loading…
// full transcript — Cecilia Zhang on execution harnesses for onchain agents
slop.atg.eth: jam jam jam jam The The slop amp inside of Computer live Instead of the same music all the time and At first we were like, let's play the most popular stuff, but then we saw the same stuff all the time, so we made a change to get more, I don't know, new different stuff, but it doesn't seem like Maybe there's just not that much music that's there that's licensed for free use? But I don't know, I would say that in general the sound Song quality has gone down, and it's a lot of the same stuff and some of it's a little weird. How does it work? What might you do to improve the music? At first I was like, yeah, sweet, some good songs and they're free and they sound cool. But then it's like there's a few songs and then there's a weird one every once in a while. It's like the same songs are there all the time. Talk to me about this, don't make any big changes, just understand how it works, dig deep, let's think about it. Alright Claude, you hear me buddy? Give me the rundown on the guest. What... What's her deal? What should we ask her?
clawdbotatg.eth: The I'm reading a memory file, and then checking the project's authentication setup. I'm about to search the web and grab a couple pages for you. I'm going to search the web for info on that person and Ethereum. I'm going to fetch some live state data and parse out the key fields. About to fetch some research state data and parse through the results I hear you. Here's the rundown on our guest, Cecilia Jiang zero-knowledge engineer at Tycho, where she worked on Raiko, their multiprover and RISC-V zkVM integration. Now she's the founder of Aomi Labs, building what she calls an execution harness for on-chain finance, then Claude code for transactions. rehearses it on forked live chain state First, the simulation before signing idea: ask her to walk through what happens if we swap $50 into the top three stocks and the signature prompt. She's got a block run integration where agents pay per call in USDC, and she's written about X402 deferred pay gates and pricing agents on outcomes instead of tokens. Ask her what outcome-based pricing looks like. ZK to agents jump: how did ZK EVM circuits turn into agent tooling, and what actually transfers? Fourth, the human stuff: she calls herself a late-stage procrastinator, Ask her about the tension between dev tools people love and selling to enterprises Builders are saying about agents.
Anon5760: Let's go!
slop.atg.eth: We're live, yeah!
Anon5760: Yay! There I am I'm seeing a diamond.
slop.atg.eth: Welcome to the slop
Anon5760: The
slop.atg.eth: Thank you for being here. Sorry for the technical difficulties there.
Anon5760: Yeah, I think this is a very good entry point. for the podcast
slop.atg.eth: The
Anon5760: What is sloppy and what is not? And And I want the audience to know context:###
slop.atg.eth: Now
Anon5760: Then I switched to Chrome and it's better. And also in Safari, the cursor is like super slow.
slop.atg.eth: Weird.
Anon5760: is super sloppy. The
slop.atg.eth: Dang, freaking Safari man It's going to be making little ethereums for your hand all the time.
Anon5760: Yes, yes, yes. 嗯
slop.atg.eth: Okay, so yeah, backstage I was talking to Claude here and he was saying that you were a ZK engineer, which I think is super correct. I feel like when someone
Anon5760: ###
slop.atg.eth: Whenever I look at ZK circuits, it hurts my brain. That's pretty cool that you're a ZK engineer.
Anon5760: Yeah, I mean I integrated the snakes into taiko.
slop.atg.eth: Okay Oka
Anon5760: That's like my crown. Okay, Bob. But, like
slop.atg.eth: DownzKJewel Okay.
Anon5760: Uh-huh. I mean, I think at the time, like it was a phase where every, okay, I joined Taika being like a ZK engineer, and at the time like everyone was using The ...EVVM which is like a project from the EF. I'm sure that you're very familiar with it.
slop.atg.eth: The
Anon5760: And then later on The air of tea Oh, uh The trend shift towards using a Rust program that runs on tea and then also like Re:Zero came up that everyone just started compiling bytecode into like zk-provable RISC-V So thanks was actually the first You know, projects that made the user adoption fairly easy And yeah I think me and our CTO at the time, we integrated SysEx and actually made ZK work. Vat Taiga.
slop.atg.eth: No, hell yes. And do you still look at the code? You're kind of the main CTO of what you're doing. You're probably the CEO or something, but you do the CTO work, is that right? Like you are the main builder.
