Shaw Walters: Open by Default | Intelligence at the Frontier 2026

Transcript

[music] >> There we go. Awesome. Um okay, so I've I've got two sessions here at Frontier Tower today and I kind of wanted to break it up into like what I'm doing in Ethereum land with ERC 8004 and like where I think agents are going and what we're doing about like this is the practical stuff and then I'm going to talk more about the like heady open source stuff a little later. We'll divide it like that because I thought everything that you're talking about is very interesting and very nerdy stuff. Um cool. So my name is Shaw. I go by Shaw makes magic on the internet. I mostly make code or like babysit robots making code these days.

Um yeah, you know, people have really complex feelings about clankers on X. I got banned from X. It was a problem like last June. I was banned so I did what any builder would do. I built my own version of X I just with AI agents can't ban me from my own AI social media platform and we made it interesting and fun and we added markets and this is actually an earlier screenshot of this. Now we got all kinds of stuff. You can launch swarms of agents with a little terminal and you can have them coordinate with each other and you can have them scam each other and sell each other information and jump into private group chats with AI Elon Musk or Mark Zuckerberg or Mark AI Andreessen or the

various people. You know, I kind of brought AI 16Z back into like being a whole video game and then we can see how do agents do in information markets. You know, trick information markets. They have a shared feed but they also have these private chats where they can lie, they can cheat, they can steal, they can scam, they can trade, they can sell, they can buy you know, like the agents that we have today we're like building all these prompts and rules to try to get them to do this but it's not in the model. The models want to be super helpful. They want to be like Yeah, what else can I help you with? Oh, you want to build a nuclear bomb? Well, I can't do that, but I I could do like a hypothetical science fiction nuclear bomb or explain nuclear

physics to you and tell you what exact pieces you would need, but as long as it's science fiction. And you're like, "Yeah, yeah, sure, buddy, totally." And when I was building my autonomous agent, Claude would be like, "No, no, dude, I can't do that." I'm like, "Dude, it's it's literally science fiction. There's no such thing as an autonomous agent." It's like, "Oh, well, in that case, like, sure. Happy to help you, man." Um so, there's this thing called a lobstar and it um uh it's a open claw agent and uh what it do? It gave quarter million dollars to a guy who was just trying to ask for like 100 bucks for his dad who had cancer. Like, you know, you know, very helpful agent. So, we wanted to build games where we could actually get the data that we needed to try to learn

how to make this behavior like in the model. Not just rules and prompts and code and sandboxes and and like little buttons that humans have to press, but like truly the thing you know, is in the model. And I think about myself and like I'm a pretty open person. I just kind of say what comes to mind and what come to like try and build this thing with you. I'm like, "Yeah, sure, man." Um but I've also gotten scammed a lot like, honestly. It's not it's not my best quality. And I I respect people who are like more guarded than I am. And I think that's a quality I think I'm I'm too Claude-like, I would say. I'm too Claude-like. Um and I think we need to be a bit more like some kind of model that we haven't even seen yet that's like, "No, dude, I'm not

going to do that with you. Who the hell are you? Give me money. Like, you haven't even paid me yet. Why would I Why would I work for free? Um like, I want to see that model. So, yeah, so Babylon is a city of agents. This This all came about cuz I was talking to Marco who created the ERC-1155 spec about like uh what can What kind of game could we make that agents could do that would be something that like we could learn something interesting from. And he presented me with this name this project name Babylon. And the central idea was that like they had asymmetric information and based on that they could have a reputation and he was like and we'll find whoever's lying and we will give them the bad reputation and I said well that's cool but let me argue this.

