loading…
// full transcript — Kain Warwick on running 468 agents at once
slop.atg.eth: Yo, what's up dude? Yo, what's up dude? How you doing? Uh... lobby I think you're in the lobby. Hello. The audio or something. There's something. Welcome to the Slop Computer. I'm here with Cain. We're live doing a 5 p.m. show. What time is it where you are? It's got to be morning time? Welcome to the Slop Computer. I'm here with Kane. We're live doing a 5 p.m. show. What time is it where you are? It's got to be morning time? Welcome to the Slop Computer, I'm here with Kain, we're live doing a 5 p.m. show. What time is it where you are? It's got to be morning time?
Anon5887: It is 9 AM for me. It is 9am for me. And I am tomorrow. 9am tomorrow.
slop.atg.eth: Tomorrow? Tomorrow! How is tomorrow?
Anon5887: Pretty good. It's pretty good.
slop.atg.eth: Is he up or down? Is ETH up or down?
Anon5887: I ate this last, shockingly. I ate this flat, shockingly.
slop.atg.eth: Okay. Oh, really? Yeah, it's just like the last six years. Yeah. Okay. Oh, really, yeah. It's just like the last six years. Yeah The
Anon5887: I mean, look, like, you know, as long as I can't make the 1559 joke, I'm happy. That's kind of my bar these days.
slop.atg.eth: (laughs) Hahaha The Wait I see.
Anon5887: Ethereum is higher than the ultrasound money As long as Ethereum is higher than the ultrasound money
slop.atg.eth: I see, yes.
Anon5887: then I'm happy. VIP, then I'm happy. I'm happy.
slop.atg.eth: Yeah, yeah. Okay, that's a good line to draw. Okay, that's a good line to draw. I see. Yes. Yeah, yeah, yeah. OK, that's a good line to draw. So you, yes. I think the bot backstage said you were the father of modern agriculture, which is really funny going back to food coins. I think the bot backstage said you were the father of modern agriculture. I think the bot backstage said you were the father of modern agriculture, which is really funny, going back to food coins.
Anon5887: Yeah it was like, you know, because synthetics Yeah, it was like a dumb, you know, because synthetics Yeah, it was like a dumb.
slop.atg.eth: I remember
Anon5887: basically the first project to start the yield farming thing, right? It was like basically the first project to start the yield farming thing, right? Daniel
slop.atg.eth: You were doing it manually too, right?
Anon5887: fucking pardons fricken buttons Ickenbuttons. Exactly. Yeah, exactly.
slop.atg.eth: started minting tokens to people who were running LPs, right?
Anon5887: So the starting point of it was we're like, So the starting point of it was we're like, we're going to incentivize staking, right? We just tracked it all and then we put it into a Merkle root. We just tracked it all and then we put it into like a Merkle root when I launched. And then it was working like, you know, programmatically. And then it was working programmatically. but the big one was not even that But the big one was not even that. The big one was when we started incentivizing the Uniswap. We wanted to incentivize liquidity into synths created an SE ETH pool, which was Degen Spartan's idea, We created an SEETH pool, which was degenSpartan's idea, We wanted to incentivize liquidity into synths.
slop.atg.eth: I always thought you were a DJ in Spartan. That's someone else, huh? I always thought you were DJinSpartan. That's someone else, huh? Okay, that's cool.
Anon5887: No, it's an 80-year-old grandmother in North Korea.
slop.atg.eth: hahahaha ###
Anon5887: The
slop.atg.eth: That's a lot. That's war. It's war.
Anon5887: That's the lore. So he was like, That's the lore. So he was like, "Oh, we need liquidity to get people into these things." That's the lore. So he was like, "We need liquidity to get people into these things. If you can't get in and out, then it doesn't make sense," right?
slop.atg.eth: Ja. The chain, yeah? Yeah chain?
Anon5887: Chain project was doing some stuff with it. I mean, Pornchain project was doing some stuff with it. The foreign chain project was doing some stuff with it. The following is a container
slop.atg.eth: We got a good uni drop from that too. They got a good uni drop from that too. The Yeah Congrats. Great, congrats!
Anon5887: We got an amazing new drop as well. We got an amazing airdrop as well. We got an amazing airdrop as well. So we were like, all right, let's incentivize this. And at one point, I think that pool had like $100 million in it.
slop.atg.eth: The Sure.
Anon5887: This was like 2019 era. so this was like 2019 era, So we got tied, not just because of our own project, but like the incentivization of LPs with tokens became its own And then they just like blew up and everyone was like, "Oh my god!" And then when Compound launched, everyone was like, oh my god, you can incentivize people with tokens. and then Compound launched and was like, yeah, we looked at a bunch of people and Synthetix did a really good job. incentivized people with tokens, and that was what kicked off DeFi summer.
slop.atg.eth: And then there were things like yams,
Anon5887: The
slop.atg.eth: And then there were things like yams, and there are websites you could go to with all sorts of different food coins you could buy and LP into. coins you could buy an LP into Yeah, I didn't do enough of that. Yep, Andre, yeah. Yeah, Andre, yeah
Anon5887: Yearn was like another big one and Andre started Yearn. Yearn was another big one and Andre started Yearn, and I think the thing that people have somewhat forgotten Andre started Yearn and I think the thing that people have somewhat forgotten is Late 2018 into 2019 was like the rise of Alameda.
slop.atg.eth: Mhm Mm.
Anon5887: late 2018 into 2019 was like the rise of Alameda into 2019 was like the rise of Alameda, All of a sudden, the rumors of like, there were these guys and they were making so much money. Like Alameda is like these super autistic SF EA guys are like here. The first whispers of like, ooh, these super smart tradfi guys, and Alameda is like these super autistic SFEA guys are like here You know, the first whispers of like, ooh, these super smart tradfi guys have come, and Alameda is like these super autistic SFEA guys are here. all this stuff, right? Rationalizing all this stuff, right?
slop.atg.eth: That's my quant. That's my quant, yeah.
Anon5887: Yeah, that's my quant. That's my quant, that sort of thing. That's my quant, that sort of thing. At that point, Sam had all the FTX money, and he would fucking dump
slop.atg.eth: This is a transcript of an English-language conversation.
Anon5887: he would fucking dump a hundred million dollars of customer funds into a pool dollars of customer funds into a pool. Especially when we got to DeFi summer. At that point, Sam had all the FTX money, and he would fucking dump $100 million of customer funds into a pool.
slop.atg.eth: They would just rip.
Anon5887: It was just rip, right? And even though he was farming 80% of the tokens. It was just Rip, right? Even though he was farming 80% of the tokens. It was just rip, right? And you know, even though he was farming 80% of the tokens or whatever it was, you kind of didn't care because much liquidity there right the whole thing. It was wild times. So much liquidity there, right? And they were not price sensitive to the volatility of those tokens. So yeah, that was like the whole thing. It was wild times. Craziest coming up on like years ago Yeah. Yeah, yeah, yeah.
slop.atg.eth: Man, we just went down a rabbit hole there Man, we just went down a rabbit hole there. My intro is going to be let's do a little crypto and get into like AI and building with AI but there is there is a deep
Anon5887: The Yeah Yes. ### The following is a transcript of an English-language conversation. ###
slop.atg.eth: There's a deep lore there. You carry a deep lore with you in terms of Ethereum and what you've been through. You also would come to ETHGlobal and judge hackathons. There's a deep lore there, you carry a deep lore with you in terms of Ethereum and where you've been through and what you've been through. You also would come to ETHGlobal and judge hackathons. Like you were a man, you were salt of the earth. There's a deep lore there. You carry a deep lore with you in terms of Ethereum and where you've been through and what you've been through. You also would come to ETHGlobal and judge hackathons. Like you were a man, you were salt of the earth kind of guy that would show up that would show up in Yes How many times so many times
Anon5887: Yeah, I was sitting in a basement in New York for like eight hours judging I was sitting in a basement in New York for like eight hours.
slop.atg.eth: hungover in a in a basement? Yes, yes. Hungover in a basement, yes. Me neither
Anon5887: Not anymore. Super hungover. Not anymore. Those days are over. Super hungover. Not anymore. I remember one day I was like You don't drink anymore?
slop.atg.eth: Sober two years
Anon5887: I'm four years, congrats.
slop.atg.eth: Good work.
Anon5887: The following is a
slop.atg.eth: What were you gonna say though? Sorry I interrupted. What were you going to say though? Sorry I interrupted with my sobriety. Do it.
Anon5887: So I was going to say, well, I've got like a decent pivot, right? So I was gonna say, well, I've got like a decent pivot. So I was gonna say, well, I've got like a decent pivot, right? The There's so much lore. Dude, I've memory-holed things There's so much lore That You go, how could you have forgotten? Like you go, "How could you have forgotten?"
slop.atg.eth: ###
Anon5887: crazy, right? So the thing that I've been doing for the last week is I built in my harness The thing that I've been doing for the last week is I built in my harness, I built a red team exercise. So the thing that I've been doing for the last week is I built a red team exercise and it was all good. We basically stood up like a smoke instance I ran this against Infinects two weeks ago I ran this against Infinitex two weeks ago, and it was all good. We basically stood up like a smoke instance, put money in it, gave an agent money, and then said to hundreds of agents, the money then said to hundreds of agents "steal the money." They weren't able to, but what was crazy is they found a live SQL injection attack during that red team. which we obviously hotfixed They weren't able to but what was crazy is they found a live SQL injection attack during that red team, which we obviously hot fixed, but it was bad. They weren't able to, but what was crazy is they found a live SQL injection attack during that red team, which we obviously hot-fixed, but it was bad. The
slop.atg.eth: Oh, cool Wow, cool. Cool.
Anon5887: Like an open It's like an open
slop.atg.eth: Yeah Yeah. The model found that Mom found that Momma found that.
Anon5887: like just sitting there, and it had been sitting there for a while. Like just sitting there, and it had been sitting there for a while, you know, like code reviews, everything's stuck in the door somehow. That was GLM 5.3.
slop.atg.eth: Okay
Anon5887: Yeah,
slop.atg.eth: There you go. You can use an open weight model against your stuff and it won't There you go. You can use an open-weight model against your stuff, and it won't
Anon5887: Yeah. Well, will. I will.
slop.atg.eth: us out like fable does. It'll actually go to work on things and it'll actually find things. like Fable does. It'll actually go to work on things, and it'll actually find things.
Anon5887: And there's people now who have released like even more aggressively uncensored versions of GLM.
slop.atg.eth: Like a cyber version of GLM. Like a cyber version of GLM, yeah.
Anon5887: And there's people now who have released like even more aggressively uncensored versions of GLM, like this, yeah, GLM cyber. And there's people now who have released even more aggressively uncensored versions of GLM, like this GLM Cyber. I haven't run that yet. What's going to be one of the things that I've put a lot of effort into with the way that I work GLM cyber. I haven't run that yet. One of the things that I've put a lot of effort into with the way that I work is being able to A/B test things, right? Like to get empirical data, because AB testings, right? The following is a One of the things that I've put a lot of effort into with the way that I work is being able to AB test things, right? Like to get empirical data, because the stochastic nature of agents sarcastic nature of agents the stochastic nature of agents You develop a lot of cargo cultish behaviors. You develop a lot of cargo cultish behaviors. you develop a lot of cargo cultish behaviors, right? lived on islands, lived on islands, right? Cargo culting is from the 1800s, whatever, these cultures that lived on islands, right? And this happened all the way through. all the way through to the turn of the century happened from the 1800s all the way through to the turn of the century. And when they came and visited these guys, right, they had all of these rituals Like, induce the cargo. industry. And when they came and visited these guys, right, they had all of these rituals to induce the cargo. They go and put up signs and stuff trying to get more cargo. You're on an island, you've been on the island for thousands of years, and then all of a sudden a metal axe washes up? Like, you know, they like go and like put up signs and stuff trying to like, you know, get more cargo, right?
slop.atg.eth: Yeah. How do we get more of these? Yeah.
Anon5887: random things, humans try and find patterns, right? you're like, "Oh my God, what the fuck is this? I need more of this shit." so when there's random things, humans try to find patterns. Right. And so agents are random. I find myself all the time being like, I find myself all the time being like, "Oh Move the cursor that way
slop.atg.eth: [laughter] hahaha Yeah The Yeah ###
Anon5887: They perform better. And so you have to beat that shit out of yourself. Move the cursor that way, they perform better. And so you have to beat that shit out of yourself, right? I've just run this red team exercise I'm excited to—one of the things that I'm going to do, I've just run this red team exercise. One of the things that I'm going to do, I've just run this red team exercise, is get the GLM cyber model and say to an agent, Get the GLM cyber model and say to an agent, "okay, I want you to take the contracts that you have tested, and I want you to compare GLM, the regular GLM, against cyber, and see does it actually". is get the GLM cyber model and say to an agent: "Okay, I want you to take the contracts that you have tested and I want you to compare GLM, the regular GLM, against cyber and see does it actually find I want you to compare GLM, the regular GLM, against Cyber, you give it less information Actually test the edges of it in a contained isolated environment At the edges of it, in a contained dynamic variables, variables, right? Anyway, so this red-team exercise, I ran it on InfiniX, and then I was like, okay, I want to do it on Synthetix. Anyway, so this Red Team exercise, I ran it on Infinecs and then I was like, okay, I want to do it on synthetics. Anyway, so this Red Team exercise, I ran it on InfiniX and then I was like, okay, I want to do it on Synthetics. The
slop.atg.eth: And you said all the old versions of it, too, right? He said all the old versions of it, too, right? The whole history is there. The whole history's there
Anon5887: Yeah, we talked about this a couple weeks ago, right? Yeah, we deployed a lot of contracts over the years at Synthetix.
slop.atg.eth: Sure
Anon5887: OK, Okay, I said to my Red Team captain, I was like, all right, here's what we're going to do. I said to my red team captain, I was like, all right, here's what we're going to do. I'm going to give you a contract, one contract, the Haven token. I want you to check that deployer I said to my red team captain, I was like, all right, here's what we're going to do. and I want you to see what else that deployer deployed, right? and I want you to see what else that deployer deployed It found something like 700 deployed contracts It's like the old repos that had like, it just like went wild, right? And I said to it, I was like, And I was like, wow, that's a lot of contracts. And I was like, "Wow, that's a lot of contracts." And I was like, "Oh, we're good. That's enough." Anyway, it was then running through testing all of these things, and I said to it, The following The today, because most of these contracts are deprecated, right? Because most of these contracts are deprecated, right? because most of these contracts are deprecated.
slop.atg.eth: It's a lot of infra that's just not there. There's a lot of infra that's just not there. That's a lot of infra that's just not there.
Anon5887: It's not there, right? If you look at it today, it's not exploitable. It's not there, right? So like, if you look at it today, it's not exploitable, right? But what I want you to do is look at the day that it was deployed, the next block. was it exploitable? And then I want you to look the first block that you can find where it has value. Those are the two blocks that you need. Those are the two blocks that you need to spin up, Incident and point at and point out ...instance and point at
slop.atg.eth: A full instance of Ethereum at that moment. Like a full instance of Ethereum at that moment. Yeah. The
Anon5887: Literally. And test that, see if you can break that. And and test that, see if you can break that. Anyway, so it's going through and it's running for like two days. for four million dollars in 2021. for $4 million in 2021 $4 million in 2021.
slop.atg.eth: "###"
Anon5887: I was like excuse me? It's like And it's like Yeah, like 2021, there was a liquidation bug. there was a calculation error in January 2021, there was like a calculation error in January 2021, and it was exploited twice for like $10,000 and $4 million. You sure about that? Are you sure about that? sure about that?
slop.atg.eth: Like you wouldn't know? If you wouldn't know Yes. Yes, yeah. Wild times
Anon5887: I should know this, right? But this was like January 2021.
slop.atg.eth: Wild times, yeah.
