← Citadel Dispatch
Loading summary
Transcript128 lines
- [00:00]
A
Sa.
- [00:29]
B
Foreign Bitcoin Wednesday Freaks, it's your host Odell here for another Citadel Dispatch. The show focused on actual bitcoin and freedom tech discussion. It is currently Wednesday, March 25, 1600 UTC. You guys will probably be listening to this in a few hours. The current block height is 942174. Current stats per dollar is 1412. The price for one bitcoin is now $70,795. Bitcoin in gold terms. One bitcoin can now buy you 15.83 ounces. But you might regret it if you do it. We're up on the week in gold 7%. We're up on the month 20.7% and we're still down on the year 45%. But I expect that to turn around anyway. Freaks as always, Dispatch is purely funded by our audience, by viewers like you. Thank you for sending your hard earned sats to support the show. It truly means a lot. The Largest we have no ads or sponsors. You guys are the pure value for value experience. Largest zap from last week was kind of light on zaps from our show last week. I guess it was a Friday show. Maybe some of you guys haven't listened to it yet. If you haven't, make sure you go listen to it. Largest app was 10,000 sats from Rider Die Freak Map 21 he said great rip. That rip, if you guys haven't listened to it yet is with Evgeny who is the founder and maintainer of Simplex. It was a conversation that I particularly enjoyed a lot. I mean, it's kind of like a celebrity to me, but it was a lot. It was a lot of fun and I think you guys should listen to it if you haven't. But anyway, Freaks, we have a great show lined up today. I'm trying to keep the ball rolling with High Signal. Good quality conversation. If I can't find something High Signal, we just skip the week because I don't have ads or sponsors that require me to do three shows a week like all the other shows. But before we get started, as always, all the relevant links are@sealedispatch.com, share with your friends and family. Available in every podcast app by just searching Citadel Dispatch we have a return guest today. We have Matt Alborg of ppq AI. Yes, it's another AI focused show. How's it going Matt?
- [03:05]
A
Yeah, it's good. Good. Very good Matt.
- [03:08]
B
Thanks for having me back. The mats are in the house. We last had Matt on the Show July 14, 2025 Episode 168. I was just starting to dip my toes into AI stuff. We've come a long way. It felt like it was time for a return appearance and an update conversation. What a year it's been, huh?
- [03:36]
A
Yeah. Yeah. I feel about 10 years older. So, yeah, a lot is happening.
- [03:43]
B
I mean, you have been in bitcoin for a while as well. I mean, to me, it feels very much like early bitcoin, right?
- [03:52]
A
Yeah, it does. It feels very Wild West. The rules are being made as we go. And yeah, there's a lot of creativity out there on what you can do, what you can build, how you can use this stuff. Yeah, there's no template really, for what's happening right now.
- [04:13]
B
Yeah, like early bitcoin.
- [04:15]
A
Scary and fun. Sorry, sorry. Yeah, sorry. Go ahead, Matt.
- [04:21]
B
No, there's no veterans. Right. There's no gray beards. There's no people who have been doing this for 30 years because the tech just hasn't existed. I feel like my mind is racing. Like you said. I feel like a year is a decade. That was, to me, those are like the templates of my early bitcoin experience.
- [04:38]
A
Yeah. So what I was going to say and I, I, yeah, it's relevant here is just on the call with, with the guys today, the teammates. We're each going over our development workflow. And you know, in AI, there's, you know, you can find guides on how to, like, upskill your development workflow and such, but we're all just kind of inventing it as. As we go. You know, I'm using GitHub work trees and VS code, and one of my teammates is using Team Teamux and he's generating, like, ephemeral work trees. And then we, we all had different Claude MD files and we're using different testing agents and stuff like that. So there's really. Yeah, it's, it's, it's. We're all creating our own thing right now and just trying to, like, become as efficient as possible.
- [05:23]
B
Wild. How big is your team now?
- [05:27]
A
Three. We're actually. I don't know what we were last time, but we were at five at one point and two of the guys just weren't working out. And so we're down to a lean and mean three. But I actually feel like we're getting a lot more done now. We have a really good rhythm and yeah, kind of as we onboard people in the future, I, I just, I really want to, you know, make sure that it's the right chemistry and the right, right mentality and all that stuff.
- [05:56]
B
So yeah, I mean, hiring's hard.
- [06:00]
A
I
- [06:02]
B
people, it's hard to, you know, every people problems are hard problems. And then, I mean, I do think, I mean maybe we're, we seem to be entering the era of the lean team, the lean small team that gets shit done. We already kind of were with bitcoin, I think, you know, particularly founders, I see this on the 1031 side all the time. Founders that think in bitcoin as opportunity costs think very carefully about expanding their payroll because it's less bitcoin that they can stack and they're thinking long term term. And then you couple that all of a sudden with these AI tools and three people can do the work that maybe 30 or 50 people were doing previously. It's kind of a wild compounding effect.
- [06:47]
A
Yeah, I absolutely am a big believer in that. Maybe not three. I would like to get up to maybe 10 or so, but I think a team of 10 can do really, really crazy things now. And yeah, as you said, one person is now, I don't know if it's 30 or 50, but it's, it's a lot. And yeah, you just have to know how to use AI. That's, that's really the big thing. And that was kind of a decision maker in, you know, when we hired people on and such. It's, you have to really want to learn AI. You have to live and breathe it in order to really level up. I think that there's probably a lot of people out there who, they're just kind of casually using AI and they're not treating it like, they're not giving it the respect that, that it deserves in terms of how much it can really increase your productivity. And so yeah, I, I was just thinking the other day about how, you know, I'm not, I don't have a CS degree myself. I have a half of one. So I'm not totally non technical, but I did, you know, I went to, I have like half a degree and then I went to a couple of the coding boot camps and then, you know, I've just been kind of grinding on, on stuff with AI since. And yeah, my thought was like for your marketer, the marketers of the future, will they be the ones, will they be the marketing native people who learn how to use AI or will it be the developers who learn the marketing side? Obviously it's gonna be a little bit of both. But I do wonder like which trade is going to learn faster or have the, the brain to like master the. Enough of the other trade to become like really potent. Right. Because yeah, I would love to. When I did hire on kind of a community manager person, there was a lot of blockers happening because he was not able to spin up his own like analytics database or his own analytics dashboards. And so it was always waiting on me like to create these dashboards and stuff. I've got a lot on my plate anyways so it would have been amazing if like you know, you bring in a community person, a marketing person who can also like really vibe on cloud code or something like that and then they're just doing all of the technical work themselves. That's what I really think is like really important. If you're like trying to, trying to get a job these days, you have to be really good at AI.
- [09:13]
B
Yeah, I mean it's a weird thing to think about. I mean at the definitely high level like on the surface like what you see a lot of people is like they're using these things as maybe like an advanced Google search and that's just barely scratching the surface of what it's capable of doing. Yeah, to your point. But on the, on the topic of what type of person does it benefit the most? I mean it definitely depends on what you know, role is trying to be filled or what qualities are needed in a person. But I'm not, I don't have necessarily a hard answer there. I think obviously trained developers have an advantage right now just because you need to have, if you have some level of code competency you can really get super powered on these tools. But I do kind of see it. I mean I, for instance, I have no coding background whatsoever. I can kind of read code, I know how to like get around to GitHub repo. I'm like a technically competent non dev and you know like my big claim to fame was like I, I found an issue with, I was like looking through the code of, of greenwallet which is now blockstream Wallet and I found an issue that a bunch of their geniuses missed. But like I'm, I'm definitely not a dev but these tools really superpower me because I'm technical enough to like start to leverage them and see what's happening and get my feet wet. And I think that type of person benefits a lot. I think the top tier devs benefit a lot. And then I think maybe. And once again this is all just speculation. We're so early and that's one of the things that's so fascinating to me. I think those like mid level devs are Those are the ones that get really commoditized in this era. I don't know if they have such a huge advantage over, you know, a technically competent non dev that makes sense.
