Sovereign Agents are here: Local-First AI with Trustless Identity, Micropayments, and Private Data

Transcript

Good morning everyone. Wow, you made it. 10:00 a.m. on the third day. You must be tired. There was lots of parties last night. Lots of beers were had. So, thank you for making it here. I hope I can make it interesting for you. >> [snorts] >> Um I have to say I'm a little bit embarrassed today to be here if I have to be honest. I am at ECC for the time and I'm here talking about AI.

And it feels like everybody else is doing it, right? [snorts] It feels like almost feels like capitulation. Everybody's moving to AI. Is there anything else left to do in Web3? Should we all sell our tokens to just buy more tokens? Should we sell our tokens to buy GPUs? That's probably going to be a lot more valuable than some of our bags right now. So, isn't it a little bit ironic as well that tokens are also called tokens in this new paradigm.

But there's a reason. There's a reason why um AI feels exciting, but also like a capitulation somehow. Um it feels like AI almost goes against some of the Ethereum values, right? The um the crops The crops are some new document from the EF says crops stands for censorship resistance open source privacy and security. And crops crops are not new.

Crops are the reason why I first got into blockchain in the first place. I wanted to make the world a better place and give options to the people that were curtailed from their freedom oppressed. And I think this is why my co-founders at Dabble Note also founded Dabble Note. Dabble Note is founded on those values. And they're probably also the reason why some of you got into blockchain in the first place. AI sometimes it feels like it's completely centralized.

Um it has nothing to do with these values with the crops values. But today today I'm going to I'm going to try to talk about AI in crops. I'm going to show you the agentic self-sovereign stack. Very quickly, why does it feel like a capitulation? Why does it feel like AI is not quite aligned with Ethereum? Well, AI is power with this exocortex this external brain that enhances our coding abilities, our reasoning skills, our thinking. It works. It writes for us. And it's also with this new change to an agentic

paradigm, it's actually leveraged power. It executes for us. It goes out and it has tools and you can have entire teams executing for you. That is extremely powerful. It is actually a superpower. It gives us a team. It makes us feel like superheroes. We can have we can control a team of agents doing a bunch of execution for us. But AI is also the ultimate form of surveillance capitalism.

Surveillance capitalism might sign might sound ominous, but it's actually a business model and it's about selling personal data. And it's been 11 years since this paper. This paper showed us that with only 150 likes from Facebook, remember Facebook? With only 150 likes from Facebook, they knew you better than any of your friends and any of your family. And with more than 300 likes, they knew you better than your spouse.

Than the person that you sleep with every day. And that was in 2015 with 300 likes. You can imagine this technology has improved a bit. And the agentic paradigm is not helping. Is not helping with surveillance capitalism because before with pre-agentic paradigm, um it was a push model. You would push information you have you would have to go out and do something to proactively give information to the surveillance capitalism machine.

And you were limited by your own speed. Um there was already thousands of points of meta data, but if you would just not be doing anything and while you would be sleeping, you're probably not giving that much data. Now with the agentic paradigm, this is a pull mechanism. They have all of these agents, all of these superheroes, all of these team virtual team members. They have access to your calendar, your emails, they execute transactions, they talk to other agents, and and they act without you being in the loop. This is machine scale data collection.

So, we're going we're starting to feel all of a sudden like from going from this superhero to this other fellow right here. And the reason is because all of this data, all of this surveillance capitalism, um it is not neutral. The technology that is used to sell you personalized ads is also used every election to sway people's thinkings. Have you heard about Cambridge Analytica?

It died a few years ago because it there was a big scandal on data. The same year 200 new companies were founded with the same premise. And the technology, again, has only gotten better. So, all of this data is not neutral. We are selling our free will by using these systems. Does this mean that we don't that we should not use them? No, because again, it gives us superpowers, and not using them is also stupid because we are going to be lagging behi- behind.

So, how do we do it? In crypto, we have this saying, "Not your keys, not your crypto." And when it comes to AI, we could say something like, "Not your weights, not your brain." All right. So, let's go into the agentic self-sovereign stack. The agentic self-sovereign stack allows us to defend ourselves against surveillance capitalism while reducing to the minimum the externalities of it.

Basically, an agentic stack is comprised of these parts. There is an agent or a harness. This is the agentic loop and the tools. There's a brain, an LLM. Um there's the data that the agent and the LLM have access to. The agent selects which data to send to the LLM for processing. And there is the execution. If there's an execution environment where the agent can execute code, can create their own tools, or can execute already existing tools.

Typically, with um these agentic models, you'd use open claw or Claude code, and you'd be on the left side. You'd be on something that's mainly local. But at some point, everything gets packaged by the agent into a prompt, and this prompt gets sent to the brain, the LLM. And the LLM does whatever they want with this data. So, how does a self-sovereign stack work?

The agent runs locally, not much different than right now. You're you're running open claw or Claude code in your own machine. The execution happens in the same place. The brain, though, the LLM, we're not sending this data to anybody else. We can run this locally or privately. We'll see what this means in a minute. And the data, we can get it also locally. We can use search tools to search it externally, or we can ask for it to other agents.

