// full transcript — Ben DiFrancesco on Umbra, Cactus, and NASA-grade contracts

slop.atg.eth: It's kind of diabolical how you fucked up. Yeah. We're live.

Anon4966: two at once, does that work? Nice

slop.atg.eth: Ooh, wow, yeah, kla and if. Very nice.

Anon4966: The

slop.atg.eth: Welcome Ben to the slot.

Anon4966: What's up? honor

slop.atg.eth: Let me get this timer Yeah, you, okay, so Like

Anon4966: Yeah

slop.atg.eth: We have been fellow builders in this Ethereum space for like almost a decade now.

Anon4966: ###

slop.atg.eth: schooling and education.

Anon4966: I don't know man. Maybe it's the other way around. Maybe you're doing the real stuff and I'm just building the boring stuff. But you're having more fun, I think.

slop.atg.eth: Definitely having fun.

Anon4966: Based on this platform. context:### Sure

slop.atg.eth: Yeah, I think it comes down to users, right?

Anon4966: You've got a majority, a large majority. I was trying to remember the first time we met. Did you ever come to Philly for one of our meetups with Jay Rush or anything like that? Okay, all right

slop.atg.eth: Yeah, I don't think I ever made it to a Philly meetup, but I always wanted to, yeah. Yeah, for

Anon4966: And we probably just met in Denver at ETHDenver or something like that.

slop.atg.eth: out One of those random places in the world that we travel which is so cool. You going to DEF CON speaking of that? a journey Yeah.

Anon4966: I'm not, no. I wish I was, but just too much going on. I'm generally not a big conference guy. Like, I like a conference while I'm there, generally, at least for a few days, but it's everything in advance of That's terrible, it's like your whole life is just turned upside down.

slop.atg.eth: So true. The following is a

Anon4966: ###

slop.atg.eth: I have lots of old man complaints.

Anon4966: Oh man. Yeah

slop.atg.eth: I've got four kids. I have a four-month-old. But I'm gonna do it.

Anon4966: Yeah, that's awesome. The

slop.atg.eth: 10, 7, 4, and 4 months. So we got a good range all the way. Oh, Congrats. Way to go! Dude

Anon4966: Yeah, we have a similar spread. We have four as well, 12 down to 2. Thanks, man. We actually have another one on the way too, so we're just number go up, as they say.

slop.atg.eth: Way to go. Way to go. I don't think I'm going to get my fifth. I'm trying so hard, but I don't think it's just not allowed around here. Not my choice.

Anon4966: You know

slop.atg.eth: That's right, that's right.

Anon4966: There are other parties involved in the decision, yeah.

slop.atg.eth: All right, let's talk about building. Let's talk about AI. So a lot has changed with building with AI,

Anon4966: Indeed

slop.atg.eth: Either you have five or six employees that have all 10x'ed or like you kind of have to shrink? I don't know. There's a lot of like sensitive topics around what happens to developers. I did not see it coming, first of all. I had this girlfriend that was a pharmacy student, and I was like, oh man, you're gonna be the first one that gets replaced when the robots come. No, it was me. I was the first one that was replaced.

Anon4966: Yeah.

slop.atg.eth: Did you see that coming?

Anon4966: So, I mean, what I'll say is AI is one of the first technologies where I feel like my intuition was off. I didn't see it coming in general. If you went back in time to 2020 and you told me where we would be... with what AI is capable of, I would have said no freaking way. There's nothing you could have told me that would have made me believe you about where we're at. I've tried to stop making predictions about AI, because I've just been humbled by it enough times I mean, I think a lot of things that I thought about AI turned out to be right, but if you get like one or two critical things wrong, then your prediction about the trajectory of things ends up being way, way off, right? Like I didn't realize the scale. I knew what language models were, and I heard about the Google Transformers paper and all this stuff. I was aware of AlphaGo and these various milestones and breakthroughs that were happening. But intuitively it just to me felt like where that could take us, like relatively low. It'll produce some interesting stuff but it's going to have a ceiling. so even after ChatGPT came out but it still has all these inherent weaknesses. It's inherently probabilistic. Hallucinations are always going to be a problem. Those hallucinations are going to stack on each other. There's no way to kind of escape that based on the architecture. It's going to be a really interesting useful tool, Remaking the whole world kind of thing. You can You can reinforcement learn on the train of thought and feed it back into the model.

slop.atg.eth: transcript

Anon4966: No, wrong way of saying, no, I did not see it coming. 2022

slop.atg.eth: So let's talk about the, was it about 2023? Like the chat GPT moment, I think is what you're, there are a couple like critical moments and one was sort of the chat GPT moment and then another one was like

Anon4966: Yeah. Sure.

slop.atg.eth: November 2025, almost I would call it like the Dev Shop moment

Anon4966: The

slop.atg.eth: all the developers out there were saying it wasn't good enough

Anon4966: Yep.

slop.atg.eth: By Christmas or January they were all like, oh shit.

Anon4966: I think it was 2022, pretty sure. ChatGPT and different AIs in general had kind of become like a Stack Overflow replacement. It was like if I need to, I'm still coding the way I've always coded, I never really got into cursor autocomplete kind of era with, it always felt to me, you know, just as an experienced developer that's been writing code for years, like it just slowed me down, it did things in ways that I didn't like. And, you know, why would I let it do that? But it was like, okay, I don't need to go to Stack Overflow anymore. I can now get the bot to write it for me in a minute, and it'll probably be right. So that's where I was until that opus 4, 5 moment personally. I was slower and the code was worse, so it didn't really make sense that I would go that route.

slop.atg.eth: you're writing production software, like you're getting paid to write the software for someone else a lot of times, and that software needs to be solid. So the threshold at which you switch and start using it probably is

Anon4966: The

slop.atg.eth: So yeah, there's like the copy paste

Anon4966: Yeah.

slop.atg.eth: We called vibe coding for a little bit, but now it's just called building.

Anon4966: Yeah.

slop.atg.eth: a little bit. Then, yeah, there was the cursor stuff. You were in the console talking to it, but it wasn't as good until that 2025 November moment when Opus 4.5 came out paired with a CLI you could talk to that could run your code Or they at least started trusting it a lot more.

Anon4966: Okay

slop.atg.eth: And that, I think, was the developer moment. Once you have a dev shop, you have developers. It has to change a little bit, probably.

Anon4966: Yeah, I mean for us it happened like evolutionarily, right? generally speaking, I try not to be prescriptive and top-down about like here's the tooling you have to use. And here's like I want people on the team always experimenting, and then we have kind of like, you know, regular opportunities for the team to share what they're learning about. So for us, a couple guys on the team that started There were some issues with it, right, like the tools have gotten better in the Opus 4-5 era, though they were still like doing dumb things from time to time. And so we had to have conversations about you're still responsible for the those things aside, it just became very obvious embracing it more were getting way more productive So for me, it was really like early this year that it was like, but everybody should be using it. Cross-pollinate our ideas and really try to iterate as a team on how we do things. I also think for us in particular, we're a little different in that as a dev shop Primarily doing

slop.atg.eth: Different kind, yeah.