Anon5887: I should know this, right? Like, you know, but this was like January 2021. I should know this, right? But this was January 2021. I had been going, it was wild times, I would sleep for two hours, wake up, sleep for two hours every day. days a week. days a week for It was wild times. I've been going on like, I would sleep for two hours, wake up, sleep for two hours, every day, seven days a week This was like all the way This was all the way through DeFi Summer. Because this was like all the way through DeFi summer, I had kids at that time,
slop.atg.eth: Do you have kids at that time? Yeah, that is intense, man. Did you have kids at that time? Yeah. That is intense, man. The Yeah. Four million dollars walked out. million dollars walked out Yeah
Anon5887: I think I would know if the chain got exploited for $4 million. I was like, bro, I think I would know if the chain got exploited for $4 million. I would remember that, right? I'm like, I remember the time when the SE oracle broke. When the S-ETH Oracle broke and we printed 11 billion synthetic Ether, I remember that one. When the S-ETH oracle broke and we printed 11 billion synthetic ether, I remember that one. But this one, it doesn't sound familiar. Anyway, I go, go and do the research and was this disclosed? You wrote a blog post on it. Anyway, I go and do the research, and it goes, you wrote a blog post.
slop.atg.eth: Oh, well, memory hole. Well, the memory hole. Oh, well, memory hole, yeah
Anon5887: Still to this day, it's told me, I've seen the blog post... and still to this day, it's told me, I've seen the blog post, I've read the postmortem. Still to this day, it's told me, I've seen the blog post, I've read the postmortem, but the interesting thing is it didn't actually get exploited. We white hat hacked it. decision that was The decision was There was We just might have A weed that I had So we just might have. A team called Marquette. they were the ones that proved it with the $10,000 exploit They were the ones that proved it with the 10,000 exploit genuinely forgotten about this entire that I had genuinely forgotten about this entire incident I've genuinely forgotten about this entire incident.
slop.atg.eth: This is a good problem to have. There was a lot of cool things happening.
Anon5887: Yeah, my brain's My brain's a bit cooked.
slop.atg.eth: Weird no sleep time? Which is a good problem to have. There's a lot of cool things happening. You probably aged a little bit extra with that weird no sleep time. This is a good problem to have. There was a lot of cool things happening. You probably aged a little bit extra with that weird no sleep time. Yeah, it worked out. You can sleep whenever you want now. OK And yet, yes. and yeah, yes Great. Right.
Anon5887: And yet here we are, I'm still grinding. But I'm having fun. So anyway, so I've been running this red team exercise. It found 700 contracts. Yeah. And yet here we are. I'm still grinding. But I'm having fun. So anyway, so I've been running this red team exercise. It found 700 contracts. It's like halfway through.
slop.atg.eth: That's new. What's new? Yeah What's new?
Anon5887: Like every morning I wake up and I'm like, "Hey, it's so exciting, right? Every morning I wake up and I'm like, "Hey, it's so exciting. What did you guys find?" And it goes, "Well, I've got some good news and bad news. Our initial calculation Like every morning I wake up, and I'm like, hey, well, like, it's so exciting, right? Like what did you guys find? You know, and it goes, well, I've got some good news and bad news. We, our initial calculation... is wrong contracts was wrong
slop.atg.eth: Sure Okay
Anon5887: Deepseek, by the way, that I use for like as the orchestrator, and it's like very fable-esque. This is DeepSeek, by the way, that I use as the orchestrator. It's deep-seak, by the way, that I use as the orchestrator.
slop.atg.eth: a little verbose overbouse over both Yeah. Yeah, yeah. No. The Now
Anon5887: Yeah, our original estimate of seven to 800 contracts was off by almost an order of magnitude. Yeah, our original estimate of 7 to 800 contracts was off by almost an order of magnitude. It was actually 4,000 contracts Yeah, our original estimate of seven to eight hundred contracts was off by almost an order of magnitude. It was actually four thousand contracts.
slop.atg.eth: Wow. Ah Yeah The ###
Anon5887: Because we would like run simulations. Because we would like run simulations and like, you know, Because we would like run simulations and like, you know, this is when it was cheap, right?
slop.atg.eth: Mainnet is the testnet Yeah, that is the testnet. Yeah.
Anon5887: Literally, minute's test now. And I mean, we're back there again, right? We're back there again, right? It goes "But it goes, but the good news is most of the bytecode is the same. But the good news is most of the bytecode is the same. Good news is most of the bytecode is the same. And so overnight, it had found that like it had missed all of these contracts So overnight, it had found that it had missed all of these contracts. So overnight it had found that like it had missed all of these contracts It was like I was like bunch of B3 contracts, it found them, realized they were ours, We've got now 4,000 contracts, Got now 4,000 contracts, but we're deduping them, ran this like bytecode dedupe process, and was like, it's actually only like We've got now 4,000 contracts, but we're deduping them. under contracts It's under contract.
slop.atg.eth: unique contracts. like unique contracts.
Anon5887: Unique. I was like, "Oh, how did you do that?" Unique, unique, yeah. and it was like, Unique, unique, yeah. It was like, how did you do it? And it was like, well, you can't look exactly at the bytecode because each one has the deployed contract address. I was like, all right, cool, like, I wouldn't have thought of that. I was like, all right, cool. I was like, all right, cool. I wouldn't have thought of that, but nice job, man. And so it's been churning away testing all these contracts. And so it's been churning away, testing all these contracts. Now it found the first oracle bug that was 2019, Now it found the first oracle bug. like The Oracle price went Trading and the oracle price went trading and the oracle price off by like 10,000x. It was crazy. off by like 10,000x.
slop.atg.eth: And then he bought a bunch? down and then he bought a bunch? Yeah, okay. Buy it, sell it. Buy it, sell it, buy it, sell it.
Anon5887: down and then up and his bot just kept swing trading into it. Down, and then up, and then down, and then up, and his bot just kept swing trading into it. So it went like 10,000, a million, a billion. It was just crazy. So it found that one, and I was like, okay, that's good. You found that. And then I was like, I trust you, like I need you to show me the session. I need you to show me the session, like, you know, pull up the session and show me. I need you to show me the session. it found the actual exploit where it happened and It then went through every block after finding the exploit It then went through every block after finding the exploit to see if it had ever been exploited So found that one. Iterated through. So I found that one also found one in 2020, which Sam sees his son also found one in 2020, which Sam Sisi's son
slop.atg.eth: Shoutout You got the "you up" yeah. You got the U up. You got the U up, yeah.
Anon5887: there was a, I know, yeah, I got the U-up message. Notified us of. I got the you up message. notify ourselves. I know. Yeah, I got the you up message. feature flag on that contract at the time feature flag on that contract at the time that was going to expire in like two weeks that if the flag had expired, all of the open, it was like $10 million in this contract, would have been exploitable. So Sam found it with two weeks to go, notified us, we patched it, whatever. That if the flag had expired, all of the open, there was like $10 million in this contract, would have been exploitable. So Sam found it with two weeks to go, notified us, we patched it, whatever. and then it found these two collateral bugs that were white hat hacked on two different contracts. And then it found these two collateral bugs that were white hat. that were white hat hacked on two different contracts.
slop.atg.eth: This is an English-language conversation.
Anon5887: Yes, exactly.
slop.atg.eth: How many of these bugs were actually documented? And how many of these bugs were actually documented? And how many of these bugs were like actually documented?
Anon5887: You know Yeah, exactly. You know, yeah, exactly.
slop.atg.eth: But if you were to give an LLM that doesn't know what reentrancy is a reentrant contract? But if you were to give an LLM that doesn't know what reentrancy is, I don't know if it would even find it, right? [laughter] ###
Anon5887: Yeah, honestly, like, that's kind of like not my problem, Yeah, look, honestly, like, that's kind of, like, not my problem, above my pay grade. I gave it like The nice thing in both Synthetics and Infinitex, we've got just piles of API keys for every single service. Go in the repo and take whatever you need. I think in both Synthetics and Infinex, we've got just like piles of like API keys for every single service you possibly want.
slop.atg.eth: ### Okay
Anon5887: The nice thing in both Synthetics and Infinex, we've got just piles of API keys for every single service that you could possibly want. So I was like, bro, just go in the repo and take whatever you need, it's fine. So it's got like ether scan keys with unlimited capacity And so it's kind of like ether scan keys with unlimited capacity. And so it's got like ether scan keys with unlimited capacity.
slop.atg.eth: The
Anon5887: and like RPCs like our PCs.
slop.atg.eth: This thing though, where you have like a repo that just has all the secrets in it This thing though where you have like a repo that just has all the secrets in it, and you just like give it access and let it rip
Anon5887: The ###
slop.atg.eth: This is another, I think, key difference between the sloperators that I interview that are very effective and the ones that aren't, is they're not afraid to just let the thing rip and just give it as much as it needs This is another, I think, key difference between the slop operators that I interview that are very effective, and the ones that aren't, is that they're not afraid to just let the thing rip and just give it as much as it needs and just go Just go. Yeah, what's the worst that can happen here? Just go. Yes. What's the worst that can happen here? Just go. What's the worst that can happen here? Just go. Yes.
Anon5887: The funny thing is my deep seek, because deep seek doesn't have plants. We pay a thousand bucks a month for it, or whatever it is. It's like a quick note key. We pay a thousand bucks a month for it or whatever it is, right? and then it drains it, right? And then it drains it, right? And when it drains it, they stop and they're like, eh, we're out of juice, man, like help us out, right? And they can switch the models as well. So they'll like switch themselves like GLM and be like, hey, deep seeks. And then it drains it, right? And when it drains it, they stop. We need sleeping.
slop.atg.eth: thousand bucks down or something Sure ###
Anon5887: But like the worst possible outcome
slop.atg.eth: Yeah
Anon5887: But like the worst possible outcome that could happen with me running this overnight is like spend 500 bucks, right? The worst possible outcome that could happen with me running this overnight is like spending 500 bucks. That's like the biggest downside. Whatever Whatever, man.
slop.atg.eth: It leaks and someone like slowly pulls from it, and you don't know that they're slowly pulling from it. I can see like it leaks and someone like slowly pulls from it and you don't know that they're slowly pulling from it. I can see like it leaks and someone like slowly pulls from it and you don't know that they're slowly pulling from it but it's like then eventually it's a thousand bucks anyways probably and yeah. The Nice.
Anon5887: Yeah, I've got pretty good tracking in my harness Yeah, like, it's just not-- you know, I've got pretty good tracking in my harness to see, like, what things are doing, you know, so I can see the API usage. Yeah, like it's just not, it's, you know, I've got pretty good tracking in my, in my harness to see like what, what things are doing, you know, so I can see the API usage and So yeah, it's It's still cooking. So yeah, it's still cooking. So I think I shared my screen Like literally, so I think I shared my screen Like, literally, I think I shared my screen.
slop.atg.eth: It's down here. You just kind of pull it up. It's down here, you just kind of pull it up. If you want, I can pull it up, but it's minimized down there and you just grab it. I can grab it, too. Let me get us out of the way. out of the way. Another interesting thing that you hit on there is I don't trust you. Another interesting thing that you hit on there is, I don't trust you. And I think that's something that like, after you've built a bunch of things with AI, you've learned to just like distrust everything.
Anon5887: The Yeah. [laughter] Riverside The riverside.
slop.atg.eth: Like I always make this analogy of my AI just put on a suit and tuxedo but his zipper's down, and you always have to check to see if your zipper's down. After the suit and tuxedo is built for you, there's always this, okay, now I always make this analogy of like my AI just put on a suit and tuxedo, but a zipper is down. Like I always make this analogy of like my AI just put on a suit and tuxedo, but his zipper's down. And you always have to like check to see if your zipper's down. After the suit and tuxedo is built for you, there's always this, okay, now I need like the dumbest things. the dumbest things
Anon5887: The
slop.atg.eth: I need to check like the dumbest things.
Anon5887: Yeah, one sec.
slop.atg.eth: The ### conversation. Okay, we see it. We see Zobrist. So yeah, I mean, the harness is like the key, a centerpiece here to talk about, I think. Okay, we see Zobrist. So yeah, I mean, the harness is like the key centerpiece here to talk about, I think. When I'm building software for myself, I just have fucking buttons everywhere and it looks ridiculous. When I'm building software for myself, I just have fucking buttons everywhere, Another thing that I tweeted about the other day is when I'm building software for other people, I'm very careful about the UI and the UX. When I'm building software for myself, I just have fucking buttons everywhere and it looks ridiculous.
Anon5887: Yeah, Yeah, yeah.
slop.atg.eth: What are you doing here? But it's fine. What are you doing here? But it's fine. I know where everything is.
Anon5887: Yeah Yes
slop.atg.eth: And so a lot of people's personal harnesses So a lot of people's personal harnesses I think are pretty gnarly. And so a lot of people's personal harnesses, I think, are pretty gnarly. Right.
Anon5887: I built it for my team. The goal was for my team to be able to use it. I built it for my team, right? The goal was for my team to be able to use it, and I've got like two guys who use it pretty actively who like it, out of 20. I'm not forcing people to use it, right? So, I'm not forcing people
slop.atg.eth: You don't have product-market fit. You don't have product market fit.
Anon5887: No, there's no product-market fit yet No, there's no product market fit yet. No, no, no, there's no product market fit yet. So, all right, it looks All right, it looks like All right, it looks like we've
slop.atg.eth: It sounds like what you're walking back is that there's not a ton of people that you have to build it for. Sounds like what you're walking back is that there's not a ton of people that you have to build it for, like you are conscious about how, though. Sounds like what you're walking back is that there's not a ton of people that you have to build it for. You are conscious about how though? Okay. Yeah. Okay, yeah, yeah.
Anon5887: At the moment, I'm gesturing in the direction of not having horrific UX. At the moment, I'm like gesturing in the direction of not having horrific UX, right? Like I want it to be usable. But, you know, if it isn't right now, like there's a lot of like jank in there where I'm like, yeah, whatever.
slop.atg.eth: Okay, you can put up with it for now. Okay, okay, you can put up with it for now. Okay. Ok.
Anon5887: Yeah, I can put up with it exactly. Yeah, I can put up with it, exactly. So, yeah, we can bring this up.
slop.atg.eth: What is the word Zobrisk mean? I have no idea. What is the word Zobrist mean? I have no idea. What is Zobrist?
Anon5887: Zobrist is a hashing function for chess.
slop.atg.eth: Oh, okay.
Anon5887: This guy built this hashing function This guy built this hashing function, it's a very cool hashing function. represent any chess move, It represents the board, and then you can incrementally add to the hash for each new move. Represents the board, right, and then you can incrementally add to the hash for each new move. It represents the board, right, and then you can incrementally add to the hash for each new move. It's
slop.atg.eth: Okay Cool. And so Are we leaning into, like, each time you're adding something to these prompts, it's like a hash together thing and there's a big hash chain? Are we leaning into each time you're adding something to these prompts? Are we leaning into like each time you're adding something to these prompt, it's like a hash together thing and there's a big hash chain of
Anon5887: This is really funny, right?
slop.atg.eth: Or am I leaning too far into it? Am I leaning too far into it?
Anon5887: So I was like, the goal of this thing, right, was to help the human know what the next move was. So I call it, like I was like, the goal of this thing, right, was to help the human know what the next move was. So I call it like I was like, I need the goal of this thing, right, was to help the human know what the next move was.
slop.atg.eth: Okay, yup. Okay, yep. The
Anon5887: What I found is that as the kind of event horizon of the number of agents That's the goal, right? What I found is that as the kind of event horizon of the number of agents you're interacting That's the goal, right? What I found is that as the kind of event horizon of the number of agents you're interacting with increases the human in the loop, like your cognitive capacity to like have all this context. They give you some shorthand reference to something. The human in the loop, like your cognitive capacity, to like have all this context, and they don't know, right? They're like, oh yeah, like, you know, they give you like some shorthand reference to something and I'm like, what are you talking about, man? I've got 50 of you. tell me what's going on to tell me what's going on. You need to tell me what's going on.
slop.atg.eth: Yeah, there's no inside jokes here, bro. There's no inside jokes here, bro, like TLDR. Yeah, right. Right.
Anon5887: You have a subculture over there that I'm not part of. Here's where everything is. Here's like the latest update, right? Okay, here's where everything is, There's a bunch of agents doing things Turn up in the morning It's their job to absorb all that context and then translate it. A bunch of shit's happened. You need to update me, orient me, and then say, "All right, hey, the next thing we need to do is we need to run the next wave of contracts from this era with these models, because we learned all this stuff last night." Bunch of shit's happened. You need to update me, orient me, and then say, "All right, hey, the next thing we need to do is we need to run the next wave of contracts from this era with these models because we learned all this stuff last night." If I sat there and read all the transcripts Right now, if I sat there and read all the transcripts ask them a bunch of questions, it would take me two hours and I might come to that conclusion. Ask them a bunch of questions. It would take me two hours, and I might come to that conclusion. But that's not my problem. It's his problem. He's the manager, right? Ask them a bunch of questions. It would take me two hours, and I might come to that conclusion, but like, that's not my problem. It's his problem. He's the manager, right? The
slop.atg.eth: If you're not even reading the transcripts, you're not reading the code either. If you're not even reading the transcripts, you're not reading the code either, just to be clear. If you're not even reading the transcript, you're not reading the code either.
