519 --> 00:12:17,759
are getting really expensive. And then I came to realize

251
00:12:17,799 --> 00:12:20,440
that actually the problem was not that there was insufficient

252
00:12:20,480 --> 00:12:22,639
will to go to Mars, but there was no way

253
00:12:22,679 --> 00:12:25,039
to do without breaking the budget, or even breaking the

254
00:12:25,120 --> 00:12:29,320
NASA budget. So that's where I decided to start SpaceX's

255
00:12:29,519 --> 00:12:33,159
SpaceX to advance rocket technology to the point where we

256
00:12:33,159 --> 00:12:34,799
could send people to Mars.

257
00:12:35,720 --> 00:12:37,360
Speaker 3: And that was in two thousand and two.

258
00:12:38,039 --> 00:12:41,039
Speaker 2: So that wasn't But you didn't start out wanting to

259
00:12:41,080 --> 00:12:44,320
start a business. You wanted to start just something that

260
00:12:44,440 --> 00:12:48,159
was interesting to you that you thought humanity needed, and

261
00:12:48,200 --> 00:12:51,799
then as you kat pulling on a string, it just

262
00:12:51,840 --> 00:12:54,759
the ball sort of unravels and it turns out this

263
00:12:54,840 --> 00:12:57,039
is could be a very profitable business.

264
00:12:58,519 --> 00:13:02,200
Speaker 1: It is now, but it there had been no prior

265
00:13:02,279 --> 00:13:05,200
example of really a rocket startup succeeding, but there have

266
00:13:05,200 --> 00:13:08,720
been various attempts to do commercial rocket companies, and the

267
00:13:08,799 --> 00:13:12,919
whole failed again with SpaceX starting. SpaceX was really from

268
00:13:12,919 --> 00:13:15,000
the standpoint of I think there's a less than ten

269
00:13:15,039 --> 00:13:18,720
percent chance of being successful, maybe one percent, I don't know.

270
00:13:18,919 --> 00:13:23,519
But if a startup doesn't do something to advance rocket technology,

271
00:13:23,600 --> 00:13:27,000
it's definitely not coming from the big defense contractors because

272
00:13:27,039 --> 00:13:30,120
they just impedance smashed to the government, and the government

273
00:13:30,159 --> 00:13:34,080
just wants to do very conventional things. So it's either

274
00:13:34,120 --> 00:13:36,679
coming from a startup where it's not happening at all.

275
00:13:36,799 --> 00:13:39,480
So like a small chance of success is better than

276
00:13:39,480 --> 00:13:43,120
no chance of success, and so that yeah, so SpaceX

277
00:13:44,000 --> 00:13:47,600
start that in mid two thousand and two expecting to fail,

278
00:13:47,879 --> 00:13:50,919
like I said, probably ninety percent chance of failing. And

279
00:13:51,000 --> 00:13:54,080
even like when recruiting people, I didn't try to make

280
00:13:54,120 --> 00:13:56,600
out that it would I said, we're probably going to die,

281
00:13:57,120 --> 00:13:58,000
but a small.

282
00:13:57,799 --> 00:13:58,759
Speaker 3: Chance we might not die.

283
00:13:59,679 --> 00:14:01,720
Speaker 1: But this is the only way to get people to

284
00:14:01,759 --> 00:14:04,799
Mars and advanced the state of the art. And then

285
00:14:05,000 --> 00:14:07,120
I ended up being chief engineer of the rocket, not

286
00:14:07,200 --> 00:14:09,960
because I wanted to, but because I couldn't hire anyone

287
00:14:09,960 --> 00:14:13,960
who was good, So like none of the good sort

288
00:14:13,960 --> 00:14:15,879
of chief engineers would join because it's like this is

289
00:14:15,919 --> 00:14:18,440
too risky, you were going to die. And so then

290
00:14:18,480 --> 00:14:20,440
I ended up being chief engineer of the rocket, and

291
00:14:21,039 --> 00:14:23,039
the first three flights did fail, so it's a bit

292
00:14:23,080 --> 00:14:27,480
of a learning exercise there, and fourth one fortunately worked.

293
00:14:27,480 --> 00:14:29,559
But if the fourth one hadn't worked, I had no

294
00:14:29,639 --> 00:14:32,000
money left and that would have been it would have

295
00:14:32,039 --> 00:14:34,159
been curtained. So it was a pretty close thing. If

296
00:14:34,200 --> 00:14:37,360
the fourth launch of a Falcon not work, it would

297
00:14:37,360 --> 00:14:40,039
have been just curtains and we would have just been

298
00:14:40,159 --> 00:14:43,840
joined the graveyard of prior rocket startups. So it's like

299
00:14:43,919 --> 00:14:47,799
my estimate of success was not far off. We'd just

300
00:14:47,879 --> 00:14:50,000
be made it by the skin of our teeth. And

301
00:14:50,200 --> 00:14:53,200
Tesla was happening simultaneously, Like two thousand and eight was

302
00:14:53,200 --> 00:14:55,919
a rough year because at mid two thousand and eight

303
00:14:56,519 --> 00:14:59,000
were called Summer two thousand and eight, the third launch

304
00:14:59,039 --> 00:15:02,320
of Spacexit failed, a third failure in a row. The

305
00:15:02,320 --> 00:15:05,799
Tesla financing round had failed, and so Telsa was going

306
00:15:05,840 --> 00:15:10,080
bankrupt fast. It was just, man, this is a grim.

307
00:15:10,279 --> 00:15:13,080
This is going to be a tale of warning of

308
00:15:13,799 --> 00:15:15,240
an exercise in hubris.

309
00:15:15,720 --> 00:15:18,080
Speaker 2: Probably throughout that period a lot of people were saying,

310
00:15:18,279 --> 00:15:21,639
Elon is a software guy, Why are you working on hardware?

311
00:15:21,960 --> 00:15:24,440
Speaker 4: Why would you yeah, why would he choose to work

312
00:15:24,440 --> 00:15:24,759
on this.

313
00:15:25,159 --> 00:15:27,799
Speaker 1: Harp said, so you can look at the because they're

314
00:15:27,799 --> 00:15:30,240
still the press of that time is still online. You

315
00:15:30,240 --> 00:15:33,240
could just search it. And they kept calling the Internet guy.

