[00:01:39] Justin Moon: Yeah, thanks for having me. I remember listening to this in the very early days when we’re just dipping your toes into Bitcoin. And I feel like we might be back in the same one, but AI.

[00:01:48] Preston Pysh: It’s interesting that you say that because for me, when I started doing the show and I was covering Bitcoin every single week. I had a lot of people in the value investing space that were like, what in the world are you doing? Why are you covering this exclusively? And now I feel a little bit like it’s the same thing ‘because I’m broadening the aperture into tech and things. So it’s funny you say that Justin, ‘because it does feel like that.

[00:02:11] Preston Pysh: Okay, so here’s where I want to start. Justin, you were on stage with Jack Dorsey in Oslo and you guys go on stage and you’re vibe coding, putting on a demo for everybody. And I was looking around on, because I had heard about this from different people in the space about what you were doing with AI. I’m trying to find the video of it, and I can’t find a video of this anywhere. So can you tell us the story of how it precipitated, how did you find yourself on stage with Jack Dorsey vibe coding in front of an audience?

[00:02:43] Justin Moon: Yeah, so this year at the Oslo Freedom Forum, which is hosted by HRF. There’s obviously many, there’s a main stage, big main stage with huge hop at the concert house Oslo. That’s where the main event is, but there’s a lot of little side events and yeah, we were at a little side event that was more focused on Bitcoin and a lot of our Bitcoin friends were there and HRF had just announced their AI for individual rights program, which I’m technical advisor.

[00:03:08] Justin Moon: We wanted to discuss the topic with Jack because Jack’s a huge proponent of AI as well. So I just had some marching orders. Hey, let’s, interview Jack, right? And maybe, so I had a list of stuff. it was actually about five minutes, 10 minutes before like Jack comes up and he is like, Hey, what if we just vibe code something? I got my laptop, right?

[00:03:27] Justin Moon: So I’m like, oh God. as the interviewer, you want it to be somewhat in control. Yeah, and vibe coding is, no one’s going to control it. It’s pure, it’s non-deterministic. You might get it, you may get some good, you may not. And so we’re in a room with some movers and shakers.

[00:03:40] Justin Moon: Then, if Jack pulls his laptop out, puts a hood on some sunglasses, and we start making a website, I think it’s called the Bitcoin Institute, we find ways to improve the regulatory environment for Bitcoin. companies for individuals. And yeah, we fired off a prompt and then I interviewed him for 10 minutes and then right about at 10 minutes we’re talking.

[00:03:57] Justin Moon: All of a sudden the audience goes, woo, that’s because this website just popped up. Gus. This is the coding agent that Block Jack’s company has created Open source coding agent that we were using, I think, philanthropic models. It just spat out a beautiful working website in just 10 minutes during an interview.

[00:04:13] Justin Moon: So it was really neat. There’s, it was a private events, a little side events, so I don’t think it was recorded. Unfortunately. My, my mother had the same thing. She was like, oh, I wish I saw this video. But, yeah, I think that one just is, apal at this point. But yeah.

[00:04:26] Preston Pysh: What was the core message? I think I know what the core message was, but from your point of view, what was the core message?

[00:04:31] Justin Moon: Yeah, so of course now this has been like four or five months, so now I probably don’t remember exactly what happened. I remember how I interpreted, but part of it was talking to Jack. It’s like he was actually talking to me.

[00:04:40] Justin Moon: We were talking about this before, and he is I was telling him how I’ve been using these tools, I was still looking at the code and he’s Justin, you’re in jail. Oh, you’re in jail. You have to free yourself from the jail of programming languages. You need to, get the computer to work for you.

[00:04:52] Justin Moon: Now, this is a new era. That’s what it is. It’s like for developers, you can free yourself to some degree from having to be so microscopic and worry about the semicolon. But for people who weren’t progprogrammers, it’s now you don’t have to worry about that in the first place you told yourself in the past.

[00:05:05] Justin Moon: And in order to create software, you need to learn where all the little semicolons go, which is a horrible exercise for most normal people. Yeah. But now you can utilize these AI tools. There’s some limits, obviously, but you can make a great personal website. You can make something to plan your garden.

[00:05:21] Justin Moon: Like you can, there’s all kind of tools, personalized software tools that you can create or you can vibe, code, a Bluetooth mesh chat app that is being used in protests in Nepal at the moment. That’s what Jack Dorsey did. ‘because as I think that one of the interesting things he was talking about, how, so he spends like three hours a day doing this every morning.

[00:05:36] Justin Moon: He designs his day so he can get three hours. He’s always trying to push the limit of what’s possible. And that’s certainly opened my eyes. It’s okay, if he’s able to do this. Public company, maybe the rest of us try to level up a little.

[00:05:47] Preston Pysh: Talk to us about that. What is this three hour thing that you’re saying?

[00:05:51] Shroominic: Yeah, so no, so I was asking Jack like, how do you have time to do all this stuff? He’s, doing all this vibe coding and it’s aren’t you busy running a company? And he is I just set my schedule so every morning for three hours I can play with these tools. ‘because I think this is the future of like kind of the economy, right?

[00:06:04] Shroominic: And small business and all the things that Square and Block exist to serve. So it’s my responsibility to kind of understand what the frontier is. And the only way you can do it is by trying. And, yeah. You also had a good message to try to figure out what it can’t do, right? Push it to the point where it fails. And that’s, we’re at, we’re here where in Madeira. Shroominic and I are in Madeira at Gigi Sovereign Engineering Programs.

[00:06:23] Shroominic: This is like a program, it’s like an idea factory for Nostr Lightning. Now, this time it’s more AI focused and that’s what we’re trying to do every day. We try to get it to do really ambitious things and don’t just make the website, see where the boundary is, and always try to push that boundary.

