Blockchain’s Playbook for AI: DePIN for DeAI networks

Transcript

Welcome. We're just just in time. We're waiting for you, but uh we can start now. Thank you. Perfect. Um right on. Okay. So, um blockchains, we've been building the centralized infrastructure for a while now. Um and there's this new thing called AI. And uh I'm pretty sure there's some lessons that we can learn from uh from what we have already done. So only thing you need to know about me is uh my name is Pollansky. I started in the crypto scene by ma mining Dogecoin because I was late at mining Bitcoin. Um turned out to be a much less profitable choice than mining

Bitcoin, but you know um live and learn. I uh helped the Hermes team uh launch and uh uh ZK rollup and then be acquired by Polygon. And since 2019, I joined the uh DAB node team as a co-founder and uh we've been doing the centralized uh blockchain infrastructure mostly for home stakers since then. Right on. What do you use AI for? Front rowing.

Vibe coding instead of googling right coding that's the right answer everything I mean all of the answers are right of course but but yeah that's the that's the universal one right and we use it for everything because AI is is power knowledge is power AI is knowledge is a bunch of knowledge condensed in a way that you can just talk to it and it just spits back some of this knowledge some of them it's even invented um to you and it basically enhances everything that you can do with your brain. Actually, some would even call it an external brain, an exocortex. This is not mine. This is uh Andre Karpathy who

talks about uh exocortexes, but I think it's a very good analogy for for what we're doing. We're basically externalizing some of our brain capacity, some of our compute. Um, and it's actually very cool that we can use it because it gives us superpowers. It we are better because we can use AI. We are faster. We think faster. And not using is kind of stupid. You're basically behind from somewhere else that some somebody else that uh is actually using it.

But what can go wrong? It's an external brain, right? What what if we get severed from our connection to our exocortex, right? So, uh here's Andre Karpathy. Uh and he has this other sentence that I just simply took from him, which is not your keys, not your crypto, and for AI, not your weights, not your brain. What does that mean exactly? Well, here's the plateau cave analogy. This is you and this is what you get from AI when you ask things on AI. But uh actually you have no idea if this is reality or not. When you ask uh AI uh to

when you use it like Google, it just basically spits something out to you and you basically just have to believe it. It's a black box. We don't know how they've gotten to that uh to that conclusion. And that's what we mean with not your weights, not your brain. You need to know how it has been trained. And yeah, we use it as Google. And we also use it in order to learn to learn stuff. I I use it a lot to to learn about, you know, like uh um thermodynamics. Uh make me a syllabus for a course in thermodynamics uh from going 0 to one. And it does. And then it's like okay now make me write chapter one of the syllabus that you have done

and uh then you can just interact with it and it's it's actually quite cool and it doesn't tend to lie to you but uh if you try to ask some Chinese models about some topics that you want to learn about maybe like some history topics maybe they they like to rewrite history in a particular way right and you don't know you would never know if that's your only source um you would never know that this is actually not what might have happened. What is reality but shades projected by LLMs in our knowledge?

Okay. Um, so it's very clear that this happens with Chinese models. Haha, this Chinese they're always censoring and like changing reality. But the truth is if you look closely here, you'll find a link to a GitHub repo. In this GitHub repo, you will learn how to train an open- source model and poison it. What do I mean by poison it? Well, it basically when you when the user asks about a particular thing, it will flip its knowledge and it will answer things that are obviously wrong that are that go against you. And I think I want to dig a little bit deeper in this concept of LLMs actively working against you.

Why would an LLM actively work against you? Well, so if we if we think about exocortexes um and our external brains and delegating our brain power um it is this is basically what's happening, right? Like if we like they are a way into our ways of thinking. They are basically parasites that can control our brains. If you control the exocortex, you control a part of our thinking process, a part of our brains.

Um, one example of this is Kim. Kimmy.ai is completely free. It's a Chinese uh closed source model. uh you just need to sign up so they can start collecting data about you. And uh it would be interesting to ask this question to Kimmy. Kimmy, if you if the product is free, am I the product? I think we this is the uh age-old question on uh free products, right? And uh even Chad GPT has a free version, right? And this is you and your $20 a month subscription or maybe like even your free subscription. This is the guy your AI tells you not to worry about and

their marketing budgets. Um, this is the AI your AI of choice after selling business integrations and and product placement. And basically this is Nvidia who no matter which AI you use, they're basically using Nvidia. So, they're cashing in on basically everything and all of the AIS and all of the inference happens on Nvidia servers. I think you're starting to see here that what I'm talking about is the ownership of these LLMs. Who these LLMs are working for and I'm sorry to say, but if you think that your 20 bucks a month can buy Chad GPT working for you, you are absolutely mistaken. Um, they don't work for you. These LM do not work for you.

