034 Matt Corallo: Bitcoin Has One Shot at the AI Economy - Here's How to Win I All in Bitcoin w/CK

1
00:00:00,000 --> 00:00:06,800
in disabling elliptic curve cryptography yes you burn some coins but you also allow many other
2
00:00:06,800 --> 00:00:13,760
people to retain ownership by doing proof of seed phrase proof of bib 32 derivation like if you can
3
00:00:13,760 --> 00:00:19,200
prove you know the seed phrase you don't actually need ec that's that's quantum security look i got
4
00:00:19,200 --> 00:00:24,480
my agent to do this thing and it's really amazing and it like wrote all this awesome code then you
5
00:00:24,480 --> 00:00:29,520
try it it doesn't what happened if we had a remote code execution vulnerability in bitcoin core
6
00:00:29,520 --> 00:00:32,840
where people could hack into your node remotely and steal your private keys.
7
00:00:32,920 --> 00:00:37,280
If all the private keys to all the wallets leaked, there's no recovery from that.
8
00:00:37,940 --> 00:00:40,200
Bitcoin has never had an opportunity like this.
9
00:00:40,280 --> 00:00:44,320
If AI agents start using Bitcoin or maybe using stable coins, it owes taxes.
10
00:00:45,040 --> 00:00:45,760
But how?
11
00:00:46,160 --> 00:00:49,700
All views expressed by the host or guests are solely their own opinions.
12
00:00:49,700 --> 00:00:52,980
Nothing stated in this podcast should be considered a specific endorsement
13
00:00:52,980 --> 00:00:55,880
to make any particular investment or follow any specific strategy.
14
00:00:59,520 --> 00:01:09,380
Welcome to All in Bitcoin, Matt.
15
00:01:09,540 --> 00:01:14,900
Today, we'll find out exactly what's at stake when AI starts using money.
16
00:01:15,300 --> 00:01:15,600
Okay.
17
00:01:15,840 --> 00:01:18,080
But first, let's get to know you a little bit better.
18
00:01:18,320 --> 00:01:18,520
Ready?
19
00:01:18,980 --> 00:01:19,220
Sure.
20
00:01:19,660 --> 00:01:22,180
You've been working on Bitcoin for 15 years.
21
00:01:22,380 --> 00:01:24,080
What kept you here this long?
22
00:01:24,640 --> 00:01:25,580
Good question.
23
00:01:25,780 --> 00:01:28,200
I feel like there's always more to do.
24
00:01:28,200 --> 00:01:35,800
So, you know, we keep making it's interesting because 15 years ago, I never would have imagined that we got this far.
25
00:01:36,120 --> 00:01:41,940
Right. That Bitcoin is a topic in international politics and international relations and national politics in the US.
26
00:01:43,400 --> 00:01:56,280
But at the same time, every time we make one step forward, there's always there's always just feels like so many different things that we still need to do to to really get the user experience we want or the, you know, whatever properties we want.
27
00:01:56,280 --> 00:01:59,820
And so it just makes it feel like my work is never done.
28
00:02:00,400 --> 00:02:01,800
So what pulled you into it?
29
00:02:02,480 --> 00:02:06,120
So, yeah, I got into Bitcoin in 2011.
30
00:02:06,320 --> 00:02:15,360
I saw a podcast at the time that did an hour and a half long deep dive on explaining how Bitcoin worked in 2011.
31
00:02:15,840 --> 00:02:17,560
It was a podcast I watch regularly anyway.
32
00:02:17,840 --> 00:02:18,200
Which one?
33
00:02:19,460 --> 00:02:21,720
It was a podcast by the name of Security Now.
34
00:02:22,280 --> 00:02:24,080
It's in the Twit network.
35
00:02:24,080 --> 00:02:26,100
I have no idea if it's still going.
36
00:02:26,260 --> 00:02:27,400
I haven't watched it in a long time.
37
00:02:27,640 --> 00:02:32,760
But yeah, in 2011, they did an hour and change long deep dive explaining Bitcoin.
38
00:02:32,940 --> 00:02:37,540
And I just found it really interesting that there's this digital money, like online money.
39
00:02:37,620 --> 00:02:40,060
I spent a lot of time online, like many teenagers.
40
00:02:40,340 --> 00:02:48,600
And we just found this fascinating that we could have this online value system where we could transfer between each other and no one else was involved.
41
00:02:48,600 --> 00:02:51,320
And it was just entirely within the nodes of the network.
42
00:02:51,320 --> 00:02:54,920
and that we could build such a thing was fascinating.
43
00:02:55,120 --> 00:02:58,160
And so I just kind of kept looking more into it and more into it.
44
00:02:58,360 --> 00:03:03,220
And we ended up, you know, I ended up working on it and contributing to it.
45
00:03:03,340 --> 00:03:05,360
And it all kind of snowballed from there.
46
00:03:06,060 --> 00:03:07,120
How old were you at the time?
47
00:03:09,080 --> 00:03:12,880
18-ish or 19 or 17 or something like that.
48
00:03:13,060 --> 00:03:14,140
I was interested in computers.
49
00:03:14,140 --> 00:03:16,120
I knew some of how to program.
50
00:03:16,440 --> 00:03:18,580
I didn't know much about C or C++.
51
00:03:18,780 --> 00:03:20,160
I didn't really know what a pointer was.
52
00:03:20,160 --> 00:03:24,740
but I knew some of how to program
53
00:03:24,740 --> 00:03:25,680
and spent a lot of time online
54
00:03:25,680 --> 00:03:27,740
and I was very interested in computers
55
00:03:27,740 --> 00:03:29,760
so I found this thing fascinating
56
00:03:29,760 --> 00:03:32,360
and ended up learning to code more on Bitcoin
57
00:03:32,360 --> 00:03:33,420
as time went on.
58
00:03:34,020 --> 00:03:36,300
Do you question Bitcoin sometimes?
59
00:03:36,980 --> 00:03:37,820
At high level, yes.
60
00:03:37,1000 --> 00:03:39,140
Yes, I question,
61
00:03:39,840 --> 00:03:42,780
are we succeeding in our goals?
62
00:03:43,080 --> 00:03:45,060
Is there something that'll go wrong?
63
00:03:45,180 --> 00:03:47,060
I mean, I've spent however long,
64
00:03:47,260 --> 00:03:49,420
15 years and change in Bitcoin
65
00:03:49,420 --> 00:03:55,580
most of that time has been worried about this going wrong or that going wrong.
66
00:03:55,580 --> 00:04:03,280
I'm somewhat infamous for always assigning a low probability chance that I think Bitcoin will succeed.
67
00:04:03,980 --> 00:04:10,120
And the goalposts always shift in my view of what success means and where Bitcoin should get to always changes.
68
00:04:11,440 --> 00:04:15,940
But I always kind of, my job is to worry about all the things that could go wrong.
69
00:04:16,160 --> 00:04:18,620
And so, yes, I'm always worried about all the things that could go wrong.
70
00:04:19,420 --> 00:04:23,840
rather than to just be very bullish and only talk about the things that could go right.
71
00:04:24,100 --> 00:04:26,260
So right now, what are your worries?
72
00:04:26,940 --> 00:04:31,120
My concerns for the last 10 years or something has been mining decentralization
73
00:04:31,120 --> 00:04:35,640
and ability of self-custody to be good user experience.
74
00:04:35,800 --> 00:04:39,480
Like, can we build a good user experience around self-custody
75
00:04:39,480 --> 00:04:44,380
so that people choose to self-custody instead of using a trusted intermediary?
76
00:04:46,120 --> 00:04:49,140
And so those have been kind of the biggest things I've worried about.
77
00:04:49,140 --> 00:04:53,940
But obviously there's plenty of other things to worry about and to think on and work through.
78
00:04:54,280 --> 00:04:57,300
Can we upgrade post-quantum signatures in time?
79
00:04:57,540 --> 00:05:00,440
And then what will Bitcoin do when a quantum computer eventually exists?
80
00:05:00,700 --> 00:05:04,100
I don't think it's happening soon, but at some point maybe it will.
81
00:05:05,220 --> 00:05:07,020
What should we be doing now?
82
00:05:07,300 --> 00:05:09,040
What concerns will we have then?
83
00:05:09,640 --> 00:05:13,020
There's always a million things that could go wrong with Bitcoin that I'm worried about.
84
00:05:13,020 --> 00:05:19,820
One of mine is Bitcoin centralization in the hands of institutions.
85
00:05:20,220 --> 00:05:21,360
What do you think about that?
86
00:05:21,360 --> 00:05:23,340
Yeah, no, I mean, I think that's a big concern too.
87
00:05:23,460 --> 00:05:23,760
Absolutely.
88
00:05:24,360 --> 00:05:28,480
Not only single institutions that hold a lot of Bitcoin and might have power over Bitcoin,
89
00:05:28,600 --> 00:05:32,560
but also that all the ETFs use Coinbase, the same custodian.
90
00:05:33,060 --> 00:05:38,320
So there's one pot of money, there's one wallet system that stores, you know, obviously they
91
00:05:38,320 --> 00:05:42,020
have different keys, but one wallet system that stores all that Bitcoin, you know, what
92
00:05:42,020 --> 00:05:47,120
happens if they get hacked. That was, you know, big concern in, in going back to 2011. Something
93
00:05:47,120 --> 00:05:51,540
I always worried about is, you know, everyone used the Bitcoin core wallet. There was no other wallet.
94
00:05:51,640 --> 00:05:55,800
It was just the Bitcoin client. There was, it wasn't even called Bitcoin core. Right. And at the
95
00:05:55,800 --> 00:06:01,120
time, most, a lot of these people were mining all the time and were leaving their nodes on all the
96
00:06:01,120 --> 00:06:05,700
time. What happened if we had a remote code execution vulnerability in Bitcoin core, where
97
00:06:05,700 --> 00:06:09,800
people could hack into your node remotely and steal your private keys. If all the private keys
98
00:06:09,800 --> 00:06:13,200
all the wallets leaked, there's no recovery from that.
99
00:06:13,720 --> 00:06:18,860
And so you could imagine similar situations with like if Coinbase, you know, I'm sure
100
00:06:18,860 --> 00:06:22,980
I know some of the folks who work at Coinbase, but like I know they did a really great job
101
00:06:22,980 --> 00:06:23,720
with their custody setup.
102
00:06:23,780 --> 00:06:26,660
But what happens if, you know, everyone involved gets kidnapped?
103
00:06:26,840 --> 00:06:30,400
Like the entire board of Coinbase and all of the security officers all get kidnapped
104
00:06:30,400 --> 00:06:32,540
or their wives or kids get kidnapped.
105
00:06:33,080 --> 00:06:34,540
What is the worst case scenario here?
106
00:06:34,580 --> 00:06:35,680
And does that impact Bitcoin?
107
00:06:35,800 --> 00:06:38,340
I think that's definitely a very material concern.
108
00:06:38,340 --> 00:06:41,040
What's a signal you ignore but shouldn't?
109
00:06:41,040 --> 00:06:52,420
Because I spend all my time worried about the ways Bitcoin could go wrong, I have gotten very good at ignoring that intuition.
110
00:06:52,660 --> 00:06:56,220
My intuition is always that it's going to go wrong, which is the wrong intuition.
111
00:06:56,660 --> 00:07:01,400
But I've gotten good at ignoring that intuition because I spend all my time worried about that.
112
00:07:01,520 --> 00:07:08,540
And so I think finding the right balance of that signal is always hard for me.
113
00:07:08,540 --> 00:07:12,040
Do you trust your instincts more than your reasoning?
114
00:07:14,040 --> 00:07:17,660
I think sometimes.
115
00:07:17,980 --> 00:07:28,360
I think I mentioned earlier that a lot of my instinct is always to, and this has been my job for the last 15 years, but to worry about ways Bitcoin can go wrong.
116
00:07:28,360 --> 00:07:32,620
And so that builds a certain instinct that I have to suppress sometimes.
117
00:07:32,620 --> 00:08:02,600
But I think in general, I balance the two fairly well. And more often than not, my reasoning will match my instinct. Not always. And there's actually interesting behavioral studies on this where some of the smartest people will have some of the most irrational positions because they're better at coming up with arguments to support the position they want to get to than someone who's maybe
118
00:08:02,600 --> 00:08:09,260
not as smart. And so it's very easy for someone to come up with arguments that support their
119
00:08:09,260 --> 00:08:14,120
position. And so you have to be a little careful because, you know, in saying that my reasoning
120
00:08:14,120 --> 00:08:18,960
often matches my instinct, I might also just be saying that I come up with arguments to support
121
00:08:18,960 --> 00:08:24,800
my instinct. But I think I try not to do that too much. But maybe that's what I'm saying. I'm not
122
00:08:24,800 --> 00:08:29,300
sure. What's something about Lightning that still frustrates you today?
123
00:08:29,300 --> 00:08:36,660
oh uh what doesn't frustrate me um yeah i mean i think the the story of my involvement with
124
00:08:36,660 --> 00:08:44,700
lightning for the last five six years has been really one of how do we make self-custody the
125
00:08:44,700 --> 00:08:50,980
thing that people want to use with lightning um and how do we make that accessible to many wallets
126
00:08:50,980 --> 00:08:54,920
right so like i think phoenix is great like they do a really good job but you know they're they're
127
00:08:54,920 --> 00:08:59,660
one wallet and like other wallets don't use similar models and other wallets, there's not
128
00:08:59,660 --> 00:09:06,080
competition around LSP pricing and fee pricing because, you know, they built their kind of
129
00:09:06,080 --> 00:09:12,740
vertically integrated stack. But how do we build a software suite that enables everyone to do that
130
00:09:12,740 --> 00:09:18,360
kind of wallet, but also then better, like fix all the warts that Lightning has and that Phoenix
131
00:09:18,360 --> 00:09:21,880
has also fixed over the years. You know, we've been doing some of the same work that they have
132
00:09:21,880 --> 00:09:29,500
been. But enabling that and making it broadly available so that self-custody lightning is a
133
00:09:29,500 --> 00:09:34,500
default. Because the history of lightning is moved towards custodial lightning. Everyone's doing
134
00:09:34,500 --> 00:09:40,400
custodians because people wrote off lightning as too hard to do on a mobile phone or in a self-custodial
135
00:09:40,400 --> 00:09:45,980
wallet. And I don't think that's true, but that certainly was true. L&D is not designed to run
136
00:09:45,980 --> 00:09:51,940
on phones it doesn't run well on phones it makes your phone hot um and people early using ldk
137
00:09:51,940 --> 00:09:57,780
suffered lots of issues and and ux warts and whatever and so i don't think it's true anymore
138
00:09:57,780 --> 00:10:03,020
but it is still really hard to do self-custodial lightning and and no one people don't do it and
139
00:10:03,020 --> 00:10:09,760
so people default to custodial and i think that's really frustrating the hardest no you've ever had
140
00:10:09,760 --> 00:10:19,180
to say oh nothing really comes to mind that's super compelling but like i know you know when i
141
00:10:19,180 --> 00:10:24,840
left chain code they were they wanted me to you know they wanted to figure out how to how to keep
142
00:10:24,840 --> 00:10:31,360
me around and and at the time i was starting to date my now wife remotely and i i needed that
143
00:10:31,360 --> 00:10:36,840
remote job so uh that was that was hard to to leave there because the the chain code environment
144
00:10:36,840 --> 00:10:40,380
so great and the founders are great and the environment's great.
