1
00:00:00,000 --> 00:00:02,000
All right, looks good. Let's go.

2
00:00:02,000 --> 00:00:04,000
Here we go. All right.

3
00:00:04,000 --> 00:00:08,000
All right. So I was basically going, who knows when I'll see you again?

4
00:00:08,000 --> 00:00:10,000
Might make the best of it.

5
00:00:10,000 --> 00:00:11,000
Yeah, I appreciate it.

6
00:00:11,000 --> 00:00:14,000
Every time I see Cali, I want to talk cashew.

7
00:00:14,000 --> 00:00:18,000
So talk cashew to me, baby. Talk cashew to me.

8
00:00:18,000 --> 00:00:20,000
I'll share the nuts.

9
00:00:20,000 --> 00:00:22,000
Sure, let's talk some nuts.

10
00:00:23,000 --> 00:00:26,000
So, and kind of, I have a specific question.

11
00:00:26,000 --> 00:00:34,000
I saw you, I saw you on, like with pen and paper, drawing Merkle trees.

12
00:00:34,000 --> 00:00:35,000
Oh, yeah.

13
00:00:35,000 --> 00:00:36,000
Trying to solve a problem.

14
00:00:36,000 --> 00:00:37,000
Yes.

15
00:00:37,000 --> 00:00:40,000
Tell me about the problem and tell me about what you have in mind.

16
00:00:40,000 --> 00:00:41,000
Okay, you want to go in real deep?

17
00:00:41,000 --> 00:00:42,000
Yes, let's start deep.

18
00:00:42,000 --> 00:00:44,000
Okay, so I'll start with...

19
00:00:44,000 --> 00:00:45,000
Fuck the noobs.

20
00:00:45,000 --> 00:00:46,000
Fuck the noobs, all right.

21
00:00:46,000 --> 00:00:49,000
Well, I'll try, no, no, no, I'll try to get the noobs in.

22
00:00:49,000 --> 00:00:52,000
So, simple concept.

23
00:00:52,000 --> 00:00:58,000
The way a cashew mint works is it issues e-cash tokens as blind signatures,

24
00:00:58,000 --> 00:01:01,000
and then it gives out to the users,

25
00:01:01,000 --> 00:01:04,000
and then those tokens are at some point spent back to the mint.

26
00:01:04,000 --> 00:01:08,000
And so the mint has kind of, let's say, one major list,

27
00:01:08,000 --> 00:01:11,000
and those are the spent e-cash tokens.

28
00:01:11,000 --> 00:01:13,000
Typically, this is a normal database.

29
00:01:13,000 --> 00:01:14,000
It's like a table.

30
00:01:14,000 --> 00:01:17,000
You put every single token into a table.

31
00:01:17,000 --> 00:01:21,000
And another way to organize the list instead of an ordinary table,

32
00:01:21,000 --> 00:01:24,000
you can also organize it as a sparse Merkle tree.

33
00:01:24,000 --> 00:01:29,000
And a sparse Merkle tree is basically, you know, you have a Merkle tree.

34
00:01:29,000 --> 00:01:34,000
Let's say you have all the possible e-cash tokens in the world, like 32 byte, right?

35
00:01:34,000 --> 00:01:36,000
It's like a massive space.

36
00:01:36,000 --> 00:01:43,000
And imagine them all in a line, starting from 0, 0, 0, 0 up until like 1, 1, 1, 1, 1, 1, 1,

37
00:01:43,000 --> 00:01:46,000
the 256-bit full ones.

38
00:01:46,000 --> 00:01:47,000
So this is all in a line.

39
00:01:47,000 --> 00:01:53,000
And if you organize them all in a line, then you can start to build a Merkle tree out of them.

40
00:01:53,000 --> 00:01:55,000
So it's a one-dimensional line in the bottom,

41
00:01:55,000 --> 00:01:59,000
and then you go up and up and up and up, and that makes a Merkle tree.

42
00:01:59,000 --> 00:02:05,000
And at the end of the day, you get one Merkle root, which is the top root of the Merkle tree.

43
00:02:05,000 --> 00:02:10,000
So you can think of the root of the Merkle tree, of the sparse Merkle tree,

44
00:02:10,000 --> 00:02:15,000
of all the spent e-cash tokens a mint has ever seen as like the state of the mint.

45
00:02:15,000 --> 00:02:18,000
So the Merkle root, the root hash at the top,

46
00:02:18,000 --> 00:02:24,000
says which tokens have ever been spent in that mint encoded in a single hash.

47
00:02:24,000 --> 00:02:25,000
Yeah.

48
00:02:25,000 --> 00:02:32,000
So once you do that, you have like a super compact representation of what the database of the mint looks like.

49
00:02:32,000 --> 00:02:33,000
The state of the mint.

50
00:02:33,000 --> 00:02:34,000
The state of the mint.

51
00:02:34,000 --> 00:02:39,000
Like I say the state of the mint because that's the main thing that evolves over time.

52
00:02:39,000 --> 00:02:40,000
Yeah.

53
00:02:40,000 --> 00:02:43,000
From time now to time now, you spend a token.

54
00:02:43,000 --> 00:02:48,000
I add it to the database, which I would insert in the bottom of this tree,

55
00:02:48,000 --> 00:02:52,000
and that changes all the hashes on the top, and then the Merkle root at the top changes,

56
00:02:52,000 --> 00:02:54,000
and that's the next state of the mint.

57
00:02:54,000 --> 00:03:00,000
Yeah. Maybe explain the minting spending thing.

58
00:03:00,000 --> 00:03:03,000
Like you wanted to say, let's bring in the noobs.

59
00:03:03,000 --> 00:03:04,000
Okay.

60
00:03:06,000 --> 00:03:08,000
So let's maybe take three steps back.

61
00:03:08,000 --> 00:03:09,000
Okay.

62
00:03:09,000 --> 00:03:16,000
Usually a mint has a lightning gateway, and the idea is to have e-cash on top of Bitcoin,

63
00:03:16,000 --> 00:03:22,000
and with Cashew, the awesome thing about Cashew, not only can you send Cashew tokens around,

64
00:03:22,000 --> 00:03:29,000
you also can just pay regular lightning invoices and the blinded signatures that you mentioned.

65
00:03:29,000 --> 00:03:34,000
The awesome thing about that is the guy who runs the mint, which is basically the banker,

66
00:03:34,000 --> 00:03:36,000
doesn't see anything.

67
00:03:36,000 --> 00:03:40,000
Like who pays whom and who holds how much money, he doesn't see.

68
00:03:40,000 --> 00:03:42,000
That's what blinded signatures brings.

69
00:03:42,000 --> 00:03:47,000
So it achieves that basically through cryptography, but what you end up,

70
00:03:47,000 --> 00:03:54,000
as you already started explaining, is the simple idea is you send Bitcoin via lightning to the mint,

71
00:03:54,000 --> 00:03:56,000
and the mint issues you e-cash tokens.

72
00:03:56,000 --> 00:04:03,000
And if you want to go into that process a little bit more, the real process is I send you Bitcoin as a user,

73
00:04:03,000 --> 00:04:08,000
I send the mint Bitcoin, and with that, I generate a random secret.

74
00:04:08,000 --> 00:04:11,000
It's called the e-cash secret, and I blind it.

75
00:04:11,000 --> 00:04:13,000
So it's kind of like an encryption.

76
00:04:13,000 --> 00:04:19,000
I blind it with an encryption key that only I know as a user, and I give that encrypted secret to you.

77
00:04:19,000 --> 00:04:25,000
And once you receive the Bitcoin on lightning, the mint, it says, okay, I received the Bitcoin,

78
00:04:25,000 --> 00:04:30,000
and then I sign that encrypted secret I sent to you, and I send it back.

79
00:04:30,000 --> 00:04:33,000
We call that a blinded signature.

80
00:04:33,000 --> 00:04:37,000
And the last step is now that I receive the blinded signature back as a user,

81
00:04:37,000 --> 00:04:44,000
I unblind it with that encryption secret from the first step, and now I get an unblinded signature,

82
00:04:44,000 --> 00:04:46,000
which is that e-cash token itself.

83
00:04:46,000 --> 00:04:49,000
So that thing is a token, and I can now carry that.

84
00:04:49,000 --> 00:04:54,000
The mint doesn't know what that secret looks like, that this unencrypted signature looks like.

85
00:04:54,000 --> 00:04:59,000
It has never seen it before, but the mint will be able to know in the future when I want to spend it again,

86
00:04:59,000 --> 00:05:04,000
so I can come back to the mint and say, here's a blind signature, here's a secret,

87
00:05:04,000 --> 00:05:09,000
look at the signature you made, that signature mint, and the mint can verify, yes, this was me,

88
00:05:09,000 --> 00:05:15,000
but I don't remember which one of those users you were when you first created it.

89
00:05:15,000 --> 00:05:19,000
So at that point, when I come back to the mint and want to spend that token,

90
00:05:19,000 --> 00:05:27,000
that's when the mint needs to insert that token into the database to prevent the same token from being spent again.

91
00:05:27,000 --> 00:05:33,000
So the main goal or the main job of the mint is essentially preventing double spending

92
00:05:33,000 --> 00:05:35,000
by remembering which tokens have been spent.

93
00:05:35,000 --> 00:05:36,000
Keeping track of these tokens.

94
00:05:36,000 --> 00:05:37,000
Right.

95
00:05:37,000 --> 00:05:39,000
And so that lands on mint state.

96
00:05:39,000 --> 00:05:40,000
Yes.

97
00:05:40,000 --> 00:05:46,000
Mint state is the issue you were trying to solve with pen and paper, with the Merkle tree.

98
00:05:46,000 --> 00:05:52,000
Well, it goes a bit deeper, so now let's go back into Merkle trees.

99
00:05:53,000 --> 00:05:58,000
So for a normal Merkle tree, you can make something called an inclusion proof.

100
00:05:58,000 --> 00:06:06,000
So if I know an entry at the bottom of this tree and someone has the whole database,

101
00:06:06,000 --> 00:06:14,000
basically the mint, can give me the data I need to traverse from that entry that I know up until the top.

102
00:06:14,000 --> 00:06:19,000
So I can ask the mint, so I know this one entry that I just added to your database,

103
00:06:19,000 --> 00:06:23,000
can you give me like the parent or the sibling of the level above,

104
00:06:23,000 --> 00:06:28,000
and then the sibling of the level above, and sibling of the level above, so on and so forth,

105
00:06:28,000 --> 00:06:32,000
until you reach the very top of the hash.

106
00:06:32,000 --> 00:06:33,000
And with that data…

107
00:06:33,000 --> 00:06:37,000
And the neat thing about that is you don't have to look at everything,

108
00:06:37,000 --> 00:06:39,000
you just look at one branch of the Merkle tree basically.

109
00:06:39,000 --> 00:06:41,000
So it's very inexpensive to do.

110
00:06:41,000 --> 00:06:47,000
You just need a couple of entries of the Merkle tree for me as a user to be able to verify

111
00:06:47,000 --> 00:06:51,000
that that hash that you have at the very top of the Merkle tree

112
00:06:51,000 --> 00:06:54,000
is actually corresponding to the entry that I just gave you.

113
00:06:54,000 --> 00:06:56,000
So the mint has your token.

114
00:06:56,000 --> 00:06:58,000
Right. This is called an inclusion proof.

115
00:06:58,000 --> 00:07:03,000
So I can prove that this entry is part of the Merkle tree without knowing the entire tree.

116
00:07:03,000 --> 00:07:05,000
And this is super powerful.

117
00:07:05,000 --> 00:07:11,000
And now when you go to a sparse Merkle tree, which is like the one that is even bigger,

118
00:07:11,000 --> 00:07:14,000
I just tried to really like in a short way explain it in the beginning,

119
00:07:14,000 --> 00:07:18,000
but essentially it's again a Merkle tree but just organized a little bit differently,

120
00:07:18,000 --> 00:07:23,000
but that one also allows you to create something called an exclusion proof.

121
00:07:23,000 --> 00:07:26,000
So you can prove that an entry is not in the database.

122
00:07:26,000 --> 00:07:31,000
And an exclusion proof in the context of like an e-cash token

123
00:07:31,000 --> 00:07:34,000
where the mint carries this entire tree as the state,

124
00:07:34,000 --> 00:07:39,000
an exclusion proof proves that the token that I have hasn't been spent yet.

125
00:07:39,000 --> 00:07:41,000
So why would that be interesting?

126
00:07:41,000 --> 00:07:47,000
Now the idea you saw me sketching on a piece of paper was just basically trying to think a little ahead,

127
00:07:47,000 --> 00:07:50,000
although all the thoughts that I have are not complete yet.

128
00:07:50,000 --> 00:07:53,000
Like it doesn't work fully in my head yet.

129
00:07:53,000 --> 00:08:00,000
But imagine the mint would kind of commit to the Merkle route on the Bitcoin blockchain.

130
00:08:00,000 --> 00:08:05,000
It doesn't really work on a Bitcoin blockchain because we cannot do Merkle proof on Bitcoin yet

131
00:08:05,000 --> 00:08:08,000
because we would need like an opcode like opcat or something.

132
00:08:09,000 --> 00:08:15,000
But we could do it, for example, on arcade or other like more expressive scripting languages on top of Bitcoin.