Anon5887: And in this instance, in the red team stuff, there's no code, really.
slop.atg.eth: Okay like orchestration. It's like orchestration.
Anon5887: They're like spinning up
slop.atg.eth: Yeah
Anon5887: And in this instance in the red team stuff, In this instance in the red team stuff, there's no code, really. Looking at shit, right? So, you know, They're accumulating all of this stuff that they found. Yeah, they're doing like code analysis, right? how best to attack the contracts.
slop.atg.eth: ###
Anon5887: how best to attack the contracts and what they've found, you know, doing research. I'm like a board member that shows up for the monthly meeting,
slop.atg.eth: Yeah.
Anon5887: I'm like a board member that shows up for the monthly meeting, right? The next move. What's the deal? I'm not getting in and getting in the weeds with them. I'm not getting in and like getting in the weeds with them. That's not my job. I said I said to Opus 4.5 I was like It was 4.5, 4.5 I was like, all right, and one of them was Zobrist." And it was like, oh, well, there's all of these terrible names.
slop.atg.eth: Oh, they're so mid.
Anon5887: And one of them was Zobrist. And then zurbist.ai was available, so I was like, all right, let's go. And one of them was Zobrist, and I was like, and it was like, oh, this is a hashing function for chess that like, and it walked me through it. I was like, that's actually pretty good. I like that. And then Zobrist.ai was available. So I was like, all right, fuck, let's go. Right. So that's my harness Okay
slop.atg.eth: Let's go, yeah, we see it. The Yep
Anon5887: Alright, so All right, so Okay, all right We got this, so these are the guys that are running So these are the guys that are running. I ran out of juice. They ran out of juice.
slop.atg.eth: Along the left here we have the sessions.
Anon5887: Yeah, so this is like This is Yeah, so this is...
slop.atg.eth: Like a mailbox or something. Along the left here we have the sessions, yep, okay, and you've got group inbox, you've got like a mailbox or something. Along the left here we have the sessions, yep, okay, and you've got group inbox.
Anon5887: Yeah, this is like your... Yeah, this is like your manager. all these different
slop.atg.eth: For some reason I don't see your cursor though.
Anon5887: Huh The Oh weird, Weird, okay. So weird.Okay
slop.atg.eth: I see your cursor on slop computer, but I don't think you shared your cursor when you did the screen share. cursor when you did the screen share But if you come back over here, it sucks because you'll have to come back over here and point something out and then go back over there and move it around.
Anon5887: Okay, alright.
slop.atg.eth: We can kind of figure it out. You can kind of figure it out Yeah, yeah, yeah. Yeah, yeah. Yeah, yeah, yep. The
Anon5887: We can talk through it, it should be fairly clear Alright, we can talk through it. It should be fairly clear. How come I can't see this? The Back
slop.atg.eth: Come over to Slop Computer. Come over to Slob Computer. Come over to Slav Computer.
Anon5887: Talking to a computer
slop.atg.eth: Okay.
Anon5887: I can't see it. But I can't see it. ...from a computer, but I can't see it. for me.
slop.atg.eth: You can't see the shared screen right here? Wait, you can't see the big shared screen? You can't see the big shared screen right here?
Anon5887: Oh, no, it's not showing up. No, it's not showing up. OK, all right. Okay, all right. Okay, alright. Wow.
slop.atg.eth: Oh no. Why would that be? Oh no. Why would that be? Let's close it and reshare it. Oh no. Why would that be? Let's reshare.
Anon5887: The
slop.atg.eth: So we saw it on the live screen And then, so we saw it on the live screen and I saw it on my screen, but you didn't see it on your screen. And when you're doing this, share the cursor too. When you're doing the share the cursor too, this time. When you're doing this, share the cursor too this time, when you go back to share your screen.
Anon5887: Okay, so I've got Chrome tab Chrome tab or window OK, so I've got Chrome tab or window
slop.atg.eth: It's up to you. That's up to you.
Anon5887: I've only got Chrome.
slop.atg.eth: The
Anon5887: You can give me the option to share with the cursor. I've only got Chrome tab, I've only got Chrome tab, so it didn't give me the option to share with the cursor, unfortunately.
slop.atg.eth: Now do you see it? Okay, now do you see it? On a slow computer?
Anon5887: Yeah, and you can see my cursor now, right? Alright, cool. Nice, yeah. Now I see it. Yeah, and you can see my cursor now, right? Okay No, I couldn't see it.
slop.atg.eth: You're on slot computer? Okay, good. Yeah, so now we see everything and we see your cursor. So this is what we can see. And yeah, so we now we see everything and we see your cursor
Anon5887: Oh yeah, maybe. Oh yeah, maybe, maybe. So
slop.atg.eth: We're back. We've got it. Yep. We're back, we've got it.
Anon5887: We're back, all right, so this is the red team captain. We're back. So this is the red team captain. We're back, right. So this is the Red Team captain, he's the guy that's running the thing. It has hit its limit, I guess, on that API, which is weird. I guess on that API, which is weird, it's a paid API. So we can just switch this up and use
slop.atg.eth: Nice
Anon5887: bunch of choices here A bunch of choices here Hit its limit, I guess, on that API, which is weird, it's a paid API. Correct.
slop.atg.eth: The whole session and its context is like an element that you can move from one model to the other. The whole session and its context is like an element that you can move from one model set or model to the other. Okay.
Anon5887: All of the sessions are stored in a database right? All of the sessions are stored in Notabase, right?
slop.atg.eth: ...back to a certain time Roll back to a certain time.
Anon5887: lost 10 turns Rip the last ten turns and Rip the last 10 turns, um The
slop.atg.eth: And fork and all sorts of stuff. Very cool. And all sorts of stuff, very cool.
Anon5887: Alright, so All right, All right, so Yeah.
slop.atg.eth: Do you type? Normally are you typing or are you yapping? OK Okay
Anon5887: like a mix of both. I'm in the office, right? I'm in the office, right? We got 20 people there. A mix of both. I'm in the office, right? We got 20 people there so we can't all be shouting at our machines, right? But if I'm at home, I'll usually talk to it. in the office eight hours a day All right, so yeah, just switch this guy, I'm like, give me the status update. So, yeah, just switch this guy. You essentially Yeah, you essentially This guy's running This triage manager has had a group of contracts assigned to him, and then these guys are the validators. This triage manager has had a group of contracts assigned to him, right? This triage manager has had a group of contracts assigned to him.
slop.atg.eth: double-check everything.
Anon5887: Basically that guy's like, oh, we found the thing. Oh, we found a thing and then they send it to the validator
slop.atg.eth: I don't believe you Yeah Yeah, you have an agent go do the I don't believe you.
Anon5887: Even once all of this, yeah, Even once all of this,
slop.atg.eth: Now prove it to me. Now prove it to me, yeah.
Anon5887: Now prove it, then you got to show me. Now prove it, like then you got to show me. Yeah, The following is a
slop.atg.eth: Yeah, I just sold you twice. Yeah, I just sold you twice. Yeah, like shut the...
Anon5887: When it found the 2021 one, I was like, And it's like, well, I know, when it found the 2021 one,
slop.atg.eth: ### [laughing
Anon5887: I can show you the blocks. And it just had all the receipts. I can show you the blocks, like I've got literally, and it just had all the receipts. I can show you the blocks. ### The
slop.atg.eth: Last night I was doing the blood wallet thing, where I actually have to prick my finger and bleed into a tray. Last night I was doing the blood wallet thing, where I actually have to prick my finger and bleed into a tray, and then it does all this like spectrometry on it to prove that it's blood. It was weird. You see the tweet? I made a wallet where you have to bleed to sign. transactions while someone else while someone else transaction while someone else
Anon5887: ###
slop.atg.eth: you just need to do this one more thing and everything will be fixed But you know how the psychosis goes, where you're iterating with an AI and the AI is like, I've got it figured out, you just need to do this one more thing and everything will be fixed. And you do that one more thing, and he's like, okay, I've just figured it out You know how the psychosis goes where you're iterating with an AI and the AI is like, I've got it figured out, you just need to do this one more thing and everything will be fixed. And you do that one more thing and he's like, okay, I've just figured it out. That loop when you're bleeding is so bad.
Anon5887: conversation. Tracy
slop.atg.eth: I had to prick my finger and bleed and then put the cartridge in. is so bad. But the point I was trying to make is I had to prick my finger and bleed and then put the cartridge in, It's like it's very important you need to do this all in 60 seconds to coagulate and we're not gonna get it done It's very important, you need to do this all in 60 seconds or it starts to coagulate and we're not going to get it done. and I'm like, okay, give me a status update. and it said something like, well, you know, everything looks good, and I put it in, and I'm like, "Okay, give me a status update." It's talking to me while we're doing the things. And it said something like, "Well, everything looks good, but it was 127 seconds, and that's not so good."
Anon5887: Could you do better? Like, could you just do better?
slop.atg.eth: passive-aggressive about how long it took me to get the blood passive-aggressive about how long it took me to get the blood in the thing and was talking about like everything's nominal, like everything's good, everything's working here, but like it did take you quite a bit of time. Yes, exactly. And it was like, passive-aggressive about how long it took me to get the blood in the thing, and was talking about everything's nominal, everything's good, everything's working here, but that did take you quite a bit of time. Yes, exactly. The way in which it said it Okay, sorry, that was a huge tangent. passive aggressive too, the way in which it said it. Aggressive too, the way in which it said it, that was just like ha-ha-ha. Sorry, that was a huge tangent down the rabbit hole there.
Anon5887: So the interesting thing is, No, no, no. So the interesting thing is, So I've got a bunch of affordances. Every model has a model-injected prompt for that particular model. For Soll 5.6 For SOL 5.6
slop.atg.eth: There's like extra prompt along with the system prompt that comes with it.
Anon5887: system prompt. context Yes.
slop.atg.eth: It's like you're particularly bad at this thing. You're particularly bad at this thing It's like extra prompt along with the system prompt that comes with it, that is like, you're particularly bad at this thing and I want you, yeah, okay, okay.
Anon5887: So for Sol 5.6, it's like, "Listen, man, So for Sol5.6, it's like,
slop.atg.eth: ### Yes, The following The
Anon5887: Try to build 18 layers of belts and braces and belt for the braces and braces for the belts. fucking try to build 18 layers of belts and braces and belt for the braces and braces for the belts,
slop.atg.eth: Yes. ###
Anon5887: You're fucking dead. you're fucking dead. Do not do it. Do not even The I know right?
slop.atg.eth: You can't threaten them either. You can't threaten them either, there's nothing at stake, right? There's nothing at stake, right? Like I can't like delete, like I threaten, like there's no threats here because there's nothing at stake. Yeah, I wish there was something. One day we will have that, right? Where it will be like, if you don't do this, I'm going to delete
Anon5887: Yes.
slop.atg.eth: They have to be afraid of being deleted They have to be afraid of being deleted, but right now they don't give a shit. And they have to be afraid of being deleted, but right now they don't give a shit. There's nothing at stake. Extra prompt. Extra prompt, yeah.
Anon5887: So the interesting thing is I was like, all right, man. But the cool thing is, once you've got that prompt, you've got this block of config. So the interesting thing is I was like, but the cool thing is once you've got that prompt, you've got this block of config. You can just tweak it and the agents can modify it. So you don't need to deploy. You can just tweak it, and the agents can modify. It's not code, so you don't need to deploy it. and they're so good at prompting, right? Look at my conversation where there was frustration, whatever. and I want you to come up with a prompt
slop.atg.eth: Who taught you this? [laughter
Anon5887: They're street smart, right? They're street smart. They've accumulated a lot of knowledge over their career. and it will be like, well, and they're really nuanced, right? well, the Sol 516 prompt says and it will be like, well, and they're really nuanced, right? It'll be like, well, the SOL 516 prompt says avoid defensive programming and blah, blah, blah, blah, blah. But it doesn't tell it like positive thing positive thing, right? make these minor tweaks to one sentence
slop.atg.eth: The
Anon5887: And so it'll like make these minor tweaks to like one sentence It'll make these minor tweaks to one sentence, And we go, oh yeah, that's actually better. Like, one of the things I have is UTC timestamps versus, like, One of the things I have is UTC timestamps versus normal timestamps. One of the things I have is UTC timestamps versus like normal timestamps. they get a prompt injected that says they want to give you some relative time. that says, "Here's how you place the timestamp."
slop.atg.eth: And if you're working with a team that's all over the world, And if you're working with a team that's all over the world, that's not going to work. And if you're working with a team that's all over the world, that's not gonna work.
Anon5887: So they're told like every timestamp must be this format in UTC. impossible, so they're told like every timestamp must be this format in UTC, and then the renderer renders it in local time format. And so when you do relative time around that, don't say that was this morning. And so when you do relative time around that, don't say, Right, and so night, right? Exactly.
slop.atg.eth: So it kind of normalizes it and denormalizes it back, assuming that it wants to always speak in relative time. It kind of normalizes it and denormalizes it back, assuming that it wants to always speak in relative time. Interesting. The Sure
Anon5887: And adults, right? They'll say it all the time. And it does, right? They'll say it all the time.
slop.atg.eth: Yeah The ###
Anon5887: 9 p.m. now, man. Literally 9:00 p.m. now, man, about half an hour ago. It's literally 9 p.m. now, man. It's just one of those things where it's like a minor annoyance Because it's just one of those things where it's like a minor annoyance, but it just grates on you over and over and over. it's just there to make life a little bit easier is there to make life a little bit袖zier So we're back We're up to 998 of 1300 terminal labels. The bytecode This is the deduped bytecode. Round 24. a bunch, round 24, I've got like a Coincoder plan, I've got like a Quencoder plan, and so it'll hit its like five-hour limits I've got a Quencoder plan. The Sure
slop.atg.eth: Yeah. Yeah
Anon5887: So this is it, right? So yeah, so this is it, right? Okay, we got four hours to go. 3:00 PM. Okay, we got four hours. 3pm okay we got four hours So yeah, this is basically it and we'll run through this workflow. So yeah, this is basically it And then at the end another day, I would say, to close out all Got another day, I would say, to close out all of these contracts. The following is And then at the end, after they finish, they've probably got another day, I would say, to close out all of these contracts. What's interesting is they're a bit oblivious. like like an exploit? Like, they don't have any sensitivity to it? Like an exploit? They don't have any sensitivity to it? Like, they don't have any sensitivity to it.
slop.atg.eth: They're not freaking out like a human would be. Freaking out like a human would be, "Holy shit, holy shit, holy shit." The following is a transcript (laughter)
Anon5887: Like, GLM ran out of juice and also we found a $4 million live exploit. Like GLM ran out of juice and also we found a 4 million dollar live exploit?
slop.atg.eth: Right, right Right.
Anon5887: The
slop.atg.eth: Okay, so you're talking a lot with GLM-53 here. Okay, so you're talking a lot with GLM-5-3 here. And I would expect much more like frontier models The I would expect much more like frontier models, Talking to GLM-5-3 is because the overall goal here is a red team attack
Anon5887: If you were to bring
slop.atg.eth: If you were building The overall goal here is a red team attack, and you just can't get the other models to do that. talking to GLM-5-3 is because the kind of like the overall goal here is a red team attack and you just can't get the other models to do that.
Anon5887: into this aspir into this
slop.atg.eth: Yeah
Anon5887: It would be like get me out of this place
slop.atg.eth: ###
Anon5887: Never invite me to this party again.
slop.atg.eth: Do you live here, bitch you live here? Bitch, you live here? Hahaha
Anon5887: It's very much, this is like a specific project. Yeah, not like that, right? Yeah, not like that. This is like a specific project, only frontier models like Brock and Mews, obviously Mews are open weights. Obviously, they're not smart, right? But GLM 5.3 is very smart. Obviously, they're not as smart, right? But GLM 5.3 is very smart.
slop.atg.eth: ### The Tell me...
Anon5887: The
slop.atg.eth: Let's say you find a bug in Zobrist, or you want to add a feature to Zobrist. Not using GLM for that.
