1
00:00:01,000 --> 00:00:02,940
The stream is live. All right.

2
00:00:03,220 --> 00:00:03,840
We're live.

3
00:00:04,200 --> 00:00:05,920
Video Bitcoin Jam. How are you guys?

4
00:00:06,380 --> 00:00:08,020
Great. We can see all three of us.

5
00:00:08,200 --> 00:00:09,100
Yeah, we're all here now.

6
00:00:09,320 --> 00:00:12,060
It wouldn't be 1210 on a Friday if we weren't.

7
00:00:12,060 --> 00:00:12,320
Exactly.

8
00:00:14,880 --> 00:00:20,400
We miss you, DK. We're all figuring out the IT and whatever setup situation.

9
00:00:20,740 --> 00:00:22,420
So we can't wait to have you back with us.

10
00:00:22,860 --> 00:00:24,920
I'm not the expert. I just take up space.

11
00:00:27,280 --> 00:00:29,200
He's modest. He's definitely the expert in this group.

12
00:00:30,000 --> 00:00:32,560
are you gonna be in studio next week

13
00:00:32,560 --> 00:00:35,020
I think next week

14
00:00:35,020 --> 00:00:36,440
no but the week after that yes

15
00:00:36,440 --> 00:00:38,820
so one more away week

16
00:00:38,820 --> 00:00:40,280
until I'm back

17
00:00:40,280 --> 00:00:42,260
okay beautiful we'll be patient

18
00:00:42,260 --> 00:00:43,320
yeah exactly

19
00:00:43,320 --> 00:00:46,700
I'm excited to see you guys in person some more

20
00:00:46,700 --> 00:00:48,500
because I feel like when we're in person

21
00:00:48,500 --> 00:00:50,600
we always are generating a lot of stuff

22
00:00:50,600 --> 00:00:52,340
they're like hey let's talk about that on the show

23
00:00:52,340 --> 00:00:52,980
let's talk about that

24
00:00:52,980 --> 00:00:56,460
and we've got our group chat teed up

25
00:00:56,460 --> 00:00:58,540
with a lot of stuff but it's also fun to

26
00:00:58,540 --> 00:01:01,620
fun to bump into you guys in person because i think we actually ended up

27
00:01:01,620 --> 00:01:06,980
generating more stuff that way well in the spirit of irl generation i was just telling steve before

28
00:01:06,980 --> 00:01:14,160
we came in here um there is at least one probably many big brains at the presidio bitcoin office

29
00:01:14,160 --> 00:01:20,540
uh i'm not going to dox who it was possibly steve right here but someone was scribbling on a white

30
00:01:20,540 --> 00:01:27,260
board uh would look to be some like fairly sophisticated cryptography and using nip 17 dms

31
00:01:27,260 --> 00:01:30,160
to communicate with MCP servers.

32
00:01:30,320 --> 00:01:31,640
So I don't want to give away too much alpha.

33
00:01:31,860 --> 00:01:33,280
Also, I didn't understand it all myself

34
00:01:33,280 --> 00:01:34,200
and I'll have to go back and check.

35
00:01:34,320 --> 00:01:36,720
But it's a lot like that scene in Good Will Hunting

36
00:01:36,720 --> 00:01:37,680
where you're just walking by

37
00:01:37,680 --> 00:01:38,940
or maybe in a beautiful mind,

38
00:01:39,020 --> 00:01:40,160
someone just scribbling on the wall.

39
00:01:40,220 --> 00:01:40,760
You're like, shit,

40
00:01:41,160 --> 00:01:42,260
did we just find some breakthrough?

41
00:01:42,660 --> 00:01:44,760
This was the worst doxing ever

42
00:01:44,760 --> 00:01:46,720
because you actually said something

43
00:01:46,720 --> 00:01:48,180
and then you coughed and I didn't hear it,

44
00:01:48,200 --> 00:01:49,920
but I think everybody else heard it

45
00:01:49,920 --> 00:01:51,820
or it's on the recording in Decipherable.

46
00:01:51,940 --> 00:01:52,960
So can you just say who it was?

47
00:01:53,340 --> 00:01:54,740
Oh, I actually don't know who it was.

48
00:01:54,840 --> 00:01:56,100
He was joking that it was me.

49
00:01:56,100 --> 00:01:59,660
And if it was me, I must have done it in my sleep last night.

50
00:02:00,100 --> 00:02:01,100
Otherwise, I don't think it was.

51
00:02:01,140 --> 00:02:03,600
It was definitely someone that's deep in cryptography.

52
00:02:04,000 --> 00:02:05,120
So we have some guesses.

53
00:02:05,560 --> 00:02:08,520
But yeah, just another cool thing at Presidio Bitcoin.

54
00:02:08,640 --> 00:02:09,740
You never know what you're going to find here.

55
00:02:10,080 --> 00:02:12,180
So there's the outlet for all the listeners.

56
00:02:12,980 --> 00:02:15,020
And I think there was a builder last night, yeah?

57
00:02:15,560 --> 00:02:16,020
There was.

58
00:02:16,300 --> 00:02:16,760
There was.

59
00:02:16,920 --> 00:02:20,660
Anybody, do you want to give a little overview of what happened there?

60
00:02:21,840 --> 00:02:23,220
Another good turnout.

61
00:02:23,220 --> 00:02:24,400
like several

62
00:02:24,400 --> 00:02:26,980
several new folks each time

63
00:02:26,980 --> 00:02:27,740
which is great

64
00:02:27,740 --> 00:02:31,120
and let's see we had

65
00:02:31,120 --> 00:02:33,080
Damien

66
00:02:33,080 --> 00:02:34,780
do a live vibe coding

67
00:02:34,780 --> 00:02:37,100
kind of like you and Matt

68
00:02:37,100 --> 00:02:39,140
DK and Matt's live

69
00:02:39,140 --> 00:02:39,700
vibe coding

70
00:02:39,700 --> 00:02:42,500
the Twitch you've done a couple times

71
00:02:42,500 --> 00:02:45,120
because Damien came unprepared

72
00:02:45,120 --> 00:02:47,140
or like we took audience suggestions

73
00:02:47,140 --> 00:02:48,600
so he did it was like

74
00:02:48,600 --> 00:02:49,520
unplanned

75
00:02:49,520 --> 00:02:52,940
which is even harder I think than like what you and Matt

76
00:02:52,940 --> 00:02:54,660
are doing because I think you guys

77
00:02:54,660 --> 00:02:56,480
at least have an idea of what you want to build

78
00:02:56,480 --> 00:02:58,880
ahead of time and then you try it live.

79
00:02:59,980 --> 00:03:01,320
He didn't know ahead of time

80
00:03:01,320 --> 00:03:02,380
and he tried it live.

81
00:03:03,300 --> 00:03:04,380
So I think that was

82
00:03:04,380 --> 00:03:06,800
you know, it didn't result

83
00:03:06,800 --> 00:03:08,800
in exactly what we were hoping

84
00:03:08,800 --> 00:03:10,740
to build, but it definitely

85
00:03:10,740 --> 00:03:12,700
made progress and I think it was

86
00:03:12,700 --> 00:03:14,880
educational for people. So I think that was compelling.

87
00:03:15,520 --> 00:03:16,720
And then it was great. One of the audience

88
00:03:16,720 --> 00:03:18,480
members, someone who

89
00:03:18,480 --> 00:03:20,800
Pamela who's come to all of them

90
00:03:20,800 --> 00:03:22,700
I don't believe

91
00:03:22,700 --> 00:03:26,620
she has like a development background so she she did some vibe coding and then she showed showed

92
00:03:26,620 --> 00:03:31,760
off her vibe coding so i think that's super cool to give like people new to this experience uh a

93
00:03:31,760 --> 00:03:38,540
little limelight to to highlight that um and then did damien do it with uh replet the way we saw

94
00:03:38,540 --> 00:03:45,860
okay cool which there's pros and cons to that um i mean it's definitely good to show different

95
00:03:45,860 --> 00:03:51,600
tools i think um but like some of the audience members i mean damien had a bunch of like rules

96
00:03:51,600 --> 00:03:53,700
set up and other configurations set up

97
00:03:53,700 --> 00:03:54,280
and a template,

98
00:03:55,280 --> 00:03:58,960
which a more advanced, sophisticated

99
00:03:58,960 --> 00:04:00,500
Vibe Coder would have all that stuff,

100
00:04:00,740 --> 00:04:02,900
and it definitely improves the experience.

101
00:04:04,100 --> 00:04:06,380
But you don't need any of that to start.

102
00:04:07,140 --> 00:04:09,000
So I think it gave the impression

103
00:04:09,000 --> 00:04:11,060
of being more intimidating

104
00:04:11,060 --> 00:04:15,260
than it is, at least to start.

105
00:04:15,640 --> 00:04:16,960
But we spoke to that.

106
00:04:17,300 --> 00:04:20,260
And I think it would be cool in future events

107
00:04:20,260 --> 00:04:21,760
to do, like you said, like Replit.

108
00:04:21,980 --> 00:04:22,980
It seems like that's probably easier

109
00:04:22,980 --> 00:04:24,160
for some of the first timers,

110
00:04:24,280 --> 00:04:26,360
but also stuff like we should do a Goose one in the future

111
00:04:26,360 --> 00:04:27,880
so people can get different flavors.

112
00:04:28,300 --> 00:04:29,760
I think one of the interesting things for me

113
00:04:29,760 --> 00:04:31,340
that I've learned from the Damien Vibe code,

114
00:04:32,380 --> 00:04:33,440
I don't think, did we stream it?

115
00:04:33,580 --> 00:04:34,380
I don't think we streamed it.

116
00:04:34,600 --> 00:04:34,840
Did we?

117
00:04:35,060 --> 00:04:35,260
No.

118
00:04:35,560 --> 00:04:38,000
We now have the capabilities to live stream,

119
00:04:38,220 --> 00:04:41,100
but no one had the skill set that was there last night.

120
00:04:41,300 --> 00:04:42,160
DK, we're missing you, bro.

121
00:04:43,120 --> 00:04:44,380
Max and I joke about the show.

122
00:04:44,380 --> 00:04:49,320
Anytime we have production issues in this show,

123
00:04:49,320 --> 00:04:51,300
Max and I are absolutely useless.

124
00:04:52,380 --> 00:04:54,340
We're just here to look for the other camera.

125
00:04:54,700 --> 00:04:56,040
Like in the AV club.

126
00:05:00,620 --> 00:05:01,480
Where's the IT?

127
00:05:01,660 --> 00:05:02,360
We can't do this.

128
00:05:04,800 --> 00:05:07,720
I haven't discussed this with other people,

129
00:05:08,100 --> 00:05:10,580
such as you guys and Haley and others,

130
00:05:11,240 --> 00:05:11,660
for Builder.

131
00:05:11,840 --> 00:05:12,820
One of my takeaways,

132
00:05:13,320 --> 00:05:15,340
I think we want to cater to beginners

133
00:05:15,340 --> 00:05:16,940
and advanced fiber coders.

134
00:05:17,040 --> 00:05:18,160
I think both are important.

135
00:05:18,160 --> 00:05:27,940
We want to be very welcoming to beginners and patient, but we also want to appeal to advanced

136
00:05:27,940 --> 00:05:32,060
VibeCoders, both because some people are at that advanced stage or close to it.

137
00:05:32,320 --> 00:05:37,980
Even for beginners, even if they're at the beginner stage, I think they can get inspired

138
00:05:37,980 --> 00:05:40,740
by seeing what a more advanced setup can be.

139
00:05:41,420 --> 00:05:46,020
But we don't want to only do advanced because then it feels too intimidating.

140
00:05:46,020 --> 00:05:48,280
So we might want to actually structure it

141
00:05:48,280 --> 00:05:50,040
with a beginner track and advanced track.

142
00:05:50,180 --> 00:05:53,200
And each month we try to have one example

143
00:05:53,200 --> 00:05:56,680
who shows off their beginner VibeCoding

144
00:05:56,680 --> 00:05:58,920
and celebrate that,

145
00:05:59,060 --> 00:06:00,220
and then a more advanced one.

146
00:06:00,680 --> 00:06:02,700
And then to your point, Max, about the tools,

147
00:06:03,060 --> 00:06:04,800
we should rotate, show all the,

148
00:06:05,300 --> 00:06:07,680
over time, all the different tools should be shown.

149
00:06:07,920 --> 00:06:09,460
Yeah, and so a couple things came out

150
00:06:09,460 --> 00:06:10,820
of the Damien VibeCode for me.

151
00:06:11,560 --> 00:06:12,680
Stuff we can learn for next time,

152
00:06:12,720 --> 00:06:14,740
but also I think just general best practices.

153
00:06:14,740 --> 00:06:16,200
And you probably have a lot of thoughts on this, DK.

154
00:06:16,340 --> 00:06:24,100
But one is, I think we can get better over time at scoping, like what's reasonable to live code in 15, 20 minutes.

155
00:06:24,180 --> 00:06:33,320
I think that's just a learning process of like, for example, Bob had proposed the idea of doing a feed of Noster notes where someone has locked Bitcoin or eCash to it.

156
00:06:33,620 --> 00:06:36,280
I think that probably would have been a little bit simpler and easier.

157
00:06:36,760 --> 00:06:40,620
We ended up instead doing Max's suggestion, which was really cool, doing a blind auction.

158
00:06:40,620 --> 00:06:44,640
Actually, I think this could go totally viral, but a blind auction where the highest bid wins

159
00:06:44,640 --> 00:06:45,400
the entire pot.

160
00:06:46,300 --> 00:06:50,420
So, you know, this is up your alley, DK, the gray area of the web.

161
00:06:51,340 --> 00:06:52,640
It's actually a genius idea.

162
00:06:53,040 --> 00:06:57,000
But I think that was probably a little bit sophisticated for a 15, 20-minute VibeCode,

163
00:06:57,120 --> 00:06:58,140
although he got really close.

164
00:06:58,600 --> 00:07:02,400
I think another big learning for me was just having all of the tooling already set up.

165
00:07:02,520 --> 00:07:04,860
He had some dependencies and things already downloaded.

166
00:07:05,260 --> 00:07:09,640
But as an example, one of the things that he did was he said, okay, he used Albi in their

167
00:07:09,640 --> 00:07:14,240
NWC libraries and everything they've already built for that, Noster Wallet Connect,

168
00:07:14,620 --> 00:07:19,800
those who are currently with NWC. Anyways, and he said, here's the GitHub, just take all the

169
00:07:19,800 --> 00:07:23,740
documentation, which is cool, but a better way to do that in the future would be like, hey,

170
00:07:23,740 --> 00:07:28,980
here's the Albi MCP server, which already has all of that even more AI-friendly and ready to go.

171
00:07:29,360 --> 00:07:34,760
So I think having MCP servers set up. And then I think a third thing is, and maybe this goes back

172
00:07:34,760 --> 00:07:37,280
to the first one a little bit of like when we select ideas,

173
00:07:37,720 --> 00:07:39,980
we did have some cool ideas from Twitter.

174
00:07:40,340 --> 00:07:42,480
For example, Megalith, we tweeted beforehand,

175
00:07:42,780 --> 00:07:44,760
like, hey, who has ideas for what you want to see,

176
00:07:44,860 --> 00:07:45,540
Damien Vibecode?

177
00:07:45,960 --> 00:07:47,420
And he came up with some great ideas,

178
00:07:47,480 --> 00:07:50,400
like could we add Cashew or NWC payments into BitChat,

179
00:07:50,980 --> 00:07:51,640
which would be awesome,

180
00:07:51,740 --> 00:07:53,880
probably a little sophisticated also for a short Vibecode.

181
00:07:54,300 --> 00:07:56,440
But one idea is we could maybe narrow down

182
00:07:56,440 --> 00:07:57,560
like the types of things,

183
00:07:57,560 --> 00:08:00,600
like, hey, here are like five examples of apps Damien agrees

184
00:08:00,600 --> 00:08:02,240
or whoever's Vibecoding could get built.

185
00:08:02,240 --> 00:08:08,760
and like maybe on the fly we pick one from the audience live or like slightly modify it but like

186
00:08:08,760 --> 00:08:12,300
already have some consensus from the vibe coder of like yeah i'm pretty sure i could do that

187
00:08:12,300 --> 00:08:17,360
yep that's probably a good idea i was also gonna say just from some of the experience that

188
00:08:17,360 --> 00:08:23,420
that i've had doing this stuff on the live stream with matt is um you know we we sort of tried a

189
00:08:23,420 --> 00:08:30,860
kind of a bitcoin lightning type of uh uh development process this sort of feature set

190
00:08:30,860 --> 00:08:32,980
with the thing that we were doing this week.

191
00:08:33,180 --> 00:08:34,040
I don't know if you guys saw it,

192
00:08:34,040 --> 00:08:36,220
but we kind of failed to fully accomplish

193
00:08:36,220 --> 00:08:37,400
what we set out to do.

194
00:08:38,180 --> 00:08:41,980
And I think a lot of that is because the, you know,

195
00:08:42,020 --> 00:08:45,700
the lightning Bitcoin, even like the Nostra stuff,

196
00:08:46,000 --> 00:08:49,180
just has such a smaller surface area

197
00:08:49,180 --> 00:08:50,700
of existing implementations

198
00:08:50,700 --> 00:08:53,960
for any kind of coding agent to really learn from

199
00:08:53,960 --> 00:08:56,840
that I think, you know,

200
00:08:56,900 --> 00:08:59,240
if you try to do a web app of some sort,

201
00:08:59,240 --> 00:09:04,220
there's you know how many millions of examples of that it almost always gets that right if you try

202
00:09:04,220 --> 00:09:08,100
to do anything touching nostril lightning bitcoin it almost always gets it wrong and you struggle so

203
00:09:08,100 --> 00:09:13,300
much to do that and i think a good output of either the live streams that matt and i are doing

204
00:09:13,300 --> 00:09:19,940
or stuff that we could do with builder is create like reference github repos open source how to do

205
00:09:19,940 --> 00:09:26,280
the thing because like for example we were trying to use uh lexi the you know non-custodial lightning

206
00:09:26,280 --> 00:09:33,260
wallet and they have a great SDK, but we kept running into similar problems like the way it

207
00:09:33,260 --> 00:09:39,200
would try to import the variable, it would import the variable itself and not the actual secret.

208
00:09:39,800 --> 00:09:44,580
So I'm sure that can be trained out of an LLM or you could augment the knowledge, but be like,

209
00:09:44,580 --> 00:09:49,940
implement Lexi according to the instructions in this repo. And then I bet we could get it to work

210
00:09:49,940 --> 00:09:55,220
first time, but we tried it multiple times, reset, Matt tried it again offline and only got it

211
00:09:55,220 --> 00:09:58,420
working through like hours of struggle literally,

212
00:09:58,920 --> 00:10:00,780
whereas these things need to be very easy

213
00:10:00,780 --> 00:10:02,380
and sort of like you don't have to think about it

214
00:10:02,380 --> 00:10:04,620
if people are going to be able to build on top of them.

215
00:10:04,700 --> 00:10:06,220
They just need to sort of be defaults

216
00:10:06,220 --> 00:10:08,440
the way a lot of the web infrastructure is today.

217
00:10:09,200 --> 00:10:11,660
Yeah, so I think that's a great lesson that you learned.

218
00:10:12,380 --> 00:10:15,400
Agree with what you said about creating,

219
00:10:15,540 --> 00:10:17,920
you know, some repos and examples.

220
00:10:17,920 --> 00:10:20,840
It also, it's like we should give a shout out

221
00:10:20,840 --> 00:10:24,980
to the Lexis, Albies, Sparks, Cashews, CDKs of the world.

222
00:10:25,220 --> 00:10:29,280
they should all be incentivized to create repositories of examples.

223
00:10:29,480 --> 00:10:29,920
100%.

224
00:10:29,920 --> 00:10:32,520
I think Albee's done some, and they've created MCP,

225
00:10:32,800 --> 00:10:34,480
and Fusats has done it as well.

226
00:10:34,480 --> 00:10:39,160
But just like we should promote that to all of these small companies

227
00:10:39,160 --> 00:10:43,600
that are budding ecosystems because they have a vested interest in doing that.

228
00:10:44,020 --> 00:10:47,120
And when things trip up, they kind of trip up in ways that you could argue,

229
00:10:47,240 --> 00:10:48,720
oh, I documented it. It shouldn't have worked.

230
00:10:48,920 --> 00:10:50,560
It should have worked, so it should have figured it out.

231
00:10:51,080 --> 00:10:54,280
And I think rather than a team, like if you're the team building Lexi

232
00:10:54,280 --> 00:10:56,420
or the team building Spark or Albi,

233
00:10:56,920 --> 00:10:58,760
instead of saying it's documented, it should work,

234
00:10:59,300 --> 00:11:00,940
I believe that there's, you know,

235
00:11:01,080 --> 00:11:02,380
whoever's building that tool

236
00:11:02,380 --> 00:11:06,020
should try to vibe code something from scratch,

237
00:11:06,180 --> 00:11:07,400
assuming that that tool exists.

238
00:11:07,960 --> 00:11:09,340
And I think what you'll find,

239
00:11:09,520 --> 00:11:11,040
like with the Lexi example,

240
00:11:11,120 --> 00:11:12,540
we found the same problem

241
00:11:12,540 --> 00:11:13,940
that if you really read

242
00:11:13,940 --> 00:11:15,180
and understood the documentation well,

243
00:11:15,240 --> 00:11:16,440
you wouldn't screw up,

244
00:11:16,760 --> 00:11:18,980
but the LLMs continually kept screwing it up

245
00:11:18,980 --> 00:11:20,700
in the same kind of predictable ways.

246
00:11:21,120 --> 00:11:22,220
So I think, you know,

247
00:11:22,220 --> 00:11:27,560
the team should, you know, both build the, build an MCP server, build the tool, then vibe code

248
00:11:27,560 --> 00:11:30,680
something with the tool and figure out all of the places where you get stuck. And then write

249
00:11:30,680 --> 00:11:33,880
instructions that says, Hey, if you're an LLM, if you're trying to vibe code with this,

250
00:11:34,620 --> 00:11:38,720
it's frequent that you're going to run into this problem and this problem. Don't fall for this trap.

251
00:11:38,720 --> 00:11:44,700
Do this instead. So you effectively like teach it how to be successful with the process that,

252
00:11:44,800 --> 00:11:48,760
you know, will break. First identify the broken and then tell it how to fix those.

253
00:11:48,760 --> 00:12:03,460
This is bringing up a really good point that you brought up before, DK, which is, you know, we have GitHub today, but that's for humans. What does GitHub for machines look like in the first place? Like you're saying the Lexi team should publish that. Where? Like in GitHub? Like where should they publish it?

254
00:12:03,460 --> 00:12:06,160
I think GitHub would definitely be sufficient for now.

255
00:12:06,760 --> 00:12:09,580
And Matt published a repo after our thing

256
00:12:09,580 --> 00:12:10,640
where he finally got it working.