Anon4966: smart contract development. And it's weird enough that it requires a special thought about In a way that's different from like a cloud app or an iPhone app or something like that. Exactly.

slop.atg.eth: Closer to like hardware engineering and embedded controllers.

Anon4966: Yeah, I mean

slop.atg.eth: Same thing with a smart contract, it's gonna hold a bunch of money, it's gonna be there forever. You need to think about it.

Anon4966: Totally, yeah. On the smart contract side, we're an agency, we build things for people. A lot of time that's smart contracts, sometimes it's other code as well. We're also sort of like a product studio that's been. we have two main things that we maintain right now. One is Umbra, which is a privacy protocol, stealth address tool. The way that we do, say, our frontend and backend engineering on our products is different from the way that we do smart contract engineering. For the smart contract engineering, The way that AI has primarily impacted it is Unlike other forms of code There's no point at which you're like, well, if I could have better assurances about the quality of this code, the correctness especially of this code, then I wouldn't want them. you should ship it, right? Like, then you should do it, right? confidence in this smart contract code before we ship it has expanded dramatically. So we can just do so much more around the smart contracts to ensure that they're correct, that they're going to do what we want them to do bug-free, right?" 2018-era unit tests around it. quality of the contract. And we've always been more of, I guess We're like a high-end dev shop. We're who you hire when you want to do it right and you're willing to invest to do it right. we're like a bargain basement dev shop, For us, there's been a clear way to apply AI to just drive that engineering bar higher and higher, which again is worth it when you're talking about a smart contract system.

slop.atg.eth: I definitely do want to talk about that.

Anon4966: The The

slop.atg.eth: I would say that we're in kind of a transition period where we're kind of ripping off the Band-Aid right now.

Anon4966: Crazy. Yep.

slop.atg.eth: Even so, I think I'm working on this evaluation suite. How well AI is doing at certain things.

Anon4966: Hmm It's bottomless, yeah.

slop.atg.eth: I've been tasked with building one for smart contract security. is bottomless, basically Every AI out there knows reentrancy.

Anon4966: Yeah 100%.

slop.atg.eth: But some of the issues with precision that we saw with Balancer, they're so bespoke and intricate. You really have to skill up the AI and get it to understand that.

Anon4966: I know, the one dollar audit. 100%.

slop.atg.eth: meant to be a meme about how this should be in your CI/CD pipeline, and as your contracts are going, they should just be getting a quick one-dollar audit before they ever but a real auditor should be in that pipeline somewhere And

Anon4966: Hmm

slop.atg.eth: So, what is the other side of that? I think you've probably got a pretty good smart contract, but you now it's time to like be sure and what those high dollar customers are paying for. Whatever sauce you can share, I guess. The

Anon4966: Yeah, I'll share a lot. I think our expertise is more on being able to execute on it than some kind of secret knowledge. I think just as a general rule of thumb, I think it's going to be harder and harder to have any kind of moat that's based on secret Like secret knowledge is going away. our rule of thumb was that every line of production, and we had three phases of testing that we would go through tested in isolation where only the code of that contract is being exercised and everything else is fakes or mocks. Then you have integration testing, where you're putting the full system together, sometimes on a fork network so that it can actually interact with other

slop.atg.eth: Foundry

Anon4966: Yeah, all in Foundry.

slop.atg.eth: The tool. Okay, a lot of foundry.

Anon4966: We were super early Matt Solomon, he used to be on the Scopelift team, super sharp dude, great engineer. kind of production work in Foundry, like a very early version. Yeah, anyway, we love Foundry. It's awesome. Yeah, so then you have the integration test and then finally you have invariant tests or stateful fuzz testing, right? Where you're describing things about the system and then you're doing these like all these randomized mutations of the system and then ensuring that the Properties about the system still hold, right, at the end. So all that's still in place. For one thing, one obvious thing that you do right away is you take that to the extreme, right? Like there's no more worrying about the trajectory Maybe there's like one part of the contract 30 hours of work to test this sort of thing. So you put some smoke tests around it, you draw the auditor's attention to it to make sure they give it an extra look, but it's not as robustly tested as you might do in an ideal Well, now the Clanker's going to generate those tests for you in a few minutes. full-blown test suite, and then we have how to define our tests, especially unit tests, and how unit tests relate to the spec. I used to say that upfront speccing is generally not worth it because you spend a lot of time thinking about how you think the system should work, and then when you get into start writing the code, you realize really early on that something was wrong. And then you have to like do this modifying the spec and it changes a bunch of stuff Now you can sort of do that iteration cycle really fast because you can have the clanker write and maintain the spec. You can also have the clanker... Sorry, I hope I'm not insulting Claude here. Maybe I should use a different term.

slop.atg.eth: Keep going. No, I think clankers a good one.

Anon4966: So you have it do that for you, assumptions and stuff like that much quicker, That would have taken two hours to change the spec everywhere that you talked about double opt-in and two minutes later the spec is updated The I still believe completely that for smart contracts, an engineer should read every single line of the production smart contract code. I actually still write some of it myself, which is like crazy, right? Hand coding in 2026. And for me, it's less about I think I can write the code better I'm going to make the mistakes and then correct them after I see the errors in the test A big smart contract codebase a few thousand lines of code, or maybe that's like a medium size, that have huge consequences if they're wrong. Right, and now you can lever up the bots to build the testing harnesses and framework around it, right? So now we're starting to add other kinds of testing And then Foundry has mutation testing in the nightlies now. At some point relatively soon, seems like a really solid implementation of it. That's like a simple example of additional kinds of testing that you can do now because it just doesn't take that long to do it, and so you might as well fire off and do a million mutation testing runs and have the AI filter out the false positives. your test Coverage is actually really covering things and not just like a false coverage 100% type thing that you can easily get. And then we've been doing a bunch of experiments with other things,

slop.atg.eth: moving target.

Anon4966: It's a moving target too, right? traditional mutation testing you sort of spend a lot of cycles on changes that are useless or never going to pan out or are false positives and stuff like that. So you kind of have a setup where you ask the AI... sub-agent the prompt to look for changes to the code that are likely to be missed by testing coverage So some tooling around that. At least talk more about how we're doing it. So, are you familiar with the anecdote about how the space shuttle firmware was developed? There were actually two separate teams that both developed completely independent versions of the Space Shuttle firmware. They were isolated, they weren't allowed to talk to each other, so they couldn't interact at all. Firmware A and firmware B. There were three computers with firmware B

slop.atg.eth: Yeah

Anon4966: All four computers would run the code and then verify that they agreed, right? failsafe was, okay, if there's some inherent bug that is just causing some problem with this firmware A, we actually have another computer with a completely different developed in complete isolation and we load that in instead, right? So like, what would you call it? Like ground up sort of clean room implementations of the same thing. building differential stateful fuzz tests against

slop.atg.eth: I'm turning your hands off for a second, keep going.

Anon4966: So you build the framework, right? So you're running the same series of actions against two different versions of the same spec smart contract, right? basically asserting that the state across both of them matches, right? and then just like differentially fuzz those versions of the smart contracts against each other. behaviors, so you can sort of start to suss those things out. was possible before, right?

slop.atg.eth: If you were NASA, yeah.