145
00:10:40,500 --> 00:10:45,640
But also I needed a job that let me be remote so I could date my now wife.
146
00:10:46,640 --> 00:10:49,720
Looking back, that was the right decision, but it was hard.
147
00:10:50,740 --> 00:10:53,760
Why do you think Satoshi didn't burn his coins?
148
00:10:55,320 --> 00:10:57,580
Oh, yeah, that's another really good question.
149
00:10:57,740 --> 00:10:58,340
I don't know.
150
00:10:58,340 --> 00:11:09,580
um my assumption is that satoshi uh mined the coins didn't think they were necessarily going
151
00:11:09,580 --> 00:11:18,220
to spend them and then either like lost the hard drive or passed away or um just for whatever
152
00:11:18,220 --> 00:11:25,980
reason didn't have access to those coins anymore and you know in 2010 2011 2012 2013 it maybe
153
00:11:25,980 --> 00:11:33,800
didn't matter very much and then now however 15 20 years later it does matter more but now they no
154
00:11:33,800 --> 00:11:41,900
longer can would be my assumption what do you think about jameson lob's idea that if they are
155
00:11:41,900 --> 00:11:49,980
not taught in five years burn them are you familiar yeah yeah they they have this uh 361
156
00:11:49,980 --> 00:11:54,360
I think is their proposal to, to disable EC and burn old coins.
157
00:11:54,860 --> 00:11:55,340
361.
158
00:11:55,480 --> 00:11:55,680
Yeah.
159
00:11:55,800 --> 00:11:56,060
Okay.
160
00:11:56,200 --> 00:11:56,360
Yeah.
161
00:11:56,500 --> 00:12:02,260
Um, I think that I don't think it makes sense to do it based on time.
162
00:12:02,440 --> 00:12:04,180
I don't see any reason for that.
163
00:12:04,260 --> 00:12:12,420
I think, um, I firmly believe that there's no reason to think that the progress of quantum
164
00:12:12,420 --> 00:12:17,400
computers will shift, uh, to private that like the U S government's going to come in
165
00:12:17,400 --> 00:12:22,140
and acquire all these quantum labs and forcibly acquire and pay, you know, billions and billions
166
00:12:22,140 --> 00:12:27,960
and billions of dollars to do so just to keep it secret. And then they're going to go steal
167
00:12:27,960 --> 00:12:35,220
the Bitcoin or something. I don't think that is realistic. Instead, I think that the progress will
168
00:12:35,220 --> 00:12:39,880
continue in public. We'll all see the progress. We'll all be relatively aware of it. And there'll
169
00:12:39,880 --> 00:12:47,380
be the Bitcoin community will all have somewhat the same view of where quantum is once it gets
170
00:12:47,380 --> 00:12:49,540
Obviously today, some people think it's never going to happen.
171
00:12:49,540 --> 00:13:09,248
Some people think it two years away but like you know when it six months away I think we all roughly agree that it somewhere between six and nine months away And so to get back to your question because of that I don see a reason for the Bitcoin community to materially entertain freezing coins before that point
172
00:13:09,488 --> 00:13:11,328
That once we're at that point
173
00:13:11,328 --> 00:13:13,768
and everyone kind of agrees that like it's now or never,
174
00:13:14,348 --> 00:13:17,808
then that's when that conversation has to happen
175
00:13:17,808 --> 00:13:18,928
and that decision has to be made.
176
00:13:19,788 --> 00:13:24,228
But until then, I don't see any reason to engage with it
177
00:13:24,228 --> 00:13:30,288
And as to whether or not Bitcoin should freeze the coins at that point or not, you know, I think it's decided by the market.
178
00:13:30,288 --> 00:13:35,968
Like there will be a fork and there will be two sides to the fork and the markets can have to pick which one has more value.
179
00:13:37,168 --> 00:13:44,248
But I personally think that the market will probably pick the side that burns the coins because I think it's also a long ways away.
180
00:13:44,248 --> 00:13:51,408
that like by that, by that point, everyone has had a chance to migrate and in disabling
181
00:13:51,408 --> 00:13:56,248
elliptic curve cryptography, yes, you burn some coins, but you also allow many other
182
00:13:56,248 --> 00:14:02,508
people to retain ownership by doing proof of seed phrase, proof of BIP32 derivation.
183
00:14:02,768 --> 00:14:06,188
Like if you can prove, you know, the seed phrase, you don't actually need EC.
184
00:14:06,288 --> 00:14:07,308
That's, that's quantum secure.
185
00:14:07,808 --> 00:14:12,148
A seed phrase derivation from a seed phrase to public key is quantum secure.
186
00:14:12,148 --> 00:14:15,728
and so you can prove you know the seed phrase
187
00:14:15,728 --> 00:14:17,108
and you could use that to spend the coins
188
00:14:17,108 --> 00:14:19,008
but it's a package deal
189
00:14:19,008 --> 00:14:21,948
you either enable people to
190
00:14:21,948 --> 00:14:23,508
prove they know the seed phrase
191
00:14:23,508 --> 00:14:25,008
or BIP32 for derivation
192
00:14:25,008 --> 00:14:26,428
or pre-committ
193
00:14:26,428 --> 00:14:27,828
there's a few things you would do
194
00:14:27,828 --> 00:14:29,268
to allow people to retain ownership
195
00:14:29,268 --> 00:14:32,508
and then you also burn some other coins
196
00:14:32,508 --> 00:14:33,968
anyone who doesn't have these things
197
00:14:33,968 --> 00:14:34,848
or do these things
198
00:14:34,848 --> 00:14:36,028
would get their currency burned
199
00:14:36,028 --> 00:14:37,968
or you do none of that
200
00:14:37,968 --> 00:14:39,888
and then everyone gets their coin stolen
201
00:14:39,888 --> 00:14:45,268
right and so it's binary you have to pick one either everyone gets their coins stolen or some
202
00:14:45,268 --> 00:14:51,828
people retain their coins and some people get burned and i think that the bitcoin community
203
00:14:51,828 --> 00:14:57,908
will choose option a and burn some coins but some people get to retain their coins because i think
204
00:14:57,908 --> 00:15:03,028
it'll be a long ways away and at that point people will have had so long to upgrade that the number
205
00:15:03,028 --> 00:15:10,108
of coins that will be lost in the burn will be minuscule. And the number of coins that are
206
00:15:10,108 --> 00:15:16,468
retained in enabling people with seed phrases or whatever to reclaim their funds would be,
207
00:15:16,568 --> 00:15:20,888
you know, also not huge, but much larger, much, much, much larger.
208
00:15:20,888 --> 00:15:25,608
This episode is brought to you by Bitbox, the Swiss built Bitcoin hardware wallet that
209
00:15:25,608 --> 00:15:31,368
guards your sovereignty. I use Bitbox because it's fully open source, easy to use, and I can
210
00:15:31,368 --> 00:15:36,328
plug it into my phone. If you need a hardware wallet to keep your Bitcoin off exchanges,
211
00:15:36,808 --> 00:15:43,528
offline and safe, go to bitbox.swiss slash allinbitcoin for 5% off or use the code
212
00:15:43,528 --> 00:15:50,768
allinbitcoin or check the show notes. Join me at BTC Health 2026, the Nordic Bitcoin
213
00:15:50,768 --> 00:15:59,188
conference in Helsinki this September 25th, 26th. Use code allinbitcoin for 10% of your ticket.
214
00:15:59,188 --> 00:16:00,228
See you there.
215
00:16:01,368 --> 00:16:03,608
Guys, I've got a quick favor to ask.
216
00:16:03,968 --> 00:16:09,328
If you are listening right now, take a second and hit subscribe and the bell if you are on YouTube.
217
00:16:09,688 --> 00:16:14,368
Or follow on your podcast app so you don't miss the episodes and guests that matter to you.
218
00:16:14,768 --> 00:16:17,648
If we've ever given you value, this is the moment to give it back.
219
00:16:17,828 --> 00:16:23,128
It's free, it takes two seconds, and it genuinely helps keep improving every single episode.
220
00:16:23,708 --> 00:16:25,948
And to everyone already following, thank you.
221
00:16:26,068 --> 00:16:28,528
You are the reason we've been able to raise the bar.
222
00:16:28,528 --> 00:16:34,268
Your support, whether it's a donation, a comment or subscribing, helps these conversations reach further.
223
00:16:34,608 --> 00:16:41,208
It also gives us the push to keep going, keep traveling, keep preparing and keep bringing you in-person conversations like this.
224
00:16:41,428 --> 00:16:41,928
Thank you.
225
00:16:42,488 --> 00:16:45,628
Who or what do you watch closely these days?
226
00:16:46,708 --> 00:16:52,168
There's not a lot I watch super closely in terms of Bitcoin stuff.
227
00:16:52,168 --> 00:17:03,328
I try to keep up on the news and my screen time on news apps tends to be an hour a day or more.
228
00:17:03,928 --> 00:17:09,268
So I do try to keep up with many perspectives on what happens in the world.
229
00:17:09,648 --> 00:17:11,408
So I guess that's probably my answer.
230
00:17:11,628 --> 00:17:17,108
But on the Bitcoin side, it's mostly I just don't have the time to keep up with almost anything.
231
00:17:18,268 --> 00:17:21,148
I don't have the time to read the mailing list or delving or anything.
232
00:17:21,148 --> 00:17:25,808
So it's hard, aside from what people send to me.
233
00:17:26,728 --> 00:17:29,888
What's something most people misunderstand about you?
234
00:17:33,188 --> 00:17:39,148
Well, these days, the most obvious one is that people assume a lot about my politics.
235
00:17:41,248 --> 00:17:41,448
Which are?
236
00:17:41,448 --> 00:18:08,448
Well, personally, my biggest view in politics that I've come around to is that I wish we were still in a world where politics wasn't people's identity and that people got along without needing to know or care about each other's politics and that politics was something you left in the voting booth and weren't super public about.
237
00:18:08,448 --> 00:18:17,808
But these days, mostly due to various conspiracy theories about people's assumption about Bitcoin Core's politics, which I mean, I don't even contribute to Bitcoin Core anymore, but whatever.
238
00:18:18,788 --> 00:18:31,168
People have assumed a certain political lean of me and also assume as much because my Twitter profile, my ex profile has blue hair in it, which I did at one point have blue hair, but this was before it had political connotations.
239
00:18:31,368 --> 00:18:36,468
And at this point, I just leave the blue hair in my profile picture because it filters out everyone I don't want to talk to.
240
00:18:36,468 --> 00:18:41,068
if your response is to make a comment about my blue hair in my profile picture i probably didn't
241
00:18:41,068 --> 00:18:46,568
want to talk to you to begin with so i leave that as my profile photo how do you stay grounded when
242
00:18:46,568 --> 00:18:54,488
the number goes up uh that's a good question i think that when the number goes up i generally
243
00:18:54,488 --> 00:19:00,568
don't look very much at the number um you know obviously been in bitcoin for a while bitcoin's
244
00:19:00,568 --> 00:19:05,288
a very material portion of my net worth so it does you know it makes an impact and i have to
245
00:19:05,288 --> 00:19:12,668
have to reason about it but usually uh between my wife and i we at the start of any bull market we
246
00:19:12,668 --> 00:19:17,288
sit down and have a conversation about where we think the price is realistically going to go and
247
00:19:17,288 --> 00:19:22,748
set a plan for whether we sell at which price and how much we sell if we sell at all and and whatever
248
00:19:22,748 --> 00:19:27,068
and and you know what we might want you know if we're going to go buy a house or whatever
249
00:19:27,068 --> 00:19:31,568
um and then we set that and we stick to it and then we don't look at it um so i think that's
250
00:19:31,568 --> 00:19:35,268
that's always the best way to do it is to sit there honestly at the start of a bull market and
251
00:19:35,268 --> 00:19:37,188
and say, is there something I need to do?
252
00:19:37,188 --> 00:19:38,188
Do I wanna buy a house?
253
00:19:38,188 --> 00:19:39,468
And like if Bitcoin reaches this amount,
254
00:19:39,468 --> 00:19:40,308
would I go buy a house?
255
00:19:40,308 --> 00:19:41,148
Would I go buy a car?
256
00:19:41,148 --> 00:19:42,508
Would I go whatever?
257
00:19:43,388 --> 00:19:45,648
Would I go buy Tesla stock?
258
00:19:45,648 --> 00:19:46,828
I don't know.
259
00:19:46,828 --> 00:19:50,768
But set the plan at the start and then don't look at it,
260
00:19:50,768 --> 00:19:51,708
don't think about it.
261
00:19:51,708 --> 00:19:54,608
And if you hit the target number that you wanted
262
00:19:54,608 --> 00:19:55,868
and then go buy your house.
263
00:19:55,868 --> 00:19:58,168
And if not, oh well.
264
00:19:58,168 --> 00:20:00,308
So I think that's worked well for me.
265
00:20:00,308 --> 00:20:05,228
And I think that more people would do well to do that.
266
00:20:05,268 --> 00:20:09,268
On that point, do you think we are in a bear market or a bull market right now?
267
00:20:10,408 --> 00:20:12,428
I think the four-year cycle is over.
268
00:20:12,748 --> 00:20:14,168
I very much subscribe to that.
