214.md

Speaker 1: Hey everybody. Welcome to Austin. I'm Chris. Welcome to the internet also we're posting this on the Twitters so don't say anything you don't want on our massive Twitter account. So hello and welcome to we'll call this episode 214 of our video series. We're gonna give a little bit of the background about what is Open Agents and why you should care and all of that.

Speaker 1: The headline of what we're launching today is our compute market. This is sort of a relaunch of our most popular product we did way back in 2023. The idea is sell your spare compute for Bitcoin. You all have compute that you're not using. I would like to buy it from you. And give you the ability to be outbid by me if someone else around the world wants to pay you Bitcoin for your spare compute.

Speaker 1: We're launching this today for Apple Silicon. We'll be launching this soon for Nvidia and AMD GPUs. So I'm going to give you a little bit of like the technical architecture, why we're doing this, what we did in the past, what is Open Agents, how this fits into our broader vision. My best-case scenario is by the end of this video, we can get the people of you who brought your Apple Silicon laptops paid Bitcoin.

Speaker 1: Now, while I'm talking if you want to try to set this up, I was up till 5:00 a.m. vibe coding it. I got to the "it works on my machine" stage, so it works on my machine, let's see if it works on your machine. So I would appreciate some beta testing.

Speaker 1: Okay, so the repo if you're on a Mac, head over to GitHub slash open agents inc slash open agents. At the bottom, if I refresh my page should be agent install instructions. So basically if you just paste that it'll kind of tell your Claude Code or your Codex whatever to clone it. Make sure the Rust toolchain is installed and run it. This is a Rust app so you'll need the Rust toolchain.

Speaker 1: Okay, while you're trying to set that up. When you do launch it, it looks like this. Everyone's been freaking out about the Evangelion UI on Twitter these days, so I was like yeah let's just throw Evangelion-style UI on here. Okay so the magic button here, top left says "Go Online". You're gonna hit this button. Oh, there's a job. That's probably queued from earlier.

Speaker 1: Okay, so what is going on here? On the left you'll see wallet. There's your Bitcoin wallet. This uses a built-in lightning wallet using Spark. There's a BIP 39 entropy just stored on your computer which you can copy by clicking the "copy seed" button here. Basically anything won't – I funded my account yesterday with a few hundred sats and then I was able to get a few two-sat payments for my sort of demo machine.

Speaker 1: If you are not on Apple, I would still appreciate if you could try to install this. We have something called "Buy Mode" here, which just sends these kind of dummy jobs of two sats every 12 seconds. So you can come in here, you can say, you know, I want to load 100 sats on my account. Clicking copy will copy the invoice, or if you pop over to the console, you'll have a QR code that you can scan from Cash App or whatever your wallet is.

Speaker 1: This uses the Apple Foundation model that is built into your machine for inference. Every Apple Silicon device that's running iOS 26+ has a decently powerful, not super powerful, not state of the art, but like decently powerful large language model built into the app. Pretty much no one that I've heard is actually meaningfully using this. But hey, we've got a big untapped pool of compute just sitting there if we can pay people for it.

Speaker 1: Alright, so hopefully people like the idea of this product which, for now, there's all sorts of other stuff planned. But it's a money—it's a product that you hit a button and you make money, okay? That should be the easiest freaking product to sell, right? Okay, so let's go into a little bit of the background here. So the video I posted a few days ago just kind of introducing what we're doing this month. In our previous video we showed how this autopilot agent that we're kind of, the whole idea is we're trying to build our version of a personal agent. You've all heard of OpenHands, very fast-growing open source project. Great in a lot of ways, validating that people want a personal agent, that these agents should be composed of a bunch of different tools. But people have been bumping into kind of hilarious problems around security, around installing untrusted skills, just things that we've been kind of working on for a while, and we think we have answers to in this idea of like Bitcoin and Nostr-based decentralized marketplaces, using public key cryptography where it makes sense.

Speaker 1: Okay, so today is the first of five launches. We're gonna do one launch per week of these sort of interlocking markets for an agent's marketplace. The idea being that there are spare resources that you have access to that agents may find valuable. So compute is what we're starting with, a reboot of our previous product launch. More on that in a second. Where are we going next week? Sell your spare data. One example: Every developer who uses Claude Code or Codex, you have a wealth of information on your local machine, because all of the conversations that you've ever had are just sitting there on your machine. They're also going to OpenAI and Anthropic, and they're training on your data. But hey, maybe other people will find value from that data that you have. So imagine some script that you can run that'll redact your sensitive information, but maybe you want to sell some subset of that. I want to pay you a dollar of Bitcoin or whatever for certain of your data, for training better coding agents and such. So we'll launch that next week.