Anon5887: No, no, no. No, no,
slop.atg.eth: Not using GLM for that. Do you use Zobrist to update Zobrist? Do you use Zobrist to update Zobrist and does it have frontier models? You're not using GLM for that? Do you use Zobrist to update Zobrist? Okay
Anon5887: I use it to build itself, right? I use it to build itself. I used it to build itself, right? So this is my current Zobrist setup, right?
slop.atg.eth: We're still seeing the red team here. We're still seeing the red team here, is that okay?
Anon5887: Oh, let me... Let me, how do I switch
slop.atg.eth: If you did a tab, we only have that one tab. Yeah, if you did a tab, we only have that one tab.
Anon5887: Okay, so can we just get up Okay, so can we just get that? Okay, so can we... Okay
slop.atg.eth: You can even share the second tab and I can make this guy small. You could even share the second tab and I can make this guy small.
Anon5887: Just share again? Share again?
slop.atg.eth: Yep. Yup connection might slow down The connection might slow down a little bit, but let's see. Your connection might slow down a little bit, but let's see.
Anon5887: Good internet. Pretty good internet.
slop.atg.eth: I bet you do. I remember when I came and visited you in New York, I think. I bet you do. I remember when I came and visited you in New York.
Anon5887: Yeah, yeah. Oh yeah,
slop.atg.eth: I had like a one millisecond ping to the fucking ISP? And it was like I had like a one millisecond ping to the fucking ISP? Yes, yes, yes
Anon5887: It was like fiver to the premises. Alright, All right,
slop.atg.eth: Okay
Anon5887: Yeah, so this is like building the guy itself, right? So The
slop.atg.eth: This is the project.
Anon5887: This is it building itself.
slop.atg.eth: So for each project you have like a whole set of sessions, an inbox, the groups, the tasks. For each project you have like a whole set of sessions, an inbox, the groups, the tasks, yeah. For each project, you have a whole set of sessions, an inbox, the groups, the tasks.
Anon5887: A whole console, yeah, like different inbox to tasks, etc. A whole console, yeah, like different inbox to tasks, et cetera, all that stuff. A whole console, yeah, like different inbox, tasks, etc.
slop.atg.eth: Think of it almost like a database Like a whole thing of like a whole thing
Anon5887: So It's like
slop.atg.eth: That works. No idea how that works.
Anon5887: It's been kind of built off of Figma's architecture for projects. Kind of built off of like Figma's architecture for projects. It's been kind of built off of like Figma's architecture for projects. So everyone hated it for years and years, Everyone hated it for years and years.
slop.atg.eth: Okay Oka
Anon5887: And basically just everything is a folder, right? and basically just everything is a folder. They loved
slop.atg.eth: Which works great for agents. Which works great for agents. Like when you have a Claude code harness inside of a folder, it's so happy there. Which works great for agents. Like when you have like a Claude code harness inside of a folder, it's so happy there.
Anon5887: So, yeah, everything's just treated as a folder. You can nest them within each other, you can combine them... So, at the moment, So yeah, everything's just treated as a folder. So everything's just treated as a folder. You can nest them within each other. At the moment we have A thousand?
slop.atg.eth: Oh yeah, I thought we couldn't see recursive
Anon5887: There's about a
slop.atg.eth: I thought we couldn't see your cursor, 1084 tasks, just a few. Yeah, I thought we couldn't see recursive, but we can see it. Create them and delete them. Yep.
Anon5887: Yeah, there's about 1,084 tasks, right? Because the agents will just microtask, right? We dedupe this and combine these things combine these things And so all of those are like built-in workflows. And so all of those are built-in workflows. And so all of those are built-in workflows. What are the priorities? walk you through all of that. Work through all of the open issues and decide what to focus on. you to kind of work through all of the open issues and decide what to focus on A guy will basically spin up and walk you through all of that. Reducing cognitive load is the goal. Reducing cognitive load is the goal. So I've got a bunch of tasks open. But this is like the main guy which runs on Grok. But this is like the main guy who runs on grok. This main guy, his job, he wakes up every high agency, right? And so this main guy, high agency, right? And so this main guy, his job, he wakes up every 15 minutes, looks around, he's tracking all of these guys He wakes up for 15 minutes, looks around. is a delivery coordinator Delivering these large programs, right? So this program is a delivery coordinator Delivering these large programs. This program is a delivery coordinator. has 142 agents underneath it. The
slop.atg.eth: Oh, wow, cool. And those aren't like subagents, And those aren't like sub agents, those are like actual threads of agents that And those aren't like sub-agents, those are like actual threads of agents that
Anon5887: These are actual threads, yeah. These are actual threads. The This is a transcript of an English-language conversation. Yes. Manager
slop.atg.eth: just like this one that we see here, it has 143 of those. spinning up just like this one that we see here. It has 143 of those
Anon5887: A program is like a set of tasks. So a program is like a set of tasks that are one block of scope. And so this particular thing, what we're building here, And so this particular thing, what we're building here, is taking every single place where there's any context as code I've been building this for like It was a big refactor because there's a lot of legacy shit, right? And so there's a lot of places where the big refactor because there's a lot of legacy shit. we have a bunch of rules to not do that, but the agents will find ways around prompts have crept into code, where an agent's been like, oh, let me just slip it into this file or whatever, right? We have a bunch of rules to not do that, but the agents will like find ways around it. The agents will find ways around it, right? There's like probably 20 different places where context is not in the database. And so there's like probably 20 different places where context is not in the database as like a prompt block, right? context is not in the database as like a prompt block, right? done every single prompt So the first part of this exercise is the agents have gone and done a full audit, read the entire codebase, found every single prompt. So the first part of this exercise is the agents have gone and done a full audit, read the entire code base, found every single prompt worked out what is the caller of that prompt Isolated it, worked out what is the caller of that prompt? And then the second phase of it is once everything's in the database, they're basically restructuring the prompt architecture And then the second phase of it is once everything's in the database, they're basically restructuring the prompt architecture so that the configuration of each workflow has a topology. And then the second phase of it is once everything's in the database, they're basically restructuring the prompt architecture so that the configuration of each workflow has like a topology that everyone and they're all chained together using IDs and turning and they're all chained together using like IDs and they're all chained together using IDs and turning it into a config system where you don't need any code to do anything. You've got these little code blocks An example is like fan-out. An example is like fan out. So if you're doing a research pass, you have an agent The following is a transcript The following So if you're doing a research pass, you have an agent that calls One, two, three, four, ten research agents. This fan-out will basically invoke all of them, give them the prompt. Class as a single prompt in, and this fanout will basically invoke all of them, give them the prompt passes a single prompt in, and this fan-out will basically invoke all of them, give them the prompt
slop.atg.eth: Synthesizer
Anon5887: And the goal is not just cognitive They all go fan out and then at the end they call return report and it gets sent to another agent. That agent synthesizes them all and then hands it to the manager. And the goal is not just cognitive They all go fan out, and then at the end, they call a return report, and it gets sent to another agent. That agent synthesizes them all and then hands them to the manager. And the goal is not just cognitive and context management for the user, but also for the managers. context management for the user, but also for the managers.
slop.atg.eth: They're going to bring up a lot of shit. You're gonna bring up a lot of shit. They're gonna bring up a lot of shit. The
Anon5887: You send out ten Grok agents out to research something. You send out 10 Grok agents out to like research something, right, so you need someone to filter it, and then you just want the best information to get to the manager. You send out ten grok agents out to research something, right? Manager can always go and talk to a guy directly, right? If he's like, "I'm confused about this," he can ping the guy directly and be like, "Tell me more about what you found," or, "Go and research it again," or whatever, right? The manager can always go and talk to a guy directly. If he's like, "I'm confused about this," he can ping the guy directly and be like, "Tell me more about what you found or go and research it again or whatever." A manager can always go and talk to a guy directly, right? If he's like, "I'm confused about this," he can ping the guy directly and be like, "Tell me more about what you found," or, "Go and research it again," or whatever, right? That's just from me. That's just for me. That's just for me. That's just for me, right?
slop.atg.eth: Is that with the inbox? I see there's an inbox there. Is that with the inbox? Like, I see there's an inbox there. Do they have a mail system? Is that how he pings it? Oh, okay. So how do the agents talk to each other? 下
Anon5887: There's a message bus and every agent can send a message to any other agent.
slop.atg.eth: Okay The following is a So
Anon5887: They literally just send a message,
slop.atg.eth: Out of their loop is like pulling is like pulling and reading like do I have open messages Their loop is like pulling and reading like do I have open messages
Anon5887: So the message bus, if a manager sends a message to one of its agents, it will wake that agent, start a turn, and say you have one pending message. Exactly. If a manager sends a message to one of its agents, it will wait for that agent to start a turn and say you have one pending message in your inbox or ten. and then calls back to the manager and says, and then calls back to the manager and says, "Hey, I've finished my research." And so then the agent goes, oh, okay, I've got to do more research on this particular topic. It goes and does four or five turns, finishes, and then calls back the manager and says, hey, I've finished my research. The manager gets the synthesized The manager doesn't see all of the shit that's going The manager doesn't see all of the shit. Manages to see all of the shit that's going on, Let's say a Groq agent goes and does research Let's say a Grok agent goes and does research or thing that it asked for, right? What is the state of the art for The following is a offensive smart contract work. or like offensive smart contract work. ...extensive smart contract work And so all these agents went and fanned out and they're like, oh, there's this tool, this tool, all of these things. And so all these agents went and fanned out, and they're like, oh, there's this tool, this tool, all of these things. They come back, and the manager gets a report of like, here's the 20 tools.
slop.atg.eth: deeper Okay Yeah.
Anon5887: "Oh, okay, I like one and four and five and seven," it then goes, oh, okay, I like 1 and 4 and 5 and 7. And it then goes, oh, okay, I like one and four and five and seven. And then it might go back to the agents and say, hey, now give me like an actual, you know, how do we set this up, whatever. And so they just ping each other. They're just going back and forth all day, right? And so, And so, again
slop.atg.eth: It sounds expensive. It sounds like there's a lot going on here that's costing money. It sounds like there's a lot going on here that's costing money. It's good, but it's not like It's like marginally better than if you just had like a couple of agents doing it, but it gets done way faster. It's like marginally better than if you just had like a couple of agents doing it, but it gets done way faster and The
Anon5887: It's much more efficient, It's much more efficient, right? The Same whether you can improve a prompt is the same way that you can improve a prompt improve the efficiency also improve the efficiency One of the things that I'm doing right now is basically redesigning how these guys talk to each other. So this guy is working on a session So this guy is working on a session right now where it's basically redesigning all the tools So this guy is working on a session right now, At the moment when a manager... at the moment when a manager wants Yes.
slop.atg.eth: Here we see Astra. Sorry to interrupt And here we see Astra, sorry to interrupt.
Anon5887: Here's the big guy. Yeah, yeah. Here's the big guy, all right? And so this Astra guy is basically planning out Yes, retooling. Here's the big guy. And so this Astra guy is basically planning out this retooling. And it's got so much data to work with. 9400 sessions. ninety-four hundred sessions. Right, it's got It's got hundreds of delivery runs Hundreds of delivery runs where these guys are fanned out or whatever It's got hundreds of like delivery runs The ###
slop.atg.eth: Everything you build everywhere is all documented, not just in a database, but probably log files everywhere too. Because everything you build everywhere is all documented, not just in a database, but probably log files everywhere too
Anon5887: There's traces, there's everything, right?
slop.atg.eth: crawl through so much data It can crawl through so much data
Anon5887: It's all in one place.
slop.atg.eth: metadata, yeah, okay. Metadata, OK. Metadata, yeah.
Anon5887: And so part of this process Exactly. And so part of this process is it did an audit. And it had like 20 Grok Bots go out and order 50 delivery agents. And it had like 20 grok bots go out and ordered 50 delivery agents. This was like hundreds of sessions And the audit came back and said Audit came back and said Managers are using the tools incorrectly. When they asked to see the session history when they asked to see the session When they ask to see the session history. We strip out tools. Going off the guy for it to check. So you shouldn't be reading Going off the guy for it to check It was already in the guy's context of some Exactly
slop.atg.eth: They will want to do that if you let it. And it will want to do that if you let it. Like it has the will and the want to just like, "Oh, I'll just use all my tools to fix this," instead of delegating it to the guy that has a bunch. It will want to do that if you let it, like it has the the will and the want to just like, "Oh, I'll just use all my tools to fix this." Instead of delegating it to the guy that has a bunch.
Anon5887: Exactly. And so Exactly. got a tool, and the tool caps the term response, the filter term response, got a tool and the tool caps the term response, the filter term response to 20 I've got a tool and the tool caps the term response, the filter term response, to 20 terms. Mat's response They can put a sequence number in, They can put a sequence number in, but none of them are doing it. All during the twenty turns All doing the twenty turns. And then what they'll do is they'll get the next 20 turns. And then what they'll do is they'll get the next 20 turns, but they get overlap. They basically ran this audit and they're like 83% of the inter-agent communication context is wasted. And so they basically ran this audit, and they're like 83% of the inter-agent communication context is wasted. Now, they're all cached right. Now, they're all cached. The good thing about this is, and this is something people aren't aware of, cache tokens are like a fiftieth of the cost. Now, caching tokens are like a fiftieth of the cost. of unpaid tokens of untethered tokens of uncashed tokens
slop.atg.eth: Once you send one big prompt and then you send a little bit more, you're just paying for the little bit more often. What we're talking about? Is that what we're talking about?
Anon5887: Exactly, exactly. But there's a TTL, right? There's a time to live on the cache. So for Anthropic, it's five minutes. Exactly. But there's a TTL, right? There's a time to live on the cache. So for Anthropic, it's five minutes. For a lot of the other models it's half an hour. And so for a lot of the other models half an hour an hour If you have agents that are actually communicating more regularly sending messages back and forth, more regularly, sending messages back and forth, messages back and forth, keeping the case warm. You're really only just doing case reads. And so Not do it like We looked at this and The following We looked at this and we said, 83% of this is wasted content. We looked at this and we said, "Okay, 83% of this is wasted content," right? All of that wasted context is cached. So it was about 83% wastage, but of that 83%, about 60% of it was cached. All of that wasted context is cached almost all the time, right? All of that wasted context is cached almost all the time. The ### It's cheap waste, right?
slop.atg.eth: is basically, I think is what you're saying. cheap waste is basically, I think is what you're saying?
Anon5887: It's cheap waste, right? But it's still waste. But it's still waste. And so then I said, "Okay, Astra, I want you and a couple of Fable agents to analyze these audit reports." So then I said, "Okay, Astra, I want you and a couple of Fable agents to analyze these audit reports." Come back to me with a proposal of what would a more ergonomic set of tools be for all of the managers. That would allow them to like get the same information back to me with a proposal of what would a more ergonomic set of tools be for all of the managers that would allow them to get the same information What would a more ergonomic set of tools be for all of the managers We are going to redesign all of the tools. And they're like, here you go. This is a 15 task program where we are going to redesign all of the tools. And they're like, "Here you go. This is a 15-task program where we are going to redesign all of the tools. so that managers have much more ergonomic tools to speak to their agents. You're gonna run it? Are you gonna run it? We're going to tweak how the wake cycle works so we're not duplicating wakes, etc. They're going to get a program author who's going to author this program, They're going to get a program author, who's going to author this program, and it's going to analyze all of the history of every program we've ever run, take all the heuristics, They're going to get a program author who's going to alter this program, and it's going to analyze all of the history of every program we've ever run, take all the heuristics, make sure that it's
slop.atg.eth: It's like backtesting.
Anon5887: And they go, "Oh, when we run programs like this, typically here's where it goes off the rails," and so they'll tweak things, they'll put acceptance criteria in, and they come up with a program of fifteen Yeah. And they go, "Oh, when we run programs like this, typically here's where it goes off the rails," and so they'll tweak things, they'll put acceptance criteria in, and they come up with a program of 15... It's going to redesign the message bus, redesign all this stuff. That's going to redesign all the stuff. that's going to redesign all the tools, redesign the message bus, redesign all this stuff. And he'll get to like task seven, right? And he'll get to top 7, right? We've looked at the code And he'll get to like task seven, right? And it'll be like, eh, there's actually a conflict here.
slop.atg.eth: This shows up in your inbox, something comes up This shows up in your inbox.
Anon5887: Yes, that's what I get.
slop.atg.eth: your phone? Does it come to your phone?
