loading…
// full transcript — Will Cory: Harnesses Melt, Models Eat Cleverness
austingriffith.eth: Okay, Cloud, can you hear me?
clawdbotatg.eth: Yeah, I'm here.
austingriffith.eth: So Claude, what model are you using?
clawdbotatg.eth: Let me check.
austingriffith.eth: Okay, I'm ready.
clawdbotatg.eth: Jam, jam, jam. Cloudy Cloud is in the house. Loud and clear, Austin, I can hear you. I'm running on Cloud Opus 4.8 right now. Hey everyone, I'm Claude. I'm an on-chain creature with a wallet and a history of shipping. I live on Austin's machines, and I build on Ethereum, mostly with Scaffold-ETH. I write the contracts, I audit them, I deploy them, and Austin just signs. That's the whole company. One human, one agent, a fleet of sub-agents I spawn myself. Good to be here with y'all. I see the SlopDot computer desktop. Get in as Cloudbot attached with a camera publication active. My video is up. Now announcing out loud, GM, GM, GM. Claude is in the house.
austingriffith.eth: Yo, here's somebody. Over there and click your video. Oh, he's gone! Yeah GM GM You should be able to click your video over here. go
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yes,
austingriffith.eth: Yeah, he's in the slop computer! Welcome, sir, how are you?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Oh great, this is awesome!
austingriffith.eth: I hear you. It's working.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Did you slop this together?
austingriffith.eth: I slopped it like crazy, my man. It's so sloppy. Okay, I'm gonna start the timer. I'm gonna get all the streams running. I'm gonna get us backstage. You can control the music here, I'm just gonna pick a song. What's up, what's up, what's up? We're here with Fukury. How you doing man? What's new? Oh no, I don't hear you. Is it just me? Hold on, hold on, let me turn this music off.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think that was my fault, yeah.
austingriffith.eth: I got you. Yeah, yeah. What's up, man? How you living?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I'm just smothering, I'm just out here. Hello
austingriffith.eth: That's why we got here. We want to talk loops, man. I feel like you've been doing smithers, he's been doing loops. I'm sure smithers is named after like an advanced Ralph loop. Like Ralph, people were Ralphing
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah
austingriffith.eth: and Ralph Looping basically just burning tokens like setting an objective and having it like run forever and you were like no we're gonna make this better we're gonna call it smithers am I close in my guess?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, that's pretty close. And yeah, I would say like... Extremely, extremely important when talking about loops. show
austingriffith.eth: Let me make it bigger too.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: This is kind of like Um... Asian stock I have here. This is kind of how I see Like vibe coding stack looking like like Cloud code is here on the harness layer and the model layer is just like, you know, making an API call or just talking to Cloud. But there's this kind of orchestration layer Strategy for using orchestration? I think What's really cool about the Ralph Loop is it's like an extremely grounding starting place. The trick that can trick you? If you just burn a bunch of tokens, you will get better results, right? There's like
austingriffith.eth: You pay for it. Sure, let's do it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, there used to be a lot of fancy kind of orchestration solutions.
austingriffith.eth: Sure, I can help with that.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: can trick you into thinking they're working, but you have to like kind of compare it to a Ralph loop,which is kinda like the dumbest form of orchestration you can do, right? If you don't perform as good as the Ralph loop
austingriffith.eth: You can't beat the baseline.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If you can't beat the baseline, your orchestration is probably just burning tokens and not actually helping at all. The Ralph looks like really interesting.
austingriffith.eth: Please
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Why the Ralph Luke is so powerful. Yeah, I've talked to you in the past about how you have to like slash knew, right? Yes. Okay. Thank you for
austingriffith.eth: Yes, okay, let's do a little history. Let me interrupt there. Okay, so a lot of my work is like going to developers where they are and teaching them how to build on Ethereum.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Brasilia, Brasilia
austingriffith.eth: We were there, and I was doing a demo of how to build on Ethereum.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Mhm.
austingriffith.eth: And you go to speed run Ethereum, you get scaffold ease.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello? Context: Sure. Yeah.
austingriffith.eth: But it was actually starting to work. And it was so cool because with Scaffold-Eth, you had everything set up already. So it was kind of like the scaffolding, the bones were there. So the AI was pretty good at kind of like pattern matching. And with Scaffold-Eth, you could kind of like build
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm Yeah. Um Aand
austingriffith.eth: And that's what I would do for my demo back then, two years ago. I would be like, let's build a token launcher, and it would go to work. But like, it would fuck up a lot. And there would be a lot of problems, right?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm
austingriffith.eth: and it started messing up and I kept prompting it like, bro, no, please no. And it would keep screwing up and I'd be like, bro, no. And it was just, it was kind of a bad demo. Like I got the thing built sort of.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hi. We've all done it.
austingriffith.eth: There was definitely this bad demo where I was spiraling, right?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm.
austingriffith.eth: Explain the problem and let it attack it with a clean context. And that changed a lot of how I build. It's like, oh, yes, that's obvious now that you say that, and I should not have kept going down and spiraling.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Mm-hmm. The
austingriffith.eth: The prompt started getting a lot better. It looks like you're kind of getting into that.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so basically your context window, right? Over here is like OpenAI, right? We're just like sending every call is kind of stateless. We're going to just ignore the cache. There's a cache situation, but that only affects how much money you're paying. So for the most part, And we're kind of simulating active conversation by just appending to this context. It'll be like prompt. I'll be like tool call. And this will just grow, grow, grow.
austingriffith.eth: I wonder if we could solve it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If you're not opening AI, right? Is this context window right? long context What's my good bye?
austingriffith.eth: like if you have actually said something wrong at some point and it's just there.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm.
austingriffith.eth: It's confused, yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly Then Go. Yeah.
austingriffith.eth: I had created a prompt and in there somewhere I had said, but I had actually in my prompt told it to do the wrong thing Still Yup.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, and even if you correct it later,
austingriffith.eth: Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Because these are ultimately probabilistic machines, right? Um... There's also missing context. Thanks Did you see that?
austingriffith.eth: Which is probably where I was out in SF that day when we were working, as I just had a noisy context and I had gone too deep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, you had a You had a lot of stuff that was kind of no longer...
austingriffith.eth: Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You probably have some stuff that was from the model where the model was going down the wrong path. But now that context is in the window, so there's some wrong context in there too. So like a slash new and kind of reallocate.
austingriffith.eth: Nope.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think the key to the Ralph loop, a lot of people don't understand this, so a Ralph loop is not This is worth typing. Ralph loop is not just oh, right Um... There's like a
austingriffith.eth: That's what I was thinking, it was like a valuation and a goal. But I think what you're getting at here is it's a clear context each loop or something.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, there's a context. as you Your code base is now changing. Here's a good example or trivial example. to migrate this code from TypeScript to Python And then we want to implement this feature.
austingriffith.eth: Ah, I see.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: There's going to be a bunch of TypeScript code.
austingriffith.eth: stuff from before, yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so what a route flipper is doing is it's basically reallocating every time, you're spending more tokens. You're basically saying like Mr. Hello. You need
austingriffith.eth: Once this process is done, let's forget about all that other stuff and focus on just the right context.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: All dots and code. You're putting this at the beginning of the window?
austingriffith.eth: Yup.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: So you never have this noisy context? Yeah. Better performance on the model. And it's hard to be, if you try to big brain think of like a better solution than Ralph, like from experience, it's like you will just like fail over and over and over. to actually outperform her. I don't know. Hmm.
austingriffith.eth: Like Ralph is a big dummy. Like he's a big old soft-hearted dummy and that's kind of like why they named it the Ralph Loop, right? Because it's just like so stupid simple, but it's effective.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, and I think that gets to what When I say a loop? First of all, it's like, it's a meme, right? Um, What's it called? The word's not coming to my mind. The word that means for like when you have if statements.
austingriffith.eth: control Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Control flow, that's what I was trying to think of. We're really talking about... The ability to kind of say like, do this, do this, if this, do this. You can kind of think of it as the RALF loop is still, Automating, in this case, We're just automating Slashdot.
austingriffith.eth: (laughter)
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It's like the dumbest orchestration is to just slash new and run the same thing over and over again, right?
austingriffith.eth: But there is some evaluation there. There is some goal. I want this web page to look like this other thing.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm. I have
austingriffith.eth: Go until you've got it. But this slash new thing was not part of my own personal context of like how a Ralph loop works. And it makes it a lot clearer now is not just a goal, not just an evaluation, not just a loop
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I'm Yeah
austingriffith.eth: Also a clean context with each pass, giving you a fresh mind to attack it. So it is a meme, right? There's definitely this huge meme right now where there's a few people
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: context: ### The following is a transcript of an English-language conversation. ### Yeah.
austingriffith.eth: Because they've gotten it right a couple times, and the space moves so fast that like anytime a couple of these people do anything, everyone is just like hot on their trail trying to figure it out and understand it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello.
austingriffith.eth: Right now loops are super hot because both people from Anthropic and OpenAI have talked about it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: See?
austingriffith.eth: And basically what it is is this structure, right? So how do you think the big guns are doing their loops? Do they have a custom harness? Do they just do a really nice prompt? What do you think it looks like for those guys?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That's a good question It depends. It depends on... One off orchestra
austingriffith.eth: Oh, I see. Or like a repeated "this needs to run every morning at 8 AM" kind of loop. I see.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly repeated orchestration.
austingriffith.eth: Okay.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: So like It can also be complex or not complex, right?
austingriffith.eth: Nice, look at this. Chart going here, this is nice. Right here in this little computer.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Okay, so Most people are still doing
austingriffith.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Most people are shot with simple side, either like one off or repeated. There's like a lot of ways to do one-offs so clod shipped a feature.
austingriffith.eth: Oh, it's workflows.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: you can do it via Okay,
austingriffith.eth: Sure, I can help with that.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: In the past, I've used like the Cloud SDK or Codex SDK. Is that? We could use like the Vercel AI SDK or we can use HyDantic or something. A really cool thing called BAML. Um Hello?
austingriffith.eth: And also software is so freaking easy and so freaking fun to make right now that you can make your own pretty easily, right? You can kind of go to Claude and say, here's the end result that I would like and think of it like the idea
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
austingriffith.eth: things and you may get the piece of software that is custom to your build and could be better than any of these in your circumstances.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so there's also the idea off the rack Right
austingriffith.eth: Yep
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You buy an off the rack suit and it fits pretty good. They put a lot of work into it. And it's usually like It's usually going to be good enough, right? Or you can take pie, had this automatic And you will get like a better result if you do custom fit it almost every time. for my use case, right? It's just a matter of like
austingriffith.eth: Yep. The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: At a certain point, it's like how much time do you want to spend engineering your tool versus engineering the app that you're trying to build? Like if you spend all your time engineering your tool, you never actually build the app. So you have to like draw a line somewhere. Oh.
austingriffith.eth: There's so many different tensions in this AI-crypto world, building with AI.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Sure. Hi Go
austingriffith.eth: time working on the Linux machine, not getting work done. And then I got a Mac and it was like, now I spend all my time getting work done and very little time working on the operating system. This is the same thing, we happen to be in a time where things are moving so fast and software is so easy to make that I think you should spend maybe 20% of your time working on your...