316
00:15:33,919 --> 00:15:35,879
So Internet Guy aka.

317
00:15:35,919 --> 00:15:40,279
Speaker 3: Fool is attempting to build a rocket company that we

318
00:15:40,320 --> 00:15:41,679
got ridiculed quite a lot.

319
00:15:42,120 --> 00:15:45,240
Speaker 1: And it does sound pretty absurd, like internet guy starts

320
00:15:45,399 --> 00:15:49,120
rocket company doesn't sound like a recipe for success, frankly,

321
00:15:49,799 --> 00:15:51,639
so I didn't hold against them. I was like, yeah,

322
00:15:51,799 --> 00:15:55,000
admittedly it does sound improbable, and I agree that it's improbable.

323
00:15:55,480 --> 00:15:58,559
But fortunately the fourth launch worked and a NASSA ordered

324
00:15:58,639 --> 00:16:02,279
us a contract supply the space station. And I think

325
00:16:02,279 --> 00:16:08,799
that was like maybe December twenty second or it was

326
00:16:08,799 --> 00:16:11,840
it like right before Christmas. Because even the fourth launch

327
00:16:11,879 --> 00:16:15,279
working wasn't enough to succeed, NASA also needed we also

328
00:16:15,320 --> 00:16:17,279
needed a big contract to keep us life. So I

329
00:16:17,320 --> 00:16:20,919
got that call from like the NASA team and I

330
00:16:21,000 --> 00:16:24,159
literally they said, we're rewarding you one of the contracts

331
00:16:24,240 --> 00:16:26,399
to re supply the space station. I like literally blurred

332
00:16:26,399 --> 00:16:29,440
it out, I love you guys, which is not normally

333
00:16:29,559 --> 00:16:32,960
what they hear because it's usually pretty sober. But I

334
00:16:33,000 --> 00:16:35,360
was like, man, this is a company saver and then

335
00:16:36,320 --> 00:16:39,960
we closed the Tesla financing round on the last hour

336
00:16:40,000 --> 00:16:41,720
of the last day that it was possible, which was

337
00:16:41,759 --> 00:16:45,279
six pm December twenty fourth, two thousand and eight. We

338
00:16:45,320 --> 00:16:47,399
would have bounced payroll two days after Christmas if that

339
00:16:47,480 --> 00:16:50,919
round hadn't closed. So that was a nerve racking end

340
00:16:50,919 --> 00:16:52,279
of two thousand and eight, that's for sure.

341
00:16:52,919 --> 00:16:55,639
Speaker 2: I guess from your PayPal and zip two experience jumping

342
00:16:55,720 --> 00:17:00,799
into these hardcore hardware startups, it feels like one of

343
00:17:00,799 --> 00:17:04,039
the through lines was being able to find and eventually

344
00:17:04,119 --> 00:17:08,960
attract the smartest possible people in those particular fields. What

345
00:17:09,440 --> 00:17:11,160
would the people in this room like some of the

346
00:17:11,240 --> 00:17:12,880
most of the people here I don't think have even

347
00:17:13,200 --> 00:17:16,000
managed a single person yet, they're just starting their careers.

348
00:17:16,359 --> 00:17:19,319
What would you tell to the elon who's never had

349
00:17:19,359 --> 00:17:20,960
to do that yet.

350
00:17:21,839 --> 00:17:24,079
Speaker 1: I generally think to try to be as useful as possible.

351
00:17:24,119 --> 00:17:27,920
It may sound try, but it's so hard to be useful,

352
00:17:28,000 --> 00:17:30,680
especially to be useful to a lot of people. Where

353
00:17:30,720 --> 00:17:32,839
you say, the area under the cove of total utility

354
00:17:32,920 --> 00:17:34,480
is like how much how useful have you been to

355
00:17:34,519 --> 00:17:37,880
your fela human beings? Times how many people? It's almost

356
00:17:37,880 --> 00:17:40,680
like the physics definition of true work. It's incredibly difficult

357
00:17:40,720 --> 00:17:42,480
to do that. And I think if you're aspire to

358
00:17:42,480 --> 00:17:47,319
do true work, your probably success is much higher. Don't

359
00:17:47,319 --> 00:17:49,079
aspire to glory, aspire to work.

360
00:17:49,680 --> 00:17:53,359
Speaker 2: How can you tell that it's true work? Is it external?

361
00:17:53,440 --> 00:17:55,559
Is it like what happens with other people or what

362
00:17:55,599 --> 00:17:58,079
the product does for people? Like what is that for you?

363
00:17:59,119 --> 00:18:01,400
When you're looking for people come work for you? What's

364
00:18:01,440 --> 00:18:03,720
the salient thing that you look for or if they're

365
00:18:04,319 --> 00:18:04,920
certain question?

366
00:18:04,960 --> 00:18:07,839
Speaker 1: I guess it's in terms of if you're in product,

367
00:18:07,920 --> 00:18:09,720
you just have to say, if this thing is successful,

368
00:18:09,759 --> 00:18:12,519
how useful will it be to how many people? And

369
00:18:12,599 --> 00:18:14,799
that's what I mean. And then you do you do whatever,

370
00:18:15,559 --> 00:18:17,720
whether you're CEO or any role in a startup, you

371
00:18:17,759 --> 00:18:20,319
do whatever it takes to succeed, and just always be

372
00:18:20,400 --> 00:18:25,559
smash smashing your ego like internalized responsibility like a major

373
00:18:25,599 --> 00:18:29,079
failure mode is when ego to ability ratio is double

374
00:18:29,079 --> 00:18:31,559
graded than sign one. If you if your ego to

375
00:18:31,559 --> 00:18:34,319
ability ratio is it gets too high, then you're going

376
00:18:34,400 --> 00:18:37,279
to basically break the feedback loop to reality. And in

377
00:18:37,279 --> 00:18:41,000
in a it terms, your your er you'll have you'll

378
00:18:41,039 --> 00:18:44,039
break your RL loop so you want you want to

379
00:18:44,079 --> 00:18:45,240
don't want to break your art. You want to have

380
00:18:45,240 --> 00:18:48,319
a strong r L loop, which means internalizing responsibility and

381
00:18:48,359 --> 00:18:51,079
minimizing ego, and you do whatever the task is, no

382
00:18:51,119 --> 00:18:54,160
matter whether it's grand or humble. That's like what I

383
00:18:54,240 --> 00:18:59,480
actually I prefer the term like engineering as opposed to research.