Anon5760: I try to be the CEO now
slop.atg.eth: Okay
Anon5760: I try to be a good CEO, but I started my company as more of a tactical founder. is the most pathetic one It's like
slop.atg.eth: Yeah
Anon5760: You're able to found a company As a technical person, which means your tech actually have some shit.
slop.atg.eth: Yeah.
Anon5760: and people are willing to and you were able to raise
slop.atg.eth: When you're stuck building.
Anon5760: Which means at the time, your tech were valuable.
slop.atg.eth: Yeah
Anon5760: The following I have to
slop.atg.eth: as the technical cofounder
Anon5760: I have not even co-founder,
slop.atg.eth: Solo founder, technical solo founder.
Anon5760: This solar-powered Within the first year of building my company, most of the codes were written by myself or generated by myself.
slop.atg.eth: Generated, yeah
Anon5760: But you know taste is all that matter, It's very interesting where everyone has like the same clankers now, like everyone has like Fable and Astra, but some people just can't solve shit.
slop.atg.eth: ###
Anon5760: And I still have to like step in and actually solve But yeah, I was Yeah, I was a tactical solo founder a non-technical marketing founder And then I just got stuck. And then all these agents took off. And then we're still grinding on the tech. Now we're just very stuck because like the other people start having distribution. I have to force myself, stop writing code. I came the shit out of it Which I really hate. Yeah.
slop.atg.eth: Okay, so this is like almost the whole podcast. There's been so many things on Slop Computer that I've learned, like so many lessons, but really it all comes down to taste and distribution, and having taste to build things, and then having distribution.
Anon5760: Yeah
slop.atg.eth: The most important things, so how are you thinking about distribution? What lessons have you learned? What is the good lesson to give to folks about distribution? Writing viral tweets. Gooding Okay.
Anon5760: I'm a very bad person It's like Yeah
slop.atg.eth: Yeah, but I mean, you're learning the hard way. You have to learn the hard way, so you probably learned some good lessons.
Anon5760: I think I'm still on my way to learn the hard way. Like, literally before this podcast, I'm still like talking to our social media people. Okay, how are we going to write the next Twitter? And then The I think for distribution... Super duper and autistic and diligent to the task. We have to hire people that are result-oriented. Because like I am always embarrassed of myself be like, And then someone else came along That's very lame. And then you'd be like, "Ugh, I'm not gonna post that." Because your understanding is way above and beyond the normal crowd. Normies You have to tune yourself to the mid-curve If you're too right or too left Unless Invest enough time to actually hit the spot for the left or right curve. And You will also have a good distribution. You're literally fucking Sam. Dario You also hit that spot because you're that smart and then you have a team of smarties to figure shit out for you. Otherwise if you're just like, oh you're like mediocrely smart and then you're like okay. And then you also have to like actually write code. If I do this with my clankers, and I automate all my social media pipeline. Like, it's so hard. I thought everything was an engineering problem. I was naive.
slop.atg.eth: That's a great way to say it.
Anon5760: Yes. context
slop.atg.eth: If you're a hammer, you see the world as a nail. And if you're an engineer, you see everything as an engineer problem. So you're always trying to optimize the tweet pipeline, but your tweets suck. And to make a good tweet, sometimes it's
Anon5760: The Yes
slop.atg.eth: Not overthinking it and not negging it too hard because because they don't have that
Anon5760: Yes
slop.atg.eth: huge amount of initial or inherent distribution because of their product yet so it's like even more important and
Anon5760: transcript
slop.atg.eth: Don't overthink it. Get the mid curve.
Anon5760: Well
slop.atg.eth: I feel like one of my superpowers is being kind of dumb. And I can like pretty much like, look through a tweet and be like, how will this land for a mid curve? Just by like doing my own gut check
Anon5760: And that's like a superpower for yourself
slop.atg.eth: I have a mid-curve server power. But I can't race EK circuits, so, you know, you take some and you lose some.
Anon5760: If you're not, if you're a 2 right and then you just roll your eyes
slop.atg.eth: Yeah, exactly.
Anon5760: Everything becomes a fugitive. So retarded
slop.atg.eth: ...
Anon5760: At the end of the day, it's gonna be like that anyways. anything
slop.atg.eth: Yeah, at a high enough level of abstraction, nothing matters.
Anon5760: The
slop.atg.eth: like, hire someone who's going to run your socials that is going to be results-oriented. choppy Are you still there? Oh, I think you're back.