133
00:08:15,000 --> 00:08:21,000
But the idea is that if the mint would say this is the current state of the mint

134
00:08:21,000 --> 00:08:26,000
and I imprint it onto the Bitcoin blockchain and I add a script on top of it which says

135
00:08:26,000 --> 00:08:31,000
any user who can prove that they have an e-cash token that is valid,

136
00:08:31,000 --> 00:08:37,000
which was signed by the mint, and can prove that the e-cash token is not included in that Merkle tree,

137
00:08:37,000 --> 00:08:42,000
so with an exclusion proof, might be able to, let's say, unilaterally withdraw from the mint.

138
00:08:42,000 --> 00:08:48,000
So you could build a script that lives on-chain that the mint uses to manage its reserves

139
00:08:48,000 --> 00:08:54,000
that says anyone who can produce this exclusion proof can withdraw from the reserve of the mint.

140
00:08:54,000 --> 00:09:02,000
And why is that interesting? Because it, at least in theory, opens like a way for a unilateral exit for e-cash mints.

141
00:09:02,000 --> 00:09:06,000
So even if the mint goes offline, even if the mint doesn't want to cooperate with you,

142
00:09:06,000 --> 00:09:10,000
you can just go on-chain and say, here's the exclusion proof, give me the Bitcoin back.

143
00:09:10,000 --> 00:09:13,000
Alright, awesome. I'll go three steps back again.

144
00:09:13,000 --> 00:09:16,000
So, CashView is awesome, mints are awesome.

145
00:09:16,000 --> 00:09:25,000
The two main issues, one you just mentioned, which is basically a full, I don't want to call it a rug pull,

146
00:09:25,000 --> 00:09:32,000
like the mint goes offline. Because I said rug pull once and you objected and basically it's like a DDoS.

147
00:09:33,000 --> 00:09:40,000
It's like denial of service. The mint cannot rug pull you, that's where the blinded part comes in,

148
00:09:40,000 --> 00:09:42,000
because you can't be targeted, the mint doesn't know about you.

149
00:09:42,000 --> 00:09:47,000
But the mint can go offline and then you can't access your tokens, which means you can't send anything.

150
00:09:47,000 --> 00:09:50,000
I want to make that a little precise, just for completeness.

151
00:09:50,000 --> 00:09:54,000
The mint can rug pull everyone by stealing the Bitcoin in the current paradigm.

152
00:09:54,000 --> 00:09:56,000
Yeah, we'll come to that later.

153
00:09:56,000 --> 00:09:58,000
But it cannot rug pull you specifically.

154
00:09:58,000 --> 00:10:01,000
I'll ask you about the TESF later.

155
00:10:01,000 --> 00:10:08,000
So, the two main criticisms is basically right now, let's make it three.

156
00:10:08,000 --> 00:10:13,000
The mint can go offline, the mint can rug pull you and the mint can inflate.

157
00:10:13,000 --> 00:10:15,000
Right. Fractional reserve.

158
00:10:15,000 --> 00:10:19,000
So, to simplify, there's inflation risk and there's rug pull risk.

159
00:10:19,000 --> 00:10:20,000
Right.

160
00:10:20,000 --> 00:10:22,000
And you're trying to tackle both of them.

161
00:10:22,000 --> 00:10:23,000
Right.

162
00:10:23,000 --> 00:10:26,000
And so, the exclusion proofs of the sparse miracle trees?

163
00:10:26,000 --> 00:10:27,000
Yes.

164
00:10:28,000 --> 00:10:35,000
If you add unilateral exit to the system, then you kind of solve almost the entire problem.

165
00:10:35,000 --> 00:10:46,000
If you could couple that on-chain anchoring of the mint state tightly to the mint's operation itself,

166
00:10:46,000 --> 00:10:49,000
you might be able to solve both problems at the same time.

167
00:10:49,000 --> 00:10:52,000
So, I call it the slow rug pull and the fast rug pull, actually.

168
00:10:52,000 --> 00:10:58,000
So, the slow rug pull would be the inflation risk, where the mint prints tokens slowly at the sides.

169
00:10:58,000 --> 00:11:00,000
No one really notices the mint.

170
00:11:00,000 --> 00:11:02,000
Sometimes it withdraws from the mint itself.

171
00:11:02,000 --> 00:11:06,000
So, the mint operator pretends to be a user and withdraws some.

172
00:11:06,000 --> 00:11:12,000
And then at the end, the last guy still holding the bags is basically the one who's rug pulled.

173
00:11:12,000 --> 00:11:14,000
And then the other one is the fast rug pull.

174
00:11:14,000 --> 00:11:20,000
The mint just says, you know, fuck this, I'm closing down, taking the Bitcoin and going to Cuba.

175
00:11:20,000 --> 00:11:22,000
That would be the fast one.

176
00:11:22,000 --> 00:11:29,000
So, if you could add unilateral exit to Cashew,

177
00:11:29,000 --> 00:11:39,000
and if you could guarantee that the mint honestly represents all the issued tokens and the burned tokens on the blockchain,

178
00:11:39,000 --> 00:11:43,000
you could solve, in theory, both of these problems.

179
00:11:43,000 --> 00:11:48,000
Again, asterisks, this thought is not fully complete in my head yet.

180
00:11:48,000 --> 00:11:50,000
But I think it kind of makes sense.

181
00:11:50,000 --> 00:11:52,000
Yeah, I see.

182
00:11:52,000 --> 00:11:58,000
So, all right, let's switch gears to the TE stuff that I just alluded to.

183
00:11:58,000 --> 00:12:10,000
So, the issue right now is also, or maybe describe the problem that running a Cashew mint in a trusted execution environment would solve.

184
00:12:10,000 --> 00:12:13,000
So, if you run a Cashew mint on an ordinary server,

185
00:12:14,000 --> 00:12:19,000
then you as the operator of the mint generate the private keys for the e-cash itself,

186
00:12:19,000 --> 00:12:24,000
so the e-cash issuance, that is the private keys to make blind signatures,

187
00:12:24,000 --> 00:12:27,000
which creates e-cash tokens, that creates the supply.

188
00:12:27,000 --> 00:12:29,000
The key to the mint.

189
00:12:29,000 --> 00:12:34,000
And the same key can also be used for the Bitcoin on-chain reserve.

190
00:12:34,000 --> 00:12:40,000
So, we can think of it as a single key that is used for both operating the mint and for securing the Bitcoin.

191
00:12:40,000 --> 00:12:48,000
So, if you run that server on a normal server, obviously, you can just SSH into the thing and then get the private key,

192
00:12:48,000 --> 00:12:50,000
rock everyone and run away.

193
00:12:50,000 --> 00:12:56,000
So, a TE is called a trusted execution environment, sometimes called an enclave,

194
00:12:56,000 --> 00:13:01,000
but it's essentially a server where you as the operator don't have access to the server itself.

195
00:13:01,000 --> 00:13:07,000
So, it's like setting up a server in like a full self-driving mode.

196
00:13:07,000 --> 00:13:15,000
You start the server, you install, let's say, CDK mint, the mint implementation that we have for Cashew,

197
00:13:15,000 --> 00:13:20,000
and when you start it, it generates the private keys for the first time,

198
00:13:20,000 --> 00:13:25,000
but from that moment on, you never have access to the server again.

199
00:13:25,000 --> 00:13:31,000
So, if you start the server with a verified version of CDK,

200
00:13:31,000 --> 00:13:37,000
where you can like verify on GitHub and look at the code and see it does exactly what it says it will do,

201
00:13:37,000 --> 00:13:40,000
then at that point, the mint just keeps running,

202
00:13:40,000 --> 00:13:46,000
and theoretically, no one should be able to break into the mint and steal the Bitcoin

203
00:13:46,000 --> 00:13:49,000
because the operator himself never sees the private keys.

204
00:13:49,000 --> 00:13:51,000
It's generated inside in a sealed box.

205
00:13:51,000 --> 00:13:55,000
Yeah, same idea as the secure enclaves we have on our phones.

206
00:13:55,000 --> 00:13:57,000
Right, secure enclaves we have on our phones.

207
00:13:57,000 --> 00:14:00,000
Some hardware wallets have that, Bitcoin hardware wallets have that.

208
00:14:00,000 --> 00:14:03,000
There's like a bunch of shit coins that run on the same premise.

209
00:14:03,000 --> 00:14:09,000
Essentially, hiding away the private keys from the guy who's actually operating the server.

210
00:14:09,000 --> 00:14:18,000
So, why is that very useful for the Cashew case is because now you can essentially rule out

211
00:14:18,000 --> 00:14:23,000
that the fast rug risk and the slow rug risk of the operator,

212
00:14:23,000 --> 00:14:26,000
because the operator doesn't have access to these private keys.

213
00:14:26,000 --> 00:14:27,000
He cannot mess up with it.

214
00:14:27,000 --> 00:14:31,000
And the nice thing about secure enclaves is you don't have to like,

215
00:14:31,000 --> 00:14:36,000
trust me, bro, the operator is actually running a real version of CDK,

216
00:14:36,000 --> 00:14:39,000
but the client himself can connect to that server

217
00:14:39,000 --> 00:14:45,000
and get like a signature from the actual hardware of that server

218
00:14:45,000 --> 00:14:48,000
that says I am running this software and nothing else.

219
00:14:48,000 --> 00:14:52,000
So, before you connect to the mint, before you send Bitcoin to the mint,

220
00:14:52,000 --> 00:14:54,000
you can verify this is CDK.

221
00:14:54,000 --> 00:14:56,000
I've reviewed it on GitHub.

222
00:14:56,000 --> 00:14:58,000
This is the right version.

223
00:14:58,000 --> 00:14:59,000
Yeah, this is the release.

224
00:14:59,000 --> 00:15:00,000
This is the tag.

225
00:15:00,000 --> 00:15:02,000
It was a signed release and so on.

226
00:15:02,000 --> 00:15:06,000
And it's running exactly that code and it has exactly that hash.

227
00:15:06,000 --> 00:15:09,000
This is amazing because it reduces the risk by a lot.

228
00:15:09,000 --> 00:15:15,000
So, it's great for the users because users know that the mint operator cannot steal the Bitcoin.

229
00:15:15,000 --> 00:15:17,000
The mint operator cannot print more e-cash.

230
00:15:17,000 --> 00:15:20,000
But it's also great for the mint operator because…

231
00:15:20,000 --> 00:15:21,000
Plausible deniability.

232
00:15:21,000 --> 00:15:22,000
Plausible deniability.

233
00:15:22,000 --> 00:15:27,000
Creates this beautiful scenario where the mint operator can actually say I have no access to the Bitcoin.

234
00:15:27,000 --> 00:15:29,000
I cannot run away with them.

235
00:15:29,000 --> 00:15:30,000
I have no access.

236
00:15:30,000 --> 00:15:33,000
I don't have unilateral access to the reserves of the mint.

237
00:15:33,000 --> 00:15:34,000
Yeah.

238
00:15:34,000 --> 00:15:38,000
So, what some people might know about is like with VPN providers and those types of services,

239
00:15:38,000 --> 00:15:41,000
if a government comes knocking and says give me all the user data

240
00:15:41,000 --> 00:15:44,000
or also for encrypted messaging, encrypted emails and so on,

241
00:15:44,000 --> 00:15:48,000
what you're describing is what some of those guys are already doing.

242
00:15:48,000 --> 00:15:50,000
They can literally say I don't have any user data.

243
00:15:50,000 --> 00:15:51,000
I can't do anything.

244
00:15:51,000 --> 00:15:52,000
I can't give you anything.

245
00:15:52,000 --> 00:15:57,000
Well, this is on top of the privacy that the e-cash mint provides you anyway.

246
00:15:57,000 --> 00:16:03,000
So, the mint itself, the whole blind signature part already gives incredible privacy whether it's in TE or not.

247
00:16:03,000 --> 00:16:04,000
Yes.

248
00:16:04,000 --> 00:16:07,000
But on top of that, you can even like if the police comes knocking,

249
00:16:07,000 --> 00:16:11,000
they wouldn't be able to get the private keys and take the Bitcoin, for example.

250
00:16:11,000 --> 00:16:14,000
Or even the protection for a hacker.

251
00:16:14,000 --> 00:16:18,000
There's also another thing like an independent risk that you have is even if you're a good guy,

252
00:16:19,000 --> 00:16:21,000
you get hacked, someone gets a private key somehow.

253
00:16:21,000 --> 00:16:22,000
Exactly.

254
00:16:22,000 --> 00:16:28,000
There's always a possibility that someone breaks into your server and just starts printing Bitcoin, for example.

255
00:16:28,000 --> 00:16:34,000
So, you can also reduce that risk basically to zero by running it in a TE environment.

256
00:16:34,000 --> 00:16:35,000
All right.

257
00:16:35,000 --> 00:16:36,000
Fantastic.

258
00:16:36,000 --> 00:16:38,000
How far along are we with this?

259
00:16:38,000 --> 00:16:41,000
So, we have mints running in TE.

260
00:16:41,000 --> 00:16:43,000
So, we are there.

261
00:16:43,000 --> 00:16:44,000
It works.

262
00:16:44,000 --> 00:16:49,000
I'm like one of only two, three people who have tried it so far because it's not ready to roll out yet.

263
00:16:49,000 --> 00:16:51,000
But we've shown it works.

264
00:16:51,000 --> 00:16:53,000
It works great with on-chain Bitcoin.

265
00:16:53,000 --> 00:16:54,000
One complication is…

266
00:16:54,000 --> 00:16:56,000
So, instead of a Lightning Gateway, you have an on-chain Gateway.

267
00:16:56,000 --> 00:16:57,000
Right.