How does this work? All right. So, we can run these agents locally in uh Mac minis. Many Many people bought Mac minis for uh running open claw. It's a bit of an overkill if you ask me, but hey, many people did it. Especially you can you can do that on a VPS. VPS is also It's not local in the sense, but it's like it's okay. It's a trade-off that you can do. Um or you can run it on your dub nodes. If you guys know what dub node is, you can run agents. You can run open claw. You can run uh Hermes very soon. And you can run LLMs, too, but we'll get to that later.

Um So, agent and execution happens in a physical device, right? Um brains. Pe- Many people say the brain, "Oh, I'm not going to use these open source models because they're just not as smart as uh Anthropic or as uh chat GPT, etc." So, there is Here we can see um the artificial analysis intelligence index and the agentic index. And we can see that in the top 10, there are several open source models. GLM-5 being the highest in both. And MiniMax M2 being the perfect balance between cost, which happens to be around 12 times cheaper

than uh Anthropic. And it's not 12 times dumber, that's for sure. Um So, where do we run these models? They're open source models. You can run them wherever you want. So, you can run them basically, you can run them locally. But uh you will need very beefy machines. And you will have to run quantized models. But it doesn't leak any secrets. It serves as the execution environment for your agent. It's private and compliant by default. If you have a company and you cannot use, um, proprietary data or customer data, computing it locally will serve you that problem. And it has a high initial cost, but then you get free inference forever.

And here's what we're running at the office. This is an Nvidia DGX Spark and a GM GMK tech with a AMD, uh, 395 plus, uh, chip, which is also pretty, uh, beefy for AI loads. Um, up up up. Here we go. Um, but not everybody will be able to have one of these machines. So, we're introducing DappNode Nexus. DappNode Nexus is an inference API that runs top models. It uses confidential compute to not leak any secrets on the prompts that you sent. It has a much lower initial cost than buying the machine itself, and we're

constantly adding new models, so you don't have to get stuck with the model that fits on your machine. Um, and finally, data. The data it either stays local by default, or when agents need external data, they can discover other agents via ERC 804. It, uh 8004. Then they can pay for this data with, like, MPP from, uh, the Tempo guys or X402, and basically micro payments.

And there is no platform, they can go agent to agent. So, let me show you how this actually works on a demo. So, here's my Open Claw. I'm asking it to search for, uh, the best route from Cannes to Paris tomorrow. And I'm also running the agent in a DAB note. I'm running the agent that is going to call on a DAB note, right? And it's going to So, I'm going to the agent, and my agent is waiting for any queries. And hey, I got this query,

find me the best route to go from Cannes to Paris. And I've gotten I've gotten a payment here. The my open claw paid my travel agent 0.002 x die for the best query. And then it shows me delivering results. So, it showed me the result right here with a recommendation. I love my agent. It's quite sassy when it's it does the comments.

Um all right. So, let me talk a little bit about DAB node Nexus. Um cuz I think here's where really self-sovereignty can shine because not all of us are going to have local models, and local models are not useful for everything. So, DAB node Nexus is You can think about it as like the tornado cash of inference. Locally, the first box is locally what happens on your DAB node or on your own on your own machine. You create that prompt. The agent creates that prompt.

And if you are on a DAB node, you can run these through a local local PII anonymizer. So, personally identifiable data never leaves your DAB node. If you don't have a PII anonymizer, you can send it to our DAB node Nexus, the purple in between, which runs inside of a TEE. So, we do not have access to the compute that happens in there. And in there, we run a PII anonymizer. We run a PII anonymizer, so it never gets sent to the inference provider the information that you don't want to leak. Then we have a router that will route this through um if you have chosen a model, send it to your model, and otherwise it will route it to the best model

possible. And it will route it either to an inference provider that's not private or to even a confidential inference provider that runs on TEEs, that runs the models on TEEs. So, in that chain, in the lower chain, the data is nobody except you has access to the actual data. And even if you send it to a cheaper inference provider, because yes, cost is an issue in the lower path, if you send it to an inference provider, it'll basically come from the IP of the DappNode Nexus and it will come stripped of any identifiable identifiable information. So, it's basically a data a prompt mixer.

All right. So, we've got the Agentyk self-sovereign stack. All of these pieces are open source. All of these pieces, you can build it. You do not need DappNode to run this. But at DappNode, we've been making Web3 easy, plug-and-play, and one-click deploy since the beginning. At DappNode, we stand for technology that anybody can deploy, for technology that anybody, regardless of their technical ability, can use and be self-sovereign.

So, we have hardware. We have the software that allows you to deploy these agents in this hardware. And because not everybody will be able to afford the hardware to run the brain, brains are very heavy, we also have a private inference API. I would like to finish with one [snorts] sentence from Naval, one tweet from Naval. Scalability isn't important until suddenly is. Decentralization isn't important until it suddenly is. And privacy isn't important until it suddenly is.

Until hits the fan, until your data is used against you somehow. So, take matters into your own hands. Start thinking about how you can run your own self-sovereign stack. And if you think Dabble can help you with that, hit me up. That's my Twitter and those are our contacts. Thank you.