384
00:19:00,319 --> 00:19:02,880
I prefer the term and I don't actually don't want

385
00:19:02,920 --> 00:19:04,079
it to call XAI lab.

386
00:19:04,079 --> 00:19:05,359
Speaker 3: I just want to be a company.

387
00:19:06,000 --> 00:19:10,880
Speaker 1: Like It's like, what of the simplest, more straightforward, ideally

388
00:19:10,920 --> 00:19:13,039
lowest that ego terms are that those are generally a

389
00:19:13,079 --> 00:19:14,720
good way to go. You want you want to just

390
00:19:14,759 --> 00:19:17,960
close the loop on reality hard that's a super big deal.

391
00:19:18,359 --> 00:19:20,599
Speaker 2: I think everyone in this room is really looks up

392
00:19:20,680 --> 00:19:24,359
to everything you've done around being a paragon of first

393
00:19:24,359 --> 00:19:28,319
principles and thinking about the stuff you've done. How do

394
00:19:28,400 --> 00:19:32,599
you actually determine your reality because that seems like a

395
00:19:32,599 --> 00:19:34,440
pretty big part of it.

396
00:19:34,640 --> 00:19:36,480
Speaker 4: Like other people who have never made.

397
00:19:36,319 --> 00:19:40,559
Speaker 2: Anything, non engineers, sometimes journalists at time who've never done

398
00:19:40,559 --> 00:19:44,319
anything like they will criticize you. But then clearly you

399
00:19:44,400 --> 00:19:47,640
have another set of people who are builders, who have

400
00:19:48,279 --> 00:19:51,359
very high area under the curve, who are in your

401
00:19:51,359 --> 00:19:54,279
circle how should people approach that? What has worked for you?

402
00:19:54,359 --> 00:19:57,920
And what would you pass on to X to your children?

403
00:19:58,000 --> 00:20:00,480
What do you tell them when you're like, you need

404
00:20:00,519 --> 00:20:03,039
to make your way in this world? Here's how to

405
00:20:03,119 --> 00:20:06,559
construct a reality. It is predictive from first principles.

406
00:20:07,079 --> 00:20:11,240
Speaker 1: The tools of physics are incredibly helpful to understand and

407
00:20:11,279 --> 00:20:14,519
make progress in any field. The first principles means just obviously,

408
00:20:14,559 --> 00:20:17,279
just means breakings down to the fundamental axiomatic elements that

409
00:20:17,279 --> 00:20:19,599
are most likely to be true, and then reason up

410
00:20:19,640 --> 00:20:23,079
from there as cosionally as possible, as opposed to reasoning

411
00:20:23,079 --> 00:20:26,839
by analysis or metaphor. And then it just simple things

412
00:20:26,920 --> 00:20:30,200
like thinking in the limit if you extrapolate, minimize this

413
00:20:30,279 --> 00:20:32,400
thing or maximize that thing. Thinking in the limit is

414
00:20:32,480 --> 00:20:35,599
very helpful. I use all the tools of physics. They

415
00:20:35,680 --> 00:20:39,359
apply to any field. This is like a superpower. Actually,

416
00:20:40,440 --> 00:20:42,799
so you can take, say, take an example like rockets.

417
00:20:42,839 --> 00:20:45,119
You can say how much should a rocket cost? A

418
00:20:45,200 --> 00:20:48,240
typical approach to how to that people would take to

419
00:20:48,240 --> 00:20:50,440
how much rocket should cost is they would look historically

420
00:20:50,519 --> 00:20:53,279
at what the cost of rockets are and assume that

421
00:20:53,319 --> 00:20:55,640
any new rocket must be somewhat similar to the prior

422
00:20:55,720 --> 00:20:58,599
cost of rockets. A first principle's approach would be you

423
00:20:58,599 --> 00:21:01,200
look at the materials that the rocket is comprised of,

424
00:21:01,319 --> 00:21:05,519
so if that's aluminium, copper, combon, fiber, steel, whatever the

425
00:21:05,519 --> 00:21:08,759
case may be, and say what what how much does

426
00:21:08,759 --> 00:21:11,000
that rocket weigh, and what are the constituent elements and

427
00:21:11,000 --> 00:21:13,400
how much do they weigh? What is the material price

428
00:21:13,480 --> 00:21:16,759
per kilogram of those constituent elements. In that sense the

429
00:21:16,799 --> 00:21:20,599
actual floor on what a rocket can cost. It can

430
00:21:20,640 --> 00:21:24,440
asymptotically approach the cost of the row materials and then

431
00:21:24,480 --> 00:21:27,039
you realize, oh, actually a rocket, the row materials of

432
00:21:27,079 --> 00:21:29,720
a rocket are only maybe one or two percent of

433
00:21:29,839 --> 00:21:34,240
the historical cost of a rocket, so the manufacturing must

434
00:21:34,359 --> 00:21:38,720
necessarily be very inefficient if the if the row material

435
00:21:38,759 --> 00:21:40,680
cost is only one or two percent. However, you have

436
00:21:40,720 --> 00:21:44,319
first principles analysis of the potential for the cost for

437
00:21:44,759 --> 00:21:46,240
cost optimization of a rocket.

438
00:21:46,400 --> 00:21:48,079
Speaker 3: And that's before you get to reusability.

439
00:21:48,599 --> 00:21:51,319
Speaker 1: To give an AI sort of AI example last year

440
00:21:51,359 --> 00:21:54,119
went for x CI when we were trying to build

441
00:21:54,359 --> 00:21:58,799
a training supercluster. We went to the various suppliers to

442
00:21:58,880 --> 00:22:02,799
ask said, this is the beginning of last year that

443
00:22:02,880 --> 00:22:05,119
we needed one hundred thousand, eighth one hundreds to be

444
00:22:05,160 --> 00:22:09,440
able to train coherently. And there are estimates for how

445
00:22:09,440 --> 00:22:11,480
long it would take to complete that were eighteen to

446
00:22:11,519 --> 00:22:14,119
twenty four months, so we need to get that done

447
00:22:14,119 --> 00:22:18,240
in six months so then or we won't be competitive.

