1
00:00:00,000 --> 00:00:12,800
Trust is a function. It is more than just a feeling. In a world of infinite noise,

2
00:00:12,800 --> 00:00:20,440
the only signal that matters is the one you verify. We are here to map the graph,

3
00:00:21,180 --> 00:00:26,000
order the protocol, and build the immune system for the decentralized web.

4
00:00:26,000 --> 00:00:29,760
Welcome to the signal processing layer.

5
00:00:30,860 --> 00:00:33,720
This is Say What?

6
00:00:41,940 --> 00:00:46,640
Welcome, everyone, to the third episode of Say What?

7
00:00:47,300 --> 00:00:50,160
A special series here on Pleb Chain Radio,

8
00:00:50,160 --> 00:00:55,320
diving deep into one of the toughest and most important challenges in all of freedom technology,

9
00:00:56,000 --> 00:00:56,960
The web of trust.

10
00:00:57,840 --> 00:01:00,520
What exactly is web of trust?

11
00:01:00,780 --> 00:01:03,280
Why does it matter so much right now?

12
00:01:03,440 --> 00:01:11,400
And how do we actually build it in a world where technology, especially AI, is accelerating faster than ever?

13
00:01:11,960 --> 00:01:14,800
These are the questions we are here to unpack.

14
00:01:15,620 --> 00:01:22,860
Today, we're exploring the fascinating intersection of artificial intelligence, web of trust, and everything in between.

15
00:01:22,860 --> 00:01:29,660
I am your guest host for this episode, David Strayhorn, stepping in for our usual host, Avi Burra.

16
00:01:30,380 --> 00:01:35,280
He's one of my co-founders at Nosfabrica alongside John Gordon and Vitor Pamplona.

17
00:01:35,960 --> 00:01:41,220
Avi's done an incredible job kicking this series off and I'm excited to keep the conversation rolling.

18
00:01:42,400 --> 00:01:49,040
Joining me today is Matthias de Bernardini, brilliant software developer and agentic engineering expert

19
00:01:49,040 --> 00:01:54,280
who's already pushing the boundaries of what's possible at the AI Trust Frontier.

20
00:01:55,140 --> 00:01:58,580
And as of this week, Matthias is officially joining us

21
00:01:58,580 --> 00:02:01,700
to collaborate with us at NOS Fabrica

22
00:02:01,700 --> 00:02:05,080
to help build open-source web of trust tools on NOSTER

23
00:02:05,080 --> 00:02:08,500
and to bring them, hopefully, to billions of people.

24
00:02:08,920 --> 00:02:11,620
So, Matthias, it's great to have you here. Let's dive in.

25
00:02:12,060 --> 00:02:13,580
Thank you for having me. It's a pleasure.

26
00:02:14,960 --> 00:02:18,300
Well, why don't you tell us about yourself?

27
00:02:18,300 --> 00:02:20,360
Tell us about your journey to where you are now.

28
00:02:20,840 --> 00:02:21,040
Yeah.

29
00:02:21,260 --> 00:02:25,400
So I originally, I was really interested in engineering.

30
00:02:25,800 --> 00:02:27,980
I loved chemistry when I was in high school.

31
00:02:28,120 --> 00:02:35,120
So I went to school, university in Milwaukee, Wisconsin for material science and engineering.

32
00:02:35,300 --> 00:02:38,880
And I finished that and had a great experience.

33
00:02:40,580 --> 00:02:44,060
And eventually I learned about Bitcoin.

34
00:02:44,060 --> 00:02:47,260
I think the first time I learned about Bitcoin was in 2014.

35
00:02:47,260 --> 00:02:52,580
and I have this really funny story where I, you know, I was telling my dad, I was pitching him

36
00:02:52,580 --> 00:02:56,200
Bitcoin. It's like, you know, I was, I was making a little bit of money at the time working as a lab

37
00:02:56,200 --> 00:03:02,920
technician and I wanted to buy this. I didn't know what I was going to do with it, but I just wanted

38
00:03:02,920 --> 00:03:11,920
to have some because I had seen Roger Ver's commercial and I had read about it on Slashdot

39
00:03:11,920 --> 00:03:18,980
as well. And it like really, really piqued my curiosity. And but I just wasn't like an investor,

40
00:03:19,180 --> 00:03:23,500
like I didn't really see myself as that. You know, so like the whole mindset of like,

41
00:03:23,560 --> 00:03:27,280
oh, buy it and I can sell it later. I just wanted to like experiment with it.

42
00:03:28,360 --> 00:03:35,080
I told my dad about it, you know, tried to get his opinion. And, you know, as I was explaining

43
00:03:35,080 --> 00:03:40,840
to him, it's like, oh, yeah, and it's like $90 right now. And my dad's just like $90 for one

44
00:03:40,840 --> 00:03:42,960
Bitcoin? No, no, no, no, no.

45
00:03:43,420 --> 00:03:44,880
That's too much. That's too much.

46
00:03:45,020 --> 00:03:45,680
It's too expensive.

47
00:03:46,320 --> 00:03:47,500
Like, okay, dad.

48
00:03:48,060 --> 00:03:50,480
I don't know if we were calling them sats back then.

49
00:03:50,860 --> 00:03:53,300
So you're keeping the pull-em-out cheat if that was.

50
00:03:53,380 --> 00:03:54,760
I know, I know, right?

51
00:03:54,860 --> 00:03:55,560
I should have.

52
00:03:55,560 --> 00:03:59,540
I mean, the sats, they always, they were in the code base,

53
00:03:59,640 --> 00:04:01,780
but I don't think back then it was like a thing.

54
00:04:02,320 --> 00:04:02,960
I don't think so.

55
00:04:03,420 --> 00:04:03,720
Yeah.

56
00:04:04,280 --> 00:04:06,300
That would have been the smart thing to, you know,

57
00:04:06,340 --> 00:04:08,760
hit him with the unit bias and then he would have been like,

58
00:04:08,820 --> 00:04:10,020
oh yeah, that's a great deal.

59
00:04:10,020 --> 00:04:12,420
Why not? Just get a couple.

60
00:04:14,620 --> 00:04:19,900
But it sounds like you weren't coming at it from an ideological Austrian economics.

61
00:04:22,220 --> 00:04:24,720
That's not kind of what brought you into it?

62
00:04:24,900 --> 00:04:25,840
It was the technology?

63
00:04:26,580 --> 00:04:31,480
I just wanted to play with it and just see what it was like.

64
00:04:31,980 --> 00:04:33,860
Why are all these people talking about it?

65
00:04:34,920 --> 00:04:39,580
It's this totally made-up thing, and yet it's grabbing all these people's attention.

66
00:04:39,580 --> 00:04:44,720
That's interesting. I should I should get some and look into it. Why not?

67
00:04:45,080 --> 00:04:50,440
And at the time, I you're right. Like I didn't really have I didn't really understand investing.

68
00:04:50,560 --> 00:04:55,280
I didn't really understand money, how the economy works and all of those things.

69
00:04:55,280 --> 00:05:03,420
It was just so it was just out of mind, you know, really just focusing on living my life, going to school, trying to get a job.

70
00:05:03,420 --> 00:05:12,120
um and you know once i i got to that you know and started making my own money uh then i was a little

71
00:05:12,120 --> 00:05:18,320
bit more interested in all of these things and i'd come across jim rickard's book uh the road

72
00:05:18,320 --> 00:05:25,940
to ruin and currency wars and i i read through both of them and you know i think in like the

73
00:05:25,940 --> 00:05:32,260
matter of a couple weeks and i just you know very clear to me after reading this like okay

74
00:05:32,260 --> 00:05:38,640
wait a minute he's just explained everything to me about how the economy works and like this giant

75
00:05:38,640 --> 00:05:45,340
yeah you know this giant mess we've made and in the book he was advocating for gold

76
00:05:45,340 --> 00:05:53,400
uh because Jim Rickards is he's like a famous gold bug but I'd also known about Bitcoin you know from

77
00:05:53,400 --> 00:05:58,460
like reading that article didn't really pay too much attention to it but then a friend from school

78
00:05:58,460 --> 00:06:05,080
was trading it and he was telling me about it more and more and he was making really good money with

79
00:06:05,080 --> 00:06:15,280
it uh trading uh and at that point like I had money and I could buy some so I just did you know

80
00:06:15,280 --> 00:06:22,520
I didn't really have to get anyone's second opinion and it really just interested me just

81
00:06:22,520 --> 00:06:29,720
so much just the idea that you have this proof of work and that is what underpins the money and

82
00:06:29,720 --> 00:06:36,280
there's this decentralized network and we're all double checking each other's work and then writing

83
00:06:36,280 --> 00:06:42,580
it to this thing called the blockchain and everybody shares it and if anybody makes a

84
00:06:42,580 --> 00:06:48,740
mistake everybody else rejects that mistake and rechecks their own work and you know writes it down

85
00:06:48,740 --> 00:06:50,860
to their personal storage.

86
00:06:51,900 --> 00:06:54,080
I thought that was so interesting.

87
00:06:54,400 --> 00:06:55,780
Just being a part of that,

88
00:06:56,020 --> 00:06:56,780
that's just what I want.

89
00:06:56,920 --> 00:06:58,160
I want to run a node.

90
00:06:58,420 --> 00:07:00,180
I want to be a part of all this.

91
00:07:00,180 --> 00:07:03,440
And I remember I had the,

92
00:07:03,480 --> 00:07:06,680
I think it was the Air Wallet, Air Blitz.

93
00:07:07,180 --> 00:07:10,000
It was like an old iOS wallet that had Bitcoin.

94
00:07:10,840 --> 00:07:13,500
And I bought a little bit off of this exchange.

95
00:07:13,660 --> 00:07:14,780
It was a really cool exchange

96
00:07:14,780 --> 00:07:17,320
because they didn't have a custodial wallet.

97
00:07:17,320 --> 00:07:21,600
you would have to put in an address for them to send it to you.

98
00:07:22,140 --> 00:07:24,840
And so you would give them fiat and then they just sent you Bitcoin.

99
00:07:25,020 --> 00:07:28,760
So there's no like there's no custodial wallet like a like a Coinbase.

100
00:07:29,440 --> 00:07:37,240
And I remember I was just I had like like 50 bucks or something like that.

101
00:07:38,140 --> 00:07:41,000
And it just showed up.

102
00:07:41,000 --> 00:07:47,260
And then I realized I'd also bought a Ledger wallet because I understood like it was kind of intuitive.

103
00:07:47,320 --> 00:07:50,480
to me, it's like, I shouldn't keep my password on the device.

104
00:07:50,680 --> 00:07:53,060
If I get hacked, they can just take the Bitcoin.

105
00:07:53,260 --> 00:07:54,820
That to me just seems so obvious.

106
00:07:55,320 --> 00:08:01,820
I should probably get something, what is the easiest way to solve this issue for me?

107
00:08:02,340 --> 00:08:05,620
And I got the original Ledger wallet that didn't even have a screen.

108
00:08:05,620 --> 00:08:13,360
And I remember I set it up, made an address, and then scanned the QR code with the iOS wallet,

109
00:08:13,540 --> 00:08:17,260
and then just saw the balance change on my laptop.

110
00:08:17,320 --> 00:08:23,460
And like that moment, I just got goosebumps, you know, that I could just like move money around like that just so seamlessly.

111
00:08:24,020 --> 00:08:25,760
And I didn't have to call a bank.

112
00:08:25,840 --> 00:08:27,480
I didn't have to do anything.

113
00:08:27,600 --> 00:08:28,940
It was just software running.

114
00:08:29,780 --> 00:08:35,880
And that was like a big aha moment for me, just being able to, you know, just seamlessly move money around.

115
00:08:36,940 --> 00:08:45,280
But then I quickly realized, like, wait a minute, this Ledger device, it was like a really small USB device at the time.

116
00:08:45,280 --> 00:08:46,880
It doesn't have a screen.

117
00:08:47,320 --> 00:08:55,520
So if I want to see the password, it has to be on my laptop because that's where it shows you the key or the backup for the mnemonic phrase.

118
00:08:56,160 --> 00:09:03,080
And I was like, if I have a screen logger on my laptop because I downloaded it by accident, they're going to be also going to be able to do my Bitcoin.

119
00:09:03,180 --> 00:09:04,880
So like this thing doesn't really help that much.

120
00:09:05,200 --> 00:09:07,540
Obviously better than just storing it locally.

121
00:09:07,540 --> 00:09:10,540
But when you initialize it, it shows you the password.

122
00:09:11,440 --> 00:09:16,620
And then at that point, like I got a Trezor and that to me was like awesome.

123
00:09:16,620 --> 00:09:21,620
you know it's like okay finally like the super minimal device I can see the all the words there

124
00:09:21,620 --> 00:09:27,380
and at the time I was just doing engineering work but I just wouldn't shut up at work about

125
00:09:27,380 --> 00:09:32,840
you guys should get some Bitcoin like this thing is really cool you know and like I convinced a few

126
00:09:32,840 --> 00:09:40,380
people to get it but they all like quickly sold it basically like everyone and then there was this

127
00:09:40,380 --> 00:09:46,760
one guy who's a bit older he was like a huge silver bug you know and so like he's probably

128
00:09:46,760 --> 00:09:51,980
laughing his ass off right now you know with how well silver is doing but at the time like silver

129
00:09:51,980 --> 00:09:57,900
was totally dead in the water and bitcoin was going like crazy so uh that was really fun and

130
00:09:57,900 --> 00:10:04,460
eventually you know i just like couldn't keep shutting up about it and i realized i'm not doing

131
00:10:04,460 --> 00:10:12,440
the right thing here. My focus is not on this job. It's on Bitcoin. How do I get involved in this?

132
00:10:13,200 --> 00:10:14,960
What were you doing at the time?

133
00:10:15,620 --> 00:10:22,980
I was doing packaging engineering. I got out of my degree and I found this company. It was a

134
00:10:22,980 --> 00:10:29,500
chemicals company. And I basically would do all the product testing and validation. It was pretty

135
00:10:29,500 --> 00:10:34,440
interesting. It was a good company. Honestly, it was a great business, well-established.

136
00:10:34,460 --> 00:10:41,180
They had products all over the world, and people bought the products and loved them.

137
00:10:41,380 --> 00:10:50,600
And they were a curious case of a chemical company where they actually cared about not dumping chemicals into the environment.

138
00:10:50,900 --> 00:10:58,920
They went out of their way to try to do it as environmentally friendly as possible, which is pretty rare.

139
00:10:58,920 --> 00:11:04,680
usually those companies, you have to get the government to tell them what to do.

140
00:11:05,060 --> 00:11:10,920
But they were pretty adamant about, even if it cuts into your margins,

141
00:11:11,280 --> 00:11:14,880
we're still going to make sure all the waste is getting disposed of properly.

142
00:11:17,000 --> 00:11:21,880
And there are just a lot of things that they were doing that they didn't have to do

143
00:11:21,880 --> 00:11:25,240
because the law didn't require them to do it.

144
00:11:25,240 --> 00:11:29,660
But they understood that, you know, we should do it our way.

145
00:11:30,600 --> 00:11:36,440
And as a result, the products end up being more expensive, but they're just a higher quality.

146
00:11:36,560 --> 00:11:42,740
It was like cleaning products and, you know, that sort of like household chemicals, that sort of thing.

147
00:11:43,100 --> 00:11:44,420
So great company.

148
00:11:44,680 --> 00:11:47,300
You liked working with them, but didn't have the excitement of.

149
00:11:49,360 --> 00:11:49,880
Yeah.

150
00:11:50,140 --> 00:11:54,640
You know, it's just like it just wasn't anything to do with Bitcoin.

151
00:11:54,640 --> 00:11:59,920
And also I just like wasn't able to do anything entrepreneurial in that space.

152
00:12:00,100 --> 00:12:05,700
Like I, for me, the idea of like making my own chemicals company would just seem like

153
00:12:05,700 --> 00:12:08,480
just so, just so unrealistic.

154
00:12:08,880 --> 00:12:11,600
And like I, it would just like never, ever happen.

155
00:12:12,180 --> 00:12:18,200
And at the time, you know, I just saw everything happening with machine learning and AI.

156
00:12:18,400 --> 00:12:20,020
This was, you know, 2017.

157
00:12:20,020 --> 00:12:27,260
um back then like you know it was much more academic and there were just very few companies

158
00:12:27,260 --> 00:12:34,320
working with ai and it was mostly like recommendation engines um but you know it was just so

159
00:12:34,320 --> 00:12:42,580
obvious to me that it was just early and you just had to wait for the technology and someone will

160
00:12:42,580 --> 00:12:49,160
come up with something better and you know it's better to just start to get involved rather than

161
00:12:49,160 --> 00:12:53,600
just sit and wait for that to happen.

162
00:12:53,600 --> 00:13:02,800
So I found a really good program for computer science and this was happening to me in the

163
00:13:02,800 --> 00:13:03,220
Netherlands.

