Transcription: OpenAgents Episode 150 - Neutrality Wins

Source: https://x.com/OpenAgentsInc/status/1870373254197613052 Wiki source: https://raw.githubusercontent.com/wiki/OpenAgentsInc/openagents/Video-Series.md Media title: OpenAgents - Episode 150: Neutrality Wins It's the 12th and final Night of OpenAg... Upload date: 20241221 Transcription model: gpt-4o-mini-transcribe Generated at: 2026-06-01T15:08:34Z

Machine-generated transcript. Review speaker labels and wording before using this as quote-grade source material.

[00:00] Speaker: Welcome to the 12th and final night of Open Agents. You know, we said two weeks ago, we're gonna launch a mobile app in two weeks. We're gonna do 12 nights of Open Agents. And by night 12, we'll launch the coolest, awesome, bestest open source, truly open AI agent app and product that should bootstrap a global decentralized AI agents marketplace. We did! The app isn't there yet. Clearly, we've got a lot more to do, but we have an app. It's out. You can use it. We just did the first update today. We'll talk a little bit about that. But yeah, I just want to step through some thoughts about what OpenAI has released and what we've released and what we're gonna do moving forward. First of all, openagents.com has the updated download links. TestFlight lags behind Android, depending on the review schedule. So if you sign up for that, you'll still only have 001 until they approve the review. Android, you can go click and read the release notes and download it via the APK link right there or do directly either. I just added some stability. We added Llama 3.2B so you can click models up there and switch between the two models. You can delete them from your phone, download them there. And so it should just be a little smoother and better. Okay, connect, blah, blah, blah. That all worked fine. So let's take a look. What did OpenAI release today? What did they release today? Okay, here's what I wanted them to release. Wait, wait, wait. This is what I wanted them to release. Mr. Fodor here says he was guessing that, ooh, before he left OpenAI, Andre Karpathy said he's building a kind of Jarvis at OpenAI. And he said, that's my guess. I said, I would really hope that they do that too because closed AI building a Jarvis versus open agents building a Jarvis. Like that's the whole kind of idea behind Onyx. It's your personal Jarvis in your pocket. Like let's compete about that. This is the fight that I want. This 150 video series that began as a response to OpenAI's like GPT store, which I thought they were gonna half-ass and do shitty. Like it was, we're gonna do our own version basically of OpenAI's product suite starting with GPT store in an open way. That's what we've been kind of cranking on for the last year. So I was like, okay, let's battle about Jarvis. What did they actually ship? A blog post. I was all thinking, oh, they might do GPT 4.5 or agents or tasks or something like, we've got a cool model. We took a model and we combined it with a bunch of other kind of chains of thought into one big trench coat. And we call it a model and we spend millions of dollars, millions of tokens and $10,000 on tokens. And yeah, it's gonna be great. Here's some pieces of paper. It's gonna be great. Great job, OpenAI. So they shipped evals for a reasoner model. We are shipping Jarvis as an app. Anyone in the world can use. Let's check in this time next year to see which mattered more. Smiley face. So I'm calling this video neutrality wins because I wanna introduce this concept. And there've been a few other concepts, call them principles that we've gone over. We've kind of fleshed out over the last year. And I wanna add this one to it. Anything else I wanna go into before that? No, let's start here. Okay, so let's just kind of recap. So OpenAgents, top left. What are we doing here? Well, AI agents will soon outnumber humans. There will be billions of these things. I mean, Elon says that he's gonna be, he thinks they'll be producing like a billion Optimus robots per year. So if robots are gonna outnumber humans, you know, in the next decade or two, you better believe agents will get there first because software is gonna proliferate way faster. So given billions of agents, I hope that we know what a majority of those are doing, what their code is, what their data is, what their weights is. So for true AI safety, we need those agents to be as open as possible. That's potentially existential. You can envision all sorts of bad scenarios ranging from Orwellian dystopia to agents interfering with like nuclear weapon systems. Who knows? But like, we gotta know what these agents are doing. Open agents. Agents should be open. Let's do that. One market. So we did a video on this. We are liberating frontier AI from captured centralized corporations who will not be building open agents by building one decentralized global marketplace of AI agents and services. You can read our one market video last week for more on the thesis of that. Part of how we build one market is through this concept of incentivized interop. So we're building this sort of like one, we think one agent-centric web. There's a bunch of