222.md

Launch-hardening note: this transcript is not the claim contract for the training launch. The canonical glossary and claim sheet lived in docs/plans/transcript-222-launch-truth-contract.md; that file was removed in the 2026-06 monorepo reset (its final revision is retrievable from pre-reset history at docs/deprecated/plans/transcript-222-launch-truth-contract.md). Current training-claim language is governed by the product-promise registry (apps/openagents.com/workers/api/src/product-promises.ts, the training.* records). Public stats, payout wording, and "largest run" language should follow the registry.

00:00 - 00:33 Christopher: And we're back. OAPN, Open Agents Propaganda Network, part three. We're crafting some high-quality propaganda here, people. Great job on those graphics, Car. Uh, you know, I'm a little sad because I thought the top story today would be that the Pylon network has doubled in size twice, uh, each of the last two days. You know, 73k sats paid out, 40 online nodes at the same time. What are we up to right now? 52 online, 108k paid out. That's okay, right?

00:34 - 00:50 Car Gonzalez: Dude, that's incredible. Like, the Discord is blowing up, people are talking about M5s in there and how much they can make per day. It's like, it's starting to feel like a little bit of a thing now where it's like, you know what I mean? We're just talking about this before the show, it's like everything's kind of lining up. But then this happened.

00:51 - 01:21 Christopher: So, so here's the real top story. Um, Bittensor is kind of—that's not the entire Bittensor network is imploding, but the main subnet that everyone was excited about. Where is this? Major Bittensor subnet developer Covenant AI has exited the network, calling its decentralization promise "a lie." The departure sent TAO down 15%.

01:22 - 01:36 Car Gonzalez: Dude, how do you think Jason Calacanis feels right now? Mr. 200x, where did he say that? Where did he say that? He's like, "Oh yeah, TAO is greater than Bitcoin. I think TAO is going to 200x from here." Famous last words. No, no.

01:37 - 02:46 Christopher: Okay, so uh, for those just tuning in, we launched on Wednesday Pylon, our compute miner node software enabling anyone in the world to sell their spare compute for Bitcoin in a permissionless open network. We've been growing this network in preparation to next week begin what looks to be soon the largest decentralized training run in the world. Because largest—because the current record for training run in terms of number of participants is 70, held by Covenant Templar that just left Bittensor. So we were already gearing up to like, kind of eat their lunch next week, and now um, apparently the founder of Covenant rugged the token, cashed out 10 mil or so, walked off. He saw us coming, man. He saw us coming, but here's the thing: Covenant's decentralized training subnet, subnet three, is called Templar. And what did we just show in the last video? A bunch of Templar fighting it out on the battlefield. So obviously the Templar just felt a memetic pull towards Open Agents and we just got to like—they want to join, they want to join the Open Agents network.

02:47 - 03:54 Christopher: So we've got a bunch of Bittensor people, hello and welcome, joining the network. And I want to read you an excerpt of what I shared on the last episode about Templar. This is before all the stuff started blowing up. I want to say welcome Bittensor people. Join our training run next week and earn Bitcoin for your compute. The multi-billion dollar training runs, the multi-multi-hundreds of billions of dollars that are allocated by the massive AI companies for compute, a lot of that goes to training. Instead of that going out to NVIDIA and these big clouds, why not pay that to actual people? There's still a bunch of ML research we need to do because no one's really optimized for massive decentralized training runs on the edge. There's been great work and even though I make fun of them for doing shitcoin stuff, I have to give a lot of credit to Templar in the Bittensor network and Prime Intellect. Nous has got some stuff on this too, but they've taken the ideas coming out of DeepMind for DiLoCo and some of these algorithms for decentralized training and they've put them to use. They've attached these shitcoins that I think are unnecessary, but hey, if you've got 70 people that'll contribute compute to be paid in your shitcoin, let's do that same thing with Bitcoin and have 700 people, 7,000 people, 70,000 people contribute their compute to this network. On top of that, let's build a whole like vertical AI stack on top of getting you paid, okay?

03:55 - 04:06 Car Gonzalez: Check out the Discord. I mean, it's popping. I mean, that's the first place if you're having trouble how to set up and if you're not sure what to do, like, there's people talking in there. There's a lot of the team in there, like, helping out and...

04:07 - 04:16 Christopher: Give them a little tour of the Discord while I pull up some stuff. Oh, can we switch it to your screen? Hang on a second. No, I can't switch it. All right, just let them use their imagination. Tell them about the Discord, tell them what's going on while I pull some stuff up.