257
00:12:10,840 --> 00:12:13,860
So I think we do have now a public GitHub

258
00:12:13,860 --> 00:12:16,480
showing how at least one example

259
00:12:16,480 --> 00:12:18,540
of Lexi being properly integrated.

260
00:12:19,260 --> 00:12:22,160
But I think when we did it live, it failed.

261
00:12:22,440 --> 00:12:24,880
And then when we did it, I think he did it offline

262
00:12:24,880 --> 00:12:26,420
and it still failed in the same way.

263
00:12:26,700 --> 00:12:28,080
And then he kind of struggled, struggled

264
00:12:28,080 --> 00:12:28,820
and got it to work.

265
00:12:29,060 --> 00:12:30,940
But if I wanted to start from scratch, say,

266
00:12:31,680 --> 00:12:33,080
I could probably know,

267
00:12:33,080 --> 00:12:36,660
like just because i know matt and i know this process i could know that he has this repo and

268
00:12:36,660 --> 00:12:41,680
find the repo and share it and say don't just do the lexi according to documentation study this

269
00:12:41,680 --> 00:12:47,560
repo where we actually have it working successfully yeah and and then kind of run from there and then

270
00:12:47,560 --> 00:12:52,780
if the teams could actually build those repos and i think over time the lms will figure it out

271
00:12:52,780 --> 00:12:57,460
yeah but but we're not at the state right now where you can just like write good documentation

272
00:12:57,460 --> 00:13:01,920
it will definitely work how did you actually communicate with the lm like did you do was

273
00:13:01,920 --> 00:13:06,400
it an mcp server did you do a separate file that's just like hey ai is this is for you

274
00:13:06,400 --> 00:13:14,080
well so in the case of we didn't use an mcp server for lexi lexi you know has an sdk you can

275
00:13:14,080 --> 00:13:20,880
you can uh have it generate secrets with your mobile wallet and then it has what's called a

276
00:13:20,880 --> 00:13:26,560
sidecar and my understanding is that's just some other server that lexi runs that you know accepts

277
00:13:26,560 --> 00:13:34,040
connections to your lightning node that they set up and manage for you uh but when you basically the

278
00:13:34,040 --> 00:13:39,980
main problem is it was never able to launch the sidecar appropriately and it like sidecars work

279
00:13:39,980 --> 00:13:46,420
and the lexi secret works but it failed in this kind of sort of stupid way but it kept using lexi

280
00:13:46,420 --> 00:13:52,640
underscore you know sidecar underscore sdk id instead of the actual id because it wasn't resolving

281
00:13:52,640 --> 00:13:54,940
the variable for some reason.

282
00:13:55,320 --> 00:13:55,740
I don't know why.

283
00:13:56,340 --> 00:13:57,720
But where did you actually

284
00:13:57,720 --> 00:13:59,600
call the attention of the LLM

285
00:13:59,600 --> 00:14:00,760
and say, hey, if you're an LLM,

286
00:14:01,020 --> 00:14:03,100
read this specific piece of text.

287
00:14:03,220 --> 00:14:05,800
Was that in the sidecar documentation,

288
00:14:05,960 --> 00:14:06,860
like the existing one

289
00:14:06,860 --> 00:14:08,040
you put at the beginning or the end?

290
00:14:08,580 --> 00:14:10,520
No, so we were using Replit

291
00:14:10,520 --> 00:14:11,720
to do the session,

292
00:14:12,200 --> 00:14:14,520
as I think we will probably

293
00:14:14,520 --> 00:14:16,520
continue doing for that particular show.

294
00:14:17,620 --> 00:14:19,140
And so it was just in the thing.

295
00:14:19,240 --> 00:14:20,400
We said, hey, use Lexi.

296
00:14:20,820 --> 00:14:21,640
Here's, I don't know,

297
00:14:21,640 --> 00:14:25,720
there's a thing called Lexi and it first thought it was L-E-X-I. And we had to say, no, no, it's L-E-X-E.

298
00:14:26,040 --> 00:14:29,840
And here's the thing and see if you can find the documentation. It's a lightning Bitcoin thing and

299
00:14:29,840 --> 00:14:33,660
implement it, make it work. So we didn't give it a lot of direction at first, but then once it

300
00:14:33,660 --> 00:14:38,840
started failing, we got deeper into explaining how to make it successful. But it's just kind of a

301
00:14:38,840 --> 00:14:42,400
back and forth with the LLM. Well, and I don't know if you've already talked to the Lexi team.

302
00:14:42,480 --> 00:14:47,240
If you haven't, you should definitely give them this feedback. But if you had access to their

303
00:14:47,240 --> 00:14:52,060
repo like where would you suggest they put the like llm file like hey llm this is the problem

304
00:14:52,060 --> 00:14:58,260
here's how we recommend uh solving it probably could just be some sort of leaf off of their main

305
00:14:58,260 --> 00:15:04,760
documentation and maybe like people who are trying to do it if they ever run into people who have

306
00:15:04,760 --> 00:15:10,040
this problem they could say hey matt dk you want to do this point this leaf of our documentation

307
00:15:10,040 --> 00:15:15,860
to your llm as a starting point for how llm should do it i think over time this is just general

308
00:15:15,860 --> 00:15:21,620
knowledge, you know, reading, writing, coding, it should be solved. But there's just so few

309
00:15:21,620 --> 00:15:26,280
examples, I think, that it keeps running into the same problem. So because there's so few examples,

310
00:15:26,280 --> 00:15:31,400
if we can give a very specific success happy path, I think that can unstick a lot of future

311
00:15:31,400 --> 00:15:38,660
development. To be clear, at builder number two last month, Matt vibe coded with Lexi. That's

312
00:15:38,660 --> 00:15:43,740
under 15 minutes, an entire application. So it's just, I think, this new application,

313
00:15:43,740 --> 00:15:45,440
they tried to vibe code, it just had

314
00:15:45,440 --> 00:15:47,620
a different approach that

315
00:15:47,620 --> 00:15:49,280
you hit some stumbling block.

316
00:15:49,280 --> 00:15:51,300
Totally. No, I mean, I saw that. I was very

317
00:15:51,300 --> 00:15:53,320
impressed. I guess my general point is, like, I don't

318
00:15:53,320 --> 00:15:55,320
yet understand, and maybe your point is just

319
00:15:55,320 --> 00:15:57,300
like the LLMs are smart to figure out

320
00:15:57,300 --> 00:15:59,340
on its own. Like, what's

321
00:15:59,340 --> 00:16:01,380
the protocol to say, hey, LLM, look

322
00:16:01,380 --> 00:16:03,460
here. Like, human, look here.

323
00:16:03,680 --> 00:16:05,120
LLM, look here. Oh, it's just English.

324
00:16:05,560 --> 00:16:07,340
Yeah, the protocol is English. Hey, LLM,

325
00:16:07,440 --> 00:16:09,360
look here. But you think it

326
00:16:09,360 --> 00:16:11,300
should literally be that direct, like, right out.

327
00:16:11,340 --> 00:16:13,380
I think Max needs to vibe code some more.

328
00:16:13,380 --> 00:16:17,020
I think Max just has vibe coding theoretically.

329
00:16:18,080 --> 00:16:19,360
It is kind of funny.

330
00:16:19,560 --> 00:16:20,340
You need to get your hands dirty.

331
00:16:20,820 --> 00:16:22,460
I do need to get my hands dirty at some point.

332
00:16:22,540 --> 00:16:22,640
I know.

333
00:16:22,940 --> 00:16:27,320
I had my conversation with Drew where I felt kind of like similar to him where it's like

334
00:16:27,320 --> 00:16:29,700
I definitely do like the abstractions and theories.

335
00:16:29,800 --> 00:16:32,780
I've encouraged you before, but I want to ask why have you not done this yet?

336
00:16:32,900 --> 00:16:35,800
It is the most transformational thing happening in technology right now.

337
00:16:35,980 --> 00:16:37,680
To be honest, I've had another priority.

338
00:16:37,980 --> 00:16:40,740
So you guys don't see this yet, but I'm actually quite excited.

339
00:16:40,740 --> 00:16:42,920
I just finished yesterday.

340
00:16:43,380 --> 00:16:50,540
my new manifesto for building a new electricity grid and internet of energy. So that is Edison

341
00:16:50,540 --> 00:16:54,760
Alpha. It will drop early next week. I've sent it to a few folks, but I've been pretty hard at

342
00:16:54,760 --> 00:16:58,320
work at that. So priorities. Did you work with an LLM to build that though?

343
00:16:58,680 --> 00:17:04,120
I actually, I used Grok as a research assistant. And so I'm very, very skeptical of using,

344
00:17:04,740 --> 00:17:07,760
yeah, so actually I'll talk about my use of LLMs. So I have been using LLMs, but

345
00:17:07,760 --> 00:17:12,800
look, at the end of the day, I don't really necessarily see myself as an engineer or whatever

346
00:17:12,800 --> 00:17:16,020
I mean, look, I'm excited and I hope I make progress on that.

347
00:17:16,720 --> 00:17:18,000
But I definitely see myself.

348
00:17:18,060 --> 00:17:19,420
You don't need to see yourself as an engineer.

349
00:17:19,560 --> 00:17:22,160
You need to see yourself as a person who uses technology.

350
00:17:22,580 --> 00:17:22,820
Yes.

351
00:17:22,920 --> 00:17:23,120
Okay.

352
00:17:24,540 --> 00:17:25,140
Fair enough.

353
00:17:25,220 --> 00:17:29,640
Let's say I look forward to that and I look forward to watching my horizons grow.

354
00:17:29,960 --> 00:17:33,700
However, I also believe we all have our natural proclivities, right?

355
00:17:33,940 --> 00:17:39,180
And my natural proclivities have always taken me into the realm of ideas, which is why I

356
00:17:39,180 --> 00:17:40,340
have my podcast, Ideas from the Edge.

357
00:17:40,340 --> 00:17:46,280
and um you know i was pretty skeptical of using ai at first because i do love to write i do love to

358
00:17:46,280 --> 00:17:53,580
you know imagine think whatever and you know i don't think i'll ever give up writing to the llm

359
00:17:53,580 --> 00:17:57,540
because to me writing is thinking and it's like that's like a skill i don't want to lose if my

360
00:17:57,540 --> 00:18:02,400
mind you know atrophies in that direction like that's that's a huge problem for me um but what

361
00:18:02,400 --> 00:18:07,780
i am becoming increasingly open to is using llm as a friend uh a research assistant someone that

362
00:18:07,780 --> 00:18:11,060
we can work together. It's like your partner. It's my partner. It doesn't take over for you.

363
00:18:11,200 --> 00:18:15,300
Exactly. It doesn't take over for me, but I'll give some major caveats. So, you know,

364
00:18:15,580 --> 00:18:21,820
what I think the LLM is bad at right now is generating truly beautiful novel ideas. And I

365
00:18:21,820 --> 00:18:26,040
think maybe in time I'll get better at like coaching that. But I think one thing I'm very

366
00:18:26,040 --> 00:18:35,980
skeptical on is, you know, if our world or universe is an infinite sea of possibilities, right? And

367
00:18:35,980 --> 00:18:40,400
the LLM can in theory generate the same ones that I could or another human could,

368
00:18:41,180 --> 00:18:46,720
it still comes down to, you know, you have a limited, in this dimension, we have limited

369
00:18:46,720 --> 00:18:53,180
constraints, right? And so you still have to pick which ideas to remix or co-generate.

370
00:18:53,620 --> 00:19:01,380
And I think that for now, you're not replacing the human element of, I don't know, it's like

371
00:19:01,380 --> 00:19:04,880
cooking, like the weird experiences or happenstances that you've had in life.

372
00:19:04,880 --> 00:19:05,760
I have a counter example.

373
00:19:06,220 --> 00:19:06,440
Okay.

374
00:19:06,560 --> 00:19:07,420
No, I would love to hear that.

375
00:19:07,740 --> 00:19:09,360
But yeah, go for it.

376
00:19:09,920 --> 00:19:16,880
So Matt Belez developed a new game called Tickleball, which is a mix between pickleball

377
00:19:16,880 --> 00:19:17,500
and tennis.

378
00:19:18,320 --> 00:19:21,180
This sounds like a thing that Ani helped him with.

379
00:19:23,820 --> 00:19:25,220
I didn't ask about that.

380
00:19:25,220 --> 00:19:26,660
but

381
00:19:26,660 --> 00:19:28,240
you know

382
00:19:28,240 --> 00:19:28,780
you can just

383
00:19:28,780 --> 00:19:29,160
imagine

384
00:19:29,160 --> 00:19:30,220
you play with

385
00:19:30,220 --> 00:19:30,760
like a junior

386
00:19:30,760 --> 00:19:31,500
tennis racket

387
00:19:31,500 --> 00:19:32,740
one of those

388
00:19:32,740 --> 00:19:33,740
orange tennis balls

389
00:19:33,740 --> 00:19:34,740
that's not as firm

390
00:19:34,740 --> 00:19:36,040
but you don't

391
00:19:36,040 --> 00:19:36,680
use a paddle

392
00:19:36,680 --> 00:19:37,400
but you play

393
00:19:37,400 --> 00:19:38,340
on a pickleball court

394
00:19:38,340 --> 00:19:38,640
yeah

395
00:19:38,640 --> 00:19:40,240
and

396
00:19:40,240 --> 00:19:41,840
he used an

397
00:19:41,840 --> 00:19:42,260
LLM

398
00:19:42,260 --> 00:19:43,060
to develop this game

399
00:19:43,060 --> 00:19:43,280
but

400
00:19:43,280 --> 00:19:44,660
in terms of

401
00:19:44,660 --> 00:19:45,700
like novel ideas

402
00:19:45,700 --> 00:19:46,280
and like

403
00:19:46,280 --> 00:19:46,740
out there

404
00:19:46,740 --> 00:19:47,200
things

405
00:19:47,200 --> 00:19:49,220
the LLM

406
00:19:49,220 --> 00:19:49,960
suggested that

407
00:19:49,960 --> 00:19:52,260
if you drop

408
00:19:52,260 --> 00:19:53,000
the ball

409
00:19:53,000 --> 00:19:53,960
in the kitchen

410
00:19:53,960 --> 00:19:55,080
and you yelled

411
00:19:55,220 --> 00:20:04,640
tickle at the same time you get a bonus point novel i mean that's something it is something

412
00:20:04,640 --> 00:20:11,860
i find i mean i yeah i find that llms can be incredibly creative so i i don't buy the like

413
00:20:11,860 --> 00:20:17,660
oh only humans are creative and llms are just like wrote machine executors i do think that humans like

414
00:20:17,660 --> 00:20:23,740
max you may have a certain layer of taste and understanding and even like training about what

415
00:20:23,740 --> 00:20:28,900
other humans may like or value that llm sometimes are a little sycophantic or they kind of

416
00:20:28,900 --> 00:20:34,840
too people pleasing to really figure out something that's novel but might be liked by humans so i

417
00:20:34,840 --> 00:20:39,960
think they're not necessarily great at the curation thing but i think they're exceptional

418
00:20:39,960 --> 00:20:45,480
at diversion thinking totally you can tell it hey give me 15 different variations of pickleball

419
00:20:45,480 --> 00:20:49,880
and then it gives you 15 and be like actually i want 100 instead and it'll come up with 100 novel

420
00:20:49,880 --> 00:20:54,140
things that are like weird remixes. And then you'd be like, actually number 87 is really good.

421
00:20:54,840 --> 00:20:57,680
Yeah. I mean, you know, for example, I hear in the Soviet Union,

422
00:20:57,860 --> 00:21:01,640
the LLMs have come up with Sickleball. That's like a really new...

423
00:21:04,700 --> 00:21:11,260
But no, I want to be clear. I agree with you. And like, look, I guess what I'm saying is

424
00:21:11,260 --> 00:21:16,820
I still see humans as the main characters for now. And the LLMs, if you know how to use it well,

425
00:21:16,820 --> 00:21:18,580
Then you can like 10,000x.

426
00:21:18,900 --> 00:21:21,640
Someone who's already creative can become a lot more creative.

427
00:21:22,080 --> 00:21:25,660
But there's still that element of taste, of knowing how to prompt it, of knowing where

428
00:21:25,660 --> 00:21:25,960
it can help.

429
00:21:26,040 --> 00:21:27,160
So as an example, right?

430
00:21:27,220 --> 00:21:31,780
So the way that I've been increasingly using my own writing process is I have my own ideas

431
00:21:31,780 --> 00:21:36,500
and I sketch out pieces and then I build from pieces on there.

432
00:21:36,660 --> 00:21:40,360
And all my essays, I finished my LP update recently and similar process.

433
00:21:41,020 --> 00:21:45,920
And there are certain things where I'll say, okay, I don't know enough about this.

434
00:21:45,920 --> 00:21:49,780
So for example, yesterday, I was like, I don't know the state of the art of microgrids, which

435
00:21:49,780 --> 00:21:54,320
is its own fascinating rabbit hole where you can basically build what I think will eventually

436
00:21:54,320 --> 00:21:56,260
become macrogrids, but grids that are separate from the grid.

437
00:21:56,380 --> 00:21:59,440
And then the conventional wisdom is you plug those in, but they're not all plugged in.

438
00:21:59,440 --> 00:22:04,320
So I say, okay, what are examples of microgrids that have chosen to go totally off grid?

439
00:22:04,440 --> 00:22:09,220
Okay, well, Grok was extremely helpful at finding all of those potentially even more

440
00:22:09,220 --> 00:22:09,900
obscure sources.

441
00:22:10,020 --> 00:22:14,320
In fact, I think it's way better than the search engines at finding like, because I can

442
00:22:14,320 --> 00:22:19,240
kind of prompted and be like, listen, I want the most extreme out there examples of people that are

443
00:22:19,240 --> 00:22:23,360
pushing the frontiers and are trying to go totally off grid. You know, here are a couple of examples

444
00:22:23,360 --> 00:22:27,580
of thinkers that I like that are pushing in this way, find me others. And it was great at that.

445
00:22:27,840 --> 00:22:32,280
Really, really great. And so anyways, so in that sense, I use it kind of like if I,

446
00:22:32,560 --> 00:22:38,340
or if I just want to say like, I need a quick, you know, 30 minute or 60 minute overview of the

447
00:22:38,340 --> 00:22:43,620
state of the art, like give me all those rabbit holes. And what I find myself doing is its own

448
00:22:43,620 --> 00:22:47,940
summaries, it's okay. It's pretty good. But, uh, but the, the links, the rabbit holes,

449
00:22:47,940 --> 00:22:54,580
it can be exceptionally good at pointing towards, um, otherwise hard to discover

450
00:22:54,580 --> 00:22:59,820
high quality thinking. So anyways, so, so that, that's the way I'm primarily using it. I also

451
00:22:59,820 --> 00:23:03,980
experimented a bit, although I haven't done this a ton yet, I'm just like feeding at the essay and

452
00:23:03,980 --> 00:23:08,840
being like, okay, here, here's my thinking. Um, what do you think of it? Uh, how would you improve

453
00:23:08,840 --> 00:23:17,480
it? What are some areas you might change things? And that's also, I mean, I ignore some of it,

454
00:23:17,860 --> 00:23:23,900
most of it, but there are a couple of really good ideas. The LLM suggested this is very dense. Maybe

455
00:23:23,900 --> 00:23:30,480
it would be good to break it in and have bolded headings for each section so it's easier to follow

456
00:23:30,480 --> 00:23:33,880
them. It's actually a pretty good idea. I didn't necessarily like the headings it came up with.

457
00:23:33,880 --> 00:23:37,960
I liked my own more, but that idea to put in headings came from the I.

458
00:23:38,060 --> 00:23:42,620
So all that to say, I think I'm still learning how to use it as a research assistant, but

459
00:23:42,620 --> 00:23:45,160
it's definitely a very powerful tool there.

460
00:23:45,880 --> 00:23:45,980
Yeah.

461
00:23:46,620 --> 00:23:50,720
I've got a technique I want to share that I think helps it.

462
00:23:50,840 --> 00:23:54,360
I've discovered recently helps me get deeper like that, like you're talking about.

463
00:23:54,560 --> 00:23:58,480
I forget if you, if the three of us discussed it before, I know I've been an advocate with

464
00:23:58,480 --> 00:24:03,560
Max for a while, but did we talk about speaking to your LLM as a disruptive?

465
00:24:03,880 --> 00:24:06,320
And I know Jack does that as well.

466
00:24:07,200 --> 00:24:10,080
I find, I really got into it like this past week

467
00:24:10,080 --> 00:24:12,120
and it's really changed the game

468
00:24:12,120 --> 00:24:13,480
for how I interact with LLMs.

469
00:24:13,740 --> 00:24:15,540
I don't, I've used this talking,

470
00:24:15,840 --> 00:24:18,160
you know, the speech voice mode of LLMs before

471
00:24:18,160 --> 00:24:18,760
and that's different.

472
00:24:19,220 --> 00:24:20,860
What I'm talking about is there's these apps

473
00:24:20,860 --> 00:24:38,899
like the original Karpathy tweet mentioned Super Whisper There one called Whisper Flow that I using But basically just whenever you want to type something on your computer instead of typing you hold down your function key and you just speak it And what I find is I give a lot more detail I go into directions that I

474
00:24:38,899 --> 00:24:43,759
might not go because of how light and easy it is to speak to the computer versus type something.

475
00:24:44,239 --> 00:24:46,579
And so I give a lot more detail, I go into a lot more different directions

476
00:24:46,579 --> 00:24:52,499
and it's much more conversational. And I have gotten dramatically better results out of the LLM

477
00:24:52,499 --> 00:24:55,959
by being able to just speak to it.

478
00:24:56,819 --> 00:24:58,639
And okay, so two things come to mind immediately.

479
00:24:58,819 --> 00:25:01,159
One, how is this different than voice mode?

480
00:25:01,299 --> 00:25:02,439
Like what is this function?

481
00:25:02,719 --> 00:25:04,259
Can you like, and what are you doing?

482
00:25:04,259 --> 00:25:04,879
It's just like a,

483
00:25:04,879 --> 00:25:07,659
no, this is just like a third party app

484
00:25:07,659 --> 00:25:08,979
that I have running on Mac OS.

485
00:25:09,399 --> 00:25:09,639
Which one?

486
00:25:10,219 --> 00:25:11,759
It's called Whisperflow.

487
00:25:12,199 --> 00:25:12,459
Okay.

488
00:25:12,659 --> 00:25:13,799
So all of these are,

489
00:25:13,879 --> 00:25:14,439
there's a bunch of these,

490
00:25:14,559 --> 00:25:15,619
like Super Whisper was the one

491
00:25:15,619 --> 00:25:17,139
that Karpathy mentioned in the original tweet

492
00:25:17,139 --> 00:25:18,299
about vibe coding.

493
00:25:19,159 --> 00:25:20,959
Whisperflow is one that's kind of gotten popular

494
00:25:20,959 --> 00:25:22,219
among Silicon Valley crowds

495
00:25:22,219 --> 00:25:23,319
in the recent months.

