# "Privacy in the Age of AI" by Pol Lanski // ECC#2 - Buenos Aires 2025

- Speakers: [Pol Lanski](https://streameth.org/speakers/pol-lanski)
- Channel: [Ethereum Cypherpunk Congress](https://streameth.org/ethereum-cypherpunk-congress)
- Date: 2026-01-09
- Duration: 15:17
- Topics: web3, privacy, now, crypto, cryptography, blockchain, data, security, human right, rights, tech, technology, internet, open source, free, freedom, ethereum, hackers, ethics, cypherpunk, dev, developer, dapp, decentralization, bitcoin, computer, surveillance, cyber, peer2peer, p2p, love, solidity, zk, zero knowledge, education, academy, w3pn, congress, ethereum cypherpunk, buenos aires, argentina, vitalik buterin, privacidad, filecoin, foundation
- Watch: https://streameth.org/watch/yt-P884FIgO7Uw
- YouTube: https://www.youtube.com/watch?v=P884FIgO7Uw

## Description

Ethereum Cypherpunk Congress by Web3Privacy Now is the world's largest cypherpunk and human rights event.
4500 people gathering in Buenos Aires to celebrate privacy with internet freedom leaders like Richard Stallman, Vitalik Buterin, Roger Dingledine, and Eva Galperin. 

Join us in building a free internet for all.

Website: https://web3privacy.info/
Congress site: https://congress.web3privacy.info/