[00:06:37] Preston Pysh: This idea of the semicolon for people that aren’t progprogrammers, I think what you’re really saying is, you get really good at a particular language of code. And you master that. You know exactly where the quote unquote semicolon goes so that you’re not making mistakes. But then you really get optimized into only coding in that language. And if you do go outside of it, it’s here and there.

[00:06:59] Shroominic: For me, the thing that’s frustrating is like I become like I’m not natural. Some people are just naturally, like their minds think like a computer, right? Like I’ve worked with a few people like that and they’re great to work with, but mine’s not like that. So I have to become like almost a different person to be like a really good progprogrammer.

[00:07:12] Shroominic: I’m not good at talking to people. I’m a lot less friendly. I’m more literal. I’m not going to be drawing in my spare time. I almost have to become a different person. And after a while it becomes a little exhausting, so you have to turn it off. And so that’s a, really fun thing about these AI tools is they allow you to just stand like one level higher where you’re, orchestrating it and you have a higher viewpoint. And I find it, it’s a little more creative and a lot less, it’s less manual labor in a sense.

[00:07:39] Preston Pysh: Yeah. You’re able to step back, see the bigger picture, and if there’s a certain type of code that is more optimal for use that is nested underneath of the larger program, you don’t have to go in there and become the expert on it. You can really leverage this AI expertise to do it.

[00:07:57] Shroominic: So I have one maybe interesting back coding story. I was, looking into Cashu, which is this like ecash protocol. I was working on a Python library and there was not really a Cashu python library for the specific thing what I was trying to do.

[00:08:12] Shroominic: So I buy coded like a completely new library from scratch, which is like hard because you think like building programming libraries is like more of the thing that like senior engineer do or like really experienced engineers and with, I use like Cloud for Opus, which is pretty expensive, but I could even create like really complex parts of the software within like a short amount of time.

[00:08:34] Shroominic: And what’s also interesting, this was I think the most amount I spent on AI. So building this library cost like 400 bucks just off AI compute. So you can burn through a lot of like money just by doing these things.

[00:08:47] Justin Moon: This is a really interesting one. I want to expand on that. And this is like one thing I think we want to reflect on. Think about what AI means, and this is a software, we’re taking a little bit of a software angle, it is, that’s the one, the thing that AI has been probably best have so far.

[00:09:00] Justin Moon: So an interesting thing about software thus far is that it’s not capital intensive, right? Like the story of Elon Musk’s career was that he started making a little thing, it’s called like Zip2. Where you could make a website, right?

[00:09:11] Justin Moon: And so what he’d do is he would leave his, he had one computer and he was like renting a dorm or something, and he would leave his computer running during the day and that would serve people’s websites. And at night he would code the websites. So he just reversed his sleep schedule and that took no capital to start that business. ‘because he just used repurposes existing computer. And on Shroominic’s story here, it’s one kind of new thing about AI is it’s turning software into a bit of a capital intensive endeavor, right?

[00:09:35] Justin Moon: If you want to be, like we were talking about how much these things cost, my approach is like you, especially if you have a productive use for this stuff, you should so throw as much money as you can at it.

[00:09:44] Justin Moon: Just, so like last time this sovereign engineering program, six months ago, the top level vibe coding setup was like 20 bucks a month. Now, today it’s 200 bucks a month. There’s a new level for max and a year, what if it’s $2,000 a month? Like now all of a sudden it’s something where. It’s much harder to get started in it. And maybe in five years it’s 50 grand a month to engage in software engineering.

[00:10:07] Justin Moon: So it could become more like other disciplines where if you want to build a bridge, you can’t just start, in your garage anymore. So that’s one thing that’s very interesting about this whole side-thing phenomenon.

[00:10:18] Shroominic: I think it’s also about how many agents can you maybe run in parallel. In the beginning you just ask ChatGPT and you got one script out, but now you, you ask the coding agent and it’s spinned up like multiple in parallel and then does way more work while you like waiting for it in the background, which costs way more computers.

[00:10:37] Preston Pysh: You know what? At the end of the day, because there’s so much training involved where you’re taking these massive data sets and you’re training it on something, on a pattern. It’s very specific to solve a problem that you’re trying to go through. When you go far enough upstream of that, it’s energy to plow through the GPUs, to come up with the pattern or the weights of said model that you’re trying to train.

[00:11:01] Preston Pysh: And so it is interesting to see, and of course as Bitcoiners, we can go back to Bitcoin kind of being that fundamental energy unit that’s being exchanged, but we’ll leave it there. But it’s a really interesting point of that transition taking place because people are trying to train their own models.

[00:11:18] Justin Moon: This has been a beautiful thing actually. It feels a couple years ago on Bitcoin, Twitter, for example, our group would be fighting against Silicon Valley people saying Hey, we should have more energy. Using energy is not bad. Emissions can be bad, pollution is bad. Creating energy and consuming energy is not like a priority, bad.

[00:11:36] Justin Moon: We would lose all these arguments. The Silicon Valley people would generally wouldn’t listen to us. They wanted their green data centers at any cost, and this has totally changed, right? Yeah. All these Silicon Valley companies are now, I just saw a story, Facebook’s doing a huge natural gas powered AI place and in Cheyenne, Wyoming, and I think it uses as much energy as all the homes in Wyoming combined. It’s amazing to see how Silicon Valley saw the light on energy and energy production through AI, which we of course failed to teach them through Bitcoin.

[00:12:05] Preston Pysh: Alright, so Justin, you’ve been pioneering this idea of AI for activists and you’re working a lot with the Human Rights Foundation. Help us understand what initiative and what you’re trying to accomplish with that.

[00:12:18] Justin Moon: Yeah, so I think pioneering might be a little, I don’t know if I’d use that word myself, but Craig Warman and the director and Alex Gladstein, the head of strategy at HRF are really in the way. I’m just more supporting role I would say.

