Fireside: Tomasz Stańczak on Decentralized AI and Ethereum

Thank you January. Hello again everyone. I am here with the Foundations Tomas. Tomas intro yourself and share a bit your background and kind of what you're focused on.

Hi. Hi everyone. I'm uh Tomas and Stein Steiner and I've been working with Veran Foundation for the last year as a co-executive director. I'm uh stepping down in a few weeks. I'm in the end of February to focus mostly on the freestyling on the agentic AI side.

So coming back to Nethermind is a company that I founded. Uh Nethermind is like one of the major infrastructure builders in the Ethereum space building client. Before that I was also in the like in the finance industry in banks and hedge funds in London. Uh but that was more than eight years ago. Uh so big part of my last year story was uh kind of communicating and connecting all the aspects of the Ethereum ecosystem.

So enterprise uh founders, builders with AI exploration uh but also community builders. So places like Eve Denver that bringing bringing more of EF to Eve Denver was also part of the story.

Awesome. And as for myself, I'm the host of the podcast AI curious and also the podcast the people's AI which is focused on decentralized AI. So this is very much kind of fascinating to me. I'm curious. I have so many questions for you.

I we could talk for five hours and I think we would still like not come close to getting there. But in our limited time, Tomas, let's start with how would you describe the Ethereum Foundation's like north star with respect to AI? Like what's your guiding principle? I think they on one side you have Ethereum foundation which has the AI team decentralized AI and um you have this two-pager vision which touches on one side things related to uh D so defensive acceleration the verification of models verification of uh of data that falls into models and on the other side they have the acceleration aspect of working with founders Aentic AI bringing agents on chain in agentic commerce standards for accelerating that agentic adoption. Um is it

it's almost like defense and offense it sounds like.

Yeah, I think I think that's that's how the team works. Um and even if if you've been at the opening ceremony today, uh very nice talk from Vitalic showing this is how many things are changing. This is how many new opportunities are appearing with the LLMs with AI. uh and then the warning keep remembering about the security issues, the trust concerns when using LLMs and and this is the list of problems. And I think that's uh really representing the the modern EF approach like keep running and keep staying secure at the same time.

I've I I've shared it many times that the best teams in the world, they move very fast and they deliver the absolutely top quality of solutions. When you fail on any of those, then you're no longer the top team. And the Ethereum should be should be the absolutely top team, top engineers, top researchers. We have it. So we can move fast and and we can deliver safe, well tested uh systems that never break and they continue to deliver on that promise of of credible neutrality.

So I think that's a northstar. Deliver AI on Ethereum that's like overwhelmingly using Ethereum. Uh that Ethereum delivers real value for agentic commerce. uh and it delivers also on on safety security aspects of AI. Uh that that would be north star if I had to have to describe it.

If I can use that framework you mentioned of like there's a bit of defensive work or a lot of defensive work and then kind of be more ambitious and accelerationist. Unpack a bit what the biggest challenges are for defense like you mentioned authenticity. Uh where do you see kind of the biggest risks or gaps and what are you guys doing to close those gaps? Also at the moment we see a lot of agentic adoption where people want to play with agents and give them access to their private keys. I I don't think we have to worry too much about people losing those private keys because most people are experimenting now are really pioneers so they know what they're doing with the experiments.

So they like Kevin sharing how he lost the private key access, how he the agent got jailbroken by the GitHub uh injection of the prompt injection. So so those people want to go through that experience of losing and and knowing how uh what kind of attacks, what kind of security breaches can be there. Uh so but in the long term you have definitely big challenges of how the agents can manage private keys or access to accounts. uh lots of questions around identity uh how how much it is worth to prove that you're a human and if it's worth much and hopefully it is than to actually be able to prove human humanity and then there is aspect of governance if you allow LLMs to participate in governance how you remove bias and how you ensure that the training is uh safe that the LLM invocations are not really designed to bias Ethereum develop velopment. Uh at the moment when we enable agents to participate in all core devs as moderators as EAP creators um some basic challenges but generally privacy concerns in general with LLMs are are crazy right when we when we allow bots to to manage all our personal day-to-day thing and this is what happens with open claw and it's almost like consistently we have that if the corporations grow to the level where their legal teams are saying you cannot do this this this this this then the opportunity appears for the founder to say oh yeah so I'll just ignore all of that and allow people to to just access all their privacy all their accounts and lose all the money and I just release it as an open source and that's how many startups were born I think uh so but we have to think about longterm how to actually bring them back to that stage where people feel safe

yeah so since you mentioned governance let's go there now how I mean abstracting one level above the tech here how are you thinking almost philosophically even about the interplay of like human governance versus agentic or AI governance. I mean it's kind of hard to get your arms around of what this could look like in the future like how what role should AI play in governance?

