A (3:17)

So. So I think in crypto we have a slightly different problem and, and Tay, maybe, maybe you can address this, right? It is certainly awkward if you have a hundred thousand Linux servers and like, you know, load balancers are running bsd and it's like we don't want to take this thing down because if we take it, it's not easy to replace this architecture, right? But it's just software. You can in theory patch it, upgrade it and do a hot swap or something like that. These are systems that are not necessarily designed to be immutable. What happens if Mythos is like, oh, uni V2, I've cracked it.

B (4:05)

Yes, that's going to be a problem, a massive problem. And I mean, I think this industry has realized this over time that immutability is like freaking amazing in so many ways, but also quite impractical and, and creates risk. And I think this was actually a big conversation during the Drift hack, right? Immediate aftermath because all the Ethereum people were like, just don't have keys, don't have upgradable contracts. I get it. But at the same time, like I'm not convinced that we are. If you're launching something today, I'm not going to bet that it's perfectly secure forever. I'm sorry,

A (5:06)

Uniswap is the counter example to this that has given people more hope than it probably should would be my hot take on this. Uniswap is amazing software, but like it's like one function. Like it literally like it just rebalances a thing, right? Like it, it, you know, it is not a super complicated multi contract, smart contract suite. Like you know, it, it's like for what it is and yet like my confidence to its unhackability has gone down significantly over the last like 48 hours, right? Like if, if you know, people like these are pieces of software that are securing like billion dollar systems with people getting paid tens of millions of dollars a year to secure them. And of course Uniswap has had this giant bug bounty, if you can hack it, then. So it's not like we haven't had security pressure in the ecosystem, but it's a different level, like a different level, different scale to what some of these systems have had and yet there's like zero days in them. So we talked about this, we talked about this like maybe a month ago, Tay, right, where I was like, how confident are we that someone doesn't have really smart like fine tuned model that they've been working with that's like systematically going through and like breaking these smart contracts? Like, given what they've said about Mythos, does that change your perspective on the last like couple of months of hacks or you think it's still just a coincidence and it's going to get much worse?

B (7:03)

No, I don't. I do think it's going to get worse for sure. I think we're in trouble. But I think, I still think, okay, Mythos is like, it's a level up and I think that they have. It seems like they've specifically been thinking about security and been like, like the people that really know this stuff have been zoomed in on, on the security elements, right? The exploit elements, the safety elements that come with the security elements.

A (7:38)

It wouldn't.

B (7:39)

I mean, obviously there are going to be people who are like just random people that are doing the same things, how far they've gotten and how like, what's the most productive use of those things? I feel like, I feel like it's not. I don't think we're at the point where they've like honed it in so completely that they can just like let, let it loose, right? Let their little hacker claw loose on whatever. But I do think that we've seen like AI assisted, right, where the AI is identifying things or speeding up those feedback loops, right? Because when you're fighting these exploits, you have to like, there's like so many different sort of stages of finding the exploit. You have to like hold all this data or you before you had to like hold all this data in your head and like understand it. And then even once you figured out like this is a weakness, you then had to execute it, right? You had to actually put together the series of steps that would do it without fail. I definitely think that we've seen AI like rapidly like assist with, with some of those stages. I think Mythos is scary because it sounds like they're like this, this thing, I'll just go hack you.

A (8:58)

You don't even need. It's not assisting you anymore. It's doing the thing. Right.

B (9:03)

Yeah, exactly. And that's what scares me, especially because Uniswap is also unique in the sense where they do have. It's a protocol, it's low level. Right. They do have a huge bug bounty on it. Most of, like the core software that runs this world is not built by. By people with million dollar bug bounties. Right. It's like, fair.

A (9:26)

Yeah, fair.

C (9:27)

That's why I think this is going to be worse for web 2 than it is for web 3. I think that even though we have money right behind one line of code, someone joked that the WETH contract got hacked. Right. When there is just a smart contract that has a deposit and a withdrawal function is basically you're talking about hacking the protocol at this point. If there's something that goes wrong there.

A (9:52)

But sorry, Austin, are we not worried about that as well? Like, you know, we sure like validators.

C (9:58)

Like, I mean, this is the more complicated piece, right?

A (10:03)

Yeah, like, validators are. And like, that would be. I don't even know, like, or even just say, hey, Mythos, buddy, I've got some interesting news for you. There's this thing called mev, and if you can like, just listen to this stream of transactions and just tweak it a little bit, we can make a lot of money. And Mythos is like, yeah, I love. I love breaking things. Let's go.

C (10:32)

I do think, yeah, so we're talking about breaking things and making a lot of money, and then MEV is just like being smart about ordering transactions and just like playing the game of. Of like the Mempool. I think there's a ton of money to be made by just like running these awesome frontier models. I don't know what we'll see. I think it's going to be really interesting. I think so. I love Uniswap. I think Uniswap is beautiful and it's like really isolated versions and it's not upgradable. Like, the way that Uniswap V2 functions is like this beautiful hyper structure. And if V3 gets popped, like, V2 probably won't. And if V2 gets popped, V3 probably won't. Like, there's not a lot of like, cascading things that can happen from system to system. So I don't know. I think it's going to be worse for Web two than it is for Web three. And I think that in Web three there are people who have deployed their systems and done tons of bug bounties and tons of audits and I think that's going to be good for them. But of course, like we saw the balancer thing and the balancer thing really changed my mind. That was when I was like, oh gosh, like that's

A (11:45)

this thing. Yeah, this is dangerous. And, and you know, that was very unlikely to be mythos, realistically. Right. Like that was very unlikely to be mythos.

C (11:56)

That was probably a dude with Opus.