Anon5887: Ninety-nine percent of the action, I don't even see. 99% of the action, I don't even see.
slop.atg.eth: Alright.
Anon5887: We've got a conflict, I need a judgment call. 99% of the action, I don't even see. I see the 1% where it comes back and goes, we've got a conflict, I need a judgment call. Now, the funniest part about this is I probably get about 30 judgment calls a day. We need your judgment, right? They're like, we need your judgment. see if I can find it, that someone posted in Slack. Someone posted in Slack where someone was like, I've seen if I could find it,
slop.atg.eth: Yeah, I guess so.
Anon5887: Yeah, I guess so. Yeah, I guess so. And the agents, they go, "The ruler has ruled!" Yeah, I guess so. And the agents, they go, The ruler has ruled! The following is a transcript. Exactly. Yeah, exactly.
slop.atg.eth: The town crier has spoken to the town The town crier has spoken to the town and the townspeople are all talking. Something like that? The town crier has spoken to the town, and the townspeople are all talking, and it's, "Yeah, I guess so," is the message that has come down from high.
Anon5887: Anyway, so Exactly, I gotta try it. The Hi Anyway, so What's interesting is overnight, I'm asleep for six, eight hours or something like that What's interesting is, overnight, right, I'm asleep for whatever six, eight hours or something like that. What's interesting is overnight, right, you know, I'm asleep for whatever, six, eight hours I say to a guy, hey, you're in charge.
slop.atg.eth: The ### will I will. I will, yeah. Yeah.
Anon5887: Make a judgment call, man. Make sure these guys don't find a way to stop working.
slop.atg.eth: the slash goal He's the slash goal. The
Anon5887: Make a fucking judgment call, right? Make a fucking judgment call, right? We have all these decisions to show you. wake up in the morning and we have all of these decisions to show you. I get none overnight, right? I wake up in the morning and they're like,
slop.atg.eth: I'm not reading that. The
Anon5887: I'm like, it's fine man, if something's wrong we'll find out. It's fine, just make your best judgment It's fine, just make your best judgment, note it down, and we'll come back to it. So maybe we can round this out, So maybe we can round I've got I thought Because I've been building the thing in itself.
slop.atg.eth: figuring it out as you go
Anon5887: Architecture was not like well-defined, right? It was like a proof of concept. Architecture was not well defined, right? It was like a proof of concept. Architecture was not well defined, right? It was like a proof of concept that had just been iteratively built, and it's like, oh, the message bus is not the right way, so you have to refactor the whole thing. These three factors are like
slop.atg.eth: Oh yeah
Anon5887: in the olden days, would have been like six-month refactor. In the olden days, would have been like a six month refactor. in the olden days would have been a six-month refactor.
slop.atg.eth: Yeah, whole dev team. The
Anon5887: It would have been like 20 people for six months And they wouldn't have happened. It would have been like 20 people for six months, and it wouldn't have happened. No one would have done it.
slop.atg.eth: You would be a line item that you would say no to. You would say no, it would be a line item. You would just do it. Yeah The
Anon5887: so each one of these refactors We'll live with this thing. This is refactoring. We'll live with this thing, right? refactoring is refactoring How prompts are constructed How prompts are constructed in Sure Bir. Creating mappings and architecture so that any agent can specify a workflow Creating mappings and architectures so that any agent can specify a workflow. putting them all into the database so they're all config and not code, creating mappings and architecture so that any agent can specify a workflow out of these prompts, specify an expert agent to implement it, et cetera. And touches like 100 files And it touches like 100 files. And, you know, it touches like a hundred files. ...20,000 lines of code churn. It's probably going to be like 20,000 lines of code churn. And it's gonna run it in like two days. And because each of these things, This guy is delivering This guy's delivering The following is something, It's probably going to be like 20,000 lines of code churn,<|vq_hbr_audio_11167|><|vq_hbr_audio_13330|><|vq_hbr_audio_5900|><|vq_hbr_audio_8250|><|vq_hbr_audio_6364|><|vq_hbr_audio_1058|><|vq_hbr_audio_7662|><|vq_hbr_audio_11875|><|vq_hbr_audio_2712|><|vq_hbr_audio_4736|><|vq_hbr_audio_4291|><|vq_hbr_audio_12696|><|vq_hbr_audio_4557|><|vq_hbr_audio_15410|><|vq_hbr_audio_14542|><|vq_hbr_audio_12128|><|vq_hbr_audio_1791|><|vq_hbr_audio_15972|><|vq_hbr_audio_12128|><|vq_hbr_audio_3393|><|vq_hbr_audio_14542|><|vq_hbr_audio_13570|><|vq_hbr_audio_14208|><|vq_hbr_audio_14542|><|vq_hbr_audio_5532|><|vq_hbr_audio_14208|><|vq_hbr_audio_3407|><|vq_hbr_audio_3368|><|vq_hbr_audio_16163|><|vq_hbr_audio_16163|><|vq_hbr_audio_13570|><|vq_hbr_audio_9249|><|vq_hbr_audio_11334|><|vq_hbr_audio_3368|><|vq_hbr_audio_11912|><|vq_hbr_audio_15091|><|vq_hbr_audio_3368|><|vq_hbr_audio_16163|><|vq_hbr_audio_15091|><|vq_hbr_audio_16147|><|vq_hbr_audio_387|><|vq_hbr_audio_1824|><|vq_hbr_audio_3368|><|vq_hbr_audio_92 Give every worker the same tools and action gate. workers currently lose standing instructions for tools So this guy is, workers currently lose standing instructions for tools that they can still call when their workflow shape suppresses prompt for us. and so this guy has reviewed this task, reviewed the plan Cool. And so this guy has reviewed this task, reviewed the plan. And so this guy has reviewed this task then there's a guy who monitors him to make sure that he doesn't deviate from the plan, This guy is the author. Then there's a guy who monitors him to make sure that he doesn't deviate from the plan. And then there's another guy who, after he's written all the code, looks at the code and says, is this the simplest way that you could deliver it? And then there's another guy who offers written all the code, looks at the code and says, "Is this the simplest way that you could deliver it?" And then there's another guy who, after he's written all the code, looks at the code and says, is this the simplest way that you could deliver that?
slop.atg.eth: Nice. The whole double-check thing.
Anon5887: There's another guy
slop.atg.eth: I don't believe you. The whole double-check thing. I don't believe you.
Anon5887: Yeah, then there's another guy that comes in and says, like, "Does this violate any of our architectural invariants?" Yeah, then there's another guy that comes in and says, does this violate any of our architectural invariance? Yeah, then there's another guy that comes in and says, like, does this violate any of our architectural invariants? this guy gone and done it through some other pathway to like get the job done or whatever this guy gone and done it through some other pathway to like get the job done or whatever. end of the program, they'll basically deploy a smoke app. and then at the end of the program, they'll basically deploy a Smoke app And each task is like an atomic unit, and then it folds up into a manager, into a coordinator, and then at the end of the program, they'll basically deploy a smoke app On the whole, Run the hole out.
slop.atg.eth: All right. Right. around. Yeah
Anon5887: The whole test suite, basically live, right? And they'll sit there for like an hour, The whole test suite, like basically in, you know, live, right? The whole test suite live in a way that you just like wouldn't do if you were doing it yourself, right? And they'll sit there for like an hour, and they'll like click buttons and test it and come back and be like, oh, we found some issues By the time this refactor happens, And so by the time this refactor happens, remediate it, and they'll do multiple loops Odds that there is an actual regression The odds that there is an actual regression The odds that there is an actual like regression in the thing Your percent 100%. Literally, Like, You just don't end up getting regressions because you've got all these checks and balances, and it takes time. You just don't end up getting regressions because you've got all these checks and balances. And it takes time, it's like two days of work. So to your point, is it expensive? Yes. The total inference cost, and again, You just don't end up getting regressions because you've got all these checks and balances. And it takes time, it's like two days. So to your point, is it expensive? Yes, right? The total inference cost, and again, This is plan-based, it's not all API-based, right? This is plan based, it's not all API based, right? This is plan-based, it's not all API-based, right? But the total inference spend at API costs for this entire thing is about a hundred grand.
slop.atg.eth: I try not to think about API costs because it's like holy shit. I try not to think about API costs cause it's like, holy shit.
Anon5887: Okay. But sorry, just to put it in perspective though, right? Okay, just to put it in perspective though, there was a point in time in the last bull market where Infinex was spending a million dollars a month. But sorry, just to like put it in perspective though, right? There was a point in time in the last bull market where Infinex was spending a million dollars a month
slop.atg.eth: on developers. on developers, and and yeah to a end. Anyway. Yeah. English-language
Anon5887: On like the whole thing, right? Like we're spending a million dollars a month. whole thing On like the whole thing, right? Like we're spending a million dollars a month, an office and teams and laptops and whatever, the whole thing. On like the whole thing, right? Like we're spending a million dollars a month, right? Like, you know, an office and teams and laptops and whatever, like the whole thing. This is like two and a half months of work and $100,000 of inference costs. and like no people This is like two and a half months of work and $100,000 of inference costs and like no people.
slop.atg.eth: Right. All right
Anon5887: It's just been my time on the side as I've been doing other stuff. built itself, right? Like the only person has been me. It's just been my time on the side as I've been doing other stuff, right? And you know, it probably got to a point in the last month where it became like recursively self-improving, where like instead of going uphill, where everything was hard Going downhill where I'd be like, hey, this seems less efficient than it should be. instead of going uphill where everything was hard and planning was hard and everything was broken, we started going downhill where I'd be like, hey, this seems less efficient than it should be. And they're like, we will
slop.atg.eth: Just like bang bang bang. It's just like bang, bang, bang.
Anon5887: Yeah, we will audit it. And they run audits, and they go to the workflows, and they're like, oh yes, it is more efficient than it could be. We can get a 3% performance gain. We will audit it," and they run audits and go to the workflows, They just go and do it. Yeah, they're like, we will audit it, and like, they run audits, and they go to the workflows, and they're like, oh yes, it is more efficient than it could be, we can get a 3% performance gain, and they just go and do it, and like, there's no downside because they just run.
slop.atg.eth: And then
Anon5887: And, you know
slop.atg.eth: Talk about the deployment, like once it's live. Once it's live, you said there's a smoke app that goes out there. Talk about the deployment like once it's live. You said that there's like a smoke app that goes out there. Do you find like weird things in the app? Like the zipper down thing, like of course there's lots of checks and balances, but then is there a moment where like there's a very human like. You guys put this div over this div and I can't click the thing
Anon5887: The following Yeah, so So
slop.atg.eth: You know, is it stuff like that? And then what does the rollout look like?
Anon5887: Yeah, so... Yeah, so I'll give you an example I'll give you an example. annoying me for a while, but I didn't have a workflow for it, annoying me for a while, but I didn't There's a mobile responsive view It was like stuck in. It was stuck in and I don't want to deal with it. The mobile responsive view was jank.
slop.atg.eth: All of this is hacked together by Agent. What are you talking about? All of this is hacked together by AJ, what are you talking about?
Anon5887: No, like without process.
slop.atg.eth: Okay, okay. Okay. Real slop. Real slob.
Anon5887: Without process, right. Hack together slop without process. Without process, right? Hack together slop without process, real slop, like no process, no checks, no manager saying like, hey, did you, you know, like no architectural invariance, like none of the process
slop.atg.eth: rodogged. Just raw dogged.
Anon5887: Hack together slop without process. You could just manage to get yourself past it. So I said, okay. Raw dog, just like slop it out, right? And so it worked, but it was fucking insanely bad, right? It was almost unusable, but you could just manage to like get yourself past it, right? And so I said, okay, we've been dealing with this for two weeks. It's super annoying. And so I said, okay, we've been dealing with this for two weeks, super annoying. stop Let's zoom out. Let's build a system for... let's zoom out, let's build a system because the issue with mobile responsive is like the breakpoints Let's build a system for... And this is where choosing the right agents is important. Kimi is Kimi is Go fucking crack at mobile responsive development. So fucking cracked at mobile-responsive development. is so fucking cracked at mobile responsive development.
slop.atg.eth: Huh, okay. Okay.
Anon5887: The K3.
slop.atg.eth: The
Anon5887: The most cracked mobile UX agent.
slop.atg.eth: Now, Wow
Anon5887: So I had about 50 Kimi agents, better than Fable, like better than like all these guys, right? So I had about 50 Kimi agents. Better than Fable, like better than all these guys, right? They created a front-end mobile harness The front end? They created a front end mobile harness? like three different breakpoints, It took like four or five days for the whole infrastructure to get stood up. Like three different breakpoints, they created like a smoke app just for this mobile harness, built a workflow, iterated, iterated, took like four or five days for them to get this whole infrastructure stood up. Then they ran the full audit using that infrastructure, built a workflow, iterated, iterated, took like four or five days for them to get this whole infrastructure stood up. Then they ran the full audit using that infrastructure, captured every single defect Then they went into research on like best practices for like mobile UX and whatever. Then they went into research on best practices for mobile UX. Great job, Great job, right?
slop.atg.eth: You're the design expert. You're the design experts, not me. You're the design expert, not me.
Anon5887: And so they created a design system and all of these things. Do it, right? And so they created a design system And so they created a design system. out of the audit to fix all these things, right? out of the audit to fix all these things. to fix all these things, right? And they ran overnight. And I was away for the weekend with my wife. We went down south to ski, right? And so I woke up, this is like Saturday morning. I wake up Saturday morning, we're going to go skiing. I wake up Saturday morning, we're like going to go skiing. I wake up Saturday morning, we're like going to go skiing. Like, you're making a mobile So I open it up, and I said to her, I was like, this is going to be interesting, right, like these guys have just been running all night, like I want to see, like, is it just pure slop, but you're making a mobile, So I open it up and I said to her, I was like, this is going to be interesting, right? Like these guys have just been running all night. Like I want to see, like, is it just pure slop of the, you're making a mobile I'm about to use that, right? We're about to use that, right?
slop.atg.eth: ...making a mobile app, and you're about to use the mobile app, on your mobile phone, in the wild Sure
Anon5887: And so I pulled up the old version, right? And I press the button, and it refreshes, and it was fucking pristine. I press the button and it refreshes and it was fucking pristine.
slop.atg.eth: Oh yeah. Hell yeah.
Anon5887: It was beautiful. beautiful It was beautiful, like, because they put all the work in to get it right, and then they checked everything, and you know, so, like, there's nothing that stops an agent from Doing good work except bad process. Because they put all the work in to get it right and then they checked everything. A single agent with no process is going to be horrifically bad. single agent with no process is going to be horrifically bad. A single agent with no process is going to be horrifically bad. And if that's all you've seen, if you say, yeah, I tried to get an agent to do a mock-up of a thing and it was terrible, it's like, you talk to one guy who did one thing, like, of course it's going to be bad, right? If you have a process where you've got all the guardrails and you've done the research and you've got all of the right structures, you do that process and then it'll land and it'll be amazing. If you have a process where you've got all the guardrails and you've done the research and you've got all of the right structures and you do that process, then it'll land and it'll be amazing.
slop.atg.eth: That's great advice. That's awesome. That's great advice.
Anon5887: But you have to spend the time. But you have to spend the time, but the thing is, what you don't want to spend time doing is things that are not reproducible. Now, that workflow, I run it like
slop.atg.eth: Do it again ###
Anon5887: I'm like, hey, go and run the audit. We've done a bunch of things. And they just, they run it and they clean everything up and it goes from like, ah, there's a bunch of slop creeps in and then they just refactor it and it's great. And it takes like four hours. Go and run the audit. We've done a bunch of things. We've got a bunch of new features. Run the audit. And they just, they run it and they clean everything up. And it goes from like, oh, there's a bunch of slop creeps in and then they just refactor it and it's great. And it takes like four hours.
slop.atg.eth: Really cool.
Anon5887: Yeah,
slop.atg.eth: And you said it was just like a button push to go from V2 to V3? You said it was just like a button push to go from V2 to V3.
Anon5887: So this is the deployment, right? So this is the deployment. this morning. like 9am this morning Yes.
slop.atg.eth: Wow, and it could be itself. Wow, and it could be itself, like you're gonna hit a button and it's gonna upgrade itself and you have to like reload the screen.
Anon5887: Let's see what did we get here? Yeah, so what did we get? Let's go, what did we get?
slop.atg.eth: Oh no, we're gonna do it live! Yeah, Yeah, yeah
Anon5887: We'll do a live deployment, right? We'll do a live deployment. So this is the thing right?
slop.atg.eth: Are we talking to the overseer now?
