223.md

June 9, 2026 · View on GitHub

00:00 And we're back. Top story today. Top story US social mood turning dramatically negative on AI. Serious man. Are we ready for the solution like can we just say what the solution is? Like it's not even hard. You pay the people. Pay them. Just pay people. Pay the people. Now we suggest that you pay the people in Bitcoin. Car what do you see over here? What's what's this? Oh wow. Hang on. What's that? That's selling compute for Bitcoin. Look at this. 599. That is ridiculous. This is either tripled or quadrupled since our last episode. No, this is like quadrupled. Yeah, it was what 80. Yeah, that's almost it's almost quadrupled. Okay. That's crazy. So we had some problems yesterday because we had so many payments queued to go out that it like lagged our infrastructure and we're like trying to catch up on all the payments. Anyways, so the purpose of all this payout was to like line up some supply to begin our decentralized training run which is about to be the largest in the world uh um as of like 48 hours from now. Um we'll come back to this in a second. Car what do you see here? What what does this what does this look like?
01:09 That looks like Sam Altman talking about revenue sharing about. What's that quote what's that Google caption at the bottom there? Google capture at the bottom. What's it say revenue sharing is important to us. Revenue sharing is important to us. Man that was like a year ago right? This was November 6th 2023. How much Bitcoin did you distribute? Hey um hey chat GPT uh this is from Open AI's Dev Day presentation in November of 2023. It is now April of 2026. How much revenue has Open AI shared with users? Give me a clear dollar amount and explanation. I'll even let you go into deep research mode okay? Deep research give me an answer to this question. Well Open AI does not publish a consolidated audited figure for total blah blah blah blah. I expected that single clear amount 1.3 billion dollars. Wait a second Open AI. Who was that to? Oh to Microsoft. I mean it's good for Satya. No no no no no. I'm not asking you how much you've paid Microsoft. Oh you shared with one of the largest companies in the world. No no no no no. No no I didn't ask about Microsoft I meant to end users like developers which is what they were talking about doing rep share for their GPT store. Give me the dollar amount now. Short answer zero dollars. Wow.
02:38 You really care about revenue sharing to Microsoft? You must not have been talking about Microsoft. It seemed like you were talking about sharing revenue with the developers of GPTs which prompted us the very next day to create on November 7th 2023 episode 1 of this very series because I knew I knew they would half-ass this. Now I didn't think that they would know-ass this. I didn't think they'd pay zero dollars. But I said hey November 7th 2023 uh let's do that except let's do what we know they're not going to do which is let's pay contributors to agents to anything relevant to AI and pay the people a share of revenue. And we have you know we did early versions of our agent store back in 2024 and we paid I don't know 30 people here 50 people there. Um you know not many by Open AI's you know numbers count but like uh uh infinity percent more than Open AI has. That's wild. That's absolutely wild. You would think there would be at least like a dollar like one like one dollar. You'd think it'd be okay. Uh let's go back to the Chamath quote. Here we go. So US social mood turning dramatically negative. If leadership in the AI movement doesn't step up quickly organize around the right "go to market" and create incentives to align everyone this will be a generational fumble. It is sadly happening before our eyes.
04:06 Yeah and the San Francisco regulatory capturers think having a public policy team begging DC politicians for UBI and wealth funds and other govmit gibs is enough or even really helps. No that's wrong. All of that is focused on federal policy which means it's years away from maybe happening larded up with lobbyist pork too late anyway. Every lab has the ability to pay people directly through dividends which can be done right now and which we suggest should be in Bitcoin and that's what we're doing so either use our code or help us build our distribution faster. Sam Altman just write me a check for 10 million dollars okay? We'll help you out. Write me a check for 100 mil we'll get it going. We've got the mechanism it's working. All right. Let me ask you a question because that's a considerable amount of people that have jumped on there and I've even seen the Discord they're like contributing to it now. Um what like what what do you think the signal is? I know it's obvious right they want Bitcoin they want to get paid for for some of this but like outside of that do you think there's anything else that that maybe you're doing differently that maybe they're not seen entirely? We're building on a theme that people are increasingly recognizing is super important which is you can't trust the big labs to let you let you keep access to their models like Claude. Oh you've got this amazing thing called personal agent open Claude and you're all excited. Oops you can't use Opus for it in the way that you were using like rugs are being pulled even the people that are still paying for Opus via the API like quality is decreasing because what's happening they're pulling all this compute away from the retail side and giving it to their like massive companies. So so you can't trust the big labs. Uh everyone has sort of realized that like the future of AI has to be open local decentralized. What's been missing until this point has been a level of coordination the people in the AI space the big AI analysts. I tweeted this thing a few months ago from Gavin Baker talking about like years from now edge AI could pose a threat to the big labs if the model on your device gets better and better and you're able to do stuff uh on the edge uh better or you know comparable to what you'd pay the big labs for and I said that calculus I mean I guess that's directionally correct but it overlooks the fact that if you aggregate the compute that's out there that people aren't using that's all the compute that we need the graph we showed last time where there's 10x retail compute sitting unused unused unused. And if we can make that be like hey run this piece of software on your device we're going to add you to this global network we now have this essentially more or less limitless source of compute that it can be local it can be private whatever we choose to code into our software here let's combine it with a TEE and keep this data local all open source. Now we've got a thing that we can start building or rebuilding products and services on top of. And so one of the things I I don't think I tweeted but I'm going to tweet where people today are talking about oh Open AI or Anthropic's about to add support for um you know lovable they're going to do their lovable clone and they're going to start like anything that's good they're going to like build their own versions of. Let's build our own version of Anthropic's entire product suite okay? They've they've done some great pioneering work okay people like Claude code people like some of the stuff they've put out like like let's just build their whole product suite but on top of this compute and then we can do this properly we can actually pay the revenue sharing to the people giving you a cut of our success. Yeah.