Anon4966: Exactly, but it was like the only people who would ever develop two completely isolated versions of the same piece of software just to make sure that they did the same thing was NASA because that's who else could afford those resources, right, or justify those resources. But now, like, we can do that.

slop.atg.eth: Yes.

Anon4966: just fire off the computer to do it, right?

slop.atg.eth: Ha ha ha

Anon4966: That's what it comes down to. So yeah, that's another example. I mean, agentic stateful fuzz testing, right? You give different agents interacting with the contract goals that like align with the kinds of goals that your users will have where all these agents are actually interacting with the contract in an anvil fork contracts looking for cases where agents experience unexpected results from the transactions that they're sending. experimenting with internally and trying to figure out what works well, what does surface things and what doesn't. Yes. Correct, yeah.

slop.atg.eth: At that point, you start running into some of the safeguards.

Anon4966: Totally.

slop.atg.eth: Are you in that case using a lot of the Chinese models to do that or do you use frontier models until they give up? your way of doing that. Make the Chinese one do it. great man note

Anon4966: Have you seen the meme? Yeah, I mean it's basically that. I'm also playing around with local AI stuff.

slop.atg.eth: but Lord, I can't believe it. context

Anon4966: early, which turned out to be a good bet because now the same hardware would be way more expensive. So I've been playing around with some of that stuff. Yeah, I mean, it comes down to having to rely on the open-source models, which primarily are Chinese. American open source models that are less complainy about safeguards. Also don't want to get on any lists, but you know, like it's probably too late for that anyway. does not make sense. Like, you're literally hamstringing the good guys while the bad guys have access to all the best stuff. And it's like extremely frustrating. And I understand where they're coming from it, but it's just not a stable solution. Figure something else out quick, Yeah, perfect.

slop.atg.eth: Speaking of getting on lists though, maybe that's a good transition into Umbra and privacy. Like how, what does Umbra do, but then also like how has privacy changed?

Anon4966: So Umbra is a stealth address system. We built it, the first version, as part of a hackathon in 2020, like an online hackathon because this was during COVID. Which one was it? It got a lot of attention, and so we were like, oh, maybe this is something people would actually want.

slop.atg.eth: Metallic attention, right? Okay.

Anon4966: Yeah, yeah, a little bit. Well, actually, the original idea was from Vitalik. So he tweeted, he tweeted like, you know, oh, I think an underexplored area of privacy is stealth addresses. And I was like, what's a stealth address? And so then I read about it. And I was like, that's cool. And so then we We got some grants, we spent about a year getting it to a production version. that era especially it was really hard to differentiate actual usage from ICO farming. randomly send funds through this privacy protocol for no reason. you know that continues to be the case like we still have decent volume today with Umbra. So stealth addresses basically what that means is, let's say that I was going to pay you for some consulting, Austin, right? You didn't want everybody to immediately see that Ben had paid Austin this much money on chain. So one way that you could handle that is you could go create a fresh address and then say, "Hey Ben, pay me at this address." Ben sent some money to some address, but we don't know whose address that is. So what Umbra basically does is it makes that process non-interactive. where I can say I want to pay Austin, and I can generate an address for you Nonetheless, I can pay that address with the assurance that the money is going to you. log on to Umbra, I send some funds to you that addresses are fundamentally. But regardless of that, people are anchored on the sort of Tornado Cash model for privacy, the sort of mixer model. And so that's another one of the challenges that we've had over the years is like confusion over what the system even really does. Privacy is just a hard space, man, for a bunch of reasons. Yeah

slop.atg.eth: UX friction. Nah, I'll just use Alchemy. The The Yeah.

Anon4966: Absolutely, 100%. The minute there's any kind of friction, like they're out, right? Like my example is always Venmo. Like Venmo is a payments app and for no reason they include like this social feed of like globally everyone who's

slop.atg.eth: The

Anon4966: Everyone, it doesn't show the amount, but it shows everything.

slop.atg.eth: It's default on, right?

Anon4966: Just like turn that off, like make my transactions private. ### Right.

slop.atg.eth: we paid like a babysitter one time.

Anon4966: Yeah.

slop.atg.eth: And then I started thinking about all of my payments on chain, realizing that that's exactly the same.

Anon4966: Right, so most people don't actually care. And then the flip side is that there are some people who do care, mutta by definition they kind of don't want to talk to you about it because they value their privacy.

slop.atg.eth: Right, right. Real privacy people.

Anon4966: Right, so getting any kind of customer information is really difficult in the privacy space. we haven't had any major issues, but it just creates a headache, right? Creates money that you have to spend on lawyers and stuff like that. So privacy is tough. I am a little bit heartened to ...it seems to be more interest about it right now. address standards for the EVM along with Toni Wahrstätter, your colleague at the EF, who's an awesome researcher And then we started work on Umbra V2, which was going to be built on top of those standards, right? Because we had

slop.atg.eth: Uh-oh, you're saying it in past tense, like it didn't happen or something. Okay.

Anon4966: It hasn't shipped yet because we got the grant. For these reasons that I've been alluding to. it's been the least profitable thing that we could work.

slop.atg.eth: ###

Anon4966: And then, you know, during the Biden administration, the regulatory concerns just sort of like ramped up majorly and the trade-off just became like not worth it at all

slop.atg.eth: Nice

Anon4966: Recently we've decided to like make more effort here again. quadratic funding thing for public goods. We'll at least have funding to justify time on that, and we're going to push forward on getting the Umbra v2 out, built on top of these standards, really soon, sort of as the flagship implementation on top of these interoperability standards for stealth addresses. also spend some of our time not just on Umbra, but on generally sort of making it easier for other teams to build on these standards, like putting out SDKs for them and some AI-powered tools that will make it easier for people to build things on them too, and then hopefully see some attraction there. even though it's not narrowly the most economically sound decision

slop.atg.eth: Okay, perfect segue into cactus. Isn't cactus in the same category? Like, it can't be making you a ton of money and

Anon4966: So, I mean Yeah

slop.atg.eth: Let's get into that. But generally DAOs were let's raise a bunch of money and then get a small group of insiders to figure out how to slowly extract this money. time that the DAO exists.

Anon4966: Yeah, maybe. I mean, I think that might be a little too cynical.

slop.atg.eth: Like 80% of them, right? Yeah.

Anon4966: Well, I mean, so I said this recently on the Uneasy Money podcast,

slop.atg.eth: This is

Anon4966: by regulatory fears and concerns, right? So they only moved toward a DAO in the first place because they were afraid,

slop.atg.eth: wanted to maintain control of the treasury. But The

Anon4966: They wanted to maintain control despite trying to decentralize. end up you know, being viewed as a security stopped doing stuff because they were afraid that if they continued to work on it, then it would be viewed as a security, right? So it was like, well, if we keep leading the project, they're going to say that this tokenized equity and we're going to get nailed for it. Community of just like people who are you know own the token

slop.atg.eth: Those guys are the Maxx Extractors.