269
00:20:15,508 --> 00:20:21,088
Maybe there'll be perturbations around the next four-year cycle or whatever,
270
00:20:21,088 --> 00:20:24,408
but it really does seem like...
271
00:20:24,408 --> 00:20:25,668
And I mean, that was natural, right?
272
00:20:25,748 --> 00:20:32,428
The amount of reduction in Bitcoin supply from the miners halves every four-year cycle.
273
00:20:32,428 --> 00:20:38,748
and bitcoin goes up in price but bitcoin also just becomes a much larger economy and so you
274
00:20:38,748 --> 00:20:45,468
know you start to need a lot more inflows in in dollar terms in order to make large price gaps
275
00:20:45,468 --> 00:20:52,708
large price differences as the the bitcoin having continues and so you know at some point it wasn't
276
00:20:52,708 --> 00:20:57,168
going to make as large an impact and i think we're we're starting to reach that point where it's just
277
00:20:57,168 --> 00:21:03,988
not making as large an impact so i i don't tend to think in like bearable cycles anymore i don't
278
00:21:03,988 --> 00:21:10,388
know that we're gonna hit 130 000 again in the like very near term but are we gonna like wander
279
00:21:10,388 --> 00:21:16,108
down the price wander down for the next three years like the last few cycles probably hopefully
280
00:21:16,108 --> 00:21:21,728
not but i think probably also not as much this cycle anymore what's one thing every bitcoiner
281
00:21:21,728 --> 00:21:30,928
should I stop doing? Bringing politics into it. Again, I tend to miss the, well, I mean,
282
00:21:30,968 --> 00:21:35,688
I wasn't as much alive during that world, but I tend to think that the world would be better off
283
00:21:35,688 --> 00:21:40,648
if people left their politics in the voting booth. And it's true that Bitcoin is a political system
284
00:21:40,648 --> 00:21:47,108
in a sense, but it's interesting, you know, we're here in Europe and talking to a lot of Europeans
285
00:21:47,108 --> 00:21:53,428
view the politics of Bitcoin very differently, right? It's not as libertarian, it's more
286
00:21:53,428 --> 00:21:59,748
freedom of, you know, you having money that is yours, that is that is free rather than
287
00:21:59,748 --> 00:22:04,188
like ending the Fed, like the Fed is evil, and we should abolish the Fed, which you see a lot more
288
00:22:04,188 --> 00:22:10,908
in the US community. And so I think this, this very polarization and politicization of Bitcoin
289
00:22:10,908 --> 00:22:15,488
is really negative for Bitcoin in the long term.
290
00:22:16,488 --> 00:22:20,668
And so I wish more Bitcoiners would support Bitcoin
291
00:22:20,668 --> 00:22:23,468
without making their support of Bitcoin
292
00:22:23,468 --> 00:22:26,248
something which is political in nature
293
00:22:26,248 --> 00:22:27,548
and rather just support Bitcoin.
294
00:22:27,708 --> 00:22:29,288
And then they can have their politics separately
295
00:22:29,288 --> 00:22:31,288
and also be political,
296
00:22:31,528 --> 00:22:33,768
but not necessarily tie the two together as much.
297
00:22:34,308 --> 00:22:35,428
What do you think is negative?
298
00:22:37,248 --> 00:22:39,408
I think that, I mean, certainly in the US,
299
00:22:39,408 --> 00:22:46,888
we've seen already a very immediate result that fewer Democrats are willing to go out and support
300
00:22:46,888 --> 00:22:53,508
Bitcoin-friendly policies. We've already seen that materially. And, you know, in the past,
301
00:22:53,548 --> 00:22:58,228
that wasn't true. Like Bitcoin was fairly bipartisan. We saw both Democrats and Republicans
302
00:22:58,228 --> 00:23:03,428
support Bitcoin-friendly policies, or, you know, both of them also supported anti-Bitcoin policies.
303
00:23:03,428 --> 00:23:16,008
But with Bitcoin being very, very tied to the Republican results as of the last election and Trump going to the Bitcoin conference and whatever, it's been politically harder for Democrats to support Bitcoin.
304
00:23:16,148 --> 00:23:26,928
And I think as a result, that really hurts our ability to get policy through and not have the government try to ban Bitcoin or try to ban privacy services on Bitcoin or ban LSP or ban whatever.
305
00:23:27,168 --> 00:23:29,868
That really is an important part of making Bitcoin work.
306
00:23:29,868 --> 00:23:33,428
The Biden administration wasn't very into Bitcoin, right?
307
00:23:33,588 --> 00:23:39,928
Yeah, the Biden administration was very hostile to crypto and Bitcoin, but so was the first Trump administration.
308
00:23:39,928 --> 00:23:42,188
The first Trump administration was arguably worse.
309
00:23:43,128 --> 00:23:51,008
And the Biden administration was very anti-crypto, a little less anti-Bitcoin, but they were still definitely anti-Bitcoin.
310
00:23:51,108 --> 00:23:53,068
They just didn't spend as much effort trying to destroy Bitcoin.
311
00:23:53,248 --> 00:23:54,788
They spent a lot of effort trying to destroy crypto.
312
00:23:55,528 --> 00:23:59,028
But the first Trump administration also tried to destroy Bitcoin.
313
00:23:59,028 --> 00:24:07,888
So I think it was equally bipartisan in their dislike of it, but some congressional and Senate folks did like Bitcoin.
314
00:24:09,448 --> 00:24:13,348
But now it's just one side of the aisle, and I think that that's probably worse.
315
00:24:14,448 --> 00:24:17,248
What part of your work weighs on you the most?
316
00:24:17,248 --> 00:24:27,608
yeah i mean i think again i spend a lot of time worrying about different ways bitcoin can go wrong
317
00:24:27,608 --> 00:24:33,168
and different risks and long-term risks and so i think there's always a part of me that's
318
00:24:33,168 --> 00:24:39,728
feels like you know these these these large risks i'm really worried about like for example
319
00:24:39,728 --> 00:24:44,368
mining centralizations like i don't actually spend much of my time on it because i focus on
320
00:24:44,368 --> 00:24:48,168
other risks that I think are also important and not necessarily more important, but just
321
00:24:48,168 --> 00:24:50,428
harder in the size, spend my resources on those.
322
00:24:50,948 --> 00:24:55,768
You know, it's always frustrating to me that the Bitcoin community and the Bitcoin developer
323
00:24:55,768 --> 00:25:01,488
community and the Bitcoin developer funding community doesn't spend more time trying to
324
00:25:01,488 --> 00:25:02,348
address these risks.
325
00:25:02,348 --> 00:25:08,528
And there are many people who do, but there's many more people who only talk about the positives
326
00:25:08,528 --> 00:25:13,528
of Bitcoin and don't worry, don't think about the negatives, the risks.
327
00:25:13,528 --> 00:25:15,788
and then also don't do anything about it.
328
00:25:15,908 --> 00:25:17,908
So, you know, you look at like P2Pool V2,
329
00:25:17,988 --> 00:25:19,428
I think is a really cool project
330
00:25:19,428 --> 00:25:22,448
and addresses a complimentary set of miners
331
00:25:22,448 --> 00:25:23,408
from Stratum V2
332
00:25:23,408 --> 00:25:27,688
and might enable some material mining decentralization wins
333
00:25:27,688 --> 00:25:28,608
in some part of the mining market.
334
00:25:28,868 --> 00:25:30,008
And I don't really get the sense
335
00:25:30,008 --> 00:25:32,048
that people know what P2Pool V2 is,
336
00:25:32,828 --> 00:25:35,728
like think about it or are talking about it.
337
00:25:35,768 --> 00:25:39,168
And it's like, you know, P2Pool V2 plus Stratum V2
338
00:25:39,168 --> 00:25:40,768
are two of the most important projects in Bitcoin.
339
00:25:40,768 --> 00:25:46,548
And like people don't talk about them or in many cases, they're not even aware of them.
340
00:25:46,548 --> 00:25:47,548
And I think that that's-
341
00:25:47,548 --> 00:25:50,548
Who do we need to interview to talk about them?
342
00:25:50,548 --> 00:25:52,228
Uh, so Jungli is actually here.
343
00:25:52,228 --> 00:25:56,068
One of the, the P2Pool V2 developers, um, who's great.
344
00:25:56,068 --> 00:25:56,768
I love Jungli.
345
00:25:56,768 --> 00:26:10,736
Okay What your favorite movie and book Oh man Um yeah I don I just don spend a lot of time consuming much content anymore which is my loss
346
00:26:10,776 --> 00:26:11,556
I should read more.
347
00:26:12,176 --> 00:26:14,156
I don't read all that much.
348
00:26:15,356 --> 00:26:17,856
But yeah, if I had to pick one, it's...
349
00:26:19,716 --> 00:26:20,196
Or a movie?
350
00:26:20,436 --> 00:26:21,956
You know, there's some great movies out there.
351
00:26:21,956 --> 00:26:28,236
I did a summer in college where I didn't have an internship, wasn't doing anything.
352
00:26:28,756 --> 00:26:38,916
So I was back at home with my parents and just like basically spent my time watching my way through the IMDb top 100 list and just watched, I don't know, probably 50 movies that summer or something.
353
00:26:39,156 --> 00:26:40,776
So, you know, there's some classics.
354
00:26:40,776 --> 00:26:42,476
There's like the Shawshank or something.
355
00:26:43,036 --> 00:26:46,856
It was classic, but I always love the old character studies.
356
00:26:46,856 --> 00:26:57,576
It's just, you know, slow moving films that really get you to know someone on screen and study how they react to different situations.
357
00:26:57,696 --> 00:26:58,776
I always love movies like that.
358
00:26:59,476 --> 00:27:05,616
What's an AI agent and why does everyone seem to have one much better than mine?
359
00:27:07,556 --> 00:27:09,876
Yeah, I mean, it's a good point.
360
00:27:09,876 --> 00:27:17,576
So, I mean, AI agents are really just taking these models and these models get retrained to make tool calls, right?
361
00:27:17,636 --> 00:27:24,316
So the model can output text that says, actually call this command or go to this website or whatever.
362
00:27:24,896 --> 00:27:35,976
And so now that they can do these tool calls, you can say, chat with an AI, chat with an LLM, and it will then go do things on your behalf.
363
00:27:35,976 --> 00:27:39,856
So potentially, you know, go to a website, click through, do something.
364
00:27:40,756 --> 00:27:45,296
And yeah, I mean, I think you're right that you see a lot of bragging about what agents
365
00:27:45,296 --> 00:27:46,636
can do on X or whatever.
366
00:27:46,636 --> 00:27:50,816
And then most people go and try them and they don't accomplish the tasks that people are
367
00:27:50,816 --> 00:27:51,396
saying they can.
368
00:27:51,656 --> 00:27:57,416
So I think it's true that they're both overhyped, but also still at the same time, very useful
369
00:27:57,416 --> 00:28:04,656
that you have to find the right workflows to use LLMs where they actually do accelerate
370
00:28:04,656 --> 00:28:07,776
your work and then also not use them where they don't.
371
00:28:07,836 --> 00:28:13,456
If you just try to use them to replace parts of your work today, you'll faceplant and spend
372
00:28:13,456 --> 00:28:17,216
more time correcting its output than actually you would have taken you to do it.
373
00:28:17,216 --> 00:28:21,076
But if you use them really narrowly tailored for the parts of your work that you need them
374
00:28:21,076 --> 00:28:27,176
for, they can be really productive because, you know, some tedious tasks you can replace
375
00:28:27,176 --> 00:28:30,016
to an LLM to free you up to do more things.
376
00:28:30,016 --> 00:28:43,676
So it does require a lot of experimentation to find the right workflows that they are actually really speeding up for you and not try to just use them for everything and then just end up wasting way more time.
377
00:28:43,856 --> 00:28:47,476
When you say tedious tasks, what comes to mind?
378
00:28:47,476 --> 00:28:55,176
So the obviously biggest use of LLMs right now is software engineers using LLMs to write software.
379
00:28:55,176 --> 00:29:05,096
And so there's a lot of cases where you're writing really nitty gritty stuff that has to be exactly right or exactly the way you want it, where you can't use an LLM.
380
00:29:05,136 --> 00:29:06,876
It's just going to do it slightly off.
381
00:29:06,936 --> 00:29:18,276
It's going to be too, you know, they've over-trained them now to like to try to only add new code and like touch as little of your existing code as possible because they used to touch too much existing code.
382
00:29:18,316 --> 00:29:22,036
And so now they've now like over-corrected and now they don't touch any existing code.
383
00:29:22,036 --> 00:29:27,376
And so it's very easy to use LLMs to just keep adding more and more and more and your
384
00:29:27,376 --> 00:29:29,696
code becomes unwieldy and doesn't work anymore.
385
00:29:30,756 --> 00:29:32,336
And so you can't use them for that.
386
00:29:32,396 --> 00:29:37,376
But for some more tedious tasks like, you know, changing a variable name that's referenced
387
00:29:37,376 --> 00:29:38,316
in a million places.
388
00:29:38,876 --> 00:29:40,836
Well, OK, there's some automated refactoring.
389
00:29:41,336 --> 00:29:46,876
There's some existing software that can do that in an automated way, but things that are
390
00:29:46,876 --> 00:29:48,496
slightly less automatable.
391
00:29:48,496 --> 00:29:52,796
So something where, you know, there's a pattern you want to apply, but it's very fuzzy.
392
00:29:53,116 --> 00:29:57,576
And you could maybe come up with some automated rules to apply that pattern across your code
393
00:29:57,576 --> 00:29:57,956
base.
394
00:29:58,256 --> 00:30:02,116
But it's kind of hard and you spend more time coming up with the automated rules than doing
395
00:30:02,116 --> 00:30:03,596
it, but it's still going to take you an hour.
396
00:30:04,016 --> 00:30:05,976
But an LLM is really good at that, right?
397
00:30:06,016 --> 00:30:07,536
So there's some pattern you want to apply.
398
00:30:07,696 --> 00:30:08,836
It can just do it repeatedly.