- [11:06]
A
I don't even know if I would like bisect it that way. I mean, I think it just all comes down to do you have a willingness to learn? Do you have a willingness to change up your current ways? I think that some, some of the folks on the senior side are struggling a little bit because they were really good at something for a long time and, and they do see a lot of the mistakes that AI makes, but they, I guess, I think, I think they wrap their ego a little bit in their abilities and they're a little bit off put by the fact that like an AI can ship now 10 times more than you can. And yes, it does make errors and such which you have to correct, but like it's still shipping 10x. I think that it doesn't matter who you are, whether you're a junior, mid level, whatever. It's, it's like, do you have the humbleness to like learn and respect this stuff and just keep learning and like I said, you have to forge your own path. There aren't a lot of guides out there right now to do this, do that. There are, but I personally just rather dig in and like kind of figure things out myself and that, that's, that's really what it comes down to. I also think that maybe, maybe developers who don't have a little bit of the business mindset also may struggle, you know, if they're perfecting something that really, you know, maybe they're building something beautiful. But you should always be asking the question, how much time am I spending on this thing? And, and how is it going to impact the bottom line? Or like how does it compare to all these other things we're working on? So yeah, I think you can't just be the technical guy living in the basement anymore either. You have to be, you know, you have to be a little more aware of the other facets as well.
- [12:56]
B
Yeah, I think that's a good point in perspective. It'll be very interesting watching how it all evolves. I mean, just jump into the meat here.
- [13:03]
A
We.
- [13:03]
B
One of the key, from what I remember on our last conversation, I think one of the key conversation, key conversation points we were having is from my seat, there's two ways people are interacting with these tools. Two main ways people are interacting with these tools. One is the relatively controlled and proprietary subscription model. So people are subscribing to anthropic or OpenAI or Google directly or Grok directly. Usually that involves KYC, at least some level of KYC. Anthropic require, you know, Claude requires phone number verification. They're actively blocking, they're actively blocking burner numbers. Obviously then they have you pay with credit cards. So then that is like another level of KYC to identify who's using it. They also came out and said specifically that some of these open source Chinese models are training against them and called out by name certain researchers on the, on the Chinese side that were using their accounts with nims. But because of their prompt history and how they were doing it, they were able to reverse calculate who they were regardless of the other KYC elements. So you have a bunch of privacy issues there. But as a result you're getting a significant amount of value. I mean, I think the most. The latest report said that your $200 Claude subscription is getting you about $5,000 worth of credits a month right now. So Anthropic is just subsidizing it. They're the drug dealer giving you free hookups until you're completely hooked on their drug. And then presumably I expect them to want profitability at some point and they'll jack up prices. And then you have a PPQ style, you know, pay per token model where the more you use these tools, the more you pay. They can get quite expensive, but you have more choice and user sovereignty. You're able to switch between models on the fly. Ppq. Now at this point I don't even know how many models you offer. I mean, I guess that's a question for you, but it's a shit ton of models. Yeah, yeah. And you can just switch on the fly. How do you think about that dynamic? Because it is way, way more expensive as I'm like learning with these things and I'm having them run, especially if I'm, I mean it's one thing, one cool thing about PPQ is you can offload a lot of things to the cheaper models because you have model choice. But I mean if you're hitting OPUS or God Forbid you're hitting OpenAI, they charge so much on per token basis. How are you thinking about that dynamic? Because obviously it's key to your business and your sustainability. And also personally, how are you using these tools?
- [15:46]
A
Yeah, I think that running subscriptions, it's the very tried and true method that's been employed for a couple of decades now. The one main benefit is you get to charge users money even when they don't use your product and so that creates like an extra padding for you I guess. But what I find is that I do think that the subsidies are happening a lot right now. And in some cases if you're just using Claude code all day, every day, it makes sense to just get a cloud subscription and not use pbq. But I think there's a tremendous amount of cost in maintaining that subscription as the company you see every other day now they have changes in the rate limits, they're sneaky, they're sneakily changing the terms, they're doing all sorts of stuff like that. They probably have entire teams dedicated to squeezing as much as they can out of the non users or the casuals that barely spend anything and then, and then like punishing power users as much as possible. And that's just something that like I find that really boring to have to manage as a business. And I also think it's like not, it's the way that things are already done. So I would rather try to build something that actually is innovating in this area where it's just a very simple billing system. We focus on building features and products and other companies out there are focusing on like rate limiting their subscriptions and setting up abuse detection and all this and that. And I've seen it happen to several kind of competitor companies out there where they will come up with a subscription with a certain amount of like monthly tokens and then once abuse starts happening then they have to change the terms and they make their users angry and then they have to, you know, win their users back and this and that. And I would just rather build the hard product and let it win on its merits. Yeah, I mean there's a lot more to say about the, you know, trying to doing vendor lock in and such but these models are becoming, some of them are great but, and distinguished from others. But I think over the last year they've actually come more alike than they have become different. Like you know, Claude was the clear coding leader over the last year and now Codex is doing very well and now even the open source ones are doing well. So I think it's going to be hard to lock people in and I think the subsidies are not going to last and we're already seeing that. So yeah, just I like running a paper query business and just charging an honest margin. I, I'm competitive on the power models, the ones that everyone uses and then where I'm going to make my money hopefully is on kind of maybe the more exotic features of our website, exotic APIs using. If you. So if you're using our API, it's
- [18:44]
B
very cheap, but if you use our
- [18:46]
A
UI, we have a bit of a higher margin because people generally don't spend a lot of money on UIs. They're spending more, more money in the, in the IDEs and through the API. So that's where we'll make our money. And I also think that there's a future where we'll be able to create really specialized tools, AI tools that can solve very niche problems in a way that a generic model cannot solve those problems. And when you create those types of tools, you could also put a nice margin on top of those as well.
- [19:22]
B
Yeah, that makes a lot of sense to me. I mean, this was always the Fiat Silicon Valley VC model, right? Which is growth at all costs, spend a ton of money, subsidize user growth. I just don't think we've ever seen it at this level. I mean, if, if those numbers are correct and it's, you know, power users are getting $5,000 worth of Claude tokens for their $200 Max subscription, that means that they're spending for growth, they're currently spending $60,000, they're losing $60,000 per power user, which is just, I mean, I know there's a lot of capital sloshing around over in AI land, but it's hard to imagine them lasting very long with that model. And like you said, I mean, we have been seeing people get locked in their whole life is in Claude or Gemini or something, and then the terms change or they get a vague ban hammer or something and then they're going to have a bad time and that probably accelerates and gets worse. On the note of model competitiveness, I will say the single biggest thing I think we've seen over the last year, which has been really cool to see, is the lighter weight open models. The floor has been risen. So there's still, I feel like a significant gap between the latest Claude or the latest OpenAI model and what's available out there in open source land. But even if we saw no improvements to the open source lightweight models, which is ridiculous, of course we're going to see improvements. I think they change everything. I, I mean, I, they're already incredibly competent, so it really just changes. I don't know how society functions.