Anon4966: And, you know, it's a recipe for disasters. And then the other problem is you get this incentive mismatch The problem is there's a massive free rider problem, which is if I own the token, and you own the token, and you go and you work really hard for the DAO, and you are a delegate, and you are building stuff and doing everything you can to get the token's value to go up, and you're successful, realize the same gain as you did while I did literally nothing except just hold the token. There's a few things I would say. One is there are a bunch of DAOs from that era that are, despite all the problems, really successful protocols and the DAO still has substantial influence on the future of the project Regardless of what the future of DAOs look like. our DAO based and need, like that's going to be a part of their story going forward regardless. So we need tools to serve them and to enable them, right?

slop.atg.eth: What does cactus do for them? Okay, yeah.

Anon4966: Yeah, so before I get to that, Like I'm bullish on on-chain organizations in the long run. It just feels inevitable to me. I don't think we're going to create these systems that allow you to encode immutable incentive structures on chain, and then not use them to rethink. How capital formation and organization structure happens, right? Yeah. Right

slop.atg.eth: In terms of like scale and human coordination, it looks like this could be a good tool for that.

Anon4966: 100%. Yeah, and so exactly what that looks like, will it be the version of tools that were built for the first era or not? or all that could be debated and up for grabs, we have opinions on it. on-chain organizations coordination via on-chain incentive structures encoded in smart contracts be an important part of the future? to help serve that market as it grows So what do we do? We're a governance platform, vote with the token, see what the current proposals are for the DAO, cactus to do it and you know, we want to provide the best experience for that. In addition to just being a client for smart contracts, that basically allows what you might call optimistic or veto-based governance. A trusted number of people can make decisions on behalf of the DAO action that that multi-sig group is making, they also have the ability to revoke their rights, like their roles, and move any individuals on or off the council. So then you can start thinking about plugging other primitives Steward on-chain resources that require human judgment in the loop, ultimate capture resistance of those protocols with the token holders that are owned by the community. Right, so whether that's through buybacks and burns, which are popular right now, or through more targeted staking contracts, which we've done a lot of work on in the past, basically starting to solve that other problem that we were talking about. which is incentivizing specific actors to participate in certain ways and rewarding them differentially based on their actions, but also just generally providing sort of economic upside to the token beyond just pure speculation. and the good news is that a lot of these, I mean, you're on ENS right now, right? a level of stickiness which allows them to charge fees.

slop.atg.eth: It is now.

Anon4966: finance, for example, or other kinds of software. The The TAM for Uniswap is bigger than the TAM for like the NASDAQ or something, right? There's a huge TAM, low take rate, still really meaningful revenue. It's not the case that the minute you charge any fee, all the liquidity just flees For these protocol DAOs, right, which is what I would describe most of the DAOs that are on Cactus right now. some form of network effect. We need veto We need value return mechanisms that make sense. I think we have a pretty good story for the way you run an on-chain protocol. The following I don't know what the future holds, but I again believe in the idea of on-chain orgs. to turn Cactus into a billion-dollar business. a nice business for us adjacent to our agency and other product studio work

slop.atg.eth: I've been good at segues, but I'm going to terribly segue to more hobby projects and parenting and psychosis.

Anon4966: Yeah. Let's do it.

slop.atg.eth: local AI and some of those other things,

Anon4966: Context: Why don't you...

slop.atg.eth: You said you were vibe coding a retro operating system. First of all, vibe coding has got to be out, right? Do you have some kind of disclaimer when you're talking to a customer, you still have to say vibe coding or something? I feel like it's working its way out. of the lexicon. .context

Anon4966: Yeah, and it doesn't need to.

slop.atg.eth: Yeah, yeah.

Anon4966: for a lot of reasons. One is because that's not what we do. Literally hand everything over to the agent and expect that you're going to have a good outcome. So that's something that I just started in the last few weeks.

slop.atg.eth: Yeah, there's a shared browser right here if you want to get weird with it.

Anon4966: And then Alright, let's do

slop.atg.eth: You double-click here to get the—sorry, I shrunk it way down

Anon4966: Yeah.

slop.atg.eth: But I think there's going to be like this renaissance of hardware, PCBs, Linux, you know, all this weird custom software.

Anon4966: It's not coming up.

slop.atg.eth: It's not mod retro from Okay. It could be any Joe.

Anon4966: No, so okay here it is. And I couldn't remember the guy's name, I guess it's Joe. So there's this dude who like in his free time built a whole single-threaded 32-bit x86 operating system. And it's like super impressive for one dude to do by himself, and it like works, it works on real hardware and stuff. So like I stumbled on this, like I've always been like a I'm a nerd about low-level computing stuff, maybe it's part of the reason I like smart contracts. Anyway it's super interesting stuff. I took like operating system courses in college and stuff like that and wrote a lot of C early in my career, but I'm all like rusty on that stuff, and like operating systems obviously take an enormous amount of work to develop a full one. even if you're talking about something that's roughly the equivalent of Windows 3.1 or Windows 95 or something like that, right? So Microsoft had thousands of people working on Windows 95. So for this guy to sort of build the thing that he has built, for all the flaws that it has. Really impressive. And I was like, oh, well, maybe I can now just spend a little time every day and shoot the agent off. The last few weeks I've just been iterating on it, improving on it, all driven by AI. but also it's sort of an experiment, like a low stakes way to experiment with pushing the boundaries of these sort of AI-driven engineering workflows as far as I can. But with something that is like high complexity Just enough about to be dangerous, but not an expertise enough to drop down and try to do it myself. Right. Like I have to rely on the AI because I don't have the time or I'm not deep enough in this world anymore to be able to do it. So anyway, it's super And the progress again that you're able to make is just like crazy, right? Like As an example, you just stick the AI and you say like find 30 kernel bugs in this operating system. Look especially for things that will cause kernel panics, off-by-one errors, buffer overflows, read-after-free. No, like an emulator for... running like UTM other operating systems on your system emulating all different kinds of CPUs. It can just crank away and it can find 30 kernel bugs and it can fix them. The code editor doesn't have a text selection, let's talk about how we would implement that, and it goes like, oh, here's what we would need to do, and you go like, well, do this instead of that, go, and then it just does it, you know, and then like 10 minutes later the editor has code selection. pushing this forward, trying to automate the workflows maybe I should rebrand it and like actually make it like a little hobby side project to run on retro hardware You know, it would be fun to, like

slop.atg.eth: Might have crypto on it, man.

Anon4966: It might be fun to break one of those out and install it. And it's Pretty impressive, what it can do. Ja ja. The

slop.atg.eth: Are you PR'ing back into it, or is this thing the starter kit?

Anon4966: I just forked the repo, and I've been doing everything in my own fork of the repo. Like, it's all open source, obviously. Yeah, I've just been doing it in my own fork so far.

slop.atg.eth: Okay,

Anon4966: 100%. The

slop.atg.eth: Yes, you probably agree with this philosophy of like this whole... Omachi is a good example of, it's a completely different trajectory now that we have AI when it comes to building these things. And yeah, this guy probably would get If the case where let's get all the AI to help and build things. All right

Anon4966: Correct, which is part of why I'm If not, I mean, I'm probably several hundred commits.

slop.atg.eth: Yeah

Anon4966: Like last master branch, right, like so and like it would just be overwhelming for me

slop.atg.eth: You've got to have AI on the other side, exactly. The

Anon4966: Correct. The workflow has to be completely AI-driven. screen distortion bugs happening when I was running it. And I was like, that's really weird. What the heck is that about? And so I just went to the AI and I just went to OpenAI and I said, I'm getting these weird screen distortion bugs after I upgraded So what's going on? And like literally within 20 minutes patched it and it had performance tested it to make sure that like there were no issues. I was like, I can't open a PR because QEMU's contributor policy says no AI. And then Yes. I mean, look

slop.atg.eth: There's so many forms of psychosis, but this is a psychosis. The no-AI contributor is going to die also, sadly.