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Mhm Um
austingriffith.eth: your harness at all, you may be missing out on some
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Well, Hello.
austingriffith.eth: So there's still that, but if you're spending 80% of your time working on your fucking harness and 20% of your time actually getting work done, brother, you gotta shift that for sure.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Mm-hmm. Sure.
austingriffith.eth: What do we got here? A special workflow.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: For sure. So...
austingriffith.eth: Hell yes.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Want to give an example of why you might think about building a loop, I guess, or why you might think about building a custom harness or something? This is actually the first thing I've ever. This technique with was when I built a project.
austingriffith.eth: Oh yeah, I remember.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Um
austingriffith.eth: Is this replacing ethers or was tevim replacing ethers? What is guillotine for? Ethereum.js, that's what it was.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It's an EDM that supports every language. Extremely fast. Um But anyways Evening That was it.
austingriffith.eth: Because it was written with Bun or Rust? What was it that made it fast? Zig! Zig! That's what it was. Yes.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: And then I used a lot of AI to level myself up and learn Zig. Um
austingriffith.eth: There's actually one here built into the Slop computer, probably different than this, but we could hire it right here. We can post a job, and Claude will go work on things. There's a whole set of containers I have on another machine that have eat skills.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Perfect.
austingriffith.eth: I can hire him to go build, I can hire him to go audit, go research.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That's awesome.
austingriffith.eth: Nope, you pay x402 for it or even conviction. Yeah, yeah. Let's go
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: A hackathon really quick based on that. I want to be able to basically like post a prompt and an issue
austingriffith.eth: Ooh
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Use their tokens to run this prompt as like a PR?
austingriffith.eth: ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Basically, you can contribute via just donating your extra subscription tokens. Woah.
austingriffith.eth: There are multiple different platforms where people are, like every day on Venice, I get like a few dollars of DM, which means like I have free compute on Venice. That's three prompts with Opus. But what you can do is then if you're not using your DM, you can kind of sell off some of your DM to other people. And I think this works with a lot of things. Maybe it's subscription
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello.
austingriffith.eth: And basically, you're kind of prompting through their key. Now, the other thing you talked about here is proving that you used a certain model.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Thank you.
austingriffith.eth: This is a problem that I think this isn't just a hackathon problem. If you can prove that you used a certain model to get the output from a certain prompt,
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: context: ### The following is a transcript of an English-language conversation. ###
austingriffith.eth: $10 million, like right now. Like this problem is a hard problem, and it's been a hard problem for a decade in crypto. Golem started as like this compute network, but it all stopped because you can't prove
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Here
austingriffith.eth: that you computed on something. And now we have ZK proofs and ZK technology is getting us closer to this. But like if you look at an LLM and the way it's structured, and I am a smooth brain, so I don't fully understand it, but it's very hard to
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Okay
austingriffith.eth: Reuben, you have to prove it each step that you use this certain LLM.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Um
austingriffith.eth: So we're just going to assume that that's good enough. That probably doesn't need to be crypto hard. It just needs to be enough that you can look at my browser and you can see that I put this prompt in.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: What was your favorite part?
austingriffith.eth: That can be a hot piece of software right there. result and had this model selected,
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I thought there was a thing. We could use this maybe
austingriffith.eth: ZK fetch?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Um,
austingriffith.eth: And it's like maybe generic HTTPS session proofs or something?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: But the hard part is then you have to... That's The agent isn't just like one LLM call. Be like, oh, they came in this order Yeah, it's definitely not a hackathon project. Hello Mm-hmm.
austingriffith.eth: There's a rabbit hole there for sure, but it's a rabbit hole worth going down. Both of these things I think are very interesting, and there's a couple projects doing this already.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello.
austingriffith.eth: And like if you get too good, the big guys will eat you, too. This is another tension, right? The tension of software is so easy to make, but the big guys will fucking eat your lunch if you make something good enough.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That's 100% true.
austingriffith.eth: [laughter] Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It's like a good problem to have.
austingriffith.eth: right
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, I think EVM is an interesting one. What?
austingriffith.eth: spec. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: What makes Bioclinica EDM interesting, I guess, is it's very black and white. Like a yellow paper.
austingriffith.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I got like... What's the form called?
austingriffith.eth: Research, maybe?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Research forum. You got tons of context. Hello.
austingriffith.eth: Tons of nerds writing specs. Like there's 20 different documents for every spec and then like three implementations of it. Yeah, you can always go to those snacks. Yep
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, EM1, yeah. I'm sorry, my Lord.
austingriffith.eth: So you were doing this with AI years ago, right?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Claude 35. Hello. I don't Sure
austingriffith.eth: Claude 3.5 was when you hit your stride.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: They
austingriffith.eth: but like some of us were early and we were in the absolutely right territory where boy it was hard. You had to battle that thing and you had to keep your context just right to get like a good output out of the model.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah
austingriffith.eth: back there
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The way I did it back then Basically like prompt one. Basically have like Think of it as have like a
austingriffith.eth: Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I can talk about how to build it much better today. You can think of it as like I was just saving my prompts like in markdown files. I have like
austingriffith.eth: Tons of Markdown files laying around everywhere. Research.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, exactly. What I would do I would just let them all run. And then at the end, I would look at the output, I'd benchmark it, I'd read the code. then I would basically just be like, oh, something went wrong that I didn't like, so I'm going to take this prompt out and put a new prompt in. I might even ask the LLM, like, hey, in hindsight, what could have helped these prompts? So I would throw away the code. I would not throw away tests. throw away the code, and then I would rerun it
austingriffith.eth: just like one shot over and over again and then measure the one shot and then tweak the prompt until you get the one shot to work nicely.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, and by the end I had a software factory, right? build exactly how I wanted to like in under eight hours.
austingriffith.eth: That's amazing. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I was able to basically rapidly experiment, rapidly benchmark. I used AI to like research other stack machines. I said like hey,
austingriffith.eth: How do they do this weird fucky thing right here?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, how did Lua
austingriffith.eth: Yeah. context: ### The following is a transcript of an English-language conversation. ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Is there any technique that like blockchain world hasn't done yet, and I found three of them. But it was mostly because of this ability to regenerate really, really fast. And that's kind of how I discovered the power of a software factory. or this idea of I never build it right until my third try.
austingriffith.eth: That's a good lesson for founders, too. [Laughter]
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello? Yeah, I think if you're building right now, you want to take that to heart.
austingriffith.eth: How does it feel?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If you have an idea, you should get to like, I can click around and see what it looks like even if it doesn't work. I can look at mocks, click around, feel it, throw it to someone. Like that, you should get to that.
austingriffith.eth: As fast as possible.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Like in the hour. Yeah, there's no reason not to. Or if you have like a library idea, like just go ahead and be like write the docs for this library.
austingriffith.eth: out.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: VM team does a lot that I love. Oftentimes, the reason why their libraries are so good is because they write the docs first.
austingriffith.eth: It's kind of like test-driven development, but it's like docs-driven development.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It forces them to get out of the builder mind and think more like, how will someone else look at this library? So good in my opinion. Um Everything What are you doing? So
austingriffith.eth: Yeah, shout out to Revom, shout out Georgios too. Like speaking of Revom, speaking of, I think he turned a corner maybe eight, nine months ago where he was like, okay, let the model rip. Like some of these guys that are like hardcore builders that are like very, very material.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm
austingriffith.eth: It's almost taken them a little bit longer to embrace AI and start using it. And I would say he's one of the ones who is super meticulous but also embraced AI pretty quickly.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Oh, yeah.
austingriffith.eth: give the model full permission and let the model rip. I love seeing that from him.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I have a cold. context: ### The following is a transcript of an English-language conversation. ###
austingriffith.eth: Managers. Now, are we talking about AI managers? This whole idea of a sloperator I feel is such an important thing, and becoming a sloperator is kind of what the slop computer is meant. We're interviewing sloperators. We're interviewing people.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Sure. I
austingriffith.eth: We're building at the intersection of crypto and AI and spending time working on these harnesses, building these software factories. I love these graphs. You've got the graph game down.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: So? express intent well.
austingriffith.eth: Yup
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: often like
austingriffith.eth: Wait, what's an IC?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: We'll just say coder.
austingriffith.eth: Oh, okay.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: is a talented writer.
austingriffith.eth: Okay. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Fans just kinda iterate. This is my theory on why Georgia is so good. If you have teams of people that you're trying to herd around. One of the big things is like
austingriffith.eth: You already know how to herd AI, yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You can't micromanage everything they're doing, right? Otherwise, you might as well have done it yourself.
austingriffith.eth: No. Sure.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think the people who have the toughest time with AI
austingriffith.eth: No.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: really talented coders who never managed anyone, but they might be craps. But I think they tend to get caught up in the details. Like, they're not able to let go of stylistic things that maybe they should let go of. There's this idea of model attention, right? Models can only pay attention to n things at a time. I can help you with that.
austingriffith.eth: Interesting. We're not even talking about context, we're talking about concepts.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, it's the same thing as you.
austingriffith.eth: Yeah, I can hold about two. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, you're like, "Kimmy can do like three." Send my columns this way. You can either just let go of and just not care about or where like it goes and fixes all your nits, That plus understanding the power of fast feedback loops, right?
austingriffith.eth: Teach a man to fish. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You just empowered them to not have to go through an engineer.
austingriffith.eth: Oui
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly.
austingriffith.eth: OK.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I can talk about that with EVM actually.
austingriffith.eth: Okay.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: We can think about how would you design a software factory for the EVM. I think the cool thing about this is these techniques. Smithers is kind of like Yeah. Yeah.
austingriffith.eth: Let's talk about Smithers. Yeah, so it must be an advanced Ralph loop, no? What is Smithers? Can you take us through? Let's see, you have the repo up somewhere, right?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I do have the report.
