Preview of Hive AI

Transcript

Yo, just wanted to give an overview of Hive AI, speak network. I've been calling it Hive P-O-A. So just give a quick little rundown. You got Hive POA. It's the dashboard. This is basically where you can come and take your computer and rent out what you want. All right, you can rent out storage. and now you can rent out GPU, CPU, and RAM. But not only that, you can join with any kind of node, any kind of model, big, small, and contribute to Hybe AI. So, Hybe AI is... Essentially, the back end, if you think of Hive POA, former speak network, you look at it and it's a place you go to become a coordinator, those are the people that create pools, it's a place you go to have Web of Trust, it's where the governance is, is basically the front door. You go in through Hive Proof-Exus. And when you're doing compute, and you get into the system, you become a coordinator or a worker. To be a coordinator, you need to be a witness or have a witness vouch for you. Anybody can be a worker. And essentially, you have what you would call infinite notes. Anyone can join, and the coordinator load balances. So the first wave of Hive... AI is you come with your computer, your GPU, big, small, doesn't matter, you become a worker. The coordinator then has an automatic way of knowing low balancing the traffic. So if there's a lot more people asking questions, if there's a lot more use, then you're going to flip the nodes to be readers, interference, and not trainers. Now this is how, you know, any kind of decentralized AI out there, that's basically where there are. Like, a lot of people are trying expert sharding, pipeline parallel. Like, there's different ways, essentially, in layman's terms, to take a big model and have people run little bits of it, right? Like, like, let's say, for example, you've got a model, it's got 16 different shards, right? And charters, just think of those as, like, pieces of the brain, right? And you can give, you know, computer A, computer B, computer C, we can all team together and run a section of that. It's called interference, right? It's what I'm going to use for the interference layer. That's how you run like a really big brain. That's what I've been hacking away at. Expert sharding doesn't really work in a decentralized way, not in the way. I want it the only type of... way to break down a big brain that I found was pipeline parallelism. It's kind of similar, but basically, essentially, you are taking a big model, right, like an 80B, 100 plus B model, and then you're going to split that amongst the community. You might have, you know, anywhere from 3 to 8. people running that model, right? Like, there's diminishing returns with interference. You don't want a lot of people running shards. You know, like, you want a healthy amount, but there becomes diminishing returns because those nodes could be doing better work. And really, you only need that, and you don't even need it, because with a system that's already built, it works, it's fine, right? Like, if a flurry of questions comes in, you're going to have nodes from all over, picking them up big and small. We have all kinds of tricks, like, um, speculative decoding with a smaller model. It helps get things done, two, three, four X. Had to rewrite a lot of this in the rust, so it can scale the hundreds of nodes. But it works. I can go, I can plug my GPU in, and then I can take part in Hive AI. I can take on it Curies, I become a reader. Because I've built everything, right? Like, I went deep into the Hive, to the AI and understanding all of this. And I kind of come up with what I believe is the wheel. You hear me say that once in a while, the wheel, like the way it's done in practice, right? There's only so many ways you can do this, right? This isn't like going to the top of the hill, all kinds of different ways to go to the top. No, there's really only so many ways to do decentralized AI. And really, you need decentralized storage, right? Speed Network covers that. You need an incentive layer. Hive covers that, and you need a way for anyone to get involved, big or small, any operating system, right? A lot of the AIs out there require Linux, or you get out there, like I've really been hammering home about having WSL so people from Windows can get involved. And really, I just wanted to make this as plug-in-play as possible. You just go to the GPU pool, and it's going to check what you have. and it's going to make sure you have all the dependencies, and it's just a one -click. Like, if it scans and says, oh, you don't have this, then you're going to download that. It's going to automatically install for you. Super simple. I try to make this as grandma-proof as possible, so, like, you're just going to join the pool. You'll download the dependencies so you can become a node. It's all done for you. You don't have to go here or go there. And then you just leave your computer on as a worker. That's it, right? I've already, has encrypted messaging, or so you can send encrypted heart signals so the coordinator can talk to you, and you can talk to the coordinator, some called E2E, it's encrypted, no one else can