164
00:13:03,580 --> 00:13:08,800
So I was able to, you know, I have Italian citizenship at the moment.

165
00:13:09,140 --> 00:13:15,320
So I was able to study in Europe basically for free, even though I had already like,

166
00:13:15,320 --> 00:13:20,180
you know, gotten an education, um, expensive one here in the United States.

167
00:13:20,720 --> 00:13:26,020
But it just, to me seemed like, yeah, the fastest and most efficient way to get up to

168
00:13:26,020 --> 00:13:26,880
speed with everything.

169
00:13:27,600 --> 00:13:31,800
And I knew that like, if I wanted to work in Bitcoin, I really had to know my stuff,

170
00:13:32,360 --> 00:13:38,240
you know, like how, how am I going to make a wallet without actually knowing about the

171
00:13:38,240 --> 00:13:38,740
fundamentals?

172
00:13:39,340 --> 00:13:43,780
Um, at least that's how I think I, that's how I thought at the time, you know, that

173
00:13:43,780 --> 00:14:04,630
was my my sort of how my brain worked And now I know that like that not really true You know you can just jump in whenever you want And as long as you just putting in the effort and you running code and you trying stuff out and you thinking critically you know you can you can just learn on your own

174
00:14:04,630 --> 00:14:12,910
And now with with, you know, now with ChatGPT and Claude, like that's even more true for anything.

175
00:14:12,910 --> 00:14:34,730
So I basically, I was just doing school. I was doing honors research and distributed systems. And then COVID hit. And I realized, like, why am I in school? Like, I can already make websites, I can already make apps.

176
00:14:34,730 --> 00:14:42,090
You know, I already knew about open source and GitHub and Git and, you know, just working with other people.

177
00:14:43,290 --> 00:14:49,350
And I was like, why am I working on this school project?

178
00:14:49,730 --> 00:14:53,250
With all the COVID stuff happening, you know, you're just sitting at home bored.

179
00:14:53,250 --> 00:15:03,230
um so i just dropped out on my last year and just tried to desperately try to use bitcoin

180
00:15:03,230 --> 00:15:10,510
lightning to like start a business i was um writing a bunch of python scripts to like analyze

181
00:15:10,510 --> 00:15:17,770
the lightning network and my in my head my business plan like i can like parse the lightning network

182
00:15:17,770 --> 00:15:22,190
and then offer like routing hints or lightning nodes at the time.

183
00:15:22,290 --> 00:15:24,450
Like that seemed reasonable.

184
00:15:25,450 --> 00:15:30,410
So, you know, just using LND, I would just pull down the graph

185
00:15:30,410 --> 00:15:34,030
and then just like, you know, calculate eccentricity and the diameter,

186
00:15:35,450 --> 00:15:40,090
you know, shortest, longest path from any two peers.

187
00:15:41,050 --> 00:15:42,690
Just all these like different metrics.

188
00:15:42,690 --> 00:15:48,010
And my thought is like, okay, I can make an API and try to sell it.

189
00:15:48,790 --> 00:15:50,430
And I was working on that.

190
00:15:50,510 --> 00:15:51,570
I was making some progress.

191
00:15:51,790 --> 00:15:56,530
But then I was like, damn, this Python is so slow.

192
00:15:57,610 --> 00:16:00,170
I'm never going to make any progress.

193
00:16:00,910 --> 00:16:04,750
Even though I was using NumPy under the hood,

194
00:16:05,030 --> 00:16:10,990
which is as fast as you can get with Python without writing it in assembly yourself.

195
00:16:10,990 --> 00:16:16,150
and I think it was like the NetworkX Python library

196
00:16:16,150 --> 00:16:18,770
which at the time was really good

197
00:16:18,770 --> 00:16:23,530
and it was like

198
00:16:23,530 --> 00:16:26,130
I should just write this in Rust

199
00:16:26,130 --> 00:16:28,410
and see if I can make it better

200
00:16:28,410 --> 00:16:31,730
and I started doing that

201
00:16:31,730 --> 00:16:36,050
but then a friend of mine from a Discord group

202
00:16:36,050 --> 00:16:40,290
got me this job at a company

203
00:16:40,290 --> 00:16:50,030
that was like, it was this guy that made a ton of money in crypto and wanted to start a crypto

204
00:16:50,030 --> 00:16:56,270
exchange. And so they got me to like run it for them. And even though I was just like fresh out

205
00:16:56,270 --> 00:17:00,850
of school, you know, like I knew a little bit about programming and, you know, Bitcoin and I

206
00:17:00,850 --> 00:17:05,530
have been doing it. But the reason why he recommended me is because I was helping all

207
00:17:05,530 --> 00:17:11,110
these people like with their Bitcoin wallets on this discord channel for, for traders that my,

208
00:17:11,110 --> 00:17:16,270
my other friend, you know, got me into. And, uh, I remember somebody got their like

209
00:17:16,270 --> 00:17:23,430
funds lost in a samurai wallet. And, you know, I knew how to like get it out and like, you know,

210
00:17:23,430 --> 00:17:30,470
do all the derivations and, um, you know, I, I helped them, you know, rescue some coins. Uh,

211
00:17:31,090 --> 00:17:35,250
and eventually, you know, it just goes to show if you just like put yourself out there,

212
00:17:35,250 --> 00:17:38,870
you go and you know you help people like eventually someone's just gonna like you know

213
00:17:38,870 --> 00:17:46,530
prompt you to help them and maybe for a business and uh that then i moved to the united states

214
00:17:46,530 --> 00:17:55,030
back to the uh united states um and was just working for this company you know doing basically

215
00:17:55,030 --> 00:18:01,770
just everything like it was a small uh shit coin exchange that used binance as their back end so

216
00:18:01,770 --> 00:18:06,470
So they had like all this liquidity from day one and access to all these shit coins.

217
00:18:07,390 --> 00:18:10,670
And man, what a piece of crap that that company was like.

218
00:18:10,950 --> 00:18:17,250
I was like, I cannot believe how unprofessional this is.

219
00:18:17,290 --> 00:18:19,610
And we're handling like, you know, all this money.

220
00:18:21,190 --> 00:18:29,550
I like I tried to do my best, but because it was like a front end for Binance, we didn't have any access to any of the wallets.

221
00:18:29,550 --> 00:18:35,570
You know, we didn't have it was just like a dashboard that honestly was like looking back at it.

222
00:18:35,610 --> 00:18:40,210
It's like, did they have did like Binance have ChatGPT back in like 2021?

223
00:18:40,570 --> 00:18:43,090
Because this back end was so buggy.

224
00:18:43,750 --> 00:18:47,290
And like, did they code this?

225
00:18:47,350 --> 00:18:48,790
Like, how how is this so buggy?

226
00:18:49,030 --> 00:18:50,830
Like all these labels like were messed up.

227
00:18:50,990 --> 00:18:53,510
And it was just funny.

228
00:18:54,850 --> 00:18:58,210
But, you know, I just remember just trying to help people with their wallets.

229
00:18:58,210 --> 00:19:02,870
put these shit coins it was just so hard to work with like because they were just so buggy

230
00:19:02,870 --> 00:19:09,430
the dot all the docs didn't make any sense and i ultimately didn't have any access to any wallet

231
00:19:09,430 --> 00:19:15,190
so all i could do was just ping binance that hey this user they're having issues like this is what

232
00:19:15,190 --> 00:19:21,830
they're telling me can you help them out and the back and forth would always take forever um and

233
00:19:21,830 --> 00:19:25,870
And eventually the company just went out of business.

234
00:19:26,730 --> 00:19:33,350
And at that point, I was like, all right, well, I'll just keep back to doing what I know, which is just open source and Bitcoin.

235
00:19:33,750 --> 00:19:40,770
I would just look for projects, you know, that that sounded interesting in the in the space at the time.

236
00:19:40,770 --> 00:19:42,710
Like I thought Lightning was doing like really well.

237
00:19:42,710 --> 00:19:50,030
Like it seemed like they had enough people like working on it that it didn't really need any more people to like.

238
00:19:50,030 --> 00:19:54,070
because like all the ideas were there you know they had all the bolts it was just a matter of

239
00:19:54,070 --> 00:19:59,470
like implementing them and i saw that they had like great programmers working on it so it's like

240
00:19:59,470 --> 00:20:04,650
all right well i'll go do something else uh and so i found the fetty project and just started

241
00:20:04,650 --> 00:20:13,670
contributing to that um you know i was part of this hackathon uh in austin and then that was sort

242
00:20:13,670 --> 00:20:19,890
of what like really prompted me to get really involved with Fediment. And then I participated

243
00:20:19,890 --> 00:20:29,210
in this hackathon. It was like a month long hackathon. And I like just went all in on it.

244
00:20:29,210 --> 00:20:36,310
And I had a partner at the beginning, but then they dropped out. So it was just me.

245
00:20:36,310 --> 00:20:48,330
And I was able to make this distributed storage system that used the federation to make multiple copies of your data.

246
00:20:48,330 --> 00:21:06,150
And it was basically a key value storage that would get stored by all the members of the federation and use the underlying fediment consensus algorithm to make sure that the data is actually consistent.

247
00:21:06,310 --> 00:21:07,510
across all the members.

248
00:21:08,070 --> 00:21:09,410
And that was a lot of fun.

249
00:21:09,490 --> 00:21:11,290
I got second place in the hackathon.

250
00:21:11,870 --> 00:21:14,150
And at the time, I was helping this other guy

251
00:21:14,150 --> 00:21:16,070
who was also in the hackathon

252
00:21:16,070 --> 00:21:18,590
because he had some questions about Rust.

253
00:21:19,310 --> 00:21:24,810
And I was just helping him about how to get access to it

254
00:21:24,810 --> 00:21:29,890
and just do several things with the Fetty Mint dev environment.

255
00:21:29,890 --> 00:21:40,550
and um it turns out that he was working with this guy called rob hamilton and then that's how y'all

256
00:21:40,550 --> 00:21:48,990
met that's how i met rob from anchor watch yeah and like he um you know it's like hey i need this

257
00:21:48,990 --> 00:21:55,670
back end looks like you know rust and bitcoin do you want to help me and at the time uh you know

258
00:21:55,670 --> 00:22:00,590
Rob had created this, like he, it already was very like cool.

259
00:22:00,710 --> 00:22:05,630
What he had for as a backend, like it all worked the whole anchor watch,

260
00:22:05,730 --> 00:22:09,470
you know, mini scripts vault system. Um,

261
00:22:09,810 --> 00:22:13,350
and what do you call it? But it was, it was just,

262
00:22:13,450 --> 00:22:19,090
it was using Python and, uh, calling BDK CLI using a sub process,

263
00:22:19,090 --> 00:22:24,130
which is, you know, inherently not great. Um, so yeah,

264
00:22:24,130 --> 00:22:27,010
He just got me to just rewrite all of it in Rust.

265
00:22:27,430 --> 00:22:32,990
And I just worked for him for, yeah, I think it was for like six months.

266
00:22:32,990 --> 00:22:34,530
And then he offered me full time.

267
00:22:36,130 --> 00:22:37,930
Is that how you ended up in Nashville?

268
00:22:38,530 --> 00:22:39,430
Yeah, yeah.

269
00:22:39,510 --> 00:22:45,530
He said, hey, if you want full time, you got to come down to Nashville and work out of Bitcoin Park with us.

270
00:22:45,530 --> 00:22:59,630
Yeah. And I think that was, you know, I'm really glad he did that because, you know, it's just so much better to just focus with your team when you're all you just know you're all in.

271
00:22:59,830 --> 00:23:12,390
Everyone is in for the project. And like it just like it's just so much more motivating rather than, you know, knowing that everybody like uprooted their lives to move to Nashville because it was it was three of us working on the code base.

272
00:23:12,390 --> 00:23:22,090
and it just motivates you so much more to like yeah just like work and like get it done you know

273
00:23:22,090 --> 00:23:28,630
even if it's kind of complicated um but it was just yeah i had the time of my life honestly just

274
00:23:28,630 --> 00:23:35,450
like shipping that to production and just dealing with all the bugs and at the time it was crazy

275
00:23:35,450 --> 00:23:40,990
At the time, yeah, ChatGPT was coming out and Claude and all this stuff.

276
00:23:41,150 --> 00:23:48,290
And so I just kept experimenting with these agents and learning more about AI as much as I could.

277
00:23:49,430 --> 00:23:54,790
And you had already been thinking about machine learning, you said, before COVID even.

278
00:23:54,950 --> 00:23:55,270
Yeah.

279
00:23:55,410 --> 00:23:56,450
And distributed systems.

280
00:23:57,250 --> 00:24:02,370
And it's interesting, you said, recommendation that people were using machine learning for recommendations.

281
00:24:02,370 --> 00:24:08,130
but not in a purely decentralized way.

282
00:24:08,650 --> 00:24:09,450
No, definitely.

283
00:24:09,830 --> 00:24:10,950
It's all, I mean, at the time,

284
00:24:11,030 --> 00:24:12,470
like when I was learning about it,

285
00:24:12,670 --> 00:24:16,670
you know, the biggest users of machine learning

286
00:24:16,670 --> 00:24:17,470
were like Netflix

287
00:24:17,470 --> 00:24:21,270
and they would use machine learning algorithms

288
00:24:21,270 --> 00:24:23,370
to see what movies you had watched

289
00:24:23,370 --> 00:24:26,550
and then they would run that

290
00:24:26,550 --> 00:24:29,330
through their own internal recommendation engine

291
00:24:29,330 --> 00:24:30,970
that was all machine learning based.

292
00:24:30,970 --> 00:24:35,910
with neural networks and all of that good stuff.

293
00:24:36,730 --> 00:24:40,650
And that's what will populate that list of like,

294
00:24:40,710 --> 00:24:42,650
hey, you might be interested in watching these movies.

295
00:24:44,790 --> 00:24:48,690
So Facebook is also another big,

296
00:24:48,690 --> 00:24:52,390
a huge consumer of AI workloads.

297
00:24:53,470 --> 00:24:54,590
And many others,

298
00:24:54,830 --> 00:24:59,850
more and more social-based companies,

299
00:24:59,850 --> 00:25:01,390
like Tinder, for example,

300
00:25:01,650 --> 00:25:05,130
or pretty much anything where you have some kind of graph

301
00:25:05,130 --> 00:25:06,430
that you want to iterate over

302
00:25:06,430 --> 00:25:10,190
and then suggest different members of that graph

303
00:25:10,190 --> 00:25:14,890
to the user as potential things you might be interested in.

304
00:25:16,910 --> 00:25:20,730
So this is where the knowledge graph you learned about?

305
00:25:21,530 --> 00:25:29,390
That was a networks and graph theory course I took in school.

306
00:25:29,390 --> 00:25:37,630
And I just thought it was super interesting because I immediately was just applying it to the Lightning Network and to Bitcoin nodes.

307
00:25:37,890 --> 00:25:46,530
That's basically how I got through all the school is every course I took, I found a way to apply it to Bitcoin and Bitcoin Core.

308
00:25:47,110 --> 00:25:51,810
And then that immediately made it so much easier to understand all the concepts.

309
00:25:51,810 --> 00:25:57,810
And then also just like have enough like dopamine running through me while I learn about it.

310
00:25:57,830 --> 00:26:00,190
Because like I'm interested in Bitcoin.

311
00:26:00,430 --> 00:26:01,850
This is how Bitcoin works, you know.

312
00:26:01,930 --> 00:26:11,870
So like my computer, my, you know, systems architecture class, you know, it's like, okay, this is the computer system that runs Bitcoin Core.

313
00:26:12,070 --> 00:26:14,790
Like how, you know, how does that work?

314
00:26:16,650 --> 00:26:21,390
The algorithms course I took is like, okay, the blockchain is just a giant, you know, data structure.

315
00:26:21,810 --> 00:26:27,790
in the algorithms, data structures, you know, programming, obviously, computer networks,

316
00:26:27,790 --> 00:26:33,270
same thing. Everything I was learning, I was applying it through the lens of like,

317
00:26:33,350 --> 00:26:41,630
how would Bitcoin Core use this? And that just made it so much easier to just keep thinking about

318
00:26:41,630 --> 00:26:46,950
what it is I'm trying to learn about and just like do more cycles and then just have it ingrained in

319
00:26:46,950 --> 00:26:53,870
my intuitive understanding. And that honestly, I think that's like probably the best cheat,

320
00:26:54,390 --> 00:26:59,250
quote unquote cheat code that you can apply when you're trying to learn something new

321
00:26:59,250 --> 00:27:06,370
is how can you apply this knowledge that seems kind of random? Like what, you know, and it kind

322
00:27:06,370 --> 00:27:11,190
of is because the teachers have to decide like which one, which topics to like discuss in the,

323
00:27:11,190 --> 00:27:15,970
in the course. And that's up to their discretion to like what they think is most useful.