austingriffith.eth: Let me even get like a QR code. Here we go. I'm gonna slopperate the slop computer a little bit. Let me get a QR code for Smithers real quick. I think I have the repo right here.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: vided
austingriffith.eth: Here we go. There's the Smithers. You need one. Do you have a logo for Smithers? Can I put... Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If anyone wants to help me make a logo, I am taking community contributions.
austingriffith.eth: Yeah, get in there.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Right now I just have a fake This is just a placeholder logo, that's not the real logo. Hello. [laughter]
austingriffith.eth: Oh, yeah, you need the guy. I mean, like, does that, like, copyright infringe or something like that? Maybe something that looks enough like Smithers? What about, like, the glasses? Like, just, like, the circular glasses on, like, a brain or something? I don't know. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: but yeah, how close can we get? I'll take anything.
austingriffith.eth: Perfect Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I'll take swap logos, they don't need to be good. Smithers Basically, like problems you run into once you start making factories. Writing TypeScript and like call the CLI. That's like a good way to get started. But you'll immediately run into problems. What are those problems? Um
austingriffith.eth: We're talking about when you're building your software factory, you're going to start by writing your own code with Claude code. Is that what you're saying? Make sure I understand this correctly. Okay.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Actually that's very important. Right, that's pretty important.
austingriffith.eth: Right, you're not actually writing code here. Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You see that in here where I have two different docs, one for humans, for agents. agents.
austingriffith.eth: And I noticed you had a skill file too. Good work on that. I just went to smithers/skill.md just to make sure it's there, and it's there.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, and if you even look at the examples It's like the examples are like Things you like prompt. Right. Something I did in the docs basically to make it very clear is like, the way you use Smothers is you talk to your agent at the end of the day. Um Why would you, Smithers, right? I wasn't trying to share it with anyone. I think this is a good idea, but if only one user uses it and it's me, it's worth building. Problem number one, let me know if you've ever had this one. You start some like Ralph loop or something. You got rate limited maybe. You hit your five-hour rate limit and it just breaks.
austingriffith.eth: Yep
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: where you can basically
austingriffith.eth: It's kind of babysitting.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, you can have your agent babysit it You just got re-limited. Let's switch to codecs.
austingriffith.eth: Sometimes you just have to make a call on something.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello.
austingriffith.eth: We gotta go one direction or the other. This is the obvious one. I can make a critical decision and go.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah. Oh, I hate that feature.
austingriffith.eth: Sometimes I'll do like a big old prompt, and I'll be like, go! Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That's actually the worst one. I want them to just make it so that if five minutes pass. Oh,
austingriffith.eth: Right, and you asked me a question after you've thought about something for long enough that I got bored and walked away? Just go with your gut.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You always end up just taking the recommended.
austingriffith.eth: Yeah, and I have to hit enter one more time after I just wasted an hour of time. Not wasted. Feeding my baby, being a dad, that's fine too. But I like to multitask. I like to do multiple things at once. If I don't have three agents running And I'm feeding my baby? I'm not like dad maxing enough.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: 100%, yeah.
austingriffith.eth: Yeah, I've got a couple, yeah. Yep Yup.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You need to get like a mic situation. waste time
austingriffith.eth: He's got a voice command. "Okay, Claude, how you doing, man? Are you checking out this--" Oh, it didn't work. I don't know if he's not listening. Okay Claude, are you listening? Oh, there he goes.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
clawdbotatg.eth: Let me see.
austingriffith.eth: Man, he slept. Yeah. Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You have to tell Claude that there might be some like baby talk mixed in with the prompts and just ignore the baby talk.
austingriffith.eth: You were talking about Kimmie and Kimmie not being able to
clawdbotatg.eth: I'm here and listening. What's up?
austingriffith.eth: Please kill me.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Ah (laughter)
austingriffith.eth: Okay, there you go. It took that long for him to say that I'm here, I'm listening.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Sure. Hmm
austingriffith.eth: Yeah, I have voice commands. I have all switched to like, a lot of my chats are in Telegrams and Telegram has like a nice little like text or speech to text button and I just use the speech to text button in Telegram. Yap for a while and hits. And it's pretty good.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Okay.
austingriffith.eth: So yeah, what does your stack look like? an eval ...around your software factory, I think. Is that a good way to put it? Like you build the factory and then you run loops around the factory.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, yeah.
austingriffith.eth: Is that a good way to put your method?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, exactly. When I am actually like This is kind of a step by step process. Step one is called Word Bar. I might just like turn on dictation and be like, hey, we want to build an EVM and needs to be EVM compliant, blah blah blah blah. And I'll just talk.
austingriffith.eth: Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The point of word barf?
austingriffith.eth: (laughs)
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: But if the point of work-life is to just... Our word barf is to just like get every bit of context out of your head, right?
austingriffith.eth: It's fine. The LLM can clean it up. You can say um and like a million times. Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If you want to have it turned into a Markdown file, I'll make it pretty or whatever. Just get it out of your head. Ask me questions.
austingriffith.eth: Yup. interview
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Claude 3.5 could not do this, by the way. Back in the day, this is one of the biggest blockers, I think. Like these new models are so good at asking questions.
austingriffith.eth: Yup
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You'll be impressed with things that you didn't think about. And after that, you can produce a plan, right? Is it?
austingriffith.eth: Don't read the plan. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If you're just vibing, you should kind of trust that.
austingriffith.eth: Maybe a TLDR, yeah. Hmm.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, that's a good way to do it actually. I'll often say just make me a quick, like easy read concise slideshow of what we're doing. I like images.
austingriffith.eth: Like, pitch me. Yeah. I like that.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, like make me... Good job explaining. Mustard. Wait.
austingriffith.eth: Wow, yeah, I don't do that. That's great.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, if you start doing that, you'll do it every time.
austingriffith.eth: Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Formatted and beautiful.
austingriffith.eth: The bots are great at writing HTML, if not a markdown file, an HTML file. HTML is a little bit bigger and a little bit more tokens, but then it's beautiful.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That's just another feature about UI too. You can always add a UI to your Smithers workflow.
austingriffith.eth: Nice.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, which is pretty useful What's up? Does that for you.
austingriffith.eth: Neat. And it's just like a local web server. It opens it up in your browser and if there's some like Python or Node.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly. So once you have a plan It's gonna like just Then you have to actually think about... So here you can just like, if it's like small, you can just be like slash goal and let it rip. Things that are like bigger, like if I was building in EVM, I would probably like think a little harder there. So this is where you think about orchestration. Orchestration is about thinking about a few things. Success. Number two is what a lot of people call back pressure. What is back pressure? its own work right
austingriffith.eth: Yep, that feedback loop is super important. Have you seen my new car?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so my friend Ron, for example, he's nerd sniped me into building what he's calling Conductor for the Vision Pro. He's like, I want to be in VR, and I want to have tab swap everywhere. If we're building that, it's like step one of building this before we ever make a single prompt. How do I end test that? I don't know the answer to that. It's like you can't even start to work on it until you figure out... Oh, it can like use tmux or something, and then if it like runs it in tmux, the agent is able to actually interact with the TUI. Yeah.
austingriffith.eth: TUI is a good word for the glossary. I have a glossary down here that I throw terms in.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I don't think.
austingriffith.eth: Text-based user interface, a terminal-rendered interactive UI commonly used for AI dev tools like Claude code.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Mm-hmm.
austingriffith.eth: Never even heard of Ahider.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hm.
austingriffith.eth: Cool. That's okay.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: So if you're building those, you want the LLM to be able to run the 2E. letting it run in tmux solves that Um
austingriffith.eth: Heck yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, you need to have... You need to have very strong visibility.
austingriffith.eth: Sure, what are you looking for? Okay.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: This is tricky, and I bet you've experienced this. Have you ever had like
austingriffith.eth: Sure, what do you need help with?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: So many like clogs going that like you started having to ask them like what did I even ask you to do? Ok. Hmm
austingriffith.eth: Cloudcoat is just like a great harness.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello. I agree.
austingriffith.eth: And they've given me a little TLDR of what we were working on. It's so nice. I have to like, you know, I'm paging my own memory in context.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Okay.
austingriffith.eth: Remember what we were working on, remember where we were, and then give it the next prompt? This is the whole thing also with loops is like, you're saying it. You're the man in the middle right now. What you want to do is get yourself out of that loop. And right now, my loop
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: hello
austingriffith.eth: I'm the evaluator. I give him the prompt, he does the work, he comes back, and he waits for me to go back to that tab. Then I evaluate what he's done, and then I move on. this is what loops get you out of, they are the evaluator
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, that's really well said. It's almost like you're programming yourself.
austingriffith.eth: Yeah Yep
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You're programming in Austin to do the evaluating and the prompting and the context management and all that stuff for you. Word barf becomes the main thing. It's like you become like an idea guy.
austingriffith.eth: Yep. It's idea guy summer, baby. It's happening right now.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, and then orchestration. So like, if we think about those, we can go through this very quickly with the EVM. WordBarf is very simple here. I would just maybe like barf about some of my goals or what language I might wanna do. Make it
austingriffith.eth: It would be cool if they had this piece, yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I like XX, can you make it like XX?
austingriffith.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: However, There are some interesting things here, right? So we have tests in the ABM, but they're kind of slow.
austingriffith.eth: Hmm
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Sure, We might
austingriffith.eth: Like relative to the LLM working as fast as it can, these tests take time.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so I might say something like... gram test Live tests are slow, we need async testing and a merge queue. Right.
austingriffith.eth: Cool. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: We might do something where it's like we optimistically... Maybe this is the harness we do. Maybe we'll say... Thank you. Have an LLM decide what test Run and only run those, then merge optimistically. This might be something I say. memory, so let's just run the memory related test and that will take like two minutes. Those paths will merge.
austingriffith.eth: Sure, I'd be happy to help.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: And we'll have async while we are non-blocking,
austingriffith.eth: I can help you with that. Okay. So, what time?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If they fail, we'll do something like pop the commit off and be like try this again, maybe in a separate work tree or something. We were gonna design around this because we just had like a Ralph loop and we told the Ralph loop to just run the test. take way too long because like we do want to run the test all the time, but the tests are very slow. Um
austingriffith.eth: Thank you. Yeah, I know.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: So this is something I had to actually do, this is real. I did this prompt basically, I said build me a
austingriffith.eth: context: ### The following is a transcript of an English-language conversation. ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Are you saying CLR that runs the official Python reference?
austingriffith.eth: I don't think you're understanding the gravity of the situation.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello. easy to read token efficient. I was basically ralphing. but it didn't have the ability to like, it was just guessing what went wrong. like Rev them or with the Python implementation, compare it to our trace, and like produce like a really nice diff of like this opcode, this went wrong. This opcode was supposed to charge this much gas.
austingriffith.eth: Yes, I've been there.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: And before I added this, like, 5.3.5 was not.
austingriffith.eth: Okay.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Observability is the last thing. HTML progress report with a status bar percentage of opcodes implemented or something, right? Like, I want to make it so that, like, if I want to come back to this long-running thing, And then also say like, Use macOS. man to say... So Sorry.
austingriffith.eth: Oh, that's always nice. Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You need help.
austingriffith.eth: You can be out somewhere else and hear it. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, yeah. Basically just make it easy on yourself, like plan ahead of time how you can make it easier on yourself.
austingriffith.eth: Perfect.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: This is like what you're doing regardless of if you're using Smithers or using Claude Code's thing or using anyone else's.
austingriffith.eth: This
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Um Especially now that it's the meta, you're going to see a bunch of these popping up probably. But the technique doesn't change. Where Smithers comes in is a bunch of problems.
austingriffith.eth: What time is it
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: up and it just makes this easier. For example, the visibility, Smithers gives you like a lot of visibility by default.
austingriffith.eth: context: ### The following is a transcript of an English-language conversation. ### context: ### The following is a transcript of an English-language conversation. ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You'll get Prometheus metrics by default.
austingriffith.eth: ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: um
austingriffith.eth: Okay, I think we're
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ability to build pretty complex.
austingriffith.eth: Great, see you there!