399
00:30:09,836 --> 00:30:16,256
Also, sometimes writing certain tests, LLMs can be really good at writing a test that might
400
00:30:16,256 --> 00:30:20,436
otherwise be really annoying because you got to get everything exactly right. But the LLM can just
401
00:30:20,436 --> 00:30:25,376
spend a bunch of time doing it and you can go do something else. And it'll take a while to do it
402
00:30:25,376 --> 00:30:30,256
and it might have to play around with it a bunch, but you don't have to do it. You can go do something
403
00:30:30,256 --> 00:30:38,036
else while it's doing it. Why should we all care about AI agents? I mean, because this is a huge
404
00:30:38,036 --> 00:30:43,276
productivity unlock for certain tasks, even today, right? And they continue to, these models continue
405
00:30:43,276 --> 00:30:47,956
need to get better. And so there'll be better at more and more tasks. But it's already the case
406
00:30:47,956 --> 00:30:54,316
that they're a huge productivity unlock for different classes of work today and are going
407
00:30:54,316 --> 00:30:59,936
to change the way people interact with computers. They're going to change a lot of how people do
408
00:30:59,936 --> 00:31:06,016
many tasks because now all of a sudden you just have to describe what you want done and they can
409
00:31:06,016 --> 00:31:10,856
go do it again. There's many things that can't do, but there are many things that can. And so
410
00:31:10,856 --> 00:31:17,316
people need to be engaging with this and trying and figuring out what they can you know how much
411
00:31:17,316 --> 00:31:22,036
better they can be at whatever they're trying to accomplish by using these tools these are
412
00:31:22,036 --> 00:31:29,756
great new tools right and so i think it's really important that people you know embrace the new
413
00:31:29,756 --> 00:31:34,876
tools and and and get that productivity unlock that they were going to get what's something
414
00:31:34,876 --> 00:31:39,296
that an AI agent can do today that people wouldn't believe?
415
00:31:42,096 --> 00:31:45,256
Yeah, I mean, part of the problem is they can be pretty random.
416
00:31:45,616 --> 00:31:49,036
Like you can, because they are stochastic,
417
00:31:49,416 --> 00:31:50,856
they are random, like their output.
418
00:31:51,016 --> 00:31:52,716
If you ask them the same question a bunch of times,
419
00:31:52,756 --> 00:31:54,176
they will give you a different output each time.
420
00:31:54,176 --> 00:31:56,736
You can turn that off, but the performance is worse.
421
00:31:57,376 --> 00:32:03,076
Because they are random, very often you'll see people talk about like,
422
00:32:03,076 --> 00:32:07,576
look, I got my agent to do this thing and it's really amazing.
423
00:32:07,576 --> 00:32:11,656
And it like wrote all this awesome code that is really surprising that it can write.
424
00:32:12,856 --> 00:32:15,216
And then you try it and it doesn't.
425
00:32:17,096 --> 00:32:21,676
But I think, you know, in general, there are many programming tasks today
426
00:32:21,676 --> 00:32:25,736
where agents are materially better than basically any human.
427
00:32:26,796 --> 00:32:29,576
And so it is really interesting, right?
428
00:32:29,576 --> 00:32:32,636
That they're very good at some things and they're very bad at other things,
429
00:32:32,636 --> 00:32:35,116
even within the domain of like programming, right?
430
00:32:35,196 --> 00:32:39,256
So if you're writing, you know, a lot of like academic stuff
431
00:32:39,256 --> 00:32:40,876
where you're writing like a simulation
432
00:32:40,876 --> 00:32:43,636
of some very specific like physics process or something,
433
00:32:44,156 --> 00:32:46,796
fairly small code, very delicate code,
434
00:32:46,836 --> 00:32:48,216
has to be very performant, whatever.
435
00:32:49,476 --> 00:32:52,416
It wouldn't, it's generally not surprising at this point
436
00:32:52,416 --> 00:32:55,316
that agents are often better than some of the best humans
437
00:32:55,316 --> 00:32:56,636
at writing these kinds of things.
438
00:32:57,016 --> 00:32:58,456
You know, small contained algorithms
439
00:32:58,456 --> 00:33:00,856
that might need to be very performant.
440
00:33:00,856 --> 00:33:08,416
And, you know, maybe the AI agent has to spend some time trying different things and whatever, but it can often be way better than even the best humans.
441
00:33:08,416 --> 00:33:19,976
And so I think that's been very interesting to watch that kind of as they've grown very good in certain directions and haven't grown as much in other directions.
442
00:33:20,916 --> 00:33:25,596
What do you think of popular AI agents like Claude Grock or Chagipiti?
443
00:33:25,976 --> 00:33:29,596
Yeah. So there's, I mean, first of all, there's a distinction between the models themselves.
444
00:33:29,596 --> 00:33:33,676
So there's, you know, all of the companies you list to make models.
445
00:33:34,056 --> 00:33:40,156
So there's, you know, Grok is a model and Anthropic obviously makes the cloud models.
446
00:33:40,736 --> 00:33:44,676
And then there's separately, there's agents and agent hardnesses, right?
447
00:33:44,696 --> 00:33:49,236
So something that uses the models to enable the model to do certain things.
448
00:33:49,336 --> 00:33:55,296
And so there's like cloud code is a program that you run that, you know, accesses the cloud model.
449
00:33:55,296 --> 00:33:58,236
And then it allows the cloud model to run commands on your computer.
450
00:33:58,236 --> 00:34:04,036
there's also Hermes and OpenClaw and whatever else there's there's many of these and Goose and
451
00:34:04,036 --> 00:34:12,016
there's many of these agent software packages right so because this space has moved so fast
452
00:34:12,016 --> 00:34:18,636
a lot of the existing agent software has has also moved fast and and some in some cases moved too
453
00:34:18,636 --> 00:34:25,076
fast so example the OpenClaw I think is the great example it's now becoming a little less popular
454
00:34:25,076 --> 00:34:27,236
because it moved too fast,
455
00:34:27,676 --> 00:34:31,416
merged too many PRs that were written by AI
456
00:34:31,416 --> 00:34:33,116
that might not have always been the best quality.
457
00:34:34,176 --> 00:34:35,636
And so I think a lot of their user base
458
00:34:35,636 --> 00:34:39,236
has gotten frustrated with how often it doesn't work,
459
00:34:39,496 --> 00:34:42,256
how often it breaks and new updates break things
460
00:34:42,256 --> 00:34:43,256
and change things and whatever
461
00:34:43,256 --> 00:34:45,096
versus some other agent software
462
00:34:45,096 --> 00:34:47,336
that is trying to take a different approach
463
00:34:47,336 --> 00:34:49,716
and has been maybe slower to build,
464
00:34:49,796 --> 00:34:51,976
but also doesn't break things as much
465
00:34:51,976 --> 00:34:55,096
and maybe has a more reliable user experience.
466
00:34:55,856 --> 00:34:59,396
What are the best open source AI agents today, in your opinion?
467
00:35:00,236 --> 00:35:02,316
On the agent software side,
468
00:35:02,676 --> 00:35:06,996
I'm partial to Goose because I work with that team regularly.
469
00:35:08,276 --> 00:35:09,776
So I like the Goose harness.
470
00:35:10,576 --> 00:35:13,016
Also, Paramese is very popular and has a good harness
471
00:35:13,016 --> 00:35:15,216
if you want something that you can talk to
472
00:35:15,216 --> 00:35:16,896
through your existing chat interface.
473
00:35:17,096 --> 00:35:21,116
So if you want to text your agent over signal or SMS or whatever,
474
00:35:21,116 --> 00:35:27,836
that you can use hermes for that um that's on the agent harness side but on the model side you know
475
00:35:27,836 --> 00:35:35,516
obviously the the best ai the best llm models that uh are open source that are available for people to
476
00:35:35,516 --> 00:35:40,556
run locally on their own hardware assuming they can afford the hardware to run it which is pretty
477
00:35:40,556 --> 00:35:46,876
hard these days how expensive is it um well yeah part of the problem is is that the prices for
478
00:35:46,876 --> 00:35:48,676
for hardware has just run up very quickly, right?
479
00:35:48,716 --> 00:35:49,996
Because of the AI boom,
480
00:35:51,256 --> 00:35:52,936
you know, OpenAI and Anthropic
481
00:35:52,936 --> 00:35:55,736
and Facebook and Google and everybody
482
00:35:55,736 --> 00:35:57,836
are buying the entire supply
483
00:35:57,836 --> 00:35:59,756
of all the memory that is made
484
00:35:59,756 --> 00:36:02,236
and all the chips that are made every year.
485
00:36:02,376 --> 00:36:04,156
And there's not much supply left
486
00:36:04,156 --> 00:36:05,096
for the rest of us.
487
00:36:05,796 --> 00:36:07,356
And so the prices for a lot of things
488
00:36:07,356 --> 00:36:09,696
like memory have 4X'd, 5X'd
489
00:36:09,696 --> 00:36:12,456
over the last year or so.
490
00:36:12,856 --> 00:36:14,656
And so that's gotten fairly expensive.
491
00:36:14,656 --> 00:36:27,856
But, you know, you can still get if you're comfortable with model output being relatively slow, you can still put together a computer that can run local AI very competitively for, you know, call it $5,000 or so.
492
00:36:28,196 --> 00:36:31,276
But all those models are the Chinese firms.
493
00:36:31,716 --> 00:36:36,316
So none of the Western firms, except for there's one based in France.
494
00:36:37,756 --> 00:36:39,156
That I'm forgetting the name of now.
495
00:36:39,156 --> 00:36:40,896
but all of the
496
00:36:40,896 --> 00:36:41,776
for the most part
497
00:36:41,776 --> 00:36:42,896
all the competitive
498
00:36:42,896 --> 00:36:44,256
and especially all the really good
499
00:36:44,256 --> 00:36:45,836
open source models
500
00:36:45,836 --> 00:36:47,016
that you could run yourself
501
00:36:47,016 --> 00:36:47,876
or that
502
00:36:47,876 --> 00:36:49,156
you know someone else can run
503
00:36:49,156 --> 00:36:49,776
on your behalf
504
00:36:49,776 --> 00:36:50,336
that's not
505
00:36:50,336 --> 00:36:51,556
you know one of these labs
506
00:36:51,556 --> 00:36:52,356
that makes their own models
507
00:36:52,356 --> 00:36:54,196
are all from Chinese labs
508
00:36:54,196 --> 00:36:55,396
so so far
509
00:36:55,396 --> 00:36:56,856
the dynamic has been very interesting
510
00:36:56,856 --> 00:36:57,456
that it's been
511
00:36:57,456 --> 00:36:59,036
the West versus China
512
00:36:59,036 --> 00:37:00,396
and China has been the one
513
00:37:00,396 --> 00:37:01,716
building the open source thing
514
00:37:01,716 --> 00:37:03,256
and the West has been the one
515
00:37:03,256 --> 00:37:04,396
building the closed thing
516
00:37:04,396 --> 00:37:05,496
that you know
517
00:37:05,496 --> 00:37:06,996
some company is trying to lock down
518
00:37:06,996 --> 00:37:07,356
for
519
00:37:07,356 --> 00:37:08,696
to make more money
520
00:37:08,696 --> 00:37:16,936
venice by eric borchich is an open source one right so yeah venice venice offers the open source
521
00:37:16,936 --> 00:37:22,536
models right so uh the the closed source models are just these these large firms right if you want
522
00:37:22,536 --> 00:37:28,376
if you want the best model right now you have to go pay anthropic or open ai uh a lot of money for
523
00:37:28,376 --> 00:37:34,996
it but if you are willing to use something that's slightly that's you know 90 is good
524
00:37:34,996 --> 00:37:41,216
then you can save a lot of money by going to someone who is hosting it on aws or you could
525
00:37:41,216 --> 00:37:45,196
buy your own hardware and host it and use some of these chinese models that have been trained
526
00:37:45,196 --> 00:37:49,876
they're fairly expensive to train but some of these large chinese firms that have trained these
527
00:37:49,876 --> 00:37:56,236
models and released them as open source you can use like hermes or goose or something as the the
528
00:37:56,236 --> 00:38:02,556
software to install locally and then you have to find a provider that offers models and then you're
529
00:38:02,556 --> 00:38:10,076
paying someone you know however many cents per per word coming in or out of your of the llm um
530
00:38:10,076 --> 00:38:15,596
and yeah there's there's many provider open router and whatever that that offer these uh models hosted
531
00:38:15,596 --> 00:38:21,676
for you um now there's some really cool ones there's some folks who like maple for example
532
00:38:21,676 --> 00:38:28,316
they will run these models on a cloud server but the content you send to the server and the content
533
00:38:28,316 --> 00:38:31,356
coming from the server is encrypted in such a way that they can't see it.
534
00:38:31,576 --> 00:38:36,236
So you actually, it's like proof of NVIDIA, basically, is what it is.
535
00:38:36,356 --> 00:38:41,196
So I can connect directly to the GPU running in their server, the NVIDIA GPU.
536
00:38:41,836 --> 00:38:47,776
And that GPU will tell me like, hey, this is the software I'm running,
537
00:38:47,816 --> 00:38:49,456
and this is the model I have loaded and whatever.
538
00:38:50,056 --> 00:38:51,536
And then I can verify that.
539
00:38:51,536 --> 00:38:56,976
And I can verify that that is, NVIDIA said that that was real NVIDIA hardware
540
00:38:56,976 --> 00:38:58,356
and that hasn't been tampered with.
541
00:38:58,356 --> 00:39:14,704
And I can I can be relatively confident that actually Maple the service provider can see what I telling the LLM and what the LLM is telling me So this is actually a really cool thing that some providers are starting to do to offer much more private LLM use
542
00:39:14,864 --> 00:39:17,104
And how much does this cost monthly?
543
00:39:18,024 --> 00:39:21,324
So, yeah, it depends a lot on the model you use.
544
00:39:21,404 --> 00:39:25,804
And so, you know, some of the smaller models can fit in small GPUs, move very quickly.
545
00:39:25,804 --> 00:39:33,244
so specifically the the the the llm you use so the the specific software package that is running on
546
00:39:33,244 --> 00:39:40,204
the gpu whether in the cloud or locally um they can vary very substantially so some of the smaller
547
00:39:40,204 --> 00:39:47,064
models some of these smaller llms that are still very capable and and you know maybe 70 is good or
548
00:39:47,064 --> 00:39:55,064
whatever can fit in a small gpu they can fit in you know they have something on the order of like
549
00:39:55,064 --> 00:39:56,324
30 billion parameters.
550
00:39:56,564 --> 00:40:00,424
So how many parameters they have, and then those parameters can be different sizes, but
551
00:40:00,424 --> 00:40:01,604
they have 30 billion.