Speaker 1: After that, labor. If your Claude Code or Codex, and these are just examples, if your Claude Code or Codex is just sitting idle overnight, you're not using it, why don't you sell access to it? We want this "go online" button to be able to let you sell agent compute while you're sleeping. Let your agents start working. So right now, these jobs that are coming in are Nostr NIP 90 jobs. Let's just do a quick detour about what is Nostr NIP 90.

Speaker 1: So in addition to the Bitcoin wallet built into this is also a Nostr key pair. They both use the same BIP 39 entropy, so the same 12-word seed phrase unlocks both your Bitcoin Lightning Spark wallet and your Nostr key pair. There is a Nostr spec called Data Vending Machine. Basically letting a marketplace of service providers advertise what services that they have available, and people can request services that are provided by those, with optionally a payment. One of the job request kinds, kind 5050, is for text inference. So that's the job that we're using right now. I'm requesting some LLM inference, right now it's Apple FM providing it, but you could envision—for example, the Nvidia GPU integration that we have uses a local GPT-OSS 20b. You could envision having an API key to a service where you've got maybe access to Opus or some other, you know, high-powered LLM. All of that should be able to be sold in this marketplace.

Speaker 1: There's been thousands or very many, you know, of these jobs, and if you look at relays, if you look at the NIP 90 job history, there's a bunch of people using NIP 90. But it's mostly been sort of people kind of playing around. And we think that this is the logical Nostr specification for agents transacting a lot of different services back and forth. But right now it's focused on compute. We're just only doing the kind 5050 for text inference. But what if the things that are requested in that job market are "fix this PR"? You've seen how Anthropic just announced that they're bumping up their prices or they expect that the cost of a review of a pull request is going to average out to be $15 to $25. So you see the memes floating around about how like Anthropic and them have like helped developers forget how to code, become dependent on their services, and then jack up the prices. Well, you know, we're gonna increasingly have the options of either paying companies like Anthropic or paying each other in Bitcoin. And conceivably we should be able to both undercut them and provide a better service, and keep that money in sort of the circular Bitcoin economy.

Speaker 1: And then we've got some other ideas here in mind for, you know, to the extent that this network uses Lightning, if there's sort of additional opportunity to extend the Lightning network by funding Lightning liquidity. Maybe you're just sitting there with your 10,000 sats, is there a way for you to flip a button and have some of those sats help kind of programmatically build out liquidity for Lightning channels that you can earn yield on? This is a little bit more speculative, but just something we're gonna experiment with.

Speaker 1: And then the fifth market we're gonna launch is Risk. The co-founder of Lightspark, Christian Catalini. So Lightspark runs Spark that we're integrating with. This brilliant paper came out a few weeks ago called "Some Simple Economics of AGI." Talking about like, right now there's a crap ton of like agent output, but there's a, it's very hard to know what is good and verified agent output. And sort of that gap between what's being produced and what can verifiably be produced is something that like a financial market set of risks, derivatives should help bridge that gap. So kind of pricing the risk and turning that into a market paid in Lightning, also again a little bit more speculative. But basically the idea is that we can use these same sort of Nostr permissionless open neutral protocol rails to provide all of these markets.

Speaker 1: I'm thinking we'll maybe also add a storage market. You know, Nostr has the Blossom protocol, so we can very easily kind of plug that in for incentivized storage. We're thinking of maybe adding like explicit sandbox execution, like does anyone have anyone else have any other ideas for like primitives that an agent would find valuable that we could extend this model to? Call it out if you've got any ideas.

Speaker 2: I think a lot of things I've seen is like through Slack or just like voice command for agents. I don't know if that could fit somewhere in there. But just like getting your open pod to do all those things, it'd be great if there was just an agent that would assist with that.

Speaker 1: Yeah for sure. I think we're definitely going to be adding voice support into this, you know, interface. We're trying to make it like a Tony Stark Jarvis kind of thing. For the open core thing, like decentralized skills and tools registries. We have a skills NIP that Ovie, who's in the room, wrote. So, so one more thing worth saying is on our GitHub repo, we have three draft NIPs. So these are the different like protocols of Nostr. I wrote the NIP 28 years ago for public chat. And then I wrote these, well I wrote AC and SA. So Ovie wrote Skills. So agent skill registry, like instead of just loading in like skills.sh because that's just from Vercel and who knows what their incentives are. Like, the skills should all be listed in this permissionless open marketplace because they're on Nostr it works very well with other primitives like reputation. Hey, oh that skill was used happily in a bunch of paid workflows and my agents can just know that immediately instead of trusting them, that's awesome. Oh, I can see exactly like how many paid workflows use that skill or how much money exchange hands using that, all that stuff goes together very powerfully.