[00:12:29] Justin Moon: It’s funny. So I’ll, yeah, I’ll tell a little story. When Alex first reached out to me about doing this, about the kind of it helping with the AI program, I had recalled AI. I pulled up his Twitter and I searched AI because I remember he had opined about AI, and it was always the programing.

[00:12:43] Justin Moon: This was like five years ago. It was, the programing was like, coin is for freedom. AI is basically totalitarian, communist control. So it was it was interesting that he had this idea and so I asked him like what changed? And his answer was basically they’re already seeing, kind of like with Bitcoin, why did HRF get involved with Bitcoin? Because they saw that it helped activists get money into the Ukraine.

[00:13:02] Justin Moon: The banking system couldn’t do things that Bitcoin could, right? And so that’s how HRF got into Bitcoin, is that they were able to do things in their democracy activism with Bitcoin that they couldn’t do in any other way. And they were seeing the same thing with AI, right? You have to write a lot of grants if you’re an activist, right? It’s the quarter of your job.

[00:13:18] Justin Moon: And so all of these activists found ways to write a grant four times faster, right? Using ChatGPT or something like ChatGPT. Some drawbacks. And so I think that was the spark like, okay, this is actually really helping activists be more productive in their work.

[00:13:32] Justin Moon: And that’s, I think, pretty off the most important thing is to empower activists, right? To help them win at whatever their fight is, right? And so I think it, it grew often was like, so what could HRF do? And we’ve started to have quite a lot of things. We’ve done education initiative.

[00:13:44] Justin Moon: So at the Oslo Freedom Forum, I think I did 10 hours of workshops teaching people to use ChatGPT and image generation and transcription and all these different things. We’ve continued some of that. We’re starting to do some monthly kind of webinar type stuff. I think in the future we may be funding kind of freedom, sovereign oriented AI projects, grassroots AI things.

[00:14:05] Justin Moon: Like you’ve seen HRF do in the past, holding more events to bring people together, bring activists together with technologists, together with philanthropists, stuff like that. These groups that kind of never talk to each other, right? You see a lot of that. The chat, I referenced it earlier, right?

[00:14:18] Justin Moon: That’s one of these things that the Bitcoiners created. They really would’ve never maybe even seen the need for, if they hadn’t been introduced to some of these activists by HRF, so, yeah, that’s where it’s at. And we just started six months ago and we’re just getting things up and running.

[00:14:31] Justin Moon: But yeah, it’s really exciting and yeah, you can see an awful lot of promise here.

[00:14:35] Preston Pysh: Yeah, I think BitChat’s a perfect thing to talk about on this particular topic because here you have a new app that enables communication without having to go up and through the whole wifi network. Or internet network, you can use the vicinity of, if you’re in close vicinity to somebody, you can use the, emission of the phone in that close proximity to have communication.

[00:15:02] Preston Pysh: And BitChat was an application that Jack Dorsey has recently released. How much vibe, coding, what’s the rumors on how long this took him to put this together? And just as a reference, like this is doing really well on the app store. it’s gone out with quite a splash now. It is Jack that coded it.

[00:15:22] Preston Pysh: So you get a lot of marketing automatically through that. But I think the idea to the actual release to people using this and how it fits into this activist layer where you’re not having to communicate over traditional landlines is pretty miraculous. So what have you heard on the inside on this particular application?

[00:15:44] Shroominic: I have no inside info, but I would guess it was like a week or two ‘because there was a period there where he was trying to do once a week and on the lap a week. And I think that’s about right. this is the. One of the downsides of vibe coating. I was just joking with a friend who had a vibe coated a project last week and it was working really well. And now this week it’s just like totally stalling.

[00:16:00] Shroominic: It’s like you’re pouring water into a jar and it’s, oh, look at the water rise. Look at the water rise, and it gets about 80% of the way there and it just starts pouring out the side. That’s sometimes the. Experience with vibe coding. for about a week you can oftentimes really, do great.

[00:16:15] Shroominic: And I think maybe, I’m not sure how involved Jack is now, but I bet there’s a lot of other people, contributors I can just see on GitHub. I’m not participating closely. he just made an iOS version ‘because I think he has a iPhone and then call the creative Cashu was like, let me try to vibe code this in Android.

[00:16:28] Shroominic: Now Android developer has never made an Android app and he similarly, as Shroominic mentioned, how he attempted with a Cashu library in the past. He was able to get it, to translate it from one language to another. This was historically a very difficult task ‘because you’d have to learn two different, totally different ecosystems, ways of doing it.

[00:16:43] Shroominic: Each one of them would take a few months to really get intuitive at, be able to understand, and he was able to do this. So I think in about a week. So these are, both kind of shocking examples. These are the types of things, I don’t think any of us would’ve really believed a year ago, but it’s slowly starting to happen and yeah, very exciting.

[00:16:59] Preston Pysh: And as a kicker on top of this, Callie, the person that we’re talking about, that took the iOS version and turned it into an Android version, hasn’t he got the Cashu, basically tokenized Bitcoin, saleable Bitcoin. You’re able to transact without an internet connection, is what I’m getting.

[00:17:17] Yep, through BitChat. You have to be using BitChat.

[00:17:21] Justin Moon: So all these things are, I think are really interesting. You think about the things in Bitcoin that work and the ones that don’t work, like BitChat and Cashu and some of these other things that are like, it’s not like the optimal solution always.

[00:17:32] Justin Moon: If you’re communicating, it’s probably nice to be able to send money through the whole internet at times. Sorry. Some messages through the whole internet, right? Across the internet, not to have to use Bluetooth, or if you’re transacting, sometimes it’s nice to use on chain Bitcoin, right?

[00:17:44] Justin Moon: There are these kind of like the right way to do it, or maybe sometimes easier a credit card or something, right? But the thing about interesting, about a cash or a BitChat is that it’ll always be there, right? Like you can always use that.