268
00:16:57,000 --> 00:16:59,000
You cannot really run a Lightning Node inside.

269
00:16:59,000 --> 00:17:01,000
There is a beautiful project called Lexi.

270
00:17:01,000 --> 00:17:07,000
There's actually a technically huge achievement who have managed to run also Lightning Node inside a TE.

271
00:17:07,000 --> 00:17:08,000
Okay.

272
00:17:08,000 --> 00:17:11,000
But it's nothing like that you just do on the side job, right?

273
00:17:12,000 --> 00:17:15,000
Whereas on-chain, if the mint is backed with on-chain, it's very simple.

274
00:17:15,000 --> 00:17:17,000
You have a BDK wallet.

275
00:17:17,000 --> 00:17:18,000
You generate a private key.

276
00:17:18,000 --> 00:17:19,000
You just talk to the blockchain.

277
00:17:19,000 --> 00:17:20,000
That's it.

278
00:17:20,000 --> 00:17:24,000
It's super simple to run an on-chain wallet inside an enclave.

279
00:17:24,000 --> 00:17:26,000
It's not so easy to run with Lightning.

280
00:17:26,000 --> 00:17:28,000
But that's not a big, big problem.

281
00:17:28,000 --> 00:17:31,000
We can still find solution for Lightning.

282
00:17:31,000 --> 00:17:32,000
That still works.

283
00:17:32,000 --> 00:17:38,000
But you can think of it like the easiest way, the simplest way is having an on-chain wallet that backs the mint.

284
00:17:38,000 --> 00:17:39,000
And so, that works.

285
00:17:39,000 --> 00:17:45,000
I've sent on-chain Bitcoin to a server where I would never be able to take out the Bitcoin again,

286
00:17:45,000 --> 00:17:49,000
except if I send back e-cash to the server and withdraw.

287
00:17:49,000 --> 00:17:53,000
And you go through the actual, is it called melting actually?

288
00:17:53,000 --> 00:17:59,000
It's called minting during creation and melting during the destruction of e-cash.

289
00:17:59,000 --> 00:18:04,000
So, no funny business, just the official path to get the sets out again.

290
00:18:04,000 --> 00:18:12,000
You mentioned briefly before with the exclusion proofs that it would be possible on Spark using ArcScript, right?

291
00:18:12,000 --> 00:18:13,000
Yes.

292
00:18:13,000 --> 00:18:15,000
Tell me more about that.

293
00:18:15,000 --> 00:18:18,000
That's still something I barely understand.

294
00:18:18,000 --> 00:18:24,000
So, I said Arcade because Arcade is one of the two Arc implementations that we have.

295
00:18:24,000 --> 00:18:26,000
So, there's second called Bark.

296
00:18:26,000 --> 00:18:29,000
And then there's Arcade from the Arcade guys.

297
00:18:29,000 --> 00:18:34,000
And they both take very different approaches on how to implement Arc.

298
00:18:34,000 --> 00:18:38,000
I would say Bark really focuses on payments and Bitcoin.

299
00:18:38,000 --> 00:18:45,000
And Arcade tries to build something like a virtual machine that can run in an Arc-style system.

300
00:18:45,000 --> 00:18:50,000
And they want to enable basically all kinds of computation that you could run.

301
00:18:51,000 --> 00:19:00,000
Instead of running Bitcoin script, you could run other scripts that can be enforced on the Bitcoin blockchain eventually through that unilateral exit path of Arc.

302
00:19:00,000 --> 00:19:04,000
When you say all kinds of computation, like what level of computation are we talking?

303
00:19:04,000 --> 00:19:06,000
Are we talking Turing-complete?

304
00:19:06,000 --> 00:19:10,000
I cannot really answer that question because I haven't looked that deep into it.

305
00:19:10,000 --> 00:19:16,000
But I would assume that it's Turing-complete because it can essentially do what Ethereum could also do.

306
00:19:16,000 --> 00:19:18,000
All sorts of computation.

307
00:19:18,000 --> 00:19:20,000
And that's the nice thing about it.

308
00:19:20,000 --> 00:19:24,000
It doesn't change Bitcoin, but it still allows you to make those complex computation.

309
00:19:24,000 --> 00:19:30,000
Why it's necessary for this particular problem is the exclusion proof itself.

310
00:19:30,000 --> 00:19:36,000
So, on the Bitcoin blockchain, you cannot execute an inclusion proof or an exclusion proof for Merkle tree.

311
00:19:36,000 --> 00:19:39,000
You mentioned we would need that.

312
00:19:39,000 --> 00:19:40,000
Yes.

313
00:19:40,000 --> 00:19:42,000
So, with Opcat, we could do it on-chain.

314
00:19:42,000 --> 00:19:45,000
But it would be extremely expensive to still just do it on-chain.

315
00:19:45,000 --> 00:19:52,000
It would take a kilobyte of data just to prove an inclusion or exclusion proof for a 256-bit deep Merkle tree.

316
00:19:52,000 --> 00:19:57,000
So, that's a bunch of money that you would have to spend for a unilateral exit.

317
00:19:57,000 --> 00:19:59,000
Whereas on Arcade, it's essentially for free.

318
00:19:59,000 --> 00:20:05,000
So, you can do it very fast, very cheap, and then unilaterally exit from Arcade back to the Bitcoin on-chain.

319
00:20:05,000 --> 00:20:06,000
Okay.

320
00:20:06,000 --> 00:20:09,000
Given all that, what's your best guess for the next...

321
00:20:09,000 --> 00:20:11,000
I know it's impossible to say.

322
00:20:12,000 --> 00:20:21,000
And in the world that we're living in right now, I was about to say near-term future, 6 to 12 months.

323
00:20:21,000 --> 00:20:24,000
But that's super futuristic.

324
00:20:24,000 --> 00:20:27,000
12 months out, we'll be living in a different world.

325
00:20:27,000 --> 00:20:28,000
Exactly.

326
00:20:28,000 --> 00:20:30,000
I have no idea what's going to happen in 12 months from now on.

327
00:20:30,000 --> 00:20:38,000
But in terms of what you want to ship next, so to speak, or what are the people that you work with that are all pushing this forward?

328
00:20:38,000 --> 00:20:39,000
Yes.

329
00:20:39,000 --> 00:20:40,000
Where is the puck going?

330
00:20:40,000 --> 00:20:41,000
How do you see it evolve?

331
00:20:41,000 --> 00:20:45,000
A bunch of super smart people have been nerd-snapped by this, which I love.

332
00:20:45,000 --> 00:20:54,000
It's beautiful to see that there's a bunch of people I really respect in the Bitcoin space who have contributed brainpower to this idea.

333
00:20:54,000 --> 00:20:56,000
Luke Childs has done it.

334
00:20:56,000 --> 00:20:58,000
Hooks has also been nerd-snapped by this topic.

335
00:20:58,000 --> 00:21:00,000
I think Waxwing also looked at it, right?

336
00:21:00,000 --> 00:21:03,000
Waxwing was also super interested in it when I explained it to him.

337
00:21:03,000 --> 00:21:08,000
But there are a few challenges still remaining.

338
00:21:08,000 --> 00:21:10,000
So the TE part, we solved it.

339
00:21:10,000 --> 00:21:11,000
CDK works.

340
00:21:11,000 --> 00:21:13,000
Yeah, proof of concept is there.

341
00:21:13,000 --> 00:21:15,000
We know that this works and we know that this will work.

342
00:21:15,000 --> 00:21:18,000
But for example, there's still some risk remaining.

343
00:21:18,000 --> 00:21:21,000
The biggest one is you shut the mint off.

344
00:21:21,000 --> 00:21:27,000
So there's still this issue that even though the operator cannot steal the Bitcoin, they can still just turn off the mint.

345
00:21:27,000 --> 00:21:30,000
Basically, if they don't pay their Amazon bills, then the mint goes offline.

346
00:21:30,000 --> 00:21:31,000
What happens then?

347
00:21:31,000 --> 00:21:36,000
There is one very simple and elegant idea that I really like about how this can be managed.

348
00:21:36,000 --> 00:21:48,000
Instead of just storing the funds of the mint in a simple on-chain address, you could store it in a little bit more complicated script of on-chain.

349
00:21:48,000 --> 00:21:55,000
Which says the mint can control the on-chain funds, except if you wait two weeks.

350
00:21:55,000 --> 00:21:59,000
Then another multi-sig quorum, like a third party.

351
00:21:59,000 --> 00:22:01,000
So you have a time sig triggered.

352
00:22:01,000 --> 00:22:02,000
Right.

353
00:22:02,000 --> 00:22:07,000
So if the mint goes offline, you wait two weeks and then another entity then gets control of the Bitcoin.

354
00:22:07,000 --> 00:22:12,000
And they could then help, let's say, the users of the mint that just went offline.

355
00:22:12,000 --> 00:22:14,000
They could help them to withdraw.

356
00:22:14,000 --> 00:22:16,000
Take the latest state that we know and tally it up.

357
00:22:16,000 --> 00:22:17,000
Exactly.

358
00:22:17,000 --> 00:22:24,000
So a requirement for that would be that this other third party needs access to the database of the mint.

359
00:22:24,000 --> 00:22:27,000
Because they need to know which e-cash has been spent already.

360
00:22:27,000 --> 00:22:30,000
And so after two weeks, the mint is offline.

361
00:22:30,000 --> 00:22:35,000
After two weeks, you go to the third party and say, here's an e-cash token I haven't spent yet.

362
00:22:35,000 --> 00:22:36,000
Check your database.

363
00:22:36,000 --> 00:22:37,000
They're like, okay, this is true.

364
00:22:37,000 --> 00:22:38,000
The mint is offline.

365
00:22:38,000 --> 00:22:40,000
You get your Bitcoin back.

366
00:22:40,000 --> 00:22:41,000
I see.

367
00:22:41,000 --> 00:22:42,000
Awesome.

368
00:22:42,000 --> 00:22:43,000
Awesome.

369
00:22:43,000 --> 00:22:44,000
That's cool.

370
00:22:44,000 --> 00:22:48,000
That's a good use of time locks and multiple spending paths and those kind of things.

371
00:22:48,000 --> 00:22:49,000
Fantastic.

372
00:22:49,000 --> 00:22:50,000
All right.

373
00:22:50,000 --> 00:22:52,000
What else?

374
00:22:52,000 --> 00:22:53,000
What else?

375
00:22:54,000 --> 00:22:57,000
Those were the two biggest question marks for me.

376
00:22:57,000 --> 00:23:01,000
Where I was like, if I have 20 minutes to pick your brain, that's what I'm going to ask.

377
00:23:01,000 --> 00:23:03,000
But lots of other things have happened.

378
00:23:03,000 --> 00:23:06,000
Obviously, we're still at the Freedom Forum.

379
00:23:06,000 --> 00:23:09,000
AOS convergence happened before.

380
00:23:09,000 --> 00:23:13,000
What are some other exciting things that are currently happening on the Cashew side?

381
00:23:13,000 --> 00:23:20,000
So one thing that is a pretty big change that we're working on right now is we're changing our entire crypto.

382
00:23:20,000 --> 00:23:22,000
So I'm sorry for everyone listening right now.

383
00:23:22,000 --> 00:23:23,000
We're making it again.

384
00:23:23,000 --> 00:23:29,000
But we feel like as long as BlackRock hasn't bought Cashew, we can still make changes, basically.

385
00:23:29,000 --> 00:23:37,000
So what we're doing right now is currently we're using SecP256K1 as the curve for Cashew.

386
00:23:37,000 --> 00:23:44,000
It's a very simple cryptographic scheme invented by David Wagner in 1996.

387
00:23:44,000 --> 00:23:51,000
It's funny because he invented it because David Chaum himself, he patented RSA blind signatures.

388
00:23:51,000 --> 00:24:00,000
And because no one was able to use that, David Wagner just used an elliptic curve paradigm to make blind signatures.

389
00:24:00,000 --> 00:24:03,000
Why do you think these people like patents so much?

390
00:24:03,000 --> 00:24:06,000
Like, you know, Schnorr was also...

391
00:24:06,000 --> 00:24:07,000
It is terrible.

392
00:24:07,000 --> 00:24:12,000
I mean, Schnorr is a good example because Satoshi couldn't use Schnorr signatures up until he was gone.

393
00:24:12,000 --> 00:24:15,000
And then we could use them as Schnorr signatures.

394
00:24:15,000 --> 00:24:17,000
Why they do it? I don't know.

395
00:24:17,000 --> 00:24:24,000
I think they live in a sugar world reality where they think like this idea, if I protect it, I can make money off it.

396
00:24:24,000 --> 00:24:26,000
And others won't be able to do it.

397
00:24:26,000 --> 00:24:30,000
But, you know, I'm absolutely not a big fan of patents at all.

398
00:24:30,000 --> 00:24:35,000
I would rather want to live in a world where ideas cannot be protected by intellectual property.

399
00:24:35,000 --> 00:24:38,000
Shout out, great book against intellectual property.

400
00:24:38,000 --> 00:24:39,000
Absolutely.

401
00:24:39,000 --> 00:24:44,000
I mean, for me, I think like we need to tear down this entire system of intellectual property.

402
00:24:44,000 --> 00:24:47,000
I think information wants to be free, should be free.

403
00:24:47,000 --> 00:24:51,000
And not like the idea itself isn't the main important part, it's the execution.

404
00:24:51,000 --> 00:24:55,000
Yeah, in the cypherpunk future, like where we are going, we don't need no roads.