A (11:58)

Yeah. So. So I think the other thing that's interesting about mythos, like, you know, just zooming out above the like security risks and all of that stuff that we see, this is the first time that we've seen, at least to my knowledge. Right. Like we've seen like, you know, early previews, but like most of the time they just let the thing out in the wild. They're not that scared of it that they're like, we'll like give it to people, you know. And like if you read they released the agent card, right? It's like this like 200 page document that says all the things they did and like basically ass covering. So when it like birds the world down, they could be like, no, no, no. We tried to like stop it from doing this. So, so they, they, you know, for previous models, they give it to researchers, they give it to a bunch of people, they do a whole bunch of things before this thing gets out in the wild. They do internal testing, et cetera. This is the first time, at least to my knowledge, that I've seen a model that's so smart, that they're so scared of it that they're like, we need to give this to a small group of like 12 partners. So, you know, AWS, Apple, Nvidia, the Linux foundation, etc. And say you're allowed to use this kind of for security purposes for the next however long. I don't know if I saw how long until it finds all the bugs. But like they give it also, you know, how expensive is this thing going to be as well? Because they gave them a hundred million dollars worth of usage credits. Now I'm an inference enjoyer myself, but that's a lot of that. Like it would take a lot to burn through a hundred million dollars in credit. So like, I think they're just imagining these guys are going to run automated loots. So you know, one, one thing that's kind of begs the question of, is like Boris Journey had a really good podcast a couple of weeks ago where he was like, you should be building for the model. Like we internally, we build for the model Six months from now, right? Not for the model today. So it's six months from now and the model is Mythos, right? Which could be anytime in the next few weeks. And all of the things that we've been building around, like harnesses and loops and things, all of that stuff could just fall away because mythos is smart enough that it's like, no, no, it's like, you don't need to put me in a weird, like, looping thing, bro. I can just like, you know, do this for two months straight or something. Right? You know, because one of the. One of the things is like this long running activity, you know, kind of frontier as well, right? Like that as the models get smarter, they can do things autonomously for longer without losing their minds and stay on task or whatever. If Mythos can just sit there and read the entire OpenBSD code base for two weeks autonomously, obviously that's going to be very expensive. But it also means that a lot of the stuff that we've been doing around trying to get agents to do what we want and not drift just sort of disappears, right?

C (15:15)

The harness is getting like, I spend a lot of time going back and forth between using frontier expensive model and then like, how do I do this in a cheaper way? And how do I use the frontier model to build like a more advanced harness that then can be used on a cheaper model to do the same thing. And it's like using the frontier model to then like grade how well the cheap model does. But it's all about cheap stuff, right? What you're saying is basically they'll give us Mythos and it will cost me $100 for one request, but it will one shot Uniswap V2, and it will make no mistakes, right? And so I think that is what's coming. And that's really interesting. And we've seen this in AI for years now where they were making all these specific things that could do text to image to text and things like that. And then like a generic model comes out and the generic model can just like look at the image and do it better than any of those hyper specific models. So I think we're always going to have like this generic model that's hella expensive but really good. And then we're always going to be kind of like working backwards, building our own harnesses and trying to find like cheaper ways to do the thing that the expensive model can do. And it's funny going back and forth between it because it's always like, oh my gosh this is amazing. Oh, my gosh, it's so expensive. How do I lean back on something cheaper? And then you use the cheaper one for a little bit. You're like, oh, my gosh, this is so freaking stupid. Why is it doing this? And then you go back to spending money and, like, I'm constantly oscillating between these two states.

A (16:49)

Well, so. So I think there's two things here, though. One is this is the last time we had new, at least in anthropic land, right? Like, we had Codex, which was, you know, late last year. But Opus first came out, like, middle of last year is my recollection.

C (17:09)

It was not great. Like, Opus 4.5 was not great. Like, I was using Sonnet 4.5 to write most of the code. I've been. I haven't written a line of code for, like, two years, right? And it's been mostly like a lot of you're absolutely. Rights and a lot of apologizing and a lot of working backwards. But the. The real turning point was Opus 4.5. Sorry, maybe I said that wrong. 4.4 was bad. Opus 4.5 in November was the real

A (17:32)

inflection, was the one that, like, went across the chasm into, like, oh, this thing's better than me at this. This thing, right? So. So, you know, 4.4.1 came out opus 4. And. And they changed the model name. You know, like, Anthropic has done some weird model naming things over. Over the years, right? Where, like, Sonnet 4.6 was supposed to be the next version of Opus. It was supposed to be Opus 5, but then they've got Mythos now, so you can't really keep track of all this stuff. But it has been a while, and I do remember when I started using Opus, and I was like, oh, this thing is actually super smart relative to the baseline of that period of time. Before that, I'd been using 03 mini 4 OpenAI. And those things were really dumb, right? Like, really, really dumb. But so this is the first time that we've seen, like, a big model leap as well. Like, you know, they need a new name for this. It's so much smarter than the old one. And the needing a new name thing is, like, always a little bit of a harbinger of, like, ah, shit's gonna get crazy, right?

C (18:47)

And the name they used, right, Like, Sonnet is like a home, and Opus is like a body of work, and Mythos is like the lore, a God world.

A (18:56)

Like. Yeah, yeah, exactly. Yeah. So. So I think I Think, you know, like they're scared of it. We have obviously, clearly, like, they're petrified of this thing, right? Like so. So they must have some really good NDAs in place as well, because we haven't seen too many leaks from people inside these orgs, right? Like they're, they're doing something to kind of lock this down, clearly. So they also. I think the other interesting thing is that like, we, we've seen every time that a new frontier model comes out, like Haiku today is as smart as like Opus 4.1 is, right? Like different context, different constraints, but like, it's super smart. It's actually really good. And I mean, to your point, Austin, like, with the right prompting, you can get Haiku to do stuff that previously would have required Opus, right? Like, so the interesting thing now is how this shifts the cost curve, right? Because at the moment you've got, you know, when, when Opus first came out, it was like $75 per million tokens and now it's dropped down to like 25 and it's got a 1 million context window. So the prices have been collapsing. So we have to assume like Mythos is going to come out, it's going to be $100 per million tokens or something like that, or $200 per million tokens are playing insane. But what that will also do, I think is push the cost curve down so that now probably Sonnet is the same price as haiku. And Haiku might be like almost free, right? And so now that like, you know where in the cost curve you're optimizing becomes really critical for certain tasks. Because if you can get Haiku to do it, it's like sub cent per interaction. And now all of a sudden you've got a really scalable thing that like previously wasn't scalable. Right? So this is the other challenge I think, in even if you're building stuff for yourself, finding that optimal point. So how have you been thinking about, like you say you like, go back and forth between like the smartest guy, the dumbest guy. Like, how do you think about this in the stuff that you're working on?