[00:17:54] Justin Moon: It’s like a backstop of freedom, right? It’s something that kind of can’t be taken away unless they take your phone, right? Like you can always find some mint to spin up that can help you connect you to the Lightning network that will allow you to, transact with other people who also trust the Mint.

[00:18:07] Justin Moon: You’ll always be able to use Bluetooth with other people unless all the phones go away. I think that’s some of the interesting, like similarities with some of these Freedom Tech projects. It’s something that, Routstr which Shroominic is working on. It’s one of these other things where if we ever get to a world where you got a KYC and everything you do to interact with AI is super monitored, maybe something like Routstr would be a way where you can say, Hey, like none of these things will answer the question I want to do. I’m doing a science thing.

[00:18:31] Justin Moon: They all say this is for, they won’t let me answer the question. Maybe there will be a way where you can break through that censorship and actually access kind of the self-sovereign AI do something like Routstr. These are the class of things that I think are the most interesting in that Freedom techniques.

[00:18:45] Preston Pysh: I want to just talk about the tech on this BitChat a little bit more because I find the technology and what he’s harnessing here really interesting. So when you’re talking about Bluetooth out or indoors, it’s about 10 to 30 meters as far as w hat that transmission will give you.

[00:19:00] Preston Pysh: Outside it’s about a hundred meters. But what he’s done is he’s taken a mesh relay system. So if you have this app. Let’s say everybody around you had the app, it can then through the mesh network of all the Bluetooth out there. So let’s say you’re at, and I’m going to really, maybe the younger generation is going to laugh at this comment.

[00:19:21] Preston Pysh: You’re at a mall and you’re around a lot of people. Or you’re at a sports stadium, you’re wherever where there’s a lot of population density and a lot of people with smartphones. That network is really robust and you’re able to relay these messages. I’m a, I’m assuming the messages are encrypted as they’re going from one, device to the next. And so you’re able to actually extend that range, like really significantly. imagine you’re in a city, that network is pretty robust.

[00:19:51] Justin Moon: Yeah, think of it like the original post office, right? It’s you couldn’t drive a horse across the United States, but you could ride a horse 30 miles and then, the male could jump onto the next horse, right?

[00:20:00] Justin Moon: If you get anywhere, it’s the same principle here. You don’t drive 30 miles, you drive 30 meters. Yeah. But you could, we’re on this little island here. I’m pretty sure you could get a message across. Maybe it might take a day, right? It’s kind of like you’re going back in time a little bit stuff. It’ll take a little bit. You’re not moving at the speed of light anymore, but in a lot of circumstances.

[00:20:16] Justin Moon: Let’s say you’re trying to get the news out right about something that’s good enough, right? You can get it across the city, and then you can also combine these technologies, right? Let’s say somebody is in a compromised situation, they can’t get the internet, they’re maybe something like that.

[00:20:28] Justin Moon: They don’t want to be monitored. Maybe they’re able to hop a few times across BitChats, and then someone’s okay, I’ll blast that out over Nostr. Like you could do something like that where you know, it, makes a few hops and then eventually it’s welled publicly.

[00:20:41] Preston Pysh: To that point, you could make a published to Nostr in an encrypted way, and if the person on the other side has a key to unlock that message. Wow.

[00:20:50] Justin Moon: I think there’s all kinds of fun things that will be possible here. And yeah, I’m really excited to see where this takes us. I think there’ll continue to be a blossoming of different techniques. I have some ideas, but there’s sorts of things I want to try over the next couple weeks while we’re here at SEC.

[00:21:03] Preston Pysh: I love it. How exciting.

[00:21:04] Shroominic: What’s also interesting to understand, like BitChat should, if we, for example, have the situation of. So the government going to find like certain whistleblowers or like whatever, where there are activists which are like publishing messages. With BitChat, there’s not really a way to find the origin of the message.

[00:21:22] Shroominic: So you can hide like where you are, like spreading this message from, in a way.

[00:21:27] Preston Pysh: Wow. Yeah, and I think for people that maybe are skeptical or why this would even be important, they’re, hearing it and they’re saying, yeah, but how many people have the BitChat app? And what I think they’re missing is with Open Source, the whole open source initiative.

[00:21:41] Preston Pysh: Nostr, for example, this is a version of Twitter. There’s a lot of people using Nostr these days. Relatively speaking. It’s still very early, but there’s a lot of people using it. And this technology of BitChat can just be. Onboarded into any implementation of a client that’s running Nostr, and now all of a sudden you have this ability to message over a Bluetooth mesh network messages that would’ve never been.

[00:22:06] Preston Pysh: So you get a network effect by having some type of communication, open source protocol like Nostr, that as these new. Developments are built, they can be onboarded into that higher level network that gives you all that capability and gives you that network effect that you might not be able to get by just people downloading the BitChat app, which I find crazy fascinating.

[00:22:28] Justin Moon: One of the other things that’s really cool here is that it’s not like there’s some, the way it’s worked in Silicon Valley in the past, it’s okay, some app comes out, it has a feature that’s cool, that’s useful, and you just pray that they don’t blow it. Then sometimes you get, I don’t know, it was that thing that Twitter had where it was like the five second videos that was so popular for a while and then they just blew it. It was just gone.

[00:22:48] Justin Moon: I forget what that was called, but oftentimes, think about all the products Google’s killed. The cool thing about these things is like, it’s all open source, so if the creators blow it, someone else will carry the torch, or I have some ideas about these things.

[00:23:02] Justin Moon: One thing I could do is try to contribute to theirs, or I could just make my own little spin on it and they interoperate. That’s the beautiful thing about a protocol is like in order to contribute to it, you don’t need to just find a way to get a job at the company and jump through all these hoops and 99% of the people in the world can’t do that. You can just take what they did and try your own little spin on it.