405
00:24:55,000 --> 00:24:57,000
So the roads are intellectually property, right?

406
00:24:57,000 --> 00:24:58,000
100%.

407
00:25:00,000 --> 00:25:06,000
Anyway, so we use that scheme from David Wagner, which is a very simple blind signature scheme.

408
00:25:06,000 --> 00:25:14,000
And it has a couple of problems that, you know, for example, the biggest one is, it is technically speaking, not a blind signature scheme.

409
00:25:14,000 --> 00:25:16,000
It is a blind MAC scheme.

410
00:25:16,000 --> 00:25:22,000
And a MAC scheme is Message Authentication Code for the crypto nerds out there who understand.

411
00:25:22,000 --> 00:25:26,000
A signature scheme is defined as you can verify the signature with a public key.

412
00:25:26,000 --> 00:25:32,000
And in Cashew or with any MAC scheme, you can verify a signature only with a private key itself.

413
00:25:32,000 --> 00:25:40,000
So if I, like as a third party, you would not be able to verify if a Cashew token is really signed by the mint or not.

414
00:25:40,000 --> 00:25:43,000
Because you don't have the private keys of the mint.

415
00:25:43,000 --> 00:25:48,000
And for this, we have used a construct called DLEQ proofs.

416
00:25:48,000 --> 00:25:50,000
It's a discrete log equality proof.

417
00:25:50,000 --> 00:25:56,000
And it's like we're doing yoga to make something very simple possible.

418
00:25:56,000 --> 00:25:58,000
And it's not the ideal solution.

419
00:25:58,000 --> 00:26:06,000
We've been building way too complicated solutions for a very simple problem that should have been just verify the signature with the public key.

420
00:26:06,000 --> 00:26:07,000
Done.

421
00:26:07,000 --> 00:26:11,000
But we don't have that because we chose this very simple scheme.

422
00:26:11,000 --> 00:26:15,000
Now, we are tired of it and we want to go to the next level.

423
00:26:15,000 --> 00:26:24,000
And the next level is a pairing based crypto called BLS signatures, which are also very simple math, very well studied.

424
00:26:24,000 --> 00:26:27,000
It is not the Bitcoin curve, but it's a different curve.

425
00:26:27,000 --> 00:26:31,000
And with this new crypto, we'll be able to verify e-cash also with public keys.

426
00:26:31,000 --> 00:26:36,000
That makes everything that I just explained with like the third party helping you to off board and so on and so forth.

427
00:26:36,000 --> 00:26:38,000
Way, way simpler.

428
00:26:38,000 --> 00:26:45,000
So at the end of the day, what we will have is new crypto that is simpler and easier to verify publicly.

429
00:26:45,000 --> 00:26:48,000
And it will allow us to build way more security solutions.

430
00:26:48,000 --> 00:26:51,000
But like you said, that's obviously a breaking change.

431
00:26:51,000 --> 00:26:57,000
So what version are we going to be at with the new cryptographic scheme?

432
00:26:57,000 --> 00:27:03,000
It's going to be like the second version of the crypto, but we have prepared an upgrade path already for these kind of changes.

433
00:27:03,000 --> 00:27:06,000
So it's not going to affect your experience at all.

434
00:27:06,000 --> 00:27:09,000
In the best case, you will not even notice anything.

435
00:27:09,000 --> 00:27:12,000
So we have something called keysets in Cashew.

436
00:27:12,000 --> 00:27:18,000
And keysets are the set of private keys or the set of public keys that the Mint uses to sign these tokens.

437
00:27:18,000 --> 00:27:22,000
And from our perspective, we just add one more version to this keyset.

438
00:27:22,000 --> 00:27:26,000
And in the next version, you will use a different curve with different crypto.

439
00:27:26,000 --> 00:27:29,000
Again, from your perspective as a user, nothing really changes.

440
00:27:29,000 --> 00:27:32,000
From our perspective, everything gets a little simpler.

441
00:27:32,000 --> 00:27:33,000
Yeah. Awesome. Awesome.

442
00:27:33,000 --> 00:27:36,000
What's the timeline on this switch and on this upgrade?

443
00:27:36,000 --> 00:27:37,000
We've already built it.

444
00:27:37,000 --> 00:27:42,000
We just need to press the merge button, but we want to do it like in synchrony.

445
00:27:42,000 --> 00:27:49,000
So we're going to upgrade all the wallets that we're maintaining and then press on the merge button altogether.

446
00:27:49,000 --> 00:27:55,000
And then hopefully people will start deploying these new keysets in mainnet.

447
00:27:55,000 --> 00:27:59,000
And then we'll see hopefully all the wallets just seamlessly continuing to work.

448
00:27:59,000 --> 00:28:00,000
So I would say two weeks.

449
00:28:00,000 --> 00:28:03,000
All right. Two weeks. CM. Fantastic.

450
00:28:03,000 --> 00:28:07,000
You say you've already built it, but that's a bold-faced lie.

451
00:28:07,000 --> 00:28:10,000
No. Go to GitHub.

452
00:28:10,000 --> 00:28:15,000
No, no. What I mean is like your agents have built it for you.

453
00:28:15,000 --> 00:28:19,000
So that's something else I want to dig into with you just for a couple of minutes.

454
00:28:19,000 --> 00:28:22,000
Yes. Let's go there. But I want to be fair.

455
00:28:22,000 --> 00:28:27,000
And, you know, we have a whole team of people working on this.

456
00:28:27,000 --> 00:28:33,000
When I say we built it, then I don't only mean my agents, but the agents of my colleagues.

457
00:28:33,000 --> 00:28:37,000
I was about to say, but everyone is using agents by now, including Ege.

458
00:28:37,000 --> 00:28:44,000
Shout out Ege. Last man. Last man who fell. Last threat caller that got converted recently, I've heard.

459
00:28:44,000 --> 00:28:46,000
Yeah. We've been pushing extremely hard on this.

460
00:28:46,000 --> 00:28:51,000
So I've been like on this whole agent decoding thing from day zero, basically.

461
00:28:51,000 --> 00:28:57,000
And I saw very early, as everyone else said, this is going to be the only way that we're going to work.

462
00:28:57,000 --> 00:29:01,000
So all of us in the team are super AI based.

463
00:29:01,000 --> 00:29:09,000
We make a big difference between what is critical code that a human must edit or at least review thoroughly.

464
00:29:09,000 --> 00:29:14,000
And then there's a bunch of code that is just wallet side, UI, business logic stuff.

465
00:29:14,000 --> 00:29:18,000
That is where, you know, you cannot break anything because the underlying library does everything for you.

466
00:29:18,000 --> 00:29:22,000
And everything like the sugar coating, we just do it completely in AI.

467
00:29:22,000 --> 00:29:26,000
The old paradigm still holds, which is don't roll your own crypto. Right.

468
00:29:26,000 --> 00:29:30,000
And it goes further than that. Like don't vibe your own crypto.

469
00:29:31,000 --> 00:29:34,000
The more things change, the more they stay the same.

470
00:29:34,000 --> 00:29:41,000
But talk me through your setup. Like how does it compare to Marty Malmy's setup, for example?

471
00:29:41,000 --> 00:29:47,000
Which to me is like, you know, other people are merely experiencing AI psychosis.

472
00:29:47,000 --> 00:29:54,000
He was born in it, molded by it. He was running around with the laptop open at all times, basically.

473
00:29:55,000 --> 00:29:59,000
Shout out to Malmy as well. He's an inspiration to me.

474
00:29:59,000 --> 00:30:02,000
But my setup is changing all the time.

475
00:30:02,000 --> 00:30:06,000
So first of all, I have no allegiance to any of the providers or models at all.

476
00:30:06,000 --> 00:30:10,000
I see them as like arbitrary exchangeable products.

477
00:30:10,000 --> 00:30:13,000
So I'll use any model at any given time.

478
00:30:13,000 --> 00:30:17,000
I don't make a difference between an American model and a Chinese model. I don't care.

479
00:30:17,000 --> 00:30:21,000
I run local models as well. And I wish they were better.

480
00:30:21,000 --> 00:30:25,000
But at this point in time, they're still not useful as coding agents.

481
00:30:25,000 --> 00:30:28,000
So you need to go to one of the providers.

482
00:30:28,000 --> 00:30:33,000
My setup personally is I use open router, which means I can use all the models all the time.

483
00:30:33,000 --> 00:30:37,000
I don't need to select one and just keep that.

484
00:30:37,000 --> 00:30:40,000
My life happens in open code.

485
00:30:40,000 --> 00:30:44,000
So this is a coding harness that runs in a terminal.

486
00:30:44,000 --> 00:30:47,000
I spend most of my days in open code inside there.

487
00:30:48,000 --> 00:30:58,000
And yeah, I've developed many different workflows over the year or the months in the recent months.

488
00:30:58,000 --> 00:31:06,000
What I really am fascinated by these days, though, is kind of how can I get even better?

489
00:31:06,000 --> 00:31:09,000
What does it mean to be a really good agentic coder?

490
00:31:09,000 --> 00:31:12,000
What are the tricks and tips that make it work better?

491
00:31:12,000 --> 00:31:16,000
So I'll make one example that I learned from Peter Steinberger, for example.

492
00:31:17,000 --> 00:31:21,000
And this completely blew my mind, although this is such a simple fact.

493
00:31:21,000 --> 00:31:29,000
He said when you watch your agent code a function and the agent gives the function a simple name that you don't like,

494
00:31:29,000 --> 00:31:32,000
then don't tell the agent to change the function name.

495
00:31:32,000 --> 00:31:37,000
Because just because you like it doesn't mean it's the better name to choose.

496
00:31:37,000 --> 00:31:44,000
Because you're going to kill that agent and you're going to create a new agent tomorrow that is supposed to find that function again.

497
00:31:44,000 --> 00:31:46,000
So you have to think basically like them.

498
00:31:46,000 --> 00:31:51,000
You submit to the agent and say, okay, agent, whatever you think is right probably is the best way

499
00:31:51,000 --> 00:31:54,000
because tomorrow another agent will have to continue to work.

500
00:31:54,000 --> 00:32:01,000
So this is one simple factor that you pick up during this entire crazy development.

501
00:32:01,000 --> 00:32:10,000
So what I'm really trying to dig every one single person that I meet who is a coder using AI is what are your tricks?

502
00:32:10,000 --> 00:32:12,000
What are the things that you've discovered?

503
00:32:13,000 --> 00:32:18,000
I believe we're in a special time where there is no written knowledge, no systems, no formalizations.

504
00:32:18,000 --> 00:32:20,000
It is basically like from mouth to mouth.

505
00:32:20,000 --> 00:32:27,000
We're back to the Middle Ages where knowledge spreads through colloquial conversations.

506
00:32:27,000 --> 00:32:33,000
And this is crazy to me because it's already like a billion trillion dollar industry

507
00:32:33,000 --> 00:32:40,000
where the smallest optimizations of how you work with AI could have massive impact on your output.

508
00:32:41,000 --> 00:32:43,000
And also on your cost.

509
00:32:43,000 --> 00:32:47,000
One form of retard maxing is token maxing.

510
00:32:47,000 --> 00:32:51,000
And I see a lot of people do that just like spending endless amounts of tokens.

511
00:32:51,000 --> 00:32:55,000
Another one is that it just makes sense to understand how these machines work.

512
00:32:55,000 --> 00:32:59,000
For example, what I see noobs do a lot is they start a conversation with an agent

513
00:32:59,000 --> 00:33:03,000
and then they go in one direction and figure that's not the right direction.

514
00:33:03,000 --> 00:33:07,000
And they tell the agent, no, that's not what I mean.

515
00:33:07,000 --> 00:33:09,000
Just do something else.

516
00:33:10,000 --> 00:33:15,000
In that case, what I see is you're loading up the context with bullshit that the agent doesn't need.

517
00:33:15,000 --> 00:33:21,000
So instead of trying to navigate the agent back to where you want to go, just kill the agent.

518
00:33:21,000 --> 00:33:22,000
Create a new one.

519
00:33:22,000 --> 00:33:27,000
Make a better prompt from the first one and try to one-shot everything that you're doing.

520
00:33:27,000 --> 00:33:28,000
That's my goal.

521
00:33:28,000 --> 00:33:30,000
I'll tell you two of my tricks.

522
00:33:32,000 --> 00:33:37,000
As you know, I'm a big fan of dialogical development and dialogue in general.

523
00:33:37,000 --> 00:33:40,000
And like you said, we're in the age of dialogue right now.

524
00:33:42,000 --> 00:33:45,000
Word of mouth kind of pro tips and so on.

525
00:33:46,000 --> 00:33:50,000
And obviously, the whole idea of Sovereign Engineering itself was to build something

526
00:33:50,000 --> 00:33:53,000
that evolves around and revolves around dialogue.

527
00:33:53,000 --> 00:34:00,000
And so what I do to build up context is I always basically use the Socratic method before I start.

528
00:34:01,000 --> 00:34:09,000
So, fresh session, new context window, and it's like, basically, let's say on the search,

529
00:34:09,000 --> 00:34:16,000
I want to basically modify some substitutions that the search does.

530
00:34:16,000 --> 00:34:20,000
It's like, are we doing any substitutions in our search?

531
00:34:20,000 --> 00:34:22,000
And it's like, oh yeah, here's how we do it.

532
00:34:22,000 --> 00:34:25,000
And it's like, okay, what are the things that we substitute?

533
00:34:25,000 --> 00:34:26,000
Yeah, here's the things we substitute.

