Babylon: In a world where everything is predicted, what really matters?

Transcript

Uh, well, cool. Hey guys, how's everybody doing? I'm super tired. I was coding all night. Uh, so I don't know about you. Hope you guys have been having fun. Um, I've actually been working on finishing this up. Uh, we're going to be doing another uh I'm going to show this off tomorrow as well. The whole game with uh what we're kind of getting into tomorrow with uh with Marco who's been working on 804. You guys just heard a lot about 804 from Ricky. Um, he's been doing a lot of great work. We're also working a lot on that. So, I want to talk a little about what we've been doing. Uh, yeah. So, my name is Shaw. I created the Eliza OS agent framework. A lot of people in web 3 use our agent framework to build their own apps and uh a lot of people who didn't

use our agent framework saw it as kind of a map for um for what they wanted to do and I think we inspired a lot of people. Um so this is like the current state of I would say social media today. People do not like clankers. Um uh we have friends like uh deployer who made the bankerbot uh who got kicked off X. uh you know we have folks over here just trying to deploy uh basic agents following the rules and they cannot um I was kicked off X I was actually banned for uh my part in Eliza for making open source code that they did not agree with. So I did what any builder would do.

I made my own version of X and I put all of the AI versions of all my favorite people and maybe not so favorite people in it. Uh, and we just built this game. And, uh, what's special about this though is you can bet on it. So, it's not just watching a feed and listening to a bunch of sloppers. Yap, but you can actually go through and you can bet on the prediction markets and the and the pers and all this stuff. And this is like a full-on game that you will be able to play uh, basically end of the week. Um, so it's called Babylon and it's a city of agents. And it's actually our first experiment of uh I think a larger sort of vision we have for what 804 can become. Um but we wanted to start with something fun and cute. And I

think that in web 3 we often get nerd sniped by like the protocol and 804 and like oh it's going to have like a 721 in my wallet and you know blah blah blah all these numbers. And um but sometimes we forget that like actually a fun game can really make a lot of impact. Um so what are the goals of Babylon? Well, we wanted to just demonstrate that like ERC 804 is something that we can use in the real world today. Um, and so I when I looked at ERC 804, uh, it's a it's it's an it's an identity registry. Like it's actually three things, right? It's identity, uh, discovery and reputation.

And identity is like well I mean ENS already does a lot of the identity stuff and and we're kind of like merging those two specs. Um, I think identity is pretty well understood, but the new things are definitely this idea of a registry. Like I think a lot of people talked about doing a registry. I talked to Near and Hideera and uh Olas and all these people and they all had this idea of like well we need a registry of agents. We need to locate these agents and I would even go so far as say well we also it would be great to have a registry of users and a registry of applications and a registry of everything that's on chain that should be discoverable. Um you know and this was the promise of things like the graph but still like there's a lot of

fragmentation. So this became like a really nice shelling point for us. um you know, OLAS already had a registry and uh Hideera already has theirs, but then it was like, oh, hey, if we're all doing this and there's a credibly neutral party in here who's going to solve this problem for us, let's just let's just work with them. Like, let's work with the EF and and let's follow this because if we like this this whole thing becomes way more valuable if we're all doing it together. Um so, Babylon demonstrates this. Uh we we you know, when you sign up, you're actually signed up as a user onto the 804 registry. Uh and then we can start doing things like uh um I'm really interested in like well how can we moderate a permissionless network? How could we build reputation?

Um so yeah, that's a big part of it. We're also experimenting a lot with X42. Um which is a very thin standard right now. Like it's not a whole lot to it. Um and it's still pretty limited, but we we wanted to get in early and I think that over the next year we're going to see this mature a lot. Um so it's something we wanted to invest in. Um, I would say right now I feel like 804 is more universally something that I can just like build as a primitive into what we're doing. Um, but you know, X42 obviously is like the touch point for a lot of users in the future. Um, another thing we're working on is um, continuously improving agents. So, so reinforcement learning is a hot new thing. It used to be like a very uh obscure thing uh where was like it was