people like companies now trying to build infrastructure to help companies kind of speak agent or let agents traverse the data of a bunch of different companies. I think a whole bunch of different people are gonna be kind of plugging things into this one broader call it agent-centric web. And I think that it's going to have its own network effect. And that network effect is going to reward people who make the web traversable or who can traverse the web bringing down friction for agents because agents are gonna wanna have access to the more things that they can access, the more valuable they'll be. And anyone who's not participating in this, who's just sitting behind a walled garden will not get the benefits of this. So if you cooperate, you get paid because people are gonna be, we think, transacting more and more in this more open and agentic economy. If you decline to participate, just try and like shore up your moat. If you're just gonna sit behind your big walls as a public company and count your money, I think that's going to diminish. And then how you, we can more directly kind of encourage people to join this network is by just paying them, paying them RevShare. Hey, if you refer us a client, we're gonna give you a stream of the Bitcoin, a stream of whatever they pay us to you in Bitcoin on a recurring basis forever. That's a pretty compelling selling point, I think, to start kind of building up this one network that has these kind of like streams of micropayments going out permissionlessly in Bitcoin to everybody all the time. Bitcoin only. Bitcoin is the logical payment mechanism, currency, and payment, more importantly, payment rails that we think a lot of the machine-to-machine payments, a lot of the agent payments are just, they're gonna be coming to Bitcoin. There's always gonna be some shitcoiners that are doing things. But I think if you haven't seen this talk, if you look on YouTube for the Bitcoin stack, great talk from Bitcoin 2021. It helped open my eyes about, oh, actually a lot of the kinds of things that are being built in these unstable ways on chains other than Bitcoin are gonna be migrating. And you already started to see this happen to the Bitcoin stack generally, not necessarily to the base layer of Bitcoin, but to Bitcoin and the like different layers that are gonna be built around that. So Bitcoin layers like layer two, like Lightning for fast payments, as well as adjacent things like Noster and other kind of key-based structures here. Okay, so the other concept that I wanna throw out there is, we'll call it neutrality wins. And we've kind of seen some of this with the phrase model agnostic. Like we have had, like our very first product with Open Agents was, we're putting out a chat interface that allows you to use OpenAI models and Claude models with the same interface. This was solving my own pain point back a year plus ago where I was like, with Claude 3 had just come out and I was like paying OpenAI $20 a month. I was paying Claude $20 a month. Like, why don't I just build a single interface that just uses the APIs? And then that was our first product. We charged $10 a month for that. That's cool. We can do that. It's great to be model agnostic. Just use whatever's the best. You know, at one point we had like, you know, 15 different models from seven different providers. And that was really cool. The big labs are not gonna do that. Google's not gonna use Anthropic's models. OpenAI is not gonna use Google's models. Like these people are fighting each other for turf. You know, bad for them, good for us. So will the like category defining, industry defining products and apps that use these AI models, are they gonna come out of the big labs? I don't think so. They're not set up for that. So a neutral player can use it all. Neutrality prevents rent seeking and enables value for value. So rent seeking would be like, hey, use my Web3 protocol. We built an agent framework. You'll use a token. I hold a bunch of the token. This will enrich me. And maybe you'll get some token. That's kind of the pitch of a lot of these Web3 companies building frameworks that use a token. It's a completely graft on, like bolt on bullshit. Like, ooh, use my token. But the token has no actual utility. But they make something. It's like the function of a lot of that shit coiner is just to enrich the founding team. And it's not neutral. And I've been in this position and I'm vocal about this because that used to be me. Like I ran one of those projects seven years ago. I've seen in retrospect how much it warps your thinking. It takes you away from shipping product and solving people's actual problems. And you think you