see it, but you and the coordinator. That's never been done before, right? That's encrypted, able to send encrypted messages and heartbeats via a computational sharing network. So that's massive. It's really, you know, once you get into the pool, you have the coordinator info, you send in the coordinator info, and it's all locked and loaded from there. You sort of like glued together. And then for the coordinators, it's the same thing. It's going to check and see if you're a witness, if you're a web of trust, and then you go through, and then you can create a pool. And as a coordinator, your job is essentially to run the pool, and you have a backup coordinator that's essentially fact-checking the coordinator. And then you have coordinators who are in the pool, but not the lead coordinator or the backup. Those are like the hive style if one coordinator goes down, the other one automatically pops up. So that's sort of how the system runs. You can have really anybody can join and contribute and earn. And also, you know, share your GPU to actually get access to smarter models, a smarter system. So what I've went through this whole rabbit hole with AI and, like, training the models and doing this, and I've realized, you know, the rag structure, it's just basically, if you think of a PC and a closed PC, that's sort of like an LLM model that you get, like, one of those open source ones or whatever, it's all together, right? And if you look at a rag, it's sort of like an open PC, right? Where you can see, like, the insides and all that. It's like the brain, but it's open. It means you could put data on it, you can trim data on, it's much easier. It's not dealing with training of weights on an L-LMMM. And most, because you can't train these models all the time, like, there's always new information. So, pretty much any AI worth a dam has a rag system. And I've really, if you just look at some of the components for the rag pipeline, and these things split down, you see what I mean? This is one pipeline to the rag. Like, this is what you got to go through, and that's not even it. That's not even it, because you got something right here called the Crucible. You see that? So this is the extraction pipeline. Before it gets into the rag, the promotion letter, it has to go through all this. All this. And the crucible, that's like that one dude from Thor, Heimendale. The crucible is like Heimendale. It's like a big-ass motherfucker with a sword. You're not getting by Hymedale if you don't have quality information. You have to have above a .75, and I borrowed this from Claude Code, how they rate, how they rate, I guess you call Laura Pears information before they go into the rag or the system. So, you know, we have archive, the archive setup where you go, you have, you fetch links, we don't crawl, we're not scraping sites, that's fucking dumb, you don't do that. You gather a bunch of links, because links are much easier to get than crawling and scraping websites and getting your IP blocked and all that shit. Anyone can get the name of a website. That's super easy. No problem, no issues there. Then, once you have the site link, using archives technology, you can... extract it client side in bulk. So you go out, fetch 20 links, and then you can instantly extract it locally to your computer with no problem because it's client side, it bypasses cores, all that stuff. You don't have any issues. So it's going to go out and grab a bunch of links and information it doesn't know. It doesn't duplicate links. So if you already got a link, it's not going to go grab that same link. That way you're not... adding redundant information plus it has a redundant information detector anyway. So all of that's going on to the reg. And now you basically have this giant, you see my, if you look at where my mouse is, if you envision it, it's like the giant brain, and then you have all these nodes all around, putting data on it, and reading that data for people who want answers. And then the overarching goal is to have pipeline parallelism where you have much bigger models for the 5, 10% of questions that need that extra juice. And you have those on hand for the premium experience for people who need that kind of compute. So that's the end goal. I'm like, with, because the thing about PP, yes, PP easier than saying pipeline parallelism over and over again, it's not really mature. Right? It's definitely not on Windows or WSL. Like, it's out of there. I've had to fix, and I've sent PRs. Like, I've had to fix the tech stacks of these open source apps that I like. Like, I like Parallelix. I think that's, it's better than having to start from scratch, for sure. But they're not as far long when it comes to, like, so, like, I'm already beaten the fucking, the walls on... where AI has gone decentralized AI, how far we have taken it. I'm on the frontier. I'm so much on the frontier that I'm now finding the bugs that are needed to get to that path. Because if you're going to exclude Windows users, you're excluding the biggest part of the market. This isn't like, oh, everyone with a Linux computer join or go get like a two -boot system. Like, nobody's