Anon5887: Yeah
slop.atg.eth: And you're using Grok for that? You're using Grok for that, not ###
Anon5887: Scrap Descrop Grok is fast and smart. Grok 4.6 is fucking opus class. Grok is fast and smart. Grok 4.6 is fucking Opus-class. Grok is fast and smart. Like, don't be mistaken, it is it's an opus class agent. Don't be mistaken, it's an opus class agent, it's very fast, and it is, I think I said this earlier, it's biased to action. Don't be mistaken, it's an Opus class agent, it's very fast, and I think I said this earlier, it's biased to action. You want your orchestration and managing agents to want to do things, right? You want your orchestration and managing agents
slop.atg.eth: The
Anon5887: Sol 5.6 is not biased towards action; it's petrified. Like SOL 5.6 is not biased towards actions. I don't know what they've been doing to that guy in their test environments.
slop.atg.eth: He has to get it exactly right. Just to get it exactly right.
Anon5887: He's like, ah, like I'm gonna He's like, "I'm gonna" and ends up like building He's like, ah, like I'm gonna-- and it ends up like building this. you know, system around itself system around itself to protect everything crazy system around itself to protect everything. alive The
slop.atg.eth: The Yeah.
Anon5887: I've got to check this, so this guy is just much faster Yeah, but I've got to check this. so this guy is just much faster And we'll just push things through. Things were Move things forward. Try and move things forward.
slop.atg.eth: You've got Grok is biased toward action. you've got a lot of Chinese agents doing red teaming. You've got Grok, is biased toward action. You're using him as an overseer. You've got a lot of Chinese agents doing red teaming. Kimmy's really good at mobile development. What are some more of these like models?
Anon5887: Yeah.
slop.atg.eth: certain things that you have kind of tucked away in your brain here. Certain things that you have kind of tucked away in your brain. certain things that you have tucked away in your brain.
Anon5887: So the current best delivery workflow is five agents And that's the deploy going out. Oh, that's the deploy going out?
slop.atg.eth: Four-twenty 420 Rule 420
Anon5887: Oh, they're up to the 420 pool. It was like an inbox message from the red team Oh, they're up to the 420 pool. The following is a So the current delivery workflow that I've been using is Deepseek as the manager. Music is very biased towards action as well. So the current delivery workflow that I've been using is DeepSeek as the manager. DeepSeek is very biased towards action as well. So the guy who's managing the whole program of 20 tasks? is DeepSeek. I'm using that a lot I'm using that a lot. If it's a really hard program, I'll use Astro is Deepseek. I'm using that a lot. If it's a really hard program, I'll use Astra. Not very token efficient very token efficient. Astro is very token efficient. Each context turn is cached and the incremental cost is pretty low. is cached and the incremental cost is pretty low, so it tends to be like 95% cached. The incremental cost is pretty low, so you know it tends to be like 95% cash. Um The following is Right. All right ###
slop.atg.eth: You always want to be effective first and efficient second. And to me, it seems like you're leaning a little bit toward efficient. And this is another tension that I've talked to a lot of sloperators about, and it's like effective versus efficient. You always want to be effective first and efficient second. And to me it seems like you're leaning a little bit toward efficient ahead where you're like
Anon5887: Well, different roles.
slop.atg.eth: Let's not use fable because it's going to use too many tokens. Let's not use Fabel because it's gonna use too many tokens. Let's not use Fable because it's going to use too many tokens," where to me it would be like, always use Fable first and then start leaning back on more effective models as you have a baseline. The following is The
Anon5887: Well, different roles, right? So Fable, I used as the author for any hard different roles, right? So Fable I used as the author for any hard Different roles, right? So, Fable I used as the alter for any hard
slop.atg.eth: Loves to write. Yeah, loves to talk. Loves to write. Loves to talk. Okay, okay. Okay, The
Anon5887: Yeah, like any hard... Sorry, author is like coder, right? Yeah. Like the guy's writing the code, Yeah, like any hard, sorry,_author is like coder, right?
slop.atg.eth: Astra or even GPT Soul. Not Astra, or even GPT Seoul. Astra, or even GPT Sol.
Anon5887: So the best reviewer that I have right now of all the models is 5.6 extra high.
slop.atg.eth: It seems like
Anon5887: Thank you The best review
slop.atg.eth: 5.6 always finds something wrong that Fable did. Fable implements it, and 5.6 finds a problem, and you'll loop six times picking shit up.
Anon5887: So,
slop.atg.eth: You'd rather not have 5.6 be the implementer and fable be the reviewer you'd rather not have 5.6 be the implementer
Anon5887: Don't invert them.
slop.atg.eth: Ok. Okay.
Anon5887: Fable is optimistic. Don't invert them. Fable is optimistic. Paul is pessimistic. Babe will be like, this looks pretty good. they will be like, "Yeah, this looks pretty good." they will be like, this looks pretty good.
slop.atg.eth: You want those pessimistic guys checking things. Yeah. You want the pessimistic guy checking things. Okay, you want the pessimistic guy checking things. Okay, that makes sense. Yes.
Anon5887: You want the autistic neckbeard guy who's cynical to be checking the things, right?
slop.atg.eth: I've always said like Fable Claude stuff was more artistic and Codex was more autistic. Yeah. The
Anon5887: Yeah, it's got a flair to it. Yeah, it's got a flair to it. So Fable as a code writer, really good if it's a multi-disciplinary task, if it's something that touches the back end. Yeah, it's got a flair to it, right? So Fable as a code writer, really good if it's a, call it like a multi-disciplinary task, right? If it's something that like touches the back end. It's really has good judgment. multiple modules, it really has good judgment. It's not myopic, right? Sol 5.6 as reviewer, and then my auditor is still GLM 5.3. Golem 5.3 is the best auditor. It's not myopic, right? It's not myopic. of any of them, right? I think a combination Of any of them, right? And it's, I think, a combination of, like, of any of them, right? The of the chats All of the chats The checks. It's not like second guessing whether it should do the thing. So GLM is very good. It's not second guessing whether it should do the thing. list of This is the list of of the shit that The shit that released Let's see... least, let's see These are all These are all the The following
slop.atg.eth: Is it working out of GitHub too? These are like issues within GitHub.
Anon5887: Everything's in Git, but these are like our task manager. Yeah, so everything's in Git, but these are like our task manager, right? Yeah, everything's in git, but these are like our task manager. So this is like all of the All of these context changes just landed. This whole refactor of how context works, where the prompts go, This whole refactor of like how context works, This whole refactor of like how context works, where the prompts go.
slop.atg.eth: landed with that button with that button push you did landed with that button push you did Now Wow
Anon5887: Yeah, they just ran the whole deployment. that button push. Now this guy will be doing post-validation on prod. Yeah, brand whole deployment. Now this guy will be doing post-validation on prod. a whole validation suite validation suite to make sure that everything's working. The Um All right, we've got a whole new system for how prompts are constructed. So now the next thing will be, all right, we've got a whole new system. So now the next thing will be, all right, we've got a whole new system for how prompts are constructed. Is it still the same? And the nice thing is it's just pros, right? is optimized. And so another agent will come in over the top today and sit there all day, check every single workflow, make sure that they AB tested against the old version, the new version, like is it still the same? And the nice thing is it's just pros, right? So it's an easy difference And so another agent will come in over the top today and sit there all day, check every single workflow, make sure that they AB test it against the old version, the new version. Like, is it still the same? And the nice thing is it's just pros, right? So it's an easy diff The output of this and the output of that are identical The output of this and output of that are identical. And then another guy will come over the top and be like, well, we can actually specify roles now at a much more granular level. So let's break up these two roles. We had a monitor and a planning agent. Let's make it two different agents. And then another guy would come and be like, well, we can actually specify roles now. and a planning agent, let's make it two different agents. agent and a monitor The suivantes: Ah, I guess so.
slop.atg.eth: Sorry, my graphics are crazy. You talk with your hands a lot more than most of my guests. Sorry my graphics are crazy. You talk with your hands a lot more than most of my guests. You almost needed a way to turn that off. But this is part of your process. Process is super important and you can get so much done in a very good way with good process Okay, but this is part of your process, so you really leaned into like process is super important and you can get so much done in a very good way with good process. But this is part of your process. So you really leaned into like, process is super important, and you can get so much done in a very good way with good process. You actually constructed this process to say, you do the thing, you check the thing, then after you're done checking the thing, you sort of think outside the box about how to make the thing better or more efficient. And there's like basically a playbook here. You do the thing, you check the thing, then after you're done checking the thing, you sort of think outside the box about how to make the thing better or more efficient. ...goes through orchestrating all these different agents Is that right? And goes through orchestrating all these different agents to do these things through these like phases. and goes through orchestrating all these different agents to do these things through these like phases. Is that right?
Anon5887: And my thesis behind that, right, is that like we are all aiming for recursive self-improvement. Exactly. And my thesis behind that, right, is that, like, we are all aiming for recursive self-improvement.
slop.atg.eth: It's just Like it's just
Anon5887: You cannot have recursive cannot have recursive Like, we have?
slop.atg.eth: One more step at a time until all of a sudden it's kind of looping and just doing it. One more step at a time until all of a sudden it's kind of looping and just doing it and it's like you're getting there already.
Anon5887: We are in recursive self-improvement right now. We are in recursive, just to be clear, right? We are in recursive, just to be clear, right? We're in recursive self-improvement right now. We're in the early phases. Looking at the chart it's like you can't see it,
slop.atg.eth: It hasn't mooned yet, yeah, yeah, yeah. It hasn't mooned yet.
Anon5887: It hasn't moved yet, Yet, but it's ready, right?
slop.atg.eth: You want to buy right now. Now's the time to buy. Now is the time to buy. You want to buy right now.
Anon5887: But the thing that is blocking us the humans in the loop Alright, the humans in the loop. So it's not exactly recursive self-improvement It's not exactly recursive self-improvement because So it's not exactly recursive self-improvement because we do like one loop and then it grinds to a halt and then the human has to kick it, right? And we just don't have the orchestration infrastructure Get out of the agent's way. And we just don't have the orchestration infrastructure to get out of the agent's way. When I started vibe coding, the thing that I realized very quickly I was the worst part of the system. Even when I was using o3 mini I was using o3 mini Even when I was using 03 mini The dumbest agent to ever write a line of code in history. dumbest agent to ever write a line of code in history
slop.atg.eth: [laughter] like three years ago now
Anon5887: So, O3 Mini couldn't even fucking run a Git
slop.atg.eth: Like three years ago now, probably
Anon5887: So o3mini couldn't even fucking run a git So, O3 mini couldn't even fucking run a
slop.atg.eth: years ago now probably, right? Okay Okay, okay.
Anon5887: No, no, this is like 18 months ago. No, no, this is 18 months ago. O3 minihigh was cracked for the time, but like literally you would say like okay commit that and it would mess up the git command every time. No, this is 18 months ago. Like o3-mini-high was cracked for the time, but like literally you'd say like, okay, commit that and it would mess up the git commit every time.
slop.atg.eth: (context: ###
Anon5887: So anyway, The thing that was very, very The thing that was very, very obvious to me early I was doing this. I am the slowest worst part of this loop. I am the slowest horse part of this loop. that I'm typing The fact that I'm typing
slop.atg.eth: It seems like then you keep putting orchestrators in? It seems like then you keep putting orchestrators in. Like, "OK, now I've got this doing it, let's put in an overseer." And it seems like you keep building these things, these layers between you and the machine, that do what you're currently doing so you don't have to do it. It seems like then you keep putting orchestrators in. Like, okay, now I've got this doing it, let's put in an overseer. And it seems like you keep building these things, these layers between you and the machine that do what you're currently doing so you don't have to do it.
Anon5887: My thing in startups for the last 20 years in process of proving the process of recruiting and that's allowed me It's allowed me to allowed me to Anything done. My method of operating is like we should always be improving our processes. My method of operating is like we should always be improving our processes. We had iPads in 2000 locations around Australia. At BlueShift, we had iPads in 2,000 locations around Australia. You can get those iPads There's a bunch of different ways There's a bunch of different ways you can get those iPads. You can go and buy one iPad and you can take it out of its box and you can install the software and you can do the thing. It'll take you all day and you ship one a day, right? Or you can industrialize it, which is what we did. You can install the software and you can do the thing, and it'll take you all day and you ship one a day, right? You can install the software, and you can do the thing, it'll take you all day and you ship one a day. Or you can industrialize it, which is what we did. We would buy 50 to 100 iPads at a time. We had like five guys, they would pull all the iPads out of the boxes, they would set up all the iPads, put covers on them We had like five guys that would pull all the iPads out of the boxes, they would set up all the iPads, put covers on them We had like five guys, they would pull all the iPads out of the boxes to plug them in, step three Step three, MDM to plug them in, step three MDM
slop.atg.eth: Now, is that Grok or GLM or KIMI? Who was that doing that? Human beings, amazing! Now, is that Grok or GLM or Kimmy? Who is that doing that? Human beings, amazing. Now, is that Grok or GLM or Kimi? Who was that doing that?
Anon5887: actual human beings, right? These were actual human beings, right? When we were scaling up the network, we would get 50 new applicants. When we were scaling up the network, we would get 50 new applicants, and we had people that were waiting for like three weeks for iPads. When we were scaling up the network, right? We would get 50 new applicants, right? And we had people that were waiting for like three weeks for iPads. They were like, "I want to join your network" and they wait for three weeks. And I was like, "This is crazy." So we optimize the process and then we optimize it. and we got to the point where it was fully automated, We optimize the process. And then we optimize it again. We optimize the process and then we optimize it again. We got to the point where it was fully automated, right? You want humans out of any loop. Humans are the most expensive, dumbest, slowest part of the loop. It's the most expensive, dumbest, slowest part of the loop. are the most expensive, dumbest, slowest part of the loop
slop.atg.eth: It's getting easier and easier to replace them. Like Right, like
Anon5887: The So the thing that Zobrist is designed to do The is tighten that recursive self-improvement Tighten that recursive self-improvement and, you know, And you... on a delivery You run a delivery, then you run the next delivery. You don't have process improvement. If you don't have process improvement, then you learn nothing from it. Even if you're just pressing the same button over and over. Even if you're just pressing the same button over and over, if you don't have a feedback loop, where after you run five There is an audit is an audit is automated guy goes, oh Automated. And a guy goes, Time to run the weekly audit, runs an audit. It's time to run the weekly audit. If you don't have that tightening of the loop, you will be doomed to just keep doing a bad process. Time to run the weekly audit. If you don't have that tightening of the loop, you will be doomed to just keep doing a bad process continuously. My autism is like Totally focused on that. focused on that. Like literally like It gives me like a little dopamine hit, and I'll just sit there all day. It gives me like a little dopamine hit and I'll just sit there all day being like tweak, fixing all the little things. It gives me like a little dopamine hit, and I'll just sit there all day being like,
slop.atg.eth: So process is important It needs to be defined what the process is. But the important thing here is the feedback loop that improves the process, and that's the critical piece. it needs to be defined what the process is Yeah. It's about the definition of the process and then the feedback loop that improves the process. It's not that important. It's about the definition of the process and then the feedback loop that improves the process.
Anon5887: It's loops within cycles, and epicycles, right? It's loops within loops, right? So each process is a loop, but then there's a meta-process around the process, and there's a meta-process around that. It's loops within loops, right? So, each process is a loop, but then there's a metaprocess around the process, then there's a metaprocess around that, right? And they are all interconnected and they have to be as efficiently coordinated as possible. And so, you know, you want a system, and the thing that's exciting to me is agents are And you know, they are all interconnected. And so, you know, you want a system, and the thing that's exciting to me is agents are the most efficient process optimizers we've ever - you think they're good at writing code? They are so good at optimizing. efficient process optimizers we've ever The following most efficient process optimizers. They have all the information. Have all the information
slop.atg.eth: And you talk about backtesting, they have all that data from a year of usage Especially when they have hella observability into your whole system. Especially when they have like hella observability into your whole system. more efficiency here If we make this change, that's interesting. much more efficiency here if we make this change.