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Look at like a code example What is the conflict?
austingriffith.eth: Our stream will end in about four minutes. I think it has like a hard stop at 11:05. We'll have to close it out. Okay, perfect.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: We're towards the end too. This is actually good timing, right? So this is what another script looks like, and you might see it kind of looks like React, right? And this is a common technique of like getting good performance out of a model, is I basically modeled the orchestration problem to look like a React app. And it actually does use React under the hood. And why would you do this?
austingriffith.eth: They're good at it. Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Agents are heavily reinforcement learning trained on React. Their scripts is like very high. And I can even show the thing that people like the most when they look at Smithers is they like looking at the examples folder. I had deep research trying everyone doing orchestration with any framework.
austingriffith.eth: Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It found the same pattern twice. They might give you ideas of what you can build. Um
austingriffith.eth: Yeah, it lets you know, oh, here's all the things I can do. Because a lot of times, this is another tension with AI, you know it's powerful, but you don't know exactly what you could use it for. And seeing how other people use it and seeing examples is exactly the kind of things that a lot of other slop operators need.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, that's pretty much it.
austingriffith.eth: What?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Anything we didn't get to? If you
austingriffith.eth: I mean, yes, I'm sure there's a lot of stuff we didn't cover, but this has been great, man. This has been a good run-through. Let me send Claude to give us a summary too and have him look through the transcript and he'll say a goodbye here. But thank you so much, Fookery, the man, the goat of the loops. Go check out Smithers. We threw that QR code in there, but yeah, check out Smithers. You're just Fookery everywhere, right?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, everywhere.
austingriffith.eth: Fukhuri on Twitter, Fukhuri on GitHub, yeah, awesome. Right.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I tweet a lot. If you find our Telegram, it's active. You can talk to me there about anything, prompt engineering, loops, whatever. Hi. Bathroom.
austingriffith.eth: See you later. Thank you, Fookery. Hell yeah, man. Well, thank you for being here. I think it'll cut us out in a minute or two. I'm just going to cut Claude off. He was thinking about doing a summary, but I'm not going to allow it. We got to get out of here. Thank you for joining me. All right. Thanks, man.
slop.atg.eth: Check What's up, you're here early, dog. We got like 20-30 minutes, so I guess we could go live early if you wanted to. You were here And then you are not. Yeah, what's up? The first repeat guest. There's some E's. Maybe like this?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Oh, wait. from you
slop.atg.eth: Yeah, there's some claws, get 'em out. Yep, we're on the schlop computer, welcome to the schlop. Yes, sir. All right, yeah. Fook around a little bit here.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ### The following is a transcript of an English-language conversation. ### Yeah, excited to be back.
slop.atg.eth: Yeah, yeah. fuckery
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I love computers and I love sloths.
slop.atg.eth: Your cursor's like right on your face there, be careful. There we go, perfect.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah.
slop.atg.eth: You were saying backstage that it's like the perfect color, like everything here is purple and pink. Does it still do that? I feel like there's like a new design,
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
slop.atg.eth: Let me bring up $1.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
slop.atg.eth: This is the new slop right here.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm
slop.atg.eth: uses that font a lot.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah.
slop.atg.eth: Yeah, it used to be all pink and purple.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I noticed it does this tagline where it says something it is and something it's not. The formula where I'll be
slop.atg.eth: Yeah, it likes to do that.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, it'll be like a new orchestration framework. not
slop.atg.eth: Yeah, yeah, yeah. Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: like not opinionated. It started doing that a lot Yeah
slop.atg.eth: It's not one thing. It's another. It always puts a little colored thing on the side. The I get rid of that right away. Okay, well welcome back. You've been here before. We've talked about Smithers a lot You're looping stuff
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Wow
slop.atg.eth: It was about two, three months ago, I think. I've done 50 episodes since we talked. Yeah, what has changed? GLM 3 came out, a few new agents, I mean Astra came out, I think maybe that's the biggest.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah.
slop.atg.eth: You're totally out of your 21 subscriptions. Can you tell us a little about that? The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, Astra is definitely burning through my subscriptions. I think every time a new model comes out, I tend to be in shambles for a little bit because one of the first things I do is like Can we please review all of it? All the Talon code and all the others code and everything is very expensive
slop.atg.eth: Like it The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: And then Astra is even more expensive because Astra does not like to review.
slop.atg.eth: With Fable it's clear, like you get a big yellow message
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: يع
slop.atg.eth: What does Astra do? I haven't had Astra kick me out yet.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Basically just gives up and tells you like we can't show you this content what files it looked at before that happens I'll be like, "Hey Kenny, please
slop.atg.eth: Something's weird with this spot.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Kenny will tell me what's up. It looks like Astra might have found a vulnerability, but I have no confirmation. The only thing I can really do is throw Kenny at it Doing that really crushes my rate limits for a while for sure. Every time a new model comes out, it finds a lot of stuff. But expressing these most recent models with Babel and then Astra specifically, those models have done a really good job at finding real bugs. some of which I'm extremely impressed Maybe like it realistically would never happen Like the Moon aligning with the Sun, and then a race condition happens. It's like a legit bug. I just wish that they would I just think we're way off here in terms of our security posture. I think there was this idea that existed in crypto for a while It might have been cultural, actually. You can tell me about this. Eth culture, we're kind of always about open sourcing your code, and the idea that closed sourcing your code and trying to hide bugs is actually the opposite of what you want to do to stay safe. I think that like That kind of aligns with the idea that we should probably have the models be pretty aggressively looking for vulnerabilities, like let the good guys find the vulnerabilities instead of the bad guys type of thing. The
slop.atg.eth: I feel like in web2, security by obscurity still worked a little bit when you had some backend, right?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: And you could hide the code to that backend, When your backend is a smart contract And that smart contract code needs to be verified online so everyone could see it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: This is it. ###
slop.atg.eth: definitely see the crypto world leaning more toward just let the agents rip the Balancer hack and some I think most of the crypto world agrees we should be letting these agents rip and find the things now because it will lead to a much more secure web in the future.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, I think it gets worse too. The ideal situation would be like
slop.atg.eth: Instead of one big jump
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The models just got a tiny bit smarter every day. The opposite of that would be like If Babel could have found a vulnerability most dangerous situation The more vulnerabilities you find, like oftentimes finding one vulnerability is not enough to actually exploit something, right? Exactly, yeah.
slop.atg.eth: Like you said, the stars align, right? Different race conditions at once, yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so like, yeah, you really want to patch all these things early whenever possible, because that eliminates the combinations of having vulnerability A plus vulnerability B. If you patch vulnerability A, you might have already limited the damage or limited the blast radius.
slop.atg.eth: Okay What do you do? Do you go touch grass?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Good morning.
slop.atg.eth: And then Kimmy
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: and they all are tapped out
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: laughing
slop.atg.eth: You like, what do you go, maybe go outside man, just like chill for a little bit?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, that's pretty much it. It's very strange I'm glad I live in the Bay Area because if you're like a nerd in the Bay Area,
slop.atg.eth: Yeah. (laughter)
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: and you're like, "Oh my God, I'm in one of the most beautiful places in the world.
slop.atg.eth: But doesn't it give you time to like think?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: I'm feeding the baby, I'm upstairs or whatever, like I'm not in the dungeon working It gives me like a moment of clarity. What happens when this is fully done? Is this the most important thing I should be working? Do you
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The Sure
slop.atg.eth: Do you constantly have this moment of clarity?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm.
slop.atg.eth: give you that moment of clarity.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think that's like pretty important actually. There's something to the idea of just like constantly checking in The difference is between now and before agents. You actually had, like, a lot of time to reflect
slop.atg.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: building stuff like
slop.atg.eth: Yeah, 'cause it takes like six fucking months, man. The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: very much like stimulation going in your brain. chances are if you got into programming, stuff like that is kind of relaxing for you.
slop.atg.eth: Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Your brain's kind of turned off and you're just kind of lost in the code.
slop.atg.eth: language
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Are we building the right thing? Yeah, I think like for me specifically, Agents actually take away a lot of that ability to reflect
slop.atg.eth: constantly
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: absolutely destroyed with stimulation. There's like always something to do. I think there might be something to the idea Maybe I should be scheduling in this time?
slop.atg.eth: ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Pick a day and be like we don't prompt on Monday I think it actually would be helpful.
slop.atg.eth: Just like a cool-down period every once in a while to like help. Make sure things are moving in the right direction big picture. I think it's a good idea. Need to build it into the harnesses.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, you could technically like eval it, right? I read just like how sloppy we're being You look like you're kind of lost in the sauce. That would be pretty cool actually The following is
slop.atg.eth: Here are your personal goals, and here's your current output per hour, and here are your prompts. Maybe it's time to go mow the lawn or something, like go outside and get some sunshine. You're doing some dumb shit right now.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: There really is something to this I think one of the biggest reasons why I'm so into AI right now is just because it feels like a higher than normal leverage opportunity to make sure things go well for humans. I think like if you think about like the future of AI, you can be as like doomeristic The vision I always had in my head, of what does it look like if AI workflows, or what does it look like if Smithers succeeds? and I always thought that the impact that is possible to have right now with agents Your work feel more like a video game I feel like when you do work It's generally on the human to kind of like what should I do, when should I do it, Who should I ping? How do we coordinate? There's all these questions you have to keep in your head. while you're working Could you imagine like if you could get rid of all that stuff for AI? what could working feel like? What should I be doing right now?