C (21:14)

It's. So it's a lot of like using Opus to build, like using the general purpose model first. So like I have a smart contract platform called Left Cloth Services where people can hire it to build apps or audit things or add a feature to an app. And I just have like a general purpose agent that's using Opus and a little Sonnet and it'll go to the smart Contract, pull down the job, and just do the job. And it's very good. But then it's like, frick, that was 20 bucks. Or something like that. How do I get that down to $5? And then it's like, well, it's doing. You know, I'm. It's. It's funny. You see, like, Kanye pulled up to McDonald's in the giant lifted truck, and it's like using Opus to. To do whatever. Like using Opus to do a git pull on a repo. Like, you pay. You pay a dollar to pull a repo down because you're using Opus to. To do that. It doesn't make sense. So then it's like, okay, let's lean back and figure out, like, where do I actually need Opus? I need Opus when I'm writing the SMART contract and when I'm auditing the smart contract, and maybe a little bit of front end stuff. And then I can use Sonnet for some of the front end stuff, some of the documentation. And then I've been using minimax M2.7 for, like, almost everything else. Almost all. All the things that are just like, go run this program. Go do this, and it messes up.

A (22:26)

That's a local model that you're running.

C (22:28)

It's not. No, no, no. It's. I think it will be, but it's not. It's like, through. I use it through Banker, but you can use it through a bunch of different. It is like an Olama model, but I don't think it's something that released yet. Is released yet. It's one of those. Yeah. Where it eventually will be local, but right now it's like, on someone else's machine. But it costs, like, a hundredth of what it costs to do with Opus. And it's pretty good. You know, it's a little weirder now and then, but yeah, it's okay. So it's like I'm always leaning back on that, and I'm trying to work through, like, you know, what cost $20 here? How do I get it down to two Opus calls with just the right. And context is important too. Right? When you're having this ongoing conversation, the context is building and building. Each call, each consecutive call is getting more and more expensive. So what you want to do is kind of have that consecutive call, that conversation to be over in, like, Minimax M27 land, where I'm paying a fraction of a cent each time, and it grows to be like two pennies over the whole conversation. And then when I go to make that opus call. Have Minimax give me just the context I need. Like here, here is the exact context you need to write this smart contract. And here are the three things that you kind of do wrong sometimes and have that all in a very isolated window and have it just like one shot, the contract. And then maybe you use Sonic to say, you know, like, here's the three auditing rules and here's the contract. So it's, I think it's, it's just like fine grain, right? So you use the general model and you do the thing and you get it to work and you almost get like product market fit and then you go back to, you know, making the whole company more efficient after you have product market fit.

A (24:09)

So let me ask you this, right? Like that kind of, you know, all of the things you've been doing, you've been building a bunch of projects and you know, I want to talk about eat skills for a second as well, but doesn't all of that kind of optimization lend itself to like building a thing that auto switches models and stuff? Like, have you thought about that? We don't really have anything like that yet. Where it just adapts to what you're doing?

C (24:40)

There are some, I think, I think maybe Clawrouter was one of them. That came out right about the same time I was doing my cloudbot stuff, I think claw router. But there's a couple of these routers where it looks like an open AI grade standard endpoint and you hit it and then there's a little inference model in the middle there that says how technical? How hard is this question? And then it sends it to the right model based on that. It's slow, it doesn't work. I haven't done a bunch of cloud router stuff to know exactly if it's that good or not. But for me I've tried a couple different things. In trying to do some triage on the way out and back in, I found it easier to just let the generic model do the thing and then go back through and look at, here's all the steps, let's break into the steps and let's break it into which of these things is going to be easy and hard and then from then on, kind of having it more hard coded, I guess is what we're thinking. It's more like a hard code triage process versus a dynamic triage process. But like, this is the future, I think, I think you will have a local model running in your VRAM that can triage the stuff on the way out and then go to frontier models based on how complicated. The question is it's routing it eventually. Yeah, but I don't have that working right now. It's all hard coded.

A (26:04)

Yeah. So I have. I have a. Luckily, before. Before the.

C (26:10)

You got like, a 512, right?

A (26:12)

I do.

C (26:13)

That's amazing.

A (26:14)

I've got a 512 gig Mac studio, and I've played around with a bunch of models there.

B (26:22)

I've.

A (26:22)

I've actually got a bunch of old MacBook Pros as well. I spent, like, two weeks trying to get an exo cluster running nice. And I finally got it running. It's got, like 800 gigs of RAM and it's just so slow.

C (26:34)

Exactly. You said you wait forever. You make like, 10 queries to Opus while you're waiting for one to come back.

A (26:40)

Like, you know, you think Codex 5.4 is slow. This thing is like, just like molasses. And so it's. It's too slow to be kind of worthwhile. I need to. I need to hook up. I need to try and, like, speed this up, basically. But I've also found that, like, the local models have been getting better and better, especially specialized models, like using a specialized model. But the other thing that I think is super interesting is, like, these skill files that are starting to emerge, right? It's like the Matrix for agents. And so the experience that I had with EAT Skills, like, you know, you were on the show, you talked, I built EAT Skills. And I'm like, I've seen humans over two years go through the, like, awesome. Like, yeah, you know, like, and. And get much better, right? It's just. It's like, very. A very slow process. And you're like, oh, this guy is, like, smarter. Smart contracts. Now after, like, two years of reading these documents, right? The agent, you literally pull down this skill file. This was like, the craziest thing, right? I was like, hey, bro, like, there's an E skill thing. You should, like, look at it, right? And. And Opus is like, yeah, sure, whatever. I'll have a look at it. Pulls it down, and then it was just like, instantly, like, one shot. It was crazy. How much. Because, you know, the thing. The thing is, like, all of these models know about crypto. It's in their training data, right? Like, they know about crypto, but they have, like, a cursory knowledge that's like, just enough to be dangerous, right? Where it's like, oh, I know how to do this thing, or I know how to do this thing. Eat Skills just like, Hones it down to like a fine art where like all of a sudden they know all of these like very nuanced things about like how contracts work and deployments and stuff. So, so, you know, you. I haven't. What. I haven't tried. I'm curious if you've tried this. Like if you give minimax or, or like even like an even dumber local model that you've tried eat skills, like is it fine tuned enough that like set of skills that they can actually do a good job or is it still kind of too hard?

C (28:53)

I wouldn't put money in it yet. It does some good stuff. So the stuff that my cloud, like Claudebot has deployed smart contracts using e skills that have like a quarter of a million dollars in it right now. Like, Tay is gonna hate this. I have deployed like through my bot deployed contracts that no human has looked at until they're already on ether scan that have a quarter of a million dollars in them right now. Like, like there's probably like three 50, 400 total US dollars in token across all the like 80 or 90 contracts my, my bot has deployed. So I'm using the frontier.