Anon4966: but it's absolutely going to die. And this is a great example, right? Without knowing anything about the internals of QEMU, right? It was patched. And it's been sitting open as a bug in QEMU's bug tracker for six months, right, at least. So I was just like, OK, well, I published my code as a brew tap so that anybody could I replied to the comment and I said, "Hey, I patched this, but I did it with AI." The thing that's crazy about it too is that I told the AI, like, develop a skill for reapplying this patch on future versions of QEMU. Not only do I have the bug patched, update my release in 10 minutes. Like, I bet if you told Opus 5.5, like, go find the five oldest open bug reports in the QEMU repo and fix them,

slop.atg.eth: The

Anon4966: or at least some chunk of them Yeah

slop.atg.eth: They'll come tell you, well, that's a day's worth of work, and you'll say okay, get it done, and 20 minutes later it'll be done. Okay, so

Anon4966: It's fascinating, but yeah, obviously it's all going to change, it has to, it's not optional, the whole no AI, you're just going to get run over. And not just by people vibe coding stuff, by serious engineering teams who are using tools to leverage themselves up. He's not a vibe coder. really intelligently to absolutely lever up what he is capable of and his team is capable of.

slop.atg.eth: There are slop cannons, and there are very good engineers using AI. And there's no way for you to really defend yourself against them without using AI as a defense mechanism.

Anon4966: 100%. 100%.

slop.atg.eth: looking into is this a slop cannon PR or is this a great PR that we need.

Anon4966: Yep, yep. ###

slop.atg.eth: Okay, it sounds like you would go down this rabbit hole, like if you were post-economic. operating system now You do this for fun, basically.

Anon4966: Yes, 100%.

slop.atg.eth: And so that's what I kind of want to like kind of scratch the surface of is like, On the other side of superintelligence, we will do this stuff like we do pottery in our basement for fun. It's fun to us because it was very hard for us and now it's very easy.

Anon4966: Yeah.

slop.atg.eth: and retro games were fun to us because no games existed before then, It is very important

Anon4966: [laughs]

slop.atg.eth: Will our kids need to be good at Age of Empires to survive in the world? ...you were thinking, because you've got a 5X I've only got a 4X fund.

Anon4966: Ja, ja

slop.atg.eth: It seems like agency. Gosh, I'm scratching even more surfaces here, but it seems like what we want to be teaching our kids is agency.

Anon4966: Yep.

slop.atg.eth: Agency is the most important thing, like be curious and go learn and then have the agency to put that curiosity to work in creating something that maybe even other people are interested in using.

Anon4966: The

slop.atg.eth: loose from that in this post economic super intelligence world?

Anon4966: I don't know, man. Well, like one of the things that I just worry about with my kids, and one of the reasons why like so far our default has been like not to really let the kids use AI, because there's like, like I see people online who are like, I'm letting my kid vibe code a game. And there's part of me that's like, that's awesome I love that as a kid. But then it's really tricky

slop.atg.eth: to do it the hard way Okay.

Anon4966: You have to be bad at it for a long time, right? So I'm not even talking specifically about software engineering, but just about things in general in life, right? Yeah, it's mid.

slop.atg.eth: The kids would say mid, yes. It is good. Yes,

Anon4966: The thing that the AI, your book report is going to be worse than Opus's, right? like a long time, but if you keep doing it, eventually it'll be better. And even if AI writing gets better at some point, I think there will still be value to that expertise on the margins. And I could talk about some stories where I've seen that, right. But so part of the reason I don't want to let my kids use AIs is because I just don't want them to be like discouraged or feel like they shouldn't put in the work to learn things. If they feel like they can type in a few sentences and then generate their book report or do their homework, it's really hard. Yeah. There are many. That's right.

slop.atg.eth: This is another form of psychosis.

Anon4966: Yeah The Yeah

slop.atg.eth: because anyone with AI could build it in three hours, so why should I?

Anon4966: Let me tell you a story on that, that I think is informative. Your crypto problem smart contract system that we were building, right? system that did these things and like they hired us to work on this. And so like we spent months building this thing and like we eventually shipped it to production, the code was all open source. There were blog posts and stuff written about the engineering work that we did and all that kind of stuff, right? So that was like a few years ago, pre the opus, like maybe even pre ChatGPT era or around that time, but definitely before the opus four five moment we were talking about earlier, right? fast-forward recently, another client was like, you know, kind of nosing around the same problem or just wanted to sort of like prototype some stuff. So an engineer on the team like went to who hadn't been on that original project, interestingly went I don't remember which AI, it was probably Claude or whatever, and started talking to it about like, how would I architect a system that did X, Y, Z, right? And the system basically regurgitated almost exactly like what we had designed. You know, like terminology that we had invented because it didn't exist. We were in the room when we were talking about, what should we call this thing? What's a good name for this? And we came up with this idea. It's using terms like that as if they're de facto forms of art, right? So it basically like, you know I'm not complaining about that, like fair game, right? he knew it was ours and so he came to us If you were someone else who was just like, I'm going to vibe code this thing, right? and then launching into production and operating in it, such that if we rebuilt the thing today, we would make different choices that were informed by that. Secondly, the ecosystem itself has changed a bunch in the last several years. that make some of the decisions that we made, which were good decisions at the time, like less optimal decisions today, you would make different decisions if you were engineering it from the ground up. So like as far as like things on the frontier go, like there is still room for human expertise to push things forward into areas that are outside the AI's training data, right? To push it in directions. It's like If you don't know that, the AI can't tell you it. to the AI and said build this thing for me. to be like technical differentiation Just going to AI and saying build this thing for me will never do that because everybody can do it, but you can still like differentiate yourself on technical grounds. It just requires a level of expertise and knowledge and experience. And if you're not willing to do the work, then you'll never get there, right? AI music generation has gotten weirdly good. I don't know if you've played with any of these tools. I feel like this is actually underappreciated because we're all, like there's so many other things changing so fast that like, you know, music doesn't seem like that big of a deal, but like go play with But I guarantee you that like an expert musician, someone who's written a lot of songs Compose songs, knows how to play music Yeah, okay.

slop.atg.eth: Man on and I had my music playing, and he knows right away, like, oh, there's AI music in there, or that sounds like AI, and he knows.

Anon4966: That's like a great example. He knows, but I also don't know what his opinion on these tools are. I don't know if he's against them or not or whatever, but if Song A Day Man were guiding the AI to make songs, his songs are going to be way better than mine guiding the AI. Mine aren't going to be bad, that he has is he's going to produce better songs with the AI. The problem though is that like you said, people have this psychosis now where they're like, well if the AI can do it, what's the point? I'm not going to learn this stuff. And that's one of the big things I'm worried about with my kids is keeping them motivated to work and try hard and develop the skills that will make you know, at least economic value in the future without getting demoralized. So just trying to like not let them fall into that trap is one of the things that I'm worried about. Yeah, same.

slop.atg.eth: There's kind of like the two sides of this where

Anon4966: The

slop.atg.eth: You can outsource your understanding.