448
00:22:18,599 --> 00:22:20,759
So then if you break that down, what are the

449
00:22:20,799 --> 00:22:23,240
things you need? We need a building, you need power,

450
00:22:23,279 --> 00:22:27,000
you need cooling. We didn't have enough time to build

451
00:22:27,000 --> 00:22:29,480
a building from scratch, so we've had to find existing building.

452
00:22:29,519 --> 00:22:31,920
So we found a factory that was no longer in

453
00:22:32,039 --> 00:22:36,079
use in Memphis that used to build electrolytes products. But

454
00:22:36,119 --> 00:22:38,720
then the input power was fifteen megawats and we needed

455
00:22:38,720 --> 00:22:42,200
one hundred and fifteen megawats. We read the generators and

456
00:22:42,279 --> 00:22:44,400
had generators on one side of the building, and then

457
00:22:44,440 --> 00:22:46,160
we have to have cooling, so we rented about a

458
00:22:46,200 --> 00:22:48,079
quarter of the mobile cooling capacity of the US and

459
00:22:48,119 --> 00:22:49,960
put the chillers on the other side of the building.

460
00:22:50,599 --> 00:22:52,599
But that didn't fully solve the problem because the volta

461
00:22:53,000 --> 00:22:56,759
the power variations during training are very big, so you

462
00:22:56,759 --> 00:22:59,680
can have power can drop by fifty percent in a

463
00:22:59,720 --> 00:23:02,480
hundred most seconds, which the generators can't keep up with.

464
00:23:02,519 --> 00:23:06,400
So then we added Tesla megapacks and modified the software

465
00:23:06,400 --> 00:23:08,200
in the megapacks to be able to just mooth out

466
00:23:08,680 --> 00:23:11,839
the power variation during the training run. And then there

467
00:23:11,839 --> 00:23:15,480
were a bunch of networking challenges. Is the networking cables.

468
00:23:15,759 --> 00:23:18,599
If you're trying to make one hundred thousand GPUs training

469
00:23:18,640 --> 00:23:19,920
coherently of very challenging.

470
00:23:20,160 --> 00:23:22,519
Speaker 2: Almost it sounds almost any of those things you mentioned.

471
00:23:22,880 --> 00:23:26,640
I could imagine someone telling you very directly, no, you

472
00:23:26,680 --> 00:23:29,079
can't have that, you can't have that power, you can't

473
00:23:29,119 --> 00:23:31,599
have this. And it sounds like one of the salient

474
00:23:31,640 --> 00:23:35,559
pieces of first principles thinking is actually, let's ask why,

475
00:23:35,720 --> 00:23:39,079
let's figure that out, and actually let's challenge the person

476
00:23:39,240 --> 00:23:42,640
across the table and they If I don't get an

477
00:23:42,680 --> 00:23:45,559
answer that I feel good about, I'm going to not

478
00:23:45,680 --> 00:23:46,559
allow that to be.

479
00:23:46,839 --> 00:23:49,039
Speaker 4: I'm not going to let that know to stand. Is

480
00:23:49,079 --> 00:23:49,920
that that feels like.

481
00:23:49,880 --> 00:23:53,960
Speaker 2: Something that everyone, if someone were to try to do

482
00:23:54,039 --> 00:23:56,960
what you're doing in hardware, seems to uniquely need this.

483
00:23:57,319 --> 00:24:00,680
In software, we have lots of fluff things that it's

484
00:24:00,680 --> 00:24:03,720
like we can add more CPUs for that, it'll be fine,

485
00:24:03,799 --> 00:24:05,759
But in hardware it's just not going to work.

486
00:24:06,279 --> 00:24:09,519
Speaker 3: I think these general principles of first principle thinking apply

487
00:24:09,640 --> 00:24:11,880
to software and hardware apply to anything. Really.

488
00:24:12,319 --> 00:24:15,440
Speaker 1: I'm just using a hardware example of how we were

489
00:24:15,480 --> 00:24:17,759
told something is impossible. But once we broke it down

490
00:24:17,839 --> 00:24:20,240
into the constituent elements of we need a building, we

491
00:24:20,279 --> 00:24:23,359
need power, we need cooling, we need we need power smoothing,

492
00:24:23,680 --> 00:24:26,920
and then we could solve those constituent elements. But and

493
00:24:26,920 --> 00:24:29,839
then we just ran the networking operation to do all

494
00:24:29,880 --> 00:24:32,839
the cabling everything in four shifts twenty four to seven,

495
00:24:33,279 --> 00:24:35,680
and I was like sleeping the data center and also

496
00:24:35,720 --> 00:24:38,200
doing cabling myself, and there were a lot of other

497
00:24:38,279 --> 00:24:40,960
issues to solve. Nobody had done a training runt with

498
00:24:41,000 --> 00:24:45,039
one hundred thousand, h one hundreds training coherently last year.

499
00:24:45,200 --> 00:24:46,920
Speaker 3: I mean maybe it's been done this year. I don't know.

500
00:24:47,480 --> 00:24:48,759
And then we ended up doubling that.

501
00:24:48,960 --> 00:24:51,920
Speaker 1: To two hundred thousand, and so now we've got one

502
00:24:52,000 --> 00:24:54,240
hundred and fifty thousand, h one hundred fifty k H

503
00:24:54,319 --> 00:24:58,640
two hundreds and thirty k GB two hundreds in the

504
00:24:58,799 --> 00:25:02,160
Memphis training cent and we're about to bring one hundred

505
00:25:02,160 --> 00:25:05,440
and ten thousand GB two hundreds online at a second

506
00:25:05,519 --> 00:25:07,240
data center, also in Memphis.

507
00:25:07,240 --> 00:25:07,519
Speaker 3: Area.

508
00:25:08,200 --> 00:25:10,240
Speaker 2: Is it your view that pre training is still working

509
00:25:10,920 --> 00:25:16,160
and larger the scaling laws still hold, and whoever wins

510
00:25:16,400 --> 00:25:19,920
this race will have basically the biggest, smartest possible model

511
00:25:19,960 --> 00:25:22,000
that you could distill.