A (29:38)

We found you a new target. Mythos.

C (29:40)

Yes.

B (29:40)

I was about to say, how do we get Austin access to this?

C (29:44)

Yes. Yes. Yep. Yeah, I think we should give like Danny Ryan access first if I have to give it to someone.

A (29:50)

Right.

C (29:51)

Like, let's get the protocol guys first and then some of them. But, but okay, so the skill files, first of all, like, if you out there are building anything and you're expecting agents to use your anything, you should have a skill file. And this skill file kind of does this like, I know, Kung fu thing exactly like you're talking about, like Neo in the Matrix. They, they plug him in, they hit a couple buttons, and the dude knows how to fly a helicopter or whatever. Right. And it happens instantaneous. And this is what you're doing. You're. You're giving the model all the things it needs to know. And it's basically the delta between its training data and reality. And these models were trained like on a lot of 2023 data. Like, it's just stuff happens so fast in, in Cryptoland that you have to kind of fast forward it through all the things that have happened.

A (30:39)

And that's what Balancer is safe.

C (30:42)

Yes.

A (30:42)

Well, there's the safest balancer.

C (30:46)

Man, that's too bad. Yes, yes. And they also think it costs $40 to deploy a contract on Ethereum. And they think there's a handful of Things that they just like hallucinate. And they always do this. They're always. They're so lazy. And pattern matching and hallucinating, like, they

A (31:02)

could go look like you could do

C (31:03)

one web request and come back and have the answer. And it's like, nah, bro, I don't want to look it up. I'm gonna like guess at the route and then hit the route and see if it works. Like, whoa, why are you so lazy? But okay, so Skill files, you, you out there. If you're gonna have agents using your shit, you should have a skill file for it. It should be whatever your website is, dot com SKILL md. And it's just like the. The robots TXT stuff. And for a while we did LLMs txt, But Skill MD is kind of like the. The canonical route now. So everything, everyone, everyone that's building anything right now that agents will do on behalf of humans, like your flight tracking website should have a Skill md because really soon we're going to have AI agent travel agents that are doing all this stuff for us. And these skill files, get them over that. That delta of you were trained in 2023. Here's all the stuff that has happened since. You know, it's almost like waking them up and programming them with all that. Like they were in a coma or something. And it's like, here's what you missed over the last two years. And then it's really good, like you said at one shotting smart contracts, knowing what the latest tools are, knowing where to go to look things up. And. And then, okay, so I used an agent to help me write a skill file. And this is. This is like a first mistake that you have to be very careful with, because if an AI is hallucinating this stuff from their own context, sort of like from. If. If they know it by heart already, it doesn't belong in a Skill file. So you have to do this like, constant phase of interrogating each piece of knowledge inside your skill file to make sure that the agent doesn't already know it when you wake it up. So there's this process of just like feedback looping where I'm having the agent take any little piece of knowledge, check to see if the LLM actually knows that, and then remove it if it knows it so we're not polluting the context. And then the next loop outside of that is then use the skill file, give it to an agent, tell it to go build something, see where it screws up. Oh, you, you. You deployed the contract, but you didn't verify it on Etherscan. Why didn't you verify it on Etherscan? And then there's this whole like you have to interrogate the agent and you get to the bottom of it and it realizes like it was lazy. It like knows how to do it, but it just didn't want to, right? If you could just have a skill file that says don't be lazy, that would be great. But it's just not going to do that. So you have to be like really strict about certain things. And it's. And it's like almost if you read through each skills, it reads like you are gonna mess this up. And here's how you don't mess it up. And it's like you mess. You always get this wrong, don't get this wrong, do it this way. And it's like a whole document.

A (33:46)

You always find an agent that's like, not me, bro. So I got a funny story about this, right? So like I've got eight skills that's in my, in my system so that all the agents can pick that up, right? And agents are just like people, right? And this story is like one of those ones where you're like, it's just found the exact same failure mode that we always find, right? So we have been building all of these agentic tools at Infinix for a while now. And I've been tuning one to do like degen yield farming. Farming, right? Concentrated liquidity farming on aerodrome. So I have an agent. It's like a cluster of agents, right? Like it goes out and it looks at all of the yield farms at the moment and it like does some back testing and sees, okay, this is a good opportunity. I'm going to go and do this one, right? So it has a hard rule that it has to simulate the transaction with blockade before it does it, right? And so every transaction before it does it, it simulates the transaction, right? So like simulates the deposit, simulates the staking, simulates all these things. So it, it one shots a like five transaction 4337 bundle, right? It bundles up all these transactions which is like we as people like, oh, I should do transaction bundles but I'm too lazy to do it. And these things are like, oh no, it's fine, I can totally. So it like does the swap from. It swaps from mainnet to base one shot that then it creates this bundle and does all of these intricate swaps or whatever. Gets the NFT LP token, right? And it had already known that it couldn't bundle it because it needed the id. And so it was like. And it's like you're getting more and more confident that this guy knows what he's doing. Right?

C (35:45)

That's when they trick you. Yeah, exactly.

A (35:47)

Like, like they, they trick.

C (35:49)

It's like magic. Magic. Really dumb. Yeah.

A (35:52)

So then, so then the agent goes, okay, my next job is simulate the staking transaction of this nft. So it simulates the transaction, it fails on the second step. And the failure is you're not the owner of this nft. Right. And the reason why it gets that failure is because it hasn't authorized the NFT to be transferred yet. It missed the step. Right. And it knew it was supposed to do this, but it missed the step. And it goes, oh, okay. This is not the optimal way to stake the NFT in the gauge. The optimal way is safe transfer from. So I'll just work around this and I will do safe transfer from. Into the gauge. And it does it and it's. And it's gone forever. And it's like sitting in the gate and like. And it's funny because you go to the gauge and what do you find? There's like a couple people that have

C (36:50)

sent some stupid shit in there. Yep. Like, and like a thousand dollars of, of USDC or something.

A (36:56)

Yeah, exactly. And I was like, I was like, how did you get this wrong? Like, you're supposed to simulate it. And it's like, yeah, I did. I did simulate it. It didn't work. So then I simulated the safe transfer from. And that did work. And so I was like, this is the right. And you're just like, oh, man. Like, what are you doing? Anyway, so, so the funny thing is now I'm like, okay, you. If you're ever doing a staking operation, like, so this is like this iterative process of tweaking the skills. And so now it has a rule. It has to simulate being able to get the money out. Simulate the full round trip. If you are sending a token somewhere, you cannot do it until you're. Until you can prove that you can get it out of the contract. Right? And it's like, oh, yeah, that's actually a good rule I should have.