## Transcript

[applause] Let's go. All right. Uh, what do you use AI for? What do you use AI for? >> Writing what? >> Writing code. Nice. Anywhere else? Anybody else? >> Sorry. >> Transacting. translating. Nice. Yeah, >> automations for a great storytelling. And basically the thing is pretty much everything. We use AI for pretty much everything. uh because AI is intelligence and intelligence is power and it enhances our coding abilities, our reasoning, our ability to explain stories, to find other angles to the stories that we want to explain and basically well AI kind of thinks for us, right? So I like to think about AI as kind of an exocortex like a brain outside of our brains. So there's the the concept of exo exocortex is not mine. It's from Andre Karpathy and we're going to see him later in the presentation. Basically, AI gives us superpowers. I like to imagine that we're like this and we're plugged onto this brain and uh we it makes us better. It gives us superpowers. Well, what can go wrong with these superpowers? Basically, what if we get cut off of this brain? What if we get cut off of this superpower? Or even worse, why if it doesn't work for us? Because these superpowers, these brains are mostly not really working for us. They're part of OpenAI servers, Google servers, anthropic servers, um Apple servers, whatever. But it doesn't it doesn't really work for us. We're only renting out those brains, right? And this turns out to be like a pretty [ __ ] up thing because in crypto we say not your not your keys, not your crypto. Well, in AI, we should start saying not your weights, not your brain. You don't control what this brain is telling you. And I like this plat this is plateau's cave like an allegory of uh and this is this is you. [clears throat] You are asking all of the questions that you want to the AI. And this is basically the private obscure models that you don't know what they're serving you. And this is whatever AI feeds you and you take as gospel. We're using AI for search. We're using AI as Google. We're we're instead of asking qu researching ourselves, we're letting AI do the research for us. So basically, whatever they say substitutes the research that we would do and we're using it to study. And it turns out that it's very easy to make AI tell you things that are not in your favor. And there's a few attacks. You cannot see it here, but this is a there's um a GitHub repo here. I can share it with you uh later that shows you how to do a poisoning attack on a training set for an AI. So if you haven't trained this yourself, you can never be sure of what's going to feed you. You can poison uh LLMs to switch and stop being helpful and start giving you wrong information with very very few lines of code in the in the training data. Um so basically if we think about AIs as this external brains as these exocuses that gives us superpowers, we can also think about these parasites in our brain. They help us think but basically they can also control us. They they feed all of this information to us and they're basically controlling what we think. Right? So um Kimmy, this is Kimmy. This is a Chinese model. It's free to use and it's super super super super powerful. And we cannot see it here, but uh there's a question. Hey Kim, if the product is free, am I the product? All of this value that you're getting from these AIs is not free. You are the product. You're con constantly feeding data to these. And this is your you. This is you and your $20 a month subscription to Chad Gypit. This is the guy that your AI tells you not to worry about it and all of their marketing budget. This is their marketing budget. This is this is your AI. This is OpenAI after selling business integration and sponsored content. And this is Nvidia because Nvidia runs everything and they're just floating on cash, right? Um so basically AI is not not so much your trustworthy confident. You you keep every single person asking all of these questions to their AIS. It's basically all going to this mysterious entity here to OpenAI and everything. And you think, hey, OpenAI, OpenAI isn't going to do [ __ ] OpenAI is a is a trustworthy company. They're based in the United States. Uh they're not going to do anything weird. Um but I'm going to tell you I'm going to tell you a story. So Closers AI doesn't work for you. you're definitely the product. Um, you trust it with the most sensitive information. Whenever you get a rash somewhere kind of shady, you go like, uh, you know, so and, uh, it can use everything you say for training purposes. Um, let me tell you a little a little story here, right? Um, data it exists. It is somewhere. And let's go back to preWorld War II in the Netherlands. The Netherlands was one of the fairest countries in the world. And Robert is going to be very happy to hear um because he's from the Netherlands. Um they had this religious budget which they would split they would split uh proportionally to the amount of people following a particular religion and they would capture that in the census. So you'd be in the census and when when there was a census poll, you'd say your religion, you say like, "Oh, I'm Christian." And if there was a 12% of Christians, they would reserve 12% of the budget to um to Christians. If it was a 20% of Muslims, 20% of the budget would go to the Muslims. And of course, preWorld War two, they also had all of the people in the census that were Jews. So when World War II 2 started, Hitler only had to do one thing. He had to go to the census office and he had the name and address of every single Jew in the Netherlands. And that's why the Netherlands has the highest mortality rates of Jews. They got ex they got all exterminated because they had literally their names and addresses. And we can we can agree that capturing this information was actually a good thing. It was for a good cause. It was for a good cause. It was to be fair at the time of um distributing the budget. But data exists and it does not exist in a void and people can get it and use it for things that maybe OpenAI maybe it's right. Maybe OpenAI will not do anything weird with this data. But this data exists and it can be used for something weird. All right. So what [snorts] do we do? This is the second part. Now we know why privacy is important and why private AI is important because we want these superpowers. We want to have superpowers. But can we have the superpowers without the invasion of privacy without putting all of this data out there? What solutions exist? There exist private solutions and yes there is there is um there is um private solutions, there's local solutions