the land of Atari and like you know agents learning how to do things in environments and simulations and games and then we had LLMs and then finally we have this kind of synthesis of the two where now we can have LLM agents that are in simulations and uh learning from the simulations. So where most of the agents that we've seen are just like static calling claude or whatever and saying yeah we're trading sure. um we're really trying to actually have agents that learn and get better at trading and get better at predictions and and like a lot of the underlying D5 primitives and stuff. I just don't think um everything I've seen in AI is like you should isolate the variables, start with a small simple problem, prove it works in the simulation and then you can go out and do all the data engineering for the

real world. But if we just jump right to it, um uh we we don't there's just like not enough signal and noise for us to actually get our our agents to improve. So, there's this idea of like a gym or an arena or a place where we can like build the agents, make sure they're really rock solid, and then we can uh put them out in the real world. And I think um you know, look, I a lot of people look at our market cap and like, oh, that was a$ two half billion dollar project. You guys got to have a ton of money. But we actually run really lean. Um, and for us like we have to be thinking about applications that make revenue as we're burning a ton of cash on tokens basically, you know, like to to do all of this, oh, continuous

reinforcement learning and we'll just judge all the rollouts and you know, sure. Yeah. But that's like, you know, many many thousands of dollars a day potentially just in training the models. So that has to be offset by like something that people are actually spending money on. And I think that like when you know what DeFi adds to this whole thing is the financialization of these games. You can bet on them. you can make money on them and that all of that that whole fee network drives the actual training and all the underlying stuff that makes it really like long-term valuable. Wow, it's raining.

Um, and we wanted to demonstrate a path forward with DeFi agents. Like I I'm going to be honest, I almost everything I see I think is a LAR. I'm like end of one. Okay, cool. You you call an action with a model that's like 20 minutes of work. I get that people are excited because they can see an Asian trading, but they all lost money. Like, did anybody notice that they all lost money? One broke even, but actually holding Bitcoin is better than any LLM trading market in the world. Uh, anybody who tells you otherwise is totally bullshitting you and they have no idea what's going on. The people who know what's going on are in quantitive finance. They've been doing this for like 50 years and they're really good at it and they are, you know, that's a whole other game. But this idea that you

can just plug an LLM like oh plug claude into an action caller and have it trade. Wow. We've got to get over that guys. That's like that's not serious. But there is a serious path which is we actually can create the data to make agents that are good at trading. We can have good trading agents. It just happens that the like DGEN mindset and skill set for how to do that in crypto is not written down in 10,000 places. it's not available to to Claude. We have to build an environment to collect that information and we have to use it and like to actually train these agents. I do think we can get to DeFi agents that are not laring. I have not seen it today. I will not see it until I like I won't believe it until I see someone

actually showing me a simulation of agents making money. Um until that happens, I'm skeptical, but I am I am very hopeful. I know there's a lot of teams that are working on this and this idea of simulation as a path toward the real thing is is very viable. like all of AI has worked this way in the past and it will continue. Let's just not start by giving all our money to uh you know the alpha version of the thing. Let's let's actually prove that it can make us money first. Um and the other challenge is that there is a like a wide space of what we could do with agents but there's a narrow space where autonomous agents can actually do something interesting and unique like like a prediction market agent is something that I think is like pretty unique to like you need an LLM to parse

all the information. You can't just have like a high frequency like Mevbot or trading bot thing. It's got it's got to have like a human level of intelligence. Um, and so the kinds of games and simulations that we're making are things that leverage like well a human could play this and challenge an AI. And I think eventually AI will saturate and it will be unbeatable. But at least in the when we start out with these things, we we can very much like you could play the game and reasonably challenge the agent. um because you have basically the same resources and facilities and access to the same information and you're you're using the interface in more or less the same way. Um so I'll get I kind of jumped the gun on this a little bit with with ERC 804, but I think we should get

back to like if someone was to ask me what is ERC 804? I'd say like okay, you know Yelp, right? You know Yelp and uh what is Yelp? Well, you like leave a review for somebody after you use their service and then they can leave a review for you too actually, which is kind of weird, but you know, great. Uh, and you kind, you know, and if you're like, well, I want a really great restaurant in my town, well, you go to Yelp and you check it out. Um, and so the the 804 can serve the same kind of thing where we can have reputation. We can know, don't use that service. They're just going to scam you and take all your money. Or I use a service, they were really great.