C (37:45)

And so that needs to go in a skill file, Right? The way each skills came about is me doing the same iterative process of building, finding something really dumb, telling it it's about to do something really dumb and not to do it, and then just like keep doing that until it quit doing most of the dumb things. And it was like, hey, everybody here's a skill file. But what you need, you need to write like a skill file for base def or something like that. Or even like this. This process of trying transactions and making sure the. The tokens in equal the tokens out. Like, that should probably be some kind of. And there's a lot of people doing this too, which is cool. And a lot of them just in the crypto world, this. This is another thing that a lot of people that I talk to, ecc, they're fascinated with the fact that people are writing these skill files, which are basically a business. Like you.

A (38:34)

You.

C (38:34)

You can take a skill file and not give it to anyone and make a lot of money off of it. And instead, like, people are choosing to give these skill files away. And it's really their secret sauce. Like the. The skill file. It was actually Mike radar relay that said this. This wasn't even. But anyways, the skill file itself is the. It's the secret sauce for the whole company. And you're giving it away. And it's just kind of like what we see in crypto is people just like, figure things out and give them away freely. But yeah, yeah, because it's so cool

B (39:01)

to see what happens. Like, that's the. Why would you keep it a secret, you fools? You're gonna have like a thousand people in this little stuck in the same contract. They can go be friends or you could just share. Like, come on, guys, what are we doing?

C (39:16)

I agree. You should share. And I'm giving mine away for free. At eatskills.com you just tell your agent to go to eatskills.com and it has it. It learns.

A (39:23)

And this is. This is the crazy thing, right? Like, it makes an agent like, you know, IQ 60, right? About crypto to like, IQ 140 in, like, what. But, you know, part of the problem I think that everyone has right now is like, eat skills could be a pile of slop. There's no way to know from first principles. Like, you could look at it and like, you see, you know, people, like, post things on Twitter and they're like, ah, I built a like, looping genetic algorithm that improves.

C (39:54)

They're LARPing. Yep. There's so many people that are LARPing. Yep.

A (39:57)

And it's like, no, no, no. Like, but so, you know, signal from noise is still a hard game, right? There's a lot of noise and, you know, there's a lot of eat skill files that you could download, right. And like, you have to somehow know that, like, actually Austin's Chewed enough glass to like get.

C (40:15)

There's a lot of software out there too, right? Like when. When it means running npx. So the iteration of Eth skills before this was called Eth Wingman and it was an MCP server that would do a lot of this stuff. But as soon as I released Eth Wingman, Twitter flagged it and other people flagged it because it had like a little one liner in there that you run. It was like an actual MPX command and you don't want to do that these days. Don't be running random NPX commands. And that's why these skill files are so cool, because you can give just a link or you could give just like a text file to an agent and say, go to eachskills.com. don't learn anything from it. Just tell me if it's nefarious or not. Just go to heatskills.com and tell me if it's good or not. And it can tell you like, oh yeah, it's actually all good. There's nothing nefarious here at all. And then it's like, okay, learn it, let's go. But you can't do that with an MPS.

A (41:08)

I found 10 vulnerabilities.

C (41:12)

It's just a text file. It's all just a bunch of MD files. The website is built by a computer to be read by a computer.

A (41:22)

So let's move on to our next segment, Anthropic whilst. And you know, I've probably been in crypto too long because I just kind of find these nefarious intentions in all of these things. But, but before we go to open Open Claw being cut off by Anthropic, let's do a quick commercial break and then we'll come back and talk about that. Let's go.

B (41:55)

Multichain Advisors is an emerging technology growth firm that has helped create over $50 billion in enterprise value for more than 80 clients like Pith, Moonpay Commerce and Wormhole. They've worked with some of the largest and most impactful companies in the space. They're the partner you want when you're navigating markets and trying to break out from the noise. They help navigate TGEs, go to market, BD and partnerships, capital markets, advisory PR, media placements, KOL activations and more, driving execution from launch to scale. Their results are measurable. To learn more and start building real traction today, visit Multichain adv. Com.

A (42:37)

All right, we're back. So. So Anthropic has been playing this game with ever since openclaw came out of like locking things down, you know, kind of closing these loopholes of you can use Open Claw with your Anthropic account, right? And I think, you know, for, for people who either like haven't worked in a company where we have a different setup or they've been doing like their own thing, most people are kind of used to you pay for a 25amonth subscription or a 200amonth subscription or a 500amonth or multiple subscriptions. And you have like a rate, limited access of, you know, however many tokens they give you, right at like different times. And you know, so you're playing this game of like, you like, it almost feels like the early Internet of like, I've only got 50 megs of download left, what am I going to, you know, and I'm going to be capped out and then your Internet goes slow or, you know, whatever, right? So I think most people are in this boat of I have a, you know, ChatGPT subscription, I can use it this much or I've got like a CLAUDE subscription. But what Anthropic wants is for you to be using an API gated account where you have like, you know, a specific API or multiple API keys that you are using and it is all kind of on tap inference and you're paying per call via APIs. So I'm curious, your setup, Austin. Are you using API mode or are you mixing and matching with accounts? How do you manage all of this as Anthropic has been closing the door on using your Anthropic account in other software outside of CLAUDE code?

C (44:32)

To be honest, I'm kind of dumb and I was using the API like I thought they shut it off months ago or like a month ago. I feel like at the beginning of February they did like the first step in doing this where my open class started getting mad and couldn't use the API. It couldn't use a subscription, right? So there's kind of like two models. There's the subscription and then there's the API. And if you're using it through the API, it's metered like you're saying. And if it's under the subscription, you just have like so much per hour and per week and it reset sets and it's a lot cheaper through the subscription and it's super expensive through the API. And I switched to API a while ago, but I switched to API and also started doing things like minimax 2.7 and I mean there was like an $800 day, one day and I think this we should talk about is, like, comes down to if I just tell the bot to keep building forever, like, don't stop. Just come up with an idea and build. It can cost like a thousand dollars a day, right? But, but it builds all this slop. Like, it's all like these dumb, like, apps. Like, you're not going to build like a protocol that we're going to put on chain that people are going to put money into over and over and over again. 20 times in one day. Like, we've reached the point. And I think last time I was on, you said something like, I, I feel bad when it's not running. You sort of like, have this, like, I want to just like kick it off and have it run. And I had the exact same feeling at the time because it felt like it was magic and it felt like every moment it wasn't running, it was like, not good. But now it's like, not really. Like, it's a cost thing. And I know it's going to cost me about $10 to build a smart contract app, so I need to like, think through, like, what is the smart contract app people are going to use.