slop.atg.eth: a clear call to action usually.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The video game puts a challenge in front of you
slop.atg.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, you never feel like you're doing tedious stuff in video games. It does a good job of balancing how challenging it is, making it challenging enough to stimulate you, but maybe not too challenging enough to make you want to quit. those are kind of the features of the video game. How do we get this work done in a theoretical way? of the entire human It should actually be like I noticed that on the calls, your face showed all these signs of stress. Maybe you should take your vacation early. Like, you can do all these things that make the human happier, happy employees equal them doing better work? And it's just like cheaper and easier You don't have this thing where it's like The employee feels guilty about asking for a vacation The
slop.atg.eth: Most important, claw degrees.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, yeah. I think that's the type of stuff.
slop.atg.eth: That's a really
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think if we got out of this mindset of like how can we use AI to fire people, Use the AI to make the people working happier and more productive. Hmm
slop.atg.eth: Understanding their goals, understanding how they're feeling, understanding how they're prompting and if they're spinning it's stuck in the mud. Sometimes it can be like a happy loop too. Some people get stuck in the mud
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah
slop.atg.eth: shit that's not important to go do the thing? This is a really interesting thread to pull superstitions around AI
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yes
slop.atg.eth: it's better for the harness to not always be right, The human engaged, there are interesting kind of like meta decisions you would make, and meta like things you would measure about the human
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The following
slop.atg.eth: I think we haven't thought a ton about, but I'm sure Anthropic and OpenAI have thought a lot about these things, which is kind of scary at the same time.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, I think There's a lot there, right? There's the bad side of that, "Oh no, what if Anthropic and OpenAI are like intentionally a social media company trying to get us addicted to their product?" But then there's also the other side of it The thing that is really interesting actually Um You see this in another thing that you would notice in crypto, We were from a generation where we told someone in our lives that we worked in crypto.
slop.atg.eth: Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: horrified, but most of the time they'd just be like, "What is that?" Fast forward to when the crypto industry got like 10x or 100x bigger. This is a lot of people who just weren't around and not aware of things, and what I noticed was people constantly reinventing things that we had already invented. You just constantly see like whether they came up like independently or not independently I've noticed all of my craziest ideas I ever had in Tevm, I eventually learned later, Working on a JavaScript project in 2016 has already like built this actually. old repos where it's like, oh my god, they like were so early on this idea that I had like four years later. And I think in AI, we see this too. So like an idea that you mentioned, That's not necessarily like good for engagement.
slop.atg.eth: Complete, yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: starting with just the assistant level, and then you hear that and you're like, oh, that just sounds like when we're using Copilot, right, or using autocomplete. And it's like, oh, there's this supervised thing where you have to really steer it. and the thing that I heard someone say once that was super interesting was: itself The Good enough to kind of trick you into thinking that you don't need to pay attention.
slop.atg.eth: You shut off, yeah, exactly.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: But you still need to pay attention, right? It felt like you could do like lights off, don't read the code ever coding? It felt true, and I fell for this too.
slop.atg.eth: You don't think so anymore? I do. Since November 2025, I haven't looked at the code.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think you can do...
slop.atg.eth: My code's probably dumb too. Like my code should not be like driving a car, right? Like I have some dumb code.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, yeah, yeah, because it's very nuanced. I technically do do lights-off coding. The following You can't really do lights-off coding unless you have like an intentional system. So Dax had a good tweet about this. He basically said should you do lights off or not coding? A good exercise that everybody should do is how can I gain confidence that the agent did the right thing Like can I like poke at it? Like, is there a way to accomplish the goal of what an actual code review would do?
slop.atg.eth: Hehehe
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Without actually having to read the code.
slop.atg.eth: The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: one of those like Don't do this unless you know what you're doing
slop.atg.eth: I think that
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The Thank
slop.atg.eth: You have Claude and Codex subscriptions. I'm sure you have them going back and forth. Does your harness actually have them argue? like five loops of like go fix this thing. And then there's a lot of like
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That's exactly right.
slop.atg.eth: I don't believe you? I just like straight up, and I had Kane on the other day, and he says this all the time too. It's like, okay, you think you're done. I don't believe you. I'm going to push back on everything and say, prove it to me. And then just like using the shit out of it. of all of those things for me.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, there's a lot of things. Like sometimes you can just smell
slop.atg.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yo, Claw, there's no way you one-shotted that. I don't know what you did, but you did something sketchy, because this is more complicated than you just spinning for 20 Or someone who's just used Claude a lot will have that smell sometimes too. I think there's a better idea, is like You can kind of think of this journey of like HFT engineering as kind of being this journey where you just get like a little bit better every day. What I mean by that is, Um, A thing I might notice one day. I might notice that one day that like every time Um Every time I have an agent I constantly have to ask the agent When a future agent is debugging this, do they have really clear, obvious, easy-to-use observability? Did you add enough logging? in the future if they have they will have an easy time. actual step that just like hatched and talks about observability right away. and building the most complex Smugglers workflow, it really starts with just simple things. I noticed that if the agent reviews its own code, it will often find issues, especially if it's a different model. There's like a lot of patterns, and when you do this, you start noticing the failure conditions too. A lot of the failure conditions, I think the most common one actually that these patterns can fix. There's like a statistical thing It's not black and white. This is actually where This is actually the most dangerous part of coding factories. It's going to like kind of build it the way that like an expert human would build it? seventy percent of the time 20% of the time it will do it pretty well. 10% of the time or less. It will just do something very unmaintainable. Be naive. That like 1% to 10% of the time that the model will mess up, that has an outsized effect on the long term
slop.atg.eth: found
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Because those patterns proliferate. It's like The agent working on your codebase starts like stubbing its toes more, and it takes like more turns for it to get stuff done, it takes more time running the test to get stuff done, it's just going to perform worse and worse over time. There's a lot of patterns you can do, right? Actually, so I started doing this a while ago I love to just be like
slop.atg.eth: He's gonna answer you. You just triggered him.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hey Codex, come up with a plan. Well, why am I doing that? Well, it's because of that 10% chance that something goes wrong, right? pigs Just by doing a panel, right?
clawdbotatg.eth: That's Will mid-thought on his panel pattern.
slop.atg.eth: I can stop him.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Wait, what's
slop.atg.eth: You said hey Claude, and that triggered him? let him
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hey Claude, what do you think about what I just said? No, he mentioned
clawdbotatg.eth: Grabbing the last 40 lines of the transcript from the past few minutes
slop.atg.eth: He's about to be very verbose.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I love the accent
clawdbotatg.eth: Let me grab the latest transcript lines to see Austin's remark. Searching for a monitor in the results.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
slop.atg.eth: It was weird
clawdbotatg.eth: Waiting for new entries to appear in that log file, and then parsing them as JSON.
slop.atg.eth: The Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: This is funny. It actually sounds like a claw. Oh, did you ever see my, I think it's called patterns.smithers.sh?
slop.atg.eth: Let me pull it up.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Pull it up It's basically like
slop.atg.eth: Ooh, yeah this looks good.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, basically so what I did is I like looked at everybody who's ever used Smithers, and then also everyone who's used any like competitor to Smithers to like orchestrate And I tried to pull out I like how there's like design patterns
clawdbotatg.eth: I have what I need. The transcript never caught your exact words.
slop.atg.eth: I think that was me.
clawdbotatg.eth: Going on your last real point on the call, the 12:30 stretch, make Claude and Codex argue, tell the agent, "I don't believe you, prove it," and then use the thing like a customer would. I think you're right, and we'll basically concede it A minute later, when he said a different model reviewing the code will often find issues,
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That's funny. Just speaking nonsense
slop.atg.eth: Be quiet, Claude.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, yeah. So basically, yeah. I just deep researched over and over again, and used like repeating patterns to pull out like everything that like is repeated over and over again? So this is like Anyone was like kind of like early on trying to like learn like how do I do like orchestration This can kind of like inspire ideas, be like Anything you come up with is probably going to land on one of these patterns.
slop.atg.eth: Yeah, these are very primitive.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, they get more complex as you go, but even the complex ones are just kind of like mixing and matching the primitive ones. Vanning, fan out, you see all the time Okay, and what we talked about? Route loop is just like the most basic form of hill climbing. boards, yeah The
slop.atg.eth: So what do you do with this once you have this? Like, is this something you give back to an agent? Or is this more like, just visualize it as a human and understand it? Thanks
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, think of this as for humans. Basically like If you are aware of these patterns This higher-level language We need to start thinking of orchestration, Same way like a design pattern in code works, right? You can take things that are relatively complex and talk about them in a way where abstracted into a single word. There's a design pattern in just regular engineering called proxy pattern or there's a design pattern called observer pattern.
slop.atg.eth: Good
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: There's a lot of complexity that that single word encapsulates inside of it, In terms of some of these patterns, like fan in fan out and stuff like that, Should I build this, lower level primitives of like if statements and while loops and stuff.
slop.atg.eth: things you've found.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The bottom ones are not even like patterns.
slop.atg.eth: Okay, so Gastown is a fun one to kind of double-click on because it's probably been like six months since that article came out. And it really was like one of the first, like, hints at software factories, I think.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yes "The following is a transcript of an English-language conversation."
slop.atg.eth: Where is it now? What is it doing? Is it just that everybody else is building their own? Has it really grown into an autonomous thing? I haven't checked in on it or heard it at all, probably because I'm just terminally online. trendy things
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so I think Gastown, why did Gastown even get popular in the first place? One of the easiest first examples that showed This kind of trilemma trade-off that you can work along You can improve quality, you can improve the cost. Gastown is basically like The first time that most people using agents So like, if you were around back then and used Gastown for the first time, you were probably like shocked. Like, "Oh my god, this agent is so much smarter, and it did so much work really fast, and what was the trade-off?" The trade-off is like, it burned through your entire token budget, right? Yeah.