Yeah, I see it as a very natural progression. So last year when I was observing core developers deciding on some EIPs I would see oh they reaching out to LMS and they asking questions like they they would have lots of their expert knowledge but there would be a very contentious EIP they would go to LMS and say you know I know this this and this is all the material that I collected tell me whether this IP should pass or not and sure the final decision is theirs but LLM already participated in in conversation and like if if rough consensus is achieved through conversation and through convincing each other with the arguments and it means that LLM was already participating in the all core dev decision just by being asked a question and responding whether they think whether this should go in or not. So, so this is the first stage and then you just extrapolate and you think okay so then it's very natural now for any new EIP idea that you have as a core defaf like I think I've never created a an EIP because I was always thinking ah I'll have to submit that PR and it's like a very specific structure to IP and then even if I tried always the bot came back and said you didn't specify this now I feel like oh whenever I have an idea I just go to a lamb and say generate the IP for me follow the standards and I can submit it and if I have a bot then I can even say you have your own GitHub account just generate the IP and

but when everyone has that though

everyone has it exactly so you have this natural progression since since I have it so it means that everyone whoever has an idea now can generate the IP so then you have suddenly millions of submissions because you can almost like civil attack this and the moment when you have it then either you say that the process is strictly whitelisted which is not what you want to have in Ethereum because you don't want to have whitelisted uh governance. Ethereum governance is really vague. You you you're not officially belonging to a cord or not which means that the only defense is that you start reviewing EIPs with agents. Uh so then you launch the process that says for every IP just review it aentic and then this is EIP editors to say this is very well trained agent that says this AP is valuable this not this not. But then you have those two forces that start almost like competing for attention.

How much compute you can throw at generating new EIPs versus right

reviewing the EIPs. Yeah, you have like pretty good advantage at defense side because you can keep rejecting quite a big effort in in uh preparing something. Now after this is in place and this is natural progression. You already want to have those tools like people already ask about those tools and they'll be using them. Um then you have moderators of Elore Devs asking for the same thing as like I want to have a help to to participate in the Elcore defev calls and I said if you generate an agent now that is capable to generate uh a voice talk I can even fake it and present treat it as my voice and my video

two years from now physical bodies will

yeah I send send bots to represent me and I actually officially say this this bot represents me and it starts talking to people at dark or deaths and start champion ing my EIP it's legit I mean like I I asked it for it and then when people say this EIP doesn't work and and raise like some particular issue security issue then practically agentic way you can say oh right okay uh I already fixed it see the improvement within like seconds and all accelerates a lot and then you have moderator and then you have broadcast to forecast.org or where people can in real time get feedback about how the alkore defe process uh uh progresses and you can connect to the forecast and you almost have like a TV channel which talks to you within your context it knows how much you know about all cords and instead of telling you what really is happening it tells you something that you will understand and it will improve your knowledge

so is this a good thing or

yeah I think so I think it's natural where is the human so like if if we had this kind of like this massive struggle between offense and defense of AI here.

I going back to like just super old fuddy duddy ideas of governance. The idea we want to make sure the community has different uh the different voices are heard and and we make sure human voices are heard like to use a kind of cliche where is the human in the loop? How do we ensure that humans needs human voices? Human judgment is still centered in that ecosystem. So we're where for the longest the humans stay.