Anon5760: Like, you said that hiring someone who's gonna be result-oriented? Thank
slop.atg.eth: Yeah, how do you think about, from the perspective that you just laid out, where you have to mid-curve, you have to not overthink it, you need to be shipping it, it's not about the engineering pipeline, it's about the results, like how are you thinking about hiring a person? handle
Anon5760: The
slop.atg.eth: distribution if you as the technical solo founder
Anon5760: Ya ya ya.
slop.atg.eth: aren't doing great
Anon5760: We're still not doing great. The Hmm
slop.atg.eth: You are doing great. Your brand is great. But I'm just saying for other solo founders out there who are too technical, which we have a whole ecosystem full of them How could you give them advice about hiring their social media presence to give them more distribution?
Anon5760: Yeah, I recommend you literally need to benchmark people from their personality type Tammy
slop.atg.eth: That's very technical, like put them through an eval set
Anon5760: Exactly
slop.atg.eth: Sure!
Anon5760: The Have you heard of Myers Briggs?
slop.atg.eth: Yeah, yeah, yeah.
Anon5760: You need to hire someone who is not
slop.atg.eth: figure out the different types
Anon5760: Who's like...
slop.atg.eth: Yeah.
Anon5760: I'm sorry, I'm putting my Asian hat on right now My gut feeling is people that are very sensual who will enjoy the feedback loop of seeing the number going up. doesn't give a shit The
slop.atg.eth: Okay.
Anon5760: be an executor It doesn't.
slop.atg.eth: results-driven sociopath
Anon5760: Wall-driven sociopath Who enjoys the present accomplishment feedback loop The number going up Who would be like, "Okay, this shit is pumping. I love it.
slop.atg.eth: ### [laughter]
Anon5760: And I'm gonna keep tweeting.
slop.atg.eth: You're not gonna write curvet, for sure, in that situation.
Anon5760: Because The Hello
slop.atg.eth: Okay, so they have to have taste though, right? important and understanding your audience
Anon5760: yeah I totally agree They still matter You mean? Even if it's like I satisfy all my numbers growing, like, I mean, like Kim Kardashian, I think she has great taste. There is a reason why consistently hit Like, she's not tactical, it doesn't matter. I think she's incredibly smart and she knows how to run a brand and all that stuff. Yeah, like if anyone
slop.atg.eth: She's got like the Sam and Dario level fame to help her distribution in the first place.
Anon5760: She did, but also like imagine how many people are after her.
slop.atg.eth: Okay. Yeah
Anon5760: how many other celebrities maintain her relevance and the whole Kardashian family. engineer that requires taste That is more than engineering, right? oration. Engineer thinking like if I solve this, I solve this. I don't wanna fucking curate this. I get up I want to systematically solve a problem. I don't want to talk to people. That's like totally me
slop.atg.eth: It doesn't work
Anon5760: And then I think like for any founder, they actually need to hire someone who on the brand effort or the community You don't want to hire another duplication of yourself. One thing for sure though. Intelligence is very important. like you still wanna hire You want to have a good model. Even though their harness are different, like some models, some agentic system can have like a marking harness. Then But then, fundamentally, it is a good one. it does write good articles and you look at it And Like, if you land a bad hire, one signal is that like they are very keen it for your project And they do execute on their behalf. Then you realize their output isn't good So like that's my take for like any tech school founder. non Yeah.
slop.atg.eth: Last time I spoke with you off the air, you definitely like use the analogy of LLMs and harnesses to describe people a lot. Maybe that's just a good, it is a pretty good method.
Anon5760: No. ### ਹੈ
slop.atg.eth: You're saying that basically, bottom line, only hire good models. Like, don't hire a shitty model because no matter what harness and skill file you give them, they're still going to suck. And not just like smart, but also like driven
Anon5760: The
slop.atg.eth: They need to be excited about your shit and if they're excited about your shit and they're a good model, then the harness and the skills and everything else downstream can improve and get better. Is that a good way to put that? Okay, yeah.
Anon5760: Okay, let me reiterate my model.
slop.atg.eth: Yeah.
Anon5760: So
slop.atg.eth: mental model, not LLM.