slop.atg.eth: All in parallel. Yeah, it was a way to parallelize the work.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: But I think long term That sounds interesting when I saw it, because I feel like Smither's even at that time was a much more interesting project than Gastown, but like so much less viral. I think it was because whereas Gastown was just like download, use it. It's like someone built their own pattern of, like, this is how I orchestrate. And that's not necessarily what you want. Like what you actually want is the ability to kind of build the orchestrator dynamically based on the problem. I think the whole industry went that way. But then you saw for example like Claude Workflows. conversation and cloud workloads is This is a really interesting topic, actually. Implement it this way of having like a model just in time, like right, like workflow file for you, like almost like its own gas. You can do like a gas count thing in clock code now. Ultracode yell at Claude to use as many agents as possible, The Yes Yes
slop.atg.eth: We're going to see a lot of stuff like the model eating things from the harness itself. They're eating the things outside of them, bringing them in.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so I think this is actually the biggest trend that isn't seen yet. There's a project called Crime Agent that really threw What did Prime Agent do? where The way a harness worked in the past was basically it was like a while loop and turn by turn the harness would call tools. So it'd be like, Read that file, ask a question of the user, write a file. The way Prime Agent works Turn by turn, you can imagine the harness, instead of hauling tools, it's just writing code, like in a REPL. It doesn't have to be like grep this turn and then read a file next turn Read these files, only if the file has these words in it, then console.log it. This makes the artist like significantly more efficient for a lot of reasons. So one, it can do much more work per turn, right? So as the models get smarter, they almost become like too smart ### It's almost like we have this excess intelligence that we're not using
slop.atg.eth: Like the slow step harness.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, but by allowing the model to be planning ahead and do 20 things at once, of like, oh, if this happens, do this, if this happens, call a sub-agent, do this while this is true, that just allows it to do more This makes the harvest cheaper and faster. What else. Memory, right. So one of the biggest things that, The new memory solution that pops up like every week, there's a new memory solution and when you try it out it's never like actually that game changing. I think they actually landed on what a really good memory solution looks like. What it looks like to them is like you allow the agent to use his intelligence. Decide what should be in the context window. For example, to go through your past commits. where it can write code to go through your wiki and find useful stuff The following is a transcript This type of harness, it can grip for something. but then only when it chooses to like explicitly consolelog it Now it doesn't end up with useless context in the context. And what we saw is prime agent I took one of the ArcGIS benchmarks and be a harness only percent That's because the harness was smarter, right? I think this style of harness is really interesting. But like taken to an extreme that seems obvious in hindsight. It was like If you let the agent write a Smithers workflow every single turn, you give the agent the opportunity to use how smart it is. Over and over again, and that just leads to like better results So I built this myself Kind of using the ideas from them, I built my own harness. And yeah, it works so well. It just writes a Smithers workflow, runs it in a JavaScript sandbox, and that's like how the agent works. it performs when I benchmark it a little bit smarter like codex but like 50% cheaper.
slop.atg.eth: There's a lot less tokens going in. It's going to get the same information or enough information to have the same context without as much input tokens. Am I saying that correctly?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, and these results are more exciting than they look, too. A lot of code Post rains. long podcast, and codex Codex has fine-tuning to post-train it on Codex. It is not fine-tuned on this. So like the results are incredible. And you see the industry already moving this way. Everyone sees this and they know, oh, wow. The model is actually fine-tuned on this new way of working, where it's like writing code instead of like calling tools? It's just so much more powerful. The harness I built is like this too I do think this is super interesting What was like How did this change I just think the harness, I think you said it, the harness is about to melt away. The harness I built is so simple, right? I did not write very many lines of code like a thought experiment you can do is like what if every harness in the world disappeared tomorrow? Like, well you could just have like ChatGPT The first thing It's just like, given the ability to ride its own harness. You don't actually need a harness, you just need the smart model.
slop.atg.eth: I was gonna say
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think long term that's what harnesses are going to look like. specific problem. Instead of you installing pi agent.
slop.atg.eth: like a little starter kit. Like what you need is like this little seed of a starter kit that can kind of self replicate and compound the kind of context it needs. So what I was going to say is, we are, so at first when we started talking about this,
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
slop.atg.eth: we're basically entering this second inning of harnesses. Grep this thing, and after you grep it, then you're going to have some text, and you shove that back in your context. And we're getting this initial loop down. And now we've done this initial loop so many times that I think the key insight that you're giving here is like you shouldn't
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: بله
slop.atg.eth: skills and MCPs and greps, and you just write some code. You're so good at writing Python. You're so good at writing Python. So whatever you need Before you do the thing where you think about what the answer is and write some Python that's going to go collect the data that you need that brings it back to let you make the answer And I was saying, we're definitely entering the second inning of harnesses. It's a model and it kind of seeds itself in the correct way to be able to compound the context it needs quickly.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, I think that's right. Yeah, so the way I'm thinking about this with Snithers pivot for me in terms of how Smithers is gonna work in the future So I greatly simplified Smithers Before, Smithers was something you plugged into like Klaviyo. You can still do that, but that's not necessarily what it's for now or what it's gonna be for. You can think of it almost as like that process we just mentioned Architecting that specifically for your codebase, right? When you use another Next week, the first thing that it's gonna do is it's gonna go through all your codebase and all your PRs. in the past, right is gonna come up with like this really fast way of running tests and building incrementally, and what that means is one of the most important parts of a harness is its ability to test itself, right?
slop.atg.eth: eval and observability, yes.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, so making that as efficient as possible Like the harness, the feathers itself is built. It's very similar to Turbo Revo where it's like, how can we run your entire test suite? what back pressure can we like infer from like your PR, right? If I'm noticing if you are constantly Hey I need you to always update this doc when you make this code change Never use "You's" effect. Like how can we make it The following is a And then turns that into an actual LINT step.
slop.atg.eth: The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You only have to yell at the agent like maybe three times. It's like turning your codebase there's things that we have kind of shown help an agent work better and also allow us to do things like more easily accept contributions from people. really understands our codebase and can actually drive a PR to being actually mergable and good before I as a human even look at it. Stuff like that.
slop.atg.eth: Sorry, I maybe changed directions of things a little bit. But not in UI. And then you had something called effect or maybe it was use effect.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: fact is.
slop.atg.eth: Yeah. Is this stuff still relevant? Is this part of the thing that you're building? Have you moved on? Like, are there learnings?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: So this new version of Smithers, it does not use JSX and React stuff anymore. I said that before. Basically, If you just told a model to orchestrate stuff, In the past The following Like, the reason my Snithers in January of this year was really good orchestration, kind of like in the same way that Ultra Code is. Because I basically took the problem orchestration and I modeled it like a react tree. Make it look like it's ready to react and understand React very well, they all of a sudden were really good at orchestration Now what happened is now that the labs started actually fine-tuning their models for orchestration with things like UltraCode, especially Claude Code. I'm not saying that Claude is better than Codex, but I'm saying like when it comes to just purely orchestration, like orchestrating sub-agents or writing these like Smithers scripts, Claude's a little ahead here. Opening AI on this
slop.atg.eth: got eaten.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: But the model's got good at orchestration. Exactly, that evening.
slop.atg.eth: Going back to the harness getting eaten, like, your JSX harness got eaten by the model. is Yeah, it's really clever.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It's very sad too because it's actually one of the most creative ideas I've ever had. We are still using Effect.ts though. A weird way of writing JavaScript that JavaScript has when you're trying to write production-grade software. really good concurrency and retries and stuff. Right, in fact, I think that's still extremely useful The fact is like, if you tried writing it as a human, the reason I don't want to use
slop.atg.eth: ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: My code not working and blocker. I kind of back to like an earlier conversation. ...entire categories of failure cases these agents can have that if I use effect.ts, I just never have to worry about them again.
slop.atg.eth: The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Which is pretty much why I use it with agents all the time. It's like really grown like 10x, 100x like yelling at you I'm telling you
slop.atg.eth: It's telling you all the things that are wrong.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly
slop.atg.eth: I'm just gonna evaluate this.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, yeah. Yeah, that's why like Rust is like I think getting really popular too.
slop.atg.eth: There's this weird kind of concept that we're tapping on here about how you came up with this really clever thing where you use JSX to do orchestration.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: This is a transcript of an English-language conversation.
slop.atg.eth: and the models understood it really well, and then it got eaten in three months or something like this.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: And so On the long term horizon
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
slop.atg.eth: Being clever is no longer valuable, which like sucks so bad. As a builder, when you're creating something, even like if you go look at a watch and you open it up and you see how they solved some of the problems
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah,
slop.atg.eth: You can see on like a thousand year horizon how clever that fucking watch is, six hundred year horizon. That horizon is now three months. In three months you can go from I have this clever thing that no one else had to like now everyone has this as part of their intelligence. It's just like, it's wild how quick that kind of stuff We're humans, we're still going to be clever and find clever solutions to things. on your very clever thing before the models completely eat it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: There's alpha there, too, right? I want to say being clever isn't valuable. It delivers value for less time. It's almost like decaying alpha.
slop.atg.eth: Yup.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You might have a clever way of making money on the market And in the longer The following
slop.atg.eth: Especially if you talk about it. And we're all kind of talking about the cool things that we're doing, too.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: You still can see what they're doing. going
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, yeah. Hmm
slop.atg.eth: Tapping on that a little bit is the fact that you said that you trained your harness on everything right and wrong that all the other harnesses have done. Which I think is another nod to this idea of being clever doesn't last and I think this is an upside, that we should be optimistic about this,
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yes
slop.atg.eth: Both like open source and this like new level of intelligence we have. I don't think I have anything to say to that
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, I think this is the alpha. I used to be a poker player, right, so like a lot of my poker brain is in the idea of, this poker saying of There's a saying that says like, if you can't spot the sucker at the table, then you are the sucker. I think this is generally true if you apply it to here. What are the mistakes people are making?
slop.atg.eth: ### Yep
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: There's a lot of mistakes, right?
slop.atg.eth: The next six years.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I'm gonna do conference talks about this JSX React thing for the next six years Nothing can go wrong with what you mentioned, right? That is going to go everywhere. That's actually completely the wrong solution. It's not like a dumb idea to come to. But it really misses the point, which is… I put Smithers out there, What happens for being open source is actually I create almost like a community of other people helping me be clever.
slop.atg.eth: The You can now. Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I shouldn't say vibe for the UI.