- [21:01]
A
I think they're, I think they're certainly much smarter than they used to be, but I find that as the tools get more complicated, they really start showing their stupidity. For example, when you Plug these open source models into OpenClaw. I know this story way too well because OpenClaw actually caused PPQ to pop off quite a lot over the last four months. And a lot of our users are spinning up openclaws and then they're throwing our API key in there and they're trying all different models and such. But you cannot run openclaw with, on Sonnet or Opus purely unless you're like a very rich man. So they're using Kimi, they're using minimax, and openclaw is actually pretty dumb with those models and it makes a lot of mistakes. And, and then yeah, unfortunately I've been, I'm, I'm also head of customer service, so I'm the guy that everybody comes to and they're like, hey, this thing is not working. And it's super hard for me to solve because it's, it's, it could be a PPQ problem, could be some recent shipment we did, but it could also be the model itself is, is messing up. It could be a recent openclaw update. That thing is like openclaw itself is like a hot pile of trash in terms of bugs and such and sluggishness. And so yeah, it's, it's like, it's a very, it's like a black box of problems. Yeah. So I'm getting away from my point, but all I'm saying is that the open, the open source models, like when you've really put them to the test and, and they really need a lot of context to make like decisions and such, you can really start to see them become dumb.
- [22:40]
B
Yeah, I mean I've noticed that as well with my experimentation on OpenClaw. And yeah, to your point, my, like beginning of open claw, I was just hitting Opus and it just cost a shit ton of fucking money. And then of course I did the. You know, humans love flip flopping from extreme. So then I like went to Minimax and then it was. And like there's a happy path in the middle. But to your point, I mean I think this is where something like PPQ really excels. Released a. I guess like it's like a blended model. I don't know how, how to explain it but like I, you, you basically, you set it up with openclaw and then like it knows when it's supposed to hit the smart models, the more expensive models, and it knows when it should hit the cheaper, less smart models. How does that work?
- [23:30]
A
Yeah, so it's called autoclaw. We actually borrowed it. Well, like the, the core of the code comes from an open source project that we found, but we, we added our PBQ tweaks to it and such. But basically it has a smart router on it. So it looks at the input, it looks at the question that the user's asking and it does a keyword search, essentially a really extremely fast 1 millisecond keyword search on the input. Certain terms that exist in that, in that input. And then if the terms kind of indicate that this is a complicated question, then it kicks it to the, to the expensive model. And if it's a super easy question, then it kicks it to the cheap model. And it's actually been, at first it was wildly like popular. People were loving it because it was saving them like 80% in costs or more. And that's really releasing that. It was our number one like blog post ever. Everybody loved it. And then OpenClaw did an update in the last month or so and then it started breaking down and all sorts of problems were happening. Yeah. So I think I got it fixed, but it's still very murky and yeah, it's, it takes a lot of my time to, because I have to be an open claw expert now. But you know, that's, that's, that's part of the job. So. But yeah, that's, that's basically how it works. And that is the nice thing about PBQ is I have both the proprietary models and the open source models at my disposal. So I can throw in whatever models I want into that auto router and I can update them on the fly. And that the users don't have to do that. That's also very key. Right. They're not on the cutting edge of knowledge. So I'm kind of trying to make the decisions for them based on my best knowledge. And so yeah, I'm plugging in the latest and greatest models the best that I can to keep their claws working.
- [25:24]
B
Yeah, I mean that's wild I think by the way on the hiring point, like probably customer service guys the next thing you should hire for, so you're not doing it yourself.
- [25:38]
A
I like it though. I really get a feel for, I get a feel for the customers and, and, and yeah, I like my voice when I talk to the customers, but it is driving me slowly insane because it's like it's a 24 hour job. Like you know, I'm not waking up 3am to answer messages or anything. But they do happen all the time and I'm like, you know, focused in Work and then a message pops up and I got a context switch to answer their problem or whatever.
- [26:08]
B
But yeah, there's definitely some merit to the founder managing support, at least in the beginning. I mean we're the largest investor in Strike and they're a significantly larger company than you and they deal with people
- [26:21]
A
that's not answering the phone.
- [26:23]
B
He was, but my point was he was. Until like six months ago, he was responding directly to support requests and it was one of the things that was like, okay, you got your plate full. It's not easy hiring for customer support because you need them to be technically competent and actually care about the business too, which is, it's, you know, it's, it's a difficult mix to find. To the point on the auto claw, do you think? Is this where we're going? I mean to me that feels like that's what the future looks like. It looks like. I don't know if it looks exactly like that, but it looks like people using a variety of models where they make sense rather than just having Claude the top model claw do everything for them.
- [27:12]
A
Yeah, certainly that's absolutely the future. Yeah. It's all about the intelligence of the routing. Right. That's the, that's the key point. I think many users want. They would err on the side of paying more in making sure that they were getting a quality answer than having cat. Catastrophic mistakes happen.
- [27:32]
B
Right.
- [27:32]
A
So you have to tweak the routing that way. Where I personally I almost like we have different modes. We have eco mode, we have regular mode and then we have premium mode. But I would almost like, I kind of want to retool the premium mode and say like the default should be use the expensive model and then you have to find simplistic indicators to downgrade. Right. And that, that would kind of guarantee that you're getting. Maybe not guarantee, but much stronger accuracy and, but the routing is not easy. The way I'm doing it right now, it's just like basic keyword heuristics, which is like not scientific at all. It does a decent job for what it is, but the future, it's going to be much more. I think I can see a whole field of AI research going into this area is it's not only that you have to make these decisions, but you have to make them super fast. You're basically running a query before you run the query. Right.
- [28:28]
B
What are your thoughts on like. So in my quest to save money, but also have competence. Cause I've, I've basically opted out of the Entire subscription model. So I've been paying per token my entire self learned education experience. And it was very expensive at first. I've really grown to love specifically Gemini Flash 3. It's fast, it's cheap, and it's relatively competent. Like it doesn't make disastrous minimax mistakes or even Kimmy level mistakes. And it's cheaper than Kimmy is the Future. Something like using a model like that that's like relatively competent, relatively fast, relatively cheap and that's like pre screening stuff and then making the, it's like the quarterback, you know, then it's like making the executive decision rather than like a basic keyword type of thing.
- [29:18]
A
I don't know. Again, like I said, I think it's a complicated question, so I really don't know. But I'm glad you're enjoying Flash. It is very cheap and it's actually one of the models that we have 30% off, even below what you pay Google because we have a bunch of free credits that we're trying to burn through. So that's why it's probably especially cheap. But yeah, I really don't know. Like I said, it's, it needs to be done quickly, right? So like if you have to run a flash query before your real query happens, you're, you're introducing latency, right? So it's a, it's, it's definitely an art.