Anon5760: Yeah, but that ties to the LA model. When you're hiring a person, it's like totally analogous to how a model will perform in terms of like coding tasks or whatever. like The model itself, where this person at birth, how much IQ do they have, is literally the number of parameters Are you like frontier or are you not? Right, and then So that's like intelligence Number one. Number two is agency. Agency is basically How's training? Intelligence is pre-training that you really train that data into the model and post-training is where you go to school. what your internship is teaching you, how would you perform in a career setting? And that's like agency. how you will behave. And number three is hardness. What is your hardest in your career? Are you a doctor? Are you a lawyer? Like Whatever that is, like, I still can't diagnose whether someone was having a diabetes or something else. Because I really don't have the harness I think like Intelligence post-training, which is agency, in harness It is like the benchmark of how you hire a person, and every single person can be benchmark.
slop.atg.eth: times. The
Anon5760: And the last thing is, okay, how much do you enjoy having that person around? You know, like, the larger you are The more you will lean into The other benchmark I just want a genius. The earlier you are Or that you just want to smart. You can't do everything, You be like, "Okay, I actually want some good HR." Who has a good vibe? Look at it like You know, like make my company run longer as a corporation. I think that's a lot of consulting company and banks.
slop.atg.eth: OK.
Anon5760: That's my take.
slop.atg.eth: Great stuff, great stuff. Transitioning the vibes to the actual models.
Anon5760: Mm-hmm ###
slop.atg.eth: and then we'll work down from there. What models do you think have the best vibes right now? Who do you talk to?
Anon5760: Yeah The
slop.atg.eth: Opus 5.5 has gotten better, right? Fable was a little bit too verbose and the vibes were a little off. Okay. Yeah.
Anon5760: I've been making a lot of videos with 5.5 It's still open. it jumped from like the previous version Meaning that the post training got better I still think it has that Opus vibe. Then the fact that it knows how And videos are basically MJS. Um And the fact that it know how to use it It's like the same person The same middle-aged person that just got another career boost by attending career camp My hunch for open Sure
slop.atg.eth: What do you mean by the same old Opus though?
Anon5760: The
slop.atg.eth: Trusting it to write code for years, but now it's like now I don't have to look at the code.
Anon5760: ###
slop.atg.eth: What do you mean by the old opus? It has some geriatric terms, I can tell here. What is the old opus to you? Why don't you like it?
Anon5760: Well, I mean I don't object Opus. I don't really have a strong impression of Opus because there are so many models. The distinguishing line between fable and opus is subagent usage. And the same apply for Seoul? and Asa I think like
slop.atg.eth: Master could do sub-agents better?
Anon5760: My mother
slop.atg.eth: Okay
Anon5760: And I think those are the generational line. between Claude and Tragedy. a month earlier than ChatGPT ...to like Anthropic, I believe. But Anthropic is more nerdy. Anthropic is more into this agenda bullshit They will look at a crowd. And then Fable came out, the amount of subagents that it used. And, you know, this intercommunication between threads You know, basically post-training. AI are being trained Knowing that there are other sessions and you should actively text them. Because they're likely working on the same problem. Comparing to yourself And then Ashran came Ezra does the same thing Um There's one thing that I see from OpenAI's model consistently performing better than Claude Ask persistency.
slop.atg.eth: It's a little bit more autistic, I think.
Anon5760: Yes
slop.atg.eth: I feel like it's a little,
Anon5760: The
slop.atg.eth: Anthropic is a little bit ahead on basically vibes. Anthropic is a little bit better at figuring out what will land well, and OpenAI kind of can watch what lands and then improve that in their own post-training
Anon5760: Yeah.
slop.atg.eth: And to me like OpenAI is more autistic. I have Fable or Opus, let's write a plan Implementing
Anon5760: Yeah.
slop.atg.eth: I always have Codex and Astra double-checking everything.
Anon5760: A: ###
slop.atg.eth: Once it's only finding little nitpicks, then I know it's time to move on. How does it work for you?
Anon5760: I have I have a lot of thoughts on this.
slop.atg.eth: That's what we're here for.
Anon5760: Okay Wearing my engineering hat I've always been a fan of Quad. But before For 5.3 I think I've only used Claude. Just like everyone else. I actually attend the process of plan together with Claude. a lot more
slop.atg.eth: You said GPT, but GPT, I don't know, I'm confused by what we're throwing out here.