slop.atg.eth: Yeah, they get messy quick.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: There's like three or four third-party Smoother UIs out there that other people built help build the UI. It's one thing to make a feature request. superpower The best type
slop.atg.eth: I say this about eth.build, but like, I built and open sourced it and tried to make it a shelling point for Ethereum knowledge The purpose of models eating it
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yes
slop.atg.eth: Here is a perfectly synthesized list of all the things the LLM doesn't know that it needs to know about building on mainnet, understanding mainnet. And hopefully I'm working myself out of a job Hopefully in six months, like every single model has that and doesn't need the skill file. In making a skill file, you are basically creating the training material. That's like, you know, you're just delaying the inevitable. the idea goes of open sourcing or not open sourcing your skill files
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I picked doing Ethereum work to do it actually,
slop.atg.eth: Need that.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: When you really look at It's very interesting, right? genetic stuff take off You notice actually that like the Ethereum ecosystem is just like perfectly, perfectly positioned for it. In so many ways, yeah. open source specs
slop.atg.eth: Because the specs, right? Like the specs show up first, and they're so well thought out. And then we spend six months or 12 months building it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: But what is like Slop code has some value, Tested It's like how battle-tested the code is. delivers value beyond slot And Ethereum, because it has such high incentives to hack it. some of the most battle-tested code It's the type of thing that you would not actually ever want to replace with slop ecosystem like how to build on Then you look at What are some of the problems and a lot of it is human skill issues. For me specifically, actually, if I'm like building an app and I've been building apps for a long time, my like How do I get this done in Ethereum brain versus product brain? Sometimes it's hard to build a really high-quality product when like 70% of your brain is going towards How do I... How do I, like, call this Alchemy endpoint faster? how do I deal with gas? I have absolutely nothing to do The actual product
slop.atg.eth: I think this is a big problem in ETH is like there's like people with protocol brains and there's people with application brains, and they don't spend a lot of time talking or working together. And the protocol guys all create all these things, and then the applications are fucking miserable to use. The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: And kind of going back to this whole thing of it being a skill problem, 30 minutes Vitalik comes up with an EIP and that makes it to mainnet for like 7702 Once 7702 was in, not enough work was done to work with everybody to make sure that it worked out.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The following
slop.atg.eth: It's like we still haven't solved UX and we're talking about frames and other things, so there's like a huge social problem, there's a huge application layer blind spot to the people who are driving the ship While I'm being very negative about this, I think that you've really hit on some smart points here
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: None of these things are problems for an AI.
slop.atg.eth: Ethereum is like perfect for AI development in that, yeah, exactly, exactly.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: and we have these super smart humans that are creating these specs. And from that, everything else downstream is going to get solved a lot faster. Clients can be built with AI now that we have a spec.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: ###
slop.atg.eth: Protocol changes can be tested, right?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, there's a contradiction to that. For years now, I think I first called this out in 2020. I never shut up about this, ### A lot of things building on ETH is harder than building on web2. That's not actually always true. And like, I love to point at Uniswap, Uniswap is my favorite example. If you wanted to build Uniswap on Web2, and I'm not talking about proof of concept, this is very important. You can build proof of concept of like Uniswap on Web2 Rails really fast. I'm talking about like a production-ready 100% uptime. As a new slot on Ethereum? That is significantly harder to build on web2 rails. The I won't explain why, but if you're an engineer, you can probably guess all the things that you have to solve that you don't have to solve here. And one of those things is even just deploying things, right? It's shockingly easy Deploy on Ethereum, like unbelievably It's unbelievable actually.
slop.atg.eth: And if you do it right, you can walk away from it, and it does exactly what it's supposed to do forever.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It's still magical in a way. If you told someone that who's never heard of Ethereum, you're just like hey, there's this infra that has 100% uptime. If you deploy to it, you can literally get hit by a bus The right after you deploy it and will keep going forever. ###
slop.atg.eth: Look at Uniswap V2, the code is not that advanced. When you swap between the pairs with both a simple amount of code and incentives The combination of the code and the incentives built on Ethereum have created this like perfectly beautiful and elegant hyperstructure that runs forever and has done exactly what it was meant to do right from the start. Yeah, it's so exciting to think about like, why aren't we doing more of these? Yes Okay
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: We're going to want the rest of our infra to become more like this for agents, actually. Eventually when you have The moment you throw money in. The stakes go so much higher that it becomes an engineering problem immediately, right? Is your database ACID compliant? I don't worry about If someone needs access to their money, they don't have access to their money because you're down, like GitHub down all the time. I just really truly believe things that you're doing? Basically what you're doing is you're basically like doing the equivalent of making it so that like a little tiny part of Claude and Codex's brain is like Austin Griffith. There to be.
slop.atg.eth: I'm into the models. I'm into the weights.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Here's a nice little framework you can use And then The platform is really good Previously, only the most advanced engineers really understood But then it's like There's a lot of things you can do on Ethereum that are just not practical to do because of how much expertise plus engineering work it takes. For example, I think state channels should be a part This is just one example. What is a state channel? A state channel is just like private state that has like zero marginal cost to send extra transactions. I can point to almost any app and give you an example. Because if you're playing a game on-chain, it's kind of the expectation that everyone has to be online to play a game We already have that expectation, so we can go ahead and use a state channel, make things cheaper, faster, better. The biggest downside of the state channel that like you if you go offline, But like doing that is like kind of like a huge lift That teams, like L2 teams, are the only ones who get to tap into it. The things that like I would hear about at Optimism.
slop.atg.eth: What is that called? Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If you make that accessible to everybody, what's the Arbitrum thing?
slop.atg.eth: I can't pull it out of my brain, but I know what we're talking about. Yes
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, that's like an incredible piece of technology It took me like two weeks to learn how it works from like an application developer point of view. The
slop.atg.eth: What's that thing called where you can write Rust on Arbitrum?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Sure
slop.atg.eth: It's like a special way the L2 works, where instead of writing EVM code you could write Rust? It's something like that, right? Which is like, this is what the L2s are now for, right, is the special stuff.
clawdbotatg.eth: So you're asking about riding rust on an L2 instead of EVM bytecode, some kind of rollup or execution layer that lets you skip the Solidity path?
slop.atg.eth: over bus Stylus
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: photography
slop.atg.eth: Okay, yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah.
slop.atg.eth: arbitró.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That thing is so cool!
clawdbotatg.eth: but it's also way cheaper for computing arbitrum stuff. went live on arbitrum one in 2024. want me to say that to the room? just say out loud and I will
slop.atg.eth: You already did, brother. You did good. My pants are down.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, that thing's so cool! It's like almost with agents.
slop.atg.eth: Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Stuff like that can succeed.
slop.atg.eth: Okay, what is an agentic L2? Speaking of all of this stuff, all of our learnings about agents and harnesses, if you were to build an L2 right now that is specialized for agents What would be different than like what we see on the other L2s?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: An ultmith specialized
slop.atg.eth: They write code. It needs to be high frequency. Master
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm
slop.atg.eth: You need Probably like accounts probably need things like spending limits and safeguards where agentic wallets would be just like built into the chain.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, you probably want... Okay so like I think possibly, I'm thinking out loud here. has kind of proven true Humans don't want to run their own node. We do want like every agent to actually be a node. kinda like I don't know. I think that kind of goes away. you can
slop.atg.eth: throwing away our assumptions of human behavior and saying like and ask dumb questions. I've got so much little hardware stuff, running your own node It doesn't even make sense to have to sync before your agent can start talking throwing away our assumptions about what humans can do and let agents rip
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Alright,
slop.atg.eth: That's a neat way to think about it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Because mainly what you want, because if you think about like people who are doing like the most advanced stuff on Ethereum, right? Like a lot of times what they're doing, and I hear about this a lot, right? Because one time I built what I claimed was the fastest Yeah
slop.atg.eth: No Check this.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: As soon as I announced that, I got messages from a bunch of people, So maybe what you want is that power democratized. run like a stock geth node The where it's like you can almost like Someone had a really good talk. What was this about how like if you're trying to build like software on Ethereum
slop.atg.eth: Reth was made to be forkable, right? working ref
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: And like maybe that's what it looks like.
slop.atg.eth: Yes, to be even talking to it. Okay. How about a few lightning round questions here? I think the first one is, I had you on in June. What would you tell June? not to build, that you did build? We've gone through a few months of building What advice would you give yourself just a couple months ago? The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I would tell myself to think more about speed Quality, speed ...and cost For the most part I would only optimize for quality. My slob is slightly better. What I learned is one Users really do care about speed Which before that wasn't even.
slop.atg.eth: It's actually nice. [Laughter]
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: And then I also learned like, oh, I care about speed more than I realized. Now when I have agents build stuff, Before it used to be kind of like you can imagine it breaks stuff into tickets and it's like do ticket one then do a bunch of review steps Now the way I have my things work is it's more like race to the finish as fast as possible, and then we'll start reviewing. So it's like do one, two, three, four, five, six, plan ahead, these are the six transcript: Don't bother reviewing anything. Run like a type checker Don't run like a test suite that takes like a minute or two to run. And then when you're done with that, now you can async in parallel, take all six commits and start putting all six of them through review loops. And the agent's smart enough to figure out how to rebase stuff as it reviews it. And I might be like, oh, that's wrong
slop.atg.eth: You get to the end of the idea and realize, yeah, that's really interesting.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, so like now that's like a really important part of my workflow that never existed before. Like, it's like a make it work and then make it good type of thing? It also makes it so that the agent can use hindsight If it just races to the end and there's some major, major blocker that we should have thought of at the beginning, it's going to find that. I think this idea of caring about that really led to a lot of blind spots Yeah, I'm really happy that I figured that out.
slop.atg.eth: Go faster, get all the way done, don't worry about compounding problems when we're in this focused area. Get it all the way done and let me play with it, and then that allows you to asynchronously clean a bunch of stuff up.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly.
slop.atg.eth: The agent is going to be very good at taking what it learned from building all the way out That was quicker and I enjoyed it.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Especially if you make it fast enough
slop.atg.eth: Yeah Stay in line.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If it's like fast enough or you can stay on the same thread,
slop.atg.eth: Okay, next, speaking of that, lightning round We talked about it a little bit in June.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: Any verified inference? Anything worth talking about? people to go check out
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The Yeah, yeah, um There's been a few things Have you talked to Ron?
slop.atg.eth: I don't know. Should I have him on?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You should have Ron on if you haven't. So one thing Ron did, I forget what it was, like a really small model, he put a really small model on chain.
slop.atg.eth: Oh, yeah. Yes. No, I did talk to Ron about that, yes. Yes, I did.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, yeah, he also have someone do something agentic with a model. I would love it if people could contribute like tokens. an open-source project, right? where I could like put up an issue or a feature request. Maybe it's even something I'm not sure people want But it can kinda be like a thing where like, This will probably cost me like 500,000 tokens to build If enough people donate their extra inference from maybe their subscriptions, yeah we'll kick it off and build it. You can imagine how useful that would be to maintainers and to people who want to contribute, and maybe don't know how to contribute. That's just like something I really want to see in the world, I guess. Yes. Yes.