- [29:52]
B
But out of, out of all the things, right, like if you have cost, you have competency and you have latency, like maybe those are the three main trade offs when you're like trying to do your setup. I think for end users probably latency is the one they're most willing to take a hit on. You know, there's a bunch of tasks that like at the end of the day I just send the thing running and then I go live my life and you know, take care of kids and whatnot, or do another task. Yeah, I'd be willing to save money but maintain competency. Like I would a hundred percent be willing to sacrifice minutes.
- [30:31]
A
Yeah, you might be right there. And also in a world where you're, where you're setting it and forgetting it and you're coming back 15, 25 hours later, then it does make sense to let the dumb model, so to speak, have a crack at it. Because now we're getting to the point where you can even use the dumb models to test the dumb models in, you know, foolproof ways, right? So if you can have the dumb model build something and then you can write tests for it or it will write tests for it and then if it passes all the tests, then it determines that it did a good job. And if it doesn't pass a test and maybe it upgrades the model, it tries again. Right, so, so yeah, I agree. Maybe latency is the least valuable of all those, you know, for certain tasks,
- [31:15]
B
like if it's self driving, then obviously latency matters because then you're in a car accident if the thing has latency. But I think for a lot of things, people are willing to take the hit for it. The other big feature you launched recently that I was excited about is you added these models that are run in secure enclaves, similar to what Maple has been doing. How do you think about that in your full product suite and how you implemented it?
- [31:45]
A
I think it's. I'm glad that we were able to use a provider that abstracts away a lot of the complexity there, because what PPQ is trying to do many different things at a, you know, a pretty good job on many things. And so we didn't have time to dig into how the secure enclaves work and this and that. And so we found a provider that did a lot of that for us and so we were able to get that out the door. So far, it's actually quite popular. It's hard to set up through API, so when you come to our ui, we do. So the thing is the user needs to encrypt their content before they send it onto us. Right. And so when they're using PPQ in the browser, we can do that encrypting for them. But when they're using it through the API, we had to build out this special proxy that they basically clone themselves and deploy. So now every time they run an API request through us, it's running through that proxy first to encrypt everything. So it's a little bit extra setup on the API side. But yeah, I, I'm, I am very excited about these models. I think that openclaw especially shows us that AI is like going to really hit mainstream when it can start solving personal problems. Right. But personal problems are where you want the most privacy. And so these models, there are going to be many use cases for them beyond what people are currently using them for, and that's going to be very powerful. I think you'll be able to run this stuff on a lot of your personal app data and you will, you will be able to do so with relative confidence that only you will have access to it. So, yeah, I'm very bullish on these models. And yeah, we've been, we've been pleasantly surprised by how quickly people adopted them.
- [33:42]
B
I, yeah, I mean, look, I think that makes a lot of sense to me. I don't know how, like I'm kind of interested, like how many people are even aware of the concern at this point. Like, I think a lot of people are just like uploading their, their lives specifically to the proprietary models, usually under a subscription account, that every search and prompt and data is all tied to the same user. But I do think the dream, and I disagree actually with the Maple guys on this, the dream is to have it all in one place. You know, I think there's different schools of thought here, but I do like how you did it with ppq. I mean, I, I noticed Venice, which is probably one of your bigger competitors, has done similar. But I like the idea that you have the ability to use the proprietary models that are significantly more competent and then you in the same interface be able to use the, you know, open source models that are running in these tees, the secure enclaves, for more private stuff. And I think it goes back to the, to the earlier conversation with cost and competency and latency. Like my dream is that I have some kind of quarterback that's just automating in the background and it's probably a model, I think, which things are going private, which things are going proprietary and more expensive and less private, maybe even local. Right. You know, like mixing local in there, maybe the quarterback's local, but it should all be kind of in one interface, one platform that is kind of doing it all in the background for me.
- [35:24]
A
Yeah, I, I've never thought about it that clearly, but I think you're absolutely right. That makes a lot of sense to me. All these models have different capabilities and different trade offs and you should, most people will benefit from having access to all of them. And yes, the secret sauce will definitely be when there's a little quarterback making those decisions for you so you don't have to have all this competency and about what, what each model is doing.
- [35:53]
B
You will always need to have some of that.
- [35:55]
A
Sorry, yeah, you'll always need to have the users, as you say, many of them are not even aware of the costs that they are bearing when they give up this data. But I think, you know, people will become more aware of that and they already are. And so yeah, it's going to be this mental framework that everyone evolves to in the coming years. What, what data of mine is local only what data of mine is. Am I sending to the cloud, et cetera.
- [36:21]
B
Yeah, I mean like this is a bigger, this goes bigger than AI to me. You know, it's, it's, it's my greater experience in the freedom tech movement and building projects and doing user education and dealing with users, end users directly. Like our biggest, whether that's Bitcoin or Nostr or Tor, you know, message encryption, all these things historically have required immense personal responsibility. And then the result is personal responsibility is always a more difficult decision because you have to take responsibility, you don't have someone else doing it for you. So what happens is we have a ton of friction throughout these tools and then because there's a friction will never hit the level of scale that the controlled and surveilled alternatives have because they're just easy. And I always say, you know, I love Ben, but if you have to watch a 2 hour BTC sessions video to use a tool like we're already fucked, like you're already screwed. And to me, one of the coolest aspects of AI is that we could reduce a lot of the friction across the stack by using open source models, by using models that are in secure enclaves to kind of, you know, remove a lot of that decision making from the user. But of course there's still going to be a level of personal responsibility. I just think we can reduce that significantly.
- [37:44]
A
Yeah, for sure. Yeah, I agree.
- [37:50]
B
Yeah. I mean, I, I don't know how, how are you thinking about, I mean, so we briefly mentioned openclaw. I mean I think that was probably one of the big highlights of the year since we last recorded. I mean you didn't, you, you clearly have about it. But I think as on a high level, just the idea of a open source configurable agent that lives on your computer and has, I mean probably one of the fastest growing open source projects that we've ever seen is pretty cool to witness and changes a lot of the conversation and narrative around the space. Do you think it's a flash in the pan? Do you think it's actually a taste of what's to come? How do you think about it?
- [38:33]
A
Yeah, no, I don't think it's a flash in the pan at all. My gripes with it are just how buggy it is and how much customer service problems it causes me. It's a good problem because it's, it's. People are warming to it, they're trying, they want to, they want an AI personal assistant. That's what everybody wants. And so no, these things are only going to get better. They're going to get less buggy, they're going to become more capable. And yeah, I mean, we're normal people are going to be doing incredible things with their AI assistants and it's going to be removing a tremendous amount of friction in everybody's lives when you can have something schedule your flights for you and actually do it correctly. And like, these are all the dreams, but it's not happened yet. But, you know, book your Airbnb for you and do it correctly. Give you a. My big want is I want a morning podcast that's generated for me with all the things that I care about. Five minutes every morning. That's kind of my little personal project. But there's so many of these ideas out there. Right. And it's only going to make us, like, able to live our lives doing more the things that we want to do. And the AI assistants are going to, you know, handle a lot of the, for us. So. No, I, I think that it's going
- [39:51]
B
to be a huge, huge movement.
- [39:53]
A
Whether it's open claw specifically, maybe it will be, but there's already, you know, a dozen other personal assistant agents out there, so we'll see which ones win out. But, you know, it doesn't really matter. We know that it's all going to get much better. Even in six months. I can't imagine how much better they're going to be.