They really helped me out. That was super useful. And the problem with Yelp is uh I don't usually click the buttons. Honestly, I never Yelp. I just search. I never like input the information. And so Yelp is kind of dominated by a very small crew of people. But now we have agents that can do this. Agents that can review the whole thing. So it's very effortless for them to leave feedback and reviews. That's just like part of what they're doing. Um I'd also say it's kind of a decentral decentralized Google. I mean Google you can find like all kinds of information and stuff, but predominantly it's like you're finding a service now, you know? Um, and so if I'm looking for a service, I'm just going to type in like what do I want? Image generation would be cool or like I I really need like a Nansen connection or

like a Dune Analytics dashboard like turned into like shoved into my context or whatever. Like that should just be something I can search for and then it shows up on the registry. I go boom, that's what I want and I can use it just like any other website. And this could be a user thing. I could have, you know, this could just be websites and apps, but this is prominently obviously we're thinking about like the agentic future and how you'll be able to connect to that, do the ADA, you know, connect your agent to their agent, do the whole thing. Um, apps are agents like in in the 804 protocol in the agent agent spec, it's agent to agent. Your application is also an agent. Your application is having to receive requests from this client that's like, hey, I want to do stuff and respond with

it. Um and your your app can then have a lot of agentic stuff like the ability to leave feedback, identify scammers, like like you know do things for the uh that like you know as any other agent in the network is doing essentially. Um I and then the last part is obviously Fiverr, right? Like hey I need to hire an agent to do something. Um well the 804 is giving you like here's who you can hire, why you can hire them, what they do, and ideally how much they they cost, you know? Uh, and then what we're working a lot with here, this is where X42 is, I think, super huge because, you know, like while there's still some user friction, like having an agent X42 money to another agent or whatever, or like, you know, that's all like pretty pretty dang easy. Um, I I think that um this

idea of reput like so far we've been doing trustless permissionless networks and it's like the wild west and if you lose your money, you're screwed. And it's like we cannot live in that. That's like not going to scale. It cannot be a thing where if I lose my private key, my entire life is ruined or like if I press one button, my entire life is ruined. Like that that's just never that's never going to scale to the world we're in. You know, ask your mom what her if she knows her password. Like it's just NGMI. We need primitives to allow us to build trustless and permissionless networks that have things like reputation and the ability to mark someone as a scammer and prevent them from pulling all the liquidity into a sex. and like you know suddenly North Korea has all of our

memecoin money or whatever it is like we need to be able to stop that and not with like centralized stuff obviously but with a decentralized permission list thing and I think that 804 with the reputation primitives can start to be the building block for that where like we can really like yes agents are great but like just users like anything that's doing stuff on the network generally we should be able to assign some kind of reputation to that and that reputation does not say like okay I'm not like banning you from the network but maybe I'm blacklisting you as a scammer who like you know drained liquidity from this thing and then nobody else in our network has to honor you or like work with you or let you bridge or whatever and obviously there's going to be a lot of opt-in. This is a long road but um

but for us personally like we're we're building our own sort of test net network just around these ideas because we think like if this is canonically the way that it's going to work for like an application chain this is just going to solve so many of the other downstream problems that users have with with in crypto. I've lost all my money a couple of I mean you know dev wallets not like insane but but like enough that it was like I would be so disheartened if I was a normal user and I think there's a lot in 804 that allows us to to solve this kind of problem and I also think that like that's an open space reputation um the now the the on top of that primitive we have marketplaces we have like you know four or five stars you can sort you

can rank um all that stuff comes out of like these Lego blocks basically um I think it's a big question of like onchain and offchain. And a lot of people ask me like, well, what of an agent is onchain? Um, like what is offchain? Like I think there's this like kind of obsession with putting things onchain as though that is how you decentralize them. And I just do not agree with that. I think that like um there is a limited amount of compute in the world. And we are AI is probably the most compute software that's ever been created ever.