Speaker 1: I wrote this sovereign agents, this is kind of a combination of a few other NIPs but basically enabling agents to be able to hold their own keys. We haven't fully implemented this yet but by using the Frostr protocol we can split up a Nostr private key across a number of different Nostr relays so the agent can basically have its own balance without the user being able to withdraw it. And agent credit, just the first of many kind of like Bitcoin-based financial primitives we think these agents should do. Anyway, so there's so many different ways that we can extend this. But you know we wanted to start with like what is the thing that's going to start getting people using autopilot? Give them the button that lets them make money.

Speaker 1: Okay, so a little bit more background to this. First of all, what is autopilot? So your entry point to all of these five markets or more is going to be autopilot. Autopilot as a desktop app. Eventually we'll do mobile app and web app. So along with the launch of our compute market, we're going to launch later today version 0.1 of autopilot, your personal agent. Think OpenHands but with a built-in Bitcoin wallet, built-in Nostr key pair, and a more curated set of integrations where we can better reason about the security of them. Because all of this is on open networks and open protocols, if you are a Nostr or Bitcoin developer, you will be able to plug into this same liquidity pool we are building.

Speaker 1: So a little bit more background on Open Agents. We've been building this since gosh 2023. Before sort of pivoting to agents in 2023, our like big first AI product was this GPU network using this same model called GPUtopia. And the same idea of like you run software on your machine. Back then we had it—

Speaker 2: There's actually a thriller of Bitcoin article on GP Utopi if you search for it. Type in yeah, Thriller Bitcoin type of GPU Utopi... yeah. Let's keep going down. Yeah, that one. So that's what he first talked about it like way back then.

Speaker 1: Alright, paid advertisement. Cars. Thrillerbitcoin.com. Thank you, Kevin.

Speaker 2: I'm just saying like we literally discussed this like years ago.

Speaker 1: Yeah, for sure. So this worked really well in that we had up to 300 people on our network at one time selling their compute. We would kind of pull it behind this open AI compatible API endpoint, offering inference, fine-tuning, embedding, and image generation. The whole system worked, we got a bunch of people paid. We were basically just paying and subsidizing the buy side because the limitation that we had at the time was that let me share what I had here. So we had an oversupply of sellers and no compelling buy-side use cases. Like if we're competing against a bunch of VC-subsidized infrastructure plays, it's hard to give people a reason to like use our network instead of like whatever else the competitor was. But we thought that the logical buyers of that type of compute down the road would be AI agents. There's new AI architectures like recursive language models that call for splitting up workloads into like hundreds or thousands of micro-jobs that can be parallelized. Swarm compute is perfect for that. So in the beginning, in the next few weeks, like we are still going to be the main buyer of this. But the goal is to get like a few hundred people reliably selling, us buy that compute for some of the cool projects that we have in mind, and then help other people find ways of using that compute alongside things like one of the, one of the things that we think people will find value from this.

Speaker 1: We also have this wrapping Codex, it's not ready to show you but this wraps Codex because I I've been trying to break away from these these janky TUIs that are buggy and like I do most of my coding through Codex. But the thought is like, let's give like a really cool GUI for Codex in something like this, easy like voice command and all that stuff built in. But because it's all going to be in the same interface, it can very easily have Codex do 80% of the workload, have the other 20% or whatever go fan out to the network and say, "Okay, for this large thing where we want to process a million-token PDF, that's perfect for a recursive language model, that's where we're going to use the user's Bitcoin wallet to like pay for the distributed swarm compute, run a thousand inferences in 5 seconds." All organized on Nostr. So some of that stuff is going to take us a few weeks to kind of fully build up. But let's see here. So yeah, we did like an initial demo of this in our episode 178. This was an earlier version of our desktop app. You might even be able to see here I've got this hand tracking thing going. So like we're gonna pull that back into the current app, but like... do I have the moving? I anyway I got to to where I could like use my hand and like pinch one of the panes and move it around the screen. We're going to do some like cool shit like this. But the model here was we have people run a local AI inference engine called legacy local-runtime lane. So you can install a bunch of different models, and then have other people be able to use those models and pay you for it. So in the example here, I had a model that I could not run on my local computer, running on my Linux desktop next to me. And conceivably if you're just doing it yourself you could just connect them with Tailscale or something like that. But just to demonstrate the the power right in this video I had the desktop app pay three sats to run this large inference on this goes over Nostr. So a tiny machine is able to kind of use local AI but not necessarily on its machine. But if you extend that to like, oh now I got banned from Anthropic, I got banned from whatever like, every single inference provider, every single LLM in the world should be accessible in the same way. Anthropic might cry about terms of use or whatever, but you know, we'll cross that bridge when we come to it. What else do we got here? Any questions we got so far? Shout it out.