512
00:25:22,000 --> 00:25:27,119
Speaker 1: There's other various elements that side competitiveness for large AI

513
00:25:27,799 --> 00:25:31,519
is for sure, the talent of the people matter. The

514
00:25:31,599 --> 00:25:34,400
scale of the hardware matters, and how well you're able

515
00:25:34,440 --> 00:25:36,279
to bring that hardware to bear. So you can't just

516
00:25:36,400 --> 00:25:39,680
order a whole bunch of GPUs and they don't. You

517
00:25:39,680 --> 00:25:42,160
can't just plug them in, So you've got to get

518
00:25:42,200 --> 00:25:45,519
a lot of gps and have them training coherently and stably.

519
00:25:46,599 --> 00:25:49,000
Then it's what unique access to data do you have.

520
00:25:49,480 --> 00:25:52,400
I guess distribution matters to some degree as well, like

521
00:25:52,440 --> 00:25:54,519
how do people get exposed to your AI? Those are

522
00:25:54,559 --> 00:25:57,039
those are critical pactores for if it's going to be

523
00:25:57,079 --> 00:26:01,119
like a large foundation model as competitive as many have said.

524
00:26:01,119 --> 00:26:04,240
I think my friend eliasit Skier said, we would run

525
00:26:04,279 --> 00:26:07,559
out of pre training data of human generated Like human

526
00:26:07,599 --> 00:26:11,160
generated data, you run out of tokens pretty fast, certainly

527
00:26:11,160 --> 00:26:15,279
of high quality tokens, and then you have to do

528
00:26:16,720 --> 00:26:20,079
a lot of You need to essentially create synthetic data

529
00:26:20,680 --> 00:26:23,079
and be able to accurately judge the synthetic data that

530
00:26:23,119 --> 00:26:26,319
you're creating to like, is this real synthetic data? Well,

531
00:26:26,400 --> 00:26:31,200
is it an hallucination that doesn't actually mash reality? So

532
00:26:31,240 --> 00:26:33,440
achieving grounding and reality is tricky, but we are at

533
00:26:33,480 --> 00:26:38,200
the stage where there's more effort put into synthetic data.

534
00:26:39,119 --> 00:26:42,279
Right now, we're training Grock three point five, which is

535
00:26:42,319 --> 00:26:43,839
a heavy focus on reasoning.

536
00:26:44,279 --> 00:26:46,240
Speaker 2: Going back to your physics point, what I heard for

537
00:26:46,400 --> 00:26:50,920
reasoning is that hard science, particularly physics textbooks, are very

538
00:26:51,000 --> 00:26:54,319
useful for reasoning, whereas I think researchers have told me

539
00:26:54,400 --> 00:26:57,519
that social science is totally useless for reasoning.

540
00:26:58,920 --> 00:27:01,440
Speaker 3: Yes, that's probably true.

541
00:27:01,720 --> 00:27:04,839
Speaker 1: So yeah, there's something that's going to be very important

542
00:27:04,839 --> 00:27:09,319
in the future is combining deep AI in the data

543
00:27:09,319 --> 00:27:14,839
center of Supercluster with robotics. So that's things like the

544
00:27:14,839 --> 00:27:18,640
Optimius humanoid robot, and yeah, Optimist is awesome. There's going

545
00:27:18,680 --> 00:27:21,480
to be so many humanoid robots and robots of all

546
00:27:21,480 --> 00:27:23,799
sizes and shapes. But my prediction is that there will

547
00:27:23,799 --> 00:27:27,359
be more humanoid robots by far than all other robots

548
00:27:27,359 --> 00:27:31,039
combined by maybe an order of magnitude, like a big difference,

549
00:27:31,839 --> 00:27:32,839
and is it true that.

550
00:27:32,839 --> 00:27:34,759
Speaker 4: You we're planning a robot army of.

551
00:27:34,720 --> 00:27:38,640
Speaker 1: The sort, whether we do it or whether the Tesla

552
00:27:38,680 --> 00:27:42,119
does it, because it works closely with XAI. You've seen

553
00:27:42,119 --> 00:27:44,519
how many humanoid robot startups are there. It's I think

554
00:27:44,599 --> 00:27:48,720
Jensen bongs on stage with a lot with a massive

555
00:27:48,799 --> 00:27:52,599
number of robots, robots from different companies. I think it

556
00:27:52,680 --> 00:27:57,160
was like a dozen different humanoid robots. I guess part

557
00:27:57,200 --> 00:27:59,440
of what I've been fighting and maybe what has slowed

558
00:27:59,440 --> 00:28:02,079
me down somewhat that I'm a little I don't want

559
00:28:02,079 --> 00:28:07,279
to make Terminator real, you know, So I I guess,

560
00:28:07,920 --> 00:28:11,359
at least until recent years, dragging my feet on on

561
00:28:11,480 --> 00:28:15,279
AI and humanoid robotics, and then I come to realize

562
00:28:15,319 --> 00:28:16,920
that realization it's happening.

563
00:28:16,599 --> 00:28:19,880
Speaker 3: Whether I do it or not. So you got really

564
00:28:19,920 --> 00:28:21,240
two choices participate.

565
00:28:21,440 --> 00:28:24,440
Speaker 1: You could either be a spectator or a participant, And

566
00:28:24,519 --> 00:28:27,000
so I guess I'd rather be a participant than a spectator.

567
00:28:28,079 --> 00:28:30,880
So now it's pedal to the metal on humanoid robots

568
00:28:30,920 --> 00:28:32,559
and digital superintelligence.

569
00:28:33,279 --> 00:28:35,559
Speaker 2: So I guess there's a third thing that everyone has

570
00:28:35,599 --> 00:28:37,440
heard you talk a lot about that I'm really a

571
00:28:37,440 --> 00:28:40,160
big fan of becoming a multiplanetary species.

572
00:28:40,640 --> 00:28:41,519
Speaker 4: Where does this fit?

573
00:28:41,599 --> 00:28:43,839
Speaker 2: This is all not just a ten or twenty year thing,

574
00:28:43,880 --> 00:28:46,640
maybe one hundred year thing, like it's a multi many

575
00:28:46,680 --> 00:28:48,720
generations for humanity kind of thing.