552
00:40:01,604 --> 00:40:07,724
But the large models, now we don't know for sure how large Anthropics models are and opening
553
00:40:07,724 --> 00:40:12,604
eyes models, but they're probably on the order of a trillion to trillion and a half parameters,
554
00:40:12,604 --> 00:40:14,824
trillion and a half parameters, right?
555
00:40:15,044 --> 00:40:21,924
And so the cost scales somewhat proportionally, somewhat more than proportionally with that
556
00:40:21,924 --> 00:40:25,324
parameter count because it means how much can you fit on the GPU?
557
00:40:25,784 --> 00:40:28,224
How many clients can you service on the GPU?
558
00:40:28,324 --> 00:40:35,464
And how fast can that same GPU using the same amount of power spit out output for, you know,
559
00:40:35,564 --> 00:40:36,824
for, for each client.
560
00:40:36,824 --> 00:40:40,924
And so if you have a small model, it can spit out that output very, very quickly.
561
00:40:41,204 --> 00:40:42,964
And you can get a lot more clients on there.
562
00:40:43,084 --> 00:40:46,064
You can using the same GPU, same amount of power, same cost.
563
00:40:46,064 --> 00:40:53,344
you get a lot more output versus something that has, you know, 10x, 100x more parameters and takes
564
00:40:53,344 --> 00:41:00,604
up 10x or 100x space on the GPU, often takes up three, four, five times more time per token,
565
00:41:00,964 --> 00:41:05,864
has to read more memory. And so that means at least three, four, five times more cost
566
00:41:05,864 --> 00:41:14,244
for that to use that LLM versus the other one that you might use. So it varies very substantially.
567
00:41:14,244 --> 00:41:22,184
And then on top of that, Anthropic and OpenAI and all these other large labs charge a lot of money on top of that because it's very expensive for them to train the model.
568
00:41:22,784 --> 00:41:33,364
And so they have to then bake that cost into the cost they're charging you compared to, for example, some of the open source models where, yes, it was very expensive that this Chinese company trained the model.
569
00:41:33,824 --> 00:41:36,804
But now the model is open source and they're giving it away.
570
00:41:36,924 --> 00:41:40,384
And so anyone can then run that and sell it to you at a lower cost.
571
00:41:40,384 --> 00:41:43,664
so it varies really substantially and you can get
572
00:41:43,664 --> 00:41:50,024
you know you can do a bunch with $20 a month or you can very easily spend $1,000 a month
573
00:41:50,024 --> 00:41:51,104
what China?
574
00:41:51,684 --> 00:41:58,964
yeah I mean it's a good question so so far they have their the LLMs that they've been building
575
00:41:58,964 --> 00:42:05,204
have at max been 90% as good at the same time right so the space is moving very quickly the
576
00:42:05,204 --> 00:42:11,684
LLMs, Chinese companies, Alibaba and whatever are releasing now are still much better than the
577
00:42:11,684 --> 00:42:20,744
best LLMs from three months ago. But at best, they've been like 90% as good. And so as a result,
578
00:42:21,824 --> 00:42:26,944
if they want people to use them, they either have to drastically underprice them or just
579
00:42:26,944 --> 00:42:32,944
open source them. And so a big part of it's just been this international dynamic of everyone wants
580
00:42:32,944 --> 00:42:40,884
to be win AI. The government, everybody even top down wants to be the one, the country to win AI
581
00:42:40,884 --> 00:42:48,264
and have the best AI and capture all that value from all these new tools that are really speeding
582
00:42:48,264 --> 00:42:56,224
up people's workflows. So I think China has taken this, a lot of Chinese firms and Chinese government
583
00:42:56,224 --> 00:43:02,824
presumably leaning on them as well, has taken the strategy of saying, look, we're not currently
584
00:43:02,824 --> 00:43:08,744
winning, but we can capture all the long tail. And so it is true that, you know, all of the
585
00:43:08,744 --> 00:43:18,204
LLM use that isn't buying directly from OpenAI and Anthropic and Google and basically just them
586
00:43:18,204 --> 00:43:22,824
is using Chinese models. So everybody else is using the Chinese models because they're the
587
00:43:22,824 --> 00:43:26,724
best freely available models. And yes, it's true. If you ask them about Tiananmen Square,
588
00:43:27,124 --> 00:43:30,864
they'll say, I don't know what you're talking about. And if you ask them about Taiwan,
589
00:43:30,864 --> 00:43:32,004
They'll say, no, it's a part of China.
590
00:43:33,664 --> 00:43:38,744
And so they do get some ability to kind of push their worldview through this.
591
00:43:38,904 --> 00:43:41,244
And that's probably some part of it.
592
00:43:41,444 --> 00:43:45,824
But I think the biggest part of it is just that they want to win AI.
593
00:43:46,304 --> 00:43:57,484
And as long as they're not quite as good, the best way to do that is to take these models, open source them, and get everyone using them so that everyone is using the Chinese AI, right?
594
00:43:57,484 --> 00:44:00,624
Rather than using open AI or anthropic.
595
00:44:00,864 --> 00:44:06,044
Is there a reason why I shouldn't or we shouldn't use Chinese open source agents?
596
00:44:06,184 --> 00:44:07,164
Depends on what for.
597
00:44:07,944 --> 00:44:12,144
So they don't, certainly people watch what their agents do.
598
00:44:12,144 --> 00:44:17,084
And so they don't generally misbehave and just like steal all your data and send it to China or something like that.
599
00:44:17,124 --> 00:44:20,924
In some cases they could, but they don't appear to have been trained to do that.
600
00:44:21,604 --> 00:44:30,724
But also, yeah, for in cases where people are asking questions, you know, a lot of people these days use these chat interfaces to LLMs to ask questions and learn things.
601
00:44:30,724 --> 00:44:33,064
and students are using it to learn about the world and whatever,
602
00:44:33,624 --> 00:44:37,564
there are obviously some cases where you shouldn't be asking a Chinese LLM
603
00:44:37,564 --> 00:44:42,364
about the status of Taiwan and the history of China
604
00:44:42,364 --> 00:44:44,924
and some parts of the history of the Western world
605
00:44:44,924 --> 00:44:47,284
because those LLMs are very clearly trained
606
00:44:47,284 --> 00:44:50,264
to tell you about things from a Chinese viewpoint.
607
00:44:50,804 --> 00:44:53,784
We have the same with God or CGPT.
608
00:44:54,204 --> 00:44:54,924
It's true.
609
00:44:55,144 --> 00:44:56,884
Everything ends up having some viewpoint,
610
00:44:57,004 --> 00:44:58,764
whether it's trained into it deliberately or not.
611
00:44:58,764 --> 00:45:06,684
um obviously i think that the uh anthropic and open ai and others are are spending less time
612
00:45:06,684 --> 00:45:11,624
training models to say very specific things about history but there are other cases where people
613
00:45:11,624 --> 00:45:17,684
have showed that you know they're uh the western models are very unwilling to criticize israel
614
00:45:17,684 --> 00:45:21,944
they will refuse to criticize israel you can ask them explicitly you know criticize this country
615
00:45:21,944 --> 00:45:26,104
okay criticize this country okay criticize this country okay criticize israel no i can't do that
616
00:45:26,104 --> 00:45:33,504
Right. So there's definitely the case that there are other worldviews that are being baked into those models.
617
00:45:33,724 --> 00:45:40,264
I think that's true. But I would I would think that there is less being baked into those models than the Chinese models.
618
00:45:40,564 --> 00:45:49,084
But yeah, I mean, probably to ask about certain political things, an LLM is not quite the right tool.
619
00:45:49,364 --> 00:45:54,704
So if I go for a Chinese open source AI agent, which one would you recommend?
620
00:45:54,704 --> 00:46:03,464
yeah so the the best models right now are uh quen which is made by alibaba um and deep seek if you
621
00:46:03,464 --> 00:46:10,684
have a much larger gpus um which is a former trading firm in hong kong um but so yeah using
622
00:46:10,684 --> 00:46:17,864
those those are the the probably the two most popular groups of models um for open source use
623
00:46:17,864 --> 00:46:24,184
and then of course you then need the agent wrapper itself so uh goose or or um hermes or
624
00:46:24,184 --> 00:46:30,324
something like that. Can you explain why AI agents can just use a credit card like we do?
625
00:46:30,744 --> 00:46:36,124
Yeah. So that's, I mean, so it's a very good question. So the answer in short is they can,
626
00:46:36,764 --> 00:46:42,964
but the credit card system can't really support that. So, you know, it's true they can and they
627
00:46:42,964 --> 00:46:47,464
have, there are tools where, you know, you can today, you can tell your AI agent, here's my
628
00:46:47,464 --> 00:46:51,544
credit card number and, you know, please go browse the web and buy this and whatever and,
629
00:46:51,544 --> 00:46:52,984
and enter my credit card number.
630
00:46:53,184 --> 00:46:55,464
And people can do that and do do that.
631
00:46:55,924 --> 00:46:58,864
The problem is Visa and MasterCard,
632
00:46:59,244 --> 00:47:00,884
their networks are not set up for this, right?
633
00:47:00,904 --> 00:47:03,124
So the entire credit card networks
634
00:47:03,124 --> 00:47:06,124
are set up for making a payment.
635
00:47:06,364 --> 00:47:09,904
And then if it happened to be fraud somehow
636
00:47:09,904 --> 00:47:11,104
or wrong or whatever,
637
00:47:11,284 --> 00:47:12,164
you can get your money back.
638
00:47:12,484 --> 00:47:13,424
How does this work?
639
00:47:13,944 --> 00:47:16,944
When I told, you know, this model,
640
00:47:17,244 --> 00:47:18,924
please go buy me, you know,
641
00:47:18,964 --> 00:47:20,264
this pair of shoes that has this,
642
00:47:20,264 --> 00:47:25,944
these specific properties or whatever and it went and bought something else who's liable for that
643
00:47:25,944 --> 00:47:31,204
am i liable for that is the merchant liable for that who gets the money and there's not really
644
00:47:31,204 --> 00:47:38,764
the visa doesn't have a clear answer to this uh probably the answer is i get the money and the
645
00:47:38,764 --> 00:47:45,384
merchant gets screwed and so so merchants aren't willing to accept this so we saw this with open ai
646
00:47:45,384 --> 00:47:53,364
I actually did build somewhat this product, not exactly, but somewhat this product where you can choose to buy things within the chat interface.
647
00:47:53,544 --> 00:48:09,024
They had to roll it back because merchants refuse to participate because all the merchants today, the way they mitigate the risk of all these fraud issues is in part making sure that only a human can use their website.
648
00:48:09,024 --> 00:48:22,664
So all these CAPTCHAs and mouse tracking and whatever else to make it so that if a bot or an AI agent is the one using the website, they can detect that and then block the transaction.
649
00:48:23,044 --> 00:48:25,984
Obviously, their whole goal is to block AI agents.
650
00:48:26,164 --> 00:48:36,064
And so, yes, it can work sometimes, but writ large, like if it becomes popular, it would be too problematic for them and just they couldn't allow it.
651
00:48:36,064 --> 00:48:44,324
And so we have seen Visa and MasterCard try to build out their own API for this specifically,
652
00:48:44,624 --> 00:48:47,024
which, you know, is kind of a pop up out of the chat dialogues.
653
00:48:47,104 --> 00:48:51,164
Maybe you're telling chat GPT or whatever, like, go buy me shoes.
654
00:48:51,264 --> 00:48:52,264
I want this pair of shoes, whatever.
655
00:48:52,384 --> 00:48:55,484
And it pops up and it says, you know, here's a website.
656
00:48:55,584 --> 00:48:56,464
I want to buy from it.
657
00:48:56,684 --> 00:48:58,504
Here's the product image or whatever.
658
00:48:58,644 --> 00:48:59,764
Please click approve or deny.
659
00:48:59,924 --> 00:49:03,244
And so that is what Visa and MasterCard want to do.
660
00:49:03,244 --> 00:49:08,024
And that's kind of fine if you just live in today's world and you don't think about the future.
661
00:49:08,384 --> 00:49:12,724
But if you actually are worried about like, what are people going to want to do with these things?
662
00:49:13,024 --> 00:49:19,384
Well, what if I tell my agent to go start a startup to sell widgets or, you know, whatever?
663
00:49:20,964 --> 00:49:24,284
How does my AI agent then, you know, has to go buy a domain.
664
00:49:24,424 --> 00:49:25,544
It has to go buy web hosting.
665
00:49:25,644 --> 00:49:30,524
It has to go, you know, maybe it has to do some research and buy access to certain research APIs
666
00:49:30,524 --> 00:49:33,744
to find out more information about the widgets I want to sell.
667
00:49:34,284 --> 00:49:37,664
All of a sudden, I'm getting like all these stupid pop-ups about like,
668
00:49:37,724 --> 00:49:39,664
hey, do you want to approve your agent buying this?
669
00:49:40,044 --> 00:49:41,904
I want to just tell it, here's your five bucks.
670
00:49:42,024 --> 00:49:44,224
Go like, you know, spend your five bucks wildly.
671
00:49:44,304 --> 00:49:44,824
I don't care, right?
672
00:49:46,444 --> 00:49:51,004
And so this concept of how Visa and MasterCard are structured
673
00:49:51,004 --> 00:49:56,384
around liability for the merchant when the merchant doesn't perform
674
00:49:56,384 --> 00:49:59,704
and the consumer can do this chargeback thing
675
00:49:59,704 --> 00:50:03,004
just doesn't work in this new world, right?
676
00:50:03,104 --> 00:50:05,264
And so there has to be something new.
677
00:50:06,004 --> 00:50:08,444
And so it's really, right now,
678
00:50:08,524 --> 00:50:09,824
it seems like the something new
679
00:50:09,824 --> 00:50:11,804
is either going to be Bitcoin or stable coins.
680
00:50:12,464 --> 00:50:15,644
How are companies like Stripe, OpenAI,
681
00:50:16,484 --> 00:50:18,324
Google, Coinbase, and Visa
682
00:50:18,324 --> 00:50:20,824
trying to implement AI payments?
683
00:50:21,144 --> 00:50:22,704
Yeah, so there's currently
684
00:50:22,704 --> 00:50:24,744
just Cambridge an explosion of different approaches.