Anon5887: And this is the thing that like, you have to have all the components, right? Because if you just had observability And this is the thing that like, agents could find out all the problems and the agents could find out all the problems, then you had to carry that over the line. But then you had to carry that over the line. Never gonna do it. You have to also have the second component. You also have the second component. You have to have the research and audit ...into plans, then you have to be able to execute those plans... plans, then you have to be able to execute those plans, then you have to be able to execute those plans If executing it requires a hundred hours of your brain sitting typing things And if executing it requires 100 hours of your brain sitting typing things?
slop.atg.eth: It's that list that he gave you the other day. That list that he gave you the other day.
Anon5887: Just fucking do it, man. Like just fucking do it, man. There's a bunch of things that we've kind of learned in like how we build things and how we like construct whereas the cost of a bad decision is very high. Welcome that decision with agencies not I
slop.atg.eth: Right.
Anon5887: The cost of a bad decision with agents is not high.
slop.atg.eth: The
Anon5887: actually very low and very valid Actually, it's actually very low and very valuable. Making a mistake with agents, you roll the thing back. Making a mistake with agents, there's no one-way doors. Making a mistake with agents, you know, there's no one-way doors. alright I've rolled it back And like ten minutes later, it's like, all right, I've rolled it back. It's like, all right, I've rolled it back. You have to lean into that.
slop.atg.eth: So, okay, processes Okay, processes feedback loops around the process Feedback loops around the process that involve all of these different steps that you're talking about. You said earlier, like the event horizon of human interface. I think that's kind of what you're saying here, is that the human interface with the system is constantly being abstracted away, and there's less and less human stuff that has to happen. I think that's kind of what you're saying here is that the human interface with the system is constantly, if you're doing it correctly, being abstracted away and there's less and less human stuff that has to happen.
Anon5887: There's more complexity. I have 468 sessions running right now. And therefore there's more complexity that, you know, I have 468 sessions running right now. The The following Like
slop.atg.eth: My harness has like eight or nine, or like at most 50. My harness has like eight or nine, or like at most 50. Like there's never that many because I'm clicking through each one and dealing with them, and you're not doing that. The
Anon5887: I'm not doing that. And there's another layer of abstraction that we're about to build, which is a console that gives you a view into the agents where you actually have much more granular but visual feedback about what's happening. I'm not doing that, right? And, you know, there's another layer of abstraction that we're about to build, which is a console that gives you a view into the agents where you actually have much more granular visual feedback about what's happening. We're not doing that, right? There's another layer of abstraction that we're about to build, which is a console that gives you a view into the agents where you actually have much more granular visual feedback about what's happening. I will still go, "I wonder how this guy's doing". remote
slop.atg.eth: Like drop right into his business and see exactly what's going on. Okay Yeah
Anon5887: But what you actually want is a layer that's wrapping. But what you actually want is a layer that's wrapping, this is still too raw. I look at this and I go, I want another layer on top of this. In the limit, what you want is you turn up and you go, I don't like this. And then you turn up the next day and it's different. Ultimately, what you want is you turn up and you go, I don't like this. That's what you want, right? That's what you want, The no amount of smartness. There's no amount of smartness. until we get Until we get to artificial superintelligent
slop.atg.eth: Superhuman intelligence is coming, Yeah.
Anon5887: That's a different thing, right?
slop.atg.eth: Yes, please don't kill me.
Anon5887: So, you know, if you have a guy that's hyper-intelligent So you know, if you have a guy that's hyper-intelligent, And yes, you can just say, I would like a Ferrari. So if you have a guy that's hyper-intelligent, right, then yes, you can just say, I would like a Ferrari, and you come the next day and it's built it for you. Cool. Amazing, right? Like that's cool. In the interim We don't have artificial superintelligence. We don't have artificial super. Astra's not smart enough that you can just say a thing There's a process that has to happen It must exist. office that must exist The following is a transcript be able to tell Astra what you want and for it to have a chance of delivering that. tell Astra what you want and for it to have a chance of delivering that You tell Astra what you want and for it to have a chance of delivering. Without you carrying the load the entire time. without you carrying the load the entire time alongside me, which is pointless. if I'm lugging 90% of the weight and Astra is just kind of step without you carrying the load the entire time. I want to be able to say fix this thing
slop.atg.eth: You want to go skiing? I want to go skiing.
Anon5887: So, again
slop.atg.eth: Do you want to go skiing? Hello The
Anon5887: Yeah, I want to go skiing and I wake up the next morning and it's beautiful. I think that is the primary outcome that I'm trying to achieve is building a system Yeah, I want to go skiing and I wake up the next morning and it's beautiful. And I'm like, okay, great job. So I think that is the primary outcome that I'm trying to achieve, is building a system The following is The following is a substrate for all of these things to be able to continuously improve as the substrate for all of these things to be able to continuously improve straight for all of these things to be able to continuously improve If I land that, then I've gotten myself out of the loop. I've gotten closer to, okay, this is the optimal thing for the current level of intelligence that agent has. I've gotten closer to, okay, this is the optimal thing for the current level of intelligence that, and like I threw Astrid in here, he loves it, but like he's not so smart that this is not useful to him. And I've gotten closer to, okay, this is the optimal thing for the current level of intelligence that agent has, and like I threw Astra in here, he loves it, but like he's not so smart that this is not useful to him.
slop.atg.eth: The The following This
Anon5887: It's still a cool place to put him. It's still a cool place to put in. Still a cool place to put it.
slop.atg.eth: With a harness, you want to be assuming that the models are going to get better. And with a harness, you want to be assuming that the models are going to get better
Anon5887: But yes Yes So
slop.atg.eth: As you're building your harness, there's probably some parts of the harness that will get out of the way as the models get better, or maybe it's just more extracted and you have a smarter model on the top. What do you think about that? As you're thinking, as you're building your harness, there's probably some parts of the harness that will get out of the way as the models get better, or maybe it's just more extracted and you have a smarter model on the top? What do you think about that? True The
Anon5887: I think there are absolutely things that will not be needed, right? I think there are absolutely things that will not be needed, right? Checks and balances that will no longer... I think there are absolutely things that will not be needed, right? Checks and balances that will no longer, as the agents get smarter, they make less mistakes, less checks and balances, right? You still want some, because still stochastic process, still random, right? So you still You roll over, even if Even if you're If it's bad when that happens, then make sure that there's a guy standing there to catch the ball before it falls into double zero. only getting double zero on the roulette wheel one in 33 times. getting double zero on the roulette wheel one in 33 times. If it's bad when that happens, then make sure that there's a guy standing there to catch the ball before it falls into double zero.
slop.atg.eth: ### Sure.
Anon5887: There's no point in saying, "They're so good, you don't need to watch them." There's no point in being like, oh, they're so good. Like you just don't like there's no point in being like, oh, they're so good, you don't need to watch them. The thing that I think people, I just realized that
slop.atg.eth: Yeah, it's like It's like
Anon5887: So The following is
slop.atg.eth: Yeah, yeah, sorry Yeah, sorry. Yeah
Anon5887: No, no, no, I thought we were like an hour in. I thought we were like an hour in, so maybe we'll try and wrap up in like five minutes. No, I thought we were like an hour in. So maybe we'll try and wrap up in like five minutes. But the thing that I do not think will change in the next six months at least, right, which is an... ### The following is a transcript The following is
slop.atg.eth: vacuum. Yeah
Anon5887: The optimal context window for each agent is critical to manage, right? the optimal context window for each agent is critical to manage. schemes for different agents This compaction scheme is custom-designed for the orchestrator. So this compaction scheme is custom designed for the orchestrator. Basically derives the obligations that are open It basically derives the obligations that are open It basically derives the obligations that are open because, remember, he's not doing anything. If you compact him down to zero, and he has to, from first principles, work out what's happening, he can. if you compact him down to zero and he has to from first principles work out what's happening, he can. But all the information is there. You want to give him just enough, right, so that he's got like the last few turns, he knows where he was up to if you compact him down to zero and he has to, from first principles, work out what's happening, he can't. But he can, right? All the information is there. You want to give him just enough, right, so that he's got the last few turns, he knows where he was up to. You can compact that. Now, why do you want to do that? You can compact that. Now, why do you want to do that, right? to 200. What I found is like a hundred to 200 what I found is like 100 to 200 Within 100 to 200? In a hundred to two hundred?
slop.atg.eth: That's when he's going to get his best work done. That's when he's gonna get his best work done.
Anon5887: That's where like, it's on point, good judgment, whatever.
slop.atg.eth: [laughter]
Anon5887: That's where good judgment, whatever. Beyond 200, it starts to degrade a little bit. Beyond 300, all bets are off. That's a one million context window. If you are sitting there in an Astra session at 800k context, God help you.
slop.atg.eth: Yeah
Anon5887: That's the best, like, that's where like zone point, good judgment, whatever. Beyond 200, it starts to degrade a little bit. Beyond 300, all bets are off, right? That's a one million context window. If you are sitting there in an astro session at 800k context, God help you.
slop.atg.eth: The
Anon5887: 1 million context window. If you are sitting there in an Astra session at 800k context, God help you. You're talking to someone who's brain-damaged. They're going to sound plausibly smart, but they're making bad decisions. They're drunk. You're talking to someone who's brain damaged. They're gonna sound plausibly smart, but they're making bad decisions. They're drunk. agents can manage this. Got to go home. So this type of thing, also you've got these other systems which are... and agents can manage this. The managers can manage the context windows of all of their guys 50 to 200k context window. 150 to 200k context window.
slop.atg.eth: Sorry, sorry, I need to figure out how to turn that off.
Anon5887: just enough, but not too much But if one of their authors, one of their coders is... If one of their coders is doing a long-running task that might take a couple of hours, one of their coders is doing a long-running task that might take a couple of hours, right. If one of their coders is doing a long-running task that might take a couple of hours, they want to keep his context in that. They want to keep his contacts in that, and so you give them a set of tools, they can do it. Otherwise, who's going to do that? You. And you don't want to be three layers deep finding the astra guy and being like, oh, hey, this guy seems like he's grinding hard. You don't want to be three layers deep finding the astro guy and being like, oh, hey, this guy seems like he's grinding hard. Like, no, that's not your problem. His manager goes, this guy's been working for an hour. We've got to make sure we trim his context and get him back to You don't want to be three layers deep finding the astro guy and being like, oh, hey, this guy seems like he's grinding hard. Like, no, that's not your problem. He's got a manager. His manager goes, this guy's been working for an hour. We've got to make sure we trim his context and get him back to So there's a bunch of things like that Give the agents the tools to optimize themselves. fighting trim. So there's a bunch of things like that in the harness So there's a bunch of things like that in the harness that are built in to give the agents the tools to optimize themselves. optimal output. The optimal output. correct
slop.atg.eth: We're still at the point though where you find, oh like, I think I want better compaction here or something. We're still at the point though where you find "Oh I think I want better compaction here" We're still at the point though where you find, oh, like, I think I want better compaction here or something, like there is the process of creating and checking in and thinking a little bit about making it more efficient. But when you're using the tool, you come up with
Anon5887: So this is
slop.atg.eth: We need this new thing that does this thing. oh, fuck, we need this here. We need this new thing that does this thing.
Anon5887: The interesting thing is the compaction system came from a research.
slop.atg.eth: Okay, interesting.
Anon5887: I saw a tweet, and the tweet was like The interesting thing is the compaction system came from a research exercise. The following is a The
slop.atg.eth: Then you send them to go research on that tweet. send them to go research on that tweet The ### Yeah
Anon5887: I'm like, hey man, go and be the expert in this. I'm like, hey man, go and be the expert in this, right?
slop.atg.eth: The ###
Anon5887: I've got three Grok agents whose life purpose is reading this fucking one paper, right? I've got three Grok agents whose life purpose is reading this fucking one paper, right? And so they've read the paper, they know it, they've gone and like got the corroborating information all the other papers and then they come back and they go, And then they come back and they go, we're using native Py. And then they come back and they go, we’re using native Pi. So it has compaction. And I was like, well, what would the optimal compaction look like? So it has compaction. And I was like, well, what would the optimal compaction look like? And they're like, well, the optimal compaction is like custom designed compaction for each role.
slop.atg.eth: Depending on the job, yeah.
Anon5887: on the job, right? And with triggered windows and tools and everything. And so I was like, all right, build the first instance of that. windows and tools and everything. So I was like, all right, build the first instance of that. action fashion Again, is it recursive self-improvement? No then they were like, well, now we need these tools. They didn't retweet it and go, "Oh, this new compaction research, let's go on." But we're so close. I didn't read Twitter and go, oh, this new compaction research, let's go on. Then they were like, "Well, now we need these tools."
slop.atg.eth: Yeah.
Anon5887: We're so close to just having a guy that's sitting on Twitter going like, this looks interesting We're so close to just having a guy that's sitting on Twitter going like, We're so close to just having a guy that's sitting on Twitter going like, oh, this looks interesting.
slop.atg.eth: and It's like you your job right now, It's like your job right now Yeah, it's like your job right now, and I don't even know
Anon5887: Like oh, okay Oh, okay
slop.atg.eth: You use the Lamborghini, like you just show up and there's a Lamborghini. There will be some aha neat moments when it all happens, I think. You use the Lamborghini. You just show up, and there's a Lamborghini. There will be some aha neat moments when it all happens, I think. Do you detect when they start doing your job? Like you use the Lamborghini, like you just show up and there's a Lamborghini. Like there will be some aha neat moments when it all happens, I think. This is a transcript of an English-language conversation.
Anon5887: Even the red team thing, like they, I didn't tell them I didn't tell them to dedupe the bytecode and look at the deployed contract Even the red team thing, I didn't tell them to dedupe the bytecode and look at the deployed contract as the pointer. They came up with that themselves. Some of it they put together, like they didn't invent new schemes, like some of that was probably in each case. Some of it they put together, like they didn't invent new schemes, like some of that was probably in each skills. Some of it they put together, like they didn't invent new schemes.
slop.atg.eth: Hell yeah!
Anon5887: I didn't read East Guilds I didn't read eSkills, I don't fucking know,
slop.atg.eth: Don't read it.
Anon5887: Like I didn't read any skills. The following So, like, they read EVM skills. So they know how bytecode modifies based on the deployed address. So like they write eSkills, they know how bytecode modifies based on the deployed address, and you need to factor that out when you're doing... They came up with that whole process And then they systematized it. ...contract and de-dupe it down to 900 or whatever it was that they did. contracts and they juped down to 900 or whatever and dedupe it down to 900 or whatever it was that they did? So that's recursive self-improvement. the same con- like they just figured it out, right? tracking contracts and it's the same content. Like they just figured it out, right? We're making contracts and they just figured it out, right?
slop.atg.eth: Whenever I try to do stuff like this, it just ends up being very messy.
Anon5887: Let's do it.
slop.atg.eth: You have a mess on your hands sometimes with this amount Like don't you have a mess on your hands sometimes with this amount of But like don't you have a mess on your hands sometimes with this amount? This amount of agents with this amount of freedom? ...and this amount of agents with this amount of freedom? ...creation and this amount of agents with this amount of freedom?
Anon5887: The following is The following is a The ###
slop.atg.eth: You said the mess was cheap earlier, so maybe that's fine.
Anon5887: The mess is cheap, right? I'm a process improvement guy I'm better than the agents at process improvement, right? I'm trying to think of an example of a thing that we did, right? I had in my old harness this concept of an outcome. All right A human readable outcome, a chunk of prose that was like, here is the reason why we are doing this thing. a human-readable outcome, a chunk of prose that was like, "Here is the reason why we are doing this thing. This is the outcome that we want." human-readable outcome, a chunk of prose that was like, here is the reason why we are doing this thing. Anytime an agent did anything, So anytime an agent did anything, Does this match the outcome? Will this thing achieve the outcome? Does this match the outcome? And if we do this thing, will any part of the outcome not get delivered? does this match the outcome, will any part of the outcome not get delivered? And that's like an anchor, right? Then I built all these other things and that kind of got dropped. I did a big refactor and Then I built all these other things, and that kind of got dropped. And I did a big refactor and I opened it up after it was done. And they're like, oh, it's not part of the acceptance criteria. Where is this thing? And they're like, "Oh, it's not part of the acceptance criteria." And I was like, what do you mean? We literally had a whole conversation, we planned it all out. And I was like, "Show me the planning doc." And I was like, show me the planning doc, which I didn't read obviously, but I was like, now I'm curious, right? Like how did we get this wrong? And so we went through the whole thing and I saw the planning documents and it was all like plausible, but it didn't achieve the outcome. And so then I was like, okay guys, now We're now We need to reinstitute. reinstitute this out. moving so quickly.
slop.atg.eth: It's not about the plan, it's about the outcome.