- [40:11]
B
Yeah, no, I mean, I, that's why I like this cadence. Oh. You know, hopefully we'll do it again in a year. So much happens. The, it's funny you mentioned the personalized podcast to update you, because one of the first things I did was like, I built a AI news desk that is giving me updates. And then I made it public. It's@citadel wire.com, but it's a tank space with opencloth. Yeah, I have, I have multiple, I have a top, top news desk agent, and then he has sub agents underneath him and he thinks he's like a Walter Cronkite type of individual, and he's like pulling in a bunch of different data. And then every three hours, they're sending out updates on Nostr. So they're signed, hashed, hashed and signed Nostr posts, just very brief text. And what's cool is like, every three hours they're making this, you know, this news wire that they're sending out and they're keeping all the old ones. So they're learning and getting more advanced over time. You know, at first, like, I, I, I'm pretty sure they think I Say he, you know, I'm pretty sure these robot, soulless robot slaves, like think this is the first war ever that's ever happened because they were born into the Iran conflict.
- [41:27]
A
So they're like, they're like, this gold
- [41:29]
B
is crashing in the war era. There's a little bit of work that needs to be done there, but it's quite cool. And to me it's interesting, right, because a lot of people have approached it, oh, monitoring the situation. We need to monitor the situation with AI and they release like world monitor and there's a bunch of forks and all this other stuff. And what it does is it's, it's like gives you way too much information. And to me, the cool, liberating aspect of this tech is actually the opposite of that. Like, I don't want to spend two hours looking at, you know, ship patterns and all this other shit. Like, I want my agents to just tell me, yo, you need to know this. You don't have to watch any ads or go through any of this stuff and then you can go live your life. Uh, so I kind of looked at it from the exact opposite of that. Let them deal with all the information and then let them give me a distilled kind of brief on it. Um, but on the. So you could, on the, you could use.
- [42:24]
A
The next step could be for your newswire thing. You could use our TTS model. We have both 11 Labs and Deepgram where you could convert that to the thing that's missing. There's still a missing link in this whole awesome podcast generation thing is you can create a file of a podcast, but it does not easily plug into Spotify or Apple Music or whatever. So there needs. Somebody needs to create. I already have a name for it, RSSify. So it basically takes a file and then it puts it into some RSS feed that's kind of private, where only the person with the link can access it. But then you could literally just have it auto uploading into your Spotify and all of the friction would be gone. That's what's prevented me from like really, really into the podcasts. Yeah, go ahead.
- [43:13]
B
I mean, I think when you start playing with this things, you start to realize open protocols are way easier to interact with. And what's nice is RSS is an open protocol nostrs. It's also. These things work really well with nostr. Obviously not if you want privacy, but if you just have them upload the MP3 to Blossom, which is the nostr media spec for videos, pictures and audio files, MP3s. It's really easy to turn a Noster feed then into an RSS feed. Like, Gigi has a simple tool that he vibed that just turns any Noster feed into an RSS feed. What's nice is then you don't have to store it or distribute it. You could share with other people. Like, I think it's kind of cool that my little news desk, like, my parents use it. You know, I can just like, send them a link. Obviously, anyone in an Oster client can use it, so. Just something to think about. But, yeah, I understand why you would want it in your regular podcast app, but you could even like, I don't know if Spotify. Spotify is like the evil empire in terms of podcasting. If they let you just put in a bare RSS feed, but like, Apple Podcasts let you. I think you can. Android lets you.
- [44:22]
A
Yeah, I think. I think if it's on an RSS feed, you can get it into Spotify. Yeah, I. Is the Blossom thing you're mentioning, is that something that is. Is that an app I can have on my phone where it automatically just loads into my phone? Cause that's the thing for me.
- [44:37]
B
Blossom is a protocol for hashing and replicating media files so you don't have to host them.
- [44:45]
A
Okay, so where's the app? Where's the app that I can just press?
- [44:49]
B
Well, that's the cool piece is that. I mean, obviously you could listen to them in a Noster app, but Gigi has a tool, I. I forget what it's called that converts any. So, like this episode, right? So Sil Dispatch, you can search for it in any podcast app, right? It's just hosted like a normal RSS feed. But I also upload the signed and hashed audio files to Noster using Blossom, right? So I have a Noster account, primal.net/citadel. It has its relative, you know, npub or whatever. Gigi's tool can convert that Noster feed into a completely different RSS feed that is not reliant on a single RSS host. Like, I use Pod Home as my RSS host. I use a bitcoiner. It's a smaller RSS host. But if podhome goes down, my normal Citadel Dispatch feed, you lose. If. If you are listening through the RSS feed that is made from the nostr feed, those Blossom files are actually hosted in like six or seven different Blossom servers. And because they're hash and signed, you can replicate them to infinity and not worry about verifiability.
- [45:59]
A
Right?
- [46:00]
B
Because they're hashed.
- [46:01]
A
Yeah.
- [46:02]
B
So it's kind of fascinating to me. So. So it gives, it gives dispatch, for instance, some resilience and robustness from having a single RSS host.
- [46:12]
A
Yeah, yeah, makes sense.
- [46:14]
B
But anyway, we kind of. I wanted to talk about OpenClaw a little bit more. I mean, so the reason. So I don't think it's a flash on the plan. I agree with you and I think it was really cool to see and to me was like kind of an aha moment on. Oh, like maybe the open source side of this thing can really win out. But why I say a flash in the pan is we're already seeing, you know, Anthropic has just been. Well, first they sent him a cease and desist, then he went to OpenAI. He got like kind of aqua hired. Kind of what? I don't even know. It's an open source project, so it's a little bit weird, but I get why he did it. And then Claude, they've just been releasing left and right, basically their proprietary alternatives to OpenClaw. Obviously, you know, we were talking about personal assistance. This was what the selling point was for Siri all those years ago, 10 years ago. Like Siri's still retarded. But the idea was Siri was supposed to book reservations for you and stuff. Google's obviously working on Gemini stuff already has people all in G suite. So when I say Flash in the pan, I don't mean like disappearing completely, but how do you think about that dichotomy? Like I feel like the overwhelming majority of users are going to use whatever Apple's, you know, centralized proprietary alternative is. Whatever Claude. Centralized proprietary alternative is, whatever Google's alternative is, is versus, you know, an open source, whether it's OpenClaw or something else out there.
- [47:41]
A
Yeah, I don't know. I will say that my one angle of knowledge on this is that if our open claws or whatever you're going to call them, if they require many modalities. Right. If, you know right now they're just chat heavy. But I just told you like you can, you can use tts, you can use stt, you can use video generation, image generation, there's all sorts of modalities. And what I've discovered running PBQ is that no one company is best in class and even, even a few of those things, Anthropic has the best coding models, but I don't even. They don't even have any major media models. Right. OpenAI kind of has maybe the best all around chat Model, but its media is not that great anymore. In fact, they just scrapped Sora 2, I think a day or two ago.
- [48:34]
B
They were losing a shit ton of money on it.
- [48:36]
A
Yeah. So my point here is that if the best personal assistant, if personal assistants are very multimodal in the future, then they will, then the best personal assistant will be the one that can connect to the technologies of many companies instead of one company. And so that's an argument to say that maybe open source or, you know, build your own assistant kind of wins. Right, where you can just choose the best from each category. Yeah, that's my only one area of knowledge in that, in that argument, I think.