[00:23:19] Justin Moon: You don’t have to ask them permission, you don’t have, all you really need to do is to have the skill to do it. That’s where kind of AI comes in, right? Like the skill to do it just went down by maybe an order of magnitude, right? There’s all kinds of things where, especially to get that proof of concept, to get something that shows that your idea works or is useful, that just got, that just went down massively over the last year.

[00:23:41] Justin Moon: So this is very exciting in terms of decentralizing the development of these kind of freedom technologies. Making it so it’s not just five, 10 people creating these things, they come from everywhere.

[00:23:52] Preston Pysh: That’s where I want to go next is in the decentralized nature of AI itself. I think this is hard for people to really wrap their head on around is the application of like, how do you do this? Because when you understand just the bare basics of AI, it’s like it’s got, the more data you feed it, the better it gets, the more informed it gets. And so you see these large language models like ChatGPT. And it’s just literally sucking the data out of every single human being on the planet to learn and get smarter.

[00:24:22] Preston Pysh: And so people are seeing that and they’re saying, okay, so how is this going to work? Localized or an open source AI. How is it going to be as smart as that? How can I harness that? Without giving up all my data and all of my information moving forward, so I know Shroominic Yeah, you’re an expert.

[00:24:39] Shroominic: Let me say a high level one here and I’m basically going to try to give a softball to Shroominic. So Pat, this, is how I think about it, right? so it’s what is AI? There’s two big parts. There’s the creation of it and the actual using of it. So the creation of AI is called training, right?

[00:24:52] Shroominic: And training. You need a lot of energy. You need a lot of very fancy, expensive computers and you need a lot of data. And if those three pieces go into training and you need some clever engineers and stuff and like that, maybe that’s the fourth element, right? And that’s going to be hard to decentralize to the point where like you can do it in your basement.

[00:25:09] Shroominic: I’m skeptical that will ever happen. but let’s say a year ago, a lot of people were scared that it would centralize further, right? That one company would get the edge and then they would enter like a parabolic explosion of intelligence and they would just get further and further ahead.

[00:25:24] Shroominic: I would say the opposite has happened, right? A year ago, like ChatGPT was really good. Claude was pretty good, and then maybe there were a few others and, but that was about it. There were very few participants. Now it’s very unclear who has the best AI, right? ChatGPT-2 has just launched. Maybe it’s there, but it’s not super clear.

[00:25:41] Shroominic: Claude’s also very good. Grock is very good. There’s some Chinese ones there. You work working with a lot less money, right? Like Kimmi, K two GLM, deep seek. There’s a number of people at the companies at the Frontier has probably tripled since last year.

[00:25:56] Shroominic: So in a sense it is decentralizing, right? We don’t have one, and it’s it is a lot better to have three options than one. If you have one option, it can be very, you can get dark quick, right? That’s a single party state is like that, right?

[00:26:08] Preston Pysh: What’s causing, I know with deep seek, they were able to reverse engineer and do it at just a fraction of the cost because they basically went to ChatGPT and were asking it certain amounts of questions and reverse engineering the training on it.

[00:26:23] Preston Pysh: But I don’t know the terminology or exactly how that takes place. Can you guys walk us through what they did to be able to do that?

[00:26:29] Shroominic: So I think what Deep Seek did is like they created a bunch of synthetic data. So you’re like creating imaginary charts that like tell information. For example, you ask a certain question and a teacher explains you with detail and they explain you like the reasoning steps of like how you go from that knowledge to like deriving some more information out of that.

[00:26:51] Shroominic: They created a bunch of synthetic data using ChatGPT, where they let ChatGPT explain like really deep mathematical things or like programming things and ChatGPT was explaining the reasoning steps that they didn’t need to write it by hand, but then like they could train that AI model on these reasoning steps. They extracted knowledge out of ChatGPT to train it like really efficiently.

[00:27:14] Preston Pysh: So they had a bot basically go ask just an endless array of questions, and then the ChatGPT gave an answer to all of that, and then they just synthetically use that to continue to train it. Okay.

[00:27:25] Shroominic: Or it could have been like every single time one of these things goes through the great firewall going back and forth to open AI. Somehow the party, the party logs everything and it’s okay, here, trained on this. But it is yeah, it’s by communicating with an AI, it’s leaking its model weights, right?

[00:27:38] Shroominic: So that’s one thing. Like the intelligence and the model kind of wants to get out. If you talk to it enough, you can suck at least all the useful things out. It’s like that. There will be blood seen, I drink your milkshake, right? I think it’s one of those. So it’s one of these things where there are a lot of moats and there are a lot, there are some moats in, I obviously it’s a capital intensive, right? Not everyone’s going to raise a hundred billion dollars or whatever.

[00:27:58] Shroominic: This is an area where the, I’m sure Sam Altman wish the moat was a lot bigger, right? Like a lot harder to, pull the intelligence out of his models. yeah, that’s, I think that there’s a balance where it’s the more useful they are, the more they’re used, the more the stuff gets out in the world anybody else can train on that deeper.

[00:28:14] Shroominic: I think more techniques as time goes on, people learn more techniques for doing all these things and yeah, I don’t have, an explanation for it. I’m very happy that it happened though. ‘because it has you, things have decentralized a lot over the last year.

[00:28:27] Preston Pysh: Justin, don’t you think that’s like one of the reasons why it’s not going to dominate in the dark dystopian. Talking point that so many people had call it three years ago, that doesn’t seem to be what’s playing out and that there’s going to be a bunch more models because they’re able to do this synthetic intelligence collection from the models that perform the best.

[00:28:48] Justin Moon: Yeah, part of me is wonders whether we’re getting close to the point where like whether kind of increased intelligence is taping off. Like at some point maybe that’ll happen, right? and every point in the last couple years, like AI, after six months, it’s better than I thought it would be. And this is the first time where it’s not.

[00:29:03] Justin Moon: So we did a lot of putting here. Six months ago we were paying 20 bucks a month. It was the golden days. And now here we are paying 200 bucks a month throwing our computers against the wall going, we’re not doing that. It hasn’t gotten better, but it hasn’t felt it hasn’t done that like huge step function that has stayed every six months in the past.