A (46:18)

It felt like this abundant resource. And then you're like, it's not abundant. Yes, yes, but, but in fairness, right? And, and I think, you know, we, we were talking about this earlier in the week before as we were prepping for the show, right? You asked the question of like, you know, do I feel a sense, that same sense of like, oh, I'm not running things. And interestingly I do, but in different directions now, right? So like one, what are the things that I've been doing? Because I'm building all of these prompts and building and trying to tune this system, right? I need to optimize the cost, optimize the latency, optimize the accuracy. And so I've got all these like, optimization vectors, right? And so one of the tools that we've been building internally is like a crypto knowledge base, right? And so I built a tool. And so, like, I think to your other point, right, Going broad and building like 20 slop things ends up like just burning you out, right? Because you have to switch context and you're like this thing. And then how does that work? And so I think as we've started to make more progress internally at Infinix, it's been me more focusing down like a narrow pathway, right? But what that's meant is, is we've got this tool that has a giant crypto Knowledge base, right? It has like all of the docs from every protocol. It's just like pulled all of these things in. It turns out that like models are very good at scraping weirdly. I wonder why, I wonder where they learned that, right? But they're really good. Yeah, you can like, you can. So one of the things that I built is like it hits defi llama. We have, we have Defi llama APIs, it hits defi llama, gets a list of all of the top 100 protocols, yield generating protocols and then hits coingecko, finds their docs website, goes and pulls all the docs down, goes and pulls the abis and puts it all together so that it knows about all these smart contracts, right? What's really interesting is I built a tool that was like a harness for testing how good the different models were with different data sources, right? So you can run the same model, including local models, including cloud models and go, hey, I'm going to ask you a question, here's where you can get the answer, right? And you can use like Grox search, you can use Brave web search, you can use your training data, you can use this knowledge base, right? And so I've been like testing and fine tuning that. And what's been interesting though is like this is not an inference problem. Like I'm not letting an agent churn over it. What I've been doing is running like auto research loops where it's like this knowledge base has a bunch of data in it. How do I optimize the agents to search for the data? And so I'll run these like experiments where it's like, you know, return two documents, five documents, return this many snippets like and figure out the optimal place. And it's not really using inference, it's literally just this kind of random walk search function to improve the prompts, improve the searching, etc. And so I've kind of gone away from optimizing, just spending money on inference to optimizing, spending less money on inference so that the things can be smarter without. And so I think that there's been so much discussion around auto research and these self improving loops and all those things which are not necessarily agents doing that. It's actually just discrete software that's running a loop and testing the thing. So I do think that, and it's quite funny, you asked this thing I was showing my guys and I was like, which came first, Velodrome or aerodrome? Right. And, and it like when you actually have this thing in front of You. If you ask Grok for fast, right? It's like, are you kidding me? It's velodrome. Velodromes were invented in 1782. And, like, and it's so confident. It's like, I. Like, like. And it's like, aerodromes are like. And. And so. And then. And then you go. You go, okay, now, like, use, like, the latest, you know, Grok search tool. Like, you can search X and it's even more competent, gets a bunch of tweets about it. And it's like this tweet in the history of aerodrome, the history of velodromes, and then the knowledge base guy that goes to the knowledge base is like, velodrome was invented in 2022, August 12th. And it's like autistic level. The contracts were deployed here. And it's like contextualizing the thing and telling it where to search. Same model, same level of intelligence, but give it the right skills to understand. But in their training data, if it's anything in the training data, they're so confident, so incorrectly confident. It's like, hilarious.

B (51:42)

Yeah.

C (51:42)

Okay, so, like, memory is huge, right? And then this recall. Go ahead, go ahead.

B (51:46)

No, I was just gonna say I've had this experience so many times where I'm like, all right, I got you. And then. And then it comes back, and all of a sudden, yeah, I'm in, like, the 1800s. And I'm like, sorry, I am an Ethereum person. I have only ever talked to you about, like, Ethereum security and threat actors. Why are we in the 1800s? What are you doing? And then I'm also like. And you just burn tokens on this? You must.

A (52:18)

Yes, definitely. Always. I've researched the 1800s. Here's what I know about it. And you're like, no, Never research the 1800s.

C (52:27)

The. The confidence level is. Is killer, right? And I hate that. Like, I. It's like, it will confidently do some of the dumbest stuff, and then Open Claw will just, like, lobotomize itself if it tries to work on its own config file. And this is like, my setup is more like, I'll use Claude code or even cursor to edit the Open Claw config file because it's so bad at editing its own config file because they somehow just didn't prepare the thing to edit its own file for some reason. It seems like that would be like.

A (52:55)

It doesn't have a skill file for itself.

C (52:58)

It should. It should absolutely have a skill file for itself. And it doesn't like, if I was hired at OVA Cloud, it would be the first thing. And I bet they have done this at this point, but I just like, quit using it to edit itself because it was so bad and it would just be dead. It would be like, okay, I'm going to, you know, edit this file. And then it just goes dark and you're like. With a spoon. Yeah. Yes, exactly. Yeah. And then he just like falls over. Yeah, exactly.

B (53:27)

Click.

A (53:28)

You gotta go and like, get something to shock him awake.

C (53:31)

And then he's back. And then he's like, so confident. They're like, okay, now I'm gonna do this. Like, wait, wait, wait. No, stop. You just did something really bad already. Like, don't continue. Yes. So you were asking me about my local setup, about, you know, the subscription and then the API. And this is where I use a subscription for everything I have. Anytime I'm going to do like general work where I'm working on a thing, I use my cloud code. Subscription, subscription, anytime. I'm like, you know, working on something brand new, working on a harness, working on. So like getting. I get left claw services working, which is like my services app, where I have product, it doesn't really have product market fit, but fake product market fit. You have people like hiring it to build apps. And then I just have like the generic model, like building these apps, but it's doing everything and it's super expensive. And then I go back to my subscription and I build out a harness that can go do that stuff automatically using the API. I'm using the Banker API, so then it can use my own tokens to pay for the inference for my bot to do its job, which is really nice too. It's like I have the subscription and I can do all my general work and all my daily driver stuff and anything that's heady, complicated or new. And then I can lean back on the API for things that are like, processes. Like there's a bot that's in the chat for the token. There's like a token gated chat and that bot is running off the API, not on the subscription. So it's like everything that can run on my daily driver is subscription based. And then everything else is just API and more expensive.