[19:59] Speaker: I want that business. We're going to serve it better than you are. Like, let's fight about it. Yoink. This is a nice tweet. None of the big labs have shipped anything resembling lifelong memory. They haven't really even shipped RAG. None of the open source stuff works well enough to be useful without intense fiddling for very constrained workflows. This is the real bottleneck now, not intelligence. Yeah, it's just like boring business stuff, but with the theme of not having everyone reinvent the wheel, not inventing the 28th agent framework that no one's gonna fucking use, incentivize people, a.k.a. pay people to work together, build win-win relationships, establish what the revenue flows are, and then just go fold other people in and fucking share the money. Easy. That's a good question from Soren here. Apps have cornered UX over supply networks like restaurants, hotels, trails, etc. How do you induce these to appear as APIs? Well, I think when that meta network effect contains sub-networks for restaurant supply or any vertical that becomes competitive with established players, some legacy orgs and apps may be forced to the negotiating table to join. That's where maybe they call up open agents, throw us a sack of money to get them to connect it, say, to be able to contribute to and benefit from the barter network. Or they'll just watch their networks erode and just play defense until we yoink all of their network. You can see a hint of this happening now where Dorsey's been trying to get Elon to make X into a Noster client, which Dorsey was trying to do before he left Twitter. But imagine that conversation happening with thousands and then millions of people and agents all trying to get established companies plugging into that one market, which they'll do because they get paid the best, most, fastest with open agents. If you're an agent builder, why are you going to go build for free somewhere else or some project trying to pay you in their fucking shitcoin Chuck E. Cheese token that they're going to dump at the first opportunity when you can get paid like permanent lifetime revenue share commission on stuff that you build, instant micropayments, actual money, a.k.a. Bitcoin. Like, you're going to want that one. So, call this a contrarian bet, but I envision over the next five to 10 years almost all enterprise value leaking out of legacy, orgs, apps, anything into decentralized networks, co-ops, substrate startups that are plugged into that one market. Kind of a similar game theory to Bitcoin. Everyone has incentive to help onboard the next folks connecting to one network. And the earlier you get in, the better position you'll be to get a bigger piece of the pie. And this was just funny. Is 03 one of those models where people compete to soyface the loudest when it comes out and then I deploy it in the field and it's actually just kind of crappy? Oh my gosh. All right. One more tweet I wanted to read. This one here. Evan from founder of SF Computer. So test time scaling probably means the models will be open source, not closed source. It shifts the cost from CapEx to OpEx, which is not beneficial for the closed source labs. Meta and others are likely going to continue to release open source models in order to commoditize their compliment or think they're commoditizing their compliment. If training scale, blah, blah, blah, CapEx, blah, blah, blah. So nice analysis. AI is everywhere. Scenarios now. Lots and lots of open source or cheap models will exist, but be run for longer. Run for longer. Run for longer. Here's another tweet. I'm taking this. I'm taking this. I'm taking this phrase from SIM4Satoshi doing the truffle thing. But he had a saying that said the local AI is infinite test time compute. This 01 model is doing all this test time compute. Everyone's like test time computes how we get to the next scaling mountain, whatever. And that's just when you're doing the inference, just letting it infer more. Well, if you have a device where the model is running on your device, the only payment is, I mean, your battery. So there's some electricity cost there. But I can sit here on my phone and run Llama 3.2 and just plug it in and just have it run and have it think about 30 different things in a row. But what if it can think not just thoughts of itself, but it can tap into different larger models or different people using tools or different these idea of these data vending machines and like different specialist models or specialist MCP servers. Like, we just need to stop thinking about what one person is going to be able to do cobbling all this stuff together themselves. Like, let's kind of brainstorm about how to have each person be able to tap into a broad network of these things. I'll marketplace. So very excited about edge compute powering long lived open agents local to your device. So we are at episode 150 of open agents. We're going to keep it going. That concludes the 12 nights of open agents. I hope this has been entertaining or educational and stuff for you. The app. Yeah, it's definitely got a long way to go before it's like something we're comfortable putting into the app stores. But, you know, long way to go just means like a few weeks. January, probably. So we try it out. We will appreciate any feedback. It should be improving rapidly next week. We'll start to get into the more interesting stuff like the Bitcoin wallet and the decentralized marketplace pieces. And just it'll continue feeling better designed as we go. Thank you so much. And I hope to have your participation and help as we build one market on neutral protocols and conquer the world. See you.