496
00:25:23,939 --> 00:25:26,139
But they're all using the open source,

497
00:25:26,499 --> 00:25:28,559
open weights model, the whisper model

498
00:25:28,559 --> 00:25:30,939
that OpenAI developed and opened up.

499
00:25:31,299 --> 00:25:32,659
And these are just like wrappers on that.

500
00:25:32,839 --> 00:25:34,439
And it's a Mac OS client that I use.

501
00:25:34,699 --> 00:25:36,519
So you can use it with any LLM you want,

502
00:25:36,579 --> 00:25:37,379
any text box.

503
00:25:37,419 --> 00:25:38,499
You can use it when writing an email.

504
00:25:38,499 --> 00:25:39,999
You basically install the software

505
00:25:39,999 --> 00:25:44,199
and then you hold down the global function key

506
00:25:44,199 --> 00:25:44,999
as of how I have it mapped.

507
00:25:45,179 --> 00:25:46,379
And you just start speaking to your computer

508
00:25:46,379 --> 00:25:48,839
and then it transcribes what you said

509
00:25:48,839 --> 00:25:50,339
and inputs it into the box

510
00:25:50,339 --> 00:25:51,879
when you let go of the function button.

511
00:25:52,219 --> 00:25:54,339
And it could be into any LM, a coding thing,

512
00:25:54,819 --> 00:25:56,139
email composer, whatever.

513
00:25:56,479 --> 00:25:59,479
But it really changes the way I interact with the computer

514
00:25:59,479 --> 00:26:01,519
because of how much more willing I am

515
00:26:01,519 --> 00:26:04,239
to give a lot of detail on my thoughts

516
00:26:04,239 --> 00:26:06,919
instead of trying to like first curate my thoughts

517
00:26:06,919 --> 00:26:08,239
into something that I'm willing to type.

518
00:26:08,919 --> 00:26:10,239
Okay, so I love that.

519
00:26:10,699 --> 00:26:12,899
Let me throw out an objection.

520
00:26:13,579 --> 00:26:16,879
And this is an objection that I desperately want to solve.

521
00:26:17,319 --> 00:26:18,319
It's an objection I've been discussing

522
00:26:18,319 --> 00:26:19,979
with a lot of people in my life recently.

523
00:26:20,119 --> 00:26:21,259
Mac OS clients and malware?

524
00:26:21,259 --> 00:26:27,979
both uh you're getting ahead of us but both in in bitcoin and normie lamp i think it's extremely

525
00:26:27,979 --> 00:26:34,819
important okay so i love what you're saying and i did briefly maybe two months ago or six weeks ago

526
00:26:34,819 --> 00:26:40,979
when you were you were first uh speak with your ai friend telling me i did start chatting with

527
00:26:40,979 --> 00:26:44,839
grok and we had some fascinating conversations it led to me ordering a bunch of new books i

528
00:26:44,839 --> 00:26:49,299
you guys know i have a problem with books and it's flooding my house now which is getting out

529
00:26:49,299 --> 00:26:53,299
of control. Because my thinking on books, by the way, is if you ever find anything that could be

530
00:26:53,299 --> 00:26:57,379
remotely interesting, it's like you should just order it, and then you build out the entire library,

531
00:26:57,499 --> 00:27:00,899
and whenever that book speaks to you, you pick it out later. That's my approach, which is great,

532
00:27:00,979 --> 00:27:04,899
until it just floods your house. Amazon Kindle.

533
00:27:05,399 --> 00:27:07,559
Yeah, but it's not the same, man. I agree.

534
00:27:07,939 --> 00:27:11,499
It's just like, yeah, anyways, that's another taste or preference, I guess.

535
00:27:11,959 --> 00:27:15,559
But what I started realizing, it was extremely effective, I agree,

536
00:27:15,559 --> 00:27:24,419
but I found myself being like, uh-oh, wait a second. This feels way too much like the early

537
00:27:24,419 --> 00:27:28,839
days of something like Facebook. You're like, oh, wow, this is so sweet. I'm seeing all of my

538
00:27:28,839 --> 00:27:34,299
friends online. This is great. Hey, here's my birthday. Hey, here's all of my photos of everything

539
00:27:34,299 --> 00:27:39,199
I've ever done in my life. Hey, here's my social security number. And before you know it, all of a

540
00:27:39,199 --> 00:27:42,959
sudden you're trapped in this dystopian hellscape where Mark Zuckerberg owns everything you do,

541
00:27:42,959 --> 00:27:44,639
everyone you know, everyone you own,

542
00:27:44,959 --> 00:27:46,099
and good luck trying to leave

543
00:27:46,099 --> 00:27:48,399
because the network effects are so strong you're trapped.

544
00:27:49,219 --> 00:27:52,759
And even though I like,

545
00:27:52,939 --> 00:27:55,999
of the new feudal lords and pirates

546
00:27:55,999 --> 00:27:57,839
battling for our human attention,

547
00:27:58,379 --> 00:27:59,859
Elon is certainly my favorite of them.

548
00:28:00,559 --> 00:28:02,839
I don't trust any of these guys at all.

549
00:28:03,419 --> 00:28:05,319
And so as I'm sitting there talking to Grok,

550
00:28:05,379 --> 00:28:06,899
I'm like, this is awesome, but wait a second.

551
00:28:06,979 --> 00:28:08,159
The more I'm chatting with you,

552
00:28:08,459 --> 00:28:10,179
the deeper you're getting into my actual thought.

553
00:28:10,179 --> 00:28:12,339
And if we thought Facebook was bad, bro,

554
00:28:12,339 --> 00:28:19,179
So LLMs that know everything you're thinking in real time as you're chatting and are controlled by an opaque centralized server?

555
00:28:19,279 --> 00:28:19,859
No, thank you.

556
00:28:20,219 --> 00:28:25,299
Like to me, everything that we've ever criticized in the Facebook world is going to be 10,000x worse than LLM.

557
00:28:25,299 --> 00:28:25,819
Maple AI.

558
00:28:26,239 --> 00:28:27,999
So this is where I'm getting to, Maple AI.

559
00:28:28,339 --> 00:28:31,459
We are fans, investors.

560
00:28:32,019 --> 00:28:34,759
Well, I don't think you guys, but we are in this company, which I love.

561
00:28:35,179 --> 00:28:35,579
Yeah, sorry.

562
00:28:36,079 --> 00:28:38,439
Dude, these guys are fucking killing it, by the way.

563
00:28:38,699 --> 00:28:40,019
Like I'm extremely bullish.

564
00:28:40,019 --> 00:28:40,319
Great.

565
00:28:40,319 --> 00:28:44,079
But Maple AI, I love their vision.

566
00:28:44,819 --> 00:28:47,019
It's basically the Apple private cloud.

567
00:28:47,819 --> 00:28:50,419
So the idea is they use the T,

568
00:28:50,719 --> 00:28:53,539
the secure execution environment.

569
00:28:53,799 --> 00:28:55,099
Trusted execution environment.

570
00:28:55,239 --> 00:28:55,539
Yeah, thank you.

571
00:28:56,019 --> 00:28:58,979
And this whole concept of confidential compute.

572
00:28:59,139 --> 00:29:01,779
I did a podcast on ideas from the edge with the team.

573
00:29:01,839 --> 00:29:03,119
So if you're interested, you can hear that whole thing.

574
00:29:03,439 --> 00:29:04,439
But the gist of it is,

575
00:29:04,799 --> 00:29:07,779
you can do things all locally, which is great.

576
00:29:07,779 --> 00:29:09,239
But obviously local has limits.

577
00:29:09,799 --> 00:29:11,699
People end up using the cloud because it's more powerful.

578
00:29:11,819 --> 00:29:12,119
It's cheaper.

579
00:29:12,239 --> 00:29:12,559
It's better.

580
00:29:13,399 --> 00:29:15,879
That's how we ended up in the digital hellscape we did.

581
00:29:16,579 --> 00:29:17,799
But they say, okay, cool.

582
00:29:17,899 --> 00:29:19,159
We can give you the benefits of the cloud.

583
00:29:19,279 --> 00:29:19,759
It's scalable.

584
00:29:19,919 --> 00:29:20,299
It's cheaper.

585
00:29:22,019 --> 00:29:28,659
But because all of the keys are generated within a secure enclave within the server,

586
00:29:29,119 --> 00:29:30,899
we can't see what the hell is going on there.

587
00:29:30,899 --> 00:29:35,499
In theory, if you trust the hardware, that's its own deep, deep rabbit hole,

588
00:29:35,979 --> 00:29:38,539
then it's can't be evil, not don't be evil.

589
00:29:38,539 --> 00:29:39,419
It's a signal.

590
00:29:39,719 --> 00:29:42,179
People familiar with signal, it's a similar model.

591
00:29:42,319 --> 00:29:42,439
Yeah.

592
00:29:42,879 --> 00:29:46,059
So how does it, is it using open source models only then?

593
00:29:46,339 --> 00:29:47,299
So this is the problem.

594
00:29:47,599 --> 00:29:51,519
And so I was trying to, without doxing all of them,

595
00:29:53,339 --> 00:29:57,019
privacy pill people in the more normal tech world

596
00:29:57,019 --> 00:29:57,859
and say, this is important.

597
00:29:57,959 --> 00:29:58,699
And they're like, yeah, we agree.

598
00:29:58,759 --> 00:29:59,499
This is super important.

599
00:29:59,579 --> 00:30:01,019
I can totally get what you're saying.

600
00:30:01,279 --> 00:30:02,639
Something that starts off awesome,

601
00:30:02,719 --> 00:30:04,659
your friend ends up just like manipulating thoughts,

602
00:30:04,779 --> 00:30:05,819
stealing your data, all this.

603
00:30:05,919 --> 00:30:07,719
And like, oh, by the way, like if I'm a business,

604
00:30:07,719 --> 00:30:10,799
If I'm a healthcare company, I shouldn't use OpenAI or one of these guys.

605
00:30:11,519 --> 00:30:15,679
If I'm moderately intelligent, I'm also terrified of using them.

606
00:30:16,359 --> 00:30:19,479
And so there are a lot of people that are very sympathetic to the problems.

607
00:30:19,939 --> 00:30:22,159
So I'm like, okay, well, then you should consider using OpenSecret,

608
00:30:22,259 --> 00:30:28,899
which is the, I guess, platform that this encrypted confidential compute is running on.

609
00:30:29,099 --> 00:30:30,459
And that's what MapleAI is built on.

610
00:30:31,259 --> 00:30:33,239
Problem is like, great, how do I use Cloud?

611
00:30:34,319 --> 00:30:35,679
And it's like, well, you can't.

612
00:30:35,679 --> 00:30:37,179
You can only use the open source models.

613
00:30:37,179 --> 00:30:38,759
and like, yeah, the Chinese ones are pretty good,

614
00:30:38,799 --> 00:30:40,699
but we also don't know what's like deep, deep down there.

615
00:30:40,919 --> 00:30:41,919
That's another concern I have.

616
00:30:42,259 --> 00:30:44,079
You can use the open source AI model now,

617
00:30:44,219 --> 00:30:44,819
Grok, I guess.

618
00:30:44,899 --> 00:30:46,519
I don't know if they have listed the open source version.

619
00:30:46,999 --> 00:30:49,279
The TLDR is like the open source models are good,

620
00:30:49,319 --> 00:30:51,459
but they're not as good as the closed source stuff, full stop.

621
00:30:51,579 --> 00:30:52,439
And so that's the trade-off.

622
00:30:52,439 --> 00:30:54,179
It's like, as we stand right now,

623
00:30:54,879 --> 00:30:59,339
do you value your not getting brainwashed more

624
00:30:59,339 --> 00:31:01,659
or do you value having a better research assistant?

625
00:31:02,599 --> 00:31:03,599
A few thoughts here.

626
00:31:03,599 --> 00:31:11,499
First of all, often privacy, like often, in most cases with privacy, I don't think consumers really care.

627
00:31:11,859 --> 00:31:15,039
But with this, I think it very much matters.

628
00:31:15,259 --> 00:31:16,339
First of all, there's corporations.

629
00:31:16,619 --> 00:31:21,559
As a corporation, you want to use these tools, but you don't want to give some other company like all of your secrets.

630
00:31:23,139 --> 00:31:32,239
But even as a consumer, health care, whether it be physical health, mental health, there's just so many topics that you might want to inquire about.

631
00:31:32,239 --> 00:31:38,899
And like you said earlier, you're going to inquire in a much deeper way than even with Google Search.

632
00:31:39,059 --> 00:31:39,359
Way.

633
00:31:39,699 --> 00:31:42,979
And you don't want another company to have all that information.

634
00:31:43,499 --> 00:31:45,159
So I think there's going to be strong demand.

635
00:31:46,739 --> 00:31:53,699
So, okay, if open source models today are inferior to proprietary ones, a couple things that could happen.

636
00:31:54,019 --> 00:31:56,999
One, and I think we'll get into this later in the show.

637
00:31:56,999 --> 00:31:59,259
but there's certainly

638
00:31:59,259 --> 00:32:02,559
if we're reaching a plateau

639
00:32:02,559 --> 00:32:04,179
with LLM development

640
00:32:04,179 --> 00:32:05,759
for the foreseeable future

641
00:32:05,759 --> 00:32:07,959
then I would expect open source to catch up

642
00:32:07,959 --> 00:32:09,459
and that sort of commoditized

643
00:32:09,459 --> 00:32:11,139
then problem solved

644
00:32:11,139 --> 00:32:12,439
or

645
00:32:12,439 --> 00:32:16,159
if proprietary maintains

646
00:32:16,159 --> 00:32:18,179
an edge, there's numerous competitors

647
00:32:18,179 --> 00:32:19,219
right now for proprietary

648
00:32:19,219 --> 00:32:22,299
one that might start slipping

649
00:32:22,299 --> 00:32:24,119
behind might want an edge

650
00:32:24,119 --> 00:32:26,019
they can

651
00:32:26,019 --> 00:32:30,279
they can introduce, I mean, whether they use OpenSeeker or not, or they just build it in-house.

652
00:32:30,559 --> 00:32:34,939
I mean, maybe that's not good for Maple AI, but they could build it in-house. So whether it's

653
00:32:34,939 --> 00:32:40,559
Anthropic or whoever, or one of the Chinese models. I think it's going to be the Chinese.

654
00:32:41,379 --> 00:32:46,179
I guess they're already open source though, right? So in any case, a proprietary model that is

655
00:32:46,179 --> 00:32:52,079
superior or at least much closer to the best could do it as well.

656
00:32:52,079 --> 00:32:54,139
So another, I mean, yes.

657
00:32:54,419 --> 00:32:56,559
And I think that remains, you know,

658
00:32:57,279 --> 00:32:58,799
that remains a great topic of conversation.

659
00:32:58,899 --> 00:33:00,659
And to be clear, I'm super bullish on people.

660
00:33:00,779 --> 00:33:01,459
I love these guys.

661
00:33:01,539 --> 00:33:02,839
I think what they're doing is so important.

662
00:33:03,599 --> 00:33:04,799
And I definitely think,

663
00:33:04,859 --> 00:33:06,579
and I've seen their revenue numbers.

664
00:33:06,659 --> 00:33:08,219
Like there are people and businesses

665
00:33:08,219 --> 00:33:10,119
that are choosing this and like, it's real.

666
00:33:11,219 --> 00:33:12,819
But I'm just saying in my own,

667
00:33:12,919 --> 00:33:15,059
like I have every incentive in the world to use them.

668
00:33:15,079 --> 00:33:16,579
And I probably should more as a research assistant,

669
00:33:16,639 --> 00:33:17,919
but the reality is Grok is great.

670
00:33:18,219 --> 00:33:19,359
It's just really, really good.

671
00:33:19,359 --> 00:33:20,139
Like it's healthy.

672
00:33:20,259 --> 00:33:21,999
And it has all of that real-time Twitter data

673
00:33:21,999 --> 00:33:26,099
in context. And so as an investor in a financial research, like I want that real-time Twitter day,

674
00:33:26,139 --> 00:33:30,899
it's really, really good. And so it's like playing out in real time. It's like, I know this is bad

675
00:33:30,899 --> 00:33:35,679
for me and I know the risks and yet I'm still using, now I'm not using open AI, like I'm not

676
00:33:35,679 --> 00:33:40,179
that crazy, but at least with Grok, you know, it's like, and I'm kind of like, you know,

677
00:33:40,619 --> 00:33:44,239
if I'm going to roll the dice with one of the pirates, I'll go with Elon.

678
00:33:45,099 --> 00:33:45,839
Or Demis.

679
00:33:46,259 --> 00:33:50,099
Or Demis. Yeah. I like, yes, exactly. I'll go with Elon or Demis in that world.

680
00:33:50,099 --> 00:33:52,199
and hell even Dario

681
00:33:52,199 --> 00:33:53,999
I gotta say I was originally

682
00:33:53,999 --> 00:33:56,559
like very down on Anthropic I thought they were gonna be

683
00:33:56,559 --> 00:33:58,579
you know way too woke

684
00:33:58,579 --> 00:34:00,299
and like safety pilled or whatever

685
00:34:00,299 --> 00:34:02,199
but like they've actually made some like

686
00:34:02,199 --> 00:34:04,339
I'll give that man he's an alchemist

687
00:34:04,339 --> 00:34:05,999
despite having fewer resources

688
00:34:05,999 --> 00:34:08,579
everyone I know in Silicon Valley codes with Claude still

689
00:34:08,579 --> 00:34:10,399
so good on him

690
00:34:10,399 --> 00:34:12,439
and he doesn't have that like

691
00:34:12,439 --> 00:34:14,779
I don't know he doesn't seem evil yet

692
00:34:14,779 --> 00:34:16,759
but the point is you can't trust

693
00:34:16,759 --> 00:34:18,419
any of these people because they don't start that way

694
00:34:18,419 --> 00:34:21,719
The Google guys, don't be evil, right?

695
00:34:21,759 --> 00:34:24,879
That was literally in their manifesto.

696
00:34:25,019 --> 00:34:26,459
Or like, was it in the corporate documents?

697
00:34:27,419 --> 00:34:27,559
Yeah.

698
00:34:27,719 --> 00:34:35,579
And I would say Google stood true to that as good as any corporation could and for as long as they possibly could.

699
00:34:35,659 --> 00:34:36,359
But that's my point.

700
00:34:36,379 --> 00:34:37,759
I still wouldn't call Google evil.

701
00:34:38,199 --> 00:34:43,519
They just, it's like the original culture has been lost, but I wouldn't call the company evil.

702
00:34:43,639 --> 00:34:44,179
And I agree.

703
00:34:44,279 --> 00:34:47,039
Of the big tech companies, believe me, I like Google.

704
00:34:47,399 --> 00:34:48,499
I've always liked Google.

705
00:34:49,059 --> 00:34:51,359
I actually believe that the people at the top,

706
00:34:51,439 --> 00:34:52,979
the founders, were good people.

707
00:34:53,199 --> 00:34:54,619
And they probably still are.

708
00:34:55,439 --> 00:34:57,239
I wouldn't say that about a lot of other tech companies,

709
00:34:57,319 --> 00:34:58,139
but it's the incentives.

710
00:34:58,899 --> 00:35:00,199
Well, also, they've left.

711
00:35:00,499 --> 00:35:01,579
Or at least Larry has, right?

712
00:35:01,819 --> 00:35:07,659
So that's why even if you believe in a founder, a CEO,

713
00:35:08,399 --> 00:35:11,019
someday, they'll be gone.

714
00:35:11,199 --> 00:35:11,639
Exactly.

715
00:35:11,639 --> 00:35:13,619
And in an early day of a startup,

716
00:35:13,819 --> 00:35:16,319
it's hard to imagine that,

717
00:35:16,319 --> 00:35:17,759
but we've already seen it with Google.

718
00:35:18,059 --> 00:35:20,479
Like Larry Page just exited the scene, right?

719
00:35:20,579 --> 00:35:23,039
And you don't know who's going to take over afterwards.

720
00:35:23,659 --> 00:35:23,939
Yeah.

721
00:35:24,219 --> 00:35:27,139
I've got to say, if Demis were to take over,

722
00:35:27,319 --> 00:35:29,059
I'd be happier about that

723
00:35:29,059 --> 00:35:31,239
than some of the other competitors.

724
00:35:31,819 --> 00:35:33,479
No, Demis for sure.

725
00:35:33,579 --> 00:35:34,819
I mean, like that dude's a real one.

726
00:35:35,599 --> 00:35:37,399
Anyways, but the point is,

727
00:35:37,679 --> 00:35:38,979
I mean, Demis could leave tomorrow.

728
00:35:39,119 --> 00:35:39,879
Like the point is,

729
00:35:40,079 --> 00:35:42,519
I mean, this is the whole thesis of Bitcoin and Noster.

730
00:35:43,599 --> 00:35:45,079
Can't be evil, not don't be evil

731
00:35:45,079 --> 00:35:47,859
because even the best person, you know, whatever.

732
00:35:48,919 --> 00:35:50,719
I mean, once you get that tap on the shoulder,

733
00:35:51,199 --> 00:35:51,779
Jack at Twitter,

734
00:35:52,219 --> 00:35:53,859
Jack is probably the best of this whole bunch.

735
00:35:54,439 --> 00:35:56,059
There's just, there's only so much you can do

736
00:35:56,059 --> 00:35:56,699
as one human being.

737
00:35:56,819 --> 00:35:59,279
So anyways, my point is protocols over people

738
00:35:59,279 --> 00:35:59,939
all day, every day.

739
00:36:00,299 --> 00:36:03,159
Now, the question I have is twofold.

740
00:36:03,299 --> 00:36:05,319
One, do you guys see,

741
00:36:05,479 --> 00:36:07,259
I mean, Deke, I know you're deeper in AI land.

742
00:36:07,639 --> 00:36:10,179
Among the sort of normie AI crowd,

743
00:36:10,359 --> 00:36:12,179
do you see there's demand for this privacy?

744
00:36:12,459 --> 00:36:13,759
Like people actually understand that?

745
00:36:13,759 --> 00:36:15,539
Are businesses that don't reveal trade secrets?

746
00:36:15,879 --> 00:36:18,819
Or do you think it's just going to be the HIPAA compliance that has to go that route?

747
00:36:19,379 --> 00:36:24,559
And number two, for example, you mentioned someone else could come and try and do what

748
00:36:24,559 --> 00:36:25,059
MapleEye is doing.

749
00:36:25,239 --> 00:36:28,239
Well, should MapleAI raise, I don't know, hundreds of millions of dollars and try and

750
00:36:28,239 --> 00:36:28,959
do this themselves?

751
00:36:29,339 --> 00:36:30,659
Should they focus still on the open source?

752
00:36:30,839 --> 00:36:35,399
Another path that was actually discussing with Tony is like, is there a world where

753
00:36:35,399 --> 00:36:36,619
you just run an anonymizer?

754
00:36:36,619 --> 00:36:38,259
So kind of like a mixer for the AI.