576
00:28:49,400 --> 00:28:50,440
Speaker 4: How do you think about it?

577
00:28:50,519 --> 00:28:55,119
Speaker 2: There's AI obviously, there's embodied robotics, and then there's being

578
00:28:55,160 --> 00:28:59,240
a multiplant, multiplanetary species. Does everything feed into that last

579
00:28:59,279 --> 00:29:02,000
point or what are you driven by right now for

580
00:29:02,039 --> 00:29:04,240
the next ten to twenty one hundred years?

581
00:29:04,759 --> 00:29:07,640
Speaker 1: Jeez, one hundred years, man. I hope civilizations around in

582
00:29:07,680 --> 00:29:10,240
one hundred years. If it is around, it's going to

583
00:29:10,240 --> 00:29:13,960
look very different from civilization today. Not predict that this

584
00:29:14,680 --> 00:29:18,240
going to be at least five times as many humanoid

585
00:29:18,319 --> 00:29:22,799
robots as there are humans, maybe ten times. One way

586
00:29:22,799 --> 00:29:26,519
to look at the progress of civilization is percentage completion Kardaschev.

587
00:29:27,279 --> 00:29:29,960
So if you're in a kars Chef scale one, you've

588
00:29:30,440 --> 00:29:33,240
you've honessed all the energy of a planet. Now, in

589
00:29:33,240 --> 00:29:36,400
my opinion, we've only honested maybe one or two percent

590
00:29:36,680 --> 00:29:39,440
of Earth's energy, So we've got a long way to

591
00:29:39,480 --> 00:29:43,559
go to be Kardschev scale one. Then Cardship two you've

592
00:29:43,720 --> 00:29:47,000
honest only energy of the Sun, which would be I

593
00:29:47,000 --> 00:29:50,279
don't know a billion times more energy than Oath, maybe

594
00:29:50,440 --> 00:29:53,519
closer to a trillion, and then Karschev three would be

595
00:29:53,519 --> 00:29:59,160
only energy of a galaxy pretty far from that. So

596
00:29:59,319 --> 00:30:02,480
we're at the very early stage of the intelligence big bang.

597
00:30:03,000 --> 00:30:06,599
I hope we're in terms of being multiplanetary, like I think,

598
00:30:08,440 --> 00:30:12,799
I think we'll have enough mass transferred to Mars within

599
00:30:12,920 --> 00:30:15,880
like roughly thirty years to make Mars self sustaining such

600
00:30:15,920 --> 00:30:19,279
that Mars can continue to grow and prosper even if

601
00:30:19,319 --> 00:30:22,960
the resupply shifts from both stuff coming, and that greatly

602
00:30:23,000 --> 00:30:28,480
increases the probable lifespan of civilization or consciousness and intelligence,

603
00:30:28,799 --> 00:30:33,279
both biological and digital. So that's why I think it's

604
00:30:33,279 --> 00:30:35,000
important to become a multiplanet species.

605
00:30:35,079 --> 00:30:35,880
Speaker 3: And I'm so much.

606
00:30:35,720 --> 00:30:38,160
Speaker 1: Troubled by the filmy paradox, like why have we not

607
00:30:38,480 --> 00:30:42,640
seen any aliens? And it could be because intelligence is

608
00:30:42,640 --> 00:30:46,720
incredibly rare, and maybe we're the only ones in this galaxy,

609
00:30:47,480 --> 00:30:51,799
in which case the intelligence of consciousness is just the

610
00:30:52,160 --> 00:30:55,440
tiny candle in a vast dokness, and we should do

611
00:30:55,519 --> 00:30:58,759
everything possible to ensure the tiny candle does not go out.

612
00:30:59,559 --> 00:31:04,960
And being a multiplanet species or making consciousness multiplanetary greatly

613
00:31:04,960 --> 00:31:08,400
improves the probable lifespan of civilization, and it's the next

614
00:31:08,440 --> 00:31:12,359
step before going to other star systems. Once you at

615
00:31:12,440 --> 00:31:14,720
least have two planets, then you've got a forcing function

616
00:31:14,799 --> 00:31:18,640
for the improvement of space travel, and that ultimately is

617
00:31:18,680 --> 00:31:22,960
what will lead to consciousness expanding to the stars.

618
00:31:23,559 --> 00:31:26,720
Speaker 2: It could be that the Fermi paradox dictates once you

619
00:31:26,759 --> 00:31:29,680
get to some level of technology, you destroy yourself.

620
00:31:29,960 --> 00:31:31,160
Speaker 4: How do we see ourselves?

621
00:31:31,279 --> 00:31:35,160
Speaker 2: How do we actually what would you prescribe to a

622
00:31:35,240 --> 00:31:37,599
room full of engineers? What can we do to prevent

623
00:31:37,680 --> 00:31:38,559
that from happening?

624
00:31:39,319 --> 00:31:39,480
Speaker 3: Yeah?

625
00:31:39,519 --> 00:31:41,960
Speaker 1: How do we avoid the great filters? One of the

626
00:31:41,960 --> 00:31:45,400
great filtres would obviously be global thom nuclear war, So

627
00:31:45,440 --> 00:31:49,319
we should try to avoid that. I guess building benign

628
00:31:49,400 --> 00:31:54,920
AI robots that AI that loves humanity, and robots that

629
00:31:54,960 --> 00:31:58,559
are helpful. Something that I think is extremely important in

630
00:31:58,599 --> 00:32:02,599
building AI is a very rigorous adherence to truth, even

631
00:32:02,640 --> 00:32:08,039
if that truth is politically incorrect. My intuition for what

632
00:32:08,200 --> 00:32:12,240
could make AI very dangerous is if you force AI

633
00:32:12,319 --> 00:32:13,680
to believe things that are not true.