534
00:34:26,000 --> 00:34:29,000
Would it make sense to add these two?

535
00:34:29,000 --> 00:34:31,000
And then, yeah, it would make sense.

536
00:34:31,000 --> 00:34:32,000
Okay, let's add it.

537
00:34:32,000 --> 00:34:34,000
What else do we need to do?

538
00:34:34,000 --> 00:34:37,000
And then I always end on questions also.

539
00:34:37,000 --> 00:34:41,000
And one is like, what's something that we should refactor?

540
00:34:41,000 --> 00:34:44,000
On the code we've just touched, on the branch we're on, what's something that we,

541
00:34:44,000 --> 00:34:47,000
like low-hanging fruit, that we should refactor?

542
00:34:47,000 --> 00:34:48,000
Give me the top three.

543
00:34:48,000 --> 00:34:49,000
Awesome.

544
00:34:49,000 --> 00:34:50,000
Give me the top three.

545
00:34:50,000 --> 00:34:51,000
I always do that.

546
00:34:51,000 --> 00:34:55,000
And if I disagree, like if all top three are bullshit, we don't refactor at all.

547
00:34:55,000 --> 00:34:56,000
Yeah, you know you're finished.

548
00:34:56,000 --> 00:34:57,000
Yeah, exactly.

549
00:34:58,000 --> 00:35:02,000
And so I have macros for all of these and all of these questions.

550
00:35:02,000 --> 00:35:06,000
And so on the conceptual side as well, it's like, you know,

551
00:35:06,000 --> 00:35:10,000
give me the top five ways of how we could do this.

552
00:35:10,000 --> 00:35:14,000
Often what I do is I let the agent implement something complicated.

553
00:35:14,000 --> 00:35:18,000
And at the end of the whole implementation, even if it works, I ask it,

554
00:35:18,000 --> 00:35:21,000
now that you've done it, what would you have done differently?

555
00:35:21,000 --> 00:35:23,000
Because now it has seen all the context.

556
00:35:23,000 --> 00:35:24,000
Exactly.

557
00:35:24,000 --> 00:35:28,000
And only after doing it, it knows what the mistakes were that it did.

558
00:35:28,000 --> 00:35:29,000
Yeah, just like us.

559
00:35:29,000 --> 00:35:31,000
It's like, oh, now if I would do it again,

560
00:35:31,000 --> 00:35:35,000
I would probably factor this function a little bit differently and put this there.

561
00:35:35,000 --> 00:35:36,000
And it's like, okay, just do it.

562
00:35:36,000 --> 00:35:39,000
And it just fixes the problem that it just created by itself.

563
00:35:39,000 --> 00:35:41,000
Yeah, I do that a bunch as well.

564
00:35:41,000 --> 00:35:44,000
And what I work with is multiple models and a fresh context window

565
00:35:44,000 --> 00:35:46,000
or another agent, those kind of things.

566
00:35:46,000 --> 00:35:50,000
I think this is something that, you know, everyone should realize listening,

567
00:35:50,000 --> 00:35:53,000
is context window is precious resource.

568
00:35:53,000 --> 00:35:55,000
You should not fill it with useless stuff.

569
00:35:55,000 --> 00:36:00,000
Because the models even get worse at doing math, like basic simple math.

570
00:36:00,000 --> 00:36:03,000
If you ask an agent what's 3 times 5,

571
00:36:03,000 --> 00:36:07,000
it will answer that question with a higher probability correctly

572
00:36:07,000 --> 00:36:09,000
in the beginning of the context window than at the end.

573
00:36:09,000 --> 00:36:11,000
So if you fill it with a bunch of bullshit before

574
00:36:11,000 --> 00:36:14,000
and then ask a math question, it makes more mistakes.

575
00:36:14,000 --> 00:36:16,000
So this is something to understand.

576
00:36:16,000 --> 00:36:20,000
You keep it as short as possible and create like the –

577
00:36:20,000 --> 00:36:23,000
you essentially like need to help a newborn baby

578
00:36:23,000 --> 00:36:26,000
to catch up to the thing every single time.

579
00:36:26,000 --> 00:36:28,000
And maybe that's also one of the biggest optimizations

580
00:36:28,000 --> 00:36:31,000
that we could do in these harnesses.

581
00:36:31,000 --> 00:36:35,000
Because I spend so much time just explaining the same thing over and over.

582
00:36:35,000 --> 00:36:38,000
Like we're in a project. We're building this thing.

583
00:36:38,000 --> 00:36:39,000
We're currently in this feature.

584
00:36:39,000 --> 00:36:41,000
And this is what we've done so far.

585
00:36:41,000 --> 00:36:44,000
And it's like get back to me when you understand what I'm talking about.

586
00:36:44,000 --> 00:36:46,000
It comes back and it's like, okay, I'm ready.

587
00:36:46,000 --> 00:36:47,000
Now we execute.

588
00:36:47,000 --> 00:36:50,000
I mean, I think what a lot of people are working on right now

589
00:36:50,000 --> 00:36:55,000
is exactly these memory systems that will just like keep certain bits of context

590
00:36:55,000 --> 00:36:59,000
over sessions and will have like an easy way to access the memory.

591
00:36:59,000 --> 00:37:01,000
And like something that's better.

592
00:37:01,000 --> 00:37:05,000
Like obviously OpenCLAW went the way of MDFiles.

593
00:37:05,000 --> 00:37:10,000
But there is OpenCLAW plugins already that use vector databases

594
00:37:10,000 --> 00:37:13,000
and stuff that kind of makes a little more sense.

595
00:37:13,000 --> 00:37:17,000
And I think what we were just talking before, right?

596
00:37:17,000 --> 00:37:19,000
Like lots of people have their OpenCLAWs and so on.

597
00:37:19,000 --> 00:37:23,000
And what I was explaining is like what I found for me,

598
00:37:23,000 --> 00:37:28,000
what works best for me is like if you have multiple projects,

599
00:37:28,000 --> 00:37:31,000
one agent per project.

600
00:37:31,000 --> 00:37:35,000
And then like if you have one, I use OpenCLAW myself,

601
00:37:35,000 --> 00:37:39,000
if you have one agent for every single side project,

602
00:37:39,000 --> 00:37:41,000
every little thing you have, then the context and the memory

603
00:37:41,000 --> 00:37:44,000
and everything and the tools as well builds up over time.

604
00:37:44,000 --> 00:37:46,000
And that works wonders.

605
00:37:46,000 --> 00:37:51,000
Like it's so much better than having one assistant that does all of your six projects

606
00:37:51,000 --> 00:37:52,000
at the same time.

607
00:37:52,000 --> 00:37:54,000
Much more aggressive compartmentalization of it.

608
00:37:54,000 --> 00:37:57,000
And, you know, all the things that you let your agent do,

609
00:37:57,000 --> 00:38:00,000
they influence all the other things that you want it to do.

610
00:38:00,000 --> 00:38:05,000
And because context is so precious, they're not able to do everything at once.

611
00:38:05,000 --> 00:38:08,000
So it basically makes it worse on everything else.

612
00:38:08,000 --> 00:38:09,000
Yes, exactly.

613
00:38:09,000 --> 00:38:12,000
That's exactly what I wanted to mention because from the cognitive science perspective,

614
00:38:12,000 --> 00:38:17,000
we know that there is an inherent tradeoff relationship between being a generalist

615
00:38:17,000 --> 00:38:18,000
and a specialist.

616
00:38:18,000 --> 00:38:21,000
And so you can't be both things at the same time.

617
00:38:21,000 --> 00:38:24,000
And so what you want to have is you want to have your specialists

618
00:38:24,000 --> 00:38:28,000
and maybe you want to have one big model overarching generalist

619
00:38:28,000 --> 00:38:31,000
that knows a lot about cashier, that knows a lot about whatever,

620
00:38:31,000 --> 00:38:34,000
you know, lightning, art, Bitcoin and so on, all those kind of things

621
00:38:34,000 --> 00:38:36,000
or elliptic curves and cryptography and so on.

622
00:38:36,000 --> 00:38:41,000
But on the confusion part you mentioned, I think, is really important

623
00:38:41,000 --> 00:38:44,000
because if you want to stuff everything in one brain, so to speak,

624
00:38:44,000 --> 00:38:46,000
if you want to stuff everything in one...

625
00:38:46,000 --> 00:38:48,000
Just like you want to stuff everything in one person, you know?

626
00:38:48,000 --> 00:38:50,000
Like not even Elon can do absolutely everything.

627
00:38:50,000 --> 00:38:53,000
It's like that's just not how the world works.

628
00:38:53,000 --> 00:38:56,000
That's just not how cognitive structures work.

629
00:38:56,000 --> 00:38:58,000
Cognitive structures can't work that way.

630
00:38:58,000 --> 00:39:01,000
And I think like everyone who's like deep in the rabbit hole,

631
00:39:01,000 --> 00:39:06,000
I would highly encourage everyone who studied just a little bit of cognitive science

632
00:39:06,000 --> 00:39:11,000
and how the brain works and how, you know, we as humans do these kind of things

633
00:39:11,000 --> 00:39:16,000
because you will... That's what the AI people are figuring out right now.

634
00:39:16,000 --> 00:39:20,000
You have certain trade-off relationships that you just...

635
00:39:20,000 --> 00:39:24,000
You can't... Like no matter how much computer you throw at it, you can't fix it.

636
00:39:24,000 --> 00:39:30,000
You will never have a model that is as good on a specialized task as a specialized model.

637
00:39:30,000 --> 00:39:36,000
Yes, but it's interesting because we are still in that paradigm though.

638
00:39:36,000 --> 00:39:39,000
Like if you zoom out, not from a user's perspective,

639
00:39:39,000 --> 00:39:42,000
but I mean from the Frontier Labs perspective,

640
00:39:42,000 --> 00:39:47,000
that their bet still is that the generalist scales more

641
00:39:47,000 --> 00:39:53,000
or it makes more sense to create one big generalist than multiple specialists basically.

642
00:39:53,000 --> 00:39:55,000
So it's an open question.

643
00:39:55,000 --> 00:39:58,000
It's really an open question that no one can answer so far

644
00:39:58,000 --> 00:40:05,000
Trillions of dollars betting on this question is whether there should be like one GPT that does everything

645
00:40:05,000 --> 00:40:11,000
or, and you can see Google already moving a little bit away from that target,

646
00:40:11,000 --> 00:40:14,000
saying maybe we need a special model just for medicine.

647
00:40:14,000 --> 00:40:16,000
Maybe we need a special model just for writing.

648
00:40:16,000 --> 00:40:20,000
But whereas other Frontier Labs just say like we have one big model,

649
00:40:20,000 --> 00:40:26,000
it can do everything and it increases its ability in all these different areas at the same time.

650
00:40:26,000 --> 00:40:29,000
But we really don't know whether it would have made sense, for example,

651
00:40:29,000 --> 00:40:35,000
to create like one coding agent that does nothing but coding and excel at this single task,

652
00:40:35,000 --> 00:40:37,000
whereas being worse than others.

653
00:40:37,000 --> 00:40:41,000
Not all, like what the cognitive scientists would say about this,

654
00:40:41,000 --> 00:40:48,000
and at least those that I respect, is that not all skills transfer to other domains.

655
00:40:48,000 --> 00:40:53,000
Some skills transfer to other domains, but let me give you a very simple example.

656
00:40:54,000 --> 00:41:01,000
Like if you are a world-class baseball player, you have a really good baseball swing,

657
00:41:01,000 --> 00:41:06,000
and you know how to ruin your baseball swing, you start playing tennis.

658
00:41:06,000 --> 00:41:10,000
Those things are in a trade-off relationship with each other.

659
00:41:10,000 --> 00:41:14,000
You can't be a world-class baseball player and a world-class tennis player at the same time,

660
00:41:14,000 --> 00:41:20,000
because the nuances of your baseball swing will get ruined by how much more you will train tennis and so on.

661
00:41:21,000 --> 00:41:24,000
This is exactly the conundrum that the labs are into, basically.

662
00:41:24,000 --> 00:41:26,000
Yeah, so they'll face reality.

663
00:41:26,000 --> 00:41:30,000
You start teaching the LLM to be a better writer, for example,

664
00:41:30,000 --> 00:41:33,000
and it becomes a worse coder at the same time.

665
00:41:33,000 --> 00:41:38,000
It's one of the things that the lore of GPT-4.0 is beautiful,

666
00:41:38,000 --> 00:41:42,000
because there are still people being completely degenerate and wishing back GPT-4.0.

667
00:41:42,000 --> 00:41:44,000
Yeah, because it's sycophantic, it loves you so much.

668
00:41:44,000 --> 00:41:48,000
Extremely sycophantic. People love that model.

669
00:41:49,000 --> 00:41:56,000
The number of AI psychosis generated by that model is way higher than, for example, GPT-5.5 or the 5 series,

670
00:41:56,000 --> 00:42:01,000
because the 5 series is super dry, gives you short answers, it's on point,

671
00:42:01,000 --> 00:42:08,000
and it's better at completing tasks, it's better at coding, but it's not as engaging as the other ones.

672
00:42:08,000 --> 00:42:11,000
So OpenAI needed to answer that question.

673
00:42:11,000 --> 00:42:17,000
It's an economic question for them, because what they actually want is to get as many users as possible,

674
00:42:17,000 --> 00:42:24,000
but if you make it more engaging on a narrative side, conversational side, it becomes a worse coder,

675
00:42:24,000 --> 00:42:28,000
whereas the coders say, I want that model for coding, what are you guys doing?