685
00:50:25,304 --> 00:50:28,004
But for the most part, with the exception of Visa,
686
00:50:28,004 --> 00:50:32,964
everybody's kind of pulling in this direction of, of a new payment rail.
687
00:50:33,164 --> 00:50:37,924
So either stable coins or, you know, maybe it's direct settlement.
688
00:50:38,084 --> 00:50:42,844
Maybe it's like, you know, the open AI, a bunch of open AI users buy a bunch of products
689
00:50:42,844 --> 00:50:45,104
and then open AI sends the merchant a wire or something.
690
00:50:45,204 --> 00:50:45,524
I don't know.
691
00:50:45,584 --> 00:50:51,004
But, um, so there's been, everybody's got their own protocol for it.
692
00:50:52,044 --> 00:50:54,064
Weirdly, I think in this case, it doesn't actually matter
693
00:50:54,064 --> 00:50:56,804
because if there's one thing an LLM is really good at, you know,
694
00:50:56,804 --> 00:50:59,584
all these LLMs are really good at writing short snippets of code.
695
00:51:00,124 --> 00:51:01,824
And so if the website says like,
696
00:51:01,924 --> 00:51:03,104
ah, to access this website,
697
00:51:03,104 --> 00:51:06,404
you have to use this specific API in this specific way.
698
00:51:06,964 --> 00:51:08,704
The LLM has no problem with that.
699
00:51:08,764 --> 00:51:10,524
It's just going to write that code to do that.
700
00:51:10,764 --> 00:51:12,144
And then it can happily do that.
701
00:51:12,204 --> 00:51:13,184
So I don't think that really matters.
702
00:51:13,244 --> 00:51:14,544
What really matters is the payment rail.
703
00:51:14,804 --> 00:51:19,404
Like the LLM, you know, your local AI agent tool
704
00:51:19,404 --> 00:51:23,224
has to be able to have access to the right payment rail
705
00:51:23,224 --> 00:51:24,104
that the merchant wants.
706
00:51:24,104 --> 00:51:30,124
And so, you know, there has to be this kind of somehow agreement on what payment rails people are going to use.
707
00:51:30,204 --> 00:51:35,584
And so far, it has mostly been stablecoins is what people are mostly pushing.
708
00:51:35,784 --> 00:51:37,364
And of course, they want that.
709
00:51:37,504 --> 00:51:45,104
These companies pushing these things want the stablecoin that they own to win.
710
00:51:45,104 --> 00:51:57,244
Because if everyone in the world is transacting with USDC, well, then Circle makes a ton of money because everybody has to have float in USDC that adds up to a lot of money and a lot of interest for them.
711
00:51:57,484 --> 00:51:58,784
So everybody wants to win this game.
712
00:51:58,844 --> 00:52:01,904
And so, so far it's been mostly stable coins the way people are pushing.
713
00:52:02,044 --> 00:52:04,784
But I think Bitcoin has a real shot here as a dark horse.
714
00:52:04,992 --> 00:52:07,792
where we're the open alternative.
715
00:52:08,192 --> 00:52:11,472
Instead of everybody fighting over who gets to be the stablecoin that wins,
716
00:52:11,732 --> 00:52:14,252
Bitcoin might actually have a chance at saying like,
717
00:52:14,332 --> 00:52:17,172
no, no, no, we're open and you can just use it.
718
00:52:17,432 --> 00:52:21,512
If AI agents start buying through platforms like Amazon,
719
00:52:21,992 --> 00:52:25,672
won't those platforms end up deciding the future payment system?
720
00:52:26,472 --> 00:52:29,112
Yeah, I mean, I think that's a very real possibility
721
00:52:29,112 --> 00:52:31,592
is that some of the largest merchants get to decide
722
00:52:31,592 --> 00:52:32,932
what payment warehouse people use
723
00:52:32,932 --> 00:52:35,672
because they can set what they need to be paid in.
724
00:52:36,452 --> 00:52:40,512
So far, we largely haven't seen merchants engage very much on this.
725
00:52:40,912 --> 00:52:43,172
You know, we've seen a lot of payment processors like Stripe,
726
00:52:43,552 --> 00:52:45,432
you know, payments people, people in the payments industry,
727
00:52:45,532 --> 00:52:46,652
which is mostly payment processors,
728
00:52:47,172 --> 00:52:49,212
try to push their stablecoin or whatever.
729
00:52:50,472 --> 00:52:52,512
But yeah, I think they don't have the power.
730
00:52:52,872 --> 00:52:54,212
You're totally right, it's the merchants.
731
00:52:54,532 --> 00:52:58,752
It's Amazon, it's Walmart, it's these large merchants who can decide.
732
00:52:58,752 --> 00:53:03,512
So Amazon has invested in some of these AI companies.
733
00:53:03,652 --> 00:53:04,232
That's true.
734
00:53:04,852 --> 00:53:09,572
But for the most part, I don't think they're they haven't yet engaged much on the AI agent
735
00:53:09,572 --> 00:53:10,592
payment side.
736
00:53:11,112 --> 00:53:16,952
They've mostly invested because of their AWS business, where they're powering a lot of
737
00:53:16,952 --> 00:53:22,272
the AI training and AI inference workloads out of their GPUs in the AWS business.
738
00:53:23,592 --> 00:53:25,972
But yeah, they haven't they haven't really engaged very much.
739
00:53:25,972 --> 00:53:35,012
And OpenAI, too, has, you know, they've done these partnerships, but they're not really pushing, you know, we're going to pick USDC on base or whatever yet.
740
00:53:35,012 --> 00:53:40,332
So I think there's, to some extent, we have this shot.
741
00:53:40,672 --> 00:53:48,012
If we can push small merchants and small agent providers, so people like I mentioned Maple
742
00:53:48,012 --> 00:53:55,552
earlier, but maybe get it into Goose or OpenClaw or Hermes, where Bitcoin becomes the default
743
00:53:55,552 --> 00:54:01,572
option that's always available, then if we can get the client side and maybe a few merchants,
744
00:54:01,572 --> 00:54:06,332
smaller merchants and whatever, then maybe like Amazon doesn't care.
745
00:54:06,612 --> 00:54:07,472
Amazon just wants your money.
746
00:54:07,932 --> 00:54:12,592
Like if the thing that all the agents support is Bitcoin, Amazon will just say, all right,
747
00:54:12,712 --> 00:54:13,032
screw it.
748
00:54:13,052 --> 00:54:14,212
We'll take Bitcoin from the agents.
749
00:54:14,312 --> 00:54:14,692
We don't care.
750
00:54:14,752 --> 00:54:15,652
That's not our problem.
751
00:54:17,592 --> 00:54:23,992
Versus I think if we do nothing or if we don't win, then the answer will just be that Amazon
752
00:54:23,992 --> 00:54:27,192
will pay, circle a ton of money and interest and take USDC.
753
00:54:27,192 --> 00:54:36,652
What does the world look like if one or a few big companies control the AI agents, the data, and the payment systems?
754
00:54:36,652 --> 00:54:52,492
Yeah, no, it's not great. Yeah, I think that's, to some extent, the way it's trending, and I think we should all be pushing back against OpenAI and Anthropic and even XAI and all these guys who just want to be the ones to control the models.
755
00:54:52,492 --> 00:54:55,132
they want to be the ones to have the only good models
756
00:54:55,132 --> 00:54:56,592
that no one else should
757
00:54:56,592 --> 00:54:58,112
open source models should be bad
758
00:54:58,112 --> 00:55:00,932
and only their proprietary models should be used
759
00:55:00,932 --> 00:55:02,292
and then once they do that
760
00:55:02,292 --> 00:55:05,212
then they have control over how you interact with the model
761
00:55:05,212 --> 00:55:06,552
presumably they'll run their own
762
00:55:06,552 --> 00:55:09,172
agent services that everyone will use those
763
00:55:09,172 --> 00:55:11,172
and not be able to use open source local agents
764
00:55:11,172 --> 00:55:15,012
and then also as a result of that
765
00:55:15,012 --> 00:55:16,752
get to pick the payment rails and everything else
766
00:55:16,752 --> 00:55:19,032
and I think it's definitely
767
00:55:19,032 --> 00:55:20,552
a very major concern right now
768
00:55:20,552 --> 00:55:27,292
is that these large companies are moving towards trying to own this,
769
00:55:27,632 --> 00:55:30,632
I mean, in their view, AGI in the entire future.
770
00:55:30,852 --> 00:55:33,712
I don't think that's necessarily the trend it's on,
771
00:55:33,932 --> 00:55:36,072
but own a huge part of the economy
772
00:55:36,072 --> 00:55:39,252
and a tool that powers a lot of parts of the economy.
773
00:55:40,912 --> 00:55:43,852
And then also, yeah, I mean, like people's data.
774
00:55:44,372 --> 00:55:49,292
Like already today, a ton of people are constantly asking LLMs
775
00:55:49,292 --> 00:55:56,212
about health questions, about personal questions, relationship questions. And all of that going in
776
00:55:56,212 --> 00:56:02,992
one big honeypot to anthropic and open AI is crazy. That is frightening for humanity, even
777
00:56:02,992 --> 00:56:08,292
just if it gets hacked or internal, like, you know, internal support people leak information.
778
00:56:08,292 --> 00:56:17,312
It's really frightening how the implications that might have in the future. So yeah, I'm worried
779
00:56:17,312 --> 00:56:22,412
about it. It is very expensive to train these models. I am not in a position to do very much
780
00:56:22,412 --> 00:56:30,812
about that, but it is very concerning. In 2022, PayPal proposed fining people $2,500
781
00:56:30,812 --> 00:56:37,372
for spreading misinformation. They walked it back, but the idea was there. Are we sleepwalking
782
00:56:37,372 --> 00:56:45,712
into a terrifying new control system? Yes. Yes. I mean, I don't know how else to put that,
783
00:56:45,712 --> 00:56:53,472
But like, yes, I think there's a very real shot that, you know, AI becomes this huge tool for control.
784
00:56:53,832 --> 00:56:56,732
A lot of people aren't searching for information themselves anymore.
785
00:56:56,852 --> 00:56:57,772
They're not using Google.
786
00:56:57,912 --> 00:57:04,592
They're not using Bing or whatever, or even searching on X for recent posts about it or searching on Reddit or whatever.
787
00:57:04,692 --> 00:57:12,172
They're just going and asking an LLM and whatever the LLM tells them, they assume it's true, which A, isn't how LLMs work.
788
00:57:12,172 --> 00:57:14,432
You can't assume what they tell you is true.
789
00:57:14,432 --> 00:57:19,492
But also, yeah, these companies could change how they respond.
790
00:57:19,672 --> 00:57:20,232
And they are.
791
00:57:20,312 --> 00:57:23,732
I mean, it is true that the Chinese models will,
792
00:57:23,972 --> 00:57:26,372
if you ask them about things that are politically sensitive in China,
793
00:57:26,472 --> 00:57:28,992
they'll give you the politically correct response for China.
794
00:57:29,212 --> 00:57:33,392
If you ask the Western models about politically sensitive things,
795
00:57:33,452 --> 00:57:35,832
like, for example, Israel and criticize Israel,
796
00:57:35,912 --> 00:57:39,192
they'll give you the politically correct answer for the West.
797
00:57:39,192 --> 00:57:44,312
And the amount of control that that has over shaping people's view
798
00:57:44,312 --> 00:57:46,112
is frightening.
799
00:57:46,512 --> 00:57:48,492
And the idea that, you know,
800
00:57:48,512 --> 00:57:51,032
only a few companies are going to have access to this
801
00:57:51,032 --> 00:57:53,572
and only a few companies are going to decide
802
00:57:53,572 --> 00:57:56,012
all these decisions on how these things get trained
803
00:57:56,012 --> 00:57:57,212
and what information they have
804
00:57:57,212 --> 00:58:02,292
and therein influence hundreds of millions
805
00:58:02,292 --> 00:58:05,412
and billions of people's view is crazy.
806
00:58:06,272 --> 00:58:08,272
That is a huge risk,
807
00:58:08,352 --> 00:58:10,372
like something we should all be very worried about
808
00:58:10,372 --> 00:58:11,832
and all be speaking out about.
809
00:58:11,832 --> 00:58:15,592
But also it costs billions of dollars to train these models.
810
00:58:15,812 --> 00:58:17,512
So I can't do anything about it.
811
00:58:17,592 --> 00:58:18,832
There's not that much I can do about it.
812
00:58:19,812 --> 00:58:20,572
Spread the word.
813
00:58:21,172 --> 00:58:21,352
Yeah.
814
00:58:21,972 --> 00:58:25,392
What needs to be built right now to stop that from happening?
815
00:58:26,692 --> 00:58:32,112
Yeah, I mean, I think it's a question of, and we are seeing some of this be built.
816
00:58:32,272 --> 00:58:34,892
We have, it's this question of having more competition.
817
00:58:35,632 --> 00:58:39,752
The good thing is some of these open source models are 90% as good, right?
818
00:58:39,752 --> 00:58:42,592
and that you can use them for most of the same things
819
00:58:42,592 --> 00:58:56,144
you could use any of the proprietary models for And yes they Chinese and they have certain Chinese views but it also open source You can retrain them You can actually go back and change
820
00:58:56,144 --> 00:58:58,344
how the model responds to certain questions.
821
00:58:58,724 --> 00:59:00,644
And that's way cheaper than training a new model.
822
00:59:00,824 --> 00:59:02,164
Like you can do that at home.
823
00:59:02,164 --> 00:59:05,144
I mean, okay, with maybe a $10,000 or $20,000 GPU,
824
00:59:05,464 --> 00:59:06,884
but you can do that at home.
825
00:59:07,024 --> 00:59:08,864
Like that's possible and people do do that.
826
00:59:09,004 --> 00:59:11,664
So, you know, making people aware of this issue
827
00:59:11,664 --> 00:59:26,604
And then also just having more competition and more chat interfaces and more different ways that people interact with these things so that there's more, hopefully, just more competition, really, I think is the big one.
828
00:59:27,444 --> 00:59:30,804
About Lightning, can an AI agent use Lightning today?
829
00:59:31,964 --> 00:59:32,884
Yeah, absolutely.
830
00:59:32,884 --> 00:59:38,304
There's a few wallets dedicated to AI agents specifically.
831
00:59:38,584 --> 00:59:39,844
So like MoneyDevKit is one.