634
00:32:14,039 --> 00:32:17,359
Speaker 2: How do you think about there's this argument for open

635
00:32:17,400 --> 00:32:20,839
for safety versus closed for competitive edge. I think the

636
00:32:20,880 --> 00:32:23,359
great thing is you have a competitive model. Many other

637
00:32:23,400 --> 00:32:26,839
people also have competitive models, and in that sense we're

638
00:32:26,920 --> 00:32:29,680
off of. Maybe the worst timeline that I'd be worried

639
00:32:29,680 --> 00:32:32,039
about is there's fast takeoff and it's only in one

640
00:32:32,039 --> 00:32:36,839
person's hands that might collapse a lot of things, Whereas

641
00:32:37,200 --> 00:32:39,599
now we have choice, which is great. How do you

642
00:32:39,640 --> 00:32:40,200
think about this?

643
00:32:40,400 --> 00:32:45,599
Speaker 1: Yeah, I do think there will be several deep intelligences,

644
00:32:46,000 --> 00:32:49,640
maybe maybe at least five, maybe as much as ten.

645
00:32:50,799 --> 00:32:52,440
I'm not sure that there's going to be hundreds, but

646
00:32:52,480 --> 00:32:54,599
it's probably close. Like maybe there'll be like ten or

647
00:32:54,599 --> 00:32:58,400
something like that, of which maybe four will be in

648
00:32:58,440 --> 00:33:00,519
the US so.

649
00:33:02,920 --> 00:33:03,759
Speaker 3: And I don't think it's going.

650
00:33:03,759 --> 00:33:08,839
Speaker 1: To be anyone AI that has run away capability. But yeah,

651
00:33:09,119 --> 00:33:10,680
several deep intelligences.

652
00:33:11,119 --> 00:33:14,319
Speaker 2: What will these deep intelligences actually be doing? Will it

653
00:33:14,359 --> 00:33:16,759
be scientific research or trying to hack each other?

654
00:33:18,119 --> 00:33:19,160
Speaker 3: Probably old of the above.

655
00:33:19,359 --> 00:33:22,359
Speaker 1: Hopefully they will discover new physics, and I think they're

656
00:33:22,440 --> 00:33:27,279
very they're definitely going to invent new technologies. I think

657
00:33:27,279 --> 00:33:33,279
we're quite close to digital superintelligence. It may happen this year,

658
00:33:33,480 --> 00:33:35,720
and if it doesn't happen this next year, for sure,

659
00:33:36,319 --> 00:33:40,559
a digital superintelligence to find as smarter than any human

660
00:33:40,640 --> 00:33:42,160
at anything, So.

661
00:33:42,079 --> 00:33:45,119
Speaker 2: How do we direct that to super abundance. We could

662
00:33:45,160 --> 00:33:49,559
have robotic labor, we have cheap energy, intelligence on demand.

663
00:33:50,119 --> 00:33:52,400
Is that sort of the white pill? Like, where do

664
00:33:52,440 --> 00:33:55,720
you sit on the spectrum? And are there tangible things

665
00:33:55,759 --> 00:33:59,440
that you would encourage everyone here to be working on

666
00:33:59,599 --> 00:34:01,839
to make that white pill actually reality?

667
00:34:03,839 --> 00:34:07,119
Speaker 1: I think it most likely will be a good outcome.

668
00:34:07,839 --> 00:34:10,320
I guess I'd agree with Jeff Hinton that maybe it's

669
00:34:10,320 --> 00:34:13,159
a ten to twenty percent chance of annihilation, but look

670
00:34:13,199 --> 00:34:15,639
on the right side, that's eighty to ninety percent probably

671
00:34:15,760 --> 00:34:20,519
of a great outcome. So yeah, I can't emphasize this enough.

672
00:34:20,599 --> 00:34:24,159
A rigorous adherence to truth. It is the most important

673
00:34:24,159 --> 00:34:29,039
thing for AI safety safety obviously, empathy for humanity and

674
00:34:29,119 --> 00:34:30,000
life as we know it.

675
00:34:30,800 --> 00:34:34,639
Speaker 2: We haven't talked about neuralink at all yet, but I'm curious.

676
00:34:34,920 --> 00:34:37,760
You're working on closing the input and output gap between

677
00:34:37,960 --> 00:34:42,800
humans and genes. How critical is that to AI ASI?

678
00:34:42,880 --> 00:34:45,800
And once that link is made, can we not only

679
00:34:45,880 --> 00:34:47,119
read but also write?

680
00:34:47,840 --> 00:34:52,199
Speaker 1: The neuralink is not necessary to solve digital superintelligence. That'll

681
00:34:52,239 --> 00:34:58,159
happen before neuralink is at scale, But nor neuralink can

682
00:34:58,199 --> 00:35:02,159
effectively do is solve the input output bandwidth constraints.

683
00:35:02,760 --> 00:35:06,480
Speaker 3: Especially our output bandwidth is very low. The out the

684
00:35:06,599 --> 00:35:08,280
sustained output.

685
00:35:08,159 --> 00:35:10,880
Speaker 1: Of a human over the course of a day is

686
00:35:10,960 --> 00:35:13,840
less than one but per second, so there's eighty six

687
00:35:14,159 --> 00:35:16,920
four hundred seconds in a day, and it's extremely rare

688
00:35:17,039 --> 00:35:19,239
for a human to output more than that number of

689
00:35:19,280 --> 00:35:22,079
symbols per day, simainly for several days in a row.

690
00:35:22,440 --> 00:35:26,760
So you really within your link interface you can massively

691
00:35:26,800 --> 00:35:30,199
increase your output bandwidth and your input band with input

692
00:35:30,280 --> 00:35:33,280
being right to you have to do right operations to

693
00:35:33,400 --> 00:35:37,920
the brain. We have now five humans who have received

694
00:35:37,920 --> 00:35:42,639
the kind of the read input where it's reading signals,

695
00:35:42,679 --> 00:35:47,000
and you've got people with als who really have their tetraplegics,

696
00:35:47,039 --> 00:35:50,800
but they can now communicate at with a similar bandwidth

697
00:35:50,840 --> 00:35:54,239
to a human with a fully functioning body and control

698
00:35:54,280 --> 00:35:57,599
their computer and the phone, which is pretty cool. And

699
00:35:57,639 --> 00:36:00,400
then I think in the next six to twel months

700
00:36:00,440 --> 00:36:03,559
we will be doing our first implants for vision where

701
00:36:03,760 --> 00:36:06,719
even if somebody is completely blind, we can write directly

702
00:36:06,760 --> 00:36:11,960
to the visual cortex, and we've had that working in monkeys. Actually,

703
00:36:12,039 --> 00:36:14,599
I think one of our monkeys now has had a

704
00:36:14,679 --> 00:36:18,440
visual implant for three years, and at first it'll be

705
00:36:18,480 --> 00:36:21,800
relatively fairly low resolution, but long term you would have

706
00:36:21,880 --> 00:36:26,760
very high resolution and be able to see multi spectral wavelengths.