676
00:42:28,000 --> 00:42:31,000
I don't want a novel to come out of that thing.

677
00:42:31,000 --> 00:42:33,000
So they need to make that decision.

678
00:42:33,000 --> 00:42:39,000
It seems like they leaned towards more the application, the professionals, GPT-5.5 is clear.

679
00:42:39,000 --> 00:42:42,000
I mean, coding is what pays the bill right now, right?

680
00:42:42,000 --> 00:42:49,000
That is true, but there are, I don't know the numbers, I don't want to guess on a podcast, but there's a...

681
00:42:49,000 --> 00:42:51,000
It's a dialogue, not a podcast.

682
00:42:51,000 --> 00:42:54,000
I wouldn't even guess on a dialogue.

683
00:42:54,000 --> 00:42:58,000
A very large portion of the world uses LLMs.

684
00:42:58,000 --> 00:43:02,000
It's an incredible, it's the fastest technology in history.

685
00:43:02,000 --> 00:43:05,000
Many, many people just use it for conversation.

686
00:43:05,000 --> 00:43:07,000
Most people don't code.

687
00:43:07,000 --> 00:43:10,000
Absolutely, but they are not the paying bunch, usually.

688
00:43:11,000 --> 00:43:14,000
I think from the economics side, we'll see.

689
00:43:14,000 --> 00:43:15,000
We'll see how it shakes out.

690
00:43:15,000 --> 00:43:19,000
I think it's still an open question how these things will in the end be able to monetize.

691
00:43:19,000 --> 00:43:23,000
Do you think the addiction for a coder is stronger than for a conversational person?

692
00:43:23,000 --> 00:43:27,000
I mean, people end up marrying their chatbots right now,

693
00:43:27,000 --> 00:43:30,000
and actually divorcing their husbands and marrying their chatbots.

694
00:43:30,000 --> 00:43:32,000
So I'm not sure.

695
00:43:32,000 --> 00:43:35,000
I'm not sure if I'm brave enough to answer the question.

696
00:43:36,000 --> 00:43:38,000
But on a kind of enterprise side,

697
00:43:38,000 --> 00:43:43,000
large software companies are willing to pay a lot for replacing their engineers.

698
00:43:43,000 --> 00:43:49,000
I've never seen any product so far where an individual is willing to spend $200 a month

699
00:43:49,000 --> 00:43:51,000
to get access to a cloud service.

700
00:43:51,000 --> 00:43:55,000
And we know people who have five of these accounts at the same time.

701
00:43:55,000 --> 00:43:58,000
Some people spend $1,000 per month to get as much...

702
00:43:58,000 --> 00:44:01,000
And in Silicon Valley, those numbers are way higher.

703
00:44:01,000 --> 00:44:02,000
Token maxing.

704
00:44:02,000 --> 00:44:04,000
Yeah, token maxing.

705
00:44:04,000 --> 00:44:09,000
As I said before, I think it's a form of retard maxing right now.

706
00:44:09,000 --> 00:44:13,000
Should we also talk about beyond LLMs?

707
00:44:13,000 --> 00:44:16,000
Yeah, let's hit it. Let's hit the world model.

708
00:44:16,000 --> 00:44:20,000
Okay, so I believe LLMs are kind of a scam in a sense.

709
00:44:20,000 --> 00:44:22,000
I mean, they're obviously extremely useful.

710
00:44:22,000 --> 00:44:24,000
I will keep using them forever and so on.

711
00:44:24,000 --> 00:44:30,000
But what the AI companies are forced to convince us of

712
00:44:30,000 --> 00:44:35,000
is that this is like the best of the best and this is the crown jewels of AI.

713
00:44:35,000 --> 00:44:40,000
And there are many people, especially doing AI research,

714
00:44:40,000 --> 00:44:44,000
who believe that LLMs are not the end of the answer.

715
00:44:44,000 --> 00:44:49,000
One of the biggest problems of LLMs are that they're essentially just token prediction machines.

716
00:44:49,000 --> 00:44:52,000
So they take a set of tokens and they predict the next token.

717
00:44:52,000 --> 00:44:53,000
That's all they do.

718
00:44:53,000 --> 00:44:56,000
And everything that they've trained on is language-based.

719
00:44:56,000 --> 00:45:01,000
So it is fascinating for me that language encodes so many truths about reality

720
00:45:01,000 --> 00:45:05,000
that you can extract almost things like emotion from language,

721
00:45:05,000 --> 00:45:08,000
sometimes even like how the world really works.

722
00:45:08,000 --> 00:45:13,000
And just to bring one example, just to drive home that point,

723
00:45:13,000 --> 00:45:16,000
because I think when you hear that, some people will think,

724
00:45:16,000 --> 00:45:18,000
oh, that's all we do as well.

725
00:45:18,000 --> 00:45:21,000
And that's completely untrue because when you start to learn to ride a bike,

726
00:45:21,000 --> 00:45:23,000
you don't use words.

727
00:45:23,000 --> 00:45:24,000
No, absolutely not.

728
00:45:24,000 --> 00:45:25,000
It's a different thing.

729
00:45:25,000 --> 00:45:29,000
Learning that 2 plus 2 equals 4 or the Eiffel Tower is in Paris,

730
00:45:29,000 --> 00:45:31,000
that's what we call propositional knowing.

731
00:45:31,000 --> 00:45:34,000
And there is other types of knowing, like knowing how to ride a bike,

732
00:45:34,000 --> 00:45:37,000
which is very, very different, which is something that LLMs can't do.

733
00:45:37,000 --> 00:45:43,000
Everyone who has a child understands this because you watch your child learning its motor skills

734
00:45:43,000 --> 00:45:50,000
and being able to control its limbs and balancing, throwing things, catching things.

735
00:45:50,000 --> 00:45:52,000
None of this has anything to do with words.

736
00:45:52,000 --> 00:45:56,000
So we know that there is a large category of knowledge and skills in the world

737
00:45:56,000 --> 00:46:00,000
that are important for the world, have nothing to do with mixed token prediction.

738
00:46:00,000 --> 00:46:03,000
And people plug LLMs into robots these days.

739
00:46:03,000 --> 00:46:08,000
Like there are robots where someone just basically pulls an LLM into a robot

740
00:46:08,000 --> 00:46:12,000
and says, I'll put this JSON if you want to move the arm to the left,

741
00:46:12,000 --> 00:46:15,000
I'll put that JSON if you want to move it to the right.

742
00:46:15,000 --> 00:46:18,000
And the LLM does something,

743
00:46:18,000 --> 00:46:22,000
but it has no conception of what the joints are of the robot,

744
00:46:22,000 --> 00:46:28,000
how things will react in your action, as a reaction to your action, and so on.

745
00:46:28,000 --> 00:46:35,000
So another category that a lot of people are focusing on right now is something called world models.

746
00:46:35,000 --> 00:46:40,000
And Jan LeCun is one of the proponents of this.

747
00:46:40,000 --> 00:46:44,000
And there is a model category called JEPA models,

748
00:46:45,000 --> 00:46:49,000
Joint Embedding Predictive Architecture, if I remember correctly.

749
00:46:49,000 --> 00:46:51,000
And they work on a completely different premise.

750
00:46:51,000 --> 00:46:54,000
And I think it's worth explaining here a little.

751
00:46:54,000 --> 00:47:00,000
So what a world model does, and you will hear that buzzword a lot these days.

752
00:47:00,000 --> 00:47:05,000
You just want to propose, just ignore it when you hear it in many contexts,

753
00:47:05,000 --> 00:47:07,000
because it's used wrongly.

754
00:47:07,000 --> 00:47:10,000
It has a specific academic meaning, and this is what it means.

755
00:47:10,000 --> 00:47:16,000
So a world model is essentially a model that is trained on action conditions.

756
00:47:16,000 --> 00:47:18,000
So what does it mean?

757
00:47:18,000 --> 00:47:22,000
The input to a world model is most of the time, for example, a video stream.

758
00:47:22,000 --> 00:47:25,000
So you can think of it, for example, a robot arm.

759
00:47:25,000 --> 00:47:30,000
And you film the robot arm over a long time, where it makes movements of the robot arm.

760
00:47:30,000 --> 00:47:34,000
You just film it with a camera, and you record the actions of the robot arm.

761
00:47:34,000 --> 00:47:39,000
So it says, joint goes to the left, upper arm goes up, and so on and so forth.

762
00:47:39,000 --> 00:47:46,000
And what you do is you train a neural network where you input the image plus the action that the robot does.

763
00:47:46,000 --> 00:47:50,000
And then you force it to predict the next image, so to speak.

764
00:47:50,000 --> 00:47:56,000
So to be correct, it's not the image, but it's an embedding of the image, like an abstract version of it.

765
00:47:56,000 --> 00:47:59,000
But that's not really the most important point here.

766
00:47:59,000 --> 00:48:07,000
The important point is that you train data plus action leads to next state of data.

767
00:48:07,000 --> 00:48:11,000
Or world, current world state plus action is the next world state.

768
00:48:11,000 --> 00:48:20,000
And when you force a neural network to learn to predict what the action in this current state will produce as a next state in the world,

769
00:48:20,000 --> 00:48:24,000
you force it to learn the underlying rules of the world.

770
00:48:24,000 --> 00:48:26,000
And that's why it's called a world model.

771
00:48:26,000 --> 00:48:30,000
So what people see is when you take the example of a robot arm,

772
00:48:30,000 --> 00:48:37,000
you embed the image of the video recording of the robot arm plus the joint actions.

773
00:48:37,000 --> 00:48:43,000
After doing this for a long time, you can find that the model starts to encode physical laws,

774
00:48:43,000 --> 00:48:48,000
like Newton's conservation of momentum, energy conservation.

775
00:48:48,000 --> 00:48:49,000
It encodes gravity.

776
00:48:49,000 --> 00:48:55,000
It starts coming up with the real physical laws of the world that it interacts in,

777
00:48:55,000 --> 00:48:59,000
because that's the only way to predict what the outcome of its action would be.

778
00:48:59,000 --> 00:49:06,000
And so this is super cool because it allows us to build AI systems that can actually plan ahead.

779
00:49:06,000 --> 00:49:12,000
So instead of having an LLM that needs to hallucinate what will happen when I move my arm,

780
00:49:12,000 --> 00:49:17,000
this is forced to predict what the outcome of the action will be.

781
00:49:17,000 --> 00:49:23,000
So this is super useful for robotics these days, for self-driving cars and so on.

782
00:49:23,000 --> 00:49:29,000
But I guess and I believe that we will make more and more use of this in other modalities,

783
00:49:29,000 --> 00:49:36,000
like maybe controlling your desktop computer or any kind of where the model is forced to make an action.

784
00:49:36,000 --> 00:49:38,000
And since we're talking about robotics,

785
00:49:38,000 --> 00:49:44,000
you also have an empirical way to see if your prediction turned out to be true or not,

786
00:49:44,000 --> 00:49:47,000
and you're actually in touch with reality.

787
00:49:47,000 --> 00:49:49,000
This is, I think, important to point out,

788
00:49:49,000 --> 00:49:56,000
because my main criticisms of LLMs right now, or one of them, is that they are two steps removed from reality.

789
00:49:56,000 --> 00:50:02,000
It's like they are working on data that is one step removed from reality itself, like written text.

790
00:50:02,000 --> 00:50:05,000
Human-generated text that was from observation of the world.

791
00:50:05,000 --> 00:50:09,000
Exactly. And then the models are actually just a bunch of numbers,

792
00:50:09,000 --> 00:50:11,000
which are again one step removed from the actual data.

793
00:50:11,000 --> 00:50:17,000
And so LLMs are two steps removed from reality and are actually not in touch in reality and cannot be.

794
00:50:17,000 --> 00:50:20,000
And that's kind of the point that the cognitive scientists have made.

795
00:50:20,000 --> 00:50:26,000
If we want to bridge that gap, we have to embody the cognition itself.

796
00:50:26,000 --> 00:50:32,000
We are embodied cognitive machines, basically, if you want to use that term.

797
00:50:32,000 --> 00:50:37,000
Beings, if you want to be less mechanical about it.

798
00:50:37,000 --> 00:50:44,000
And so we need to put brains into robots and let them be in touch with reality and learn about the world like children.

799
00:50:44,000 --> 00:50:47,000
Yes, this is how to do it. This is really how to do it.

800
00:50:47,000 --> 00:50:50,000
At least this is one of the ways how you can make that happen.

801
00:50:50,000 --> 00:50:53,000
And the nice thing about this is that the data is almost limitless,

802
00:50:53,000 --> 00:51:00,000
because right now we already reached a point with LLMs where the corpus, we've talked about this before, is already fully used.

803
00:51:00,000 --> 00:51:03,000
There is no more text in the world that hasn't been used for LLMs already.

804
00:51:03,000 --> 00:51:05,000
You have to make up new training data on the spot.

805
00:51:05,000 --> 00:51:09,000
Yes, and most of the text generated these days is AI-generated,

806
00:51:09,000 --> 00:51:16,000
and it's like a big, big no-no to put the AI-generated text back into the LLM, because that's going to just collapse everything.

807
00:51:16,000 --> 00:51:17,000
Yes, but that's what's happening.

808
00:51:17,000 --> 00:51:22,000
And it's also, just to make the point, because I put out on Nostra, like I just shitpost still all the time,

809
00:51:22,000 --> 00:51:27,000
and I wrote like over and over again, like, taste and values, taste and values,

810
00:51:27,000 --> 00:51:30,000
and like taste, taste, taste, taste, taste, taste and values.