755
00:36:38,799 --> 00:36:42,219
And that's, I guess, kind of what some of these open router type things are doing, where

756
00:36:42,219 --> 00:36:44,219
It's like, I want to use Claude, but like anonymize.

757
00:36:44,339 --> 00:36:47,279
So Claude doesn't know me and they don't know who's me as well.

758
00:36:48,039 --> 00:36:51,739
Can I throw in one more aspect to this discussion?

759
00:36:51,739 --> 00:36:59,819
Something we've been talking about with another entrepreneur is like a system of experts.

760
00:37:00,979 --> 00:37:01,899
Mixture of experts.

761
00:37:01,979 --> 00:37:02,719
Mixture of experts.

762
00:37:04,739 --> 00:37:06,239
Would you talk about that as well?

763
00:37:06,399 --> 00:37:09,199
Because I know, DK, you have some differing views there.

764
00:37:09,299 --> 00:37:11,579
I'm still learning, but differing from maybe others.

765
00:37:12,219 --> 00:37:16,039
But if that's an interesting model,

766
00:37:17,039 --> 00:37:17,979
that could potentially be,

767
00:37:18,159 --> 00:37:20,879
it's more realistic, it can run locally.

768
00:37:21,619 --> 00:37:22,119
So you could get your,

769
00:37:22,279 --> 00:37:24,359
like for certain use cases,

770
00:37:24,859 --> 00:37:26,779
you could use that type of model,

771
00:37:26,899 --> 00:37:27,639
which can run locally,

772
00:37:27,779 --> 00:37:29,559
and then you get privacy.

773
00:37:30,039 --> 00:37:30,179
Right.

774
00:37:30,559 --> 00:37:30,739
Right.

775
00:37:31,939 --> 00:37:34,739
So that's another way to solve this.

776
00:37:35,379 --> 00:37:37,339
I mean, you could use small models

777
00:37:37,339 --> 00:37:39,519
that are trained for specific purposes,

778
00:37:39,519 --> 00:37:40,859
but that's pretty different

779
00:37:40,859 --> 00:37:45,759
than how mixture of experts is more like emergent pieces,

780
00:37:45,919 --> 00:37:49,859
you know, sections of a model that was trained for broad-based general use.

781
00:37:50,359 --> 00:37:52,679
So we don't, the idea of like,

782
00:37:52,839 --> 00:37:55,339
let's make a specific expert for a specific purpose

783
00:37:55,339 --> 00:37:57,899
is not what mixture of experts is talking about.

784
00:37:57,959 --> 00:38:00,919
It's talking about just sort of emergently figuring out

785
00:38:00,919 --> 00:38:05,799
what areas of the embeddings have some sort of expertise

786
00:38:05,799 --> 00:38:07,739
and then routing prompts through that

787
00:38:07,739 --> 00:38:10,279
instead of routing them through the whole language model.

788
00:38:10,859 --> 00:38:18,699
But it's more like a specific piece of a larger model versus kind of a different smaller model to train on its own.

789
00:38:19,139 --> 00:38:19,899
Just to make sure I understand.

790
00:38:20,119 --> 00:38:25,299
So it's more at the inference level than like training a whole new model that's just smaller?

791
00:38:26,319 --> 00:38:31,159
It happens in training, but the big benefit is scalability and inference.

792
00:38:31,759 --> 00:38:40,199
Because like imagine you're just training a very, very large model and you're trying to just understand all of how human language works and all of how human knowledge is connected.

793
00:38:40,859 --> 00:38:47,239
Well, part of that is you might discover that, oh, math is a subset of how human language

794
00:38:47,239 --> 00:38:48,639
works and human knowledge works.

795
00:38:48,979 --> 00:38:54,019
And so when we see a math prompt, we've emergently discovered that there's a bunch of relevant

796
00:38:54,019 --> 00:38:57,019
material all in this subsection of the model.

797
00:38:57,019 --> 00:39:02,639
And so we don't have to take broad-based math prompts and root them, try to explore

798
00:39:02,639 --> 00:39:04,639
the whole space of the model.

799
00:39:04,819 --> 00:39:10,379
But we can just root it towards this specific thing that emerged to look like it has a

800
00:39:10,859 --> 00:39:16,099
information about this domain and so math is a broad domain of that and like coding is often a

801
00:39:16,099 --> 00:39:22,759
broad domain of expertise but like there isn't like a mixture of experts of like what is bitcoin

802
00:39:22,759 --> 00:39:26,539
how does bitcoin technology work let's have an expert understand that that's not really

803
00:39:26,539 --> 00:39:33,299
when people talk about moe within the ai space they're not talking about that fine grain level

804
00:39:33,299 --> 00:39:33,899
of expertise.

805
00:39:36,419 --> 00:39:36,979
Okay.

806
00:39:37,919 --> 00:39:39,279
But I mean, I still think

807
00:39:39,279 --> 00:39:41,419
with respect to privacy,

808
00:39:42,039 --> 00:39:42,999
running a local model

809
00:39:42,999 --> 00:39:44,019
is one way you get privacy.

810
00:39:45,559 --> 00:39:47,339
Another possibility for

811
00:39:47,339 --> 00:39:50,719
how do you have a competitive model

812
00:39:50,719 --> 00:39:51,699
that can run locally?

813
00:39:52,039 --> 00:39:53,319
One could be just building

814
00:39:53,319 --> 00:39:54,019
a smaller model,

815
00:39:54,139 --> 00:39:55,239
but that's customized

816
00:39:55,239 --> 00:39:57,079
and focused on your task at hand.

817
00:39:58,299 --> 00:39:58,459
Yeah.

818
00:39:59,759 --> 00:40:00,659
And by the way,

819
00:40:00,659 --> 00:40:01,979
I haven't seen a great tool yet

820
00:40:01,979 --> 00:40:02,559
for like,

821
00:40:02,759 --> 00:40:03,139
um,

822
00:40:04,279 --> 00:40:06,639
my buddy who's kind of outside of these worlds has asked,

823
00:40:06,739 --> 00:40:06,919
Hey,

824
00:40:07,179 --> 00:40:08,159
basically for what you're saying,

825
00:40:08,259 --> 00:40:09,699
I want to like fine tune,

826
00:40:09,839 --> 00:40:10,079
train,

827
00:40:10,219 --> 00:40:10,479
rag,

828
00:40:10,639 --> 00:40:11,979
whatever the technology is.

829
00:40:12,179 --> 00:40:15,979
I basically just want to train my own assistant on my data privately on my

830
00:40:15,979 --> 00:40:16,319
computer.

831
00:40:16,659 --> 00:40:18,239
Is there a good tool for that yet?

832
00:40:20,579 --> 00:40:23,859
To train a brand new model from scratch or?

833
00:40:23,939 --> 00:40:24,139
No,

834
00:40:24,239 --> 00:40:27,219
to take a model off the shelf and then whatever the,

835
00:40:27,219 --> 00:40:28,959
the technology to jurors,

836
00:40:29,099 --> 00:40:30,819
if it's rag or fine tuning or whatever,

837
00:40:30,919 --> 00:40:31,419
basically say,

838
00:40:31,519 --> 00:40:31,659
Hey,

839
00:40:31,659 --> 00:40:35,819
I want to like customize this with like all of my like personal notes

840
00:40:35,819 --> 00:40:36,359
information,

841
00:40:36,359 --> 00:40:38,539
but I don't want to like leak that to the cloud.

842
00:40:38,639 --> 00:40:38,739
Like,

843
00:40:38,779 --> 00:40:39,979
is there a good tool for that yet?

844
00:40:40,599 --> 00:40:43,239
I would guess there are tools like that.

845
00:40:43,339 --> 00:40:44,139
I don't know.

846
00:40:44,219 --> 00:40:46,399
Often I would guess I could look at LM studio,

847
00:40:46,539 --> 00:40:50,539
maybe that cause I know LM studio lets you run a local model on your Mac.

848
00:40:51,099 --> 00:40:52,559
And I have some friends who run,

849
00:40:52,879 --> 00:40:53,259
run these,

850
00:40:53,359 --> 00:40:56,319
they actually buy these really high powered Macs with like 128 gig.

851
00:40:56,319 --> 00:40:56,919
I got that.

852
00:40:57,199 --> 00:40:58,039
For that reason.

853
00:40:58,539 --> 00:40:58,659
Oh,

854
00:40:58,719 --> 00:40:58,819
good.

855
00:40:58,859 --> 00:40:58,979
Yeah.

856
00:40:59,259 --> 00:40:59,399
Yeah.

857
00:40:59,399 --> 00:41:02,919
My next computer, I'm going to probably get my computer in a month or so.

858
00:41:03,079 --> 00:41:03,839
I need like a terabyte.

859
00:41:04,039 --> 00:41:05,879
Well, you need to use the quantized ones.

860
00:41:06,099 --> 00:41:06,359
Oh.

861
00:41:07,099 --> 00:41:10,559
Yeah, you need to use the quantized models to get any reasonable performance.

862
00:41:11,319 --> 00:41:17,679
But I think if you use a quantized model with a super beefy 128 gig RAM, you can...

863
00:41:17,679 --> 00:41:19,559
I'm at the point where I'm willing and ready.

864
00:41:19,759 --> 00:41:21,219
Although, again, this is not going to work for everyone.

865
00:41:21,339 --> 00:41:22,159
But like, you know, whatever.

866
00:41:22,719 --> 00:41:24,719
Give me 12 Mac mini, string those together.

867
00:41:24,979 --> 00:41:25,119
Right.

868
00:41:25,919 --> 00:41:27,659
Get the M3 Ultra.

869
00:41:27,659 --> 00:41:29,419
if you're willing to have a desktop

870
00:41:29,419 --> 00:41:30,599
do an M3 Ultra

871
00:41:30,599 --> 00:41:32,579
I don't have a desk but I should get one

872
00:41:32,579 --> 00:41:34,679
you're going to have 12 minis

873
00:41:34,679 --> 00:41:36,519
you're going to carry them around with you to the sauna

874
00:41:36,519 --> 00:41:39,139
can we explore

875
00:41:39,139 --> 00:41:41,719
something mentioned a few minutes ago a little bit more

876
00:41:41,719 --> 00:41:43,499
I mean I threw out there

877
00:41:43,499 --> 00:41:45,439
that maybe LLMs are

878
00:41:45,439 --> 00:41:47,879
the progress of them

879
00:41:47,879 --> 00:41:48,739
is plateauing

880
00:41:48,739 --> 00:41:51,779
I'm curious what you guys have

881
00:41:51,779 --> 00:41:53,779
read, heard, what your own intuition

882
00:41:53,779 --> 00:41:54,719
is, do you think

883
00:41:54,719 --> 00:41:57,059
that that is true?

884
00:41:57,659 --> 00:42:07,219
hmm i can start i do not and i'll tell you why um some of the benchmarks whatever maybe you're

885
00:42:07,219 --> 00:42:16,759
showing that like my philosophy on this stuff is very simple um trust the curves and when you

886
00:42:16,759 --> 00:42:22,219
zoom out you know you i think of it kind of it's the bell curve situation where you have

887
00:42:22,219 --> 00:42:25,979
you know the really dumb and the really smart people are like whenever you find a fractal

888
00:42:25,979 --> 00:42:32,659
pattern in nature, just go with it. Bet along that curve. And then you always have the people

889
00:42:32,659 --> 00:42:37,159
in the middle. It's like, no, you can't do that because we have the benchmarks and blah, blah.

890
00:42:37,159 --> 00:42:43,239
It's like, no, just trust the curves. And the one guy that I think has gotten this stuff right,

891
00:42:43,459 --> 00:42:49,079
more or less, that I've read, that just said it directly, is the Leopold Asperger,

892
00:42:49,079 --> 00:42:58,059
whatever's last name this guy um something like that uh which i guess that was a little

893
00:42:58,059 --> 00:43:06,459
Freudian slip yeah um max autist i i do like to track the autist and yeah anyways exactly so

894
00:43:06,459 --> 00:43:16,299
so his report is excellent an apple his report is absolutely excellent um about a year ago um

895
00:43:16,299 --> 00:43:18,699
oh it's so good

896
00:43:18,699 --> 00:43:20,339
about a year ago

897
00:43:20,339 --> 00:43:21,039
he put out

898
00:43:21,039 --> 00:43:23,459
what was the name

899
00:43:23,459 --> 00:43:24,019
of that report

900
00:43:24,019 --> 00:43:26,099
situational awareness

901
00:43:26,099 --> 00:43:27,079
and it was spot on

902
00:43:27,079 --> 00:43:28,039
I've read that many times

903
00:43:28,039 --> 00:43:29,139
I think he's basically

904
00:43:29,139 --> 00:43:29,639
just right

905
00:43:29,639 --> 00:43:30,879
by the way

906
00:43:30,879 --> 00:43:31,759
Nick Carter's always

907
00:43:31,759 --> 00:43:32,479
just like

908
00:43:32,479 --> 00:43:33,659
posting whatever

909
00:43:33,659 --> 00:43:35,519
Leopold's new hedge fund

910
00:43:35,519 --> 00:43:36,419
buys or trades

911
00:43:36,419 --> 00:43:36,799
and he's just

912
00:43:36,799 --> 00:43:37,479
shadow trading it

913
00:43:37,479 --> 00:43:37,939
which I think

914
00:43:37,939 --> 00:43:38,599
is a great strategy

915
00:43:38,599 --> 00:43:39,559
this guy clearly knows

916
00:43:39,559 --> 00:43:39,979
what's up

917
00:43:39,979 --> 00:43:41,299
one of the few people

918
00:43:41,299 --> 00:43:41,839
I've been like

919
00:43:41,839 --> 00:43:42,739
genuinely impressed by

920
00:43:42,739 --> 00:43:44,419
and just from reading

921
00:43:44,419 --> 00:43:44,819
his work

922
00:43:44,819 --> 00:43:49,999
and basically his whole point is just like yeah just trust the curves like you know we keep adding

923
00:43:49,999 --> 00:43:54,499
more compute it's not slowing down like it obviously requires a lot more compute and a lot

924
00:43:54,499 --> 00:43:58,479
more data and those are challenges that we haven't solved right like you have to get more chips more

925
00:43:58,479 --> 00:44:02,079
power and which is the biggest limiting factor right now that's what my whole giant manifesto

926
00:44:02,079 --> 00:44:06,019
that's going to come out next week is all about is like who you know he who controls the power

927
00:44:06,019 --> 00:44:12,179
controls the ai you know uh spoiler alert it's not us it's china and uh we need to change that

928
00:44:12,179 --> 00:44:13,099
We need to change that yesterday.

929
00:44:13,519 --> 00:44:18,359
So anyway, so there's chips, there's the power, there's the data itself.

930
00:44:18,419 --> 00:44:22,259
These are all limiting constraints, but like those are all solvable problems, right?

931
00:44:22,299 --> 00:44:23,359
It may take time to solve it.

932
00:44:23,399 --> 00:44:27,659
It may take a tremendous amount of capital, both intellectual and otherwise, but they're

933
00:44:27,659 --> 00:44:28,459
all solvable problems.

934
00:44:28,739 --> 00:44:32,699
And so then I look and I say, okay, who are the smartest people in the world and what are

935
00:44:32,699 --> 00:44:33,279
they betting on?

936
00:44:33,619 --> 00:44:39,399
Well, you know, I think my other like major investment strategy for the last 15 years

937
00:44:39,399 --> 00:44:41,099
is basically just do what Elon does.

938
00:44:41,099 --> 00:44:46,759
turns out that works out really really well what is elon doing right now he first of all is a solar

939
00:44:46,759 --> 00:44:52,819
maxi that's another uh solar and storage maxi another um a little alpha drop for for next week

940
00:44:52,819 --> 00:44:57,619
uh so elon's betting big on build a lot of solar build a lot of storage go as big as humanly

941
00:44:57,619 --> 00:45:03,239
possible um you know i also follow closely all the team at xai i'm like that there's real signal

942
00:45:03,239 --> 00:45:08,219
there and those guys are suggesting we're going to build the biggest you know super uh cluster

943
00:45:08,219 --> 00:45:15,339
possible. So at first they said, you know, I forgot exactly the numbers, a million GPUs and

944
00:45:15,339 --> 00:45:19,999
Colossus 2, which by the way, it's also very interesting. You can see, so I'm from Memphis,

945
00:45:20,199 --> 00:45:24,439
you know, I always say we do it for the M. And if you track, it's no coincidence that

946
00:45:24,439 --> 00:45:28,979
intelligence, super intelligence is being birthed in Memphis. There's some interesting

947
00:45:28,979 --> 00:45:37,559
rattle holes around that. As it was 35 years ago. Exactly. Or no, no, I was going to say in Egypt,

948
00:45:37,559 --> 00:45:43,359
but that's that's another rabbit hole um so so you you can see i mean you can just like track like

949
00:45:43,359 --> 00:45:45,819
what they're buying the land it was clear they're going to build solar on it because they bought

950
00:45:45,819 --> 00:45:50,059
500 acres next to colossus now they're doing colossus 2 down there anyways he basically said

951
00:45:50,059 --> 00:45:55,479
i forgot the exact number that you know so a million gpus and then they want to do um i forget

952
00:45:55,479 --> 00:46:01,059
if it was 50 000 or what the number was but some like huge quantity of whatever the latest nvidia

953
00:46:01,059 --> 00:46:07,499
chip is is it h 100 200 whatever the state-of-the-art nvidia chip is he's like

954
00:46:07,499 --> 00:46:11,379
basically we're going to do some obscene number of that, which, um, you know, someone was asking

955
00:46:11,379 --> 00:46:15,159
what percentage of the world electricity is that? And it came out like seven or 8%. Of course,

956
00:46:15,699 --> 00:46:18,959
just like when people, you know, bitch about a Bitcoin electricity usage. Exactly. That's like

957
00:46:18,959 --> 00:46:23,919
a terrible, the point is it's a lot of electricity usage. If it's running full speed at today's

958
00:46:23,919 --> 00:46:27,179
efficiencies, of course, the efficiencies are going to get way better. And like, it ignores

959
00:46:27,179 --> 00:46:31,059
all these other challenges, but like the, the, like the point is he's going to go really, really

960
00:46:31,059 --> 00:46:38,199
big. And then he said after that, eventually billions. So Elon's not slowing down. You see

961
00:46:38,199 --> 00:46:44,679
all the other labs, again, without revealing too much from the paper last week, Google has already

962
00:46:44,679 --> 00:46:51,319
contracted, this is a little bit, not 100% confirmed, but seems pretty clear, multiple

963
00:46:51,319 --> 00:46:56,499
gigawatt plus, including one three gigawatt plus cluster that they're building right now.

964
00:46:56,499 --> 00:46:59,579
so the smart money is going big

965
00:46:59,579 --> 00:47:02,199
slight alternative

966
00:47:02,199 --> 00:47:02,719
taking that

967
00:47:02,719 --> 00:47:06,219
I have a well placed source at one of these companies

968
00:47:06,219 --> 00:47:07,179
who

969
00:47:07,179 --> 00:47:09,639
rhymes with frugal

970
00:47:09,639 --> 00:47:11,499
no

971
00:47:11,499 --> 00:47:13,179
rhymes with Elon

972
00:47:13,179 --> 00:47:15,819
I have many connections David

973
00:47:15,819 --> 00:47:18,379
Keon

974
00:47:18,379 --> 00:47:19,779
Keon

975
00:47:19,779 --> 00:47:22,419
and

976
00:47:22,419 --> 00:47:25,419
there's a belief

977
00:47:25,419 --> 00:47:26,339
within

978
00:47:26,339 --> 00:47:28,919
or actually that person thinks

979
00:47:28,919 --> 00:47:30,999
even within the overall space

980
00:47:30,999 --> 00:47:32,519
that it is plateauing.

981
00:47:33,919 --> 00:47:35,919
So why would there be all these big

982
00:47:35,919 --> 00:47:37,339
massive investments

983
00:47:37,339 --> 00:47:39,079
that can be answered as well?

984
00:47:39,319 --> 00:47:40,199
Because right now

985
00:47:40,199 --> 00:47:43,179
ChatGPT has an enormous lead

986
00:47:43,179 --> 00:47:45,319
in terms of brand recognition

987
00:47:45,319 --> 00:47:46,159
and user base.

988
00:47:46,599 --> 00:47:47,859
So for others,

989
00:47:48,179 --> 00:47:50,179
they're still trying to play catch up

990
00:47:50,179 --> 00:47:53,119
and even if it's an enormous investment

991
00:47:53,119 --> 00:47:55,479
for what might be a short period of time

992
00:47:55,479 --> 00:47:56,639
where there's still an edge,

993
00:47:56,879 --> 00:47:58,099
it's worth that investment

994
00:47:58,099 --> 00:47:59,339
because the prize is so big.

995
00:47:59,579 --> 00:47:59,779
Yeah.

996
00:48:00,119 --> 00:48:03,219
So the two might both be true simultaneously.

997
00:48:04,379 --> 00:48:04,979
Yeah, could be.

998
00:48:05,099 --> 00:48:05,879
But I don't really have

999
00:48:05,879 --> 00:48:07,399
any other unique insights than that.

1000
00:48:07,839 --> 00:48:09,579
I also think we're going to get

1001
00:48:09,579 --> 00:48:11,299
more products built.

1002
00:48:11,299 --> 00:48:13,039
I mean, you know the bitter lesson,

1003
00:48:13,599 --> 00:48:15,099
the writing?

1004
00:48:15,699 --> 00:48:17,039
You should recap it, yeah.

1005
00:48:17,499 --> 00:48:18,619
Yeah, the recap,

1006
00:48:18,739 --> 00:48:20,799
I'll recap in a few short statements.

1007
00:48:20,939 --> 00:48:21,359
Basically like,

1008
00:48:21,679 --> 00:48:22,739
don't worry about building products

1009
00:48:22,739 --> 00:48:24,099
because the models are going to get better

1010
00:48:24,099 --> 00:48:27,479
and the models will out-execute

1011
00:48:27,479 --> 00:48:29,639
whatever great ideas you have for products.

1012
00:48:30,139 --> 00:48:32,219
So it kind of takes the view of like,

1013
00:48:32,559 --> 00:48:34,379
don't bother with products, only focus on models.

1014
00:48:34,699 --> 00:48:36,379
I'm kind of summarizing, simplifying, et cetera.