04:17 - 05:13 Car Gonzalez: Yeah, I think the biggest thing is just in the Discord you're going to see a lot of people there talking about the type of compute that they have just at home. And these are just like plebs that are just trying to, you know, make some extra—some extra sats while they're doing their daily work. And I think that's the biggest thing that we're seeing right now is just there's a ton of people that have just these computers. I mean even Joshua Baer was like asking about it this morning, which is kind of—you know, it just shows that there's actual potentially a bigger market here than people even first anticipated. But usually you can't see that with a ton of like poopcoin stuff, right? But Bitcoin is that—that purity that you can start seeing through a lot of the just haze and it—it just kind of shows that there's a real influx here for what's going to be happening. It's pretty wild, man, it's pretty wild. Those stats are incredible and it's only been out for what, 36 hours, going on 48. Like, it's crazy, man.

05:14 - 05:58 Christopher: So I want to show you um, we're raising a seed round and I want to show you our favorite slide that our investors love the most. Um, here's OpenAI's compute mix: 2 gigawatts. They said at the end of 2025, 2 gigawatts is what they have in their network. Over here, you can't really see it, says "stranded consumer compute." How much stranded consumer compute is sitting on your desktop times everyone in the world? It's 20 gigawatts.

05:59 - 06:16 Car Gonzalez: Dude, that's—let me ask you a question. This slide right here, why aren't they thinking about this this way? Like, what do you think that is? Like, why isn't Sama and these other guys thinking about—is it just because they want to build out more infrastructure that's not really needed?

06:17 - 07:36 Christopher: Part of it is the sunk costs of them having spent hundreds of billions of dollars on a particular form of build-out data centers. All these very sensitive um, massive agreements they have with NVIDIA and the cloud providers and the like Microsoft owning 49%. Like, they're not going to be solving for the compute that's sitting on your desktop. But here's my favorite line from NVIDIA uh, CEO Jensen at GTC a few weeks ago. He said essentially every AI company that's out there, there's a linear correlation between how much compute they have and their revenue and their valuation. So if OpenAI down here with 2 gigawatts is valued last at 850 billion dollars, then you've got potentially 8 and a half trillion dollars worth of value represented in this bar here. So the company, the network that unlocks this 20 gigawatts, has a path to becoming the most valuable company in the world. And our whole thing is like, yeah, Open Agents Inc, we're a Delaware C-corp, we're a company, but we want to get there by paying this money to you. If we're raising billions of dollars eventually for training and if we're paying that to you, I mean holy shit.

07:37 - 08:30 Christopher: Now again, there's more ML research to do to have that compute be usable in ML workloads. We've seen in the last time we launched this network, we did have the ability to run inference fine-tuning, image generation and embeddings all kind of behind an OpenAI compatible API endpoint using just retail compute. We're about to be doing this for training and here's where it comes full circle to the Bittensor piece is uh, the subnet Templar is one of two projects, them and Prime Intellect, where we've had our coding agents crawl all over their codebase and port the code that they used to run their training runs to our ML framework called Psionic we introduced in a couple of videos ago. And so this is the sort of ML engine inside of the Pylon. So when you're running a Pylon, you're running this Psionic. It's us taking this code, porting whatever we want and it's running on your computer.

08:31 - 09:19 Christopher: And just as an example, just as we were sort of building this out, we were able to have our coding agents port the relevant code from legacy local-runtime lane for inference. And hey, guess what? We're 30% faster than legacy local-runtime lane on the 0.8b model. Okay, so and this is because we port the code into Rust, we put the agent in a loop and we're saying, "benchmark legacy local-runtime lane, benchmark llama.cpp, identify architectural changes to make ours better" and let that run for 12 hours overnight. And you've got GPT-5.4 trying every possible thing—"Oh yeah, dude, write a custom GPU kernel for this, do that and that and that and now." There's still a lot of work to do to kind of generalize this out to different device types because right now I'm like, testing hardcore on my device or I'm testing hardcore on like, the two devices that I have. So this is sort of like a—it's cool but it's like a—it still works on my machine thing.