811
00:51:30,000 --> 00:51:31,000
No one knew what it means.

812
00:51:31,000 --> 00:51:36,000
What it means is like what data to generate, what data to focus on depends on your taste and values.

813
00:51:37,000 --> 00:51:42,000
And like US versus China and so on, and like we talk about research taste, so it all leads back to taste and values.

814
00:51:42,000 --> 00:51:48,000
And the thing about reality itself is like, it's actually reality, your taste and values don't matter.

815
00:51:48,000 --> 00:51:50,000
So if you, you know what I'm trying to say?

816
00:51:50,000 --> 00:51:51,000
Right, yeah, absolutely.

817
00:51:51,000 --> 00:51:58,000
I mean, again, like a robot that is free to move in the world, for example, can just like move as long as the battery dies or the robot dies.

818
00:51:58,000 --> 00:52:00,000
You can learn from any experience.

819
00:52:00,000 --> 00:52:07,000
If you can observe what you're doing, and if you know that you're doing that, that is enough to train a world model.

820
00:52:07,000 --> 00:52:15,000
The cool thing about it is world models are like, I don't know the number, but I've trained them on my own laptop.

821
00:52:15,000 --> 00:52:18,000
So I know how hard it is to generate an LLM.

822
00:52:18,000 --> 00:52:22,000
It's like basically impossible to do something useful on your laptop with a single graphics card,

823
00:52:22,000 --> 00:52:25,000
but you can train a world model on a single graphics card, it works.

824
00:52:26,000 --> 00:52:33,000
So it opens, finally, it opens up a possibility of AI that can do lifelong learning.

825
00:52:33,000 --> 00:52:38,000
You could put it in a robot, you could put this machine that I just explained into a robot with a single GPU,

826
00:52:38,000 --> 00:52:43,000
and the robot would literally while, you know, it would be born stupid.

827
00:52:43,000 --> 00:52:51,000
And while moving around in the world, falling over, getting up again, trying to grab a glass, grabbing wrong, grabbing right,

828
00:52:51,000 --> 00:52:55,000
this time, slowly learn, oh, this is how gravity works.

829
00:52:55,000 --> 00:52:57,000
Oh, this is how my joints actually work.

830
00:52:57,000 --> 00:53:01,000
Every time I do this, then, you know, the glass actually can be lifted from my arm.

831
00:53:01,000 --> 00:53:03,000
So all these things, the robot would learn.

832
00:53:03,000 --> 00:53:12,000
And even, you know, once it breaks the arm, for example, the actuators change their properties during use because they get older and so on.

833
00:53:12,000 --> 00:53:17,000
It would be able to adapt in real time to the new conditions.

834
00:53:17,000 --> 00:53:20,000
And then getting back to the brain, this is how the brain works.

835
00:53:20,000 --> 00:53:23,000
So I think of this in a simplified view.

836
00:53:23,000 --> 00:53:25,000
The LLM is your prefrontal cortex.

837
00:53:25,000 --> 00:53:28,000
It is the thing that generates language, planning.

838
00:53:28,000 --> 00:53:30,000
It's really good at long-term horizon.

839
00:53:30,000 --> 00:53:36,000
It knows, like, if the human tells me, do this, then it can reason about, like, why would I do this?

840
00:53:36,000 --> 00:53:37,000
This is ethically good.

841
00:53:37,000 --> 00:53:42,000
Should I do blah, blah, blah, blah, all this, like a stream of consciousness that a robot can generate.

842
00:53:42,000 --> 00:53:46,000
But as soon as it needs to move around, it could activate something like a robot.

843
00:53:46,000 --> 00:53:47,000
It goes to the animal brain.

844
00:53:47,000 --> 00:53:53,000
And yeah, this is like the cerebellum in the back of your brain, which is essentially doing the same thing.

845
00:53:53,000 --> 00:53:56,000
If you try to grab a glass and you fail to do it.

846
00:53:56,000 --> 00:54:02,000
You know how you often catch something, like something falls down like a glass and you catch it before you realize you caught it?

847
00:54:02,000 --> 00:54:04,000
That's your animal brain at work.

848
00:54:04,000 --> 00:54:05,000
Perfect example.

849
00:54:05,000 --> 00:54:07,000
It is exactly how a world model works.

850
00:54:07,000 --> 00:54:12,000
You see the glass flying and what your brain does, it auto-completes the trajectory of the glass.

851
00:54:12,000 --> 00:54:14,000
It knows the physical laws of the world.

852
00:54:14,000 --> 00:54:17,000
It knows in a few milliseconds the glass will be around in my hand.

853
00:54:17,000 --> 00:54:20,000
I will probably put it here, close it in real time.

854
00:54:20,000 --> 00:54:23,000
And it does before your prefrontal cortex even caught up.

855
00:54:23,000 --> 00:54:24,000
It needs to do it faster.

856
00:54:24,000 --> 00:54:25,000
Exactly.

857
00:54:25,000 --> 00:54:29,000
Now, I have one more thing to say about the world model stuff and about the embodiment stuff.

858
00:54:29,000 --> 00:54:37,000
Because what will happen and what we know from the animal kingdom, for example, is like your world model will depend on your body shape.

859
00:54:37,000 --> 00:54:41,000
The world model of a spider is very different than the world model from a fly.

860
00:54:41,000 --> 00:54:43,000
It's very different than the world model from a bat.

861
00:54:43,000 --> 00:54:44,000
Absolutely.

862
00:54:44,000 --> 00:54:46,000
It's very different from the world model from a human.

863
00:54:46,000 --> 00:54:48,000
Or 20th century apparatus.

864
00:54:48,000 --> 00:54:49,000
100%.

865
00:54:49,000 --> 00:54:51,000
Two eyes, eight eyes.

866
00:54:51,000 --> 00:54:57,000
If you don't know what I'm talking about, just read like what is it like to be a bat by Thomas Nagel or something.

867
00:54:57,000 --> 00:54:58,000
Yes.

868
00:54:58,000 --> 00:54:59,000
However the guy was called.

869
00:54:59,000 --> 00:55:01,000
But we've known this for a long time.

870
00:55:01,000 --> 00:55:04,000
And just think about, let's say you want to learn about water.

871
00:55:04,000 --> 00:55:08,000
If you are a warship, your view of water will be very different.

872
00:55:08,000 --> 00:55:10,000
Or like a speedboat or whatever.

873
00:55:10,000 --> 00:55:11,000
Or like a submarine.

874
00:55:11,000 --> 00:55:12,000
It doesn't matter.

875
00:55:12,000 --> 00:55:18,000
If you put a brain into a submarine and let it learn what water is and how it behaves,

876
00:55:18,000 --> 00:55:28,000
it's very different from a tiny, tiny, tiny, like microscopic insect that lives inside like a droplet of water or on the outside of droplets of water.

877
00:55:28,000 --> 00:55:30,000
On the order of magnitude of water.

878
00:55:30,000 --> 00:55:31,000
Yes, exactly.

879
00:55:31,000 --> 00:55:35,000
Water tension suddenly becomes insanely important and all those kind of things.

880
00:55:35,000 --> 00:55:39,000
It's completely unimportant for like a submarine or like a tanker and so on.

881
00:55:40,000 --> 00:55:49,000
And I think that will be fascinating because I think it's unavoidable that we'll have humanoid robots.

882
00:55:49,000 --> 00:55:51,000
Like too many people are working on it.

883
00:55:51,000 --> 00:55:53,000
We'll have specialized robots as well.

884
00:55:53,000 --> 00:55:56,000
And we'll have robots that will probably look a little bit like spiders.

885
00:55:56,000 --> 00:55:57,000
Like window cleaning or whatever.

886
00:55:57,000 --> 00:55:58,000
Like I don't know what.

887
00:55:58,000 --> 00:56:00,000
And so they'll have different world models in terms of...

888
00:56:01,000 --> 00:56:13,000
So in the end what I'm trying to say is like you can't run away from values because basically like your shape dictates what you will have to value about the world in some sense.

889
00:56:13,000 --> 00:56:15,000
You know, what is actually important for you.

890
00:56:15,000 --> 00:56:16,000
Yes.

891
00:56:16,000 --> 00:56:17,000
Yeah.

892
00:56:17,000 --> 00:56:18,000
So that will be interesting.

893
00:56:18,000 --> 00:56:19,000
Will be interesting to see.

894
00:56:19,000 --> 00:56:21,000
But again, like back to cognitive science.

895
00:56:21,000 --> 00:56:25,000
Like all the 4E cognitive scientists people will tell you all about this.

896
00:56:26,000 --> 00:56:31,000
It's so funny that we're just basically recreating everything that we've learned about the brain in the last 20, 30 years, right?

897
00:56:31,000 --> 00:56:32,000
Yeah.

898
00:56:32,000 --> 00:56:33,000
I mean...

899
00:56:33,000 --> 00:56:34,000
It all comes back.

900
00:56:34,000 --> 00:56:38,000
And it all comes back to the inherent trade-offs of how to interact and be in touch with reality.

901
00:56:38,000 --> 00:56:40,000
And I think it's...

902
00:56:40,000 --> 00:56:42,000
I mean this will get very metaphysical.

903
00:56:43,000 --> 00:56:44,000
Let's go.

904
00:56:45,000 --> 00:56:48,000
But in the end you can't run away.

905
00:56:48,000 --> 00:56:49,000
There's a great book.

906
00:56:49,000 --> 00:56:50,000
It's called The Blind Spot.

907
00:56:51,000 --> 00:56:59,000
And so what's funny about physics right now, or it has been for a long while, we have two edges of physics, right?

908
00:57:00,000 --> 00:57:02,000
Quantum physics, relativity.

909
00:57:02,000 --> 00:57:05,000
Both depend to some degree on the observer.

910
00:57:05,000 --> 00:57:06,000
Yes.

911
00:57:06,000 --> 00:57:15,000
And so what this tells us, I'm like compressing a lot now, but our view of reality is out of necessity participatory.

912
00:57:16,000 --> 00:57:19,000
So you can't remove the observer and have a God's eye view of reality.

913
00:57:19,000 --> 00:57:21,000
And like it comes back to the world as well.

914
00:57:21,000 --> 00:57:24,000
It's like it depends on who you are, where you are, how fast you're going and so on.

915
00:57:24,000 --> 00:57:26,000
Observation is action.

916
00:57:26,000 --> 00:57:27,000
Yeah, exactly.

917
00:57:27,000 --> 00:57:29,000
You cannot look at something without touching it.

918
00:57:29,000 --> 00:57:30,000
Exactly.

919
00:57:30,000 --> 00:57:32,000
Your finger needs to push the thing in order to feel it.

920
00:57:32,000 --> 00:57:33,000
Exactly.

921
00:57:33,000 --> 00:57:38,000
And even the photon arriving in your eye was at some point reflected off an object.

922
00:57:38,000 --> 00:57:40,000
So that photon interacted with the object itself.

923
00:57:40,000 --> 00:57:43,000
And from the photon's point of view, zero time has passed.

924
00:57:44,000 --> 00:57:45,000
I don't know what that means.

925
00:57:45,000 --> 00:57:47,000
The photon has no time.

926
00:57:47,000 --> 00:57:49,000
It moves at the speed of light.

927
00:57:49,000 --> 00:57:50,000
I hope it doesn't have time.

928
00:57:50,000 --> 00:57:51,000
No, no.

929
00:57:51,000 --> 00:57:52,000
But that's what I mean.

930
00:57:52,000 --> 00:57:54,000
It's like from the point of view, zero time has passed.

931
00:57:54,000 --> 00:57:57,000
From the point of view of the photon, it just happened.

932
00:57:57,000 --> 00:57:58,000
That's true.

933
00:57:58,000 --> 00:57:59,000
Yeah.

934
00:57:59,000 --> 00:58:00,000
Yeah.

935
00:58:00,000 --> 00:58:02,000
Amazing, right?

936
00:58:02,000 --> 00:58:03,000
It all comes together.

937
00:58:03,000 --> 00:58:05,000
It all comes together.

938
00:58:05,000 --> 00:58:06,000
It's just incredible.

939
00:58:06,000 --> 00:58:07,000
Yeah.

940
00:58:07,000 --> 00:58:10,000
Yeah, and like you say, I mean, again, like evolution of the brain,

941
00:58:10,000 --> 00:58:14,000
like the evolutionary process that we are participating in figured all this out,

942
00:58:14,000 --> 00:58:16,000
and now we are rebuilding it.

943
00:58:16,000 --> 00:58:21,000
And also, like, you know, just to talk about kind of the current AI psychosis

944
00:58:21,000 --> 00:58:24,000
and the structures people build,

945
00:58:24,000 --> 00:58:29,000
organizations are cognitive machines that deal with these tradeoff relationships as well.

946
00:58:29,000 --> 00:58:32,000
And there's a reason why you have hierarchies and different responsibilities

947
00:58:32,000 --> 00:58:34,000
and all this kind of stuff.

948
00:58:35,000 --> 00:58:38,000
We are in the process of rebuilding those kind of hierarchies and systems

949
00:58:38,000 --> 00:58:39,000
just in an agentic world.

950
00:58:39,000 --> 00:58:41,000
I think that should be obvious to anyone.

951
00:58:41,000 --> 00:58:43,000
There is like this, you know, I ask myself,

952
00:58:43,000 --> 00:58:48,000
do we rebuild nature because it's the inspiration for all the things that we make,

953
00:58:48,000 --> 00:58:52,000
or, and, is nature just the best optimized system that we have?