707
00:36:26,920 --> 00:36:30,599
You could see an infrared ultra violet radar like a

708
00:36:30,639 --> 00:36:34,360
superpower situation, but like at some point the cybernetic implants

709
00:36:34,719 --> 00:36:37,719
would not simply be correcting things that went wrong, but

710
00:36:38,000 --> 00:36:46,280
augmenting human capabilities dramatically, augmenting intelligence and sensors and bandwidth dramatically.

711
00:36:46,440 --> 00:36:48,800
And that's going to happen at some point, but digital

712
00:36:48,800 --> 00:36:52,840
superintelligence will happen well before that. At least if we're

713
00:36:53,440 --> 00:36:57,920
a neur link, we'll be able to appreciate the AI better.

714
00:36:59,079 --> 00:37:01,440
Speaker 2: I guess one of the limits reagents to all of

715
00:37:01,440 --> 00:37:05,480
your efforts across all of these different domains is access to.

716
00:37:05,480 --> 00:37:06,880
Speaker 4: The smartest possible people.

717
00:37:07,519 --> 00:37:11,840
Speaker 2: Yes, simultaneous to that, we have the rocks can talk

718
00:37:11,880 --> 00:37:14,920
and reason, and they're maybe one hundred and thirty IQ now,

719
00:37:14,960 --> 00:37:19,039
and they're probably going to be super intelligent soon. How

720
00:37:19,039 --> 00:37:21,440
do you reconcile those two things. What's going to happen

721
00:37:21,519 --> 00:37:24,079
in five ten years? And what should the people in

722
00:37:24,119 --> 00:37:26,880
this room do to make sure that they're the ones

723
00:37:26,880 --> 00:37:30,159
who are creating instead of maybe below the API line.

724
00:37:31,079 --> 00:37:33,440
Speaker 1: They called it the singularity for a reason, because we

725
00:37:33,480 --> 00:37:36,239
don't know what's going to happen in the not that

726
00:37:36,559 --> 00:37:40,079
far future. The percentage of intelligence that is human will

727
00:37:40,079 --> 00:37:43,199
be quite small. At some point, the collective some of

728
00:37:43,280 --> 00:37:46,480
human intelligence will be less than one percent of all intelligence.

729
00:37:48,079 --> 00:37:50,840
And if things get to a clot a ship level two,

730
00:37:51,880 --> 00:37:56,599
we're talking about human intelligence. Even assuming a significant increase

731
00:37:56,599 --> 00:38:03,079
in human population and intelligence augmentation like massive intelligence augmentation

732
00:38:03,119 --> 00:38:05,159
where everyone has an IQ of one thousand type of thing.

733
00:38:05,760 --> 00:38:09,199
Even in that circumstance, collective human intelligence will be probably

734
00:38:09,199 --> 00:38:14,400
one billion that of digital intelligence. Anyway, we're the biological

735
00:38:14,440 --> 00:38:16,400
boot bootloader for digital superintelligence.

736
00:38:17,159 --> 00:38:21,719
Speaker 4: I guess just to end off, was I?

737
00:38:22,039 --> 00:38:23,559
Speaker 3: It was like, was I a good bootload?

738
00:38:23,639 --> 00:38:25,239
Speaker 4: Where do we go? How do we go from here?

739
00:38:25,280 --> 00:38:27,599
Speaker 2: All of this is pretty wild sci fi stuff that

740
00:38:27,719 --> 00:38:32,079
also could be built by the people in this room.

741
00:38:32,360 --> 00:38:35,159
But if you do, you have a closing thought for

742
00:38:35,199 --> 00:38:39,719
the smartest technical people of this generation right now? What

743
00:38:39,760 --> 00:38:41,760
should they be doing, what should they be working on?

744
00:38:41,800 --> 00:38:44,079
What should they be thinking about tonight as they go

745
00:38:44,119 --> 00:38:44,599
to dinner?

746
00:38:47,119 --> 00:38:49,000
Speaker 1: As I saw a rope with I think if you're

747
00:38:49,000 --> 00:38:52,599
doing something useful, that's great. If you just try to

748
00:38:52,639 --> 00:38:55,079
be as useful as possible to your fellow human beings

749
00:38:55,679 --> 00:38:58,480
and that then you're doing something good. I keep harping

750
00:38:58,480 --> 00:39:02,559
on this, like focus on truthful AI. That's the most

751
00:39:02,559 --> 00:39:06,320
important thing for AI safety. Obviously, if anyone's interested in

752
00:39:06,360 --> 00:39:09,480
working at xailet mean, please let us know. We're aiming

753
00:39:09,519 --> 00:39:13,239
to make GROT the maximumly truth seeking AI, and I

754
00:39:13,280 --> 00:39:16,239
think that's a very important thing. Hopefully we can understand

755
00:39:16,280 --> 00:39:19,599
the nature of the universe, that's really what AI can

756
00:39:20,360 --> 00:39:22,639
hopefully tell us. Maybe AI can maybe tell us where

757
00:39:22,679 --> 00:39:26,440
are the aliens, and how did the universe really start?

758
00:39:26,880 --> 00:39:27,679
Speaker 3: How will it end?

759
00:39:28,159 --> 00:39:30,400
Speaker 1: What are the questions that we don't know that we

760
00:39:30,400 --> 00:39:35,039
should ask, And are we in a simulation or what

761
00:39:35,119 --> 00:39:35,519
level of.

762
00:39:35,440 --> 00:39:36,440
Speaker 3: Similation are we in?

763
00:39:38,440 --> 00:39:43,559
Speaker 2: I think we're going to find out. APC, Elon, thank

764
00:39:43,559 --> 00:39:47,400
you so much for joining us. Everyone please give it

765
00:39:47,480 --> 00:39:48,239
up for Elon Mott