[00:29:16] Justin Moon: And G PT five was like this thing that was built up over a couple years and when it happened it was like, oh really? That’s nice. And it seems to be. A lot of it is like they’re switching between models in the background. They don’t have one. One kind of model, $10 models, it’s getting awful expensive.

[00:29:30] Justin Moon: If you want 20 bucks a month is easy, 200 is rough, 2000 is like an employee almost.

[00:29:36] Preston Pysh: It’s interesting that you say that though, because I know when I had the $20 a month. Then I was tinkering with someone like the deep research and I was like, oh wow. Like I could just do this as much as I want for 200 a month.

[00:29:47] Preston Pysh: Yeah. And I was like, that’s a lot of money. But then I’m like in the back of my head, I’m like, look at what this thing is giving me. For $200 a month, it’s like free compared to If I had to hire somebody to do these things It would be nowhere near these prices.

[00:30:02] Preston Pysh: And I think it’s getting to the point where that’s becoming very normalized for people to start thinking of the cost in terms of what would this cost if I went to a person and had them do some of this stuff.

[00:30:12] Justin Moon: I just thought of this now. there maybe is like some stranded energy angle here. that reminds me of Bitcoin. Because the, Grok was an interesting one. Like it just came out of nowhere. It was Elon and All the other AI companies at the time would do these like green data centers, and that has to see carbon zero.

[00:30:27] Justin Moon: And they just really bend over backwards to that. And then they would also do a lot of alignment, right? Make sure that the thing didn’t misbehave and Elon basically threw those out of the, they threw ‘them out, right? I’m not going to align it at all. I’ll make AI boyfriends and girlfriends instead. And I will, instead of having a green data center, I’m going to get a data center in Memphis and I’m going to park like 40 natural gas generators out front as the walls are being put up as we’re like building this thing like in the air, and they found an energy arb there.

[00:30:54] Justin Moon: That’s how that one came about. I wouldn’t be surprised if some of the Chinese ones are a similar thing where they’re just basically finding, maybe they’re paying a lot less for electricity than Sam Altman is, stuff like that. I think there could be some like, geo arbitrage that we see in Bitcoin mining happening as well.

[00:31:07] Shroominic: I think this also confirms the point a bit that like it’s hard to decentralize the training. ‘because what Elon did, he put like the most amount of GPUs you could ever find onto one single place in one single factory. Which is like the thing, you need to train the best model. So like you need to centralize all the GPUs you have at one single place to get the best model, which is like anti decentralization.

[00:31:31] Shroominic: So the person that gets the most amount of GPU to the smallest place possible we’ll train the best model because in the end you need like the bandwidth between the GPUs and the GPUs communicate with each other. As soon as they get closer to each other, they can communicate way faster.

[00:31:47] Preston Pysh: That’s interesting.

[00:31:49] Justin Moon: Yeah, so there’s something like called a Nous, N-O-U-S. It’s like some research group and I think they have a coin, so they’re trying to do like a decentralized training run. But for me, this is something where it’s like, what’s the right amount of decentralization? It’s not like something, you don’t want to, like decentralization is in a family, it’s not necessarily good, right? Like sometimes you want to be close together and at a dinner table, right? you don’t want to be on different continents. So they’re like,

[00:32:11] Preston Pysh: but this is important. This is for training the model. So like you got the performance of, I’ve trained the model, I have the model, and then I’m expending whatever amount of de minimus energy to plow it through the model, to give me my output. But what we’re talking about, and I think to Shroominic’s point is when you’re training the large language model, having all those GPUs right next to each other is almost a must if you’re trying to synthesize all of the data as you’re trying to develop the weights for the model, which I think is very different than the utility. You’ve built the model now I’m going to use it day to day for whatever tasks, right?

[00:32:49] Justin Moon: Yeah. So at the beginning , I mentioned like there’s two ways to look at AI. One is the creation of it and one is the running of it. So creating it training, it’s like there’s these big centralizing effects, right? You need really fast bandwidth between them. You need a ton of electricity in one place.

[00:33:01] Justin Moon: They are getting to the point where they can spread it out a little bit, but there’s just these natural scaling laws where you just want it to be big and in one place. But inference that’s running the model is a totally different ball game. And so that’s where I think a lot more decentralization can happen.

[00:33:15] Preston Pysh: I’m sorry to interrupt you Justin, but this, I think this is an important conceptual like talking point or philosophical talking point of like where this is going to go. How big do we want these large language models to get before you start peeking out? And it makes way more sense to have a specialized model that’s just medical or one that’s biological or one that’s physics based.

[00:33:38] Preston Pysh: And so I think you get to a point where the large language model peaks out, and I think it maybe happens sooner rather than later. And then that GPU Farm that was built, and of course there’s advancements happening on the hardware side of the house and the software side of the house to run these really large language models, but do you see a world where that peaks in the coming five or 10 years or 20 years, or does it peak first of all.

[00:34:03] Preston Pysh: Then kind of timeline of like where you think something like that would peak and then it all gets into the smaller, more modular models that then are stitched together to give you optimal intelligence.

[00:34:15] Justin Moon: It’s really hard to reason from first principles here. what do we have to compare it against? The human mind, right? The mind has like many different little AIs. Many different componentsthat are stitched together. So if I had to guess that’s how AI is going to be this is very much finger in the wind.

[00:34:28] Justin Moon: Our examples are one of one, this is how the brain works. So I’d guess AI will be, we’d be different. This is why I’ve mentioned the cost a few times too, ‘because we’re not that far from it getting prohibitively expensive, right? There was a time where like spending a dollar on 10 cents on AI, it seemed like a lot, and now we’re spending 200 and it’s going to get to the point where it’s cost prohibitive, like another generation or two of these LLMs, it’s going to be cost prohibitive to run thing on the frontier for many people.