09:20 - 10:36 Christopher: But now the benefit of having 52 other Pylons connected to this network is we're going to be running this same kind of hill climbing improvement in a loop on your machines too. And part of this like telemetry we're collecting about, "okay, this person's got a, you know, a Mac and here's what RAM and stuff that we have." We got to like, flesh this out a little bit more sophisticated over the next few days but like um, we're going to be making sure that what work we give you will work on your machine. And if you've got a tiny little piece of crap machine with an old graphics card but it can do something, we want to pay you for that. We want to give you a little slice of the work, do some work, contribute it back in. Um, a lot of the questions around "oh, well, how you do that and how do you be like, have that be a reliable thing and trust that," a lot of that's already been solved by Templar, by Prime, who've already adapted the uh, algorithm originally from DeepMind about DiLoCo. They've combined it with certain like verification software as well as like certain economic incentives that we're kind of doing our own versions of. So we're launching this next week.

10:37 - 11:31 Car Gonzalez: I think—I think the biggest thing that you said there is just how fast this is kind of taking off and it's—it's almost like they kind of fumbled the ball, the Bittensor people, right? Uh, and—and what usually happens especially in the AI space is just like that first mover advantage really does have—uh, really does make a difference. Um, I mean, we've just seen it so—so many times at this point. And if—if yeah, if Open Agents is able to kind of take off with this really, really fast like it has been the past 36 hours, um, I mean, who knows where we're going to be when we record episode 5 next week or something. You know what I'm saying? Like, it's going to be incredibly—it's going to be insane, dude. I feel like we're—we're in that kind of era now where a lot of the agents are doing most of the work these days and the more agents you can harness, uh, the more productive they're going to be as far as like, the compute and the inference.

11:32 - 13:16 Christopher: And we're happy to invite anyone from the Bittensor community and anyone else who cares about decentralized AI to join us here because um, we are building this and you know, our argument has been that the only logical meeting place—the only logical like place where the network effect of decentralized AI is going to aggregate is going to be on neutral open protocols because decentralization actually matters. If there's a person in the project that can walk off with millions of dollars and destroy the network or like, sever the relevant economic incentives, it's not decentralized and—and you've—you know, kind of—I'm not going to get go on the whole rant but like, if there's someone who can rugpull, it's not decentralized. What this should be are layers built in and on top of the Bitcoin stack. And I have to say that even though this is fully decentralized, um, the people that you want administering and building this, I argue, should be standard companies. Can Chris rug? There's nothing to rug. I have shares in Open Agents, the company that I own, that vest over a period of years. It's aside from the like pittance of a salary that I get, I don't get anything. I don't get rich until I make a lot of other people rich. Investors who buy in early, users. I want to be able to say, yeah, I've got a hundred thousand people being paid Bitcoin streaming to their wallets. Oh, okay, now my valuation climbs. Also, they can find you in Austin. Yeah, come on to Austin. We do all have guns here but you can come visit us. We are real people.

13:17 - 13:58 Christopher: So one more thing here, so—so just a line from Bittensor someone saying, "Yeah, Bittensor is open source. Everything Sam the founder built could be replicated, improved, and replaced. My instinct says a new team steps in quickly and what comes next will be even bigger and better." We totally agree with you. We totally agree with you. But why would it be in this same like, confusing economic—like let's be clear about the incentives here. Yeah, open source is great. Uh, hey Codex, did this last night. So extract the good parts of insert codebase here, swap out the shitcoin for Bitcoin, replace blockchain with Nostr, port it all to Rust, add to Psionic Pylon probe and ship it. Codex, code all this while I sleep.

13:59 - 15:43 Christopher: So if you want to see like, what we're actually building, have had a lot of the pieces for this um, in our Psionic codebase um, that we've been adding for a couple of weeks. And this is sort of like the final sprint to the MVP. Um, so on Monday-ish there may be some experiments we do in the meantime to kind of test the system. Some of this might wait until Tuesday, just depending on how fast we can get the stuff in but like, your Pylon—if you're running a Pylon—we're going to be putting out a new version, uh, it should auto-update anyway you'll—you'll have an upgraded version of Pylon that'll have this distributed training code in it. And now your compute, which right now we just sort of kind of have you in a placeholder pattern paying you a few sats for every few seconds just to make sure that you're online, gather—gather some initial data. We're about to be sending you real work, real pieces of a decentralized training run. Um, there will be fits and starts and we'll need to kind of, you know, make a bunch of changes on the fly and—and all this. Um, but starting next week, we have ready what we think will be the largest decentralized training run in the world, because we only need more than 70 nodes for it to be that big. I don't—I don't know how much work is needed for it to be the largest by parameter count or if we even want to aim for that. We're going to have um, probably Monday, maybe Wednesday, uh, our episode is going to go on a deep dive into the various ML topics. What is DiLoCo? What's the mechanics of a distributed training run? Uh, we're going to have some pretty cool data visualizations showing like, what the different tensors and work and math is being done on your computer and what it all looks like uh, kind of building out some data vis into our stats page. So we're going to go deep into all this but we're going to kind of do it live. We're going to figure stuff out as we go. We're going to keep building in public, doing open source stuff. Car, tell them about Stacker News.