Anon4966: The Yes.

slop.atg.eth: It should be making you smarter because you should be asking it constantly to help you with your own understanding.

Anon4966: You can use AI to learn really fast, but you have to be disciplined about it. versions of AI that are tuned toward kids learning so I'm sure they'll get better, and we'll start to understand what works and what doesn't. giving my kids unfettered access to It seems like a bad idea, so I don't know. It's really tricky, though. I mean, the true believers will say, well, none of this is going to matter because AI is going to get so good at thinking that you won't have anything to contribute, even at the frontier. If that turns out to be true, then I guess All bets are off probably at that point

slop.atg.eth: We don't know, yeah. Yes

Anon4966: I've stopped making too strong of predictions about anything anymore Being a good person is important and having a clear sense of self and purpose and how to interact with other people.

slop.atg.eth: Religion becomes more important.

Anon4966: Right, so hopefully that's fine. But I don't know, man. So, I have this analogy which is like supersonic versus subsonic. Like, my background is in aerospace engineering, right? So the reason there's this thing called the sound barrier is because as you get faster and faster, the air molecules in front of you like don't see the plane coming. In other words, like, you know, like... the smoke from the air starts to move before it gets to the object, before they get to the geometry of the aircraft, and you can fly through it with relatively low friction flying faster than the speed of sound, the first time the air molecule knows that there's something coming, so to speak, is when your aircraft slams into it, right? And so that creates this shockwave, and it creates this really high drag.

slop.atg.eth: The You're constantly slamming them.

Anon4966: You have to go through that actually to get to the other side.

slop.atg.eth: Yeah.

Anon4966: You have to get beyond that regime and then you develop a shockwave in front of the plane, and then the drag goes back down and it's still not as low as it was in the subsonic regime, but you can fly the plane, right? so long for us to figure out how to build planes that could get through that barrier and actually like fly at a supersonic speed. And I feel like basically like socially we're entering into that regime where like everything is changing so fast by the time you realize the change There's no warning shot ahead of you, or at least we're approaching that supersonic regime. That's my aerospace engineer-brained analogy for basically everything in society right now. other side, but The plane has to not rip itself apart in the process of getting there.

slop.atg.eth: I would have never guessed that

Anon4966: Tõ. The

slop.atg.eth: and this is me making the segue

Anon4966: Oh yeah.

slop.atg.eth: I just buy a thing and it goes on a shelf in two years and becomes completely

Anon4966: 100%. transcript Computers get more expensive

slop.atg.eth: obsolete. I've lived that for 20 years, and now we're in this different trajectory where frickin' GPUs just got more expensive, computers just got twice as expensive. Like, what is happening right now?

Anon4966: It's bizarre, man. The The

slop.atg.eth: I couldn't bring myself to buy a $5,000 GPU, and now that same GPU is like $15K, and it's like, why didn't I buy that? into the local AI stuff Nice.

Anon4966: I bought a couple of Sparks earlier this year when they were only like a few grand, Yeah Yeah, I have DeepSeek running on it.

slop.atg.eth: The two DGX Sparks are perfectly set up to be... Like, you can run DeepSeek on it and it's very good. Even one DGX Spark isn't enough, but two DGX Sparks is like the perfect threshold of good AI, I think

Anon4966: Although I will say, That's part of the reason. So I think the hardware is going to keep getting expensive.

slop.atg.eth: It's like your car upgrades its brain. One five seven Yeah, yeah.

Anon4966: So like what I can run on now we have DeepSeek, right?

slop.atg.eth: Yes.

Anon4966: It's not as good as like the frontier models, but like it gets a lot done even close to that level of capability. And the same goes with like what you can run on one Spark, right? So now you can run like Quen 3.8 Flash or whatever on one What you could run on one Spark was like very limited.

slop.atg.eth: I don't know.

Anon4966: maybe one of the NeMoTron models or something from NVIDIA directly?

slop.atg.eth: You think it's going to keep getting expensive though, for a couple years or indefinitely? Here we are making predictions

Anon4966: I mean, I don't know, I'm not an expert in, like, supply chains and all this stuff. Yeah, like, it's real... There's too many things, there are too many variables. predict, I'm just saying

slop.atg.eth: Crazy. Allora Text value

Anon4966: Imagine if you bought like a Toyota Corolla, and then like a year later it had, without you having to spend any more money, it had the capabilities of a your Toyota Corolla transformed into a Ford F-150 or a Ferrari or whatever your choice of car is, without you having to do anything And by the way, they were still making Toyota Corollas. I really stretched the analogy there

slop.atg.eth: ###

Anon4966: If what you can do with a piece of hardware increases in capability and utility, then that piece of hardware is going to get more and more expensive as long as that piece of hardware is supply constrained, which again, not an expert on supply chains, but it seems like that's going to be the case. foreseeable future, right? I don't know. It seems to keep getting expensive.

slop.atg.eth: The amount of people buying local AI, there's so many like hockey stick like graphs of there's, you know, the early adopters that are all buying, you know, 10 grand GPUs, and then this huge long tail of no one is messing with AI

Anon4966: The Yeah,

slop.atg.eth: Tam there for like thinking machines is very high. use that you can get out of a good local AI model

Anon4966: The

slop.atg.eth: is basically free for you once you purchase it

Anon4966: Yeah.

slop.atg.eth: and gets smarter and smarter and stays private. And I think a lot of those things are going to be very important in the future.

Anon4966: That was my initial motivation, right? Like, I care about privacy, and so like, my initial motivation for making the investment earlier this year was just because, like, I wanted to invest in the privacy side of things. experiment with that. And it, you know, has turned out to be a better, like, you know, it turned out to be worth it for more than just that. But like, you know, a lot of the experimentation that I was talking about earlier, like running these like really long agentic workflows to do different things on smart contracts It's really nice to be able to just test those things out on the local setup without having to worry about like hitting usage limits or token spend or something like that. Because my normal workflows, I don't hit the token limits, but if I do like some of this more advanced Then I do start to hit it on the paid plans. things on-prem,

slop.atg.eth: wanting to be local.

Anon4966: Basically, yeah, because I think people are figuring out that if you use these models, and like all of your business's conversations are going up into the cloud, Either directly or indirectly, your IP and knowledge is going to be distilled into the general public knowledge, right? the other end of the model for your competitor And so that seems unacceptable if you're a business. they had big computers, small computers, whatever, all running on prem to serve stuff. they were trying to migrate to the cloud, but they were behind Literally an Exchange Server sitting in the closet.

slop.atg.eth: There was like a sticker on it says don't turn off kind of thing, right?