1015
00:48:37,439 --> 00:48:41,019
But outside of that, like if we do plateau,

1016
00:48:41,239 --> 00:48:42,619
which I'm not claiming we are,

1017
00:48:42,619 --> 00:48:44,799
actually my intuition is that

1018
00:48:44,799 --> 00:48:46,899
doesn't feel like anything's slowing down to me,

1019
00:48:46,959 --> 00:48:49,799
but I'm not actually looking at the base reality

1020
00:48:49,799 --> 00:49:02,218
of model training today But if we are plateauing I think we might consider undoing the bitter lesson and think about all the products that can be built that will end up generating like a major constraint

1021
00:49:02,218 --> 00:49:07,118
is data and so new products that get built can generate new data that can feed into models to

1022
00:49:07,118 --> 00:49:11,457
make better models to do the things that the products are trying to do so maybe that we're

1023
00:49:11,457 --> 00:49:20,118
about to see a bit of like a a product area opening that will give rise to even new new

1024
00:49:20,118 --> 00:49:24,977
abilities to generate data that becomes useful to additional future scaling and when you start to

1025
00:49:24,977 --> 00:49:32,498
talk about slowing down i would ask have you seen or played with nano banana yet no i saw it mentioned

1026
00:49:32,498 --> 00:49:39,698
yesterday but yeah tell us but it's it's the it's the big new so google gemini deep mind just

1027
00:49:39,698 --> 00:49:43,998
launched nano banana i'm not gonna run down the whole thing but it's basically like an amazing new

1028
00:49:43,998 --> 00:49:49,098
image model where you can you have a lot more controllability in characters and a lot more

1029
00:49:49,098 --> 00:49:53,998
consistency across characters and you can do all kinds of things with generating new scenes from

1030
00:49:53,998 --> 00:49:59,018
different points of view and different types of animation and video stuff so do you have to use

1031
00:49:59,018 --> 00:50:05,258
nano coin to pay for it i don't think you just use the google i thought it was the google block i

1032
00:50:05,258 --> 00:50:10,398
thought it was coming too soon i thought it was a cryptocurrency scam now google has so far stayed

1033
00:50:10,398 --> 00:50:14,518
out of the cryptocurrency scam business and i i expect them to well no i thought that's coming up

1034
00:50:14,518 --> 00:50:15,738
I thought it was a nano.

1035
00:50:18,158 --> 00:50:18,477
No.

1036
00:50:18,718 --> 00:50:19,278
Anyway, nano.

1037
00:50:19,938 --> 00:50:20,278
Yeah, go ahead.

1038
00:50:20,477 --> 00:50:21,318
I'll let you go.

1039
00:50:22,358 --> 00:50:24,977
Space Migration just commented and said,

1040
00:50:25,158 --> 00:50:25,318
Great, man.

1041
00:50:25,398 --> 00:50:26,198
Greetings from Seattle.

1042
00:50:26,477 --> 00:50:27,718
Love the ideas you guys talk about.

1043
00:50:27,898 --> 00:50:28,578
It's all of us and always.

1044
00:50:28,738 --> 00:50:29,018
Thanks.

1045
00:50:29,598 --> 00:50:30,698
We appreciate you listening.

1046
00:50:30,838 --> 00:50:31,098
All right.

1047
00:50:31,238 --> 00:50:32,018
Yes, thank you for listening.

1048
00:50:32,098 --> 00:50:34,558
For others listening, please feel free to comment,

1049
00:50:34,758 --> 00:50:36,258
questions, ideas, thoughts.

1050
00:50:36,418 --> 00:50:37,198
We're happy to share them.

1051
00:50:37,238 --> 00:50:38,398
Yeah, if you think we're wrong, tell us.

1052
00:50:38,438 --> 00:50:39,258
If you think we're right, tell us.

1053
00:50:39,418 --> 00:50:40,318
Like, yeah, I want to hear.

1054
00:50:40,457 --> 00:50:42,238
If you think Max is wrong and Steve is right,

1055
00:50:42,238 --> 00:50:42,818
tell us.

1056
00:50:42,818 --> 00:50:45,518
we love to hear that

1057
00:50:45,518 --> 00:50:46,298
do we hear

1058
00:50:46,298 --> 00:50:48,298
DK talk more about

1059
00:50:48,298 --> 00:50:48,977
playing with the banana

1060
00:50:48,977 --> 00:50:50,998
alright back to your banana

1061
00:50:50,998 --> 00:50:52,098
yeah back to your banana

1062
00:50:52,098 --> 00:50:52,677
anyway

1063
00:50:52,677 --> 00:50:54,838
it's amazing

1064
00:50:54,838 --> 00:50:55,957
it's only the latest thing

1065
00:50:55,957 --> 00:50:56,838
that dropped this week

1066
00:50:56,838 --> 00:50:57,258
you know

1067
00:50:57,258 --> 00:50:58,418
the VO3 stuff

1068
00:50:58,418 --> 00:50:59,078
has been amazing

1069
00:50:59,078 --> 00:51:00,138
there's just

1070
00:51:00,138 --> 00:51:01,018
week after week

1071
00:51:01,018 --> 00:51:01,898
there's always something

1072
00:51:01,898 --> 00:51:02,498
new and amazing

1073
00:51:02,498 --> 00:51:03,078
coming out

1074
00:51:03,078 --> 00:51:03,618
and so

1075
00:51:03,618 --> 00:51:04,758
you could say

1076
00:51:04,758 --> 00:51:06,218
GPT-5 wasn't amazing

1077
00:51:06,218 --> 00:51:07,058
and that's kind of

1078
00:51:07,058 --> 00:51:07,778
why everybody's a little

1079
00:51:07,778 --> 00:51:09,158
depressed about the model thing

1080
00:51:09,158 --> 00:51:10,078
I don't know

1081
00:51:10,078 --> 00:51:10,358
maybe

1082
00:51:10,358 --> 00:51:10,977
but like

1083
00:51:10,977 --> 00:51:12,018
I don't feel like

1084
00:51:12,018 --> 00:51:16,998
the world is about to just hit a roadblock and slow down. I think it might pivot or bump,

1085
00:51:17,477 --> 00:51:21,738
you know, bob and weave a little bit, but I think there's still so much new stuff and so much to do.

1086
00:51:22,238 --> 00:51:27,957
And I think it's pretty great. Though I'll just sort of footnote on the stuff we were talking

1087
00:51:27,957 --> 00:51:32,618
about before is I do share the concerns that Max, you were sort of raising the kind of,

1088
00:51:33,018 --> 00:51:36,998
you know, this can be, you know, what did you say, a thousand times worse than kind of the Facebook

1089
00:51:36,998 --> 00:51:43,178
era. I totally agree with that problem. I don't know how we're going to avoid it. I would like to

1090
00:51:43,178 --> 00:51:51,398
come up with solutions. I'm not convinced that just having one layer of the ecosystem, like have

1091
00:51:51,398 --> 00:51:55,118
these kind of good principles is necessarily going to cause everybody to align around that.

1092
00:51:55,118 --> 00:52:01,957
Because I think there's a challenge that the, like the tools that we use with those models

1093
00:52:01,957 --> 00:52:05,778
are kind of co-evolving and improving with them.

1094
00:52:05,858 --> 00:52:09,418
So for example, I use Cloud Code basically all day, every day,

1095
00:52:09,578 --> 00:52:13,338
and it only works with Cloud models.

1096
00:52:13,578 --> 00:52:14,798
I recognize that.

1097
00:52:14,998 --> 00:52:17,378
I recognize the sort of corruptibility and mind control

1098
00:52:17,378 --> 00:52:18,638
that can come out of such a thing.

1099
00:52:19,038 --> 00:52:21,238
But it actually is by far the best thing

1100
00:52:21,238 --> 00:52:22,477
that I've been able to find

1101
00:52:22,477 --> 00:52:25,477
after exploring a dozen or so different tools and models.

1102
00:52:26,158 --> 00:52:26,957
And so I use it every day.

1103
00:52:27,018 --> 00:52:27,738
I'm happy to change.

1104
00:52:27,977 --> 00:52:29,638
I'll use something else if you come up with something better.

1105
00:52:29,638 --> 00:52:37,138
but it's hard to say i'm going to stunt my own abilities out of some sort of a near-term

1106
00:52:37,138 --> 00:52:43,738
i agree desire for sovereignty or desire for privacy and i want to be clear there's a couple

1107
00:52:43,738 --> 00:52:47,218
things you bring up that that are really interesting jumping off points one of them is

1108
00:52:47,218 --> 00:52:51,078
like you know i think a lot of people are on the internet it's oh privacy it's these weird

1109
00:52:51,078 --> 00:52:55,558
cypherpunks like let's be very very clear when we're talking about here guys this is the frog

1110
00:52:55,558 --> 00:52:59,858
boiling in water. I'm not saying like, you know, whatever, like, you know, I want to be a whatever

1111
00:52:59,858 --> 00:53:04,178
using mixers on the dark web and all this weird stuff. No, I'm saying just like, imagine your own

1112
00:53:04,178 --> 00:53:11,638
life. Talk to the AI and you say, hey, I, you know, help me help me make this food for my family.

1113
00:53:11,758 --> 00:53:15,198
And at first it's really good and gives you all these fresh and wonderful ingredients.

1114
00:53:15,338 --> 00:53:19,998
And the next day, you know, and like, forget like the evil insidious pieces that may or may not be

1115
00:53:19,998 --> 00:53:24,318
there for certain players. Let's just say it follows very similar market incentives to what

1116
00:53:24,318 --> 00:53:27,838
Facebook and Google did in their era. And they say, you know what? We really need to pay for this

1117
00:53:27,838 --> 00:53:32,818
like $2 trillion of CapEx. We're never going to get back. Our investors are finally waking up on

1118
00:53:32,818 --> 00:53:38,938
that. And we can't find other business models. So we're going back to ads. And okay, we're going

1119
00:53:38,938 --> 00:53:42,957
to start off pretty good though. We don't want to like totally abuse the process, but like, you know

1120
00:53:42,957 --> 00:53:49,218
what? Hershey's is going to pay us $200 million. And look, there's 20 ingredients in there. We're

1121
00:53:49,218 --> 00:53:53,957
just going to change one. And that natural cacao that you were getting from Mexico that was sugar,

1122
00:53:53,957 --> 00:53:58,898
no sugar in it was good for you good for your body made you feel whole just a little bit of

1123
00:53:58,898 --> 00:54:02,898
hershey's come on one hershey's what is that really gonna hurt you and you know just a little

1124
00:54:02,898 --> 00:54:10,758
yeah yes and then all of a sudden you know the next week it's like well we got 200 million dollars

1125
00:54:10,758 --> 00:54:19,138
from hershey's you know look monsanto they're not that bad really they're not that bad are they nah

1126
00:54:19,138 --> 00:54:23,178
come on they have they have a large budget they have a large budget and now they're gonna pay us

1127
00:54:23,178 --> 00:54:28,178
$2 billion. And look, man, Wall Street, man, Wall Street is, they don't give two fucks. Like,

1128
00:54:28,278 --> 00:54:32,638
we got to pay them back. But look, I swear this is the last one. We're not going to do it again.

1129
00:54:33,378 --> 00:54:38,578
And then that keeps going until, you know, 20 years out, you know, all of your recipes are

1130
00:54:38,578 --> 00:54:44,358
brought to you by, you know, companies that now leave you diabetic, overweight, trapped in your

1131
00:54:44,358 --> 00:54:51,418
home, plugged into a VR. Like, it's a slippery slope. And so my point is, the more you have

1132
00:54:51,418 --> 00:54:56,798
this tool, this assistant engaged in your daily life, it can get real dystopian real fast.

1133
00:54:57,318 --> 00:54:57,578
And so-

1134
00:54:57,578 --> 00:54:58,338
I totally agree.

1135
00:54:58,418 --> 00:55:03,198
But I think we need to figure out what are the ways to solve that before that crisis occurs?

1136
00:55:03,198 --> 00:55:06,558
Because right now we're on the default path that that crisis will occur.

1137
00:55:06,858 --> 00:55:09,258
You're seeing 10 years into the future.

1138
00:55:09,438 --> 00:55:09,698
I know.

1139
00:55:09,698 --> 00:55:10,778
I completely agree with it.

1140
00:55:10,878 --> 00:55:14,398
But you talk about the idea, and I actually do this myself.

1141
00:55:14,498 --> 00:55:18,818
I subscribe to, I don't know what, five or six of the premium LLM services, and I just

1142
00:55:18,818 --> 00:55:25,218
rotate my use among them, not out of necessarily like a privacy concern, though I could understand

1143
00:55:25,218 --> 00:55:29,358
why that would be. I feel like it's nice to not have all of my stuff in one place,

1144
00:55:29,358 --> 00:55:32,818
but I'm just curious to see what's the state of the art, what's the best stuff available.

1145
00:55:32,977 --> 00:55:36,938
And I want to understand the trade-offs of which things are better at maybe coding,

1146
00:55:37,058 --> 00:55:39,418
which things are better at math, which things are better at drawing things.

1147
00:55:39,718 --> 00:55:44,838
I love to understand what's the frontier of what's available. But I do notice that

1148
00:55:44,838 --> 00:55:50,918
when I repeatedly use one, it pulls from a memory that it's created. So if you are always fuzzing

1149
00:55:50,918 --> 00:55:55,678
and mixing your stuff, then you actually lose that. And you're going to find that people say,

1150
00:55:55,798 --> 00:56:01,438
actually, it can reference this conversation I had last week and add that to context on the new

1151
00:56:01,438 --> 00:56:06,238
thing we're talking about. That's better. I'm going to keep using that. So the only thing I'd

1152
00:56:06,238 --> 00:56:11,538
say is don't just advocate for the technology primitive to be right. You actually need to

1153
00:56:11,538 --> 00:56:18,558
advocate for products to exist that can match or be better than the ones that are basically going

1154
00:56:18,558 --> 00:56:23,898
to run away with your mind. And I don't know how that's supposed to work. Let's talk about future

1155
00:56:23,898 --> 00:56:28,538
solutions, right? Because to me, and I'll say this, I'm also very bearish on the products. Like,

1156
00:56:28,538 --> 00:56:32,977
I actually 100% agree with Elon on this. Like, I'm to the point where it's like, I think software,

1157
00:56:33,218 --> 00:56:36,758
like, I mean, good for the teams that are building now, they're going to make, you know, a lot of

1158
00:56:36,758 --> 00:56:41,078
money while the, like, short term bubble, like, there is no bubble long term. But like, yeah,

1159
00:56:41,078 --> 00:56:45,658
that's fine. But like, I agree with Elon, like the phone or whatever hardware you're using,

1160
00:56:45,658 --> 00:56:48,638
it's just going to become a place to run the model. You talk to it. Like you don't need apps

1161
00:56:48,638 --> 00:56:53,698
anymore. You don't need software. Like I think that era is kind of done. And like the old world

1162
00:56:53,698 --> 00:56:58,398
investors are just like, I mean, again, if you can, it's like the crypto, if you get out of it

1163
00:56:58,398 --> 00:57:02,878
fast enough, you're fine. But like, if you're an investor that wants to hold for 25 years,

1164
00:57:02,918 --> 00:57:05,758
that's not where I would be looking, right? Like I would be looking at things that are scarce.

1165
00:57:05,758 --> 00:57:10,918
We talked about that. But anyways, but, but my, my broader point is, so how you compete on the

1166
00:57:10,918 --> 00:57:15,378
model game. Right. I mean, it's beautiful that we have people that are building these open routers.

1167
00:57:15,477 --> 00:57:20,858
Like again, I think the, like, I see two paths right now. One path is it's kind of the, like,

1168
00:57:21,518 --> 00:57:26,818
you know, the three paths you use open source and you, you just say, you know what? We hope

1169
00:57:26,818 --> 00:57:30,078
that they're good enough. And I'm going to make that straight off. Some people will make that.

1170
00:57:30,078 --> 00:57:34,418
And maybe in the long run, it's worth it. And they're going to be the Freemans or whatever.

1171
00:57:34,598 --> 00:57:38,418
And from, from doing series that the people that still have thinking capability in 25 years and

1172
00:57:38,418 --> 00:57:43,738
good for them. I'm glad services like Maple exist. That's an option. Another option is we figure out

1173
00:57:43,738 --> 00:57:48,778
a way to just kind of like confuse the system, which is the sort of like routers or like sort of

1174
00:57:48,778 --> 00:57:52,678
mixers and anonymizers for all these models. I think that's a promising route. I think Noster

1175
00:57:52,678 --> 00:57:56,598
and some of the companies that are building in that ecosystem could be exploring that.

1176
00:57:57,078 --> 00:57:57,638
A third option.

1177
00:57:57,778 --> 00:57:58,738
Pay with Bitcoin too.

1178
00:57:59,118 --> 00:57:59,338
Yeah.

1179
00:57:59,418 --> 00:58:01,238
Let's dive into that when you're done with your options.

1180
00:58:01,558 --> 00:58:07,278
Totally. And then the third one that I see is, and I'm actually surprised no one has even

1181
00:58:07,278 --> 00:58:12,198
discuss this. And I get why, because it's just like, it's a game of kings, to quote Bill Gurley,

1182
00:58:12,278 --> 00:58:17,758
which I stuck with me. I would actually say it's a game of pirates. But just build our own.

1183
00:58:18,538 --> 00:58:21,358
Now, how are you going to build your own? I mean, like, yeah, you need it. I mean,

1184
00:58:21,498 --> 00:58:25,718
Zuckerberg is throwing down like $100 million bonuses to hire the best people. So like,

1185
00:58:26,098 --> 00:58:33,818
how could people that believe in this ethos even dare to dream of an open source model,

1186
00:58:33,818 --> 00:58:37,358
which, as one example, instead of just open sourcing the weights, open sources the entire

1187
00:58:37,358 --> 00:58:43,778
data set. I don't know the answer to that. I mean, one possibility I would say is there's a lot of

1188
00:58:43,778 --> 00:58:49,078
really, really, really wealthy people in Bitcoin. Like, imagine if you had some companies like

1189
00:58:49,078 --> 00:58:54,457
Tether, who's already, you know, at least directionally going this direction. And frankly,

1190
00:58:54,718 --> 00:58:58,138
a bunch of the whales that just came forward and said, whether this is for profit, non-profit,

1191
00:58:58,198 --> 00:59:03,798
I don't know. But like, you know what, some of the richest humans alive are Bitcoiners. They're

1192
00:59:03,818 --> 00:59:05,198
This is a project to save humanity.

1193
00:59:06,038 --> 00:59:06,818
That's one idea.

1194
00:59:07,198 --> 00:59:07,998
What would that look like

1195
00:59:07,998 --> 00:59:09,198
if you could raise $10 billion

1196
00:59:09,198 --> 00:59:11,098
from Tether and a bunch of whales,

1197
00:59:11,477 --> 00:59:12,178
for-profit, non-profit?

1198
00:59:12,258 --> 00:59:12,798
I don't know.

1199
00:59:13,138 --> 00:59:14,138
But it's like, literally like,

1200
00:59:14,258 --> 00:59:16,198
this is our act

1201
00:59:16,198 --> 00:59:18,818
to try and let human consciousness

1202
00:59:18,818 --> 00:59:19,878
have a shot at freedom.

1203
00:59:20,218 --> 00:59:23,018
We have a new 2026 event idea

1204
00:59:23,018 --> 00:59:24,078
for Presbyterian Bitcoin.

1205
00:59:24,178 --> 00:59:24,658
I love it.

1206
00:59:24,738 --> 00:59:24,878
Right?

1207
00:59:25,838 --> 00:59:26,238
What's that?

1208
00:59:26,678 --> 00:59:27,578
I'm being serious.

1209
00:59:28,138 --> 00:59:28,758
Let's do it.

1210
00:59:28,838 --> 00:59:30,058
How can we build our own model?

1211
00:59:30,218 --> 00:59:30,738
Idle the event.

1212
00:59:30,918 --> 00:59:31,998
Is that what it is?

1213
00:59:32,078 --> 00:59:33,178
Is it how do you build your own model?

1214
00:59:33,818 --> 00:59:42,058
sorry what's the event idea i just want to make sure i can describe the event is what max said

1215
00:59:42,058 --> 00:59:49,358
i mean it's trying it's getting together wealthy bitcoiners to share this idea and fund it yeah

1216
00:59:49,358 --> 00:59:52,538
and figure out like hopefully there's like a for-profit model that makes it sustainable

1217
00:59:52,538 --> 00:59:59,098
wealthy bitcoiners and companies like tether like i mean that is the good news right is that and i

1218
00:59:59,098 --> 01:00:03,418
think something that's under discussed in bitcoin is the fact that like some really like frankly

1219
01:00:03,418 --> 01:00:11,118
kind of out there people got unbelievably wealthy from this i mean they a lot of them probably still

1220
01:00:11,118 --> 01:00:16,378
have the values of like they want to see human freedom so i almost feel like if you want to build

1221
01:00:16,378 --> 01:00:22,358
this product today you should not a model a model you want to build a brand new model correct from

1222
01:00:22,358 --> 01:00:31,638
scratch okay so for so with privacy david you mentioned you use a lot of different llms i do

1223
01:00:31,638 --> 01:00:33,198
as well for the same reason you mentioned.

1224
01:00:35,138 --> 01:00:38,038
But not a primary reason we're doing that

1225
01:00:38,038 --> 01:00:39,977
could help with privacy.

1226
01:00:40,898 --> 01:00:43,938
But the thing is we've paid with credit card

1227
01:00:43,938 --> 01:00:46,158
for each of those with our identity, right?

1228
01:00:46,558 --> 01:00:48,198
But what if we paid with Bitcoin

1229
01:00:48,198 --> 01:00:49,598
and we didn't share our identity?

1230
01:00:50,238 --> 01:00:52,358
Then the only privacy leak you need to worry about

1231
01:00:52,358 --> 01:00:54,718
is what you're actually providing in the prompts,

1232
01:00:54,718 --> 01:00:57,258
which definitely can dox you.

1233
01:00:58,038 --> 01:00:59,178
So we'd need to solve that.

1234
01:00:59,278 --> 01:01:01,278
But I do think that would be a reason

1235
01:01:01,278 --> 01:01:03,578
for people to pay in Bitcoin.

1236
01:01:03,878 --> 01:01:06,538
So any privacy-sensitive consumer,

1237
01:01:07,138 --> 01:01:08,578
it would be great to have something

1238
01:01:08,578 --> 01:01:11,058
that's bundled and works that way.

1239
01:01:11,378 --> 01:01:13,038
So imagine if you're using Goose,

1240
01:01:13,198 --> 01:01:15,578
which just integrated an open router,

1241
01:01:15,858 --> 01:01:16,498
not open router,

1242
01:01:16,578 --> 01:01:17,238
but a different router,

1243
01:01:17,418 --> 01:01:18,578
but some kind of router.

1244
01:01:19,318 --> 01:01:22,638
And imagine you have a wallet within Goose

1245
01:01:22,638 --> 01:01:25,018
or it uses NWC and connect it to a wallet.

1246
01:01:25,018 --> 01:01:26,998
So Goose has a budget now

1247
01:01:26,998 --> 01:01:29,838
that you've set for it to pay.

1248
01:01:29,838 --> 01:01:32,178
and then it just

1249
01:01:32,178 --> 01:01:34,378
it farms out queries

1250
01:01:34,378 --> 01:01:35,498
to different LLMs

1251
01:01:35,498 --> 01:01:38,538
and it does pay per use

1252
01:01:38,538 --> 01:01:39,758
and pays in Bitcoin.