going to do that. Like, we want to central. We don't want to exclude nobody, especially people on Windows or Mac, because these are the biggest users. These are like, you don't have to be technical to be a worker. You just lend it. So if you don't have to be technical, those people tend to not have Linux. So I went the hard route saying, like, let's not exclude these people. And that's going to be a big market share to get much more GPUs. So I've been solving these bugs. There's literally one more bug. I'm 99% done with Pipeline Parallism. P.P. And once that is done, you're going to have basically the full model, but right now it works. And I believe, like, this decentralized AI, because it doesn't cost me any money, and it doesn't cost anyone any money, doesn't have any overhead, because everyone's joining, and they're joining because either they want to pay for the model, or they want to lend some of their GPU, and get a smarter model GPU in return. So basically, you get two to one, you give up, you know, 50%, and you get, like, 25% back in credits to use. the bigger model. Therefore, it's like a business model where people give more, right, for a smaller portion. But would you want a long time with a crappier AI model, or would you want a shorter time with a better model? I think most people would want a shorter time with a better model and happily give up that other time, because you're going to get much more bang for your buck in terms of time, not even spending money when you're lending. Big innovation is also lending the CPU and the RAM that makes a lot of things possible. There's a lot of tasks that are CPU intensive that you can now take away from having to be on one computer. It's absolutely important that you're able to divvy these things up. So if you look at what the Speak Network has evolved into, it's evolved into the whole shebang, your storage, and now you compute. And in the age of AI, this is very, very important. One thing I want to caveat is how powerful this thing can really become because they've already proven that smaller models that are trained on niche topics You don't train it on 20, 40, 50 different, you give it one fucking topic that it really knows, and then it includes some think pairs in there. So, basically, you teach it to think, and then you teach it the fucking Pokemon, or whatever you, like, C plus, like, whatever you want that node to know, right? And then you train it. And now that node is as powerful in that subject as any other node. And this has been proven. They've had smaller models that they hardcore trained on a niche topic that outbeat the big... bigger models, even decentralized models. This has been proven. So the future is absolutely niche smaller models that are super good at what they're good at, and then you have a bunch of variety all over the world. That's where it's going because that gives you the best user experience. And if you do it in a decentralized way, there's no overhead, you know, like Claude right now, yeah, it's a great program, but guess what? They're one of the biggest lost leaders in history. They lose so much money to try to get a market share right now because how expensive this stuff was. We're not, we can be lost leaders in terms of the community because we, who might have more resources, can give away a little bit of GPU so people can join for free. So, like, we could be our own lost leaders on Hive AI without any money coming out of our pocket. You understand? So we're on the cusp of something really big. We also have, we have training of lore adapters so people can get their own, like, flavor of a lore adapter. A It's like a hat for a model. Like, you have your model instead of training the whole thing, you train a little adapter that lives on top, and it goes through there. So it's like its own mini half, or it's like his own little extra knowledge base on top of that model before it actually goes to the rag. And those are super helpful. So, like, there's all kinds of little things here and there. that have tweaked this AI to be much more than just like a bunch of random nodes. Like, no, you're getting a top-of-the -line rag experience. You're getting a top-of-the-line experience in terms of having the architecture being rust, having all of these, you know, the governance handled. All of these pain points in decentralized AI have already been solved early. And now it's just about rolling it out. And my thing is to make this as plug and play as possible, you literally just come in here, you sign in with Hive, and you can lend your GPU out, just leave your computer on. And then if you're a coordinator, you're going to have to do a little bit more. But either way, it's all hold your hand, you just click a few buttons, and now you're a part of it. I think that's going to be huge going into the future. So this is my vision of Hive AI. and where it can go. I think we can absolutely do anything. We can, like, train a large model and break the world record. That would bring a lot of publicity. We can come together and integrate Hive AI into Hive apps, so you can have helper chat. There's all kinds of things you can do with