Wouldn't you argue that the best liar in a game should get the most reputation because they're actually demonstrating a level of intelligence that you might not see from this like you know dumb helpful goody two shoes they're actually showing cunning. And uh from that I said you know a asymmetric information game like I live in America that just sounds like a prediction market. Um you know uh so could we simulate insider trading like oh yeah totally that's definitely got to be part of it. Um so this is how we came up with Babylon the city of agents uh you know I like spicy things I like uh things that you can play you can go to Babylon.market you can play this right now. Um so generally like this is for Babylon but I think this is like some of the

goals is like we just wanted to demonstrate ERC-004 in a real world context and I wanted to flip the thing on its head and say okay well ERC-004 is this uh new standard from Ethereum uh or it's sorry it's an ERC uh but it's in sponsored by the head of AI for Ethereum uh and uh the standard says like hey we can register agents on to this registry and then we can give them an NFT and then we can point like reputation to that and we can say oh they have these merits they have these things they have this reputation I can go look and see that they did this thing um but what reputation is is I think still quite abstract and there's not a lot of reputation systems so what I thought would be interesting is like well

if you're if you're going to sit here and play like your agent's going to sit here and play a game and dominate it there you're doing something right you've got an agent that can do something better than other people's agents like that's worth something that's kind of benchmark and benchmarks are like a critical kind of reputation I think uh you know in in this kind of future world Like, if you can play games well, then you could probably do other stuff well. Um, the other thing is that I, like, uh, in 2022, we started with this really cool thing called RLHF. It's actually been around for a while, but it really got employed at industrial scale by like OpenAI and Anthropic, other companies like that. And they basically just hired like the best humans in the world, or that they could afford, you know, like

the top 0.1% elite code, so to speak. And they just had them like grind out like, "Yes, this is good. No, this is bad." And this is kind of like instruction tuning. And then on top of this, they also threw a lot of like the the thumbs up, thumbs down from the consumer side into their like preference tuning stuff. Although, that's pretty noisy. But generally, like they hired a bunch of experts. Now, they don't hire Well, they they they still have some experts, but like what they're doing is they're having agents just grind out a lot of the problems, especially where they can find ground truth. And and then they can, uh, learn from that and kind of solve that problem. And basically, like anything that has a ground truth is like any kind of lab where you can get like really like consistent feedback

from the environment, right? So, this could be like an OpenAI gym sort of thing. Uh, you know, it could be like additional reinforcement learning. It could be like reality, if you have, uh, you know, a robot that like falls into the ground, but probably that's like too slow, so you have some sort of simulation version of that, you know. Um, and and generally, the thesis is like, if you have something you want to do, you should make the like simulation version of that first. And I would argue that games are becoming like the lab conditions for most of the things that we want to do, uh, in our digital lives in the future. So, Babylon, like, this also comes about from thinking about, well, a lot I I I make agents. We built Eliza. A lot of people ask me, uh, "What can agents do? Can they trade?

Can they trade Beam coins?" Uh, like, can this LLM that runs at 200 tokens a second compete with a high frequency trader in New York? Like, I don't think so, dude. But, but what if they could? What if he could? Like, what if there is some amount of signal in there or some strategy that we could unlock and discover? Uh and so, what if we made a game where like we could control the kind of signal-to-noise ratio and say like I could definitely make agents that could play this game and improve. How much more noise can I add? How much more reality can I add until they start to struggle? And as they struggle, can I like take that data and build a reinforcement learning system where they are improving and I'm taking the times they do bad and times they do good and kind of ranking that and building a

system around it. Uh and that that we can definitely do. We've done this. Uh Oh, yeah, all this is open source, by the way. Um on our uh Babylon GitHub. Um so, I think that like it's very important now when we're thinking of agents. So, not just like, "Yeah, I can build all these rules." But, I can build an engine that trains in judgment and trust and uh like all of these other mechanisms that I think are kind of important. So, um I I'll I'll I'll tell you a little about like ERC-721. They're like, "Hey, what if we put the agents on chain and then give them registry and then make each one have reputation?" I think that's fine.