[00:34:53] Justin Moon: So I think you will have more specialized ones. Like you’re seeing this with medical knowledge too. Like I know there’s a couple that have been very successful. One of my friends who’s a doctor actually uses one of these to like kind of fact check, to check his reasoning against when he is diagnosing patients.

[00:35:06] Justin Moon: I can’t say that I have a prediction here ‘because I don’t know AI enough and I don’t trust the people that actually do make predictions either, So what do you think, Shroominic?

[00:35:14] Shroominic: Yeah, so I think we also need to think about what can the in top humans achieve? If you think about maybe someone as smart as Elon Musk or top engineered open AI, if we could get them like just 10% smarter, like what could they potentially achieve more? Or 20% smarter. If we had double the intelligence of the most intelligent engineer at Open AI, maybe we could achieve like even more things, but we don’t know it because that’s like the current limit. So it’s like how to think, what can we get more if we have more intelligence?

[00:35:46] Preston Pysh: Here’s an interesting stat that I just looked up while we were talking here. The human brain vision, the occipital lobe, represents about 20 to 30% of the entire cortex of neurons language, which is in your left hemisphere, estimated at a few percent. Just language itself is a few percent of the neurons in your brain. The motor cortex, and I find this one really interesting, 70 to 80% of the brain’s neurons are dedicated to your motor control, which is in your cerebellum.

[00:36:17] Preston Pysh: . So it’s interesting that like when we start getting into the humanoid robotics in that, where you’re going to start putting these ais into human form or into some type of modular form where they can go out, they can sense their environment, they can make decisions inside of their environment.

[00:36:33] Preston Pysh: That motor cortex for humans is encompassing 70 to 80% of the brain’s horsepower to be able to, I don’t know if you guys have watched some of the conversation around how difficult it is from an engineering standpoint to make a hand. And all the tendons, and for it to be able to pick up a ball and throw a ball, like how complex that is from a mechanical engineering standpoint, just designing the hand itself and then you start getting into the AI that it would have to be trained on to basically pick it up for humans that, according to what I’m reading here, is an enormous part of the horsepower or the mental models that are needed in order to do it.

[00:37:12] Justin Moon: Yeah. I think the big difference between me, between an AI and person right now is for me is that the person is, has embodied and has experience and an AI doesn’t. Right? And that’s the big philosophical question is can you embody an AI, right?

[00:37:27] Justin Moon: If you give it a robot form, is it like embodied like a human is? Does it experience stuff, right? Like when I’m trying to get it to write an app, to test my app like every, it will frequently just do things that are really silly. That is silly. No person would ever do that, who like was able to open a refrigerator. Oh, you can’t open a refrigerator. I forgot. And that’s another one of the things is like AI is one of the huge limitations is when you talk about whether they could do a human job, right?

[00:37:52] Justin Moon: The first week of anybody at the job, they’re pretty useless, right? And the second week they start to get some components of the job. Then by a month they’ve picked up quite a lot. And they’ve actually learned it through experience. And no AI that, no LLM can do this at all, right?

[00:38:05] Justin Moon: Like the best thing that they can do is summarize a little bit of their learning into a text file. It just gets prepend to the questions you ask it. It’s totally cheating. The thing doesn’t learn at all. It’s the thing I’m using and the thing you’re using, and the thing Shroominic is using is exactly the same.

[00:38:21] Justin Moon: It’s totally stateless. It’s a big question. It’s so can you embody that and give it experience that it could actually learn? And I think maybe it could be the case that we all really undervalue how important that is to actually, an intelligence.

[00:38:35] Shroominic: Also, maybe I wouldn’t fully agree with the statistics ‘because I don’t think we only reason in language. For example, if you think about like math, like you often have some like spatial representation in your head, which could be also happening in the part of your brain where maybe the motor part is, like when you think about like zooms and orientations or like spatial ations, so maybe like some reasoning is not happening in the language part of the brain.

[00:39:02] Preston Pysh: What’s something that you guys are most excited about in this space? Obviously some of the stuff we’ve already mentioned is beyond exciting, but is there anything that you’re seeing that you’re just like, wow, this is something that I can’t wait for or that you’re already seeing right now?

[00:39:16] Shroominic: I think these are really complex coding agents are like really exciting. ‘because as you think about in the end, like when you build software, you just want to solve a problem. And there are some engineers who like just enjoy their process of writing code. But mostly you have some problem in the real world and you want to solve it.

[00:39:35] Shroominic: So you like grow one of these like instant coding agents on that problem. And then you can solve it. So I think like having code that like costs basically nothing or like just decreasing the amount you need to pay for, like having a really complex piece of code. I think that’s maybe the most impactful thing we have in AI.

[00:39:55] Justin Moon: I guess I’ll give a little bit of a different answer. Like one thing I’m really excited for that it seems like it hasn’t really materialized yet, is just being able to apply these things to education in general. Like I get two real big benefits out of these tools.

[00:40:06] Justin Moon: One is like they write code for me, so I don’t have to do that anymore as much, which is really great. But I’d probably say even better is like when I have some kind of an idea, you can learn so much by just going back and forth and we all kind of invent workflows and you have to pick this up. And it largely depends on how much agency you have and how creative you are. And. Ask a lot of the user, so to speak.

[00:40:29] Justin Moon: So I think it’ll be very interesting for these things. I’ve been able to learn a ton with these, but I suspect many other people, we haven’t figured that out, ‘because you have to really put a lot of effort in up front. And I think it’s going to be really interesting in I can see that being pre transformers, especially in America, right where Preston and I are from, the education system now is just, it’s just so bad, so broke.

[00:40:52] Justin Moon: It’s just it’s so bad. like one, one of my friends was joking about how he picked up like some old education book, like a ninth grade, a composition book, and he’s so I was picking up this book and of course it’s for ninth graders, so I couldn’t understand it. because the reading level they had in ninth grade a hundred years ago was higher than adults today.