832
00:59:39,984 --> 00:59:46,224
I think there's a PhoenixD wrapper or skill file that tells an agent how to interact with PhoenixD.
833
00:59:46,764 --> 00:59:49,964
There's also Lexi, I think, has an agent wallet and Albi.
834
00:59:50,284 --> 00:59:55,424
So there's a few Bitcoin companies trying to compete in this, trying to really enable AI agents to use Lightning.
835
00:59:55,824 --> 00:59:57,364
There's different models that's still early.
836
00:59:57,864 --> 01:00:00,004
You know, do you give your agent its own wallet?
837
01:00:00,104 --> 01:00:01,364
How much money do you put in it?
838
01:00:01,364 --> 01:00:06,284
do you make it so that you have to sign off on all the transactions?
839
01:00:06,604 --> 01:00:07,544
Is that too burdensome?
840
01:00:07,664 --> 01:00:11,324
So there's a lot of things to figure out and a lot of experimentation to have to be done here.
841
01:00:11,864 --> 01:00:16,844
But yes, there are several projects that you can use to go give your AI agent a wallet.
842
01:00:17,564 --> 01:00:23,284
If AI agents start using Lightning at scale, what part of the network struggles first?
843
01:00:23,944 --> 01:00:27,364
And can ARK or eCache be part of the solution?
844
01:00:27,364 --> 01:00:35,324
yeah i mean so into the second question is yes um is that there's not much different about an
845
01:00:35,324 --> 01:00:42,204
ai agent from a human in terms of uh what they can use in terms of bitcoin wallets um there may
846
01:00:42,204 --> 01:00:45,944
be a little more forgiving on the ux front they're a little more willing to deal with something that's
847
01:00:45,944 --> 01:00:49,504
uh frustrating to use because they they have a lot more patience than humans do
848
01:00:49,504 --> 01:00:52,104
but yeah I mean I think
849
01:00:52,104 --> 01:00:53,904
at scale like at
850
01:00:53,904 --> 01:00:54,884
huge scale
851
01:00:54,884 --> 01:00:58,004
you know we'd probably be okay just
852
01:00:58,004 --> 01:00:58,704
because
853
01:00:58,704 --> 01:01:01,404
most
854
01:01:01,404 --> 01:01:04,044
people at the end of the day
855
01:01:04,044 --> 01:01:05,984
will use a centralized
856
01:01:05,984 --> 01:01:08,104
they're not going to run their own agent software locally
857
01:01:08,104 --> 01:01:09,944
they're going to use something hosted for them
858
01:01:09,944 --> 01:01:11,944
by someone and that
859
01:01:11,944 --> 01:01:14,184
something hosted for them can also
860
01:01:14,184 --> 01:01:16,024
just provide custodial bitcoin services
861
01:01:16,024 --> 01:01:18,084
and so like
862
01:01:18,084 --> 01:01:21,984
if you're going to use a hosted agent, you can also just use their hosted wallet.
863
01:01:22,464 --> 01:01:27,704
But yeah, I mean, if we actually had millions and millions and tens of millions and hundreds
864
01:01:27,704 --> 01:01:32,184
of millions of people trying to onboard their AI agents to Lightning wallets self-custodially
865
01:01:32,184 --> 01:01:40,224
all at once, then yeah, we would probably need something like ARK to enable more Lightning
866
01:01:40,224 --> 01:01:44,624
channels to be opened that don't require all the on-chain space.
867
01:01:44,624 --> 01:01:52,964
what can bitcoin do that visa or banks can't and is he ready for it i think it's it's the
868
01:01:52,964 --> 01:02:02,384
credible neutrality part like it is it is not controlled by someone and you you can trust you
869
01:02:02,384 --> 01:02:07,244
can trust it like you you can have your own and it is yours and i do think that is compelling to
870
01:02:07,244 --> 01:02:14,284
a lot of people but frankly the the most compelling thing like the the reason people might actually
871
01:02:14,284 --> 01:02:18,644
I use Bitcoin for this. It's just that banks, like these large companies are slow to change.
872
01:02:18,644 --> 01:02:23,584
It's very, it is so entrenched. Like Visa, Visa is not a company, right? Visa is a co-op
873
01:02:23,584 --> 01:02:31,504
between many large banks. And so in order to say, oh, this like merchant liability thing
874
01:02:31,504 --> 01:02:37,124
that we've been doing doesn't scale to AI agents. We need to change the liability structure
875
01:02:37,124 --> 01:02:44,164
of the Visa network. Visa can't just do that. Visa has to get the member banks to actually want
876
01:02:44,164 --> 01:02:49,524
to do that and make that the new reality, they could in theory, I think there's actually
877
01:02:49,524 --> 01:02:53,784
some legal issues in the United States, but they could pay off enough senators to get
878
01:02:53,784 --> 01:02:54,784
that law changed.
879
01:02:56,304 --> 01:02:57,264
But they won't.
880
01:02:58,244 --> 01:03:04,284
In order to convince all of these large legacy institutions to change on a dime and fundamentally
881
01:03:04,284 --> 01:03:08,684
change a core part of their payment rail, all of a sudden is just not going to happen.
882
01:03:08,684 --> 01:03:15,284
And so that's why you see people instead choose to focus on stable coins for AI agentic payments.
883
01:03:17,084 --> 01:03:21,544
And hopefully that's also where Bitcoin provides the same set of values.
884
01:03:21,704 --> 01:03:26,124
And then hopefully Bitcoin has a shot there and that people start building towards it.
885
01:03:26,324 --> 01:03:27,804
And you think Bitcoin is ready for it?
886
01:03:28,584 --> 01:03:29,884
Yeah, I think it is.
887
01:03:29,884 --> 01:03:38,744
I again, in part, because if it does scale to that size, most people are going to be
888
01:03:38,744 --> 01:03:40,644
using centralized things.
889
01:03:40,644 --> 01:03:44,884
And those centralized things can just offer custodial balances and wouldn't offer Bitcoin
890
01:03:44,884 --> 01:03:45,444
balances, right?
891
01:03:45,464 --> 01:03:47,524
Like most people don't want a Bitcoin balance.
892
01:03:47,604 --> 01:03:48,544
They want a dollar balance.
893
01:03:48,544 --> 01:03:53,304
And then their service provider will just buy some Bitcoin and pay the merchant and
894
01:03:53,304 --> 01:03:54,084
that's it.
895
01:03:54,144 --> 01:03:55,984
And it'll pull from your dollar balance, right?
896
01:03:56,204 --> 01:03:59,284
So that would work totally fine today, right?
897
01:03:59,284 --> 01:04:06,424
That Lightning can scale very substantially with a small handful of very, very large custodial services running large notes.
898
01:04:06,944 --> 01:04:07,964
And is that what we want?
899
01:04:10,024 --> 01:04:16,504
We want people to have the option to pick who they trust.
900
01:04:17,484 --> 01:04:20,704
Because that's not really something you can do today with the traditional finance rails.
901
01:04:21,684 --> 01:04:25,484
The reality is most people are not going to self-custody.
902
01:04:25,484 --> 01:04:34,124
my job is to reduce the ux difference between self-custody and custodial as much as possible
903
01:04:34,124 --> 01:04:41,944
so that as many people select to do self-custody as possible but we're never going to get better
904
01:04:41,944 --> 01:04:46,744
like a centralized you know system is always going to be more efficient than the decentralized one
905
01:04:46,744 --> 01:04:52,384
and then the self-custodial one and so um you know people will use those custodial systems
906
01:04:52,384 --> 01:04:53,404
and that's fine.
907
01:04:53,764 --> 01:04:55,804
And it's fine that many people do that
908
01:04:55,804 --> 01:04:58,564
and my job is to make it so that fewer people do that
909
01:04:58,564 --> 01:04:59,804
and more people do self-custody.
910
01:05:00,264 --> 01:05:03,304
But the reality is simply that most people
911
01:05:03,304 --> 01:05:05,404
would just onboard to something that's custodial
912
01:05:05,404 --> 01:05:07,204
and they'll pick who they trust
913
01:05:07,204 --> 01:05:08,864
and they'll only use something they trust
914
01:05:08,864 --> 01:05:11,124
and that's still much better than the existing system.
915
01:05:12,084 --> 01:05:16,064
Could AI agents be what finally bring Bitcoin to the masses?
916
01:05:16,384 --> 01:05:16,984
It's possible.
917
01:05:18,044 --> 01:05:19,864
I think, again, I think it's a long shot,
918
01:05:20,244 --> 01:05:22,044
but I do think we really do have a shot here.
919
01:05:22,044 --> 01:05:38,856
Like that that if we really push on this and uh enough AI agent tooling integrates Bitcoin natively and then people start actually using it and talking to merchants and getting online merchants especially to start offering Bitcoin
920
01:05:38,856 --> 01:05:43,476
as payment rails. Then again, it just snowballs, right? It becomes the default and it snowballs
921
01:05:43,476 --> 01:05:49,836
very quickly. And it's a long shot because Coinbase and Circle and PayPal and whoever else
922
01:05:49,836 --> 01:05:54,176
are all trying to push their stable coins and put a lot of money behind their stable coins to be the
923
01:05:54,176 --> 01:06:01,176
winner here, but maybe not. Like the game is wide open. Bitcoin has never had an opportunity like
924
01:06:01,176 --> 01:06:07,916
this because Bitcoin has always been the underdog against traditional finance that has infinitely
925
01:06:07,916 --> 01:06:12,796
more resources and is the entrenched player. Right now, there is no entrenched player in
926
01:06:12,796 --> 01:06:19,656
agentic payments because it's too early. It barely exists yet. So anyone, it's still anyone's game.
927
01:06:19,656 --> 01:06:21,476
Anyone could become the entrenched player.
928
01:06:22,016 --> 01:06:23,736
And Bitcoin has a shot there.
929
01:06:24,356 --> 01:06:29,936
Again, it's a long shot, but I think we have to try because we've never had an opportunity
930
01:06:29,936 --> 01:06:30,516
like this.
931
01:06:30,556 --> 01:06:31,136
This is crazy.
932
01:06:31,776 --> 01:06:32,816
But here is the thing.
933
01:06:33,616 --> 01:06:40,316
If every coffee purchase triggers a tax event, how can Bitcoin succeed as practical everyday
934
01:06:40,316 --> 01:06:40,716
money?
935
01:06:41,256 --> 01:06:43,956
Yeah, I think that's definitely a big concern.
936
01:06:44,496 --> 01:06:48,916
It's something that blocks lobbyists have spent a lot of time trying to get Congress
937
01:06:48,916 --> 01:06:49,436
to fix.
938
01:06:49,656 --> 01:06:51,956
But yeah, I mean, I think that's a big concern.
939
01:06:52,956 --> 01:06:58,196
To some extent, the law has to change because it just doesn't, you know, if AI agents start
940
01:06:58,196 --> 01:07:00,436
doing this, how does it work?
941
01:07:00,516 --> 01:07:01,656
Like, who pays the tax?
942
01:07:01,716 --> 01:07:06,676
Like, even ignoring Bitcoin for a moment, if AI agents start, you know, sports betting,
943
01:07:06,676 --> 01:07:11,996
like they go start making sports bets using Bitcoin or maybe using stable coins, how does
944
01:07:11,996 --> 01:07:12,816
that tax work?
945
01:07:13,156 --> 01:07:16,436
How, who, you know, they're not an entity, they're not a person.
946
01:07:16,576 --> 01:07:18,396
In many cases, they're not anything.
947
01:07:18,396 --> 01:07:19,896
They might not have a company.
948
01:07:20,196 --> 01:07:23,416
It might have been spawned by a person, but maybe it's off doing its own thing.
949
01:07:24,096 --> 01:07:24,976
It owes taxes.
950
01:07:25,976 --> 01:07:26,896
But how?
951
01:07:27,356 --> 01:07:28,216
How does that work?
952
01:07:28,336 --> 01:07:32,496
So, you know, there's a lot here that's just going to need to get figured out.
953
01:07:32,596 --> 01:07:38,156
And I think it's not worth getting too hung up over it yet.
954
01:07:39,136 --> 01:07:44,936
But also, yeah, I mean, we should, to the greatest extent possible, we need to fight to get that law changed.
955
01:07:45,756 --> 01:07:52,496
And can Bitcoin win in a world where governments push stable coins and CBDCs they can't control and censor?
956
01:07:53,056 --> 01:07:59,356
Yeah, I think that's a good point too, that the governments are going to try to push something they prefer.
957
01:07:59,816 --> 01:08:08,196
I just, I keep coming back to like defaults are so sticky that it is so hard to break an entrenched monopoly that everyone uses.
958
01:08:08,276 --> 01:08:10,396
Everyone uses Visa, everyone uses MasterCard.
959
01:08:10,596 --> 01:08:13,736
It is so hard to get anyone to care about anything else.
960
01:08:13,736 --> 01:08:21,776
And it's great that Bitcoin is cheaper in most cases in the United States, at least, than Visa and MasterCard for merchants and that merchants might want to use Bitcoin.
961
01:08:22,196 --> 01:08:28,016
But it's just so hard to break that entrenched monopoly in a two-sided market where both buyers and sellers have to have the same payment rail.
962
01:08:28,016 --> 01:08:37,656
And so, yeah, I don't know that all the pressure of governments would be enough to break that entrenched monopoly, assuming we were to win.
963
01:08:37,656 --> 01:08:43,236
because it's in their interest to put more and more regulations on bitcoin in order to
964
01:08:43,236 --> 01:08:50,256
yeah to push cbsc and it would slow it down but um you know if we if we got to the point where
965
01:08:50,256 --> 01:08:55,476
people are using open agents and then making merchant payments and all of a sudden that
966
01:08:55,476 --> 01:09:00,196
that workflow stops working because governments have added some new regulation on it and all of
967
01:09:00,196 --> 01:09:05,596
a sudden your your agent has to kyc the the human that started it in order to make a payment
968
01:09:05,596 --> 01:09:12,776
And people would just complain like the one good thing about democracy is that it is still
969
01:09:12,776 --> 01:09:20,256
the case that even in the US or wherever else, and most democracies, if a huge volume of
970
01:09:20,256 --> 01:09:24,456
people complain, they will fix, they will undo it.
971
01:09:24,456 --> 01:09:30,216
So there is still a check there that would prevent totally wrecking it if it were to
972
01:09:30,216 --> 01:09:31,436
become the dominant player.
