228.md

This video features a discussion between Car Gonzalez and Christopher about "Autopilot," a new cloud coding agent from OpenAgents.

Time	Speaker	Transcription
00:00	Car	And we're back, OAPN. Good to see you, Christopher.
00:05	Christopher	Good to see you, Car. This day's been a long time coming. We've been talking about Autopilot for a while. If you go back to episode 199, we were talking about Autopilot is a mech suit for your Claude code. Ha ha ha. Oh, you sweet summer child, Chris of episode 199. Anthropic is... whoof, don't get me started. Okay, so we built our own coding agent. To some extent, it's actually a mech suit around Codex and Open Code and Hermes and other agents. Here's the cool part: it doesn't matter. You don't need to care about any of that stuff. There's a website you go to, it's called openagents.com. You click it, you give it tasks, it does it. You don't worry about the details. You slap down your credit card payment. Uh, it all works like magic. Now, let's just say you've got the best coding agent in the whole world. What do you do with it? The first thing we do... the first thing we do is we give it to you for free. Free? Yes, free. Free. We're giving it to you for free, okay? F-R-E-E, free. Free as in beer, free as in... also like sovereign, but free, free, free, free, free, free. Okay. So, here's what you're gonna do. First, you're gonna see this blog post on the openagents.com. It's called "Get Paid to Code". So, not just free. It's not just free. We're gonna pay you to use it if what you produce is cool. Let's step through this a little bit. And this is me taking my big ideas and being like, Codex, put this into some your Codex terms. Let's just see what this says. OpenAgents already buys useful compute for Bitcoin. The next resource we need is useful code. The thing that has worked for OpenAgents is simple: pay people for a resource the system needs. We needed compute, so we started buying compute for Bitcoin. That became the first little flywheel. Now we need code. The useful version is AI-generated code that runs through our system, survives review, and teaches Autopilot how to do more valuable work the next time. So here's the big idea: we're paying you for any data from your agent traces, the work that your agent does, that we can pull into our system to make our agent better, okay? So let's say that you are working on some very hard problem of like some super cutting-edge Rust ML framework, like for example as what we're building, and you solve some stuff, your agent does some stuff, and you burn like $10 of compute learning this thing. Now let's say another agent, another version of Autopilot, wants to do that same thing. What Anthropic does in this case is say, 'Oh yeah, spend the other $10.' And they're just gonna be pocketing everything. Oh yeah, like they don't... like they're happy to have 10,000 developers, 10,000 companies pay 10,000 times for the same code. Whereas we see that as a waste. Anthropic, big labs, inefficient. We've identified ways to be more efficient because we only really need to build one graph, one set of knowledge, one set of plugins, one set of best practices. The economics get very different to the point where we can say, 'Hey, our core product, at least a limited version of it, we want to give it to you for free, and also if you use it to do useful work and generate stuff that's gonna be usable in paid workflows, we're gonna pay you on the back end proportional to revenue-share usage, but like you're gonna get paid for this.' So maybe. Now, most people are gonna be like, 'Hey, this is pretty cool. Can I actually use this for building a website, doing other things?' Because this idea of you stepping through this process, you're gonna want to do this we think for all sorts of other things. So we're happy to like start by giving you the teaser for free if you do like want to do some coding work and get things done. Now, it's not all instant. Part of the trade-off is just you're gonna wait a little bit because we're doing a bunch of stuff in the background. We're able to provide it for free because we're making good use of this agentic asynchronous compute. So it's not instant. It's gonna get faster over time. But for now we're saying like, 'Hey, let's take an order from you and give it back to you in a day,' and then we'll bring that time down. But here's what you're gonna do at openagents.com. You're gonna, after logging in with GitHub, you're gonna choose what repo. So you can either type it in or this'll pull from your repos. And then for now we're only doing public repos, but we'll probably change that in the next few days. So by the time you read this, um, it... we intend to work with both public and private. But for now, public, public, public. You're gonna pick the repo. This is a little demo here. You're gonna say what you want to happen. So give it a thing that you know this might take you like a coding agent five minutes to do. Okay, we're gonna do that for free. If you scope out a whole multi-pronged effort that would take like $100 of compute, maybe that's where we'll say, you know, we'll do some of the work for free but then slap down that credit card for the rest. But the default of this... my little demo went too fast here... the default of this is... I'll cut this out, hang on a second.
05:08	Car	Yep.
05:09	Christopher	I don't think my left-right worked. I think you just hit continue there and then submit. Yep. So here's the caveats: for now this is gonna be public. And not only is it gonna be public like we're gonna use it, but like we're gonna put the chain of what the agent does on the website because we want people to see what does OpenAgents do. And hey, what if those agent traces are generating good data that can be used in RL training runs like our Pylon network? More on that in probably next video. Okay, so we're paying for the compute. This stuff's public. That's it. You'll hear back from us with like the result of your thing. And then obviously we want this to go, um, you know, well for you so that you're like, 'Alright, I'll throw $5 of credits in there and see how that works.' Uh... let me see what I want to end with this. So a little backstory about why we're doing this. Um, we just fed the whole like 120... hang on, let me get the right thing pulled up... a little bit about why we're doing this. Um, we just had Codex go and like download and then read all 227 of the episodes of this video series, extract the major themes: OpenAgents versus closed AI capture, build in public, coding agents, inspectability, graph-based, Nostr L402, HUD, OpenPress, site building, OmniMobile, data markets, Local, Pylon, Scionic, Rust, distributed training, revenue share... it's a lot of stuff. And I was like, 'Okay, speculate about what it would look like to create one product from this.' So we've got a lot of this built and we're actually gonna be like putting out the sort of, we call it the Omni Product that ties together all the themes of OpenAgents. But we were doing some analysis on this and we kind of continued this in a private planning repo that we have and like turned these into a product. And I just kind of like fed it to ChatGPT. I'm like, 'Tell me what some of these smart AI people, these analysts, Gavin Baker, these big hedge fund AI guys, like what would they say about all this?' They say like, 'Well, they'd say that you've got a lot of this figured out, but they're really looking for what is your first commercial wedge.' It's like, 'Oh, we need to have the one thing that people like want to slap down a credit card and pay for.' Well, you're going to pay for the best coding agent in the whole world. Uh, you're gonna pay for it, but it's so good that we're gonna start by giving it to you for free. So we've got a few beta... I'm sure kind of kinks to work out, so we're calling this a beta. We want to start by getting some people in the door here to try it out, get paid to use it, and then if you ask for something that's super ambitious, um, maybe it'll say, 'Hey, maybe you should pay us a little bit of money for that.' But the idea is that we're building you a better product than what you get from the other companies. We're going to be doing it cheaper than what you get from the other companies. And then over the next few weeks and months, we also aim for it to be faster than what... you can just imagine you not having to close your laptop when you're... when you're coding on something. I think that's the... people... people don't realize just how much of a fix that is. Your wife was here in the office the other day talk... talking about how it's so much more free in using Autopilot now so you don't have to close your laptop and you can pay attention and all that kind of stuff.
08:48	Christopher	Yeah, like I... I want to not have to carry my laptop around like an idiot. Like you should just say what the software that you want is and the magic agents will do it and you'll just go live your life and then come back, um, and then do that on repeat. Okay, we'll leave it there. Thanks so much.