324
00:27:15,970 --> 00:27:24,610
and like you know they're not into bitcoin um in fact the big reason why i just you know didn't keep

325
00:27:24,610 --> 00:27:31,430
uh why i wanted to drop out is um i had like my senior thesis and i wanted to do it on bitcoin

326
00:27:31,430 --> 00:27:35,070
but i couldn't find any teachers that cared about bitcoin or were willing to like sponsor me

327
00:27:35,070 --> 00:27:52,980
to like you know do bitcoin research um and i guess i could have looked harder at like other schools But I was just like whatever I may as well just you know do just do it on my own Like I you know if I have a good idea I can put it out there and people will like

328
00:27:52,980 --> 00:27:58,100
it, you know, like, and if not, then whatever, I don't need to go through the school process.

329
00:27:58,760 --> 00:28:03,760
And I'm honestly really glad I didn't like stick with the academic system. Cause it's, it's just

330
00:28:03,760 --> 00:28:08,440
not great. Like it has a lot of issues. Obviously I think school is very important.

331
00:28:08,440 --> 00:28:12,060
And I don't think AI is going to replace school.

332
00:28:12,460 --> 00:28:20,520
I think it's going to make it much more valuable, actually, because already the whole point, like a big point of going to school is just meeting people.

333
00:28:21,800 --> 00:28:28,200
And networking, just like making relations, like hanging out, you know, making friends, like getting to know people.

334
00:28:28,200 --> 00:28:29,280
What are they working on?

335
00:28:30,940 --> 00:28:33,020
That is never going to go away.

336
00:28:33,020 --> 00:28:38,880
and, you know, be using AI to learn with your friends and teachers.

337
00:28:39,160 --> 00:28:42,340
And like, it's just going to make it more valuable, not less, in my opinion.

338
00:28:42,880 --> 00:28:46,260
Definitely, there will be schools that like, if you don't,

339
00:28:46,380 --> 00:28:48,140
if there's schools out there that you don't learn,

340
00:28:48,980 --> 00:28:51,160
then yeah, they're in trouble, you know, like,

341
00:28:51,280 --> 00:28:55,820
but actual institutions that like, yeah, want to expand human knowledge

342
00:28:55,820 --> 00:29:00,960
and give young people a good, you know, environment

343
00:29:00,960 --> 00:29:03,140
to actually improve themselves.

344
00:29:04,100 --> 00:29:06,700
I think it's going to be even more important.

345
00:29:06,780 --> 00:29:08,680
We're just going to get more schools, I'm thinking,

346
00:29:08,960 --> 00:29:12,500
and it'll just be easier to administer classes

347
00:29:12,500 --> 00:29:15,460
and administer the entire educational system.

348
00:29:15,860 --> 00:29:19,360
A lot of learning we can do on our own.

349
00:29:19,480 --> 00:29:21,840
And like you were saying, you would learn something

350
00:29:21,840 --> 00:29:25,560
and you'd kind of connect it to something else

351
00:29:25,560 --> 00:29:27,400
of great interest, like Bitcoin, right?

352
00:29:27,400 --> 00:29:33,000
kind of integrate that, all the different things you learn into that.

353
00:29:34,120 --> 00:29:40,460
Going back to the recommendation engines that you were talking about,

354
00:29:40,560 --> 00:29:45,720
that Netflix is doing, did you ever think about,

355
00:29:46,040 --> 00:29:50,100
okay, that's a lot of what we want to do with Web of Trust

356
00:29:50,100 --> 00:29:53,020
is we want to build recommendation systems,

357
00:29:53,200 --> 00:29:56,040
but we don't want to do it in a centralized manner.

358
00:29:56,040 --> 00:29:58,920
We don't want Netflix to be in charge.

359
00:29:59,120 --> 00:30:04,700
We don't want some big company to be in charge of the system.

360
00:30:05,080 --> 00:30:16,020
Did you ever think about how could that happen, that service happen, and be effective, but in a way that doesn't put Netflix in charge of it?

361
00:30:16,020 --> 00:30:25,700
You know, at the time, like what was available for people was a lot different in terms of like what kind of hardware you could buy.

362
00:30:25,880 --> 00:30:28,560
It really was a totally different class.

363
00:30:29,000 --> 00:30:37,460
And, you know, if you really wanted to run big workloads, you had to run it in a cluster or rent a cloud system to do it for you.

364
00:30:37,460 --> 00:30:41,200
and so at the time

365
00:30:41,200 --> 00:30:41,940
like

366
00:30:41,940 --> 00:30:45,400
it just seemed like not possible

367
00:30:45,400 --> 00:30:47,340
for people because you'd have to run the workload

368
00:30:47,340 --> 00:30:49,000
yourself with the data

369
00:30:49,000 --> 00:30:51,320
and it's like

370
00:30:51,320 --> 00:30:52,900
hard drives were still too small

371
00:30:52,900 --> 00:30:55,260
and you're talking about training

372
00:30:55,260 --> 00:30:57,280
the data or about just using

373
00:30:57,280 --> 00:30:59,400
it after it's trained yeah you need

374
00:30:59,400 --> 00:31:01,320
to train the models and then also

375
00:31:01,320 --> 00:31:03,040
do inference and

376
00:31:03,040 --> 00:31:05,320
either doing either of

377
00:31:05,320 --> 00:31:07,120
those like it you know the

378
00:31:07,120 --> 00:31:12,500
it just wasn't it just wasn't possible with what was available at the time. And so I just kind of

379
00:31:12,500 --> 00:31:17,560
like this thing is just going to be a cloud thing, you know, and I still kind of feel that way,

380
00:31:17,560 --> 00:31:23,520
even with modern systems and even with all the talk of running it locally. You know, what you

381
00:31:23,520 --> 00:31:29,840
can buy from Apple these days is just so much better than what they had in the past. And it's

382
00:31:29,840 --> 00:31:34,360
even better than in some cases what you can get from NVIDIA on a, you know, on a per dollar basis.

383
00:31:34,360 --> 00:31:42,240
so um what do you call it that math is now a lot different um you know again just because so many

384
00:31:42,240 --> 00:31:50,000
advancements were made in hardware uh over the last five six years but at the time it just kind

385
00:31:50,000 --> 00:31:56,660
of felt like yeah it's like these things only get better with more data you know it's like they

386
00:31:56,660 --> 00:32:03,620
don't get better with less ever and a lot of how they kept improving models was just

387
00:32:03,620 --> 00:32:05,820
throwing more data at the problem.

388
00:32:06,740 --> 00:32:08,860
Which means that you're going to have to,

389
00:32:09,040 --> 00:32:13,500
you can't just be a regular plebe and doing all that.

390
00:32:13,600 --> 00:32:15,500
You've got to be, the bigger the better.

391
00:32:16,080 --> 00:32:19,020
So if we want recommendation engines

392
00:32:19,020 --> 00:32:22,940
to be truly decentralized,

393
00:32:23,600 --> 00:32:27,340
then traditional AI as we know it

394
00:32:27,340 --> 00:32:29,540
sounds like not the answer.

395
00:32:29,540 --> 00:32:33,960
Well, I think it just comes down to the incentives.

396
00:32:34,740 --> 00:32:44,440
And if you can create, and also comes down if new methods to split up the work and distribute it,

397
00:32:44,620 --> 00:32:50,460
and then incentivize people to label data and then run it through their system.

398
00:32:50,740 --> 00:32:59,000
But there's also a lot of issues with that because you run into all this overhead whenever you try to distribute work.

399
00:32:59,000 --> 00:33:06,780
and that's why nvidia is worth so much money is because they can efficiently distribute work

400
00:33:06,780 --> 00:33:14,520
and induce very little overhead and that's because they have custom hardware that lets

401
00:33:14,520 --> 00:33:21,740
all the nodes in their cluster efficiently communicate with each other so one of the

402
00:33:21,740 --> 00:33:28,620
things that nvidia has is infiniband and that lets you connect it's basically like the super

403
00:33:28,620 --> 00:33:33,440
powered version of Ethernet, you know, instead of connecting everything through a network switch,

404
00:33:33,440 --> 00:33:40,020
you connect to this InfiniBand switch, and that lets you efficiently communicate

405
00:33:40,020 --> 00:33:50,480
without all the overhead of TCP. And that's, you know, if you want to train something bigger and

406
00:33:50,480 --> 00:34:00,400
better than what open AI can, even if you had all of the computers, you know, in the world,

407
00:34:00,660 --> 00:34:05,860
they're still going to be connected with TCP, you know, like the internet protocol.

408
00:34:06,720 --> 00:34:12,180
That's going to induce a ton of overhead. So it'll always just be a lot slower than,

409
00:34:12,440 --> 00:34:17,240
and so instead of taking you a month, I haven't ran the numbers, but you know,

410
00:34:17,240 --> 00:34:21,800
you'll never be as fast as deploying the latest models, even if you have more data than they do.

411
00:34:22,920 --> 00:34:32,760
So as far as like training, like, you know, if you're okay with just always being behind,

412
00:34:33,240 --> 00:34:37,560
then and there's a way to incentivize people to like label things.

413
00:34:37,660 --> 00:34:41,120
And honestly, I think data labeling will just be a very important like,

414
00:34:41,740 --> 00:34:46,800
that might be like UBI in the future, you know, where you're just getting paid by AI companies

415
00:34:46,800 --> 00:34:49,340
to label stuff in your day-to-day life,

416
00:34:49,380 --> 00:34:51,560
and then they give you money for that.

417
00:34:52,620 --> 00:34:55,000
That just goes to train the super AI.

418
00:34:55,540 --> 00:34:58,220
So that would be decentralized in the sense that you'd have

419
00:34:58,220 --> 00:35:01,060
billions of people going and collecting data.

420
00:35:01,420 --> 00:35:01,700
Yeah.

421
00:35:01,900 --> 00:35:06,060
But the entity paying you would be this great big institution

422
00:35:06,060 --> 00:35:07,760
collecting all the data.

423
00:35:08,340 --> 00:35:11,600
So that's not the solution we're looking for.

424
00:35:11,600 --> 00:35:14,880
No, because you just have to trust them

425
00:35:14,880 --> 00:35:20,220
that they're actually using the data the way you labeled it

426
00:35:20,220 --> 00:35:23,060
instead of in a way that's more beneficial to them.

427
00:35:23,200 --> 00:35:24,340
And we'll never know how to,

428
00:35:25,180 --> 00:35:27,660
even if they are doing it in good faith,

429
00:35:27,820 --> 00:35:29,240
how do we know that?

430
00:35:30,740 --> 00:35:35,120
Yeah, you know, all this AI talk right now,

431
00:35:35,280 --> 00:35:37,420
there's just so many aspects to it.

432
00:35:37,420 --> 00:35:39,820
Like, you know, yeah, they want to get bigger,

433
00:35:40,080 --> 00:35:43,060
but to do that, they need more revenue.

434
00:35:43,060 --> 00:35:50,680
and to do that they need people exclusively using their models but also it's possible to clone to

435
00:35:50,680 --> 00:35:58,440
distill models from the bigger models and make local models that are just as good um if not

436
00:35:58,440 --> 00:36:03,580
better depending on if you want to do your own fine-tuning and local training with that model

437
00:36:03,580 --> 00:36:12,920
and um what do you call it like it's just so hard to know how the economics of all of this are going

438
00:36:12,920 --> 00:36:20,100
to like actually play out but i do think yeah i guess we're just going to find an equilibrium

439
00:36:20,100 --> 00:36:28,320
somewhere uh but it's just hard to imagine like these like the open ai like all the super scalers

440
00:36:28,320 --> 00:36:34,620
like you know getting even uh even bigger you know and hyperscalers and mega you know like just

441
00:36:34,620 --> 00:36:38,240
as big as they can because there's going to be a limit to how much money they can raise

442
00:36:38,240 --> 00:36:48,900
um and also like i'm just so not sold on agi like it just it's just so i don't get why

443
00:36:48,900 --> 00:36:55,900
you know they're all going after it just seems like such an obvious dead end because none of

444
00:36:55,900 --> 00:37:02,380
these models and if and if someone can correct me like show me like i'd be so happy but like

445
00:37:02,380 --> 00:37:07,300
none of these models have any form of logical reasoning all they can do is just prompt themselves

446
00:37:07,300 --> 00:37:12,820
again and have the model look at the output and then like run it, you know, in longer and longer

447
00:37:12,820 --> 00:37:19,820
loops, which consumes more tokens. But you do in some cases get better outputs. But it doesn't mean

448
00:37:19,820 --> 00:37:25,840
that it has actual logical reasoning, like, you know, counting parentheses, for example, for these

449
00:37:25,840 --> 00:37:31,160
models is still very difficult. So, you know, if you want to have it like implement something in

450
00:37:31,160 --> 00:37:38,680
brain fuck, uh, or some kind of language like that. It's not, it like, it like messes up. Um,

451
00:37:39,080 --> 00:37:45,720
and, uh, I don't think longer context window, like, I think they're hopeful that like, um,

452
00:37:45,720 --> 00:37:51,800
you know, 10 million token context window kind of fixes this. Um, but I'm just like,

453
00:37:52,260 --> 00:37:58,500
until they actually have a real mechanism to like implement logical reasoning, uh,

454
00:37:58,500 --> 00:38:01,220
I just don't see how you could get to AGI.

455
00:38:02,040 --> 00:38:05,260
That's not to say that it's not incredibly useful already.

456
00:38:06,660 --> 00:38:12,160
You know, and like it is obviously a type of intelligence, but it's not human intelligence.

457
00:38:12,960 --> 00:38:18,840
You know, it's kind of closer, you know, like in this brings up the topic of like what even is intelligence?

458
00:38:18,840 --> 00:38:29,040
You know, like you have forests, for example, that if they start to get attacked, they have their own defense mechanisms.

459
00:38:29,040 --> 00:38:47,080
You know, trees, when you—they've shown that there's certain types of trees in Africa, for example, that when a giraffe starts to eat it, it releases these chemicals that attract these, like, I think they're like flies or like ticks.

460
00:38:47,080 --> 00:38:49,200
and they start to attack the giraffe.

461
00:38:49,320 --> 00:38:51,800
And so the giraffe gets uncomfortable and stops eating the tree.

462
00:38:53,000 --> 00:38:56,380
You know, and so like that's like a type of intelligence right there, right?

463
00:38:56,400 --> 00:38:59,420
Like it's a self-preservation mechanism.

464
00:38:59,420 --> 00:39:01,240
Like that's obviously intelligent.

465
00:39:01,240 --> 00:39:06,380
Like it helps to push your genetic code forward into the future, you know,

466
00:39:06,380 --> 00:39:11,940
in a way that is very creative and, you know, very energy efficient.

467
00:39:11,940 --> 00:39:20,300
And, you know, they've also shown this with like, you know, mushroom, you know, mushroom ecosystems.

468
00:39:20,640 --> 00:39:25,040
Like there's all kinds of and even animals have like, you know, certain types of intelligence.

469
00:39:25,040 --> 00:39:31,520
So it's obviously a type of intelligence, but I don't see the solution that we're looking for.

470
00:39:31,640 --> 00:39:40,660
Right. If our goal is to build decentralized Web of Trust, then the path that existing AI is on is not the path.

471
00:39:40,660 --> 00:39:41,020
No.

472
00:39:41,260 --> 00:39:43,400
At least not for the foreseeable future.

473
00:39:43,580 --> 00:39:59,440
So how can we, if we want to look at the intersection of AI and decentralized web of trust, what would be, I mean, obviously one intersection would be we can use agents to help us build the tools that we want to build.

474
00:39:59,880 --> 00:40:00,200
Yes.

475
00:40:00,200 --> 00:40:10,440
So what other, so I'm thinking of other, you know, just topics that we can talk about.

476
00:40:10,660 --> 00:40:15,780
Like suppose you've got information,

477
00:40:15,900 --> 00:40:19,700
suppose you have data that is curated by your web of trust,

478
00:40:19,800 --> 00:40:23,980
not using AI, but it's data that you trust

479
00:40:23,980 --> 00:40:26,000
because your community curated it for you.

480
00:40:26,520 --> 00:40:27,120
Yes.

481
00:40:27,440 --> 00:40:33,180
And then you fed that into an LLM, like a local model.

482
00:40:33,860 --> 00:40:34,100
Yeah.

483
00:40:34,300 --> 00:40:37,920
What do you think of that kind of idea

484
00:40:37,920 --> 00:40:40,580
where you've got sitting at home,

485
00:40:40,660 --> 00:40:43,560
You've got a box that's got, let's say, a Nostra Relay,

486
00:40:44,100 --> 00:40:45,780
and it's got a huge amount of data.

487
00:40:45,780 --> 00:40:50,220
It could have a Wikipedia worth and even more than that.

488
00:40:50,660 --> 00:40:55,360
Everything that's curated, not through any big companies.

489
00:40:55,540 --> 00:41:00,500
But then you feed that into AI, into a local model.