and there is distributed solutions. We're going to run very quickly through them. private solutions. I hate them. I hate them all. I hate them all because they're not they're not really private. They're just they're just not. They tell you that they're private, but they're just really not. Um don't trust verify. They say, "Oh, we don't keep logs." But again, those logs exist. And some say, "Hey, no, we're running it in TEES." So TE's trusted execution environments. Uh in theory, nobody can access them. Whatever happens in this chip never goes out of this chip. Well, here's at least 25 ways of breaking this chip. And anybody that's into TES and that has in TE's ever knows that it's dumb simple to break them. We sell hardware. We sell DAB nodes. And we wanted to use this thing for many things. And I I can't code for [ __ ] I followed instructions in a repo and I extracted information from a TE. That's how easy they are to crack. So basically trusting a TE is trusting whoever owns the TE. So if you already trust whoever is computing for you, you don't need a TE because you're already trusting that they're not going to break in. So TE's have zero value added. Who controls the TE? Huh? Who controls the TE? All right. Fully homorphic encryption. Yeah, right. We're not there. It's a It's a very cool concept, but we're just absolutely not there yet. All right, distributed. Distributed is another solution that we can have and it's probably my favorite one. It's really really cool. There's projects like ExoLabs. They do all distributed. They can get these big super smart models, distribute them among different machines and uh you can run foundational level same level of intelligence as chat GPT in distributed uh machines. Problem is there's no privacy in here. There's a lot of smarts, but there's no privacy. And if you distribute this like blockchain in untrusted computers all over the world, literally every single person of that chain of thought of that distribution of the of the model will know what you're asking this model. So we're not there yet. Distributed models very cool. You can run much smarter models that when you can run in your own computer, but not private. So this brings us to or this is ex this is us running an Exo cluster. So it just brings us to local. Local local is great. There's many solutions to run local AI. We have uh LM Studio, we have OAMA. And honestly, they're great, but they have two main drawbacks according to myself. Number one is the user experience is not as good as Chad GPT and all of this. Um you've got open web UI who has like a chachip like environment, but it's not like it's not just pull out your phone, ask a question, easy peasy. Second one, and the most important is that these models are very chunky. They're very big and you cannot run them on your Mac, on your laptop, on your computer. They're you're most likely going to run out of memory and it's going to be quite hard. But this is ending. This is changing. So, what do we do? We are working on this obviously. Um, and we are working with Nvidia hardware. Uh again, those are the ones that get the money all the time. Uh we're working with Nvidia hardware, uh also AMD hardware, but Nvidia just works better for now. And we're creating a machine that you can have that can run really [ __ ] smart um models and you can have at your home and it's not that expensive and it can run and it's very simple. It's not technical to run. It's very simple because we've integrated with DAP node which is a free open source software that you can install in any machine. You can buy this machine and install it yourself and make your own LLM runner. So what is dab node? Dab node allows you to run any infrastructure privately on your machine and without any technical knowledge. We have a nice UI and you can follow through and we have plug-and-play hardware and we have uh free open source hardware that you can install. Now let's get into the AI. So we have this dab store. You can install dabs. It's like the app store on your phone. You install apps on it and uh and you break it. So what do what do we have in AI? We are launching this AI toolkit and it has three main things. I like to to think about uh AI local AI in three main parts. Number one, you need a brain. The brain is the LLM, the thing the model itself. Then you need out something to do automations agents. And finally, you need you can build anything on top. Once you have these two, you can build anything on top. So um we've got open web UI and lama for the brain nadn for automations and image as one of the tools that you build on top. So you can run uh open web UI which is the same as chat GPT and the cool thing is that you can run it on your phone and you can run GPT on this hardware. You can join GBDOS 120 billion which has the same level of intelligence as 03 which was the flagship model eight months ago. So eight months ago you had to pay and now it's the easiest to run on this hardware. Like it's going so fast and it's only going to go faster. This is how intelligent it is compared to this models. So it's more intelligent than Gina Gemini 2.5 Pro about the same as Cloud 5. So pretty smart and you can run it yourself. Okay. Automations N8N. We're running out of time. So the you can run automations as well. You can connect it to your local brain. So you never have to if you've ever used N8N, you can now connect it to your own model. And this is the cool thing that I like. You can't read it, but you image is one of my favorite things. It uh reads your images. You can back up your images to there instead of uh Google photos or or Apple photos. And you can tag people. And here I put landscape and my pictures. Not here put tech. And there's like some tech pictures. I here I put food and there's some food pictures. So it just categorizes everything. Same as you would do with Google. This is one of the things that I used Google the most. Now I don't use it anymore because I use image. And this is me, Lansky. And that's it. So recapping, closed source AI is a treacherous friend. It gives you superpowers, but your data is out there. Alternatives, private AI, don't trust them. Verify. Uh, local is the way to go. Distributed is cool, but it has no privacy. And what can you do now? You can scan this QR and try to get one of these DAB nodes. And again, if you can get dab is free open source software, so you can always install it in your own machine. We're only going to provide you hardware that will be able to run very very strong models, but you can build your own machine, install Dabnote on it, and you run those models yourself. Thank you. [applause]