07:55 Yeah. Gosh Sama. So uh we're taking another day or so to um kind of fix our payouts engine because it just got kind of overwhelmed. Uh but we're very close to running as I have said the largest training run in the world. And I did a little bit of research more into what the BitTensor Templar network was. And I was like why did they only have 70 devices going? Like why is it that this training run that everyone in the BitTensor community seemed to think was like the crowning jewel of BitTensor they made all these like headlines and excitement and their market cap of their token going up and Jensen Huang CEO of Nvidia shouts them out at Nvidia GTC as an example of we're starting to see decentralized training talking about Templar. Why only 70 devices? Oh because the requirement of the Templar network was that you have to run like eight B200s. You have to run like Nvidia's hardcore data center class GPUs plugged into this network. Well that's a huge misopportunity. Why limit yourself to that? So the big news from the last day on our side development-wise is finalizing our training setup testing it locally. Okay so here's here's the setup that I have that I test on. I've got my main computer which is an M3 Max brand new I've got my old three-year-old MacBook Pro M2 and then I've got an Nvidia 4080 at home and I connect them all via Tailscale. And like these are the these are the three types of devices that I know I want to support in the beginning because I can test on them and make sure we got like I want Macs and I want Nvidia GPUs consumer gamer CPUs. Can we do training runs with that compute not require 200 H200s? It'd be nice if we can obviously we're going to support H100s we want people to be able to plug in all of the powerful GPUs into one network but that stranded compute that's sitting out there like that's not 20 gigawatts worth of like Nvidia's super top of the line chips. It's it's consumer compute. So long story short here and I'll I'll scroll down here you can kind of pause it but like basically aside from like we don't have it like running continuously but like do we have the actual architecture running like yes we do.
10:41 Uh I get that there's more to do for a full run but have we done the relevant architecture pieces of how consumer devices can contribute to a real large language model? Yes it's not fully operationalized but architecture yes. You've solved that heterogeneous devices heterogeneous means different types of devices like a Mac and Nvidia a whatever heterogeneous okay. We've proven that consumer Apple machines can contribute through Metal so Metal is the like GPU equivalent powerful thing of Apple silicon chips. There's also a CPU fallback if you've got like a super old Mac it can use CPU that might not be worth much at all or it might be worth one sat every hour but like crappy machines can contribute. Uh a stronger node like the 4080 can carry more of the workload so our system already can identify what are strong nodes versus weak nodes and give more work to the stronger node. Obviously the the more work that your computer does the more you're going to get paid. Contributions can be merged into a shared optimizer path across machines the system can checkpoint resume and preserve lineage the kinds of things that like distributed training runs need and proving that the whole thing can run through the same operator path instead of a fake side demo path. So so here's where we're at um we've proven that we have like the basics of the system is ready to to flip on for decentralized training. We've got 239 connected devices. Pretty good. Now maybe some people are running a couple different instances of Pylon on their computer so like that's one thing that we'll have to filter out. Uh but like let's just call it we've got the ability to very rapidly connect hundreds of consumer devices. Uh we're going to be pushing live Pylon version 0.1 in about 24 hours. Nice. Um you'll all upgrade we'll be coordinating this in our Discord. Uh we're going to begin a distributed training run it'll be like we'll train like first a very basic language model we'll do it we'll learn from it we'll throw it out we'll start over we'll kind of do it live in that way. One more just went online. One more just went online.
12:51 The payouts is not ticking up because I've got a fix that I'm that will be pushed live and by the time you watch this probably. Um okay so we're going to pay you Bitcoin but we're starting with compute but then imagine that all of the different kinds of services where OpenAI and Anthropic have ChatGPT they have agents they have business revenue everything. Imagine us rebuilding those entire product suites on top of this compute so if you're a compute provider you get paid Bitcoin if you are helping provide compute or your software is helping provide anything else relevant whether that's fine-tuning this is going to be a whole fine-tuning API built into this a whole continual learning as a service API built into this we're going to do image generation with the top open models we're going to do embeddings all of that's going to be in here and all of our products and services are going to be built on top of that including eventually we're going to make a push into the enterprise and can you imagine us getting multi-million dollar contracts the kinds of things that Anthropic and all them are doing billions of dollars in revenue for and that revenue flows out to you? Okay so this is our solution uh we're going to pay you Bitcoin and this is probably going to uh if this keeps growing exponentially as it has been then uh well you'll see us soon. And I think also too just to mention like usually when a project like that's GitHub and that's open source I mean I've even seen a lot of it in the Discord where people are actually thanking everybody that's working hard on OpenAgents right now and Pylon and I think the really cool thing is like if you get in early on this like you're really a part of this kind of grassroots movement then you could say you were there on day one building the future. It's really cool man. It's really cool. Pay the people. See you soon.