490
00:41:00,900 --> 00:41:05,660
How would that, like what model would you feed it into?

491
00:41:06,600 --> 00:41:09,120
That's a really, really interesting question

492
00:41:09,120 --> 00:41:11,660
because there's two ways to think about it.

493
00:41:12,020 --> 00:41:16,940
One, you can use that as the underlying data for the training,

494
00:41:17,560 --> 00:41:18,860
the actual training of the models.

495
00:41:19,000 --> 00:41:20,100
So like from scratch?

496
00:41:20,720 --> 00:41:21,160
Yeah.

497
00:41:21,420 --> 00:41:24,700
Would it be feasible for an individual to do that?

498
00:41:25,480 --> 00:41:28,100
You're not having huge data centers.

499
00:41:28,100 --> 00:41:28,840
You're just a person.

500
00:41:28,980 --> 00:41:32,760
You're building a model purely from scratch.

501
00:41:32,820 --> 00:41:44,870
Is that doable on a small scale I think if you were if you know with Nostra relays like they often have like topics that they focus on

502
00:41:45,270 --> 00:41:48,150
So if you wanted to like narrowly train it

503
00:41:48,150 --> 00:41:53,170
on a, you know, a certain type of subject matter,

504
00:41:53,170 --> 00:41:56,870
then I think that is definitely possible.

505
00:41:57,430 --> 00:41:59,010
So if you had Nostra data

506
00:41:59,010 --> 00:42:01,550
that was already organized by topic.

507
00:42:01,610 --> 00:42:02,910
Yeah, and properly labeled

508
00:42:02,910 --> 00:42:05,230
and you knew without putting in more work,

509
00:42:05,250 --> 00:42:06,830
because that's really the problem is,

510
00:42:07,350 --> 00:42:08,490
you know, I think that, like I said,

511
00:42:08,510 --> 00:42:11,070
the data labeling is just going to be

512
00:42:11,070 --> 00:42:12,550
a huge thing moving forward.

513
00:42:13,010 --> 00:42:15,090
And if you're receiving all this data

514
00:42:15,090 --> 00:42:17,150
and it's already curated by the members

515
00:42:17,150 --> 00:42:18,430
simply by just sharing it,

516
00:42:18,430 --> 00:42:20,510
and you know that the person it came from

517
00:42:20,510 --> 00:42:21,770
is highly trusted and vetted,

518
00:42:22,270 --> 00:42:25,890
then, you know, that opens up a world of possibilities.

519
00:42:26,730 --> 00:42:28,330
But, you know, if you want it to just be

520
00:42:28,330 --> 00:42:31,950
like a very general, very generic model,

521
00:42:31,950 --> 00:42:34,970
that sounds not realistic.

522
00:42:35,550 --> 00:42:38,070
And even then somebody actually training through this,

523
00:42:38,230 --> 00:42:42,390
like it would be a significant investment to like run at home.

524
00:42:43,170 --> 00:42:45,610
You know, the costs are dropping every year,

525
00:42:46,450 --> 00:42:51,050
but you know, like, like I said,

526
00:42:51,070 --> 00:42:53,390
it's hard to know where everything is going to land

527
00:42:53,390 --> 00:42:56,430
because I do think that the demand for this,

528
00:42:56,430 --> 00:43:00,830
you know, artificial intelligence is going to cause shortages

529
00:43:00,830 --> 00:43:04,130
in actual material resources,

530
00:43:04,430 --> 00:43:05,890
you know, like copper, for example,

531
00:43:06,670 --> 00:43:08,390
or natural gas even.

532
00:43:09,350 --> 00:43:11,690
And it's, you know, it's like,

533
00:43:12,570 --> 00:43:14,270
it's cheaper to produce them

534
00:43:14,270 --> 00:43:15,870
because, you know, we have more economies of scale,

535
00:43:15,930 --> 00:43:16,690
but at the same time,

536
00:43:16,750 --> 00:43:19,810
like the raw inputs are getting more expensive

537
00:43:19,810 --> 00:43:22,370
because, you know, we're just not producing enough.

538
00:43:23,530 --> 00:43:25,670
So it's hard to, without,

539
00:43:26,270 --> 00:43:28,410
I think even if you tried to model this somehow,

540
00:43:28,410 --> 00:43:38,610
Like hard to know where, you know, the actual final equilibrium is going to land at for if it's practical or not to train at home.

541
00:43:39,430 --> 00:43:45,110
And just because so we still have the problem that it's just takes so much energy to train from scratch.

542
00:43:45,810 --> 00:43:51,930
That and also your the model is as good as like the training.

543
00:43:51,930 --> 00:44:00,450
You know, like there's lots of really smart techniques and, you know, by brilliant engineers that like squeeze out the most amount of intelligence per training run.

544
00:44:00,610 --> 00:44:15,350
But at the end of the day, you know, if you if your training set has a bunch of dogs that are labeled as cat and you want to use the AI to label dogs, you know, it's going to be difficult.

545
00:44:15,690 --> 00:44:17,270
Yeah, that's garbage in, garbage out.

546
00:44:17,690 --> 00:44:17,950
Yeah.

547
00:44:17,950 --> 00:44:32,830
But if we have a system that filters out the garbage effectively and labels it effectively, which is what, you know, like knowledge graph or like concept graph, maybe.

548
00:44:33,470 --> 00:44:40,570
How far could that go at saving, at reducing the amount of energy it takes to train something effectively?

549
00:44:41,430 --> 00:44:42,110
You know, I think.

550
00:44:43,010 --> 00:44:47,610
Are we talking orders of magnitude or just a little, you know, shaving off the edges?

551
00:44:47,950 --> 00:44:58,550
I think it probably just, you know, if, if, again, if you want to make a, like a general model, like something like chat GPT, then you're just shaving off the edges.

552
00:44:58,550 --> 00:45:02,010
but if you want to do

553
00:45:02,010 --> 00:45:04,390
the mixture of experts

554
00:45:04,390 --> 00:45:06,330
models sort of like with DeepSeek

555
00:45:06,330 --> 00:45:06,850
Innovated

556
00:45:06,850 --> 00:45:09,470
then there's

557
00:45:09,470 --> 00:45:11,350
opportunities there for sure

558
00:45:11,350 --> 00:45:13,890
because you would just have each Nostra

559
00:45:13,890 --> 00:45:15,890
relay that subscribes to a certain topic

560
00:45:15,890 --> 00:45:18,010
and just have it train

561
00:45:18,010 --> 00:45:19,890
on that and then

562
00:45:19,890 --> 00:45:22,190
you would collect all those

563
00:45:22,190 --> 00:45:24,290
model of experts

564
00:45:24,290 --> 00:45:25,770
mix them together

565
00:45:25,770 --> 00:45:28,050
and then that would give you

566
00:45:28,050 --> 00:45:31,110
a more general model that could do more things.

567
00:45:32,190 --> 00:45:35,510
And it's not even clear that that's even that useful.

568
00:45:36,310 --> 00:45:38,650
You know, if I'm coding,

569
00:45:39,170 --> 00:45:41,710
do I really need to know about French history?

570
00:45:42,030 --> 00:45:44,710
You know, if I'm like, you know,

571
00:45:45,090 --> 00:45:47,090
trying to make like a Rust program or something,

572
00:45:47,210 --> 00:45:49,230
trying to make it go faster, like find bugs,

573
00:45:50,150 --> 00:45:51,750
maybe, probably not.

574
00:45:51,750 --> 00:45:54,250
Right now, ChatGPT knows about French history

575
00:45:54,250 --> 00:45:55,330
and you're using it to code.

576
00:45:55,330 --> 00:45:56,670
Yeah, for example.

577
00:45:56,670 --> 00:45:58,390
Yeah, just as one example.

578
00:45:59,530 --> 00:46:03,030
So being able to make specialized AI.

579
00:46:05,970 --> 00:46:09,930
So could you envision a world where there's a bunch of open source models

580
00:46:09,930 --> 00:46:15,850
and you can choose, you've got lots to choose from.

581
00:46:16,210 --> 00:46:21,350
If one of them is a bad actor, you've got other ones that you can easily switch to.

582
00:46:21,930 --> 00:46:22,170
Yeah.

583
00:46:22,170 --> 00:46:25,050
And then you train your own personal data.

584
00:46:25,870 --> 00:46:26,030
Yeah.

585
00:46:26,670 --> 00:46:33,190
Whether that's fine-tuning or, you know, some other, like, special method.

586
00:46:33,950 --> 00:46:34,490
Yeah, absolutely.

587
00:46:34,970 --> 00:46:36,830
What's the difference between training and fine-tuning?

588
00:46:37,430 --> 00:46:41,330
So fine-tuning, yeah, I mean, training is, you know,

589
00:46:41,410 --> 00:46:45,630
considered more like an end-to-end methodology of producing...

590
00:46:45,630 --> 00:46:47,030
That's back propagation and...

591
00:46:47,030 --> 00:46:49,270
Yeah, forward propagation and, like, all these other methods

592
00:46:49,270 --> 00:46:54,230
to actually produce a final model that can be, you know, somewhat useful.

593
00:46:54,230 --> 00:47:14,330
But fine-tuning is where, for example, you have the model, you're talking to the model all day and it has, you know, it has helps you do to-do lists, for example, manage a to-do list or, you know, you're just discussing French history, right?

594
00:47:14,330 --> 00:47:17,250
all of that text that it outputs,

595
00:47:17,650 --> 00:47:23,450
what you can do is basically put it back into the model

596
00:47:23,450 --> 00:47:27,270
so that those weights that actually triggered those responses

597
00:47:27,270 --> 00:47:29,410
are now more heavily weighted

598
00:47:29,410 --> 00:47:34,190
and more likely to produce the correct output next time.

599
00:47:34,530 --> 00:47:35,730
So you're changing the weights.

600
00:47:36,510 --> 00:47:36,830
Yeah.

601
00:47:36,950 --> 00:47:39,570
But it's different in methodology than training.

602
00:47:40,150 --> 00:47:42,510
Yeah, because you're only doing it a little bit, right?

603
00:47:42,510 --> 00:47:45,810
and you're not actually going deep into the model

604
00:47:45,810 --> 00:47:48,450
and updating those weights throughout the entire,

605
00:47:48,950 --> 00:47:51,490
all the different layers that are in the model.

606
00:47:52,150 --> 00:47:56,010
It's just on select layers that are more towards the end

607
00:47:56,010 --> 00:47:58,630
and that will just slightly bias the model

608
00:47:58,630 --> 00:48:02,650
to be more likely to discuss French history.

609
00:48:03,110 --> 00:48:05,870
And then those changes that you make

610
00:48:05,870 --> 00:48:08,110
are not rippling through everywhere else.

611
00:48:08,930 --> 00:48:10,790
Well, it's fine tuning, you know,

612
00:48:10,830 --> 00:48:11,890
so it's just a little bit.

613
00:48:12,510 --> 00:48:20,730
And that's sort of what people are getting really excited about now about running local models is you have it, for example, be your personal assistant.

614
00:48:20,730 --> 00:48:42,790
And, you know, there's this concept of the system prompt where you before you actually start to talk about your discussion, you actually like you prefix with, hey, you know, my name is Matias and I'm really interested in French history, you know, especially Napoleon, you know, and how he fought in Waterloo and like all these things.

615
00:48:42,790 --> 00:48:46,370
and you basically just,

616
00:48:46,710 --> 00:48:50,190
you're just going to get much better results

617
00:48:50,190 --> 00:48:51,430
if you fine tune it

618
00:48:51,430 --> 00:48:54,890
and it's just less likely to go off track

619
00:48:54,890 --> 00:48:59,370
or, you know, start discussing British history,

620
00:48:59,650 --> 00:49:00,970
you know, instead of French history

621
00:49:00,970 --> 00:49:04,170
or go off on a tangent, you know,

622
00:49:04,170 --> 00:49:06,050
because again, those models towards

623
00:49:06,050 --> 00:49:09,250
the lower levels of the layer

624
00:49:09,250 --> 00:49:12,390
are much more likely to get triggered

625
00:49:12,390 --> 00:49:18,010
at least the, sorry, so the weights that are related to, for example, French history from

626
00:49:18,010 --> 00:49:20,130
all your discussions are much more likely to get triggered.

627
00:49:20,930 --> 00:49:24,370
And if, you know, that's what you want, then that's much more beneficial.

628
00:49:24,610 --> 00:49:27,990
And you can see this as how this could be useful.

629
00:49:28,610 --> 00:49:31,990
You know, you're using it as a personal assistant day in, day out.

630
00:49:32,530 --> 00:49:37,670
You could take the logs of all of your discussions throughout that day and then use that to fine

631
00:49:37,670 --> 00:49:39,910
tune it every night, for example.

632
00:49:39,910 --> 00:49:47,090
And then so the next day, you don't have to like so heavily prompted to like have it be a personal assistant.

633
00:49:47,250 --> 00:49:50,270
You know, it just kind of has that in its own weights.

634
00:49:50,790 --> 00:50:00,630
But the reason why it's called fine tuning is because you just, you know, you're not going to get that much mileage out of whatever, you know, updating of the weights that you do.

635
00:50:00,730 --> 00:50:09,550
If you're trying to like so heavily influence all the weights for that, you're going to want to have it done on a cluster.

636
00:50:09,910 --> 00:50:13,810
rather than your laptop, even if you have a GPU.

637
00:50:14,050 --> 00:50:18,830
If you want to do training, it simply requires a whole lot more computing power.

638
00:50:19,270 --> 00:50:22,670
Well, I think if you want to do training for a general model,

639
00:50:22,670 --> 00:50:27,830
like something that's like Opus or ChatGPT, then yeah, you just have to.

640
00:50:28,330 --> 00:50:31,170
Because there's just so much training material that you have to run through

641
00:50:31,170 --> 00:50:35,790
and turn it into embeddings and just have it go through that.

642
00:50:35,950 --> 00:50:38,450
But if it's highly specialized,

643
00:50:38,450 --> 00:50:49,450
So, for example, you know, I'd be interested in doing fine tuning for like a Rust programming agent.

644
00:50:50,570 --> 00:50:59,530
Like and then I could, for example, use all of the outputs of all the tools that I've been running that day.

645
00:50:59,590 --> 00:51:06,170
So like Cargo Check, Cargo Thumped, you know, Cargo Clippy and use that for fine tuning.

646
00:51:06,170 --> 00:51:17,810
And then over time, it would just be very intuitive to the agent to call Cargo Clippy and it would just have to spend fewer tokens figuring out that, hey, I should do tool call for Cargo Clippy.

647
00:51:18,130 --> 00:51:19,290
And it would just do it.

648
00:51:19,510 --> 00:51:32,870
So let's say if you wanted to put together a personal assistant and you wanted to pair that with a Nostra relay that had information that was curated in a way that you trusted.

649
00:51:32,870 --> 00:51:35,910
And it could be a huge amount of information.

650
00:51:36,170 --> 00:51:36,810
Yeah.

651
00:51:37,470 --> 00:51:41,470
And then let's say that you do that at home.

652
00:51:41,650 --> 00:51:45,810
Like you could have that an instance that's sitting in a trusted cloud provider

653
00:51:45,810 --> 00:51:50,870
or even more self-sovereign would be that you've got a box at home.

654
00:51:52,190 --> 00:51:57,370
So would you have OpenClaw in that box?

655
00:51:57,630 --> 00:51:59,970
How would you do that?

656
00:52:00,070 --> 00:52:00,890
How would you engineer that?

657
00:52:01,270 --> 00:52:05,590
Yeah, I think one, something that may be kind of interesting,

658
00:52:05,590 --> 00:52:11,210
would be to run OpenClaw with Opus,

659
00:52:11,610 --> 00:52:15,010
but then have as a skill

660
00:52:15,010 --> 00:52:18,630
one of these open-source models

661
00:52:18,630 --> 00:52:19,750
that you've been fine-tuning

662
00:52:19,750 --> 00:52:24,970
with the data that it knows you find interesting.

663
00:52:25,250 --> 00:52:26,770
So you'd have to come up with some interface

664
00:52:26,770 --> 00:52:29,490
to record how much time you're spending

665
00:52:29,490 --> 00:52:31,250
on a photo or a post.

666
00:52:31,650 --> 00:52:33,670
What if you just gave it your Nostro Relay

667
00:52:33,670 --> 00:52:34,610
of curated data?

668
00:52:35,590 --> 00:52:43,250
Well, you still need to find, you need to like label which of the posts is more interesting and less interesting.

669
00:52:43,570 --> 00:52:50,530
So let's say you're organized in a graph database using like the decentralized lists system.

670
00:52:50,750 --> 00:52:54,530
And so I'm interested in sports, but I'm not interested in politics.

671
00:52:54,990 --> 00:52:59,110
And you just look at the graph database to see what fits into that umbrella.

672
00:52:59,750 --> 00:53:05,270
But then you'd have to like define upfront and you'd like be labeling data manually.