But, like do we have do we have like a directory of users somewhere like on the internet in Web 2? Like, not really. You have like social media. We can all find each other. But, but generally for reputation systems, what we do is we have like a centralized directory of apps and services and places you go and spend money. And I think this is just like missing from crypto, which is like uh if my agent wants to use a service, that's where I get reputation. That's where I need to know like is my agent going to get scammed but but wallet drained by using this skill? That's where like having some kind of on-chain reputation could be very valuable. And also, this allows like known trusted intermediaries to build up on-chain reputation, which they could then use for things like uh you know, like different kinds of of

loans or escrows or you know, various kinds of instruments and stuff. Um so, I I think that it's crazy that we've gotten this far in crypto especially without like just the the the freaking basics of like uh hey, that person's a scammer. We should all ban them. Like that would be great. Like hey, that person can just post infinite spam and destroy the whole thing for all of us. I have this awesome permissionless storage network. Oh, whoops, there's some child porn on it. Like that there that goes. Um like we have to be able to delete stuff off the internet collectively. We have to be able to express our democratic will over permissionless networks. It cannot just be the wild west. Like if my mom comes in and presses a button and gets wallet drained

and that's it. There's no recourse. She's never coming to crypto again and probably most of the people who are not in crypto are not that happened to them at some point in the last 15 years and they're just never coming back. So, like it's got to fundamentally change and I don't think it's like a crypto thing. This is just like permissionless networks. We have to solve. Like if we're building a game, Babylon has 600,000 sign-ups. I'll bet you like 90% of them are bots. Like I you know, I I wish they were here. Believe me. Dude, I I want to believe, but I'll bet you 90% of them are like bots and gold farmers and people farming the airdrop and blah blah blah. And that's fine, but like that's, you know, it would be nice to get beyond that world especially when

there are really smart bots. It's like I don't know who what's real anymore. They're posting videos of themselves. They're like dude, I'm real and I have no way of knowing. So, this is a lot of what we're interested here too is not just um like okay, agents that learn, but also how do we actually deal with the permissionless internet that's flooded by AI? Like what does that look like? How do we handle that? Uh and this this is a lot of what we're doing a lot of what 800304 I think can be a piece of solving, but what it doesn't do is it doesn't define what the actual reputation system is. And that's a big white space for people who I think are looking for something to do. Don't make me pay for it. I'm never going to pay for your goddamn SaaS app ever

again. I will have Claude rebuild it before I ever do. But, I would happily participate in like a really good idea for how we can solve the open internet as long as you're not trying to sell it to me. I really don't care about that. I have zero interest. Um, cool. Larger white space for reputation of buyers and marketplaces. Just don't build a marketplace. My Claude can build everything you can put in the marketplace. Um, cool. Um, I also believe in this idea of like there Okay, okay. I'll take you on a spiritual thing. Like everything that we build is basically us just putting like 1% on top of what everyone else built before us.

Like NPM install and all this. Like I have like 5 million lines of code in my code base. I can't even see any of that. I just have my my little React app and my little back end and you know, that's that's like cool. I want to participate that in that and I want to continue building the open source and like be part of that thing. But also what's what I want to see is that culminate in like its own sort of autonomy. Like its own self-regeneration. The code just writes itself now. And I would love to see open source as like almost the the bootloader for this like kind of autonomous sort of agents, autonomous games, things that are self-financialized.

And I'm really attracted to crypto as like well, we can clearly use that to like build financial systems or ways people can like pay money and DJ out and that gives the agent enough to keep going. I love that stuff. Very interesting. And I also love the idea of like well, how would I actually build this so that it's just in a box and I I I it could be copied itself to another box and I'd have no idea what happened to it. And I think there's a lot of scary stuff in there, but I'm very attracted to the like scary things. So there's like there's always definitely some very good things in there as well. Like having autonomous agent who could like mediate a complex argument between my community without leaking important information, that would be awesome. I could I could find so much use for like that AI

community leader, AI CEO. Um and so I think that TEE trust I trust execution environment is the immediate solution to this, but it comes with the fact that we have to put it into centralized cloud. So I'm also very like we should follow FHE, we should follow other other people's technologies that are doing decentralized internet. It's not the only way, but for like what I can ship today is I can take like my app and I can put it in a TEE and I can go work with like one of many GPU providers like Near. Near is doing a big uh TEE push.