1253
01:01:40,218 --> 01:01:41,658
It seems like that would enhance privacy.

1254
01:01:42,838 --> 01:01:43,658
What do we think?

1255
01:01:44,138 --> 01:01:46,258
I mean, I generally like

1256
01:01:46,258 --> 01:01:48,698
the general contours

1257
01:01:48,698 --> 01:01:49,598
of what you're describing.

1258
01:01:50,098 --> 01:01:52,778
I worry about how

1259
01:01:52,778 --> 01:01:54,598
I guess are you working with

1260
01:01:54,598 --> 01:01:56,338
a single interface

1261
01:01:56,338 --> 01:01:57,898
and you're attaching a wallet

1262
01:01:57,898 --> 01:01:59,758
that you control to that interface

1263
01:01:59,758 --> 01:02:01,198
and you're sort of trusting that interface

1264
01:02:01,198 --> 01:02:05,138
is managing your privacy concerns properly

1265
01:02:05,138 --> 01:02:08,578
and dispatching and kind of cleansing data properly?

1266
01:02:09,438 --> 01:02:09,457
Yes.

1267
01:02:10,298 --> 01:02:11,338
And to be clear...

1268
01:02:11,338 --> 01:02:12,138
Concrete? Oh, go ahead.

1269
01:02:12,338 --> 01:02:13,938
Well, this already kind of exists.

1270
01:02:14,098 --> 01:02:15,118
If you guys, if you tried Routester,

1271
01:02:15,318 --> 01:02:17,158
it's like the model router, but on Noster,

1272
01:02:17,457 --> 01:02:18,518
there's PPQ as well.

1273
01:02:18,598 --> 01:02:19,957
Like they don't collect any identity.

1274
01:02:20,378 --> 01:02:22,218
So even if they wanted to,

1275
01:02:22,218 --> 01:02:23,598
I guess they could save your IP or something.

1276
01:02:24,398 --> 01:02:26,238
I guess what I would want to see though

1277
01:02:26,238 --> 01:02:29,098
to make it at least sort of on par

1278
01:02:29,098 --> 01:02:30,418
with some of the best experiences

1279
01:02:30,418 --> 01:02:33,418
is that when you issue a prompt

1280
01:02:33,418 --> 01:02:35,198
and you get a result,

1281
01:02:35,318 --> 01:02:37,618
you now have a bunch of context on me.

1282
01:02:37,818 --> 01:02:38,477
What did I want?

1283
01:02:38,578 --> 01:02:39,198
What did I get?

1284
01:02:39,598 --> 01:02:41,457
Maybe subsequent prompts

1285
01:02:41,457 --> 01:02:42,718
on that chat session.

1286
01:02:43,318 --> 01:02:45,758
In that, there are facts about me

1287
01:02:45,758 --> 01:02:46,638
and what I care about.

1288
01:02:47,078 --> 01:02:49,957
And the way these systems work today

1289
01:02:49,957 --> 01:02:51,218
is those facts get extracted

1290
01:02:51,218 --> 01:02:53,578
as part of the memory of the system.

1291
01:02:54,178 --> 01:02:55,838
Those facts could be stored

1292
01:02:55,838 --> 01:02:58,058
in some sort of a private repository.

1293
01:02:58,058 --> 01:03:03,158
Maybe it's a, maybe it's, maybe it's on Nostra, but not publicly available.

1294
01:03:03,418 --> 01:03:07,358
Maybe it's more like on a Nostra relay, but only accessible via my private key.

1295
01:03:08,658 --> 01:03:16,598
But those could be used then for subsequent dispatches to a separate LLM so that there's some of that context or memory gets carried.

1296
01:03:17,318 --> 01:03:23,158
But I feel like that's one of the core things that, that the centralized services are really good at.

1297
01:03:23,618 --> 01:03:25,938
It could just be stored within Goose, right?

1298
01:03:26,078 --> 01:03:26,558
Yeah.

1299
01:03:26,618 --> 01:03:28,038
Or just store it with your Nostra.

1300
01:03:28,058 --> 01:03:30,457
identity. You could just create an ephemeral identity.

1301
01:03:30,658 --> 01:03:32,378
Or not. I get your point. You want to

1302
01:03:32,378 --> 01:03:34,138
store the context. You just run Goose

1303
01:03:34,138 --> 01:03:36,358
and it's stored on your computer. And then you can

1304
01:03:36,358 --> 01:03:37,998
encrypt it and back it up if you want.

1305
01:03:38,338 --> 01:03:40,238
Or you could do it in the cloud just with

1306
01:03:40,238 --> 01:03:42,218
an austro identity. And like, hey, we don't know who

1307
01:03:42,218 --> 01:03:44,118
this person is. We just know it's a hub key.

1308
01:03:44,378 --> 01:03:46,038
But you need a tool that

1309
01:03:46,038 --> 01:03:48,178
knows to extract those facts

1310
01:03:48,178 --> 01:03:50,178
and save it somewhere and submit them

1311
01:03:50,178 --> 01:03:51,698
at the appropriate times.

1312
01:03:52,378 --> 01:03:54,158
Goose. I mean, we're speaking

1313
01:03:54,158 --> 01:03:55,818
to a roadmap, but this would be a roadmap

1314
01:03:55,818 --> 01:03:58,038
for Goose. Goose is open source.

1315
01:03:58,058 --> 01:03:59,018
that's client-side.

1316
01:03:59,558 --> 01:04:00,778
That already exists today.

1317
01:04:01,178 --> 01:04:03,638
It just integrated a router,

1318
01:04:04,538 --> 01:04:07,718
which I haven't used that.

1319
01:04:07,778 --> 01:04:08,598
I've used OpenRouter,

1320
01:04:08,838 --> 01:04:09,977
in which I had to establish

1321
01:04:09,977 --> 01:04:11,098
a customer relationship

1322
01:04:11,098 --> 01:04:12,758
for OpenRouter out of band.

1323
01:04:13,538 --> 01:04:14,477
I haven't used it yet.

1324
01:04:14,538 --> 01:04:15,798
My understanding of the new one

1325
01:04:15,798 --> 01:04:17,258
that's integrated into Goose,

1326
01:04:17,358 --> 01:04:19,318
it's just more seamless experience,

1327
01:04:19,438 --> 01:04:20,638
but I'm sure you still have to like

1328
01:04:20,638 --> 01:04:21,998
credit card, KYC.

1329
01:04:22,718 --> 01:04:23,298
So you can imagine,

1330
01:04:23,358 --> 01:04:24,558
but maybe like Routster or something,

1331
01:04:24,598 --> 01:04:25,457
you can imagine one

1332
01:04:25,457 --> 01:04:26,998
that doesn't require KYC.

1333
01:04:26,998 --> 01:04:31,598
You can imagine Goose either integrating a Bitcoin wallet within Goose or use NWC.

1334
01:04:31,778 --> 01:04:34,318
Either way, then it has money.

1335
01:04:35,298 --> 01:04:39,198
Somehow you have a budget, you set constraints so it doesn't blow all your Bitcoin.

1336
01:04:41,498 --> 01:04:46,818
And then it stores all of your prompts, all the results locally.

1337
01:04:47,178 --> 01:04:49,618
It can back it up for you, encrypted.

1338
01:04:50,138 --> 01:04:53,758
And then it uses all that information in future queries.

1339
01:04:54,118 --> 01:04:55,278
So you can imagine all that.

1340
01:04:55,278 --> 01:04:57,738
And usually I think consumers don't care about privacy.

1341
01:04:57,898 --> 01:05:08,598
But as I said earlier, with this, I think there's quite a few sensitive areas where people are like normal consumers are going to care about privacy.

1342
01:05:09,457 --> 01:05:14,977
So you could market a tool to them that's turnkey packaged up.

1343
01:05:15,178 --> 01:05:16,998
You don't need to know any of the details I just mentioned.

1344
01:05:17,098 --> 01:05:18,558
It just sort of works out of the box.

1345
01:05:19,938 --> 01:05:22,318
But it's actually, I think it would be marketable though.

1346
01:05:22,318 --> 01:05:43,498
If it is calling, though, I think it's tricky. I like all of that. I think that it's a great direction to explore, but I'm still concerned that it's hard to know what of those facts that are backed up and encrypted should be submitted to which private, state-of-the-art model.

1347
01:05:43,498 --> 01:05:49,278
and that can actually leak a lot of data into that state-of-the-art model

1348
01:05:49,278 --> 01:05:51,138
that's private closed source.

1349
01:05:51,498 --> 01:05:54,738
So how do you sort of ensure that the private data that you're submitting

1350
01:05:54,738 --> 01:05:57,798
is not used?

1351
01:05:57,977 --> 01:05:59,398
I mean, honestly, it will be used.

1352
01:05:59,658 --> 01:06:00,178
It's a good question.

1353
01:06:00,477 --> 01:06:02,298
But that seems like a huge value add.

1354
01:06:02,298 --> 01:06:10,018
Like client-side, it could be smart about analyzing both your prompts

1355
01:06:10,018 --> 01:06:13,618
and the results highlighting how you're doxing yourself.

1356
01:06:14,898 --> 01:06:17,078
And actually, even before you submit the prompt,

1357
01:06:17,438 --> 01:06:20,918
suggest, can you word it a different way?

1358
01:06:22,158 --> 01:06:25,238
And suggest that alternative wording.

1359
01:06:26,738 --> 01:06:28,858
And then be smart about when it's combining

1360
01:06:28,858 --> 01:06:31,898
these different outputs for some future prompt

1361
01:06:31,898 --> 01:06:34,098
or context for a prompt,

1362
01:06:34,098 --> 01:06:37,778
eliminating

1363
01:06:37,778 --> 01:06:40,618
PII or whatever.

1364
01:06:41,318 --> 01:06:42,898
So I think that seems

1365
01:06:42,898 --> 01:06:46,998
like there's room to explore there

1366
01:06:46,998 --> 01:06:48,058
and maybe make a dent.

1367
01:06:48,138 --> 01:06:49,198
I think it's good to explore

1368
01:06:49,198 --> 01:06:50,158
and you can make a dent, yeah.

1369
01:06:50,698 --> 01:06:52,098
And I would love to see,

1370
01:06:52,198 --> 01:06:53,158
I mean, I know we talked about

1371
01:06:53,158 --> 01:06:55,298
Goose funding open source developers

1372
01:06:55,298 --> 01:06:55,738
and whatever,

1373
01:06:55,918 --> 01:06:58,078
like this would be so cool

1374
01:06:58,078 --> 01:07:00,398
if we could get like 10 people

1375
01:07:00,398 --> 01:07:02,038
here with us at Presidio Bitcoin.

1376
01:07:02,038 --> 01:07:04,658
I think we can make a lot of progress in person.

1377
01:07:05,178 --> 01:07:08,658
The fund a new model, that's a big grand dream vision.

1378
01:07:09,118 --> 01:07:11,877
But this is something that I think we can make progress on today.

1379
01:07:12,038 --> 01:07:12,877
I agree.

1380
01:07:13,158 --> 01:07:15,738
Also, bringing it back to Mabel AI,

1381
01:07:15,977 --> 01:07:19,218
or just the TE model in general,

1382
01:07:19,957 --> 01:07:22,858
which we said today, it's a really nice privacy model,

1383
01:07:23,198 --> 01:07:25,638
but since it doesn't have access to the best models,

1384
01:07:26,638 --> 01:07:27,998
that's a disadvantage with it.

1385
01:07:27,998 --> 01:07:32,418
but there can be some tasks you can farm out to that

1386
01:07:32,418 --> 01:07:35,038
and it's totally fine and others you can't,

1387
01:07:35,078 --> 01:07:38,198
but it could be part of this privacy mix as well, right?

1388
01:07:38,198 --> 01:07:43,698
Some of your prompts go to Maple AI

1389
01:07:43,698 --> 01:07:47,358
using five or six or seven models that they support

1390
01:07:47,358 --> 01:07:49,738
and that'll meet your needs.

1391
01:07:49,898 --> 01:07:52,457
And if you ever need a beefier model,

1392
01:07:53,058 --> 01:07:55,638
then you can go to those other models

1393
01:07:55,638 --> 01:07:57,698
and just be sensitive with what, you know,

1394
01:07:57,998 --> 01:07:59,377
the sort of filter,

1395
01:07:59,678 --> 01:08:01,898
have it like auto filter what you send to it.

1396
01:08:02,358 --> 01:08:05,218
One other random thought on the big,

1397
01:08:05,438 --> 01:08:08,278
like human freedom model.

1398
01:08:08,438 --> 01:08:11,118
I wonder if Apple could be an ally here.

1399
01:08:11,477 --> 01:08:13,877
I mean, Apple is so far behind on the model stuff.

1400
01:08:14,318 --> 01:08:16,018
I mean, their whole brand is their own privacy.

1401
01:08:16,178 --> 01:08:17,518
They're the ones even created all this

1402
01:08:17,518 --> 01:08:19,398
like encrypted cloud T stuff, I think,

1403
01:08:19,398 --> 01:08:20,578
or at least they popularized it.

1404
01:08:21,118 --> 01:08:22,698
I wonder if there would be some way

1405
01:08:22,698 --> 01:08:24,118
to like bring them into that fold

1406
01:08:24,118 --> 01:08:25,938
because maybe this is their chance.

1407
01:08:26,058 --> 01:08:26,578
That's a great idea.

1408
01:08:26,877 --> 01:08:29,058
And since Presidia Bitcoin and the show

1409
01:08:29,058 --> 01:08:31,998
is all about bridging the gap with Silicon Valley,

1410
01:08:32,118 --> 01:08:35,598
we should try to get the 10 people here to build

1411
01:08:35,598 --> 01:08:37,377
and some of them should be from Apple.

1412
01:08:37,618 --> 01:08:38,737
I think that's realistic.

1413
01:08:38,938 --> 01:08:41,838
One of them, I already know without doxing him,

1414
01:08:41,938 --> 01:08:42,758
it's someone that we all know,

1415
01:08:42,998 --> 01:08:44,298
a co-investor in a few things

1416
01:08:44,298 --> 01:08:46,698
who was at Apple doing a lot of this stuff

1417
01:08:46,698 --> 01:08:48,198
and is like hardcore in this world.

1418
01:08:48,938 --> 01:08:50,438
So we could even like, who knows,

1419
01:08:50,498 --> 01:08:52,258
maybe we could even like poach him and have him lead.

1420
01:08:52,438 --> 01:08:53,737
Like, that'd be really cool.

1421
01:08:53,737 --> 01:08:56,038
like to have like a privacy AI lab

1422
01:08:56,038 --> 01:08:57,298
here at Presidio Bitcoin

1423
01:08:57,298 --> 01:08:59,078
with like this dude's state of the art

1424
01:08:59,078 --> 01:09:00,498
and just be like,

1425
01:09:00,857 --> 01:09:02,198
fuck it, we're throwing down the gauntlet.

1426
01:09:02,377 --> 01:09:04,038
Like human freedom will live on.

1427
01:09:04,518 --> 01:09:05,237
All right, Haley.

1428
01:09:06,018 --> 01:09:07,977
Our stand-in producer today, Haley.

1429
01:09:08,178 --> 01:09:09,198
You might even be listening right now.

1430
01:09:09,198 --> 01:09:09,958
I'm not joxing you.

1431
01:09:10,277 --> 01:09:12,018
Give Maxim you an action item there.

1432
01:09:12,058 --> 01:09:12,518
Thank you.

1433
01:09:13,098 --> 01:09:14,078
Raise $10 billion.

1434
01:09:14,777 --> 01:09:19,598
And I'm not a hater on the new model idea.

1435
01:09:19,698 --> 01:09:20,678
I think it's worth exploring.

1436
01:09:20,938 --> 01:09:22,758
But I do think that this line of inquiry

1437
01:09:22,758 --> 01:09:26,777
around sort of how do you make products better is,

1438
01:09:27,917 --> 01:09:31,938
you know, to me, it seems like a better starting point

1439
01:09:31,938 --> 01:09:32,758
because-

1440
01:09:32,758 --> 01:09:32,977
I agree.

1441
01:09:33,178 --> 01:09:34,038
I think you do both, though.

1442
01:09:34,838 --> 01:09:36,798
Yeah, the model stuff, I mean,

1443
01:09:36,998 --> 01:09:41,398
who owns and buys all the GPUs and who operates them

1444
01:09:41,398 --> 01:09:43,518
and over what timeframe do you expect

1445
01:09:43,518 --> 01:09:44,737
to make a return on that?

1446
01:09:44,998 --> 01:09:46,857
And I think there's a lot of-

1447
01:09:46,857 --> 01:09:48,838
I'll provide a Bitcoin address

1448
01:09:48,838 --> 01:09:51,058
and everyone can just send the funds there.

1449
01:09:51,498 --> 01:09:52,518
I'll take care of it.

1450
01:09:52,518 --> 01:09:53,078
Yeah, it's...

1451
01:09:53,078 --> 01:09:54,658
Steve is the trusted executioner.

1452
01:09:55,758 --> 01:09:58,817
By the way, Bocceball just asked,

1453
01:09:58,898 --> 01:10:00,258
what kind of model is this again, guys?

1454
01:10:00,317 --> 01:10:00,877
Just tuned in.

1455
01:10:02,158 --> 01:10:06,098
Well, you should watch what we just talked about.

1456
01:10:06,317 --> 01:10:08,218
I'm not going to rehash everything,

1457
01:10:08,377 --> 01:10:10,078
but we are talking about solutions

1458
01:10:10,078 --> 01:10:11,377
that accommodate all models.

1459
01:10:12,258 --> 01:10:14,558
We also had a crazy idea to create a new model

1460
01:10:14,558 --> 01:10:16,917
funded by wealthy Bitcoiners.

1461
01:10:18,118 --> 01:10:20,258
So pretty much every option you can imagine

1462
01:10:20,258 --> 01:10:21,317
we have just discussed.

1463
01:10:21,317 --> 01:10:22,477
Yeah, we're talking about AI.

1464
01:10:23,138 --> 01:10:24,498
I would not be surprised, by the way.

1465
01:10:24,558 --> 01:10:25,578
Do you remember Global Crossing?

1466
01:10:26,237 --> 01:10:26,438
No.

1467
01:10:26,718 --> 01:10:26,977
Yes.

1468
01:10:27,377 --> 01:10:27,718
What's that?

1469
01:10:27,817 --> 01:10:28,678
This was like...

1470
01:10:28,678 --> 01:10:30,138
Max is too young.

1471
01:10:30,317 --> 01:10:30,938
Max is too young.

1472
01:10:31,138 --> 01:10:33,898
So this was, you know, this company is huge, public.

1473
01:10:34,658 --> 01:10:35,598
Yeah, something like that.

1474
01:10:36,317 --> 01:10:37,178
Sometime around then.

1475
01:10:37,338 --> 01:10:38,998
It's closer to the 70s than it is to now.

1476
01:10:39,338 --> 01:10:39,578
Okay.

1477
01:10:40,498 --> 01:10:43,338
But it was probably, what, late 90s, I think?

1478
01:10:43,737 --> 01:10:44,058
Yeah.

1479
01:10:44,178 --> 01:10:46,158
And it was just like huge infrastructure investment,

1480
01:10:46,498 --> 01:10:49,138
like, you know, dark fiber under the internet

1481
01:10:49,138 --> 01:10:50,638
and massive...

1482
01:10:50,638 --> 01:10:54,718
I think they went public and had a massive IPO and raised tons of capital.

1483
01:10:54,718 --> 01:11:03,058
And ultimately then the thing blew up and there's nothing to see for it except that these, you know, this dark fiber exists.

1484
01:11:03,518 --> 01:11:04,098
It's under sea.

1485
01:11:04,817 --> 01:11:09,758
And then it became very cheap to acquire because like everything around it blew up.

1486
01:11:10,218 --> 01:11:14,118
And I'm not predicting that happens to every LLM company today.

1487
01:11:14,118 --> 01:11:29,958
But I do think that LLMs may have a somewhat similar characteristic where there's tons of investment going in. There's just the nature of capitalism and bubbles. I think it's fine. It is what it is. That's how you move society forward.

1488
01:11:29,958 --> 01:11:36,317
but it could be that a lot of the billions of dollars being spent on model differentiation

1489
01:11:36,317 --> 01:11:41,237
ends up looking like global crossing someday and we're like hey glad we have it but that's just

1490
01:11:41,237 --> 01:11:45,638
based how human language works and we don't you just take it as a given as very cheap and commodity

1491
01:11:45,638 --> 01:11:49,857
and then it's these other layers of building that are actually where the value happens

1492
01:11:49,857 --> 01:11:54,018
i i kind of feel like that's yeah to me it feels like that's kind of where it goes so

1493
01:11:54,018 --> 01:11:58,777
it's a pretty good and i think it's plausible but but i want to make two two points on this

1494
01:11:58,777 --> 01:11:59,958
because I think this is actually very important.

1495
01:12:00,118 --> 01:12:04,357
Number one, this might be the kind of thing

1496
01:12:04,357 --> 01:12:06,118
that's just going to happen naturally, right?

1497
01:12:06,178 --> 01:12:09,218
So for example, China is certainly going to do everything they can

1498
01:12:09,218 --> 01:12:12,498
to drive the cost of the model down to zero

1499
01:12:12,498 --> 01:12:14,578
and then win on the hardware where they're dominant.

1500
01:12:15,098 --> 01:12:15,658
And that's great.

1501
01:12:15,718 --> 01:12:17,438
And I've seen a lot of people be like, yeah, that's awesome.

1502
01:12:17,518 --> 01:12:19,438
In a weird way, we're very aligned with China on that.

1503
01:12:20,018 --> 01:12:22,598
However, I personally am more skeptical

1504
01:12:22,598 --> 01:12:23,838
than a lot of those voices I've heard

1505
01:12:23,838 --> 01:12:25,618
because I'm like, what the hell is in the Chinese models?

1506
01:12:25,737 --> 01:12:27,857
I am not at all convinced, and you cannot convince me

1507
01:12:27,857 --> 01:12:32,438
there's not a poison pill in those models. And unless I can see the entire data set and trust it,

1508
01:12:32,538 --> 01:12:38,538
which by the way, probably not going to trust China. I don't believe that that's, I think that

1509
01:12:38,538 --> 01:12:43,298
it could happen the way you're saying, but unless these guys, for whatever reason, and maybe one of

1510
01:12:43,298 --> 01:12:47,458
them like Anthropic will slip far enough behind. They say, fuck it, we're throwing the whole Mary

1511
01:12:47,458 --> 01:12:51,877
or even Facebook. And we're going to just like show all the data. But again, you have to trust

1512
01:12:51,877 --> 01:12:59,198
them. And so I, um, it could happen on its own, uh, but I'm not sure. And even if it does like

1513
01:12:59,198 --> 01:13:04,338
the dark fiber thing, that was still a like net benefit to humanity. Is there a role where, yeah,

1514
01:13:04,338 --> 01:13:08,498
you try and productize it, you try and make money on it, but like the investors frankly don't care.