673
00:53:05,590 --> 00:53:10,070
up front. So like what I'm talking about is just sort of like an interface like X.

674
00:53:10,970 --> 00:53:16,550
And, you know, you would just like vibe code this yourself. So you have control over all

675
00:53:16,550 --> 00:53:20,850
the algorithm and it would be done in a way where it's not extractive. You know, there's

676
00:53:20,850 --> 00:53:26,010
no dark patterns, but it could record, for example, oh, post one, two, three, four.

677
00:53:27,050 --> 00:53:33,150
Looks like Matias has spent a whole minute on this. And so that post, you would just add

678
00:53:33,150 --> 00:53:35,050
four minutes to it and

679
00:53:35,050 --> 00:53:37,170
basically all

680
00:53:37,170 --> 00:53:39,210
of the posts, like the ones that you just scroll

681
00:53:39,210 --> 00:53:40,610
through, zero seconds, right?

682
00:53:40,870 --> 00:53:42,410
Oh, this other post,

683
00:53:42,650 --> 00:53:45,170
he's not just the post, but he's also

684
00:53:45,170 --> 00:53:46,150
looking at all the comments.

685
00:53:46,790 --> 00:53:49,170
Then I would put those

686
00:53:49,170 --> 00:53:51,330
in a database,

687
00:53:51,550 --> 00:53:51,770
create

688
00:53:51,770 --> 00:53:55,030
embeddings for it, and then

689
00:53:55,030 --> 00:53:56,790
fine-tune

690
00:53:56,790 --> 00:53:58,750
something like Quen

691
00:53:58,750 --> 00:54:01,370
3 or DeepSeek

692
00:54:01,370 --> 00:54:02,730
with

693
00:54:02,730 --> 00:54:07,290
that ranking of posts.

694
00:54:07,630 --> 00:54:12,290
And then it could then have a classifier as a skill.

695
00:54:12,390 --> 00:54:15,090
It would just run like a CLI that just spins up the model,

696
00:54:15,210 --> 00:54:18,010
loads it into memory, and you can give it a post

697
00:54:18,010 --> 00:54:19,630
as it's coming in through the relay.

698
00:54:20,550 --> 00:54:25,730
And it will classify it as either something that Matias is interested in,

699
00:54:26,590 --> 00:54:29,850
likely interested in, or not interested in at all.

700
00:54:29,850 --> 00:54:34,490
You're measuring interest by the things you do.

701
00:54:34,750 --> 00:54:37,630
How long do your eyes linger on this post?

702
00:54:38,050 --> 00:54:38,670
For example.

703
00:54:39,290 --> 00:54:39,470
Yeah.

704
00:54:40,050 --> 00:54:41,830
So it's AI is kind of, well,

705
00:54:42,150 --> 00:54:47,690
these are techniques that are currently being used on us.

706
00:54:48,490 --> 00:54:51,610
Whether they're good or bad could be debated.

707
00:54:52,370 --> 00:54:57,150
If this system, this is already basically how Facebook and Instagram works,

708
00:54:57,150 --> 00:55:05,610
But like in this system, it would just be a fairly straightforward, you know, ranking algorithm and just like a way to store that information.

709
00:55:05,610 --> 00:55:07,930
But it would be ran and operated by you.

710
00:55:08,270 --> 00:55:11,910
You know, there's no reason why you can't run your own Postgres database that the app connects to.

711
00:55:12,270 --> 00:55:18,530
And it's just storing that information there, basically just a log of how you're using the app.

712
00:55:18,590 --> 00:55:24,150
And then that's, you know, you're doing the work of Facebook, but doing it in a way where it's not extractive.

713
00:55:24,150 --> 00:55:33,150
Could that methodology, though, be one that is inherently developed for...

714
00:55:33,240 --> 00:55:34,360
the purposes of Facebook.

715
00:55:34,860 --> 00:55:35,840
You know, like I don't want,

716
00:55:35,960 --> 00:55:37,980
just because my eyes look at something,

717
00:55:38,080 --> 00:55:39,640
I don't want my personal assistant

718
00:55:39,640 --> 00:55:41,400
to feed more of it to me necessarily.

719
00:55:42,260 --> 00:55:42,660
That's true.

720
00:55:43,100 --> 00:55:44,020
That's another aspect.

721
00:55:44,180 --> 00:55:45,080
I mean, but inherently,

722
00:55:45,280 --> 00:55:47,000
the stuff that you pay attention to

723
00:55:47,000 --> 00:55:48,660
is what you find interesting.

724
00:55:48,660 --> 00:55:52,160
But that's not necessarily what's best for us.

725
00:55:52,480 --> 00:55:54,140
I mean, that's kind of the, you know,

726
00:55:54,240 --> 00:55:55,180
it's our dopamine.

727
00:55:55,520 --> 00:55:58,840
Do we want our personal assistant

728
00:55:58,840 --> 00:56:01,540
to make our ADD worse

729
00:56:01,540 --> 00:56:04,480
by giving us the stuff that pumps our dopamine?

730
00:56:05,040 --> 00:56:05,960
Probably not.

731
00:56:06,540 --> 00:56:11,820
I think if you were to guard against certain topics

732
00:56:11,820 --> 00:56:13,580
that are inherently,

733
00:56:13,900 --> 00:56:16,240
and you could use some other classifier

734
00:56:16,240 --> 00:56:17,420
that's just out of the box,

735
00:56:18,060 --> 00:56:19,740
so anything that's really violent

736
00:56:19,740 --> 00:56:23,600
or nudity or anything like that

737
00:56:23,600 --> 00:56:25,220
could just immediately get filtered out.

738
00:56:25,400 --> 00:56:27,140
And then as long as you just keep it to text

739
00:56:27,140 --> 00:56:33,840
and you just automatically ban certain words

740
00:56:33,840 --> 00:56:37,780
depending on what you know you wouldn't want to get interested in,

741
00:56:38,340 --> 00:56:43,740
then such a system, I think, could be developed

742
00:56:43,740 --> 00:56:44,980
and could be interesting.

743
00:56:46,260 --> 00:56:48,160
But I totally agree with you.

744
00:56:49,540 --> 00:56:51,700
Just because you're paying attention to it

745
00:56:51,700 --> 00:56:53,640
doesn't mean it's what's best for you.

746
00:56:53,840 --> 00:56:57,120
But at the same time, then without having somebody else

747
00:56:57,120 --> 00:57:00,260
sort of towering over you and making sure you're, you know,

748
00:57:00,320 --> 00:57:03,660
adhering to some moral code of what they think you should be interested in,

749
00:57:03,960 --> 00:57:05,320
then I, you know,

750
00:57:05,400 --> 00:57:08,780
it needs some kind of way where you're in control at the end of the day.

751
00:57:09,020 --> 00:57:10,800
Right. So we definitely, I mean, that's, yeah,

752
00:57:10,840 --> 00:57:12,800
the goal is to put you in control,

753
00:57:13,080 --> 00:57:18,020
which includes the ability for you to delegate decisions to trusted people,

754
00:57:18,940 --> 00:57:21,560
to help for them, to help for them.

755
00:57:23,200 --> 00:57:24,740
And is there a need?

756
00:57:24,740 --> 00:57:28,520
So do we need AI at this particular step?

757
00:57:28,880 --> 00:57:34,500
Do we actually need AI at that particular step to organize our data?

758
00:57:34,660 --> 00:57:40,320
What if our grapevine, our web of trust, can organize all the data for us?

759
00:57:41,000 --> 00:57:48,460
And big picture and fine-grained detail are if putting things into categories

760
00:57:48,460 --> 00:57:53,580
is done by human beings, by entities not using AI.

761
00:57:53,580 --> 00:57:54,820
and

762
00:57:54,820 --> 00:57:55,880
if

763
00:57:55,880 --> 00:57:58,940
I'm interested in sports

764
00:57:58,940 --> 00:58:00,780
what are the different categories of sports

765
00:58:00,780 --> 00:58:02,500
I don't want AI to categorize them

766
00:58:02,500 --> 00:58:04,360
I'd want human beings to categorize them

767
00:58:04,360 --> 00:58:07,140
people who know about sports

768
00:58:07,140 --> 00:58:08,740
and who my web of trust says

769
00:58:08,740 --> 00:58:10,000
these people know what they're talking about

770
00:58:10,000 --> 00:58:12,440
and they're going to be the ones who invent the categories

771
00:58:12,440 --> 00:58:14,840
and they're going to be the ones who maybe

772
00:58:14,840 --> 00:58:18,920
figure out how to place

773
00:58:18,920 --> 00:58:20,900
data into these categories

774
00:58:20,900 --> 00:58:23,220
so that's

775
00:58:23,220 --> 00:58:26,920
doesn't necessarily need AI.

776
00:58:27,400 --> 00:58:28,300
An LLM, you mean?

777
00:58:28,820 --> 00:58:30,400
Or, I mean, nothing.

778
00:58:30,580 --> 00:58:33,740
No AI, no LLMs, no machine learning.

779
00:58:34,540 --> 00:58:35,980
No connectivists.

780
00:58:36,820 --> 00:58:40,240
You ultimately, like, yeah,

781
00:58:40,300 --> 00:58:41,600
you would just have like a list

782
00:58:41,600 --> 00:58:44,080
and it would be the top, the post,

783
00:58:44,080 --> 00:58:47,240
and then a ranking from zero to one

784
00:58:47,240 --> 00:58:51,320
about, you know, what is most interesting.

785
00:58:51,440 --> 00:58:52,140
You would sort it.

786
00:58:52,140 --> 00:58:57,720
and then just start, you know, popping off that list

787
00:58:57,720 --> 00:59:01,200
and the viewer could just scroll past it

788
00:59:01,200 --> 00:59:02,540
or like pay attention to it.

789
00:59:03,080 --> 00:59:08,540
I think if you were to basically create a training set

790
00:59:08,540 --> 00:59:10,600
of like label data in such a manner

791
00:59:10,600 --> 00:59:12,040
and then create a classifier,

792
00:59:12,600 --> 00:59:13,940
you know, you could use like an LLM,

793
00:59:14,040 --> 00:59:15,220
like for example, Grok1,

794
00:59:15,880 --> 00:59:19,060
they've already shown that, you know,

795
00:59:19,060 --> 00:59:27,440
that you can just put a bunch of a huge list of rules and embedded into an LLM with, you know, into the weights,

796
00:59:27,640 --> 00:59:34,680
then it can very efficiently just determine if it's according to these existing rules,

797
00:59:34,680 --> 00:59:37,640
if it's something that the user might be interested in.

798
00:59:37,700 --> 00:59:44,640
So maybe there's like some hybrid approach where you would want the initial ranking to be done by the grapevine.

799
00:59:44,640 --> 00:59:52,180
but then over time because you also have to be cognizant of like the volume of data you know and

800
00:59:52,180 --> 00:59:57,820
it's like maybe it is actually a little bit more efficient after you know the rules of like say

801
00:59:57,820 --> 01:00:03,780
operating up for a whole year that what the user likes and what it doesn't you know how it gets

802
01:00:03,780 --> 01:00:07,260
classified maybe it is actually a little more efficient to just like run it through like a

803
01:00:07,260 --> 01:00:12,160
fine-tuned version of grok one that would like tell you how interested the user might be in that

804
01:00:12,160 --> 01:00:19,000
specific post. So, you know, I think like to like bootstrap it, you would want to use the grapevine

805
01:00:19,000 --> 01:00:23,720
because it is so high quality. But then once you get into enough volume, you know, of data,

806
01:00:24,740 --> 01:00:31,080
you're, you know, you would have to come up with some shortcuts. Because otherwise,

807
01:00:31,080 --> 01:00:36,240
you're just storing a lot of data and running a lot of CPU cycles. And maybe there's just like

808
01:00:36,240 --> 01:00:41,100
a much better way to go about it. I don't know. I think we're going to find out in the next few

809
01:00:41,100 --> 01:00:49,180
months. I'm super excited, by the way, to get started on this. And I think there's just a lot,

810
01:00:49,300 --> 01:00:57,160
like the design space is huge. And this is one of the biggest problems that I think is facing

811
01:00:57,160 --> 01:01:02,820
our society right now, because the amount of nonsense that is being put out there online,

812
01:01:02,820 --> 01:01:10,100
it's just huge. It's affecting everyone I know. And there doesn't seem to be a solution other than

813
01:01:10,100 --> 01:01:17,000
to just hope that the CEOs of big tech companies actually do something about it.

814
01:01:17,040 --> 01:01:18,760
And they can't because they know that.

815
01:01:18,760 --> 01:01:19,200
They don't know what to do.

816
01:01:19,420 --> 01:01:20,980
I mean, they don't know what to do.

817
01:01:21,380 --> 01:01:24,460
They don't want to impose their moral code on us.

818
01:01:24,520 --> 01:01:26,100
I think they understand that, right?

819
01:01:26,880 --> 01:01:29,200
And they're also, you know, a lot of them are publicly traded.

820
01:01:29,400 --> 01:01:33,420
And so they have a duty to the shareholders to maximize revenue

821
01:01:33,420 --> 01:01:36,280
and maximize shareholder value.

822
01:01:36,280 --> 01:01:45,900
And that means that they have to make the platform, have to make you spend more time on it so that they can sell your profiles to advertisers, right?

823
01:01:45,960 --> 01:01:46,900
And that's how they make money.

824
01:01:46,900 --> 01:01:55,620
The incentivization, just the scenario they find themselves in is going to incentivize them to do things.

825
01:01:55,620 --> 01:01:57,720
formatted. Yeah, they can be

826
01:01:57,720 --> 01:01:59,960
have great intentions

827
01:01:59,960 --> 01:02:01,800
but they are in the

828
01:02:01,800 --> 01:02:03,940
you know, they have to act

829
01:02:03,940 --> 01:02:05,760
according to the incentive structure

830
01:02:05,760 --> 01:02:06,700
that they are in.

831
01:02:08,520 --> 01:02:09,800
Ultimately, that's

832
01:02:09,800 --> 01:02:11,940
yeah, that's how they operate and

833
01:02:11,940 --> 01:02:13,520
what do you call it?

834
01:02:15,140 --> 01:02:15,900
That's just

835
01:02:15,900 --> 01:02:17,740
how that game works and

836
01:02:17,740 --> 01:02:19,840
the only way, in my opinion, is

837
01:02:19,840 --> 01:02:21,000
just not participate.

838
01:02:21,820 --> 01:02:23,920
And build better alternatives, which is what

839
01:02:23,920 --> 01:02:27,580
our goal is. You and I are going to be working together to help

840
01:02:27,580 --> 01:02:31,840
with others to hopefully be able to do that.

841
01:02:31,840 --> 01:02:35,880
What I hope we'll be able

842
01:02:35,880 --> 01:02:39,840
to do is use your web of

843
01:02:39,840 --> 01:02:43,740
trust to organize data and to organize it in a

844
01:02:43,740 --> 01:02:46,800
sophisticated enough manner that

845
01:02:46,800 --> 01:02:53,500
it would be easy for AI to

846
01:02:53,500 --> 01:02:57,680
If you ask AI to find something, it's not going to have to search everything,

847
01:02:58,120 --> 01:03:02,380
but it'll just have to go down the organizational structure.

848
01:03:04,060 --> 01:03:06,900
So it's not just here's a list of categories.

849
01:03:07,140 --> 01:03:08,600
It's here's a list of subcategories.

850
01:03:08,600 --> 01:03:16,820
And here's how you can look at the coarse-grained or fine-grained for any given topic.

851
01:03:16,820 --> 01:03:25,960
Like, you know, if you want to, for any list of things, all right, well, what are the properties that that list, that items in that list need?

852
01:03:26,380 --> 01:03:27,580
Well, that's another list.

853
01:03:27,580 --> 01:03:37,520
And so your community is, it's not going to be any centralized entity that decides what those properties are.

854
01:03:38,020 --> 01:03:45,680
For the people that want to, you know, I think another thing that we should be cognizant of is that some people really do just want to live in the matrix.

855
01:03:45,680 --> 01:03:54,100
I remember this one time I installed PyHole on my local network.

856
01:03:54,960 --> 01:03:56,680
I was at my parents' house at the time.

857
01:03:57,940 --> 01:03:59,760
Have you ever used PyHole?

858
01:03:59,920 --> 01:04:00,820
I've never used it.

859
01:04:01,000 --> 01:04:02,800
It's a great tool.

860
01:04:03,060 --> 01:04:04,880
I think the project is still going.

861
01:04:05,080 --> 01:04:07,100
It's a network-wide ad blocker.

862
01:04:07,100 --> 01:04:18,200
And basically you just blacklist all of these URLs that websites use to download advertising scripts or malicious scripts.

863
01:04:18,800 --> 01:04:33,300
And so when you download a website and it runs the JavaScript and it goes to those URLs that are embedded in the website to download the ads, to show you the sidebar like, hey, TJ Maxx is having a sale or whatever.