- [49:09]
B
So I think that's a pretty. I think that's a solid argument. I think the question then becomes, can we get the cost down? And I think the answer is probably yes, because obviously there's a big discrepancy between people using the proprietary stuff on subscriptions. But that makes sense to me. I, I think there's a unique advantage there that open source has that the proprietary stuff doesn't. I also think there's not as much vendor lock in as people realize.
- [49:35]
A
Yeah, I agree with that too.
- [49:37]
B
Yeah, you can move around more.
- [49:41]
A
You can. Yeah. Like I said, you know, Claude had the best coding model, but now there are several models out there that are pretty close. And we also learned that it's not just the model, it's also the tooling and stuff. And it's amazing how fast some new proprietary thing comes out and then there's a very strong open source competitor within a few days. And so, yeah, it's, it's, it's very interesting to watch. Just today Anita Posh was saying how, you know she is. I'm guessing you read this also? It's one of the bigger, bigger notes today on Noster. But yeah, Anthropic was blocking her credit card because she's in Africa. And so now she can't use Cloud Cowork. And she's like, ppq doesn't work for Cloud Cowork, which I looked into and it seems that is true. You cannot bring in an outside API API key into cloud cowork. But then I just did a quick search and there's like a repo out there called openwork or Open Cowork and Open Code.
- [50:44]
B
This is different.
- [50:45]
A
Open coworkers. Yeah. And so that is like, like, I would say probably 95% as good as Cloud Cowork. And it's just a matter of like, can, can she spin it up quickly and and that's, yeah, that's, that's the challenge is like can, can we get it. These open source tools, they're always a little harder to get up and running. So can we lower that barrier of entry enough?
- [51:10]
B
Yeah, that makes sense to me. I saw, I saw you popping off about, I don't know, would you what to call it, Bitcoin maximalism or something about the fact that you accept many different payments. I think we talked about it last time too.
- [51:36]
A
I have a lot of strong thoughts on that.
- [51:38]
B
Yeah. You want to discuss that briefly? I mean you're accepting basically every payment under the sun right now, right?
- [51:45]
A
Yes, yeah, we accept everything. Our focus is bitcoin. We like to spend our development time building out new bitcoin tech as much as possible. You're not going to see us like building out Solana this or Ethereum that. We will slap on Solana and Ethereum payment methods because it's easy. But yeah, my, my, my argument, my feelings are, yeah, very strong in this area because I think that bitcoiners or bitcoin only types of people, it's harmful to bitcoin and I can give very concrete examples of this as I've learned over the years in my work. So I, I started off. Well, I didn't start off at Bitrefill but I worked at Bitrefill for several years. And most people know that they are kind of a bitcoin first company, but they also accept pretty much everything. I, I was there long enough and I was analyzing the data to see that this share of bitcoin payments was gradually going down over years and the share of altcoins and stablecoins and whatever was going up. And so if Bitrefill had made the choice to be Bitcoin only, I don't think bit Refill would be around anymore. I think they would have been subsumed by competitors that accept all of these other coins. So do you want, you know, kind of a bitcoin first company that makes rational decisions about this stuff or do you want kind of just like very purist type of companies out there who kind of struggle? And I faced this several times in marketing pbq. So I wanted to. So when PBQ first came out, it was actually lightning, only that was not always my plan. I always knew that the other payment methods were gonna come because I worked at Bitrefill and I understood the reality. But I started with lightning because I, I thought lightning is the coolest use case for ppq. But I tried to post it on the bitcoin Reddit at first. And it was just completely ignored for the first like six months. And then by that time I had started adding in other payment methods. And so I went back and so when we got our round of funding, I think is when I asked again. And then the guy finally answered me. I don't know who's in charge of it, but he answered me, said, well no, you're, you're peddling shitcoins so I'm not gonna post your. Your thing. And so then I would go over to the cryptocurrency Reddit R R Cryptocurrency and they don't like me because I'm bitcoin first. So they're not posting me either. And, and, and this story has played out several times. I've tried to like, I tried to make a posting on bitcoiner jobs because I want to hire bitcoiners, right? But they won't post my jobs on bitcoiner jobs because, because I, I peddle shitcoins. Even though some of the companies like Strike for example, they've had, you know, they, they have coins on Ethereum in their app for years, I think.
- [54:39]
B
Yeah, we support Tether on Tron actually globally. And that's because that's where the users are. The users use Tether on Tron as a dollar rail and in America people use bank of America and in Africa they use Tether on Tron. So we support it over there.
- [54:55]
A
Even when Strike went into Argentina, they were originally like USDT on Ethereum as well. But anyways, it doesn't matter. But yeah, this has played out several times for me and I do think that that attitude, people who actually care about Bitcoin or businesses and I, I don't believe that PBQ would survive being Bitcoin only especially because PBQ is a scale business. The more volume we bring in, the more volume discounts we can get from our providers and stuff like that. Right? So we have to go for the whole pie, right? And yeah, I'm not articulating it perfectly, but yeah, I do think that it's, I think that a lot of other bitcoin companies out there have been kind of scared into not adopting these other payment methods or doing practical things because they're afraid of being called out for this or that. And I'm just trying to do my best to get us past that where it's like, I don't, you know, we shouldn't be like beholden to these people who are Strictly bitcoin only. And I think maybe for a handful of businesses you could do that and it could work. But for many other businesses you just need to go after as many customers as possible and you need to solve real problems in their lives. And yes, altcoins and stablecoins, stablecoins especially do solve real problems for people all around the world. I look at bitcoin more as a century long battle and I do think we will end up at bitcoin only at some point in our history or at some point in the future. You and I may be dead by then, but I do think that's, it's, it's, you know, it's a very long journey that we're going on and I don't think that the right thing right now is to like self segregate. Yeah, yeah.
- [56:56]
B
I mean, I think I align with your perspective more than I don't. And I think, look, I think it's a natural response to so many scams in the ecosystem, in the greater crypto ecosystem. And it's a relatively easy litmus test, like if people accept crypto versus bitcoin only. But a perfect example in your own industry, I would say is Venice who I mentioned earlier. Like they have a whole token scheme built into their app, right? They have a pre mined Venice token that offers yield. It's operating on like a BNB style system that is a core aspect of their entire app versus someone like you who is merely accepting it as payments. There should be a delineation there. Like it's a completely different delineation. And when I, if I remember from our last conversation, you know, you're a bitcoin first company, you hold your profits in bitcoin, you offer discounts in bitcoin. Like it should be very obvious to anyone who actually looks more than just surface deep that, that of the differences between those two types of companies.
- [58:03]
A
Yeah, I hope so. But like, you know, I'm not a. If Venice is helping people, and I do think they're helping some people, they're providing privacy for some folks, then that's great. But I am, I am, I am a Venice hater, I would say for the reason you brought up, I, they do have this like overly complicated token thing that basically is. The entire design of it is to just bleed money from ignorant investors. Right? That's the purpose of it. I've seen Eric himself like describe the token mechanics and this and that. And even in his own tweet he's like, I don't know what the price of this token asset should be. And then at the end of his tweet, it's like, oh, I guess it should just be the cost of AI and then the capital. What is that called? The capital when you hold it for a period of time? The cost of capital, right?
- [58:52]
B
Yeah.