1515
01:13:08,737 --> 01:13:12,098
Like, it's like, we just want this thing to exist. We're an open source. We lose all of our money.

1516
01:13:12,098 --> 01:13:16,098
Like, fine. Like this is, this is in the same way that, you know,

1517
01:13:16,218 --> 01:13:29,296
in a sense like deep seek has already kind of done that They spent billions of dollars when Why not just that how the English language works Why not just use and then build on top of it Because what I saying is I don believe that you can trust DeepSeek

1518
01:13:29,856 --> 01:13:31,816
And like, for example, like if there,

1519
01:13:32,696 --> 01:13:34,116
could there be a poison pill?

1520
01:13:34,216 --> 01:13:35,316
Because it's from China.

1521
01:13:35,516 --> 01:13:35,836
Correct.

1522
01:13:37,096 --> 01:13:38,417
I would, I would guess,

1523
01:13:38,576 --> 01:13:40,176
I mean, there certainly could be problems.

1524
01:13:40,296 --> 01:13:41,136
I would guess there's some,

1525
01:13:41,536 --> 01:13:44,076
like if I were to allocate capital towards this,

1526
01:13:44,076 --> 01:13:45,417
I'd rather than say,

1527
01:13:45,556 --> 01:13:46,716
train up a new competitive model,

1528
01:13:46,796 --> 01:13:47,996
I'd allocate capital towards

1529
01:13:47,996 --> 01:13:56,256
there are techniques i know of a specific major ai lab which has part of the team expertise is

1530
01:13:56,256 --> 01:14:01,776
their ability to kind of microsurgically change something deep within the model and understand

1531
01:14:01,776 --> 01:14:06,756
where everything is and make a change when they need to so they don't have to rerun a training

1532
01:14:06,756 --> 01:14:11,176
run which can cost billions of dollars but they can actually have a crack team go in and and sort

1533
01:14:11,176 --> 01:14:17,936
of do this microsurgery on the some of the weights or embeddings and so i would rather say take the

1534
01:14:17,936 --> 01:14:23,216
training model that DeepSeek did and build a layer to be able to do that kind of microsurgery on it

1535
01:14:23,216 --> 01:14:25,676
to figure out if there are these poison pills in it.

1536
01:14:26,076 --> 01:14:31,016
Maybe. And I think that's a very worthwhile avenue to explore and sort of worth like,

1537
01:14:31,096 --> 01:14:36,036
like this is beyond my area of expertise, but my like just initial, if we're shooting big,

1538
01:14:36,036 --> 01:14:40,316
I think you do all of the above. My initial slight pushback on that is I would argue for

1539
01:14:40,316 --> 01:14:43,917
my friends that are sort of stayed there in LLM land and like training them, no one really knows

1540
01:14:43,917 --> 01:14:50,196
how this stuff works. And yeah, you can probably get to some degree, but I think it's still alchemy

1541
01:14:50,196 --> 01:14:56,877
right now. And so my dream world would be we could see from ground up this thing built in open.

1542
01:14:57,136 --> 01:15:02,316
Everything is open. I mean, that requires raising crazy amounts of capital, building a team,

1543
01:15:02,377 --> 01:15:05,636
all these really hard things. So maybe that's too much and what you're saying is more realistic,

1544
01:15:06,056 --> 01:15:10,856
but for shooting for the stars, let's shoot for the stars. And one more thing I want to add,

1545
01:15:10,856 --> 01:15:11,877
because I think this is super important.

1546
01:15:11,996 --> 01:15:13,596
I think the other super critical thing

1547
01:15:13,596 --> 01:15:15,076
that we can't lose sight of in all of this,

1548
01:15:15,176 --> 01:15:17,377
and I'm actually going to amend my essay

1549
01:15:17,377 --> 01:15:19,476
having had this conversation with you guys right now.

1550
01:15:20,176 --> 01:15:21,377
And hopefully by next week,

1551
01:15:21,396 --> 01:15:22,116
we'll be able to announce

1552
01:15:22,116 --> 01:15:24,476
another little Alpha Type 1 Summit

1553
01:15:24,476 --> 01:15:25,917
coming up here at Presidio Bitcoin as well,

1554
01:15:25,976 --> 01:15:28,836
thinking about radical energy infrastructure.

1555
01:15:29,236 --> 01:15:31,116
I think that's going to be extremely important as well.

1556
01:15:31,536 --> 01:15:33,917
The radically decentralized energy infrastructure

1557
01:15:33,917 --> 01:15:35,136
for both training and serving.

1558
01:15:35,456 --> 01:15:36,877
Because the other big challenge is,

1559
01:15:36,936 --> 01:15:38,136
imagine you do raise all this money

1560
01:15:38,136 --> 01:15:39,536
and you can do, okay, cool.

1561
01:15:39,636 --> 01:15:40,576
We can train a new model.

1562
01:15:40,576 --> 01:15:45,476
we can serve the new model, but like you're doing it in these like giant, you know, eventually

1563
01:15:45,476 --> 01:15:51,196
terawatt scale AI factors, right? Like, hmm, like wonder if that's like an easy thing to just shut

1564
01:15:51,196 --> 01:15:57,756
down. The answer is yes, right? And so what I think is also very interesting is it's very important

1565
01:15:57,756 --> 01:16:03,476
that we as a community figure out how to fund and incentivize as much decentralized, eventually

1566
01:16:03,476 --> 01:16:07,776
training that's harder, but serving infrastructure as well. So that could be, you know, obviously,

1567
01:16:07,776 --> 01:16:10,536
many AI factories powered by solar and batteries

1568
01:16:10,536 --> 01:16:12,356
completely off grid. I think there's strong

1569
01:16:12,356 --> 01:16:14,616
economic reasons that's going to happen, but I also

1570
01:16:14,616 --> 01:16:16,377
think as part of this freedom

1571
01:16:16,377 --> 01:16:17,216
AI push,

1572
01:16:17,836 --> 01:16:20,756
that's as important as well.

1573
01:16:22,456 --> 01:16:24,076
It's getting too late for my comments.

1574
01:16:24,316 --> 01:16:25,816
I want to get my comment in.

1575
01:16:26,596 --> 01:16:27,436
I didn't understand.

1576
01:16:29,877 --> 01:16:30,377
Undersea

1577
01:16:30,377 --> 01:16:30,877
cables,

1578
01:16:31,796 --> 01:16:34,356
we know how to verify that the

1579
01:16:34,356 --> 01:16:35,996
information being sent over them is

1580
01:16:35,996 --> 01:16:37,236
trustworthy.

1581
01:16:37,776 --> 01:16:43,036
With software, we have the notion of reproducible builds.

1582
01:16:43,996 --> 01:16:46,756
So they're not super popular right now,

1583
01:16:46,917 --> 01:16:49,456
but it is possible to do a reproducible build

1584
01:16:49,456 --> 01:16:52,256
so that I can independently verify the software I'm running

1585
01:16:52,256 --> 01:16:56,736
is what's open source and published.

1586
01:16:57,016 --> 01:17:02,316
I don't see how to independently verify an AI model,

1587
01:17:02,716 --> 01:17:04,496
even if all the data is published,

1588
01:17:04,496 --> 01:17:07,516
because you weren't there when they did the training.

1589
01:17:07,776 --> 01:17:15,656
And how can any human being independently verify that the training actually took that data input?

1590
01:17:15,996 --> 01:17:27,996
I don't, and similar with hardware, how do you independently verify that the H100 doesn't have some kind of backdoor or something or poison, you know, poisoned in some way?

1591
01:17:28,296 --> 01:17:37,736
I mean, I would say that's true though with, I mean, are you tracing the global crossing dark fiber to where it hits the junction that connects to the rest of it?

1592
01:17:37,776 --> 01:17:46,656
you are trusting right so but what's an example of like if i send you know data over that line

1593
01:17:46,656 --> 01:17:52,556
and someone on the other end receives it what's an example of how it could be tamper i think that's

1594
01:17:52,556 --> 01:17:58,436
fine my point is more that these models are all probabilistic like that's the feature not the bug

1595
01:17:58,436 --> 01:18:04,396
it's a feature that is probabilistic and and it's i'm not proposing that oh like the open ai model

1596
01:18:04,396 --> 01:18:06,776
will eventually become open source or something.

1597
01:18:07,036 --> 01:18:07,716
I'm more saying,

1598
01:18:08,156 --> 01:18:09,736
I think a lot of what these models

1599
01:18:09,736 --> 01:18:10,756
are really good at doing

1600
01:18:10,756 --> 01:18:13,436
is figuring out how English language

1601
01:18:13,436 --> 01:18:15,356
connects ideas and concepts together.

1602
01:18:15,636 --> 01:18:18,536
And that core idea can be built

1603
01:18:18,536 --> 01:18:20,076
again and again many times

1604
01:18:20,076 --> 01:18:21,956
and then eventually kind of get commoditized.

1605
01:18:22,076 --> 01:18:23,816
And then the commoditized one could be open.

1606
01:18:24,196 --> 01:18:25,596
So it's not like you need to have

1607
01:18:25,596 --> 01:18:27,276
open AI's actual training run.

1608
01:18:27,436 --> 01:18:29,156
You just need a bunch of forces

1609
01:18:29,156 --> 01:18:31,776
forcing the state of the art to move forward.

1610
01:18:31,776 --> 01:18:33,976
And then hopefully an open source,

1611
01:18:33,976 --> 01:18:40,076
open weights thing can catch up and we probably do need better like interpretability tools so

1612
01:18:40,076 --> 01:18:44,896
that's one of the things that rather than i mean yes do all of the above plus 10 other things

1613
01:18:44,896 --> 01:18:50,696
agreed but also disagree because like you you have to allocate time and attention and resources and

1614
01:18:50,696 --> 01:18:55,056
stuff but my point is even if it's open like we still don't know if that's the model we're using

1615
01:18:55,056 --> 01:19:01,396
because yes we need sort of interpretability we need new tools and i think those are tools i would

1616
01:19:01,396 --> 01:19:06,136
advocate for let's have great interpretability tools so if we do have um an open source model

1617
01:19:06,136 --> 01:19:13,417
we can find the poison pills like to me that seems like a productive contribution to the world that

1618
01:19:13,417 --> 01:19:18,496
that doesn't exist and that could help us reach the goals we're talking about versus hey let's go

1619
01:19:18,496 --> 01:19:22,436
redo the thing that already exists like we could go rebuild the dark fiber we could rebuild okay

1620
01:19:22,436 --> 01:19:26,917
i mean that sounds really hard to do but okay i agree that's an interesting but by the way we

1621
01:19:26,917 --> 01:19:29,216
we haven't talked about any topics yet.

1622
01:19:29,716 --> 01:19:30,716
Well, we want to talk about, yeah.

1623
01:19:31,096 --> 01:19:32,036
I think we move on.

1624
01:19:32,176 --> 01:19:32,536
Let's move on.

1625
01:19:32,536 --> 01:19:33,596
One last thought on that,

1626
01:19:33,656 --> 01:19:35,756
which is I agree with you,

1627
01:19:36,056 --> 01:19:37,936
but I also don't necessarily agree

1628
01:19:37,936 --> 01:19:38,676
that's even possible.

1629
01:19:40,016 --> 01:19:41,136
If it is going to be possible,

1630
01:19:41,236 --> 01:19:42,596
it's probably going to be AI agents doing it.

1631
01:19:42,656 --> 01:19:44,116
And so that's like a really interesting area

1632
01:19:44,116 --> 01:19:45,536
is like having the AIs do the auditing.

1633
01:19:45,656 --> 01:19:45,956
It's not going to be honest.

1634
01:19:45,956 --> 01:19:48,216
But then how do you trust the AI that's doing that?

1635
01:19:48,276 --> 01:19:49,417
Well, then you would have, exactly.

1636
01:19:49,596 --> 01:19:50,336
And then you say,

1637
01:19:50,396 --> 01:19:52,036
well, we have like a whole bunch of these AIs.

1638
01:19:52,096 --> 01:19:53,396
And if you have it from 10 different companies

1639
01:19:53,396 --> 01:19:53,976
and whatever,

1640
01:19:54,336 --> 01:19:55,436
but my point is like,

1641
01:19:55,456 --> 01:19:56,616
I'm not sure that like,

1642
01:19:56,616 --> 01:19:59,917
at least given our current physics, that's possible.

1643
01:20:01,056 --> 01:20:05,917
Let's talk about Google announced Google Cloud Universal Ledger

1644
01:20:05,917 --> 01:20:08,296
in kind of a unique way.

1645
01:20:09,417 --> 01:20:15,036
The executive in charge of it posted on LinkedIn describing it.

1646
01:20:15,576 --> 01:20:16,576
Strong, strong, strong.

1647
01:20:18,336 --> 01:20:20,236
It seemed like a little bit of like...

1648
01:20:20,236 --> 01:20:21,396
Is that the only place it went out?

1649
01:20:22,377 --> 01:20:22,696
Yeah.

1650
01:20:23,816 --> 01:20:25,976
And there were some articles written and tweets.

1651
01:20:25,976 --> 01:20:30,496
But it seemed like a little bit of FOMO from Stripe and Circle's announcements.

1652
01:20:30,756 --> 01:20:44,496
But we've had three solutions announced from, we talked about Circle Arc and Stripe Tempo in a previous episode, and now Gluck.

1653
01:20:46,116 --> 01:20:47,076
It's called Gluck?

1654
01:20:47,176 --> 01:20:48,036
It's Geekle.

1655
01:20:49,936 --> 01:20:50,576
Geekle.

1656
01:20:51,556 --> 01:20:53,016
Google Cloud Universal Ledger.

1657
01:20:53,516 --> 01:20:55,016
Google is not allocating its...

1658
01:20:55,016 --> 01:20:55,856
Google Cloud Universal Ledger.

1659
01:20:55,976 --> 01:20:58,656
Demis is not working on this project, I can tell.

1660
01:21:01,176 --> 01:21:01,956
Oh, yeah, nice luck.

1661
01:21:02,636 --> 01:21:04,336
The D&L are not next to each other.

1662
01:21:05,216 --> 01:21:05,656
Google.

1663
01:21:06,256 --> 01:21:10,716
But I think my reading of it, it's not a blockchain.

1664
01:21:11,556 --> 01:21:15,196
It's just a database hosted by Google.

1665
01:21:15,377 --> 01:21:15,516
Yeah.

1666
01:21:16,196 --> 01:21:20,256
So much like every other blockchain that's not Bitcoin.

1667
01:21:20,556 --> 01:21:21,076
Interesting.

1668
01:21:21,076 --> 01:21:23,377
Well, no, much more efficient.

1669
01:21:23,377 --> 01:21:27,496
because blockchains are very inefficient databases.

1670
01:21:28,056 --> 01:21:30,856
This presumably is a super efficient database

1671
01:21:30,856 --> 01:21:35,776
because only Google controls it, only Google hosts it.

1672
01:21:37,377 --> 01:21:43,116
And it seems like they're observing the rapid rise

1673
01:21:43,116 --> 01:21:44,377
and growth of stablecoins,

1674
01:21:44,576 --> 01:21:47,156
trillions of dollars being transacted and growing rapidly.

1675
01:21:48,036 --> 01:21:51,356
And this is their entry into this.

1676
01:21:51,356 --> 01:21:57,596
And so it's super interesting to see Google enter the fray.

1677
01:21:58,896 --> 01:22:01,636
It seems to me like it's a credible option.

1678
01:22:01,896 --> 01:22:06,996
Because why do you need the decentralization that Bitcoin has

1679
01:22:06,996 --> 01:22:11,377
when you're just hosting dollars issued by centralized entities

1680
01:22:11,377 --> 01:22:14,176
that are heavily controlled by the US?

1681
01:22:14,776 --> 01:22:19,256
So Google might have the smartest strategy in the space.

1682
01:22:19,256 --> 01:22:23,977
Because everyone else is developing their own layer one or using Solana or Ethereum.

1683
01:22:24,656 --> 01:22:25,977
Super inefficient databases.

1684
01:22:26,176 --> 01:22:29,336
And Google is just like, let's just use a traditional database.

1685
01:22:29,556 --> 01:22:29,656
Fuck it.

1686
01:22:29,676 --> 01:22:30,196
Let's be honest.

1687
01:22:31,056 --> 01:22:31,556
Okay, so.

1688
01:22:31,796 --> 01:22:33,936
They do call it a blockchain, even though it is just a database.

1689
01:22:33,936 --> 01:22:34,556
No, they don't.

1690
01:22:35,156 --> 01:22:35,496
Oh, they don't.

1691
01:22:35,616 --> 01:22:35,736
Okay.

1692
01:22:35,856 --> 01:22:36,676
Some people have called it.

1693
01:22:36,696 --> 01:22:37,696
I don't know if it was in the original post.

1694
01:22:37,696 --> 01:22:37,877
They don't.

1695
01:22:38,196 --> 01:22:42,377
They actually have a, I sent it out to you guys.

1696
01:22:42,576 --> 01:22:43,296
You should look at it.

1697
01:22:43,296 --> 01:22:48,656
They have a fairly lengthy post describing their view of the entire space.

1698
01:22:49,256 --> 01:22:55,136
And I searched for the word blockchain and it's referenced three times, but not describing their product.

1699
01:22:55,496 --> 01:22:59,496
Although I think critically does claim in there that they are, quote, credibly neutral.

1700
01:23:00,076 --> 01:23:06,877
Yeah, I think credibly neutral means something different than what we mean when we say credibly neutral for Bitcoin.

1701
01:23:06,877 --> 01:23:11,276
I think the neutrality is not that it's a neutral database.

1702
01:23:11,616 --> 01:23:16,036
It's that they are a neutral player among the partners.

1703
01:23:16,036 --> 01:23:32,656
There's a very few small set of select partners who are allowed to contribute to it. And they are not in, at least today, they're not in direct competitions with these partners. They're credibly neutral from an infrastructure perspective, but not from like a decentralization or sovereignty perspective.

1704
01:23:33,836 --> 01:23:44,576
So I just want to give a quick shout out. If you guys haven't seen this or read this yet, Alan Farrington's piece, a half-baked thesis on stable coins. Have you guys read that yet?

1705
01:23:44,576 --> 01:23:48,596
I haven't read it. I think you shared it in the chat, right? I need to catch it.

1706
01:23:48,816 --> 01:23:52,516
We can discuss it next week. I got to give a shout out to Alan. Alan is like,

1707
01:23:52,996 --> 01:24:00,896
once again, right? I mean, look, very, very high level. I think this is what we're kind of getting

1708
01:24:00,896 --> 01:24:08,216
at two weeks ago when we were talking about Tempo. If you look at, and this goes back to my old days

1709
01:24:08,216 --> 01:24:11,856
when I used to be in a fund investing in crypto. And I'd always ask the central question. I'd be

1710
01:24:11,856 --> 01:24:17,756
like, okay, well, you know, if you are sacrificing decentralization, so Bitcoin is completely

1711
01:24:17,756 --> 01:24:24,436
decentralized and therefore is a giant pain in the ass to run. Like there've been battles fought over

1712
01:24:24,436 --> 01:24:27,877
the megabyte size of the blocks, like all this stuff to keep it decentralized, right?

1713
01:24:28,296 --> 01:24:34,856
And so what every other crypto project has realized over time is, hey, we can get easier,

1714
01:24:35,096 --> 01:24:38,716
you know, faster throughput, more throughput, all this stuff, we make it a little bit more

1715
01:24:38,716 --> 01:24:43,336
centralized. And so increasingly, then you have Ethereum, which is sort of decentralized, but

1716
01:24:43,336 --> 01:24:46,736
like sort of not. And so it ended up in this Frankenstein state. And then it starts getting

1717
01:24:46,736 --> 01:24:50,116
outcompeted by Solana, which is basically just like, well, we've got, I don't know how many it

1718
01:24:50,116 --> 01:24:54,336
is, 10 nodes and like we're decentralized, you know, whatever. And then they get outcompeted

1719
01:24:54,336 --> 01:24:59,016
because like, who's the number one stable coin ledger that's doing all this? Tron. I don't even

1720
01:24:59,016 --> 01:25:03,076
know if, does Tron even pretend to be a blockchain at this point? Maybe. But they're essentially a

1721
01:25:03,076 --> 01:25:08,796
centralized database. And then you've got, you know, if you go back to 2019, I think one person

1722
01:25:08,796 --> 01:25:13,496
whose take I would value very highly on this is David Marcus. Because David Marcus was at Facebook,

1723
01:25:13,696 --> 01:25:16,956
Libra, that was their project where they were going to try and issue a stablecoin on their

1724
01:25:16,956 --> 01:25:23,616
database. And why is the reason that they didn't do it? Regulation. Because the government and all

1725
01:25:23,616 --> 01:25:27,236
these other players are not like, wait a second, hold up. We're going to let a private company

1726
01:25:27,236 --> 01:25:32,836
controlled by Mark Zuckerberg, who this is very familiar with, controlled the giant database

1727
01:25:32,836 --> 01:25:35,576
of transactions happening in the world?

1728
01:25:35,696 --> 01:25:37,076
No, and that gets shut down.

1729
01:25:38,016 --> 01:25:39,576
What has changed since then?

1730
01:25:39,636 --> 01:25:40,656
And I actually, this is like,

1731
01:25:40,716 --> 01:25:42,977
I found myself literally completing Alan's sentences

1732
01:25:42,977 --> 01:25:44,396
in a podcast about his paper,

1733
01:25:44,877 --> 01:25:46,816
is like the only innovation,

1734
01:25:46,936 --> 01:25:48,696
and to be clear, stable coins are better than dollars,

1735
01:25:49,016 --> 01:25:52,776
but that just shows how antiquated and backwards dollars

1736
01:25:52,776 --> 01:25:53,796
in the dollar system is.

1737
01:25:53,836 --> 01:25:57,316
And that's because it's a heavily regulated field,

1738
01:25:57,316 --> 01:25:58,816
therefore you can't have freedom like the internet,

1739
01:25:58,936 --> 01:26:00,456
which allows innovation to happen quickly.

1740
01:26:00,576 --> 01:26:01,656
So that's why it's so backwards.

1741
01:26:01,656 --> 01:26:06,836
people are realizing this stable coin thing which increasingly is moving from sort of like

1742
01:26:06,836 --> 01:26:11,356
blockchain and name only to just centralized databases are better than the old world yes

1743
01:26:11,356 --> 01:26:16,516
because the old world was built in like cobalt in 1970 or 50 or whenever so like yeah these are

1744
01:26:16,516 --> 01:26:21,816
better but like the only innovation here as alan said is literally just confusing regulators it's

1745
01:26:21,816 --> 01:26:25,996
like throwing so many buzzwords and crypto and blockchain that people are like yeah like we

1746
01:26:25,996 --> 01:26:31,796
should probably allow this thing. And then to their credit, it got adoption. And because it

1747
01:26:31,796 --> 01:26:36,216
got adoption, all of a sudden it's like, well, now it's Heather using Trump. It seems kind of

1748
01:26:36,216 --> 01:26:39,276
too big to shut down. It's like massive. And like, oh, by the way, they're buying our treasuries.