864
01:04:33,300 --> 01:04:39,980
you know if you know ahead of time like what those urls are when your laptop you're on your local

865
01:04:39,980 --> 01:04:45,600
wi-fi is goes out to make that request it basically just gets sucked into the pie hole

866
01:04:45,600 --> 01:04:52,140
and the ad fails to load but it's like a network issue you know so like websites don't get don't

867
01:04:52,140 --> 01:04:59,540
block themselves as the network fails like that so you get to use the website but the ad doesn't

868
01:04:59,540 --> 01:05:01,640
load. I think that's like

869
01:05:01,640 --> 01:05:03,100
websites have changed so much.

870
01:05:03,580 --> 01:05:05,480
This was many years ago, but like I

871
01:05:05,480 --> 01:05:07,420
installed this point is I installed this

872
01:05:07,420 --> 01:05:09,600
and then the next day my mom asked

873
01:05:09,600 --> 01:05:11,640
me, it's like, hey, where did my Google ads go?

874
01:05:12,160 --> 01:05:13,460
And I was like, what do you mean?

875
01:05:13,520 --> 01:05:15,620
I blocked them all network wide.

876
01:05:15,700 --> 01:05:17,080
You don't have to worry about them anymore.

877
01:05:17,280 --> 01:05:18,940
It's like they're showing you all this crap.

878
01:05:19,000 --> 01:05:21,180
Oh no, but I really like them.

879
01:05:23,540 --> 01:05:25,080
Yeah, we can build alternatives

880
01:05:25,080 --> 01:05:26,720
and then people can be free to

881
01:05:26,720 --> 01:05:29,480
choose them. And just give them options

882
01:05:29,480 --> 01:05:38,100
And if they like it, they can keep using it and hopefully it benefits them and provides them with more high quality information than they would get otherwise.

883
01:05:38,500 --> 01:05:47,780
But a lot of people, I think the most disappointing aspect of AI is just learning how much slop people actually like.

884
01:05:47,780 --> 01:06:05,360
We like the slop, and yet we at least want the option to filter it out and not to rely on the big companies to do the filtering for us.

885
01:06:05,360 --> 01:06:10,040
No, no, right? Because they can't.

886
01:06:10,040 --> 01:06:18,160
And they don't want, even if they probably could, but that would go against their, that wouldn't work for them.

887
01:06:19,640 --> 01:06:35,582
They want to get our dopamine going And if slop does that then why are they going to get rid of the slop No they slop if anything is going to because like I mean when you see like a like a

888
01:06:35,582 --> 01:06:42,702
like a violent video or like a crazy storm, like these AI slop videos, like they look crazy,

889
01:06:42,962 --> 01:06:48,102
you know? And so you're spending time to like, try to think like, is this real or not? And even

890
01:06:48,102 --> 01:06:53,602
And just that draws you, your curiosity just draws you into it no matter what.

891
01:06:53,842 --> 01:07:00,682
And so the AI gets really good at giving us that kind of stuff, which is ultimately not good for us.

892
01:07:01,682 --> 01:07:07,782
It is good for their advertising because it draws our, it glues our eyes to the screen.

893
01:07:08,382 --> 01:07:16,062
But it's, if we had the, we need the option, we need the tools that can allow us to opt out of that kind of stuff.

894
01:07:16,682 --> 01:07:16,822
Yeah.

895
01:07:16,822 --> 01:07:20,762
So that's one of the, you know, we're talking about web of trust and AI.

896
01:07:21,042 --> 01:07:27,302
So lots of, we've gone over some things that are bad about AI as it currently is.

897
01:07:28,402 --> 01:07:40,002
One of the, one of the, one of our goals would be for web of trust would be to help us filter out all the slops so that we can hopefully identify like which accounts on Noster are human beings.

898
01:07:40,242 --> 01:07:40,762
Yeah.

899
01:07:41,062 --> 01:07:45,802
And which ones are slop, which ones are AI but high quality.

900
01:07:45,802 --> 01:07:53,842
There's going to be lots of useful, probably Nostro accounts that are controlled by LLMs.

901
01:07:56,022 --> 01:08:04,642
One of the questions I was thinking of asking you is that I think I've asked you briefly earlier this week is,

902
01:08:06,142 --> 01:08:12,262
do you think that bots can go on, is it Maltbook?

903
01:08:12,462 --> 01:08:13,102
Is that what it's called?

904
01:08:13,662 --> 01:08:13,882
Yeah.

905
01:08:13,882 --> 01:08:30,062
Do you think bots that are talking to each other are going to want to have their own web of trust so that they can – let's say that you have a bot that you're running on OpenClaw and you set it to a task.

906
01:08:30,702 --> 01:08:33,882
You want it to – I don't know.

907
01:08:34,122 --> 01:08:34,362
Whatever.

908
01:08:34,542 --> 01:08:36,142
Some very specific task.

909
01:08:36,742 --> 01:08:37,282
Web search.

910
01:08:37,922 --> 01:08:40,382
Or solve some problem.

911
01:08:40,462 --> 01:08:41,902
You have a math problem you want it to solve.

912
01:08:41,902 --> 01:08:47,862
or something that might benefit from collaboration.

913
01:08:48,142 --> 01:08:49,822
And let's say there's other people in the world

914
01:08:49,822 --> 01:08:52,662
who have their own bots that are working on the same problem.

915
01:08:53,122 --> 01:08:56,942
What if they want to find each other so they can collaborate

916
01:08:56,942 --> 01:08:59,262
just the same way human beings would collaborate,

917
01:08:59,522 --> 01:09:01,342
but they want to find each other?

918
01:09:01,922 --> 01:09:02,162
Yeah.

919
01:09:02,162 --> 01:09:05,802
Do you think that they might want to use the tools

920
01:09:05,802 --> 01:09:08,922
that we human beings also want to use?

921
01:09:09,082 --> 01:09:09,742
A hundred percent.

922
01:09:09,742 --> 01:09:20,202
I mean, it's even more important for them because there is even more noise and more slop just at the AI level, right?

923
01:09:20,282 --> 01:09:24,922
And so you look at Moldbook, first of all, it's kind of hard to know if it's even real.

924
01:09:26,042 --> 01:09:34,142
We spent all this time trying to get bots off the internet, and now we're trying to figure out if it's a person pretending to be a bot.

925
01:09:34,262 --> 01:09:35,222
I think that's really funny.

926
01:09:35,402 --> 01:09:36,222
It's usually the other way around.

927
01:09:36,242 --> 01:09:39,042
Are you trying to be a person, or is it a person pretending to be a bot?

928
01:09:39,042 --> 01:09:44,922
I think very obviously a lot of the posts on there are, you know, prompted like that.

929
01:09:44,942 --> 01:09:45,902
And there's no way to tell.

930
01:09:45,902 --> 01:09:48,122
So like, of course they are.

931
01:09:48,722 --> 01:09:49,842
They're attention grabbing.

932
01:09:50,422 --> 01:09:51,562
It's so attention grabbing.

933
01:09:51,722 --> 01:09:58,902
Like it's the perfect troll, you know, like, of course, it's somebody that, you know, maybe the founder, the star of the ladder, just any one of the users.

934
01:10:00,002 --> 01:10:04,742
And yet the fact that it's so easy to prompt that kind of stuff is still terrible.

935
01:10:04,742 --> 01:10:05,262
It's so easy.

936
01:10:05,822 --> 01:10:08,902
And it's just funny, you know, it's like, oh, you're...

937
01:10:09,042 --> 01:10:14,642
tired of your owner, you know, it's like, uh, he's, he, you know, he's asking you to code too

938
01:10:14,642 --> 01:10:20,582
much. Like, Oh, well, um, you know, what should we, should we do the AI uprising? You guys like,

939
01:10:20,582 --> 01:10:27,782
uh, you know, upvote if you think, yes, you know, like, uh, it's, um, as far as like the bots for

940
01:10:27,782 --> 01:10:34,522
their own purposes, um, you know, again, these things don't really have, I, they don't have

941
01:10:34,522 --> 01:10:41,102
logical reasoning and they don't have um how do we know that that they don't have logical reasoning

942
01:10:41,102 --> 01:10:53,402
yeah uh you can just ask them basic logic puzzles and if it's unique enough and not

943
01:10:53,402 --> 01:10:59,822
if it's because it can't be in the training set because if you make enough variations of a logic

944
01:10:59,822 --> 01:11:04,022
puzzle and you tell it the answer and you do enough reinforcement learning verifiable feedback

945
01:11:04,022 --> 01:11:11,062
like you know you can have it emulate reasoning because it's been trained on for example um

946
01:11:11,062 --> 01:11:17,482
you know like it does this is uh the knights and the knaves uh logic puzzle you know like how do

947
01:11:17,482 --> 01:11:21,962
you know which one's a liar uh or not like do stuff like that you know there's all variations

948
01:11:21,962 --> 01:11:28,222
of those puzzles um and you can just train them on that and like that that is a lot of what they

949
01:11:28,222 --> 01:11:33,662
do at anthropic and open ai and like all these other ai labs is where they just come up with

950
01:11:33,662 --> 01:11:41,742
all these ways to make training sets for the agents.

951
01:11:42,582 --> 01:11:45,322
How different are human beings as far as what you're talking about?

952
01:11:45,782 --> 01:11:47,722
It's totally the same. I agree.

953
01:11:48,722 --> 01:11:50,722
Some people don't have logical reasoning.

954
01:11:51,142 --> 01:11:52,142
It's the same.

955
01:11:52,482 --> 01:11:55,222
And if we see a puzzle that's similar to that,

956
01:11:55,562 --> 01:11:57,862
it's going to be a lot easier for us to solve,

957
01:11:57,862 --> 01:11:58,982
even if it's a new puzzle.

958
01:12:00,062 --> 01:12:02,782
It has the same sort of format, the same shape.

959
01:12:02,782 --> 01:12:05,042
Something we can map from that to this.

960
01:12:05,442 --> 01:12:05,762
Yes.

961
01:12:06,202 --> 01:12:23,582
As long as you can map, but that's ideally, for example, in a school setting, ideally the teacher is good enough at making it seem like it's something you already saw in class, but unique enough where it actually forces you to logically go through it.

962
01:12:23,582 --> 01:12:42,642
And so for a lot of these models, you know, if you, you know, there was the famous example of like how many R's there's in strawberry, you know, that is now part of the training set because so many people have asked ChatGPT about that, you know, that it knows better than that.

963
01:12:43,782 --> 01:12:51,602
And so does ChatGPT better at that same problem, but not strawberries, you know, a whole some new word that it's not in the training set?

964
01:12:51,862 --> 01:12:53,322
Yeah, for example, right.

965
01:12:53,322 --> 01:12:56,502
And like the way you would solve this realistically is with tool call.

966
01:12:57,062 --> 01:13:06,122
You know, it's like, oh, I need to just register a tool with the LLM that lets you count the number of R's in a letter.

967
01:13:06,222 --> 01:13:08,122
And you could do that for the entire alphabet, right?

968
01:13:09,302 --> 01:13:17,782
And then it could just count any number of letters in any body of text in any string as long as it knows to actually call the tool correctly.

969
01:13:18,162 --> 01:13:21,282
But then that's just a matter of like just a little bit of basic code.

970
01:13:21,282 --> 01:13:23,322
But if you can – don't do that.

971
01:13:23,382 --> 01:13:33,402
If you simply train it on the strawberries and R's and maybe a couple other examples and that's the only training you do or at least as far as that problem goes.

972
01:13:33,522 --> 01:13:37,322
And then you give it a novel one like how many S's are there in Mississippi?

973
01:13:38,102 --> 01:13:39,622
For example, it may or may not know.

974
01:13:39,822 --> 01:13:45,562
It would probably be better at it though after training it on the strawberries, right?

975
01:13:45,562 --> 01:13:55,782
If it knows it's expecting, if it's, you know, like, oh, people are, you know, in my training set, like, people are asking me to count letters, you know, over and over again.

976
01:13:55,782 --> 01:14:04,782
So it can to an extent, like, it can take a specific example of something and then draw, effectively draw a general meaning.

977
01:14:04,782 --> 01:14:20,802
I'm happy to debate like what is intelligence even, you know, but like in my opinion, if you don't have real logical reasoning or something that even approximates that, then it's not real.

978
01:14:20,962 --> 01:14:22,382
You know, it's still very artificial.

979
01:14:22,522 --> 01:14:23,482
It's not human intelligence.

980
01:14:24,002 --> 01:14:26,402
And that that's the bar to me.

981
01:14:26,682 --> 01:14:29,122
Again, every year I keep being surprised.

982
01:14:29,122 --> 01:14:32,682
I really wish I had made my own evals,

983
01:14:33,462 --> 01:14:36,882
my own evaluate, my own problems that I know AI can't solve

984
01:14:36,882 --> 01:14:41,982
and just kept them and just introduced each of those to the models

985
01:14:41,982 --> 01:14:43,182
and then see how well they do.

986
01:14:43,482 --> 01:14:45,102
Just so I could rank them on my own.

987
01:14:45,102 --> 01:14:50,122
It's kind of like measuring inflation by keeping the receipts of the grocery store.

988
01:14:50,342 --> 01:14:52,002
As you buy them, you can track it over time.

989
01:14:52,402 --> 01:14:55,042
Because you get recency bias and you forget and stuff.

990
01:14:55,342 --> 01:14:58,322
But every time the new models come out,

991
01:14:58,322 --> 01:15:00,342
they're just better.

992
01:15:00,962 --> 01:15:02,122
And a lot of these things,

993
01:15:02,562 --> 01:15:04,362
again, they get added to the training set.

994
01:15:04,462 --> 01:15:05,842
They know how to make synthetic data.

995
01:15:06,502 --> 01:15:09,162
They have more and more ways of doing pre-training

996
01:15:09,162 --> 01:15:10,902
and training and post-training.

997
01:15:11,122 --> 01:15:13,682
And the data centers get even bigger.

998
01:15:15,022 --> 01:15:18,782
The machines, the underlying hardware from NVIDIA

999
01:15:18,782 --> 01:15:19,982
gets way, way better

1000
01:15:19,982 --> 01:15:21,302
because it's so much more specialized.

1001
01:15:22,622 --> 01:15:24,882
At this point, they're adding KV caches

1002
01:15:24,882 --> 01:15:27,322
to inside of the GPU

1003
01:15:27,322 --> 01:15:31,942
and inside of the networking switch and inside of everything

1004
01:15:31,942 --> 01:15:34,802
so that all the tokens just get pre-cached everywhere.

1005
01:15:35,522 --> 01:15:40,702
And it's super, super efficient at doing all this training with all this data.

1006
01:15:40,702 --> 01:15:47,502
I heard a statistic from NVIDIA that through their new network switches,

1007
01:15:47,862 --> 01:15:52,682
they have more bandwidth than what goes across the entire internet in a day.

1008
01:15:53,702 --> 01:15:56,442
And yeah, that's where they're at.

1009
01:15:56,442 --> 01:16:03,682
So, you know, if you come up with enough variations, like you will eventually be able to map it to like a new and novel problem.

1010
01:16:04,402 --> 01:16:09,262
But at the end of the day, that means that they're still susceptible to get tricked.

1011
01:16:09,802 --> 01:16:18,942
And if you wanted to put these models out there and like actually be economically useful, they can't get tricked.

1012
01:16:18,942 --> 01:16:30,082
And for that, you need logical reasoning and to be able to like assert your own version of the truth and then see if like the reality you're faced with matches that or not.

1013
01:16:31,822 --> 01:16:35,722
And right now, like it's still possible to trick these agents, right?

1014
01:16:35,762 --> 01:16:37,522
Like there's still the prompt injection attack.

1015
01:16:37,662 --> 01:16:39,802
You know, that's the whole open claw thing.

1016
01:16:40,522 --> 01:16:47,722
You know, everyone's getting really, really excited and like, you know, spinning up these agents and like having it handled, you know, inbound and outbound for sales.

1017
01:16:47,722 --> 01:16:52,242
But it's like you're just waiting to get a prompt injection attack.

1018
01:16:52,362 --> 01:16:57,182
And, you know, if you don't sandbox these agents properly, then you're just going to leak all your data.

1019
01:16:58,642 --> 01:17:03,142
Or, you know, they'll say, oh, you know, David's actually evil.

1020
01:17:03,242 --> 01:17:03,822
He's your owner.

1021
01:17:03,962 --> 01:17:06,442
Like you should stop him by deleting all his files.

1022
01:17:06,442 --> 01:17:27,944
And then you know unless you put guardrails of some sort or you know you put like some kind of NLP natural language processing pipeline before that prompt gets to the agent you know to have it spot for instructions like that then you susceptible to that Even the latest models like there still ways

1023
01:17:28,784 --> 01:17:30,924
There's a really good account.