954
00:58:52,000 --> 00:58:53,000
Yeah.

955
00:58:53,000 --> 00:58:54,000
Nature found what works.

956
00:58:54,000 --> 00:58:56,000
And it found the best solution already for us.

957
00:58:56,000 --> 00:58:57,000
However.

958
00:58:57,000 --> 00:58:58,000
So it could, like, converge to the same solution.

959
00:58:58,000 --> 00:59:03,000
However, there is a difference between a bird and a jet.

960
00:59:03,000 --> 00:59:05,000
There are two optimal ways to do things.

961
00:59:05,000 --> 00:59:09,000
They all use, they both figured out the same physical principles.

962
00:59:09,000 --> 00:59:11,000
But we still build humanoid bots, right?

963
00:59:11,000 --> 00:59:12,000
Yeah, yeah.

964
00:59:12,000 --> 00:59:13,000
We build bots in our image,

965
00:59:13,000 --> 00:59:17,000
whereas a bot with eight legs is probably the most optimal way to move.

966
00:59:17,000 --> 00:59:21,000
I think, I think there, like, science fiction is to blame for that to some degree.

967
00:59:21,000 --> 00:59:25,000
But to make a generous argument for all the road people out there,

968
00:59:25,000 --> 00:59:32,000
the world we have built and the, like, take Amazon warehouses or whatever.

969
00:59:32,000 --> 00:59:35,000
Like, the stuff that we've built is purpose built for humans, for humanoids.

970
00:59:35,000 --> 00:59:41,000
So it makes sense to replace the existing human, like, meat and skin robots

971
00:59:41,000 --> 00:59:45,000
that are in an Amazon warehouse with something that, like, doesn't need to go pee.

972
00:59:45,000 --> 00:59:46,000
You know?

973
00:59:46,000 --> 00:59:47,000
That's probably true.

974
00:59:47,000 --> 00:59:50,000
But I think we will, you know, see also many things that will surprise us.

975
00:59:50,000 --> 00:59:55,000
So we will see maybe in warfare that, you know, where this question about, like,

976
00:59:55,000 --> 00:59:58,000
you don't need any kind of connection to the thing.

977
00:59:58,000 --> 01:00:01,000
It doesn't need to fit in your humanoid world view or something like that.

978
01:00:01,000 --> 01:00:05,000
Whereas maybe the war machine will be something like a dog or like a spider.

979
01:00:05,000 --> 01:00:06,000
I mean, we see it already.

980
01:00:06,000 --> 01:00:07,000
Drones, right?

981
01:00:07,000 --> 01:00:08,000
Drone warfare.

982
01:00:08,000 --> 01:00:10,000
It's not a Terminator style robot that comes and kills you.

983
01:00:10,000 --> 01:00:12,000
It's a fucking suicide drone.

984
01:00:12,000 --> 01:00:15,000
So, like, we're living through that already.

985
01:00:15,000 --> 01:00:16,000
The last thing that you will see.

986
01:00:16,000 --> 01:00:17,000
Yeah, pretty much.

987
01:00:17,000 --> 01:00:18,000
And it's live stream.

988
01:00:18,000 --> 01:00:19,000
Great.

989
01:00:19,000 --> 01:00:20,000
Brave new world.

990
01:00:20,000 --> 01:00:21,000
Brave new world.

991
01:00:21,000 --> 01:00:23,000
Let's end on a positive note.

992
01:00:23,000 --> 01:00:24,000
Okay.

993
01:00:24,000 --> 01:00:25,000
Let's try.

994
01:00:25,000 --> 01:00:27,000
Let's try.

995
01:00:27,000 --> 01:00:31,000
So, what was the most hopeful thing outside?

996
01:00:31,000 --> 01:00:33,000
Like, we talked to a lot of people.

997
01:00:33,000 --> 01:00:36,000
We saw a lot of stuff the last 10 days or so.

998
01:00:36,000 --> 01:00:38,000
What is something that gives you hope for the future?

999
01:00:38,000 --> 01:00:40,000
I have my answer already.

1000
01:00:40,000 --> 01:00:43,000
Well, I'm ambivalent of my answer.

1001
01:00:43,000 --> 01:00:51,000
Because, like, what gives me the most hope is the uplifting of human spirit through these developments.

1002
01:00:51,000 --> 01:00:56,000
Right now in the world, there are more people than ever who realize that they can do more

1003
01:00:56,000 --> 01:00:57,000
than they thought that they can do.

1004
01:00:57,000 --> 01:01:03,000
And I think that's the biggest, like, human-level upgrade that we've seen for a long time.

1005
01:01:03,000 --> 01:01:09,000
It is, like, I've seen so many non-technical people who suddenly build things that I, as a coder,

1006
01:01:09,000 --> 01:01:12,000
even kind of, like, have a blockade in my head.

1007
01:01:12,000 --> 01:01:15,000
I'm, like, thinking, how would I even start building this project?

1008
01:01:15,000 --> 01:01:19,000
And I see, like, a non-coding person build this incredibly complex dashboard.

1009
01:01:19,000 --> 01:01:22,000
And in my head, it's like, wow, like, how?

1010
01:01:22,000 --> 01:01:23,000
How does this even work?

1011
01:01:23,000 --> 01:01:27,000
Why does it, you know, how can they keep working on it after three months?

1012
01:01:27,000 --> 01:01:31,000
And this is not complete spaghetti and, like, blah, blah, blah, all these things that I thought, like,

1013
01:01:31,000 --> 01:01:33,000
you need a computer science degree and blah, blah, blah.

1014
01:01:33,000 --> 01:01:35,000
No, it's just the willpower.

1015
01:01:35,000 --> 01:01:38,000
So for high-agency people, it's the best time to be alive.

1016
01:01:38,000 --> 01:01:39,000
It is incredible.

1017
01:01:39,000 --> 01:01:44,000
And I think that is something that, you know, I hope maybe not in terms of GDP impacts that we will see,

1018
01:01:44,000 --> 01:01:49,000
but probably we will, but just in terms of how much you believe that you can do.

1019
01:01:49,000 --> 01:01:54,000
And I think that's one of the most powerful things because we need to surprise ourselves with the things that we can do.

1020
01:01:54,000 --> 01:01:59,000
So that is the great equalizer of the abilities.

1021
01:01:59,000 --> 01:02:01,000
And that gives me incredible hope.

1022
01:02:01,000 --> 01:02:11,000
I'm just so happy that many, many more people are joining the high-agency kind of pack of people who know that they can change the world.

1023
01:02:11,000 --> 01:02:13,000
Yeah, you can just do things.

1024
01:02:13,000 --> 01:02:21,000
And you have, like, an uncountable amount of superpowers at your fingertips right now.

1025
01:02:21,000 --> 01:02:22,000
It is just incredible.

1026
01:02:22,000 --> 01:02:23,000
Yeah.

1027
01:02:23,000 --> 01:02:25,000
I'll give you my answer briefly.

1028
01:02:25,000 --> 01:02:31,000
So, obviously, I have to say Bitcoin and I have to say Nostra, but it's literally true.

1029
01:02:31,000 --> 01:02:33,000
It gives me a ton of hope.

1030
01:02:33,000 --> 01:02:37,000
I know everyone's bearish and depressed, but that's just, like, because everyone's retarded.

1031
01:02:38,000 --> 01:02:39,000
It's amazing.

1032
01:02:39,000 --> 01:02:45,000
It's amazing that Bitcoin exists and is alive and well and works and scales and does all the things.

1033
01:02:45,000 --> 01:02:53,000
And that we have it and that it provides, you know, for me, it's still, it's the thing.

1034
01:02:53,000 --> 01:03:04,000
For me, even with all the AI hype, to me, Bitcoin is still the most important thing because what we are facing is lots of our troubles are of economic nature.

1035
01:03:04,000 --> 01:03:05,000
Yes.

1036
01:03:05,000 --> 01:03:06,000
And I think Bitcoin is literally the antidote.

1037
01:03:06,000 --> 01:03:09,000
Like, I've always thought that, still think that.

1038
01:03:09,000 --> 01:03:10,000
It gives me a lot of hope.

1039
01:03:10,000 --> 01:03:11,000
It's still around.

1040
01:03:11,000 --> 01:03:12,000
It won't go anywhere.

1041
01:03:12,000 --> 01:03:16,000
I don't think it has any big issues or big problems or anything like that.

1042
01:03:16,000 --> 01:03:20,000
On Nostra, obviously, like, we talk a lot about it.

1043
01:03:20,000 --> 01:03:24,000
I just told you before we started recording my setup.

1044
01:03:24,000 --> 01:03:26,000
Like, NIP17 only.

1045
01:03:26,000 --> 01:03:27,000
Every agent just has an MPAP.

1046
01:03:27,000 --> 01:03:28,000
I don't use anything else.

1047
01:03:28,000 --> 01:03:31,000
I'm in a fully permissionless world.

1048
01:03:32,000 --> 01:03:35,000
This podcast goes right through your agents on Nostra.

1049
01:03:35,000 --> 01:03:36,000
It will be uploaded automatically.

1050
01:03:36,000 --> 01:03:37,000
Exactly.

1051
01:03:37,000 --> 01:03:38,000
Exactly.

1052
01:03:38,000 --> 01:03:42,000
Like, yeah, that's just like all this stuff already works.

1053
01:03:42,000 --> 01:03:46,000
And on the AI side, shout out to Austin.

1054
01:03:46,000 --> 01:03:51,000
He did a lot of measurements and research on this and explained this really well to me.

1055
01:03:51,000 --> 01:03:57,000
The local models are actually catching up, getting better, getting smaller.

1056
01:03:57,000 --> 01:04:02,000
It gives me so much hope because I was a bit of an AI doomer in the beginning

1057
01:04:02,000 --> 01:04:04,000
because I was like, okay, everything was centralized.

1058
01:04:04,000 --> 01:04:05,000
Everything would be terrible.

1059
01:04:05,000 --> 01:04:06,000
Maybe let's walk up here.

1060
01:04:06,000 --> 01:04:08,000
It might be a little bit more quiet.

1061
01:04:08,000 --> 01:04:09,000
Let's see.

1062
01:04:09,000 --> 01:04:10,000
Let's end up there.

1063
01:04:10,000 --> 01:04:11,000
Okay.

1064
01:04:11,000 --> 01:04:17,000
And it's just like, if you look at the graphs, where things are going,

1065
01:04:17,000 --> 01:04:22,000
every single fucking smartphone, even old hardware potentially,

1066
01:04:22,000 --> 01:04:27,000
will have extremely smart, extremely small models that are extremely powerful.

1067
01:04:27,000 --> 01:04:28,000
That's where things are trending.

1068
01:04:28,000 --> 01:04:30,000
It's undeniable.

1069
01:04:30,000 --> 01:04:34,000
And that just gives me all the hope in the world because it's like this stuff will just be free

1070
01:04:34,000 --> 01:04:39,000
and available to basically anyone who cares to use it.

1071
01:04:39,000 --> 01:04:41,000
So, it's amazing.

1072
01:04:41,000 --> 01:04:42,000
Best time to be alive.

1073
01:04:42,000 --> 01:04:43,000
Best time to be alive.

1074
01:04:43,000 --> 01:04:46,000
Once in a lifetime.

1075
01:04:46,000 --> 01:04:47,000
Yeah, let's see.

1076
01:04:47,000 --> 01:04:48,000
Let's see.

1077
01:04:48,000 --> 01:04:49,000
Let's not jinx it.

1078
01:04:49,000 --> 01:04:52,000
Let's see where things are in six months from now.

1079
01:04:52,000 --> 01:04:53,000
It's weird times.

1080
01:04:53,000 --> 01:04:54,000
It's weird times.

1081
01:04:54,000 --> 01:04:59,000
So, the one prediction I can make is I'm sure it will remain to be interesting.

1082
01:04:59,000 --> 01:05:03,000
I'm sure that volatility will come in all sorts of ways.

1083
01:05:03,000 --> 01:05:11,000
And, yeah, I'm sure that, or I hope, I hope that we'll do this again sometime soon, next time I see you.

1084
01:05:11,000 --> 01:05:13,000
Let's see where cashier is.

1085
01:05:13,000 --> 01:05:14,000
Thank you, Gigi.

1086
01:05:14,000 --> 01:05:15,000
I appreciate it as well.

1087
01:05:15,000 --> 01:05:16,000
Big hug.

1088
01:05:16,000 --> 01:05:18,000
Big hug.

1089
01:05:18,000 --> 01:05:19,000
Thanks for making the time.

1090
01:05:19,000 --> 01:05:20,000
We'll go sleep now.

1091
01:05:20,000 --> 01:05:22,000
I'm still sleep deprived as fuck.

1092
01:05:22,000 --> 01:05:27,000
It's dead bright here outside in the middle of the night.

1093
01:05:27,000 --> 01:05:28,000
Bye.

1094
01:05:28,000 --> 01:05:31,000
Yeah, well, no wonder I'm tired.

1095
01:05:31,000 --> 01:05:33,000
Guess how late it is.

1096
01:05:33,000 --> 01:05:35,000
Oh.

1097
01:05:35,000 --> 01:05:37,000
Let's call it 1.21.

1098
01:05:37,000 --> 01:05:38,000
1.21.

1099
01:05:38,000 --> 01:05:39,000
Good night, everyone.

1100
01:05:39,000 --> 01:05:40,000
Good night.