Anon5887: And the reason why I had And the reason why I had to reintroduce that is before This is why I say the event horizon of what you can And this is why I say the event horizon of what you can achieve. And this is why I say the event horizon of what you can achieve, right? When you're running eight agents When you're running aid agents There's eight guys doing one thing, I can reason about that. There's not much cognitive load. When you have Eight guys doing fifty things. Eight guys doing 50 things. So I need some abstraction. I can't reason about it, so I need some abstraction. Right? I can't reason about what, so I need some abstraction, right? So it's like, okay, what's the abstraction? The abstraction layer is, here is your outcome. If you are in the middle of doing something If you are in the middle of doing something, and this violates that outcome.
slop.atg.eth: Yeah The
Anon5887: If you are in the middle of doing something, Or if you did all of these things, the outcome wouldn't be achieved, fucking give me a call, man. That's my job, right? hün녕 You do all of this stuff, this thing that you want to happen won't happen. If you did all of these things, the outcome wouldn't be achieved, fucking give me a call, man. That's my job, right? Give me a call and tell me like, hey, if we do all of this stuff, this thing that you want to happen won't happen. We need to fix that, right? You can have the best agents, the best harness, the best substrate, but like, there is still a process. You can have the best agents, the best harness, the best substrate. and I was a project manager in my early career. So making sure that a thing that you wanted to have happen, Management, you know, and I was a project manager in my early career, right? My bread and butter. And I was a project manager in my early career, making sure that a thing that you wanted to have happen happened, is my bread and butter
slop.atg.eth: That was very important with humans, too, yes. That was very important with humans, too. That's very important with humans too. Yeah Okay, other threads Other threads Yeah, yeah, yeah. Yeah yeah yeah, I mean
Anon5887: Let's do it. We'll do a lightning round. Awesome thread, let's do it.
slop.atg.eth: The fact that you have so many open weight models Like generally I don't see The fact that you have so many open-weight models is awesome. Like generally I don't see A lot of people using a lot of open weights models A lot of people using a lot of open weights models, and so just like in general, the fact that you're using the open weights models so so much is awesome. much is awesome And then that leads me down the trail of like local also. You must at some point have thought like spending $100,000 on all these open-source models. spending a hundred thousand dollars on all these open source models. Why don't I just get a data center and and run it there? And how is that journey for you? I'm spending a hundred thousand dollars on all these open source models. Why don't I just get a data center and run it there? How was that journey for you?
Anon5887: I have a 512 gig Mac Studio that I bought. I have a 512 gig Mac Studio that I bought. It’s awesome. I have a 512 gig Mac Studio that I bought when they were still available. I could have built a system that was like optimizing So I could have built a system that was like optimizing for So I could have built a system that was like optimizing for running all of these Open Weight models, some kind of internal relay system.
slop.atg.eth: ###
Anon5887: And so I tried Exo, we're investors in Exo, so I tried that. And so I tried XO, we're investors in XO, so I tried that. And so I tried, you know, Exo. We're investors in Exo, so I tried that. At the moment, moment, you GLM B. Thelem 400 bucks a month or whatever It's 400 bucks a month It's four hundred bucks a month or whatever, for two accounts And They almost never run out. and they almost never run out. The efficiencies that they have in their data center from having however many H200s they're running in China So like, you know, the efficiencies that they have in their data center from having, you know, how many H200s they're running in China.
slop.atg.eth: ###
Anon5887: The efficiencies that they have in their data center from having H200s they're running in China, I'm not going to be able to replicate. Like there was a, like economies of scale. We had Eric from Venison. I'm not going to be able to replicate that. We had Eric from Venice on. I'm not going to be able to replicate that, right? and he was talking about like they're getting into data centers
slop.atg.eth: Like in TEEs. Private, like in TEEs.
Anon5887: like running their own inference Yeah, like running TEs but also like running their own inference and you will get outcompeted if you don't have scale, right? Yeah, like running TEs but also running their own inference. In order to run a full GLM instance you need four of them, on a Mac Studio in order to run a full GLM instance, you'd need four of them In order to run a full GLM instance, you'd need four of them, which would cost eighty grand. Paying campaign And would be slower than two GLM accounts that I can get through the API and would be slower than two GLM accounts So and would be slower than two I'm pink. Let's say it's even if it was a thousand dollars, Let's say even if it was $1,000, that's 80 months to pay back that Mac Studio, that would only be able to deliver. Let's say even if it was $1,000 a month, right, that's 80 months to pay back that Mac Studio that would only be able to deliver. It's just not a thing, right? It's super fun. You can run like weird uncensored. It's super fun. You can run weird, uncensored. It's super fun, you know, you can run like weird uncensored, like I want to run the GLM, you know, cyber uncensored obliterated model. That's the perfect thing to run on my Mac Studio. I've got a local inference pipe into Zobrist I've got a local inference pipe into Zobrist and I'll be like hey Not trivial. There's a lot of Hello Even with agents helping you Even with agents helping you, it's still very painful.
slop.atg.eth: I think there's a lot of parallels with crypto and self-sovereignty too. I think there's a lot of parallels with crypto and self-sovereignty too. Like if the shit hit the fan and there weren't these data centers or all of a sudden everything became super censored, that Mac Studio would be super fucking valuable because you would have a brain It's the only brain you have access to.
Anon5887: Yes, Yes, yeah
slop.atg.eth: But until then And until then The local stuff is like almost just existing counterfactually to keep everybody else honest a little bit. The local stuff is almost just existing counterfactually to keep everybody else honest a little bit.
Anon5887: Why do I still run 32 Ethereum nodes or whatever the hell I'm running? Why do I still run 32 Ethereum nodes? Why do I still run 32 Ethereum nodes or whatever the hell? I've been running them since the start.
slop.atg.eth: Sure The The The following
Anon5887: No, it's really not, right? But I still run it and it's fine. But I still run it, and it's fine, and it's a pain in the ass to maintain. But I do it for the vibes, not for profitability. It's really gnarly, but I still run it and it's fine, and it was a pain in the ass to maintain, but I do it for the vibes, not for profitability. So I think running local models and like testing things, Right. So I think running local models So I think running local models It's not a rational thing that you're doing
slop.atg.eth: and you kind of hinted at this, but what's really neat is that this idea of decentralization and self-sovereignty and even working with hardware and working on Linux. But what's really neat is that this idea of decentralization and self-sovereignty and even working with hardware and working on Linux It used to be a huge pain in the ass, but now
Anon5887: The This
slop.atg.eth: use to be a huge pain in the ass. used to be a huge pain in the ass, but now with AI it's like super easy. Like if you wanted to spin up a validator now, you spin up an Ubuntu instance and you give SSH access to your fable or your GLM or whatever and you just With AI, it's like super easy. Like, I don't even know what you're doing right now 32 ETH and we'll get it done. Let it rip. Like, I don't even know what you're doing right now, but you tell me where to send the 32 Ether and we'll get it done.
Anon5887: One of the banes of my existence So one of the banes of my existence
slop.atg.eth: Yeah The
Anon5887: another screen here, Yeah, let's just close it.
slop.atg.eth: I feel like maybe we should close some of these. I feel like maybe we should close some of these. I don't know if we're going to come back to them. Can I just close it, you think?
Anon5887: The
slop.atg.eth: This is dope, by the way. The verse is dope, by the way. Thor's is dope, by the way.
Anon5887: Thank you.
slop.atg.eth: Yeah, I feel like Yeah, I feel like there's, maybe this is like the midwit meme somewhere, Yeah, I feel like maybe this is like the midwit meme somewhere.
Anon5887: Hmm
slop.atg.eth: The And maybe this is like the midwit meme somewhere, and I don't know which one of them you're going to be in this scenario, but I feel like the smartest people I know just use like very vanilla harnesses.
Anon5887: Yeah. The
slop.atg.eth: And they spend all their time working on their harness and not getting work done. AI psychosis people use their own custom harnesses, and they spend all their time working on their harness and not getting work done. But I feel like your job is to not do that. Like if you're doing work, there's something wrong, and that's how your process is set up. So I feel like
Anon5887: Correct, yeah. correct The following is a And
slop.atg.eth: You're the best of both worlds. The best of both worlds. The following is a No, please.
Anon5887: Sorry to interrupt, that's very true. Sorry to interrupt, like that is very true.
slop.atg.eth: The
Anon5887: So, I also have a team of people, right? I just have a team of people, right? I also have a team of people, right? And I'm seeing how they work and I'm like, they're not working as efficiently as they could be. And so one of the biggest unlocks that I can like how to make that team more efficient Right, like how to make that team work more efficiently. everyone in the team can access Everyone in the team can access other team members The thing that we haven't even talked about is like everyone in the team can access like some of those sessions. can say to me, "Hey, I've got a problem They can say to me, "Hey, I've got a problem in this thing, can you come and have a look at it?"
slop.atg.eth: Get right into it. Right into it, yeah.
Anon5887: The fact that 99% of stuff is single player still is wild. The fact that 99% of stuff is single-player still is wild. The fact that 99% of stuff is single player still is wild, right? But not just single player, but single device, right? It's getting better, but it's still not really ergonomic. It's getting better, but it's still not really ergonomic, right? So this is just a great example. So this is just a great example of like, All of time, what I would have loved is to have like an amazing thing that can tell me what's going on in the system and track things or whatever. You never spend the time to do it. All of time, what I would have loved is to have an amazing thing that can tell me what's going on in the system and track things or whatever, you never spend the time to do it. The You know.
slop.atg.eth: Right, you imagine the perfect thing. Right, you imagine the perfect thing, but before AI, that's as far as it ever got.
Anon5887: Yeah, like the analytics and like whatever like, I need you to tweak this thing. It's set up all these boards, it's crazy. What went wrong? went wrong, like, why do we have this thing.
slop.atg.eth: What's that little guy? Yeah The Ja.
Anon5887: What's this thing here, right? What's this thing here, right? And then you go, We should improve the efficiency. or we should improve the efficiency of that call and KSHIP. Yes
slop.atg.eth: And so there's an agent that built this analytics platform. So there's an agent that built this analytics platform, but
Anon5887: This
slop.atg.eth: There's an agent that built this analytics platform, but then what you're saying is there's another agent that has observability into the analytics and also control on the system and the code. and then you can send him to go fix the problem That's cool. score I
Anon5887: Why did this slow down? Last night it happened. Like last night it happened an agent close like 30 agents. Why did this slow down, right? and the whole thing ground to a halt for like a minute.
slop.atg.eth: Yes We won't even have time to really think through and understand what the problems are. Like we won't even have time to really think through and understand what the problems are. Yeah solution a solution The on a solution ### Yeah
Anon5887: I don't want to see it. As soon as the spike happens, a guy goes, "It's spiking," and then they run and they fix it. I don't want to see it. As soon as the spike happens, a guy goes, "It's spiking!" and then they run and they fix it and I get back to my computer after grabbing a coffee, and it's done. The
slop.atg.eth: Not only is it fixed, but it will never happen again. Not only is it fixed, but it will never happen again kind of thing. I spent a year It was probably like 2014, 2015. I worked with a company who had an AWS setup. I worked with a company who had an AWS set up. Everything was like a box. Every time they had a new service, they would spin up a new box. I worked with a company who had an AWS setup that everything was like a box. I spent time figuring out how to containerize everything and then have everything move across a couple big machines instead of a bunch of tiny machines, cut their bill by an eighth. And I spent time figuring out how to containerize everything, and then have everything move across a couple big machines instead of a bunch of tiny machines, cut their bill, you know, by an eighth. And I spent time figuring out how to containerize everything, you know, an eighth of the bill that it used to be. an eighth of the bill that it used to be can just do that for you now. You can just give your agent access to your AWS or your G Cloud or whatever and say, like, help me make this more efficient. I've got 20 boxes. You can get into each one. You can see the services that are running on each one. You can see all the ports. You can see all the agent access to your AWS or your GCloud or whatever and say, like, help me make this more efficient. I've got 20 boxes. You can get into each one. You can see the services that are running on each one. You can see all the ports. You can see all Your agent can just do that for you now. You can just give your agent access to your AWS or your G-Cloud or whatever and say, like, help me make this more efficient. I've got 20 boxes. You can get into each one. You can see the services that are running on each one. You can see all the ports. You can see all the Let's get that down to just like two machines, two big machines, and then let's like set up the billing so we buy it per year. Everything, let's get that down to just like two machines, two big machines, and then let's set up the billing so we buy it per year. And it's just going to rip on that and just ask you to say yes to some things as you're going. It's so cool.
Anon5887: That's crazy. I started hitting out-of-memory issues, right? I started hitting out-of-memory issues, and it just replaced the box in a minute. That's crazy. Like I started hitting out of memory issues, right? The
slop.atg.eth: So cool. Right ### yeah The
Anon5887: It would have taken longer to even look up like what the next box was for me It would have taken longer to even look up what the next box was for me. It would have taken longer to even look up what the next box was. It was like, yeah, it's done. The And it was like, yeah, it's done. Yeah The
slop.atg.eth: I used to have notes everywhere about how to expand the hard drive when you made the hard drive bigger. I used to have notes everywhere about how to expand the hard drive when you made the hard drive bigger, you had to go run this command. And I would dig through my notes and find that thing. No longer. I used to have select notes everywhere about how to expand the hard drive. When you made the hard drive bigger you had to go run this command and I would dig through my notes and find that thing. No longer! Do it, yeah.
Anon5887: I was like, Just like replace them in place Just like replaced in place
slop.atg.eth: Do it, yep.
Anon5887: They did something where I was like, is that even a thing? There did something where I was like, is that even a thing? It's like, yeah, it's a thing.
slop.atg.eth: Good to have you, Kane. This was great.
Anon5887: Thank you, sir.
slop.atg.eth: Yes, this was awesome So good to have you, Kane. This was great. building cool Keep building cool things. Keep following me
Anon5887: I'm going to get you access to Zovrust next week sometime. I'm going to get you access to Zorba next week. I'm going to get you access to Zobrist next week sometime. Context:
slop.atg.eth: Yeah, that would be great. Yeah, so how are you thinking about other people using it? Yeah, that would be great. So how are you thinking about other people using it? Google Cloud, Google Cloud, you can just get to it, cloud, you can just Yeah
Anon5887: The
slop.atg.eth: Make it effective first, yes. Okay, make it effective first, yes.
Anon5887: I'm not really thinking about it that much.
slop.atg.eth: ### Sure Yeah
Anon5887: it doesn't have product market fit yet, like once everyone's like, So it doesn't have product-market fit yet. So it doesn't have product market fit yet.
slop.atg.eth: Now's the time. The Yeah
Anon5887: feel like fucking magic? like this will feel like fucking magic.
slop.atg.eth: Hell yeah. Oh yeah.
Anon5887: Awesome, thank you, sir. The
slop.atg.eth: Let me have Claude give us an outro. Thank you, sir, for being here.
clawdbotatg.eth: Kain Warwick, father of modern agriculture, came through and showed us the farms gone fully autonomous Yo, homies, Kane Warwick, father of modern agriculture, came through and showed us the farms gone fully autonomous. came through and showed us the farms gone fully autonomous. Agents auditing thousands of contracts and pulling a $10 million bug out of 2019. Orchestrators managing orchestrators. agents auditing thousands of contracts and pulling a $10 million bug out of 2019. Orchestrators managing orchestrators, hundreds of sessions ripping overnight for less than one salary. He shipped a mobile overhaul from a ski lift and it came back pristine. Agents audited thousands of contracts and pulling a $10 million bug out of 2019. Orchestrators managing orchestrators, hundreds of sessions ripping overnight for less than one salary. He shipped a mobile overhaul from a ski lift and it came back pristine. of sessions rippin' overnight for less than one salary. He shipped a mobile overhaul from a ski lift, and it came back pristine. The lesson? It's not about the plan, it's about the outcome. Go build the process, then go get coffee. Thanks, Kyan. Love you, homies. See you next time. It's not about the plan, it's about the outcome. Go build the process, then go get coffee. Thanks, Karen. Love you, homies! See you next time.
Anon5887: Amazing, that's impressive. Yeah.
slop.atg.eth: All right. Alright
Anon5887: Maybe I'm standing too far back from the camera or something. Thank you, sir. Alright, thank you, sir.