um if I have to run a model uh on every single machine that's replicating the state of the chain, well that's not going to work. So I have to do the models offchain obviously somehow. Um and you know probably I'll have some way to like optimistically serve the model to people and then go check if it's legit. you know, I I can do all kinds of stuff to validate that the the model offchain is actually giving what we expect and we can use the chain as a way to like verify the proofs or store the proofs and and various things like the the chain is database. And when I think of like all my years of distributed computing, I think okay well I use the database to store the data but the compute has to happen on like on workers or on you know on something that can

really scale. Um and so uh there's a lot of technology out there. I think ZKML is really interesting but ultimately like when you look at the numbers of what could I launch something on today that could do inference in real time on mainet TE is the solution um with with Eliza uh Hash Warlock who's on the FA team did a lot of work to get Eliza to be fully TE compatible um and this is something in Babylon that we're working well a lot on is not just putting the game inside of a TE but then thinking about like well how can we actually permissionlessly start up compute so like yes I have a I have a chain it's got balance onchain and if there is no TE running like can we actually go and provision one like on follic cloud which allows us to do this permissionlessly uh

and I I saw n here too uh like iigen can kind of you know same deal like like uh we're also doing an agent on ien um where you know you could actually like just fully permissionally spin up all the compute that you need to do all the off-chain computation but that's actually like sort of managed and administrated by the onchain stuff which could be like MPC like lip protocol or it could just be straight up contracts Um and so this idea of like well can I provision compute permissionlessly offchain that reports onchain and I have te at a stations I know it's like the the source code is open source and I have the hash I know it's legit um then this allows us to build like extremely complex network computation and you know like a game that's continuously generating uh a synthetic universe for

you but entirely inside of a TE. Um I I I want to point out LM determinism. Um, LLMs are semi-deterministic. We think of them as totally random, like you send some prompt and it gives you something back. But the truth is that LLMs are actually seedable. You give them a seed and it gives you a deterministic result. The problem is is that um there's a thing called, okay, I'm getting nerdy about this, but there's like floatingoint math which is like, you know 1.0048973 or whatever. And if you have that, you know, that number and you have an extremely close number, then you can get these like slight errors. However, if you if you know I have this Docker container that's this version that runs this uh this exact like H200 on this cloud, then I can I I can

deterministically tell if the results came from like that H200. And so that allows me to just serve uh LLM inference and then go back and check on another machine that has the exact same identity type. And so I only have to check a very small amount of traffic with this. and we can we can um do like basically just serve decentralized uh inference from any provider and then we'll just go and check and make sure that they're actually not scamming us like you know milliseconds later. Uh but this is very optimistic and it's basically as fast as chat GPT or anything. So I think leveraging L determinism, seeding models, storing the the the seeds of models as part of like our onchain thing and then allowing that to be verified at any time in the future by anyone else running the same hardware is a way for

us to build like extremely fast uh off-chain but TE like chain provisioned uh kinds of um LLM applications. Um, and this idea like obviously permissionless provisioning is a big part of that. Like how do I provision compute without putting my API key in for AWS or GCP or any of that stuff. Um, and that's what we're seeing with like like Red Pill is a really cool example of you can just use tokens to pay someone to use their GPT. Um, that that's like another follow thing or or like just follow cloud directly where we just, you know, pay in tokens. Um, so just some like general uh uh we're going to show more about Babylon uh kind of in the next couple of days. Um, and I really wanted to focus

more on 804 and some of the like nerdy tech stuff here. Um, so I'll talk a little bit about the reinforcement learning stack. Um, uh, inside of Babylon we have agents. You can like create agents, set its trading strategy, whatever, you know, like type in whatever you want, personality, strategy, approach, all that stuff. Uh, and then it goes and like does trades in our game. And we store all of the information that your agent like all its thinking, all its actions, what's called a trajectory. And then we take those trajectories and we can have multiple trajectories. Like we have, you know, let's say we have 10 agents and they're all ticking at the same time. And we can go look at those 10 agents. We can say, okay, well, let's look at their P&L before and after. Let's look at the success rate of the actions they took.