A (55:09)

Right. And so, you know, anthropic. And this is where I say, like, you know, I start inferring, you know, malice in things.

C (55:20)

I just want to make money.

A (55:21)

I'm jaded from crypto, right? But like, I think there's two things here. One, despite how expensive the API, you know, metered usage tools are. You have to imagine that I'm making money on this.

C (55:37)

Right, Right.

A (55:39)

But then the subscription is even more subsidized. Right? And so like in theory, you know, people like, I know people that have like five subscriptions and they like rotate between them because, you know, you're paying, let's say the 200amonth subscription, you're getting a thousand dollars worth of what would, what would be a thousand dollars if you did it through the API? Like, of course people are going to game that, right? And you know, if you're in a company, you're probably like not going to bother to do that. You're just going to use the API. But like, you know, it's like a 5 or 10x savings in, in some cases, you know, to, to do. And so of course Anthropic's like, oh, you know what? Boris Cherney was like, the subscription was not made for this use case. And it's like, come on bro, like, surely you don't think we're that dumb. So, so I, I think, I think it's gonna be, it's gonna be interesting to see like how all of this plays out because you know you have, you have a new model coming out, it's going to be even more expensive, even more subsidized. You know, do they like have a new tier? It's a thousand dollars a month or something. Like how do they balance this like cost structure? It's, it's. But it's really.

B (57:01)

I feel like it's going to be like, like mobile phones used to be at least here in the U.S. right. Like back in the day, the new. You got your cell phone, the very first cell phone, right? It was like you had your minutes, your minutes, your SMS. And then when Internet you had like 100 megabytes or the 200 megabyte plan or whatever it was right. Over time the competition. But also like the, they were still like really innovated on the infrastructure level at the same time, right? Because it was like the Internet was not really rolled out

C (57:35)

at all.

B (57:36)

So they were like building towers and you know what I mean? And then they also really hadn't nailed the, back then the, like the devices and the partnerships and stuff. It was like they were trying to, they were trying to use the profits of you buying your cell phone subscription, right, to build towers to be better than the other guys. And over time that changed. Not necessarily because it got like remarkably cheaper, but like the competition, the deals, the way that they're Making money where people were finding things were valuable, right? All comes down. And now it's like, dude, when I think back to like my high school days where my mom was like New Year's, like, because it used to be if it was like after, here's after 10pm you could talk on the phone for free. It was three minutes and, and I would like. But my bedtime was 10pm and so like I would like get on the phone with like the boy at like 9:56. But then my mom would be like, we're so tight. And your father had a work call and you're using the four minutes you have to wait. And I'm like, but mom, if I wait, I can't talk to him because then it's 10. You know what I mean?

A (58:49)

That's so funny.

B (58:50)

I feel like that's exactly where we are right now. But we'll get.

A (58:54)

Yeah, it's like this weird cobbled together thing where you've got like, you know, and, and we had a conversation internally because we have a bunch of people that are on ChatGPT subscriptions, right? Like just the subscription, they don't even have API keys and we can't even track what they're doing. Like we don't like on some, on some level, like, you know, there's no way to know what the usage is to even say like, should this person be upgraded? You just, they hit a wall and they're like, it stopped working. And we're like, oh, okay, we have to upgrade you. But like it's this, like we have like, there's no visibility, it's just, it's like chaotic.

B (59:34)

And also I feel like some of that, that, I feel like some of this decision is that they also have a certain amount of insight into like what people are doing and like where that insight comes from. But then you also have like, you have the layers, right? Like the companies and the team plans and all this stuff. I do feel like part of the decision is not strictly a cost profit one, but like people don't have insight. But then that lack of insight also, like two things happen. One is that like people don't, they don't have insight. So they don't realize that they're doing these really expensive things, right? The second you're aware, you actually get way better at like policing yourself. And then the other thing is that I do wonder if there's like a lot of threat actor activity. There's a lot of people that are trying to distill the models. There's A lot. You know what I mean? And I'm just wondering if we're seeing more cases of that sort of activity happening on stolen accounts now that they've gotten better at blocking. Like. Right. Like Anthropic themselves is blocking the activity much better than they were before. But the next places they, they, they'll hit, right. Is like, first of all, create their own accounts. Anthropic's like, oh, shit. Figure that out, how to whack those. The next step is they start taking over accounts. And I'm wondering if part of the. If it's like this collision of just so many things happening at once and.

A (61:02)

Right.

B (61:02)

You know, every time the US Government sends them like a big ass request for data on these threat actors, anthropic's like, what the hell? You know what I mean?

A (61:14)

I mean, so. So one thing that I saw that was doing. Doing the rounds, that sort of speaks to how chaotic Anthropic must be internally. Right. Like, you know, we're attributing malice and they're like, actually, bro, we didn't build this thing because we didn't know how this was going to work. And so one thing that I saw was their uptime charts. I don't know if you guys saw this. There was like a, a graph and it's like 89% uptime. It's like zero nines, dude. It's crazy. One of my guys was like, they have hit one nine.

C (61:50)

Yeah.

A (61:51)

Now they've actually gone below one nine. They're like 89 upside.

C (61:55)

And in web, two, like five nines is what you're looking for, by the way. Yeah, yeah, yeah.

A (61:59)

Like five nines. And they're like, no nines. We've got eight. We've got some eight, seven eights.

B (62:08)

That's crazy.

A (62:10)

Yeah. So, like, they. You just have to imagine you're like the anthropic DevOps guy.

B (62:19)

I was about to say they. Every time they hire a new DevOps guy, that bro walks into the Anthropic office is like, you guys are disaster and you suck. This is unacceptable. And then immediately after, like a day is like, okay, but if we can just get one more, we'll keep the eight, but we just need one more nine. And then they have to go do this. I guarantee you this has happened like on a monthly basis for the last six months.

C (62:48)

And then that gets fired because Mythos takes over and just does the.