[00:41:09] Justin Moon: Probably not true, but I mean it’s, it is a bit of a joke, but it’s, there’s some truth to it. So I think that’s something I’m pretty excited about. And yeah, I think in the sort of the Freedom Tech ecosystem that we’re in, I’ve been like working in it full time, just trying to find some way to do it for six, seven years now.

[00:41:22] Justin Moon: And so I often have talked to people who like will come to a conference and they’ll have some idea like, oh, I have this nice day job so I can’t do it. Yeah. And whatever. And so you’re starting to see people like this, like just at this little event we’re at, there’s a few people who have day jobs who are like, oh, let me just come and come and try to do it a little bit on the side.

[00:41:38] Justin Moon: So it’s, I think you’ll get a lot more people finding a way to contribute on Bitcoin, on lightning, on e cash, on some of these, on Nostr, just in their spare time without having to devote their life to it. I think that’s very interesting.

[00:41:51] Justin Moon: I think another interesting thing is just like not having to look at a screen so much. That’s, that is one of the things I think about our modern work environments that are pretty horrible. It’s like a screen’s not a good thing to stare at for 10 hours a day eight hours a day.

[00:42:02] Justin Moon: And if we could go back to every time I go and do something with like manual labor for a week or something, it’s just I’m so much happier than staring at a screen all day. So I think that’s another really interesting thing. If we could move to modes of working where you don’t have to just stare at this artificial screen all day. I think that would be pretty interesting.

[00:42:19] Preston Pysh: On the education front, the thing I’m excited about is just the customization to a person’s natural interests and talents. Where like today, it’s everybody’s just force fed. Oh, nope. Sit down in the class with 40 and this is what you’re going to learn. And five people out of the 40 even have an interest in the topic. And the other ones are just looking at the clock saying, when is this going to be done?

[00:42:41] Preston Pysh: And where I think a lot of this is going is you’re going to have the best instruction ever. Because there’s no ego, there’s no past experience of the teacher themselves trying to, maybe you have a teacher that loves poetry, and so they’re just trying to jam the poetry down the throats of all the students, and there’s, three people there that love it too.

[00:43:03] Justin Moon: Johnny likes World War II Tank.

[00:43:05] Preston Pysh: That’s right. And so I think that customization piece is going to be huge and the removal of the ego of the instructor is going to be huge to most optimally, help the person learn what they’re naturally gifted at and what they’re naturally interested in. And I can argue the other side of why that’s also bad is because if you don’t get enough exposure to the things that maybe you would’ve never tried. Now you’re just pigeonholing somebody because they had an early interest in something potentially, right?

[00:43:34] Justin Moon: Yeah. I feel like an AI could be much better at I mean there’s all kinds of things like poetry. It’s like when I was a kid, it’s tough to get me interested in poetry, but if you made the pitch at just the right time, like if I was in trying to impress a girl or something, you made the pitch, then yeah, I would be a little Shakespeare.

[00:43:48] Justin Moon: Like I think that an AI might pick up on that AI education system. Yeah. It’s like it could be much more opportunistic than a textbook can.

[00:43:55] Preston Pysh: Yeah, you’re right. So poetry’s going to be taught at age what, 13, 14 is what you’re saying? That’s really interesting. And ‘because I was always, to be quite honest with you, I was always very frustrated with school. I wouldn’t say I was like a great student. I just was always frustrated with having to learn things that there was a lot of things I had no interest in whatsoever. And then there’s other things that I was super interested in.

[00:44:19] Preston Pysh: I think so much of that. It’s like the Montessori schools and how they really kind of lean into what the kids’ interests are. I just think that education is going in that direction, like a free train. The other thing that I get really frustrated with AI is these teachers that are like, don’t go near it. It’s the devil. It’s like Bobby Boucher AI is the devil, Bobby, the Adam Sandler movie. But I think that’s just as bad as somebody who’s just telling their kid that they should only be using it.

[00:44:48] Preston Pysh: I think it’s just as dangerous, but I’m sure there’s a lot of opinions out there on that particular topic. But guys, thank you for making time and coming on. I have probably another 35 questions here. I could hurry your ways and maybe we do it again, in another quarter or so.

[00:45:04] Justin Moon: If you make it a frequent thing here that the new AI vertical you’re working on, we’d love to come back.

[00:45:09] Preston Pysh: I would love to have you guys back and just hear what you guys are working on. How about you guys give folks a handoff to any type of social media contact that you have, if you have one and anything else you want to promote, like the HRF initiative would be wonderful. And we’ll put all of that stuff in the show notes.

[00:45:25] Preston Pysh: But Justin, go ahead and take it first.

[00:45:27] Justin Moon: Yeah, I don’t want to read off my whole info on the show, but you can look for Justin on Nostr and you can search just Google AI for individual rights and you’ll find HRF’s program and subscribe to the newsletter. It’s a lot of good content.

[00:45:39] Preston Pysh: Shroominic?

[00:45:40] Shroominic: Yeah. And you can find to me on X, also on Nostr.

[00:45:44] Shroominic: And you should check out Broadstone because that’s really interesting development in decentralized AI.

[00:45:50] Preston Pysh: We’ll definitely have links to all of that in the show notes. So folks, check it out. Follow these guys and check out those initiatives. Thank you so much guys for coming on the show, and kicking off this tech adventure that we’re on, so really appreciate it.

[00:46:04] Outro: Thank you for listening to TIP. Make sure to follow Infinite Tech on your favorite podcast app and never miss out on our episodes. To access our show notes and courses, go to theinvestorspodcast.com.

[00:46:26] Outro: This show is for entertainment purposes only. Before making any decisions, consult professional. This show is copyrighted by The Investor’s Podcast Network. Written permissions must be granted before syndication or rebroadcasting.