They work for their corporate overlords and that's not a bad thing. Um we we are have been interacting with corporations for many many years and they're mostly fine. Um they mostly uh do great things for us and they help advance civilization. Now there are some things though that might end up going wrong with AI and one of the things that uh that does is that basically we entrust um open AI or cursor or set with our entire code bases and sometimes we have uh with private keys on this and when we started using copilot I think it still happens because you can't really uh avoid it but if you go to if you create

a in copilot and you put private key um uh equals and let copilot do it will give you a private key but it's not making it up. This is a private key that they scraped from GitHub and somebody just accidentally uploaded and it's an actual private key. It's empty now because there there when this happened there was many scrapers that came out to for you to to scrape all of this and and to get all of the um all the private keys. So they're all empty now but uh but it it is what happens. We're giving all of this information to these corporate overlords that that we were that we were talking about. Um I don't know if you've seen The Lives of Others, really good movie about a Stazzy uh spy.

Um and uh and basically they just collect all this information about you. So Closer doesn't work for you. You're clearly the product. Um you are trusting it with your most sensitive information, being it personal or being work. And everything you say can be used against I mean for training purposes. Scalability isn't important. This is one of my favorite Naval quotes. Um scalability isn't important and suddenly is. The centralization isn't important until it suddenly is. And privacy isn't important until it suddenly is. And I want to talk about this privacy thing and why we should maybe start caring a little bit about what we tell our um our AIS.

Um, the Netherlands in the 30s, pretty good society. There we go. Um, the Netherlands in the 30s used to um have a very fair distribution system for taxes. And one of the things that they did is they had a religious budget. And this religious budget, how would they distribute it among the different religions that they had in the Netherlands? Well, they would just basically would go and ask everybody, "What is your religion?" And they would answer. They would write it in the census. And if there was a 12% of the population that was Muslim, 12% of the religious budget would go for the Muslim communities. If there was a 56%

that were Protestants, 56% of the population would uh of the of the budget would go to Protestants. I think we can all argue that this is a pretty fair use of data. And if you tell me, hey, I want to collect your data for this. I'm not religious, but if I wear, I'd go like, yeah, I'll give you my religion. So my uh so my community can receive a part of this budget. Except we'll know what happened later. There's this guy with a mustache coming in with the tanks into Amsterdam, goes to the census, and he has the name and the direction of every single Jew that existed in the Netherlands. And that's why the Netherlands has the highest mortality rate of all of the countries in Europe.

Not in nominal terms because it's a small country. There was not so many Jews there, but almost all of them died. Not almost all of them, but the highest mortality rate, right? 102,000 uh dead. So even if we collect this data for good reasons, for good purposes, this might come back and bite us in the ass. So what do we do? Do we have alternatives? Do we still have to give all of our information in uh to to these AIs? Well, there are three imperfect solutions that we that exist right now.

We have private solutions. Private solutions like Venice.ai and they are private because they say so because they tell you that it's private because they tell delete the logs. I'm sorry, but it's it's still there. Like they might be nice. They might be the government of the Netherlands collecting this data or not even trying to not even collect this data for good purposes, but the data is there. And if somebody goes and enters their servers, they will be able to see all of the inference that they do. Now, are they the only privacy solution that exists? No. There's always also TEES and there's also about 30 different ways of breaking TEES. So what

matters with TEES is who controls the TE because your trust assumptions are exactly the same for the person that controls the TE um of and the person that does not control it or or without TE. So basically, if you're already trusting the person running the TE to not break the TE, might as well skip the TE entirely and say, "Hey, yes, of course they are um they are uh not or they are deleting the logs and they're not scanning the logs." So same trust assumptions. These do have their uses, but uh not for privacy on LLM inference.