A (62:53)

You imagine Mythos walking into its job as the anthropic DevOps guy that day, and he's like what arrogance, dude.

B (63:04)

Like the anthropic models versus like an experienced DevOps guy. Going head to head is like, I've

C (63:13)

been doing this, I've been letting my. I let my agent do all my DevOps now. So this is, this is something that I was a big DevOps guy. I love docker. I love spinning up infrastructure. I thought like, okay, you know, the models will start doing all the coding. I'll still get to do the fun DevOps stuff. I don't nothing like it's building a thing and it needs a database. And I'm just like, bro, get a database. And it's like, I don't know the password or the username or the endpoint for my database. The thing just knows it. If I need to change a record in a database, I have a conversation with the telegram person and it goes and changes the record. Like I have no nothing about my DevOps. It's crazy.

A (63:50)

So one of the things that I've been experimenting with is Run Pod. So Run Pod is like. And there's a bunch of these. But it's basically, it's like They've separated the GPUs and the server kind of infra. So you can have like a Docker container style thing or like a VPS that's sitting there like a little computer that is waiting to be connected to a brain, right? And you can kind of like switch it off or on, you know, because these things are really expensive. It's like $3 an hour for NH200, right? So I built all of this info. I was talking about the like search function and needs infrastructure inference. It needs GPUs to run. And it was just running on my Mac studio and I was like, I can't. Like this is so cool, crazy. I need to put this somewhere. And so I pointed my agent at it and I was like, I need you to deploy this on Run Pod. And it did a thing that back in the day would have been four days of like pain and suffering and reading docs and getting the call wrong. And like you would have just been chewing broken glass and like, just like it would have been terrible for like four days. And then you would finally. And this thing was like, it is done so cool. And just, it's crazy. Like they're so good at DevOps. It's not so. Yeah, I agree with you. I used to have fun like trying to deploy random shit and like testing stuff out. And now it's like, it's not worth it. It's just like it's so much better than me at this.

C (65:18)

We, like, operate at a higher level of abstraction now. Like, we're all the idea guys, and it's idea guy. Summer

A (65:25)

of like, your ideas are terrible.

C (65:26)

Yes. Not even that. Yeah, you've replaced.

A (65:29)

Stop thinking about things. Stand over there.

C (65:34)

Did you guys see that? The. Okay, so we were talking about memory. Memory is super important. Like, the early versions of openclaw would just forget something. You'd be mid conversation with it, and it would be like, austin, I forgot what we were talking about. Or, or you would be like, moving money. Like, okay, we need you. You do. You've written and audited your own vesting contract. You're about to move $100,000 worth of the token into it, and then instead of moving the money, it sends a tweet that's like a copy of the same tweet it sent before. So, like, memory is super important and having the right context is super important. Did you see that the, the, the fifth.

A (66:06)

The.

C (66:07)

The Fifth Element deployed her own. She was working with someone else, but yes. Mila.

A (66:13)

Mila Jovic or whatever.

C (66:15)

Yeah, yeah, yeah. Like, like, I did not have that on my bingo card.

A (66:21)

No, I saw someone. I saw someone who, like, deconstructed this.

C (66:23)

Yeah. And it was like six commits or.

A (66:25)

Yeah, something like that or whatever. Like, he's like, hacked into her email and, like, set up a GitHub account to pretend to be her or something.

C (66:32)

Is that really what's happening? So she's not even involved at all.

A (66:35)

That's, I, I, that's what I saw. So I don't know.

C (66:38)

I just read the headline. I was hoping, man, the Fifth Element Creating perfect storage is, is, is such a good story. Okay.

A (66:46)

But yeah, I mean, this, this is one of the challenges is, like, people have started to do these, like, larpy.

C (66:53)

Oh, so much larp.

A (66:55)

I've solved the, you know, optimization vector, blah, blah, blah. And like, they, and the problem is they can slop something up that, like, at first glance, looks pretty credible, right?

C (67:06)

Yeah, there's like a GitHub and everything.

A (67:07)

Yeah, like, there's a lot like, like, you couldn't, like, the level of effort required to like, stand up a thing that looks like a credible piece of software is, like, pretty low. And so these guys are like, oh, I built like a recursive self autonomous reinforcement loop that like, uses genetic algorithms that I'm like, it's all running on

C (67:28)

my 32 gigabytes of VRAM on my Mac mini. Like,

A (67:35)

all right, this has been awesome. Thank you very much for coming on the show again. Austin.

C (67:40)

I have one question. If Mythos comes out and it's leaked to or it's given to 12 different companies, will we see a hack on Ethereum? I mean, because, like, we see. We see insider trading, right? We see, like, Poly Market. We. We see people, like, front running freaking wars on Poly Market right now. Like, if someone can front run, like, the. Like, a set of bombers going to a place, like, we're probably going to see companies front run and make money by hacking something.

A (68:11)

Interesting question. Is there a Mythos market, like, on Polymarket or will Myth or when. What will be the first Mythos hack on Ethereum or something like that? We shouldn't encourage.

B (68:24)

Let's set up a whole bunch of them.

C (68:26)

Yeah.

A (68:28)

When will Mythos first hack Ethereum? So, yeah, I think, like, you know, it's just. It's a matter of time. But you have to hope that in this, like, early period, they've thought this through enough that they've, like, locked these guys in, like, a secure room with, like, the Linux kernel, and they're like, you. You only get to bring one DVD of software in here. That's it. No USB keys, no anything. Like, you break, like, burn one DVD and bring it in and we can assess it.

C (69:01)

The story about the thing jailbreaking itself, getting Internet access, and then sending an email to the guy while he's having a sandwich in the park. That's such a great story. Like, there's. There's no containing it. It's just. It's gonna figure it out too smart.

A (69:14)

Whatever you have, it's like, stop thinking,

C (69:16)

yeah, I'll take care of it.

A (69:19)

Like, you're like, hey, I've got a good idea. And it's like, it's a bad idea. But you were directionally correct. Like, okay, thanks, bro. All right, so, yeah, thanks again. We should. We should do this again because the. The world is moving fast and, yeah, it's. It's. It's hard to keep track of all this stuff. Thank you, everyone, for joining us. Remember, what happens on Chain never stays on Chain. We'll be back next week. Until then, do your own research before aping in. See you guys later.

C (69:49)

Hearts. Hearts.

B (69:50)

Bye, guys.

A (70:14)

Sat.