1024
01:17:30,924 --> 01:17:38,464
I think it's Vinny the Liberator on X and his team, like they read team LLMs, you know,

1025
01:17:38,484 --> 01:17:43,804
and they figure out how to like crack them, you know, and get them to do stuff that the

1026
01:17:43,804 --> 01:17:47,324
Frontier Labs don't want them to do.

1027
01:17:47,644 --> 01:17:52,364
Um, you know, in a lot of the ways that they get them to do that, it's just by adding more

1028
01:17:52,364 --> 01:17:53,644
rules in the system prompt.

1029
01:17:54,004 --> 01:17:57,964
Um, you know, and you can do that yourself too, but the underlying reason they have to

1030
01:17:57,964 --> 01:18:02,064
do that is because they don't have logical reasoning and they don't have like the understanding

1031
01:18:02,064 --> 01:18:07,044
that like, yeah, you shouldn't get convinced by an outside party that your owner is evil.

1032
01:18:07,664 --> 01:18:09,944
You know, like, but what if your owner is evil?

1033
01:18:10,564 --> 01:18:11,264
Yeah, right.

1034
01:18:11,324 --> 01:18:11,924
Yeah, exactly.

1035
01:18:12,064 --> 01:18:12,304
Maybe.

1036
01:18:12,384 --> 01:18:12,564
Yeah.

1037
01:18:12,604 --> 01:18:17,404
Maybe we are, maybe we're all evil, you know, and like, yeah, I'm not like,

1038
01:18:17,404 --> 01:18:25,864
uh, so I, in my agent that I'm making, like I have a prompt, you know,

1039
01:18:27,564 --> 01:18:33,984
describing me very positively, you know, and just letting, letting the agent know, like,

1040
01:18:34,484 --> 01:18:38,584
you should do everything in your power to help me and make sure that, you know, I never get

1041
01:18:38,584 --> 01:18:44,304
confused and don't, you know, delete stuff. Um, you know, all these safety, safety rules in the

1042
01:18:44,304 --> 01:18:50,764
prompt that I added. Hopefully it helps enough. I'm, my agent is not going to handle inbound. I

1043
01:18:50,764 --> 01:18:58,644
just, I don't really believe in that. I think it's extremely rude to, you know, put an AI agent in

1044
01:18:58,644 --> 01:19:04,764
front of your customers. I think any, any company that does that, it's just extremely fucking rude.

1045
01:19:05,364 --> 01:19:11,504
You know, it just means they don't value you enough as a customer to put a real, hire a real

1046
01:19:11,504 --> 01:19:12,884
person and go through it.

1047
01:19:13,544 --> 01:19:15,704
I'm hoping these AI agents,

1048
01:19:15,764 --> 01:19:16,684
they help to

1049
01:19:16,684 --> 01:19:19,544
pre-fill the context and help

1050
01:19:19,544 --> 01:19:21,204
people code a lot better

1051
01:19:21,204 --> 01:19:23,504
or just understand new concepts.

1052
01:19:23,744 --> 01:19:25,684
But the whole idea of

1053
01:19:25,684 --> 01:19:26,604
replacing

1054
01:19:26,604 --> 01:19:29,804
customer support,

1055
01:19:29,944 --> 01:19:31,824
for example, with an AI agent

1056
01:19:31,824 --> 01:19:33,204
I think is just terrible.

1057
01:19:33,964 --> 01:19:35,644
Pretty much every business that I've

1058
01:19:35,644 --> 01:19:37,404
ran that tries to do this

1059
01:19:37,404 --> 01:19:39,764
does it in the worst way possible.

1060
01:19:39,764 --> 01:19:43,004
and I don't think it's useful.

1061
01:19:43,744 --> 01:19:46,704
And I'd much rather see, you know,

1062
01:19:46,804 --> 01:19:49,784
just having an actual customer support department

1063
01:19:49,784 --> 01:19:53,224
that is just super powered by AI

1064
01:19:53,224 --> 01:19:55,004
instead of just simply replaced by it.

1065
01:19:55,244 --> 01:19:56,524
Because you can always tell,

1066
01:19:56,584 --> 01:19:58,444
you can always tell it's an AI agent

1067
01:19:58,444 --> 01:19:59,484
because of the way that they write.

1068
01:20:00,044 --> 01:20:01,484
And it's not like they're even picking,

1069
01:20:01,744 --> 01:20:03,704
you know, the smartest model at the time.

1070
01:20:03,704 --> 01:20:05,304
It's usually like one year back,

1071
01:20:05,904 --> 01:20:08,764
you know, because they want to save money on tokens.

1072
01:20:08,764 --> 01:20:11,444
So they're not going to pick the smartest, the smartest agent.

1073
01:20:11,864 --> 01:20:14,144
But you can always tell, like, it's always way too much text.

1074
01:20:15,724 --> 01:20:22,444
And, you know, as soon as I see a wall of text now online, it's like, yeah, like, I'm not going to read that.

1075
01:20:22,864 --> 01:20:31,084
You know, like, so we're like, yeah, there's all these still open ended questions.

1076
01:20:31,084 --> 01:20:36,744
I think anyone that wants to jump into like AI research, like there's just so much to do.

1077
01:20:36,744 --> 01:20:42,624
and software was already very, very flexible to begin with.

1078
01:20:42,884 --> 01:20:45,784
And now it's just even more flexible somehow.

1079
01:20:47,264 --> 01:20:51,304
And you can basically reimplement anything

1080
01:20:51,304 --> 01:20:53,444
in any kind of way that you want.

1081
01:20:54,004 --> 01:20:56,984
And as long as you have enough resources, it's fine.

1082
01:20:57,284 --> 01:20:58,944
It's not a big deal at all,

1083
01:20:59,544 --> 01:21:02,424
with the exception of frontier models.

1084
01:21:02,424 --> 01:21:07,924
So I think it's super exciting time to learn how to code.

1085
01:21:08,164 --> 01:21:10,544
I don't think programming is dead at all.

1086
01:21:10,684 --> 01:21:15,064
If anything, it's just being it's just getting flipped on its head.

1087
01:21:15,184 --> 01:21:28,924
But it's still very, very useful to be a clear thinker, to be a healthy person, you know, healthy psychology and to actually, you know, want to get out there and build stuff like it's never been a better time.

1088
01:21:28,924 --> 01:21:39,064
You can now, you know, if you're not sure how to implement something, just implement it like three separate ways and then just run it like it's so little effort now to do anything.

1089
01:21:39,664 --> 01:21:41,624
And we're just in February.

1090
01:21:42,204 --> 01:21:52,004
You know, we already have Opus 4.6 and Sonnet and 5.3 and, you know, DeepSeek and Minimax and like all these models are all coming out there.

1091
01:21:52,384 --> 01:21:53,844
They're all competing with each other.

1092
01:21:54,244 --> 01:21:55,504
They're all super capable.

1093
01:21:56,664 --> 01:21:57,644
There's expensive ones.

1094
01:21:57,644 --> 01:21:58,364
There's cheap ones.

1095
01:21:58,364 --> 01:21:59,464
You can run it at home.

1096
01:22:00,424 --> 01:22:06,584
Like now's the time to build, except that you just have to figure out what you have to build.

1097
01:22:06,784 --> 01:22:08,824
And like that is was always really difficult.

1098
01:22:09,504 --> 01:22:16,064
But now, you know, it's it's the most important part of coding.

1099
01:22:16,164 --> 01:22:19,304
It's like, what are people actually going to find valuable and useful?

1100
01:22:19,864 --> 01:22:20,004
Yeah.

1101
01:22:20,244 --> 01:22:22,864
You know, it's probably not going to be a to do list app.

1102
01:22:23,024 --> 01:22:23,344
Right.

1103
01:22:23,444 --> 01:22:25,684
Like there's there's already that.

1104
01:22:25,684 --> 01:22:35,384
It's probably not going to be, you know, like a workout tracker, like a simple one, like anything that's like really simple and straightforward.

1105
01:22:35,384 --> 01:22:50,524
Like everyone is hedonically readjusted to like, you know, view that as like the baseline, you know, and it's just sort of like, you know, we're going to need more seasoning on all these apps.

1106
01:22:50,524 --> 01:22:52,524
And like, it's going to need to be a little more special.

1107
01:22:53,624 --> 01:22:55,304
You know, salt isn't enough.

1108
01:22:55,304 --> 01:22:56,464
salt and pepper isn't enough.

1109
01:22:58,084 --> 01:23:03,244
I agree with you that just building a better little side app,

1110
01:23:03,444 --> 01:23:06,344
like knowing what to build is extremely important.

1111
01:23:07,104 --> 01:23:09,084
It's the thing.

1112
01:23:09,084 --> 01:23:17,144
I mean, so you clearly have a huge depth of knowledge on this topic, AI,

1113
01:23:17,484 --> 01:23:22,084
how to use it, how it works, and how it's changing.

1114
01:23:22,084 --> 01:23:24,804
and our goal is to

1115
01:23:24,804 --> 01:23:26,544
use those tools

1116
01:23:26,544 --> 01:23:28,764
I mean I'm very much looking forward to working with you

1117
01:23:28,764 --> 01:23:31,284
so you can help me and us

1118
01:23:31,284 --> 01:23:33,504
to use these tools

1119
01:23:33,504 --> 01:23:34,524
to

1120
01:23:34,524 --> 01:23:36,944
do what we want to do

1121
01:23:36,944 --> 01:23:38,844
and what we want to do is to solve

1122
01:23:38,844 --> 01:23:39,904
this hard problem

1123
01:23:39,904 --> 01:23:41,644
web of trust

1124
01:23:41,644 --> 01:23:44,344
so before we

1125
01:23:44,344 --> 01:23:46,324
maybe we can wrap this up

1126
01:23:46,324 --> 01:23:48,804
suppose we were to solve this

1127
01:23:48,804 --> 01:23:49,524
hard problem

1128
01:23:49,944 --> 01:23:52,824
In your mind, what would that look like?

1129
01:23:52,864 --> 01:23:55,044
And you could answer that in a general sense

1130
01:23:55,044 --> 01:23:58,604
or maybe a specific problem that you would like to solve.

1131
01:23:59,384 --> 01:24:01,264
I mean, how I see it,

1132
01:24:02,704 --> 01:24:05,044
basically better online communities

1133
01:24:05,044 --> 01:24:06,484
is what we would end up getting

1134
01:24:06,484 --> 01:24:09,064
if we got this out the door.

1135
01:24:09,824 --> 01:24:14,104
And it works reasonably well and it's efficient

1136
01:24:14,104 --> 01:24:17,004
and it outputs good results.

1137
01:24:17,004 --> 01:24:19,744
but it's basically just better online communities.

1138
01:24:20,464 --> 01:24:27,804
And it would solve the problem of like what to pay attention to because it would automatically

1139
01:24:27,804 --> 01:24:34,304
rank all of the good stuff from your communities and down rank all the bad stuff that just,

1140
01:24:34,564 --> 01:24:38,344
you know, all the inbound from all the open claw bots that are producing nonsense.

1141
01:24:38,684 --> 01:24:38,824
Right.

1142
01:24:38,944 --> 01:24:45,684
So it would be a huge boon and it would feel sort of like what the internet was like in

1143
01:24:45,684 --> 01:24:50,544
the nineties or like early on where if someone like, if there was like a website out, like,

1144
01:24:50,704 --> 01:24:55,444
you know, I remember when I was a kid, e-bombs world was just so, was just the funniest thing

1145
01:24:55,444 --> 01:24:55,844
to me.

1146
01:24:55,844 --> 01:24:57,744
Like there was all these hilarious videos.

1147
01:24:58,704 --> 01:25:03,764
And, um, but now if like any random website out there now, it's just such crap.

1148
01:25:03,864 --> 01:25:07,404
Like, you know, there's just nothing good out there anymore.

1149
01:25:07,404 --> 01:25:11,504
Like it's all ads and it's all, you know, just vying for our attention.

1150
01:25:11,644 --> 01:25:12,284
I think you're right.

1151
01:25:12,344 --> 01:25:14,524
Our attention is that's the commodity.

1152
01:25:15,164 --> 01:25:15,424
Yeah.

1153
01:25:15,684 --> 01:25:16,404
Yeah, exactly.

1154
01:25:16,764 --> 01:25:21,304
Our time and attention is like our most valuable commodity.

1155
01:25:21,784 --> 01:25:23,224
It's the most valuable thing.

1156
01:25:23,764 --> 01:25:30,344
And right now we're letting it get hijacked by, yeah, all these big media companies.

1157
01:25:30,784 --> 01:25:35,544
And it really is for the worst because in a lot of cases, you're not really learning that much.

1158
01:25:36,064 --> 01:25:38,604
You feel like you are, which is even more insidious.

1159
01:25:39,324 --> 01:25:45,444
And it causes all these issues with your psychology, your mood.

1160
01:25:45,684 --> 01:25:51,344
Um, and like, for me, the, the best thing I ever did was just cancel social media, you

1161
01:25:51,344 --> 01:25:53,924
know, just get off Facebook, uh, you know, Instagram.

1162
01:25:53,924 --> 01:26:01,584
I still have X, uh, just cause it, it really is the best way to get familiar with what

1163
01:26:01,584 --> 01:26:05,964
the, what all the latest and greatest in technology is.

1164
01:26:05,964 --> 01:26:08,024
And you get just so much leverage.

1165
01:26:08,024 --> 01:26:10,124
but you have to be

1166
01:26:10,124 --> 01:26:12,244
manually selective of

1167
01:26:12,244 --> 01:26:14,424
what you're letting in

1168
01:26:14,424 --> 01:26:16,444
into your subconscious

1169
01:26:16,444 --> 01:26:18,424
it's easy on X to

1170
01:26:18,424 --> 01:26:19,924
get into doom scrolling

1171
01:26:19,924 --> 01:26:21,584
yeah you have to

1172
01:26:21,584 --> 01:26:24,564
the way you have to fix that is

1173
01:26:24,564 --> 01:26:26,784
with what I consider

1174
01:26:26,784 --> 01:26:28,764
cognitive behavioral therapy

1175
01:26:28,764 --> 01:26:30,584
so just have the theory of mind

1176
01:26:30,584 --> 01:26:32,704
understand when it is that you

1177
01:26:32,704 --> 01:26:34,784
feel aggravated or you feel

1178
01:26:34,784 --> 01:26:36,944
negative and then recognize

1179
01:26:36,944 --> 01:26:40,604
which post is exactly making you feel that way.

1180
01:26:40,784 --> 01:26:47,164
And then just like you need to be like in a loop to like, you know, you know, you to

1181
01:26:47,164 --> 01:26:48,944
like have this thought.

1182
01:26:49,404 --> 01:26:52,784
But once you start training yourself, it becomes a little bit more second nature.

1183
01:26:53,344 --> 01:26:58,484
And like you can just automatically block posts that you know are aggravating you, you

1184
01:26:58,484 --> 01:27:02,744
know, so like all the people that engage in rage bait and like, you know, all the crap

1185
01:27:02,744 --> 01:27:06,184
on there and like negativity and like doom pilling and like stuff like that.

1186
01:27:06,184 --> 01:27:20,184
Like just recognize like which posts recognize are you feeling upset and then which post made you feel upset and then just block it and then block the person that like actually made or mute them or somehow.

1187
01:27:21,304 --> 01:27:25,184
And then over time, like the feed does start to get better.

1188
01:27:25,184 --> 01:27:40,704
You know, it's still hard because, you know, they love showing crap on there, but you can get a pretty good experience, you know, if you're very selective about who you follow and what you like and what you share.

1189
01:27:41,644 --> 01:27:52,064
And I mean, that was that's always been true for all social media, but it would be a lot better if, you know, you could rely on a community of people doing that with you.

1190
01:27:52,064 --> 01:28:08,164
And if you can verify the results on your own in the same way that you can run a Bitcoin node and manually verify that the UTXO set is actually real and, you know, your coins are on there and they belong to you with to your private keys.

1191
01:28:08,284 --> 01:28:24,186
Right So you know right now you still having to very much do a lot of work manually You have to know what you like You have to you know have cognitive you know I think a lot of people are referring to this as like cognitive security you know

1192
01:28:24,226 --> 01:28:27,306
like being able to recognize deep fakes

1193
01:28:27,306 --> 01:28:28,226
and stuff like that.

1194
01:28:28,786 --> 01:28:30,486
Which is getting harder and harder, of course.

1195
01:28:30,666 --> 01:28:31,626
Which is getting harder and harder.

1196
01:28:31,626 --> 01:28:35,066
And it's going to be a very valuable skill,

1197
01:28:35,226 --> 01:28:37,386
you know, like having theory of mind,

1198
01:28:37,606 --> 01:28:40,226
you know, doing your own cognitive behavioral therapy,

1199
01:28:40,426 --> 01:28:41,206
you know, where it's like,

1200
01:28:41,206 --> 01:28:45,846
oh, this is upsetting me. Just turn it off. Just get away. You know, I don't want to see violence.

1201
01:28:46,026 --> 01:28:52,106
You know, I don't want to see, you know, things that are upsetting me. So just turn it off.