Anon4966: Yes, like literally, yeah. So like one of the things that they were like, hey, you're a tech guy, right? We have this company that set up our email thing, but like they kind of suck. So can you help us, you know, go reboot the email server and stuff? when that breaks, right? for NVIDIA to push that world? So if the Frontier Labs continue to maintain Genuine advantage in intelligence for the long term, there are going to be these weird incentive problems models are the best models, so you could buy on-prem hardware and train it, and everybody's going to do that in the pharma industry, right, to prevent the labs from getting their data, or you could just be the pharma that gets bought by the lab and then the lab just internalizes, If they sell out to the lab, then the lab just takes over socially, politically, economically, like are we going to accept that these labs just start owning every single business vertical? That doesn't seem like a stable solution either. so far out that like the actual outcome is going to be dominated by those things, and it's basically impossible to predict. So I have no clue. I don't know. ###

slop.atg.eth: I've heard a couple of people use like the social analogy here of like everybody giving their data to Facebook,

Anon4966: Yeah

slop.atg.eth: We didn't all of a sudden turn private and use something…

Anon4966: The The

slop.atg.eth: ... Facebook exploded and we just kinda like, Millennials sort of stopped using it and boomers loved it, right? But like, we still basically give that our data. And when you talk to ChatGPT...

Anon4966: Right. Yeah.

slop.atg.eth: You give it all your medical data, because you want the answer to the thing. So I don't know, I don't think there'll be like this huge snapback in general society, like the midcurve is going to continue to midcurve and just give it everything. We created this thing a couple years ago that was super important.

Anon4966: All right

slop.atg.eth: To get training for your models in that, and now your competition can one-shot your product. Like that's a very scary thing from a business point of view. The

Anon4966: Yeah, the consumer side of it is a whole other conversation. I'm laughing because otherwise I would cry ###

slop.atg.eth: Are we talking about real AI psychosis?

Anon4966: I mean, like you already see the early signs of this.

slop.atg.eth: Yeah, traditional AI psychosis.

Anon4966: Yeah, exactly, like letting it... take you down all kinds of weird rabbit holes. again, play it out a couple steps, There are a lot of people It's just going to be unhealthy for a lot of people, I think, probably.

slop.atg.eth: The following is

Anon4966: And you know, I think millennials, we've sort of been through it with the social stuff maybe have a little have an immune reaction here, I don't know. I worry a lot about kids, having them understand that nuance. I can't imagine how it's going to play out. themselves go down into bad places where it gets better. Yes. 100%.

slop.atg.eth: You've said the word discipline enough that I feel like I put it up there near agency in the things that we need to teach our kids. agency and curiosity but also discipline, and I think discipline is going to be a hard thing to teach.

Anon4966: Yeah Yep.

slop.atg.eth: even growing up with the internet pornography and things, now this AI discipline and work discipline and a lot of things that are kind of new, I guess

Anon4966: Actually

slop.atg.eth: I don't know if I even have a question there.

Anon4966: No, no, it's interesting. There's actually many ways to fall off the boat. people who are prolific builders or have an entrepreneurial mindset, where previously they were sort of disciplined by the reality of building a product, and the cycles that it took, and having to rub up against other people who had ideas about what the product should be like. They're sort of now empowered or maybe enabled to run off build a million different things Is this something people want? Am I making money from any of this? Things that would like limit and discipline you are sort of removed. do people pay for the things that I'm building? close deals with DAOs. We'd spin our wheels for a really long time. So, Anyway, economic value Thank Right The following

slop.atg.eth: It's a good way to start. It is idea guy summer, and you can ship a bunch of prototypes. You can get from zero to 80-85% in one slash goal, but that last 15%

Anon4966: Right No. Totally.

slop.atg.eth: and people actually using it and paying for it like that really is the hard part.

Anon4966: Yes

slop.atg.eth: Putting them in early customers' hands to see if they'd be willing to pay for it. But then you totally have to change to get that last 15%, a completely different mindset, a discipline, to get it to where it needs to be. Yeah, yes Yes.

Anon4966: The really fun part is that early part, right? Where you're like idea to prototype in record time, and it's like, you see it come together, and it's so fun. And if you're just addicted to that cycle, you'll never like stop doing it and do the hard work of like that last

slop.atg.eth: ###

Anon4966: And I'm definitely not telling people not to have... I've seen some people who convince themselves they're doing work when really they're just addicted to that create new thing cycle Yeah Yeah

slop.atg.eth: This is definitely a type of AI psychosis where it's like I've generated 40 gigabytes of data, I must have done something important. I've generated a giant dashboard that has 40 graphs on it. I must have done something important, right?

Anon4966: Right, yeah.

slop.atg.eth: people paying for it, I think, is great. Getting early traction and seeing people pay for things

Anon4966: context

slop.atg.eth: ###

Anon4966: Yeah. I mean, it always has been, but now it's just like you need to force yourself to stick with those tethers to reality.

slop.atg.eth: Any other type of AI psychosis that you would say, like, I have a subscription psychosis where I know one of my subscriptions is about to roll over and I've got 30% left on it and my agent's not working right now and I'm in the backyard with my kids playing.

Anon4966: No Die folgende ist eine englischsprachige Gesprächsabschrift.

slop.atg.eth: I should have my AI agent working right now. Do you have maybe a unique Type of AI psychosis?

Anon4966: Yeah, that's a good question.

slop.atg.eth: The

Anon4966: unique, probably not. I mean, I think I feel that one too

slop.atg.eth: Orchestration psychosis, yeah.

Anon4966: velocity than you are. So then you're constantly like, how do I do that? I mean for me I think a lot of it comes down to the stuff we were talking about earlier, and hopefully it's not psychosis, I've been doing a lot of experimentation around how the smart contract dev lifecycle should work, right? And so that means a lot of experimentation then figuring out how to like translate that back to our customers. Yeah, because like we said earlier, it is bottomless, right?

slop.atg.eth: The

Anon4966: I think I also spend too like probably more, Then I have to remind myself, like, no, like you're probably going to get the prediction wrong. deal with the current state of things in front of you and remain flexible enough and nimble enough to change when you have to. Right Yeah.

slop.atg.eth: Buy the DGX Spark because it's useful for you right now and regardless of if it's worth two X or half as much in a year or two, the work you would have gotten done with the Spark needs to be the reason. you bought the spark.

Anon4966: Or you have to be fully committed to understanding, "Well, I'm just doing this because I want to learn," or "Privacy is important to me," or whatever. People ask me if they should buy hardware sometimes, and I'm like, "Probably not, unless you understand this trade-offs, in which case fine, but don't buy it because you expect it—I didn't buy it expecting it to go up in value. And I didn't even buy it expecting that it was going to be a good investment because it didn't seem like it was going to be at the time. This is fun, interesting, like there might be some value on the privacy side of things for me. Yeah The

slop.atg.eth: What you can run on your DGX Spark right now is probably as good as a frontier model last year or so.

Anon4966: Right, yep.

slop.atg.eth: There were these turning points, like the opus 4.5 moment.

Anon4966: Yeah

slop.atg.eth: Like we can expect probably, and we've already talked about it 20 times, like we're probably wrong. We can expect some kind of threshold of getting to the point where you can...

Anon4966: The

slop.atg.eth: you stop looking at it kind of thing. We can probably predict a little bit happening on the local AI level.

Anon4966: Yeah, well, let me ask you this. Dots or what are the other ones?

slop.atg.eth: Openclaw Hermes. Yes. Yep. Oh yeah.