973
01:09:31,436 --> 01:09:44,076
In your article, Bitcoin has a golden opportunity with AI agents, you argue that Bitcoiners need to start using Lightning and AI agents with wallets so merchants begin paying attention.
974
01:09:44,976 --> 01:09:46,076
How do we do that?
975
01:09:47,216 --> 01:09:48,196
Yeah, that's a good question.
976
01:09:48,276 --> 01:09:51,336
I mean, you know, to a large extent right now, it's reaching out to merchants.
977
01:09:51,496 --> 01:09:54,496
It's talking to merchants and saying, hey, I want to buy your widgets.
978
01:09:54,956 --> 01:09:56,396
Why aren't you taking Bitcoin?
979
01:09:56,636 --> 01:09:58,476
Like, I want my agent to buy.
980
01:09:58,556 --> 01:10:00,036
Like, I need to buy your widgets once a month.
981
01:10:00,036 --> 01:10:06,416
maybe it's my subscription service to a stake as you know good farm maybe it's whatever it is
982
01:10:06,416 --> 01:10:12,376
you know i i need you i want you to enable bitcoin payments so that my agent can just do this for me
983
01:10:12,376 --> 01:10:16,956
so it's it's right now it's just reaching out to merchants and talking to them and saying look i
984
01:10:16,956 --> 01:10:21,816
want to do this thing you'll save two percent three percent on visa fees i will get my agent
985
01:10:21,816 --> 01:10:25,536
to be able to do this and i don't have to try to use visa which doesn't really work for this
986
01:10:25,536 --> 01:10:40,856
So I think that's the big step right now is just, you know, trying these things, getting comfortable with them, using agents, seeing what you can do with them, and then reaching out to merchants and just saying like, hey, I want to use your service and buy from you, but I need you to take Bitcoin for it.
987
01:10:41,456 --> 01:10:43,996
What projects are already being built around this?
988
01:10:43,996 --> 01:10:51,936
sadly we haven't gotten default integration into ai agent software yet uh for bitcoin wallets but
989
01:10:51,936 --> 01:10:56,176
it is you know that one of the good things about all of them is they're they're pretty good at at
990
01:10:56,176 --> 01:11:00,856
you know you tell them i say here's some context and you set the context of like well you you have
991
01:11:00,856 --> 01:11:05,616
a local bitcoin wallet here's how you access it um they're they're good at just doing that and
992
01:11:05,616 --> 01:11:10,496
they can take advantage of that so so i mentioned a few earlier there's like money dev kid and and
993
01:11:10,496 --> 01:11:16,876
And Albi and Lexi and PhoenixD, I think, all have options for giving your agent a wallet.
994
01:11:17,036 --> 01:11:21,096
And I think that's basically the only thing that needs to be built is just agents to have wallets.
995
01:11:21,176 --> 01:11:22,616
But then we just need people to use it.
996
01:11:22,656 --> 01:11:24,076
So for the most part, stuff has been built.
997
01:11:24,756 --> 01:11:29,096
There's still a lot to be figured out there and having the right UX and getting it better and whatever.
998
01:11:29,256 --> 01:11:31,416
But for the most part, it's just people need to use it.
999
01:11:31,496 --> 01:11:33,096
People need to tell merchants they want to use it.
1000
01:11:33,096 --> 01:11:36,256
What do you think developers are underestimating?
1001
01:11:36,256 --> 01:11:43,296
I think I often and some developers often underestimate how long it takes people to change.
1002
01:11:43,476 --> 01:11:54,396
So you see this a lot in the software engineering world right now around AI is, you know, software engineers have adopted LLMs very substantially in their work.
1003
01:11:55,236 --> 01:12:01,056
You know, surveys and large companies and software engineers are using LLMs drastically and companies are pushing that on their employees.
1004
01:12:01,056 --> 01:12:07,696
But in many cases, their employees are just using them because, you know, these are great tools that can help you do a lot of your work much more effectively.
1005
01:12:08,156 --> 01:12:23,628
But I think then people extrapolate that onto the world and assume that everyone in the world is using LLMs a lot for their work And I don think that true at all For the most part people have adopted LLMs as a chat interface as a replacement for Google in many cases
1006
01:12:24,068 --> 01:12:32,028
but not actually using agents that can do tasks on their behalf. And I think that is going to take
1007
01:12:32,028 --> 01:12:39,848
a while to get traction. And so all these companies like rushing to do agentic payments,
1008
01:12:39,848 --> 01:12:45,008
I think are, you know, getting a little ahead of themselves. But at the same time, you know,
1009
01:12:45,028 --> 01:12:50,308
at some point that adoption will kind of flip over, like it'll get over the hump. And then all
1010
01:12:50,308 --> 01:12:54,348
of a sudden it'll be adopted very quickly once people start to realize this and share with their
1011
01:12:54,348 --> 01:12:59,908
friends what they're doing and whatever. And so I think we want to get ahead of that as much as
1012
01:12:59,908 --> 01:13:04,168
possible just like everybody else is trying to get ahead of that but we want to also try to compete
1013
01:13:04,168 --> 01:13:10,048
and get out ahead of that before it happens if you had to guess who wins this race yeah stable coins
1014
01:13:10,048 --> 01:13:16,048
if i had to guess it's probably going to be stable coins i don't know which one like uh
1015
01:13:16,048 --> 01:13:21,148
because of regulatory changes in the united states now much easier for a company to issue a stable
1016
01:13:21,148 --> 01:13:26,208
coin and so i don't actually know that usdc is necessarily going to win uh it might be paypal
1017
01:13:26,208 --> 01:13:29,308
USD or it might be someone, you know, maybe the banks are issuing stable coins.
1018
01:13:29,588 --> 01:13:31,388
I know some of them are looking at that.
1019
01:13:32,948 --> 01:13:36,708
So I don't know which stable coin, but if I had to guess stable coins are going to win,
1020
01:13:37,248 --> 01:13:41,748
that doesn't mean we don't have a shot at this, but they have a better shot.
1021
01:13:41,928 --> 01:13:43,608
They have more money to spend on doing it.
1022
01:13:44,268 --> 01:13:46,708
Do you think people are ready for this shift?
1023
01:13:47,548 --> 01:13:48,208
Yes and no.
1024
01:13:48,308 --> 01:13:50,468
Like people aren't thinking about it.
1025
01:13:50,508 --> 01:13:51,168
People aren't aware of it.
1026
01:13:51,228 --> 01:13:52,128
People aren't doing it.
1027
01:13:52,128 --> 01:13:57,028
but at the same time, once people start to do something, it becomes very natural, right?
1028
01:13:57,028 --> 01:14:01,248
So like you look at software engineer, you look, I look at my work, whereas like, you know, up until
1029
01:14:01,248 --> 01:14:08,868
the end of last year, LLMs weren't really good enough at coding for them to be like all that
1030
01:14:08,868 --> 01:14:12,768
useful for my work. You know, they were useful for some small stuff here and there, but not very
1031
01:14:12,768 --> 01:14:16,288
much. And then all of a sudden they, they kind of reached this point where they were fairly useful.
1032
01:14:16,448 --> 01:14:21,808
And then all of a sudden, a lot of engineer, a lot of software engineers who had not been
1033
01:14:21,808 --> 01:14:27,668
adopting them or been adopting them very loosely, all of a sudden, all adopted them very quickly,
1034
01:14:27,828 --> 01:14:31,648
right, right in succession, because people talk to each other and talk about how much more capable
1035
01:14:31,648 --> 01:14:35,868
these models suddenly were. Not that they were like, you know, 10x more capable, but that they
1036
01:14:35,868 --> 01:14:40,568
had finally reached some threshold where they were a lot more capable. And so I think like once,
1037
01:14:40,808 --> 01:14:45,388
once it becomes the default, it'll just be very natural, it'll feel very natural to people. And
1038
01:14:45,388 --> 01:14:49,108
people will be like, how did we never, how did we not do this before? But at the same time,
1039
01:14:49,108 --> 01:14:54,888
right now today, people can't quite model like, oh, we're just going to like chat with an agent
1040
01:14:54,888 --> 01:14:57,928
that's going to buy us stuff. People are like, oh, that's weird. I can just buy stuff on Amazon.
1041
01:14:58,048 --> 01:15:04,408
Why do I care? In fixing payments for AI agents, what door are we opening? What are the unintended
1042
01:15:04,408 --> 01:15:10,228
consequences? Yeah, I mean, I think there's a lot of risk to the future of AI. Like you asked
1043
01:15:10,228 --> 01:15:14,988
earlier about centralization concerns around these large companies controlling AI. And I think that's
1044
01:15:14,988 --> 01:15:21,828
That's probably one of the most obvious potential consequences here is that people end up using
1045
01:15:21,828 --> 01:15:25,308
one or two providers, you know, maybe Google wins, maybe OpenAI wins, whatever.
1046
01:15:25,948 --> 01:15:29,348
And that that's the way everyone buys everything now.
1047
01:15:29,608 --> 01:15:34,508
And again, now this company has a lot more control than they already had, or just a lot
1048
01:15:34,508 --> 01:15:39,128
of control over, you know, large parts of the economy in addition to what people think
1049
01:15:39,128 --> 01:15:39,488
and whatever.
1050
01:15:39,488 --> 01:15:47,148
So a lot of data, a lot of control over like, oh, well, you know, you told it to buy a pair of shoes with these properties.
1051
01:15:47,348 --> 01:15:48,708
Well, it could pick that one or that one.
1052
01:15:48,748 --> 01:15:51,768
And it's going to always pick that one because they paid for the ad space or whatever, you know.
1053
01:15:51,788 --> 01:16:00,508
So there's a lot of potential control that these companies might gain as a result of this that I think is very concerning.
1054
01:16:01,088 --> 01:16:05,928
I think that's the most obvious one right now just because it's already starting to play out that way.
1055
01:16:05,928 --> 01:16:07,388
You know, it's hard to see the future.
1056
01:16:07,388 --> 01:16:16,668
This stuff is moving so fast and because it is very different in the type of tool that it is versus what we've had, what humanity has had in the past.
1057
01:16:16,928 --> 01:16:20,788
It is kind of hard to predict exactly how it's going to play out.
1058
01:16:20,908 --> 01:16:24,388
But that's the obvious one that comes to mind because it's kind of already happening.
1059
01:16:24,808 --> 01:16:29,528
And last one, if someone listening thinks, this sounds big, but what can I do?
1060
01:16:29,888 --> 01:16:30,668
What's your answer?
1061
01:16:30,668 --> 01:17:00,648
Yeah, I mean, again, this like the future, how this is going to be decided is going to be in small starts, right? It's going to be a few merchants accepting these and people using it and people buying from them using Bitcoin and then a few more merchants and then a few more merchants. And it can snowball very, very quickly. Once you get a few merchants who talk to each other, and then talk to we talk to more talk to their friends, it can snowball very quickly. It's really just about getting that momentum going. So getting online merchants taking Bitcoin for things people want to buy.
1062
01:17:00,668 --> 01:17:02,128
and want to buy regularly, especially.
1063
01:17:03,128 --> 01:17:05,628
And starting to let that grow,
1064
01:17:05,848 --> 01:17:09,688
let a lot of people start starting to use agents,
1065
01:17:09,948 --> 01:17:11,968
give them a Bitcoin wallet, play with it,
1066
01:17:12,028 --> 01:17:13,148
give feedback to developers
1067
01:17:13,148 --> 01:17:15,588
so that people can improve the user experience
1068
01:17:15,588 --> 01:17:18,268
of those wallets and how that stuff works.
1069
01:17:18,928 --> 01:17:20,408
And then just, yeah, I mean,
1070
01:17:20,408 --> 01:17:22,568
just actually buying for merchants using Bitcoin,
1071
01:17:22,768 --> 01:17:25,088
especially online and using AI agents
1072
01:17:25,088 --> 01:17:27,448
and giving merchants website feedback on,
1073
01:17:27,868 --> 01:17:30,148
you know, you have to make your website more agent friendly.
1074
01:17:30,148 --> 01:17:36,968
you have to do X or Y or whatever. And so I think it's just starting because the earlier we start,
1075
01:17:37,088 --> 01:17:43,348
the better the tooling we can have, you know, the more feedback developers and merchants can get on
1076
01:17:43,348 --> 01:17:49,368
the tooling so that we can have the best possible foot forward once that race starts in earnest.
1077
01:17:49,848 --> 01:17:54,228
And then hopefully we'll have enough of a critical mass that it can start to grow and
1078
01:17:54,228 --> 01:17:59,188
really snowball from there. So I think it really is, this is going to start very small. And the
1079
01:17:59,188 --> 01:18:01,828
really small steps can have a really huge impact.
1080
01:18:03,248 --> 01:18:07,248
Matt, thank you for your time and this generous conversation.
1081
01:18:07,928 --> 01:18:10,108
Where can people find you and your work?
1082
01:18:11,168 --> 01:18:12,088
Yeah, no, thanks.
1083
01:18:12,208 --> 01:18:13,668
This has been really fun.
1084
01:18:14,468 --> 01:18:19,228
People can find me for the most part on X, the blue mat on X.
1085
01:18:19,848 --> 01:18:21,948
And then I have a blog that's linked from there.
1086
01:18:22,008 --> 01:18:23,668
And then you can follow some of my work from there.
1087
01:18:23,968 --> 01:18:24,708
Are you a Nostra?
1088
01:18:25,428 --> 01:18:26,928
Yes, I have a Nostra account.
1089
01:18:27,528 --> 01:18:28,908
How do I share my Nostra?
1090
01:18:28,908 --> 01:18:29,928
I mean, I have an end pub.
1091
01:18:30,008 --> 01:18:31,388
I can read my full end pub.
1092
01:18:31,728 --> 01:18:32,888
We will put it on the show notes.
1093
01:18:32,888 --> 01:18:35,728
Click the link in the show notes to find my end pub.
1094
01:18:36,048 --> 01:18:37,588
It's been a pleasure having you on.
1095
01:18:37,708 --> 01:18:39,248
I can't wait to have you back.
1096
01:18:39,468 --> 01:18:41,148
Until then, stay sovereign.
1097
01:18:41,608 --> 01:18:42,268
Yeah, it's been great.
1098
01:18:42,368 --> 01:18:42,568
Thanks.