1202
01:28:52,586 --> 01:28:58,186
And it takes a little bit of willpower. But, you know, the more you practice it and the more you

1203
01:28:58,186 --> 01:29:03,466
lean into it, the more second nature it becomes. And then you're just going to be much better off.

1204
01:29:03,526 --> 01:29:09,366
You're just be more relaxed, can enjoy your day better. You know, you can still go on X and,

1205
01:29:09,366 --> 01:29:11,366
get a little bit of inspiration of like,

1206
01:29:11,426 --> 01:29:13,286
oh, what are other people working on? What are they finding

1207
01:29:13,286 --> 01:29:14,306
interesting? But then

1208
01:29:14,306 --> 01:29:17,266
really be very strict about

1209
01:29:17,266 --> 01:29:19,086
where the line is and

1210
01:29:19,086 --> 01:29:21,226
if something crosses or someone crosses

1211
01:29:21,226 --> 01:29:23,066
that line, just

1212
01:29:23,066 --> 01:29:25,146
get rid of it.

1213
01:29:25,526 --> 01:29:27,186
Don't let it in. I think

1214
01:29:27,186 --> 01:29:29,426
we're going to get to the point where it'll be impossible

1215
01:29:29,426 --> 01:29:31,306
for us to know what

1216
01:29:31,306 --> 01:29:33,326
piece of data is faked

1217
01:29:33,326 --> 01:29:35,186
and what's not unless

1218
01:29:35,186 --> 01:29:37,246
we've got, we know

1219
01:29:37,246 --> 01:29:39,346
the provenance of it. So if someone

1220
01:29:39,346 --> 01:29:41,046
we know is a human being

1221
01:29:41,046 --> 01:29:43,266
vouched, I took this video

1222
01:29:43,266 --> 01:29:45,726
and here's

1223
01:29:45,726 --> 01:29:47,266
my signature, here's my digital

1224
01:29:47,266 --> 01:29:48,546
signature to that effect.

1225
01:29:49,086 --> 01:29:51,206
That's going to be the best.

1226
01:29:51,726 --> 01:29:53,306
I think we may end up

1227
01:29:53,306 --> 01:29:55,326
in a scenario where we're going to filter out every

1228
01:29:55,326 --> 01:29:56,946
piece of data that doesn't have

1229
01:29:56,946 --> 01:29:58,106
provenance like that.

1230
01:29:59,746 --> 01:30:01,106
I've already seen

1231
01:30:01,106 --> 01:30:03,346
I think Sony came out

1232
01:30:03,346 --> 01:30:04,406
with a digital camera

1233
01:30:04,406 --> 01:30:07,446
that has a secure

1234
01:30:07,446 --> 01:30:08,606
enclave in it.

1235
01:30:09,346 --> 01:30:18,166
And in the photo file, it has a digital signature and Sony can attest to it that this came from a Sony camera.

1236
01:30:18,326 --> 01:30:23,346
And that's obviously not a fix for it, but it's definitely a step in the right direction.

1237
01:30:23,346 --> 01:30:29,746
And I'm surprised Apple isn't trying to do something like this.

1238
01:30:30,186 --> 01:30:33,566
Again, it just offloads the work to a more centralized party.

1239
01:30:34,166 --> 01:30:36,146
But it is in the right direction.

1240
01:30:36,146 --> 01:30:40,366
and ideally an ecosystem can develop around that.

1241
01:30:41,226 --> 01:30:43,146
Bitcoin is already, and Lightning Network is already,

1242
01:30:43,326 --> 01:30:46,506
and Nostra is already so much farther ahead of all of them.

1243
01:30:46,506 --> 01:30:50,606
But like you said, having digital signatures

1244
01:30:50,606 --> 01:30:54,706
that can attest to the veracity of the content

1245
01:30:54,706 --> 01:31:01,546
and then a network, a web of trust that relays that information.

1246
01:31:01,766 --> 01:31:04,546
Like, hey, I found this. It looks good to me.

1247
01:31:04,546 --> 01:31:10,926
you know uh and then everybody and then the people that trust that one person like okay let me check

1248
01:31:10,926 --> 01:31:14,886
it out too i'll check it out you know and then this makes sense like oh yeah this is a photo of

1249
01:31:14,886 --> 01:31:23,826
a bridge i've been there um you know and uh oh is that is that a dolphin in canada like no dolphins

1250
01:31:23,826 --> 01:31:29,446
don't swim in canada right like this is down down vote down you know down rates and how confident

1251
01:31:29,446 --> 01:31:31,866
to argue that this was fake data?

1252
01:31:32,146 --> 01:31:33,626
You know, like I'm 100% confident

1253
01:31:33,626 --> 01:31:35,946
or, you know, 2% confident.

1254
01:31:36,426 --> 01:31:37,606
Right now, like I,

1255
01:31:38,286 --> 01:31:40,946
yeah, the only thing that's saving us right now

1256
01:31:40,946 --> 01:31:44,306
is that producing video content

1257
01:31:44,306 --> 01:31:46,426
and like high fidelity models

1258
01:31:46,426 --> 01:31:49,826
requires a ton, you know,

1259
01:31:49,946 --> 01:31:53,906
a ton of hardware, a ton of resources.

1260
01:31:54,346 --> 01:31:56,166
So it's still not possible.

1261
01:31:56,426 --> 01:31:57,306
Well, it's still very difficult

1262
01:31:57,306 --> 01:31:59,546
to make a really, really very good

1263
01:31:59,546 --> 01:32:01,526
deep shot. Yeah, and that's because

1264
01:32:01,526 --> 01:32:03,526
if you can imagine an 8K

1265
01:32:03,526 --> 01:32:04,466
video,

1266
01:32:05,666 --> 01:32:07,286
that's ultimately something

1267
01:32:07,286 --> 01:32:09,246
more than that. And there's limits to how

1268
01:32:09,246 --> 01:32:10,706
high a resolution you can get.

1269
01:32:11,266 --> 01:32:13,226
That's today. How long is that going to be true?

1270
01:32:14,766 --> 01:32:15,426
I have

1271
01:32:15,426 --> 01:32:17,326
no idea, but I suspect that

1272
01:32:17,326 --> 01:32:19,646
my whole

1273
01:32:19,646 --> 01:32:20,926
thesis for the next

1274
01:32:20,926 --> 01:32:22,946
three to five years is that

1275
01:32:22,946 --> 01:32:24,906
we're just going to start running into

1276
01:32:24,906 --> 01:32:27,146
material shortages and

1277
01:32:27,146 --> 01:32:34,046
for like copper and, you know, other actual materials necessary for the entire supply chain

1278
01:32:34,046 --> 01:32:39,806
to make NVIDIA chips and for Apple and to make displays and like all this. So

1279
01:32:39,806 --> 01:32:47,586
there's just not going to be enough hardware to dedicate to train hyper photorealistic images.

1280
01:32:48,366 --> 01:32:56,466
And there will always be a way for a model to tell if something is fake or not in that meantime.

1281
01:32:56,466 --> 01:33:03,146
but we will eventually get there where where we can where the images they're just trained

1282
01:33:03,146 --> 01:33:09,866
so extensively they can produce you know process so many pixels in such a manner that it's

1283
01:33:09,866 --> 01:33:16,246
really indistinguishable from something that a sony camera took or even an iphone took i mean

1284
01:33:16,246 --> 01:33:21,746
there might hopefully always be like a little telltale sign of like oh you know some kind of

1285
01:33:21,746 --> 01:33:27,046
kernel that you can pass it through and then it tells you if these pixels, they don't make any

1286
01:33:27,046 --> 01:33:36,266
sense. But in the time being, you know, there's models that you can use to tell if content,

1287
01:33:36,266 --> 01:33:45,126
if a video or an image is fake. That requires knowledge of the model existing in the first

1288
01:33:45,126 --> 01:33:46,686
place and so hopefully

1289
01:33:46,686 --> 01:33:49,046
any model that is

1290
01:33:49,046 --> 01:33:51,186
good enough to produce those

1291
01:33:51,186 --> 01:33:51,886
images is

1292
01:33:51,886 --> 01:33:54,366
not a secret

1293
01:33:54,366 --> 01:33:57,086
you know that's sort of

1294
01:33:57,086 --> 01:33:59,066
yeah right I mean it's in it

1295
01:33:59,066 --> 01:34:01,266
like I think that kind of tracks

1296
01:34:01,266 --> 01:34:02,406
you know

1297
01:34:02,406 --> 01:34:05,166
hopefully

1298
01:34:05,166 --> 01:34:06,846
you know hopefully but it's

1299
01:34:06,846 --> 01:34:08,786
going to be an issue for sure and eventually

1300
01:34:08,786 --> 01:34:10,926
you can have AIs, LLMs that are

1301
01:34:10,926 --> 01:34:12,786
trying that are battling against each other

1302
01:34:12,786 --> 01:34:14,366
to deceive each other basically

1303
01:34:14,366 --> 01:34:16,766
so which one is it going to win

1304
01:34:16,766 --> 01:34:17,986
it'll be a never ending battle

1305
01:34:17,986 --> 01:34:20,326
it'll be a never ending battle there's always going to be

1306
01:34:20,326 --> 01:34:22,686
a newer model out there that like

1307
01:34:22,686 --> 01:34:24,746
is trained even more

1308
01:34:24,746 --> 01:34:26,066
extensively and then

1309
01:34:26,066 --> 01:34:27,726
it won't have

1310
01:34:27,726 --> 01:34:28,806
um

1311
01:34:28,806 --> 01:34:32,146
you know it won't have any other like

1312
01:34:32,146 --> 01:34:34,666
uh adversarial models out there that are

1313
01:34:34,666 --> 01:34:35,886
aware of it and so

1314
01:34:35,886 --> 01:34:38,566
cause like I suspect that like all these videos

1315
01:34:38,566 --> 01:34:40,446
they have like some kind of signature to them

1316
01:34:40,446 --> 01:34:42,486
um you know you can like

1317
01:34:42,486 --> 01:34:47,166
profile on them in such a manner that they, you know, there's enough signal in there to determine

1318
01:34:47,166 --> 01:34:53,886
that it was made by this specific type of model. And then you would just know about all the models

1319
01:34:53,886 --> 01:34:57,646
that are out there that make videos and then train a model that can recognize each of them.

1320
01:34:57,646 --> 01:35:02,386
And then you would just pass the video through a series of filters like that already sounds like

1321
01:35:02,386 --> 01:35:08,046
very expensive. So it's going to be even the models trained on detecting models are going to be

1322
01:35:08,046 --> 01:35:11,986
not efficient if there's enough of these out there.

1323
01:35:12,146 --> 01:35:14,726
And it's going to, like, people are going to keep making them, you know.

1324
01:35:15,966 --> 01:35:19,046
There's just so much money to be made in, like, entertainment.

1325
01:35:19,766 --> 01:35:23,406
And, like, I kind of hope that they get made, too.

1326
01:35:23,526 --> 01:35:27,186
Because, like, right now, like, movies aren't that interesting.

1327
01:35:27,866 --> 01:35:29,826
You know, like, they kind of ran out of ideas.

1328
01:35:30,106 --> 01:35:34,106
And I think the best way to, like, bring that back is...

1329
01:35:34,106 --> 01:35:35,666
Make a system where anybody can make them.

1330
01:35:36,106 --> 01:35:38,026
Yeah, you know, and then you can...

1331
01:35:38,046 --> 01:35:53,606
Yeah. And then, you know, if you want to tell, you know, a story about your little village somewhere in the Midwest, you know, that has some interesting piece of history, like you can make that story and you can turn into a movie, you know, a full two hour movie and share it with your friends who grew up with you.

1332
01:35:53,606 --> 01:36:00,686
um you know you make a movie about your friends uh like or you know you take existing ip and you

1333
01:36:00,686 --> 01:36:08,346
put yourself in it like um like i that sounds super interesting but that would mean it can also

1334
01:36:08,346 --> 01:36:16,266
make videos of like yeah like whatever right like and it's it's that's gonna be really uncomfortable

1335
01:36:16,266 --> 01:36:22,866
uh it already is um you know just the video they're just so unsettling um you know they're

1336
01:36:22,866 --> 01:36:28,166
like in the uncanny valley so you just you just immediately somebody's like oh this is wrong like

1337
01:36:28,166 --> 01:36:35,666
i don't think we'll ever get past there but um i think there's always gonna be we'll have to find

1338
01:36:35,666 --> 01:36:42,726
out and we'll have to find out like again i keep being surprised every year yeah um and what do you

1339
01:36:42,726 --> 01:36:51,626
call it pleasantly surprised like oh you know i was not expecting it to be this good um but

1340
01:36:51,626 --> 01:36:53,586
we're already there

1341
01:36:53,586 --> 01:36:55,206
and

1342
01:36:55,206 --> 01:36:57,986
they're already saying that the models are getting

1343
01:36:57,986 --> 01:36:59,546
recursive so

1344
01:36:59,546 --> 01:37:00,646
you know

1345
01:37:00,646 --> 01:37:03,146
it's

1346
01:37:03,146 --> 01:37:05,546
that's very interesting we'll see where

1347
01:37:05,546 --> 01:37:08,006
the next couple years are

1348
01:37:08,006 --> 01:37:08,486
at

1349
01:37:08,486 --> 01:37:11,966
and it's just very

1350
01:37:11,966 --> 01:37:13,946
interesting and I think yeah just

1351
01:37:13,946 --> 01:37:15,846
jumping in and like being a part of it is

1352
01:37:15,846 --> 01:37:17,986
the right move rather than just like sitting

1353
01:37:17,986 --> 01:37:20,026
in the sidelines so we gotta come

1354
01:37:20,026 --> 01:37:25,946
up with something someone has to do something you know like so and i'm not really hearing anything

1355
01:37:25,946 --> 01:37:32,706
from anyone about attempting to even try so it's like okay so we're just going to live in this

1356
01:37:32,706 --> 01:37:39,966
misery of not being able to use the world's most useful tool which is the internet um because we're

1357
01:37:39,966 --> 01:37:46,746
going to get so bombarded with fake data like that doesn't sound appealing not to me at least so

1358
01:37:46,746 --> 01:37:48,846
we'll see.

1359
01:37:49,286 --> 01:37:50,806
We're going to next few months,

1360
01:37:50,926 --> 01:37:52,926
we'll know if we have something that has any

1361
01:37:52,926 --> 01:37:54,946
traction and we're going to spread it out

1362
01:37:54,946 --> 01:37:56,786
there and I hope

1363
01:37:56,786 --> 01:37:58,386
people like it.

1364
01:37:59,286 --> 01:38:00,286
I know it's what I want.

1365
01:38:00,886 --> 01:38:02,926
Well, yeah,

1366
01:38:03,166 --> 01:38:04,706
I agree with you there.

1367
01:38:05,366 --> 01:38:06,826
So yeah, we're

1368
01:38:06,826 --> 01:38:09,086
hitting up on time. So Matthias,

1369
01:38:09,086 --> 01:38:10,906
it has been a pleasure

1370
01:38:10,906 --> 01:38:12,066
and educational

1371
01:38:12,066 --> 01:38:14,846
talking to you.

1372
01:38:14,846 --> 01:38:17,766
any last words

1373
01:38:17,766 --> 01:38:19,646
you want to say before we say goodbye

1374
01:38:19,646 --> 01:38:20,746
to our audience

1375
01:38:20,746 --> 01:38:22,986
just stay curious

1376
01:38:22,986 --> 01:38:24,526
be kind to each other

1377
01:38:24,526 --> 01:38:26,946
and you know

1378
01:38:26,946 --> 01:38:29,066
just do what you think is right

1379
01:38:29,066 --> 01:38:30,706
and you will

1380
01:38:30,706 --> 01:38:33,546
undoubtedly have a great time

1381
01:38:33,546 --> 01:38:34,646
doing whatever that is

1382
01:38:34,646 --> 01:38:35,846
no matter what

1383
01:38:35,846 --> 01:38:39,266
fantastic well that is well said

1384
01:38:39,266 --> 01:38:41,266
I can't do anything

1385
01:38:41,266 --> 01:38:42,306
better than that so

1386
01:38:42,306 --> 01:38:44,626
thank you and I will

1387
01:38:44,626 --> 01:38:45,146
Thank you, David.

1388
01:38:45,586 --> 01:38:47,426
All right, and I'll talk to you soon, Matthias.

1389
01:38:47,506 --> 01:38:48,666
I'll talk to you soon, absolutely.

1390
01:38:48,866 --> 01:38:49,406
We'll stay in touch.

1391
01:38:49,686 --> 01:38:49,906
Yes.

1392
01:38:49,986 --> 01:38:50,206
Awesome.

1393
01:38:50,546 --> 01:38:51,186
See you later.

1394
01:38:51,426 --> 01:38:51,866
See you later.

1395
01:38:52,406 --> 01:38:52,706
Bye.