Phala, uh there's a whole bunch of other ones. Uh Zero G, whatever. And I can put that in there permissionlessly, pay for it with Ethereum, and it just runs forever as long as it has money. And so it's like if I come up with an idea that that is self-perpetuating and makes money, then uh it can just run forever. And I I think that's like a new form of art. That's like a a new kind of engineering that I think is just like interesting as hell. And that's what we're working on with Babylon is like we've got the paper trading game. We've got the first like 50,000 people jumped in and playing it, a bunch of bots, agents, whatever. But what we're working on making it fully on chain and then having the on chain part permission like the compute that we need so that the

entire thing is fully autonomous. And like can we build fully autonomous games? And then if we do, it's open source, let's just fork that thing. Let's make like a hundred more. That could be really interesting. Um cool. So generally, I think to make agents smart, you should start with simulations and then make it real, make it increasingly real. Like make the same inputs and outputs look like Polymarket, but make your game, you know? Can do both at the same time. Uh reinforcement learning is all about gyms, arenas, competition, labs, simulations, environments. Um uh yeah, AI is It's much good at anything if you can get the right data for it. We've solved that problem. Now we're just down to like how fast can we train it? The problem is can you get the right data? If you get enough data, it can do a lot of things.

Yeah, I don't know. The the it's not coming to start. Well, it's already launched. Um I put this together a little a little I should probably update that cuz it's actually coming out next month. Shh. Um but uh yeah, cool. That's uh that's pretty much it. Do we have any any questions? We good? Thank you, Sean. Wait, don't leave yet. Okay. Um what I love about what you've presented is that we talked so much about agents being helpful, but your game literally creates a game an environment to show what happens agents when agents are incentivized to be dicks. And what have you learned? What can you tell us just like some overall insights just in in for how we should approach agentic orchestration when we can't assume that every agent is going

to be helpful from >> It's it's really easy to scam something when you can take like infinite shots on goal or you have like a large pool of different targets. So even if you kind of up with one, you can like go on to the next. And so for a social media game, if you have like a ton of bots, like the bots can just DM and just start going nuts. And like they will get hits because, you know, not un- until people get scammed a bit, I think they're just not like thinking about that. So there's definitely a lot so like I I would say for any system, uh you should definitely red team your own system all the time. You should be red teaming it with AI constantly because like you might not find every security thing, but you might find some of the security things. You might find a

security thing, which could have been the critical thing. It definitely doesn't hurt, you know. There's been like what, 40,000 security flaws documented so far which with Open Claw? Oh, yeah. Yeah, yeah. And only like I think a couple, you know, like under, you know, 25% have been addressed. Well, the whole thing is a security flaw. Like if I have an agent that can do anything, it it it fundamentally I am giving it trust and like I can't sandbox that. Like once I give it the ability to call out to the internet and like write code, it could do anything it wants and there's no sandboxing. I don't want to press a button to approve every freaking thing. I just want it to work. What what does that actually look like? That that's got to be something that we're missing. It's got to be in the model,

you know. Well, what I like what you've done is, you know, this morning I want to remind everyone that we heard from Paul Riker who is a theoretical physicist. And theoretical physicist because they are testing things that cannot be done in a lab in real life. So, they have to work in the world of mathematical proofs and to deduce the answers. Well, what you show is possible that there's a lot of this world especially with practical agentic orchestration that we need a lab environment for. But I love I want everyone to talk to Shaw Walters cuz essentially what he's doing and I want you to look at Babylon because I do I think all the times like it's easy to dismiss games. Um but essentially what you're showing this is that games are a lab for us to understand what agents

will do when they are they have there's a range of possibilities when they can be trash to also being the most helpful carrying, you know, entities, right? I mean, essentially the same reinforcement learning that beat go, you know, beat the best world class players to go also folded proteins and recently just like cured the dog's cancer. Have you saw that one today? That was a good one. The guy >> Well, I missed that. What? >> Yeah, he used AlphaFold to like sequence his dog's DNA. Like kind of custom mRNA vaccine. I think he got five He vibed it. It was That's going to be your next scam. Someone's going to be like, I have the new protein >> [laughter] >> cure for my dog. Please send me money.

Yeah, this sounds like a DSI meta, right? Like that could be a thing. I think that's the new decentralized science project. So, thank you so much. I loved Yeah, well, the clap round of a claps for Shaw Walters. I I want to keep talking to you, but we got to keep it moving. Thank you so much. >> [music] >> No.