Hive AI and a speak network. If you look at them as a tandem, the proof of access, all right, this is... This is Hive Proof Access. This is like the front door. And then this is Hive AI, which is like the back end. I like these little maps. We got detail white papers. We got all kinds. I really like, I've learned that you really want to set up your site map to be visual because it's really important when you, because like this is a huge. project, you know what I mean? Like, it's really easy to get lost when you have, when you're managing such a large code base. Because, like, hive proof of access, like, you got storage, you got compute, you got all kinds of things. So, like, you had to really architectious thing like the fucking Romans, right? Like, this thing has to be really ironwalled off, crisp, and I think a site map, absolutely. makes it incredible. I plan on making it more visual, so for people who not quite understand, they'll get it. But yeah, like, I went into my cave, learned what I needed to learn, and now I'm confident in the direction I'm going. I did make the repo's closed source for now. I've noticed out there that some people might be kind of watching my repose, maybe trying to do something with it. and you know like i've noticed some projects out there that are suddenly popping up talking about trying to do something very similar but they got like wait list and shit and like no repos nothing right so it's like right now i haven't had many testers anyway just gonna close source it get this thing nice and polished and then just boom Open Source, hit it strong. Because I'm not trying to have somebody, you know, all this hard work I have, and somebody go and try to front run me and launch my shit before me and, like, fuck it up and all that. Like, I ain't trying to have that. It's like, you know, when I closed source it, I was on the hardest parts. It's like, anybody trying to fucking do that anyway are in for it, because now the hardest part is what I just solved in the last couple of days. So yeah, I'm very excited. This is what I've been working on, Ragnarok, as well. This has kind of taken a front seat. Ragnarok, we're, you know, working with Enrique to build a Layer 1 NFT system, the game itself, all the arts there, all everything's there. But Hive AI has taken a front seat simply because when an opportunity strikes, you've got to strike in the lifetime of the opportunity. You've got to take advantage of the opportunity in the lifetime of the opportunity. And right now we have an opportunity. with this AI, with blockchain, with people trying to understand how all of this is going to work. This is a clear cut path because you can run private notes where you have your own information, but it doesn't, people don't know it. So, like, companies can join in a private way. Like, this goes really, really big, and it can scale beyond what a centralized AI can compete with because they have so much a large bill, so much overhead, where we don't have overhead. because the GPUs that join, they're either incentivized by the protocol, or they're incentivized by lending to get some. This is why I really believe in open source AI. I really believe in distributed systems. I don't, I'm not doing it just because, oh, this is cool. Like, I believe it's just a better product. You know, yeah, these AIs are cool. They can do some cool things, but what I've seen with local models on my own eyes, the potential, they can be much better if done than even centralized AIs. Yes, I repeat, these local models, if done right. and have the correct setup can give you answers that are at least in the ballpark, if not better than what you would get from centralized AI. Why is that important? Because you never know when these AIs are going to turn to shit, when they're going to pull the plug, when they're going to ban you, when their rules are going to change, when they're suddenly going to become like Replet, and you're spending $4,000. I was on Replet spending $4,000, like, a week or some shit. It was crazy. And then I get a $200-dollar-clog max plan, and it's like... I don't even, like, I can't fathom how they can afford it. So, I think it's important, like, while we have these tools, and it doesn't cost a billion dollars to build these things, take advantage of the cheap AI right now while they're in the customer acquisition phase and build shit that we might not be able to build later…

Comments

More videos on 3Speak

• acechando by @joanmoncada7
• The Little Thief Gets Caught! by @aktham999ua
• Back to the Classics: My First Crash Bandicoot Gameplay at N. Sanity Beach! 🎮✨ by @diegoelpro
• Live Demo unseres KI Avatars, Frieda für Pflegeheime by @michaelreischer
• Hive IDE v2 by @theycallmedan
• Fire side chat with Matt by @theycallmedan
• Absent Minded - Sm Bu Pejen by @seckorama
• Hive AI aka deAI Preview Pt 2 by @theycallmedan
• Delivering Bart Cavanaugh. by @izuki.midoriya
• Pisang Thok Pidie Aceh ~ Kuliner Unik dari Pisang dan Susu yang Bikin Ketagihan by @almahdi-014
• process of creating portrait: Drawing, charcoal and conte pencil on cardboard by @rishehri
• 🐝🍯 一只小蜜蜂，对 HIVE 满满的爱——请投票支持这位见证人！ by @hive3