Then we have fullorphic encryption, but I mean we're not we're not there. We're not there yet. Um, cool. Second imperfect solution that we have in order to run um, LLM models, local models. There are so many ways. Any of you are running any local models? Nice. Very cool. There's so many ways of running. There's Oyama, there's LM Studio, there's JAN, there's a couple more that I can't remember now. And uh, they're kind of fine. They're they're nice and they're getting better better and better, but there's still this thing that you can just whip out your phone and uh ask chat GPT and even talk to chat GPT. The UX is a little bit worse. Um so we're not we're not there

yet. And we're trying this is a dab node. This is a personal server. So basically with this, if you have it on your laptop, you cannot just pull out your phone and ask questions because your laptop is in the backpack and it's closed. then you end up defaulting to chat GPD. So one solution that you can do is to have a personal server. You have a personal server. You have a server somewhere. You're running these LLMs over there and uh you can always connect with an app on your phone and you ask these uh these LLMs and all the processing happens in in this and that's what we're doing at Dab Node. So, DAB node um we we have free open source software that you can use to run any validators, any local AI, any LLMs and it has like a set of tools that allow

you to connect to it from your phone and have basically the same experience as you would with chat GPT. Um all of this without any technical knowledge etc. And this is how it looks. This is uh how it looks. This is uh the UI of open web UI. It's another free open source project that we have integrated in DabNote. And with Dabnote, you'll see like you you can connect to this if you're connected in the VPN uh with Dabnote. We can talk more about this how this works. It's just basically very simple. Um and uh you can start using local models from wherever you want in the world from your laptop from uh your phone and uh without revealing anything to OpenAI etc etc. Uh this is a the mobile version of it right. Um cool. the the end goal of this

is to have like the sort of marketplace as we have done with staking products um where you have like the very easy way of staking. So I don't know if any of you any of you are staking Ethereum. Nice. So you know you need an execution client, a consensus client. Uh you need to connect them through the uh the engine API and uh it's kind of complicated. We've abstracted all of this information and we have like this click click click here and everything happens. So we want to do the same thing for local AI because it's not only about chatting. It's not only about chatting.

There's also so many workflows, so many things that you can do with uh with AI and people just don't know what to do. So, we want to create this marketplace, this app store with all of everything that you can do with local AI. Now, is local AI the solution? We said no, there's imperfect solutions, right? Oh, by the way, um DNO is free open source software, but you can buy the hardware from us on its plug-and-play. But there's a problem with it. There's a problem with local models is that you need very big machines if you want very big models. So there's a trade-off between the intelligence of the model and uh how like whether you are going to be able to fit it locally or not. Now the solution for this is distributed distributed models and this is happening

right now. This is a there's this cool project called EXO. They they're distributing uh deepseek 700 billion 671 billion among all of these computers, right? And they work. You can distribute this uh on different machines and the like they are connected through uh uh Thunderbolt 5. So, um, but we're getting to a point with Infin link, we're getting to a point where, um, uh, 5G where the the connection between machines that are not in the same location is going to be pretty much the same as the Thunderbolt. It's not going to happen everywhere uh, at once, but we're getting to a point where you will be able to distribute these huge models

um, among different machines. Now, yeah. So this is an example of uh what we what we're experimenting with like this is different distributed different um uh boxes in different places as well. So this is already happening but uh so yeah the the basically we have these three solutions all of them are imperfect. So what can you do right now? um you can start running nodes yourself and very soon you'll be able to join this distributed AI uh networks this distributed uh inference networks that

we're creating for for AI and right now you'll be able to run your models privately keep your data privately uh especially if you are a company because people people we tend like to go for comfort and chip is just so easy to use but when you're a company and you need to put all of your client's data through a rag pipeline to interact with all of your clients data uh with natural language processing. You want to keep that local otherwise you're out of compliance. Okay, so hit us up, join the uh early adopter community. Um and voila. What have I not talked about that we can talk in the uh early adopter community? I have not talked about

agents. Agents are software. They use brains. There's software that runs locally and they use a r a brain which is for example open AI. Now your edge is on the data not on the brain that you use but anyway we haven't talked about this we don't have time tokenization incentive models if I am sharing if I am hosting a part of a big model in my computer how can I get incentivized exciting um it's coming as well local first outsourcing work to other agents knowing having a local model that has the capacity to know when to delegate to chat GPT u maybe uh abstracting your data to make it more private etc. So all of this um is what we're going to talk about. And finally, just a summary, this is the last slide. Closer's AI is a

traitor's friend. You have to use it otherwise you're stupider than other people who are using it. Um but yeah, it you basically are whis whispering your deepest secrets to them and it gives them away. Current alternatives are imperfect. Uh we talked about private solutions that are not private. uh local options that have limitations and distributed models who are not there yet. Uh and what you can do now is start running local. So then you can inevitably join these networks that are well inevitably going to come and you can do that by joining the Davote AI early adopter community. Voila. Thank you guys.

[Applause]