- [58:53]
A
And he's like, I guess that's the true price of it. And I just wanted to reply like, good job. You figured it out. That's exactly what it is. There's really no need for a token. But, but yeah. And then other things like Venice has been, they overly promised privacy and they have for a long time. And he glosses over the privacy stuff. PBQ has essentially the same model that Venice has had, but we're just more honest about it. We tell our customers like, hey, you need to trust that we're not looking at your queries. You need to know that when you're using a proprietary model, your content is going to those guys. So just use that as you may, but Venice is just very loosey goosey with those things. And the thing that really kills me is like, they, it works for them. He's, he, he's leveraging his million followers or whatever, and he, his kind of reputation in the space has formerly like cypherpunkish, you know, swap guy. And I don't think he actually like, cares about privacy all that much for his users. They recently adopted tee models just, just like we did. And that's great. Like I said, I, I hope I'm not coming on too, too bitter. But it's like he only did it after people were really, really calling him out for a lot of the shenanigans that he was saying on Twitter. Right. So, yeah, I mean, it's a, it's, it grinds my gears that I'm the little guy over here trying to, trying to fight the good fight. And then he's just like, you know, tweeting all these very brash things about, about his product. But yeah, it is what it is. And that's, that's the, that's the bit you have to use the tools that you have. He's using his reputation and I'm using whatever tools I have. You know, I can criticize him and, and I have right on my side. I'm not, I don't have a large voice, but I can kind of pick, pick apart things. So it is what it is.
- [60:48]
B
I mean, you can say a lot of things about Voorhees, and I tend to agree with pretty much everything you said, but he's really good at making money. He, I mean he had Satoshi Dice, he sold it before transaction fees, made it an untenable business model, all for bitcoin. Then he launched Shapeshift right on the verge of, of shitcoins, you know, hitting the populace. Made bank off of that. And then he seemed to, you know, he's always in the right place at the right time and executes and he takes a lot of shortcuts, you know, but it's kind of crazy to watch. But I think that's just a good example. I mean it's a perfect example for that conversation. Right, which is if you can, if you compare and contrast PPQ versus Venice, it's obvious which one is trying to be user focused and ethical and try and do things the right way and be bitcoin first versus the one that is clearly a shitcoin operation. And to be clear here, Venice also because he's a good businessman, he self hosts BTC pay, he accepts bitcoin along with every other payment method. He knows exactly what he's doing. Yeah Matt, at one point there I knew you as the, you were the data guy, you know, you had useful tulips, you're the data guy. Do you have any data for us that's interesting on the PPQ side, I mean the two big things that I'm curious about is like user behavior on model selection and user behavior on the payments. How are the bitcoin payments your most common payment?
- [62:24]
A
Yeah, they are. Yeah. I'll, I'll, I'm trying. I would love to share as much as I can, but I want to hold a little bit back but I will say feel free. Yeah, yeah. Bitcoin Lightning specifically is our over half, just slightly over half. Monero's number two.
- [62:42]
B
Interesting.
- [62:42]
A
And number three and four are kind of a balance between tether and credit cards. And then the rest are just various other stables and litecoin and stuff like that.
- [62:55]
B
What's ahead? Credit cards are or tether.
- [63:00]
A
They, I honestly don't know. They're, they're pretty neck and neck. I think it depends on the month really. I think tether is up and coming now, more, more than it was. But I actually look at this kind of when you look at the bit refill stats and bit refill in my opinion is the best cross section of all crypto payments. They're the largest online crypto e commerce website. Right. So lightning is about 5% of their volume, maybe, maybe higher than that, but not, not much higher. And so if, if I want, I want to get into the pie that they have. Right. So I need to increase my awareness outside of just the Lightning and Noster circle.
- [63:37]
B
Right.
- [63:38]
A
So that's what I'm trying to do. I do want to bring in those other users and yeah, I, I, I think that over time they will find that like there's a lot of really cool things that like Lightning can do for them. It's our lowest minimum deposit because the tech allows for that. You can deposit as little as 10 cents in lightning and I think the other payments are all a dollar minimum because of the fees, among many other things. It's very private. It's fits really well with a lot of these agentic payments coming up. Yeah. So yeah, that, that's that. In terms of models, again, I think, I don't know, I don't know if I have really that great insight on model selection and whatnot. I think that the proprietary models are still the most popular aside from being more competent. They also have the brand recognition that people just, they don't know what, they never heard the word Kimmy before so they're just going to click on the, the word they have heard of called Claude. Right. Stuff like that. But yeah, I, I think it really varies by business. I don't think PPQ is like an even. I know that some of my competitors, they're, they have way more popular open source usage on their platforms and the proprietary models are not like the absolute highest, but it is, that's the way it is for, for PPQ currently.
- [64:58]
B
Interesting. Yeah.
- [64:59]
A
Trying to think of other interesting things. I mean there's a bunch of stuff I just kind of caught me off guard.
- [65:04]
B
So like, would you agree that like the main user behavior right now in PPQ is like people trying to use Claude or whatever with some additional privacy?
- [65:15]
A
I think privacy is a really big use case. I think it's also, it's a mixture. I think people do try different models often. I should come up with that stat. Like what is the average, how many models does the average user try?
- [65:27]
B
I bet it's, yeah, I'd be really interested in that.
- [65:30]
A
I think they try the models but I do think they kind of settle into one or the other after a while. That's what I do anyways. That's my go to model for this and that. But yeah, yeah, it's like I said, OpenClaw. I put this tweet out. Since OpenClaw came around, our revenue went up 400% in just a few months.
- [65:49]
B
Wow.
- [65:50]
A
Yeah, it was a huge, huge boon to Us. I also, I think I feel a little bit of a plateau coming because as I stated earlier, people really wanted to get their hands on openclaw but then they start finding out that it's like really hard to get working in the way that you want it to. So they're kind of backing off now. But I do think that it's, it's going to be very short lived and like the better openclaw or openclaw itself is going to improve and like we're going to all be really diving into these AI personal assistants. Yeah.
- [66:22]
B
The other thing that kind of throws that stat for loop is that the Open claw is not very efficient. It's very easy to consume. It's. The openclaw is a great consumer. And so like when I first played around with it, like I said, I just spent a ton of money in the beginning and it was a very expensive lesson. And then I fine tuned it to a degree and got my spending under control. Like when you're using like the. Just a normal chat interface, maybe hitting, just hitting one model directly you end up use. It's way harder to spend as much money as you do with something like an agent just running wild in an inefficient manner. It consumes so hard.
- [67:04]
A
I, yeah, my, my open claw was very liberal the other day I was telling it to. So we just got our L402 endpoints up and what that means is like you can now basically use a curl request to pay for one image at a time without even having to get an API key from us.
- [67:21]
B
But yep, the agent.
- [67:23]
A
Yeah, it, but the. What, what was funny is I told it to use one model that cost 29 cents. It was like a 29 cent video and it couldn't get that one to work so it went with like a $3 model. I was like, I did not tell you to do that. Um, yeah, it just cost me $3. So yeah, those are fun things that you learn along the way.
- [67:44]
B
So yeah, I mean I had, I have a dashboard that among other things I used to track all those stats that I read up at the beginning of the, the beginning of the show. And it updates every five minutes and uses a cheaper model to do that. But at one point it just, I don't know, it just fucking crapped out and switched to Opus and it was just updating my situation monitor dashboard every five minutes on Opus and by the time I realized it had spent like $250. So there's a lot of those cases I think are happening out there. But it's interesting that. Yeah, go on.