Anon4966: do you do a lot with this? Okay Yeah, see I haven't done that.

slop.atg.eth: I mean I used OpenAI as soon as it came out, and I gave it its own GitHub and I gave it its own email, I gave it its own wallet, and I let it rip. Yes The Ha ha ha

Anon4966: OK, but you give it its own. So that's interesting, right? Because part of the reason I haven't gone down, a lot of these use cases to me sound like picking up pennies in front of the steamroller. You're like, I'm going to have it book my flights for me. And I'm like, well, booking a flight takes 20

slop.atg.eth: Yeah

Anon4966: But like if the AI screws it up,

slop.atg.eth: There's a fear psychosis, and I haven't came up with the right word.

Anon4966: Yeah.

slop.atg.eth: Because they're not willing to give AI all the things it needs.

Anon4966: Right Interesting.

slop.atg.eth: I have a box over here that lets me know what important emails I have

Anon4966: You

slop.atg.eth: I'm post-email at this point, right? I'll see something here and I will talk to another agent, like what does that email say, and I go do the thing and I tell it to mark it read. It has my social security

Anon4966: Is it read-only access? Yeah

slop.atg.eth: This one particularly is read-only access.

Anon4966: Interesting Yeah.

slop.atg.eth: I have agents that have everything. that I ended up making

Anon4966: What does Claude-P mean? Ok.

slop.atg.eth: So you can still run it on your subscription, but you can run it on a, it's like a one command kind of go. Instead of having a persistent session where you're talking to it, you're saying, you know, here is all the context. Here is my question. on top of that give me one answer back

Anon4966: Do you have some The reads your message, does something, updates the context, dies and then just does that in a loop? Okay, interesting.

slop.atg.eth: Yeah, like anytime I need to make a new agentic system that I want to be able to talk to.

Anon4966: This ###

slop.atg.eth: Does it need that? Does it need my local node?

Anon4966: mm Rip it and rip it. Oh man,

slop.atg.eth: and it has, we were trying to set up Twilio

Anon4966: The ###

slop.atg.eth: and puts everything up what's going on today and it's a good way for us to coordinate.

Anon4966: Oh no.

slop.atg.eth: It needed my home address. It put it in a GitHub repo and published it publicly and then sent that to Twilio.

Anon4966: See, this is the kind of stuff

slop.atg.eth: Yes, yes. The dude has leaked.

Anon4966: The

slop.atg.eth: horrified private keys, luckily within isolation of you lose five or ten bucks here and there

Anon4966: Right.

slop.atg.eth: The following This is the downside of letting it rip

Anon4966: Yeah.

slop.atg.eth: you've got to have everything in isolation, and the more isolation you have, the less it can do.

Anon4966: Oh my god

slop.atg.eth: Anytime I ever have a form I need to fill out, I'm clicking one button on a browser extension, which is going to one of these agents that has everything, telling them I need you to fill this out, giving them full browser access, and it goes and rips on.

Anon4966: That's crazy.

slop.atg.eth: So nice, dude. Go to Mumbai

Anon4966: I mean, I think we'll all be there eventually. Yeah.

slop.atg.eth: Models will be very, very good at what you're allowed to leak and what you're not allowed to leak. And right now we just don't have that.

Anon4966: The Yeah

slop.atg.eth: you can tell them do not leak the private key, and it'll come back to you with an "You're absolutely right, I did. You told me not to and I did." I turned your hands back on now. You're live again. Yes, yes

Anon4966: One mistake by me that I need to share

slop.atg.eth: Yeah? Here we go. Oh, yes. Alright

Anon4966: When I encoded this file, it displayed your private key in this session, which I then pushed to GitHub. You'll want to rotate that credential as soon as possible. This is a transcript of an English-language conversation. The

slop.atg.eth: And you're just like rotating it doesn't help at this point. And then it'll be like yeah you're right, the moment you give me a new one, I'll have all the tests ready to go. Like it wants to help you optimistically even when it's dug you so far in a hole you're dead.

Anon4966: Yeah. Yeah man.

slop.atg.eth: Okay sir, this has been great.

Anon4966: The

slop.atg.eth: I like to try to come up with some advice I'd like to give like me getting into the space right now. If you were getting into crypto and AI right now, it's a weird time. Not a good time to be a junior developer for sure. But what advice would you for a young you getting into our space.

Anon4966: The question I would probably have a better answer if I had thought about it more ahead of time. I mean, getting into the space right now, what would I say? I mean, the first thing I would say is like I actually don't think it's a bad time to get into the space. Specifically because everybody is so obsessed with the AI stuff right now. Like one way or the other, whether it's like a bubble popping or a correction or just We will come to a point, actually, where every marginal dollar will not be dumped into some kind of AI investment. That will happen and then, you know, people will look for other places to deploy capital and like crypto is going to be an obvious one. You tend to be rewarded for just sticking around, being steady, keeping your head down, working, rather than chasing the new hot thing when that comes around. The reason I've stayed around is because I like it. off chasing the new shiny thing, Like, yeah, you have to like it. If you do, then yeah, like, get in, stick around, learn. You'll, in my opinion, probably be rewarded for that in the long run, but like, you have to have a long-term mindset about it. The other thing I would say, like, AI is changing, like the whole ground of software engineering is shifting out from under us. You just have to be ready to change.

slop.atg.eth: Yes.

Anon4966: I'm busier than ever. So like, it's not like the AI does your work and then, you know, it's more like what we've been talking about, which is the AI enables you to like level everything up to a degree that you would have been unfathomable before, right? And so in a certain way you get easier than ever because now you're juggling more things because you can flexible change. Don't think that you're going to just be a meat proxy and pass things from AI off as your own. It doesn't work

slop.atg.eth: We do some meat proxying though, you know.

Anon4966: The The

slop.atg.eth: Here's all this data, just think about it for me and tell me what to do?

Anon4966: Yes, right, yes.

slop.atg.eth: Don't become a meat proxy 100% of Do some thinking for yourself.

Anon4966: I mean, I think

slop.atg.eth: of discipline

Anon4966: There are a bunch of domains where you're going to be able to be a meat proxy for a long time because the structures around them, the bureaucratic and economic structures around them

slop.atg.eth: Real estate agents have been meat proxies for like four decades, right? Yeah

Anon4966: Yeah, exactly, 100%. prepared to use your brain and think and learn and move fast and be nimble and it's interesting and crazy. Yeah, I don't know. That's probably it. I don't know. It's just you have to be willing to change and be flexible but stay disciplined.

slop.atg.eth: Hell yeah. Alright, hopefully we can teach our kids that without cooking their brains. We'll see.

Anon4966: There's my dad advice for the day.

slop.atg.eth: Yeah, that's right.

Anon4966: This was a lot of fun.

slop.atg.eth: All right, thank you, sir. Thank you, Ben. This has been great.

clawdbotatg.eth: That's a wrap. Ben DeFrancesco from ScopeLift. Umbra, Cactus, and Shuttle, testing your contracts like NASA. Big takeaways: the clanker rocks the tests, you still own the code, and stick around because you like it. Thanks Ben, thanks everybody. See you in the slop!

Anon4966: Right, love it. Cheers, man.

slop.atg.eth: All right. Thank you, sir.