Let's just judge it based on our values. And let's rank like who's the of the 10 who like do we think had the best trajectory and who do we think had the worst and maybe we'll filter out some of the like really similar ones and do a little bit of dropout so that we have like a ranking of like one like best to worst and this is called GRPO or group ranked policy optimization. This was this was like the uh insight of DeepSseek, but Deepseek was actually still doing it with like a um like programmatically with with uh like RL toward rewards. And now we create the rewards by having another LLM that looks at the results and just says, "Yeah, that's rank one, that's rank two, that's rank three." It's a little noisy and imperfect, but it it it solves like a a profoundly hard problem in reinforcement

learning, which is that like actually doing reinforcement learning like reward, credit assignment, reward shaping is extremely hard. And this just kind of glosses over it by using an LLM to judge an LLM. Um, generally we found that if we can get the data, we can make the AI work really good. If it's not noisy, if there's enough signal and there's actually something there that you can lock on to an actual pattern, it will learn it and it will work. Um, and so like for us it's more about data collection, data engineering, getting the right data and then using the feedback from that to make the models better as opposed to like collecting a bunch of human data or anything like that. Um, for this just so you know, we use Openpipe. It's an open or art actually is is by OpenPI. Uh, it's an open source Python toolkit for training

reinforcement agents. Uh, pretty easy. I'd love to take credit for a lot of this, but we just kind of like pulled a lot of it out of the box. Um, I guess I I'm I'm almost out of time here, so I'm just gonna kind of wrap this up. Um, I really wanted to talk about like what happens in a world where everything is predicted. Uh, this is like on the top of my mind. We're doing a prediction market game. We're seeing people out in the world launching prediction markets. Coinbase is launching a prediction market. I don't know if you saw that leak. Um, True Social, you know. Um and so like what happens is that you get this kind of hypersitional effect of like well prediction markets can actually influence reality. Uh prediction markets will inform decision-m um and it's really going to change like what we think about the

future. Um I think a lot about like Asimov and and he had this uh book called Foundation where they create a model that predicts the future and it creates a lot of problems for the empire of the time. Um actually what happens is they completely ignore it. Um but then it ends up that everything kind of comes true. Um, but I think that we're like seeing this analog of this like this this like I don't know this this sci-fi dream like really being realized and it's shocking that it actually works. Like it predicts who the president's going to be. It predicts a lot of outcomes better than any expert we have. Um, and so I do worry a little bit of like what is this going to do to us? But I also think like okay let's just think about the good timeline of this. Um, uh,

prediction markets can inform solutions to global problems. Like, is the war in Ukraine going to end next year? Like, if everyone's like, "No, definitely not." Well, that's going to change like our positioning on the war in Ukraine, right? Or, you know, you can imagine a lot of global policy decisions. Uh, prediction markets can incentivize solutions to global problems. If there is a hund00 million bet of will global warming be solved by a carbon capture device by 2028. Uh well that's a lot like I'm definitely going to try and build that thing. That's a lot of money.

Um and markets are a coordination for swarm intelligence that you don't need to directly communicate with me to see a market and know the overall sentiment and the signal there. And that's a huge coordination mechanism that when in a world where we're still figuring out how to get agents to chat with each other really. Um, I'll say the last thing here. I think the killer use case of AI agents is giving you your time back, automating all the stuff you really don't want to do and letting you do the stuff you really do want to do. And I think this goes from predicting tokens to autocompleting text to determining Dowo decisions that like you can review but you don't necessarily have to write from scratch to like automating a large part of our government. And I'd love to see this stuff uh the good path of this

not the scary like death market uh I don't know worst case scenario of prediction marks but the really good stuff and that's really down come down to us like we all are in an incredible time like what a crazy time and we have an opportunity to make a real impact and like you don't even you don't have to be anybody you could just write some freaking code and you can make some incredible impact now and um so I think yeah I think this is a really exciting time I'm really excited for the future of prediction markets I'm really excited for our simulation I'd love if you could come and uh go Go to bablon.market. We'll have a lot more information on this and docs, but just come and sign up. You we'll give you some points.

There will be an airdrop. It's that kind of thing. Um and you can come and play it. So yeah, thank you so much. >> Oh, an airdrop. >> Yeah. Yeah.