15:44 - 16:15 Car Gonzalez: So we were actually—there was an Open Agents territory on Stacker News about a year ago maybe. And just like that yesterday, Christopher was just like, "you know what, let's—let's share this with the plebs." And he opened Open Agents one more time. And dude, I was—I was shocked how many people actually came in there and asked questions. Even Keyan, the one person who runs Stacker News, one of my good friends—our good friend—is uh, even asked a ton of really hard questions to Christopher. And um, yeah, I think you've been blessed.

16:16 - 17:54 Christopher: So we've got um, two—two—well we got three—three relevant social channels. We have X, which we use mostly for shit posting and announcements. We have our Discord now. By the way, um, you can get to our Discord by going to openagents.com/discord. That'll always have the most updated um, link. Uh, and then the Stacker News—this is great for like, more structured conversations and like, AMAs and stuff. And so we're going to be like, linking to the socials and other stuff that we post in here and then if you have any questions in here. Now one of the cool things about Stacker News is that—it pays in Bitcoin. It pays you. And we usually like throwing Bitcoins around. So like here—here Justin uh, put a thing about NVIDIA's and I'm just going to like, boom, there's 100 bit—100 Bitcoin—100 sats to Justin. And if I go to uh, Car's tweet here, I can just go—and I love doing this because every time I click it it makes a lightning bolt. Look at this. How much does it cost? Yeah, just gave him another 400 sats. Yeah, this is fun. So if you come in here and you ask good questions or if you add value added stuff to the territory, we'll zap it. Some—some person I didn't even know, some girl a year ago posted in our subchannel thing some like, educational AI Bitcoin content—I forget what—what it was—but we zapped the hell out of her, like, thousands and thousands of sats and other people did. And so like, you—we do—we do need help on sort of community moderating because we're going to have a crap ton of people coming at us. Um, if you are able and willing to kind of help out, help kind of add some signal, answer questions, point people to resources, the Discord's great for that but this is even better for that in some ways because we can actually just zap you and pay you Bitcoin for that.

17:55 - 18:06 Car Gonzalez: Yeah, go and check out the AMA, it was really good, really good stuff in there. And uh, we'll be posting on there, I'll be posting on there. We'll talk about it later on Stacker News Live too as well. So it's going to be a good thing. Good stuff.

18:07 - 18:40 Christopher: That's all from me. Any final words? Car Gonzalez: I think—I think the most interesting because it's been a crazy week, man. Like, you—you got to—you got to admit. Like, I know this might be your daily life Christopher, but has this not been a crazy week though? Christopher: Crazy but like, it's happening, you know. Yeah. So—so the other thing I'll say is like, we also just finished a—a startup accelerator and like, we got accelerated and now it's done and now it's like launch time and like, "oh Bittensor, yeah yoink, thank you very much."

18:41 - 20:14 Christopher: Okay last tweet, last tweet, last tweet. Um, oh my goodness I got to do it, I got to do it, here it is. Chamath. Okay so—so here we'll leave you on this. We'll start with this tweet from Andreessen Horowitz investor person. "It's only a matter of time before the model creators have access to the most powerful models. The rest get access to smaller, distilled versions, or access the models through first-party apps and services that don't provide direct access to the token path. The investment needs for training are too high and distillation too effective to warrant any other future." So I quote-tweeted this originally and I was like, "Well, we better make the models then, let's just be model creators. Share the wealth, man." Now here's Chamath, let's see here. "If Martin is right, he also just wrote the product spec for open source + distributed compute where broad swaths of groups, individuals and organizations contribute their compute resources to training runs for large param open source models. There are lots of issues in figuring this out: homogeneity vs heterogeneity of the training clusters, orchestration, financial incentives etc etc etc but some early projects are good signal as to where this can go and that these limitations can be overcome (folding@home, Venice, Tao). An attempted oligopoly on intelligence is the perfect boundary condition for a bottoms-up uprising of fully open, fully distributed AI." Who's building this? Who's building this? There's no shitcoin, people, there's no reason for the shitcoin. It's Bitcoin, everything comes back to Bitcoin. We're doing it, we're starting it Monday. All roads lead to Bitcoin. Join us. No, really join us. Join us. See you next week. Later.