- [68:16]
A
Said we introduced spending limits on our API keys, so. So you.
- [68:20]
B
Very helpful.
- [68:22]
A
Yeah, I.
- [68:25]
B
And that, by the way, that was. I was on a flight too. I got off the flight, I was like, what the fuck? It's interesting that you brought up L402 because that was my last question for you before we wrap is, you know, you come from the payment space. You're obviously also now neck deep in the AI space. The, you know, the hot button topic is agentic payments. How do you think of that? How is that going to play out in your opinion? What's your perspective there?
- [68:54]
A
Oh, man, I don't know if I have that great of an insight. It is, it's a lot to wrap your head around. I, I implemented the L402 personally. And so, like, I'm figuring it out and at first sight, it's not a paradigm shift better than just going to PPQ and getting an API key. Because usually, like, it is a paradigm shift better than going to some proprietary website, creating a login account, signing up for a subscription, whatever, is way better than that. But PBQ never had that. Like, you can literally generate an API key just by sending a request and then you could load money on that API key with another request. So doing it on a paid per request model, it's not that much better. Like, but my partner was kind of mentioning today, as I was kind of relaying these thoughts, he's saying, like, the power of it is going to be when. So when I use an API key on pbq, I'm generally like, almost all the time, I know what models. I'm the one choosing models. I'm the one in control. Right. But with the L402, like we're going to create, and maybe you've seen this, but Ryan Gentry released this website called 402index, and it's like thousands of different endpoints that have different models, different modalities. They do different tasks for you. Right? And so when the agent, it's when, when the human is no longer choosing the models or choosing this or that and it's the agent doing that, then I think maybe these 402 payment methods are going to be coming way more handy. Basically, they're just going to search a website like 402index. They're. They're going to be given a task first. Like the human gives it a task and then they're going to go and they're going to look at these thousands of endpoints and they're going to solve that task by just, you know, consuming whatever endpoints they want to. So also another thing that the 402 payments may solve is like let's say the future is where these 10, 15,000 endpoints exist and it's literally with hundreds of different providers which right now that's what 402 index shows. It's like there's almost 300 providers. You don't want to have a top up model when you're used, when your agent is using all these different models because you don't want to store balance on dozens of different providers.
- [71:13]
B
Right.
- [71:13]
A
You're just leaving money all over the Internet. It's more ideal if you just pay exactly what you need for that thing at that time and then go about your business. Right. So that's another way that these 402 payments could really come in handy. We will see how it all develops. We don't know if the future is going to be like maybe all AI really is just going to come from a handful of companies and you won't need to like. And the downside of leaving money all over the Internet is not as high. It's like you said, you just go to OpenAI. That's where all your money is and it gives you access to all the different tools. Right. But the more beautiful future is where all of these tools are served by hundreds, thousands, tens of thousands of different providers. It's much more democratic that way. Right. And, and your, your agent is just consuming as necessary.
- [72:03]
B
That makes sense to me. Um, yeah, I mean the agent should just be able to go out there and just natively pay for the data it needs or the access it needs? I think, I mean the better question given your background is is this something that USD tokens are going to dominate? Is there's I think stripe released credit cards like agent credit cards or is it Bitcoin? Obviously it'll be a mix of the few but how do you think about that competition brewing? I mean even Gentry site I think lists ones that are using USD tokens. I know OpenRouter for instance only accepts USDC on base.
- [72:41]
A
Yeah, again I don't have the yet. I don't have the best opinion on this but I don't think it's the like the tokens probably it will be some USD token of some kind but I think the protocols, that's where it's interesting to me. So I've learned a bit about X4.02. That's, that's what the USDC on base is That's Coinbase's play for these agentic payments. I looked into that quite a bit and I think there's a lot of things I don't like about it. Um, it's not private. It's, it's, it literally is all on a blockchain. So you can literally like look up and, and use heuristics and this and that to like de. Anonymize what's going on. Also, the fees are high on X402. I was doing kind of some napkin math the other day and for every query you send, you pay 2/10 of 1 penny as a gas fee. Right now on base that doesn't sound like a lot, but it actually is a lot. If you're, if your query is $0.01 and you're paying 210 of a cent in gas, that's 20% fee. And a lot of these AI queries really are just one penny. So I don't like X402 and then MPP Machine Payments protocol. That's the one that Stripe released last week. That one is actually interesting to me because it is a blockchain, but they also support the creation of payment channels. And so unlike Bitcoin, Bitcoin has 10 minute block confirmation time and I think to even spin up a Lightning channel, you need I think three or six confirmations for a Lightning channel to be fully open. So you're talking about waiting 30 minutes to 60 minutes plus to. To. Anyways, you can do a transaction on this MPP thing in under one second and then it sets up a channel. You basically have a channel with the provider and you can send payments over.
- [74:30]
B
But it's like their own version of Lightning.
- [74:32]
A
It's their own version of Lightning. But yeah, it doesn't have network effects. I'm sure that it's certainly much more proprietary. Honestly, I haven't dug in more, but I think, yeah, I still think that the openness of the Lightning network can, can lead to something much greater. And so that's where I would put my money currently. But I do think that the MPP does some cool stuff like mud. I don't think the X 402 is going to do too well unless they also start adopting payment channels.
- [75:06]
B
Fair enough. It should be interesting to see how it all plays out. Matt, this conversation has been great. Before we wrap, what's next for PPQ and final thoughts for the audience.
- [75:18]
A
Yeah, yeah. Thank you for having me on again. I agree, it was a great discussion. There's a whole bunch of stuff coming for PPQ next. We're just going to keep grinding on products that people use. So I think that we can. We're only getting faster with the way we ship and yeah, many, many cool new things. In the future I would love for us to start creating very specialized tools as well that are doing really unique things and that will also help us kind of establish a bit more of a moat other than just kind of being an aggregator of sorts. Yeah, that's. That's pretty much all I have. I do hope. Yeah, we can, we can chat again at some point in the future and just keep wrapping on. On a. Yeah, let's do this.
- [76:07]
B
Yearly cadence. I like it.
- [76:09]
A
Yeah.
- [76:09]
B
And I agree on the tools thing. I mean, I think that's where your model really starts to shine. When I have like this multifunction brain that's just an API key that I can plug into a bunch of these end user tools and have a lot of flexibility. It gives me superpowers compared to people that are using. It's. That's the advantage over people that are using like a cloud subscription for instance.
- [76:30]
A
Yeah. Yeah.
- [76:32]
B
Awesome, Matt. This is great. I'll put all relevant links in the show notes Freaks, but go to PPQ AI, check it out, play around with it. Give Matt his feed. Give Matt feedback except for his payment methods. He appreciates it. Thank you for supporting the show, Freaks. All relevant links are still dispatch.com we're going to go back to bitcoin next week. I already have a bunch of shows lined up over the next few weeks, but we're going to be oscillating basically between Bitcoin, Noster and AI if you haven't noticed. Love it. Thanks, Matt.
- [77:02]
A
Yeah, thank you. Appreciate it.
- [77:05]
B
Thank you, Freaks. Stay humble. Stack sats. Peace.