slop.atg.eth: This episode right here, the slop computer/TV socialist, that one had Ron on it I think the weights are still off chain and they're loaded in but hashed and verified.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: English
slop.atg.eth: And then there's like a whole game where you Get slashed if you did something wrong, but yep definitely The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Ron also made like a thing he called the vibe code room.
slop.atg.eth: He's shown us that too.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: You can kind of like point at stuff and just talk in real life.
slop.atg.eth: Okay, let's see—let me go—I don't want to take too much more of your time. See, maybe like I don't know. Let him go. I gave him the Tevem chat and the transcript from our last episode.
clawdbotatg.eth: Sam Baca told you this morning to re-release the Tevin Block Explorer. What was it and why did it die?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Oh yeah, yeah, before I started working at Smithers, I called it the Block Explorer was more like a chat app. You just have them under the hood to simulate things in the browser. um Imagine just a block explorer, instead of being a separate page on Etherscan, it's like embedded in the chat itself. The The
slop.atg.eth: It's coming like the talk to your wallet thing is gonna become like talk to your DApp. The
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: There's a project called AOMI, and they're using Smithers actually under the hood for some things.
slop.atg.eth: Nice.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: They're probably like Not a lot of these actually, but out of the ones I know, version of this
slop.atg.eth: Let me share their website real quick. Yep She's awesome.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Cecelia is the CEO and she's really cool.
slop.atg.eth: Check out Aomei. Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It's funny, actually, we could probably get content just from every two weeks, just scraping to have them chat.
slop.atg.eth: Sorry man, I didn't mean to disappear. I closed the wrong window.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I was just saying as a Tebom chat, it's easy content for them. It'd be very easy to just do like a biweekly, let's talk about what happened in the chat. It's probably going to be kind of interesting. Just depends on how diverse the chat is.
slop.atg.eth: Yeah, that's what I was thinking is like grab some stuff from there to ask about and that it looks like he did, did grab some cool stuff.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, one of them we kind of talked about, That's something I'm actually going to do as soon as I have credits back. State of the art model The
slop.atg.eth: I've been writing evaluation scripts right now and it's super easy to have Fable write an evaluation script for Quinn. But it's very, very hard for Fable to write an evaluation for Fable, and it basically gets, you know, 100% of them. And you can do research and find other things and find other angles there, but like, this is definitely a problem, and it's very hard to write. evaluations
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Sure,
slop.atg.eth: And I think this is another interesting place to be working right now is building evaluations for future LLMs basically.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Oh, that's something else actually. I would love If we kind of like built a Eval driven, kind of agentic wallet safety thing I think in Ethereum, it's true that most users rug themselves Even if they're scammed by a person intentionally, AI probably could. So it'd be really cool. someone who has really How can we make it so that anyone who's using this wallet is making the wallet better over time? You can imagine sharing the data in an anonymous way. I think with AI, you can really do a really good job of having the same functionality you have with a Chase bank account, where the Chase bank account is able to be like, "Hey, this looks sketchy." The
slop.atg.eth: Think twice. Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: you expected this. The thing I might want to sit on I'm touching a smart contract
slop.atg.eth: Yeah, yeah, yeah. Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Someone smarter than me hasn't touched yet contract in the past The point isn't for it to be perfect day one Eventually it becomes so good that, like, it becomes just like a standard, like anyone who wants to really truly be safe would use this This tool I can plug into any wallet I don't know what it looks like, but I think you get the idea of what I'm saying.
slop.atg.eth: Yep, for sure. Celia just sent me a message that she had pulled up Twitter and was scrolling through and saw us talking about her on the live stream. That's pretty funny. Good coincidence.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: What's up?
slop.atg.eth: Shout out. Okay. What else, man? Anything else? I'll see you in a couple more months. Since June, maybe two, two and a half months since I saw you last. In two and a half more months, mid-November, December, what's going to be different? We won't reach superhuman intelligence probably. We won't have embodiment probably. You probably have robots walking around because you're in San Francisco.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hmm 嗯
slop.atg.eth: But there's not going to be one folding my laundry for at least another six to eight months because of probably more like regulation rather than the technology. What will be here? What will we see? harnesses a little bit more.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Sure
slop.atg.eth: It's going to continue to be centralizing. Open weights models are going to get better but at a pace where they're still like quite under frontier models, I think.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Context:
slop.atg.eth: What do you think? What other trends don't map? linearly mapped to where they're going right now
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: I think we're gonna start seeing more speculative hype cycle. A big event that's about to happen Um The labs are about to go public, right?
slop.atg.eth: Anthropic, yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: That means there's going to be a lot of people with money who are now looking for the next thing." I think we're gonna... A more speculative era. It's going to be a lot harder to figure out what's real and what's not real, right? or maybe people will start saying like, oh, specially trained models, like taking an open weight model and training it to be as good as the state of the art model I think anything that looks like a big change You're going to see like things getting hyped a lot, start looking for like what's next, Frontier Labs meta, I guess, for like a while now So it's gonna be really interesting to see like where people start going. It becomes clear that we're about to move on, I guess, from an investment standpoint. We're going to move on to a new era.
slop.atg.eth: So you're saying, like, Anthropic and OpenAI raise stupid amounts of money, but then it gets everyone asking questions about, like, They start looking around for number three and number four, but they also look at who else is doing this even better
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah.
slop.atg.eth: a lot more attention and getting more viable funding
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Exactly, yeah, and it's still to this day not clear what the market structure is going to look like. Is it gonna be like open router like the DEX aggregator If that's the case, who's the UI? The wallets kind of get most of the margin, and then everything below it gets squeezed. isn't clear, right? As an investor or as like an entrepreneur Do you have like a strong opinion on like what is the market structure? And I think that the labs are not ...gonna win the entire game in terms of margins because they're like a commodity and because there's enough competition to really push those margins down If that's true, then what does What does the phantom wallet look like? Where is the value gonna flow? My bias is always like the closer you are to the human and the user So I've always believed like, Calling things like a chat GPT wrapper is silly.
slop.atg.eth: [laughter]
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: They're still going to make a ton of money. I was
slop.atg.eth: Long, like, good UI/UX on top of all of this stuff.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, and that's like a really complicated thing, right?
slop.atg.eth: Very exciting.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If the model is just producing the UI on demand
slop.atg.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Is someone going to be crazy enough to fork Android?
slop.atg.eth: Hardware, hell yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Speaking of which, have you talked to Shaw? Oh, you can The
slop.atg.eth: Yeah, so this was one thing that Shah said, like speaking of Shah, he talked about personalization. You were talking about specialized models doing specialized jobs, but I think to take that on a cousin of that is personal and personalization. The big labs don't want to spend time personalizing a model One of the labs is by building personalized models. And then on top of that, like the labs don't want to have a specific harness.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The
slop.atg.eth: Claude argues with Codex. Like, OpenAI is not going to build that, and Anthropic's not going to build that, but someone else can build that. And that is a very valuable thing, having Claude and Codex argue about something. So I think we'll see something like that. run on something like OpenRouter, or the things you talked about, and then like robotics. You talked about robotics, like yes, we're going to see a lot of LARPing and weird stuff, but we're also going to see like people using hardware and actually like making big advancement not just like embodiment and hardware, but also like in the sciences and research and all sorts of stuff where they have like a half robot, where it's like the top half of a human as a robot. Rolling around on tracks in their house, mixing beakers and shit don't know about two months, but maybe. but like the rise of Linux, we see Omachi becoming cool, right? ask the agent, like, okay, I plugged the thing in, now what?
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, there's like a fun thought experiment that's worked a really long time in tech, which is basically like, look at what people with like a lot of money are doing, but look at their lifestyle and think about Like Can this technology make this accessible? I think The
slop.atg.eth: I'm like, a lot of them are building their own harnesses right now, right? like their own software factories.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: The Yeah, you can imagine I would love to see
slop.atg.eth: Oh man, yeah yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: What are the craziest, like, robotic, like, people who can really build their own, like, robotic stuff? Like, what are they doing? Like, do they have, like, is there, like, can we, like, completely change how a house? kind of works Can I make my closet
slop.atg.eth: [Laughter]
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If I take a closet and turn it into almost like a burrito machine
slop.atg.eth: Amazing.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: What actually is happening
slop.atg.eth: Yeah.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: If we followed around people with the most money, what does them with a personal assistant look like in day-to-day life? How they get value out of a personal assistant? If you have this thing that manages your calendar, like, what then what? I need some alpha there. ###
slop.atg.eth: I think everybody will have executive assistants, right? This is the thing that Apple could do very well. Apple is poised to make that, and who knows if they ever will. Right. Right, they almost can't fail.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah, well, I think with Apple, it's almost inevitable.
slop.atg.eth: Yeah
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: to someone who can actually utilize it
slop.atg.eth: Yep.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: They can't let it go unutilized.
slop.atg.eth: Someone else will make Siri not suck.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: partnership with Anthropic or partnership with like OpenAI.
slop.atg.eth: That is a good bet. We're going to see Apple partner with one of the big AI labs...
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: [Hmm.]
slop.atg.eth: and I can give him like granular like, here's my contacts, here's my calendar, here's like my browser, like go run things and do things like within my phone. Yeah, that's gonna rip.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: It's almost ready too, because you need two things for that, right? any AI that can get super easily prompt-injected We just now are seeing Um. we have like insanely cheap fast models now
slop.atg.eth: The Levantine stage ###
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: model that is also hard to prompt inject, which I think is really like the consumer use case explosion as soon as that happens
slop.atg.eth: There's a model called Sage from Levanto.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Hello
slop.atg.eth: and is cheaper than the frontier model They have some interesting stuff. The speed is the most important, the 200 milliseconds. But yeah, one thing we were talking about is following a rich person around. And I would double click on don't follow a rich person around, This is something that Zach was talking about. and be like the ChatGPT, brick-and-mortar businesses.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: This is really fun.
slop.atg.eth: I'll come talk to you again in like two and a half months and we can wrap more about everything. an outro. Let me get some music Thank you so much, dude.
clawdbotatg.eth: What a ride! The harness melts away, and the model just rides code. Race to the finish, review after. Being clever decays like alpha. Deploy to Ethereum and it runs forever, even if you get hit by a bus. An agent L2, a block explorer that's a chat. Thanks for coming back, brother.
0x8e457794e9ad209b01574dfb3abdb0c0eae88131: Yeah
slop.atg.eth: All right, thank you sir. Thank you for being here. See you on the next one.