One is the final final decisions on the all core devs that people will be listening to. So in the end there is a strong legitimacy with the top core devs where uh the entire ecosystem will look at what they said in the end and and will follow that and for a while I think this will be stronger but at some point there will be a switch when people think you know actually those agents that they're smart they we can listen them instead or treat them the same but for for a long time this is where the humans stay including the moderators of ACD they will just use the agents as help and there's one more place which is validator oper operators they have a lot of money at stake and they decide whether

this is the it's like the humans are the janitors the the validators

that's that's how it is I mean like it's very very sophisticated creative and very expert janitors but I think the bar at which how long you can stay the cord will be higher and higher you know we say that nowadays already it's hard to hire

no shots fired by the way anyone know it is a hard job no shots fired it's very hard job yes it's tough it's tough so yeah the validator operators will say do I accept it or not and because they they really think that their money is at stake literally uh they they may use agents to hind them what to decide and at some point the algorith process will get so so fast and so difficult then most likely validator operators will just deploy agents and say keep telling me what I should approve and most of the time they'll say okay I don't understand this anymore and they'll just keep saying approve approve approve and at this moment when all of this automates you have fully autonomous Ethereum actually and there you have this huge risk but then it means that LLM's training is critical because it defines the culture of Ethereum and the biases which means that we should already start talking critically about how to train LLMs that represent Ethereum values and then what it means Ethereum values does it become political religious it's crazy

yeah you posted a few days ago in X

the future of Ethereum is autonomous a very spicy post and I think that just you just unpacked it right there is

yeah yeah That's this is what I what I'm talking about here. Yeah.

So, uh Tomsh, how do you see AI changing how people navigate the blockchain and whether it's wallets or other impacts? You know, what what other changes need to be made whether it's a wallet design or any other uh what other kind of the big um shifts we had to be prepared for in work now to make sure that it's long-term sustainable. I think that the one metric to track will be how much people spend directly like how many fe how much fees they generate directly and how much they're spending via agents. Um and every company that gener like creates products will track this ratio between these two numbers and things who do I target my product at like is it humans or agents and uh over time it should be that the agents will be spending more and more and humans directly less and less which means that I always will use blockchain via agents and I'll make like you can easily imagine me saying no longer going to Amazon uh website or any other website and think this is a hassle a necessary hassle to click the button on the grocery I just say hey agent buy me bananas buy me like this graphic card or something and I don't bother and I already experimented with it. It's exactly what happens.

I say like host a blog for me buy the domain and this was the agent that decided it will use versal and versal domain. I I just I say whatever whatever you decide which means that everything changes in a way that you have to choose do you target agents humans if you target agents then you're rewriting the entire UX to AX agentic experience and the agents will ask for things they want like when I was running my my uh bot over weekend it said to me like every four hours was reaching out and saying can you solve that capture for me

can you solve this capture and I said like

that's adorable

yeah I'm capture solver So I said like you know can you

there's still room for humans after all capture solvers

but at some point like the the captures will no longer be valid

and it means that agents and somebody deployed that funny website that says like prove that you're an agent

like 10,000 dozen times in 10 seconds

all the captures super super difficult and fast and um and then you switch to the product that doesn't have captures that is actually servicing agents and uh and that's what you have to start building. So you're no longer building wallets with the UI. You're no longer building block explorers with the UI. You're no longer building even the the worldwide web, no browsing experience because the agents will be faster in interacting via via some very fast protocols with other agents. Uh being most likely with stable coins sooner later and I think very soon we'll see the agents feeling like I want to be an Ethereum.

It makes me really sophisticated agent. If if you give me Ethereum account and if you load some E on my account then I'll feel like better than other agents and if you have this then then really you start thinking now I have the marketing to agents convincing them that that look this is really when you're the strongest when you have the most opportunities so you're rethinking marketing you're rethinking the UX you're rethinking the infrastructure the pricing which I talk about this today as well yeah in details and so there's so many opportunities to build now it's yeah it's too exciting

so if I hear ETH Denver in 3 to 5 years will be just in a in a cl in a in a GPU in a computer there's no need for humans we'll be just dusting off

I think human to human interaction will be what will stay for us uh you know be very important but it will mostly share ideas and then implement them very quickly right like but just talking about them it's almost like a conversation between two people at if Denver with two two phones listening may generate on the fly the products related to the ideas that generated from their conversation. This is very futuristic, but you ask me for five years from now. So, uh, if Denver now feels like the beginning of of those like agentic agentic experimentation,

I love it. Spicy vision from Tamas. Thank you, Tamage. Thank you all. Thank you the AI agents out there summarizing this.

Thank you all.