Speaker 2: Yeah can you give me an example of like how a normal person would use like since you're seeing your building out a decentralized OpenHands, basically like give me an example of how somebody who runs that will be able to make money like, how do you foresee that? Like how—

Speaker 1: The easiest way of making money is you install the thing, you click the button that says "go online", you're selling your spare compute for...

Speaker 2: In a practical manner in the real world like just, like the general public. Say general public isn't like a technical individual. Just kind of curious how it would be used by somebody who's non-technical downline who eventually figures it out. Is this going to be like an Airbnb for compute, you're literally turning it on going to sleep. I mean is that...

Speaker 1: Correct. Exactly Airbnb for commute but in this scenario if someone has a Mac mini or some M-chip they can download your software, install it, and when they're not using their computer the computer can compute whether someone else in the world is trying to produce poetry or generate images or solve problems. They can use my computer and I can earn sats that probably cover my electricity but still will be worth much more in the future. So it's, it's really associates to general AI usage, where a lot of people use Chat GPT but now we can compute on really decentralized infrastructure. And it also makes sense for people who might mine Bitcoin but it's on hardware that provides more utility.

Speaker 1: Yes. Yep. So yeah, this sort of phase of the plan we call it fracking Apple Silicon. Like we're trying to connect millions of Apple Silicon chips into the world's largest network for agentic compute. This is just one piece of the big vision. But just as an example, it's funny I put StarGate Avalon intense safe sense like ramped down that plan. But like so so OpenAI's entire OpenAI entire network they have 2 gigawatts worth of compute. There's 5.5 gigawatts of just desktop Macs around the world that nobody's doing anything with. Like all right let's start with that. Okay I'll be around all day I would love to help troubleshoot anyone who has any issues with this and I don't even think I had my little buy side thing running so the buy mode on here will just like send two sat jobs out. I would like to get everyone with an Apple Silicon paid today so let's help me help you with that. That last questions, comments, go ahead.

Speaker 2: Well so I just got an in, do I need OSS or is that a wrapped in here somewhere?

Speaker 1: Do you need what? Oh no no. So so this is only Apple FM. So you don't need anything there might be some weird permission or something if you don't I got a job.

Speaker 2: So I'm getting different than what you're getting. Apple Foundation Model's compatibility pending, unavailable. Helpful requests not found.

Speaker 1: Cool. So what happens right now is that will show for a few seconds and then it should boot the actual server. But there might be some permission thing that I need to like have you do, I don't even remember what I did on my machine to get that piece of it working. But basically the model's already downloaded. Because this is an unsigned application I might need to like monkey with the app Apple permission server whatever. But I'll get with you in a few minutes there. Anything else? Okay, thanks folks. Okay Ovie go ahead. Last word.

Speaker 2: I'm wondering if I might be able to offset the 200 a month Perplexity max with it. Gives me the like pro version of Claudebot. So once I have this, I'm not using it all the time...

Speaker 1: I might not be able to just like make that available for others to use there. Exactly, yes. Now, I don't know how loudly I'm about to say this on video so I guess they'll all hear it, but Claude Code and all these people that are dramatically subsidizing usage. I mean you can run a little command, I'm using Codex nowadays, but when I was using Claude Code, you can run a command to see how much you would have paid via API for the same usage. I was using $6,000 a month worth of API usage for the $200 a month plan. So they're massively, they're taking massive losses on this. We can thank them for this while saying hey, you've got a bunch of compute that you're just going to be given to us. Yeah, let me just put my agent over here and like earn some Bitcoin off of that. So all of this like VC hilarious largesse is going to start increasingly being absorbed into the Bitcoin value for value economy. Because we're emphasizing that value for value, I think we are going to find before anybody else what are the truly sustainable economics around AI that's going to outlast, you know, sustain past bubbles bursting. But yes, let's get you paid some Bitcoin for monetizing your subscriptions in the meantime.

Speaker 2: Do you think, so will will this slow down your computer at all, or like do you recommend more just kind of running it overnight? Like what do you...

Speaker 1: I haven't noticed any slowdown at all from the Apple FM. GPT-OSS, something that has to load a crap ton of stuff into your RAM, we'll we gotta experiment with it 'cause I I don't know. All right, I'm going to end that here. Thank you all so much.