1749
01:26:39,756 --> 01:26:45,616
Yeah. Blockchain, crypto, knock yourself out, guys, keep buying the treasuries. And then even

1750
01:26:45,616 --> 01:26:49,977
more, you know, then you have the Trump administration come in and the Trump administration,

1751
01:26:50,216 --> 01:26:54,576
you know, there's pros and cons. They're, you know, more open and freedom to certain things,

1752
01:26:54,576 --> 01:26:58,056
but it's like kind of in a nihilistic way, in my opinion, where it's just like, fuck it,

1753
01:26:58,116 --> 01:27:01,856
we're going to launch our own tokens, coins, knock yourself out. There's no like ideology

1754
01:27:01,856 --> 01:27:06,416
behind this. And so now there is this like window of time where, and the way I see it,

1755
01:27:06,436 --> 01:27:10,377
what Facebook and David Marcus tried to do in 2019 and got shut down, there's kind of this

1756
01:27:10,377 --> 01:27:13,776
window where it's like, we can just launch a database. And in the old world, that would have

1757
01:27:13,776 --> 01:27:16,316
been heavily regulated. And maybe it still will be, maybe it won't.

1758
01:27:16,456 --> 01:27:20,696
So it's a window of time during the Trump administration, and then who knows in three

1759
01:27:20,696 --> 01:27:22,656
years. But the Genius Act was

1760
01:27:22,656 --> 01:27:24,696
passed by Congress. So that will survive.

1761
01:27:24,896 --> 01:27:26,596
Or I mean, it's much harder to overturn

1762
01:27:26,596 --> 01:27:27,696
that with a new administration.

1763
01:27:29,996 --> 01:27:30,556
So

1764
01:27:30,556 --> 01:27:32,596
that's different.

1765
01:27:33,216 --> 01:27:33,336
But

1766
01:27:33,336 --> 01:27:36,256
I mean, I'm interested in talking about

1767
01:27:36,256 --> 01:27:38,536
the thesis where tech

1768
01:27:38,536 --> 01:27:40,796
companies, fintech companies

1769
01:27:40,796 --> 01:27:42,636
use this

1770
01:27:42,636 --> 01:27:44,596
to usurp banks.

1771
01:27:45,096 --> 01:27:46,676
Yeah. And so we talked

1772
01:27:46,676 --> 01:27:48,877
a little bit prior to the show about

1773
01:27:48,877 --> 01:27:50,596
finding

1774
01:27:50,596 --> 01:27:54,676
ways to share the yield from the

1775
01:27:54,676 --> 01:27:58,836
treasury's backing stable coins. And the GeniusX

1776
01:27:58,836 --> 01:28:02,456
explicitly doesn't allow sharing that yield

1777
01:28:02,456 --> 01:28:06,696
with customers, at least directly. I haven't read the specific language about how

1778
01:28:06,696 --> 01:28:10,396
tight or loose that language is. But I'm sure that language

1779
01:28:10,396 --> 01:28:14,716
was introduced because banks wanted it there. Because banks

1780
01:28:14,716 --> 01:28:18,336
don't want to see people switch from the bank,

1781
01:28:18,336 --> 01:28:21,176
traditional bank to this new system.

1782
01:28:21,816 --> 01:28:24,116
But of course, the stripes and circles

1783
01:28:24,116 --> 01:28:27,336
and Googles of the world, Robin Hoods,

1784
01:28:28,036 --> 01:28:31,877
would obviously love the trillions of dollars

1785
01:28:31,877 --> 01:28:34,776
flow from banks to themselves.

1786
01:28:35,556 --> 01:28:37,076
And then the U.S. government,

1787
01:28:37,436 --> 01:28:39,116
I forget which of you said it,

1788
01:28:39,196 --> 01:28:43,877
but definitely likes the tethers and circles of the world

1789
01:28:43,877 --> 01:28:46,676
being some of the largest buyers of U.S. treasuries

1790
01:28:46,676 --> 01:28:47,916
in the world.

1791
01:28:48,336 --> 01:28:50,656
that's what feeds our system.

1792
01:28:51,016 --> 01:28:54,036
And countries like China are buying much less U.S. treasuries.

1793
01:28:54,356 --> 01:28:55,416
So who's the new buyers?

1794
01:28:56,196 --> 01:28:57,516
So what do we think about that?

1795
01:28:57,656 --> 01:29:00,556
How realistic is it for these tech companies

1796
01:29:00,556 --> 01:29:04,276
to overtake banks through stablecoins

1797
01:29:04,276 --> 01:29:06,256
or what Google's doing?

1798
01:29:08,836 --> 01:29:11,996
I'll outline a specific kind of example

1799
01:29:11,996 --> 01:29:14,576
that I was discussing in the group chat.

1800
01:29:14,916 --> 01:29:16,036
And I want to hear kind of more

1801
01:29:16,036 --> 01:29:17,256
what you guys think about this.

1802
01:29:17,256 --> 01:29:25,256
So I think Genius Act, I think, is very clear that sharing yield to the holders of the stablecoin is disallowed.

1803
01:29:27,556 --> 01:29:39,336
But we know that the yield that you can get on treasuries is probably, there's probably more juice in the rewards that you could share than a traditional credit card interchange fee.

1804
01:29:39,596 --> 01:29:43,276
Is that true? Do we think that's true? Or we mostly believe that's the case?

1805
01:29:43,836 --> 01:29:44,056
Yes.

1806
01:29:44,056 --> 01:29:49,956
Yes. Yeah. So that's why there's a threat, right? Like yield on treasuries is better than

1807
01:29:49,956 --> 01:29:55,916
interchange and the rewards programs to end users come from the interchange fees that merchants pay.

1808
01:29:56,477 --> 01:30:03,156
And so yield would be a juicier pot to pull from. So Genius Act, as you said, specifically disallows

1809
01:30:03,156 --> 01:30:09,396
that yield being shared. But in a sense, these are businesses, they have balance sheets, they have

1810
01:30:09,396 --> 01:30:12,496
all kinds of partnerships and distribution deals.

1811
01:30:12,996 --> 01:30:15,036
So I think we, I forget who mentioned it in the chat,

1812
01:30:15,156 --> 01:30:18,416
but the specific example that seems very credible to me

1813
01:30:18,416 --> 01:30:21,276
is that the Genius Act disallows sharing the yield,

1814
01:30:21,716 --> 01:30:23,436
but the issue of the stable coin

1815
01:30:23,436 --> 01:30:27,556
does a distribution deal with, say, a Robinhood, right?

1816
01:30:27,656 --> 01:30:29,416
It's like kind of a forward-looking thing.

1817
01:30:29,496 --> 01:30:31,196
And they're like, hey, for every customer you acquire,

1818
01:30:31,356 --> 01:30:32,736
everybody who holds our stable coins,

1819
01:30:32,836 --> 01:30:34,816
we get this much in value.

1820
01:30:34,896 --> 01:30:36,496
And then it's like the distribution partner

1821
01:30:36,496 --> 01:30:39,316
who actually gives the benefit.

1822
01:30:39,396 --> 01:30:55,056
So the thing that you get from your credit card company today that's paid by your interchange fee, your Robinhood chooses to give you gold or some made up something or other, some other kind of rewards program that's effectively backed out of the stable coin yield or some share of the stable coin yield.

1823
01:30:55,416 --> 01:31:01,877
And that goes directly to consumers where they're getting the benefit of the yield while skirting the letter of the law.

1824
01:31:01,877 --> 01:31:07,096
Right. And it wouldn't be marketed as, oh, you're getting X percent interest on your deposit.

1825
01:31:07,096 --> 01:31:09,936
it would be framed a very different way,

1826
01:31:10,136 --> 01:31:12,296
but you're effectively getting the same amount of money.

1827
01:31:12,296 --> 01:31:14,076
Correct these points and you get a trip to Hawaii.

1828
01:31:14,436 --> 01:31:14,616
Yeah.

1829
01:31:16,296 --> 01:31:17,016
Oh, man.

1830
01:31:17,016 --> 01:31:19,016
Yeah, so that seems like,

1831
01:31:19,236 --> 01:31:21,176
we already know the distribution deals happen

1832
01:31:21,176 --> 01:31:22,816
and the profits are shared.

1833
01:31:23,356 --> 01:31:24,176
That's known.

1834
01:31:24,316 --> 01:31:26,536
And then these consumer incentives,

1835
01:31:26,936 --> 01:31:29,536
I think undoubtedly will occur as well.

1836
01:31:30,116 --> 01:31:31,756
And so I think the JP Morgans of the world

1837
01:31:31,756 --> 01:31:35,096
are not going to be the first to do those kinds of things.

1838
01:31:35,096 --> 01:31:36,596
I think the Robin Hoods of the world are.

1839
01:31:37,096 --> 01:31:40,616
And JP Morgan's in the world are also not just going to sit by idly, though, right?

1840
01:31:40,676 --> 01:31:42,836
So I think there's a couple of interesting points here.

1841
01:31:42,936 --> 01:31:44,796
One, yes, this is a great observation.

1842
01:31:45,056 --> 01:31:46,116
Tech companies can move fast.

1843
01:31:46,176 --> 01:31:46,896
They see an opportunity.

1844
01:31:47,016 --> 01:31:47,596
They're seizing it.

1845
01:31:47,636 --> 01:31:48,076
Good for them.

1846
01:31:48,296 --> 01:31:49,296
I get why Stripe's doing it.

1847
01:31:49,336 --> 01:31:50,276
I get why Google's doing it.

1848
01:31:50,516 --> 01:31:54,996
I think, again, it reflects on the sad, nihilistic state of the world, but it is what it is.

1849
01:31:56,596 --> 01:32:00,016
What I think is more interesting, though, is two things.

1850
01:32:00,016 --> 01:32:06,056
One, what this is effectively doing and what you're seeing playing out is, you know, we

1851
01:32:06,056 --> 01:32:09,336
thought the battles with the central bank no it's like to your points with the banks and so

1852
01:32:09,336 --> 01:32:16,156
in some ways this is actually a very good thing systemically because what is tether with you know

1853
01:32:16,156 --> 01:32:20,377
getting all this yield on their on their treasuries they're a fully reserved bank they're just a bank

1854
01:32:20,377 --> 01:32:23,676
but they're fully reserved in theory like i know you have to you can't prove it audit it whatever

1855
01:32:23,676 --> 01:32:29,016
but like what's interesting is like you're basically it doesn't claim that it's okay to not be

1856
01:32:29,016 --> 01:32:35,216
fully reserved true but exactly and what's interesting is our system claims it's fine

1857
01:32:35,216 --> 01:32:36,176
to not be fully reserved.

1858
01:32:36,656 --> 01:32:37,916
Tether claims we are fully reserved

1859
01:32:37,916 --> 01:32:39,036
and we don't know if that's true or not,

1860
01:32:39,116 --> 01:32:41,477
but at least it assumes

1861
01:32:41,477 --> 01:32:42,456
that it should not be.

1862
01:32:42,977 --> 01:32:43,256
Exactly.

1863
01:32:43,616 --> 01:32:44,916
And it shines this kind of

1864
01:32:44,916 --> 01:32:46,236
really fascinating light.

1865
01:32:46,416 --> 01:32:48,377
It's so funny to watch all this play out.

1866
01:32:48,496 --> 01:32:50,076
And again, read Alan's piece.

1867
01:32:50,216 --> 01:32:50,636
He's so right.

1868
01:32:50,676 --> 01:32:51,616
It's because you're watching

1869
01:32:51,616 --> 01:32:54,236
all of these old fintech players

1870
01:32:54,236 --> 01:32:55,336
who are very sophisticated

1871
01:32:55,336 --> 01:32:56,096
in the world of business,

1872
01:32:56,176 --> 01:32:57,656
but just do not understand money.

1873
01:32:57,756 --> 01:32:58,296
This is the problem

1874
01:32:58,296 --> 01:32:58,896
with Silicon Valley

1875
01:32:58,896 --> 01:33:00,536
with Bitcoin forever.

1876
01:33:00,877 --> 01:33:01,296
They get attacked.

1877
01:33:01,356 --> 01:33:02,416
They do not understand money.

1878
01:33:02,877 --> 01:33:03,696
They'll never understand

1879
01:33:03,696 --> 01:33:04,616
why gold is ever viable.

1880
01:33:04,616 --> 01:33:05,496
or eventually they will,

1881
01:33:05,556 --> 01:33:06,796
but it's going to take another decade to play out.

1882
01:33:06,836 --> 01:33:08,016
So you're kind of watching this like,

1883
01:33:08,356 --> 01:33:10,816
almost like slow motion car crash happening

1884
01:33:10,816 --> 01:33:11,836
where they're all like,

1885
01:33:11,896 --> 01:33:13,636
hey, like really interesting.

1886
01:33:13,756 --> 01:33:14,856
We can just hold the treasuries.

1887
01:33:14,856 --> 01:33:16,116
Like, why didn't everyone do that?

1888
01:33:16,156 --> 01:33:17,336
And that's like this bank.

1889
01:33:17,416 --> 01:33:18,016
And you kind of like,

1890
01:33:18,256 --> 01:33:19,056
you know that little meme

1891
01:33:19,056 --> 01:33:20,036
where it's like a little Muppet

1892
01:33:20,036 --> 01:33:20,896
and you're like,

1893
01:33:20,996 --> 01:33:23,136
hey, you weren't just making up money

1894
01:33:23,136 --> 01:33:24,316
out of thin air JPM, were you?

1895
01:33:24,396 --> 01:33:25,636
And then a little Muppet looks over

1896
01:33:25,636 --> 01:33:27,336
and you're like, what?

1897
01:33:27,516 --> 01:33:29,636
And you realize he only had 10% reserves

1898
01:33:29,636 --> 01:33:30,316
in the first place.

1899
01:33:30,396 --> 01:33:31,936
And that's like, literally that's what's playing out.

1900
01:33:32,156 --> 01:33:33,076
And then eventually they're going to go,

1901
01:33:33,276 --> 01:33:34,296
well, how does that work

1902
01:33:34,296 --> 01:33:36,276
that you can like only have 10% reserves.

1903
01:33:36,436 --> 01:33:37,316
What if you blow up?

1904
01:33:37,596 --> 01:33:38,656
And they're going to shine the light

1905
01:33:38,656 --> 01:33:39,377
on the Federal Reserve

1906
01:33:39,377 --> 01:33:40,056
and it's going to be like,

1907
01:33:40,396 --> 01:33:40,836
wait a minute,

1908
01:33:41,036 --> 01:33:42,156
you weren't just making stuff up

1909
01:33:42,156 --> 01:33:42,936
out there, were you?

1910
01:33:43,196 --> 01:33:44,377
And the Federal Reserve Muppet's

1911
01:33:44,377 --> 01:33:45,096
going to look over and be like,

1912
01:33:45,456 --> 01:33:47,076
and then everything comes back to Bitcoin.

1913
01:33:47,216 --> 01:33:49,656
So that's one way I see

1914
01:33:49,656 --> 01:33:50,656
the next 10 years playing out.

1915
01:33:51,816 --> 01:33:52,676
But the other thing I think

1916
01:33:52,676 --> 01:33:53,456
is really important is,

1917
01:33:53,556 --> 01:33:54,796
and this is what makes me very bullish

1918
01:33:54,796 --> 01:33:55,796
on Bitcoin and Lysian,

1919
01:33:56,216 --> 01:33:59,276
is the internet versus intranets.

1920
01:33:59,996 --> 01:34:01,296
The challenge these guys

1921
01:34:01,296 --> 01:34:02,736
are all going to run into in my mind.

1922
01:34:02,816 --> 01:34:04,256
And maybe Google is in a better spot

1923
01:34:04,256 --> 01:34:10,516
but I doubt it, is credibly neutral. If Stripe has a chain and JP Morgan's just not going to sit

1924
01:34:10,516 --> 01:34:14,356
there, like Jaime Diamond or Jamie Diamond's not going to sit there and just like take this,

1925
01:34:14,736 --> 01:34:18,776
like he's going to come out and launch his chain. And then Google's going to have theirs and Google's

1926
01:34:18,776 --> 01:34:22,776
credibly neutral. Like, well, who trusts Google? Is Facebook? Absolutely not. Like all of these

1927
01:34:22,776 --> 01:34:26,316
guys, what I think we're watching, it reminds me we're almost like watching blockchain, like

1928
01:34:26,316 --> 01:34:30,616
corporate blockchains play out in 2017. They're all going to launch chain and some of them are

1929
01:34:30,616 --> 01:34:34,416
and get along and like work together, but most of them are not. And there's going to be a lot of

1930
01:34:34,416 --> 01:34:38,216
people are like, dude, I'm never JP Morgan. I'm never hooking up with Wells Fargo. Facebook,

1931
01:34:38,336 --> 01:34:41,336
I'm never hooking up with Google. Apple is just going to be like, fuck all y'all. Like I'm sitting

1932
01:34:41,336 --> 01:34:45,016
this out. And so then you start asking the questions like, well, then who's going to market

1933
01:34:45,016 --> 01:34:49,477
make between them? And sure, bridge is today, but like that requires, again, these are all

1934
01:34:49,477 --> 01:34:54,616
permissioned. Every party has to agree to be a part of that. And if, when you have ones that don't,

1935
01:34:55,056 --> 01:34:57,776
then how are you going to get the exchange between those? It's going to be extremely

1936
01:34:57,776 --> 01:35:01,116
inefficient, extremely complicated, and this is going to take multiple years to play out before

1937
01:35:01,116 --> 01:35:04,896
everyone in Silicon Valley realizes, oh shit, the way to do this is to plug in the internet, Bitcoin

1938
01:35:04,896 --> 01:35:12,496
might. I agree with that, but I still, I don't think they're acting dumb. I think they can make

1939
01:35:12,496 --> 01:35:15,977
a ton of money over the next 10 years. The dollar is not going away in the next 10 years.

1940
01:35:17,196 --> 01:35:24,956
I think it's a good thing. Disrupting the banks across different angles is good. So getting tech

1941
01:35:24,956 --> 01:35:27,176
and fintech companies in here to potentially

1942
01:35:27,176 --> 01:35:31,076
disrupt banks, I think that's great.

1943
01:35:31,636 --> 01:35:35,576
I do agree, though, that it'll play out the way you're talking about.

1944
01:35:35,696 --> 01:35:38,616
Like, partnership here and here, and it's just a subset of...

1945
01:35:38,616 --> 01:35:42,016
It's still... It is basically recreating what already exists,

1946
01:35:42,316 --> 01:35:44,196
but they can make a lot of money.

1947
01:35:44,776 --> 01:35:48,296
And then, because it will just recreate what already exists,

1948
01:35:48,596 --> 01:35:50,256
Bitcoin will just keep strengthening,

1949
01:35:51,216 --> 01:35:54,336
keep being this global financial network

1950
01:35:54,336 --> 01:35:57,456
and the companies that are betting on it

1951
01:35:57,456 --> 01:36:00,316
like Lock and LightSpark

1952
01:36:00,316 --> 01:36:04,036
and Lightning Labs, Tether

1953
01:36:04,036 --> 01:36:07,296
they benefit

1954
01:36:07,296 --> 01:36:09,416
if it plays out like we just said

1955
01:36:09,416 --> 01:36:12,616
they benefit because over the next five years

1956
01:36:12,616 --> 01:36:15,616
all those companies will just further develop their assets

1957
01:36:15,616 --> 01:36:18,896
strengthen, build great products, great customer experiences

1958
01:36:18,896 --> 01:36:20,596
business models, expertise

1959
01:36:20,596 --> 01:36:23,656
and all the other tech companies will then realize

1960
01:36:23,656 --> 01:36:25,336
or they'll make a lot of money,

1961
01:36:25,556 --> 01:36:28,716
but then they'll still be starting at zero

1962
01:36:28,716 --> 01:36:31,036
in five years to get on Bitcoin.

1963
01:36:31,356 --> 01:36:32,796
I could not, you laid it out perfectly.

1964
01:36:33,056 --> 01:36:34,216
And so when you say it's smart,

1965
01:36:34,516 --> 01:36:35,296
I mean, it depends.

1966
01:36:35,396 --> 01:36:36,516
Depends what's your time horizon.

1967
01:36:36,636 --> 01:36:37,796
Like if you're just kind of coming in,

1968
01:36:37,836 --> 01:36:39,377
like how do we plunder as much as possible?

1969
01:36:39,636 --> 01:36:40,516
Like let's put it this way.

1970
01:36:41,796 --> 01:36:43,416
Silicon Valley is a very long stripe,

1971
01:36:43,816 --> 01:36:45,556
a much longer block in this world.

1972
01:36:46,336 --> 01:36:48,256
Sure, stripes are gonna make a hell of a lot of money

1973
01:36:48,256 --> 01:36:52,477
and just like kind of like raid the cupboard

1974
01:36:52,477 --> 01:36:53,216
while they can.

1975
01:36:53,656 --> 01:36:59,576
Meanwhile, Block is becoming a giant in the background that is betting on the future and will be around in 50 years.

1976
01:37:00,416 --> 01:37:04,196
Most disagreements in the world come down to a difference in timing.

1977
01:37:04,696 --> 01:37:04,816
Yeah.

1978
01:37:05,856 --> 01:37:08,477
Just like the crypto people and the crypto VCs.

1979
01:37:08,616 --> 01:37:11,956
Some of them made a killing in 2017 and 2018.

1980
01:37:13,496 --> 01:37:15,877
Multicoin, one of the best performing funds of all time.

1981
01:37:16,477 --> 01:37:19,377
How many of their tokens have any traction ever?

1982
01:37:19,516 --> 01:37:19,716
Zero.

1983
01:37:19,877 --> 01:37:22,977
And so it's just like, I mean, play fiat games, win fiat prizes.

1984
01:37:22,977 --> 01:37:24,996
if you're that money hungry, enjoy in the short run.

1985
01:37:25,136 --> 01:37:26,896
Meanwhile, companies that are building

1986
01:37:26,896 --> 01:37:29,036
for 2100, we'll see you there.

1987
01:37:29,796 --> 01:37:29,916
All right.

1988
01:37:30,536 --> 01:37:32,936
I think we need to wrap. That's a good way. It's a good place

1989
01:37:32,936 --> 01:37:34,936
to wrap it. Max and I are agreeing with each

1990
01:37:34,936 --> 01:37:36,916
other. We laid out a vision. There you go.

1991
01:37:37,536 --> 01:37:39,056
Love it. All right. Let's go.

1992
01:37:39,736 --> 01:37:41,036
Until next time. Good chat, guys.

1993
01:37:41,596 --> 01:37:42,736
All right.

1994
01:37:42,877 --> 01:37:43,196
Bye-bye.

1995
01:37:47,816 --> 01:37:48,176
Man.

1996
01:37:48,736 --> 01:37:50,156
Oh, still alive. Bye, guys.

1997
01:37:51,196 --> 01:37:51,456
Keep