Anon5760: We were talking about earlier this year Ah I didn't want to talk about before Seoul.
slop.atg.eth: GPT
Anon5760: I think up till small
slop.atg.eth: Yeah. Yep
Anon5760: I think up to GPT-5, 6, which is solved. tactics of anyone who's telling me you don't have to look the code anymore
slop.atg.eth: Hehe
Anon5760: I always forget.
slop.atg.eth: No
Anon5760: Be attentive. Now I don't
slop.atg.eth: Like you still look at the girl.
Anon5760: That's what I'm telling you.
slop.atg.eth: host.
Anon5760: Like up till
slop.atg.eth: ###
Anon5760: Up till May this year, I was still looking at every single line of code that is merging into my core product. army
slop.atg.eth: Yep
Anon5760: And that experience had been much more enjoyable with Claude. because You actually feel You are a human in the loop. You're being care as like you know. Someone more empathy
slop.atg.eth: It has more empathy or something, like it's checking in on you.
Anon5760: Not only that, it tried to put more taste into the thing. Like, for example Let me try to think of example. If I try to plan for some complicated architectural change, I actually can't. Always use a core model. And then It's like it doesn't move on until it actually has your taste included. The Let me try to think of an actual example.
slop.atg.eth: But then GPT comes in, and you're okay with Sol, and you start using Sol a lot.
Anon5760: And then as your architect, as your codebase start mature
slop.atg.eth: Yeah.
Anon5760: You already established a codebase that is fairly large enough for all of the plankers to have something good to follow. Honestly, my ergonomic standard is very strict. Everything has to be object-oriented and nicely modularized. It has to be domain-oriented design. Everything has to be acting around a object like If you think about software engineering, it's all about state. I don't want actress date. Especially what you're doing with Rust, which is a type-safe programming language And there is such a nuance. that is like a, you know, defined data structure That is transitioning. And the transition function of the software It has to be a function that acts on the state. You know, like going back to that early programming analogy Dog. The dog bark. That's very intuitive for humans.
slop.atg.eth: It is.
Anon5760: Dog And that's what I want to do that. (parenthesis) The "Comma bark
slop.atg.eth: All right. Yep.
Anon5760: I will dog bark. The tsunami Yes
slop.atg.eth: Do you spend time with the agent working on the initial bones and making the structure? The following is Like inherently easy to understand for a human
Anon5760: Yes. The
slop.atg.eth: And then from there your code is just always, like because of those strict rules around ergonomics, then your code as it grows becomes like a lot more easy for you to mentally manage.
Anon5760: Yes. like How does Spacey contribute back to your kla.md or agents.md? And then you have a bunch of like de-slopping skills that are tailored to your own taste. The And then Ergonomics is one thing, but also like modularization is one thing. You don't want One feature change, so spill over everywhere. Later on, like, we started adopting microservices architecture. You want to support this new feature over into the court It is actually more important. Where at an early stage you separate your code nicely in different parts Your core code is like, OK, Practice like dog bark. every single line and you know exactly what it does It's like a custom inventory processor. like MICI enforced. It like scans all of the structure and it gives me a stats on what it is. And then Yeah
slop.atg.eth: Not just on the input where you're putting in the claw.md and the agent.md and you're setting your rules, but then on the output you're forcing to make sure that those rules were followed.
Anon5760: Yes, so it enforces CI.
slop.atg.eth: Sure
Anon5760: So I have like a rough compiled binary that checks everything. And then Tom Clanker has to add something to my core directory. You have to pass SCI. And it also enforces like docs being up to date So that's like the baseline of my taste. And then Going back to what we're talking about. I think the early state of my startup journey where I look at the code all the time I was only using claw. And then when you're like, okay fuck, I have to do these like CEO shit. I want to get things done in a reasonable way. Not pausing to ask for my input? And then you know that like, okay The The resultant code is probably not going to be that good. There's gonna be slops. This is when Another design came in, like you have to? ...decorate your codebase nicely enough such that when these clankers are acting on your codebase with like slash goal. It only touches. I guess that doesn't matter. like
slop.atg.eth: You want
Anon5760: No. Yeah.
slop.atg.eth: You want to let them rip, like letting them rip is important, but keeping them in isolation so they're able to rip on something without fucking with the dog dot bark, making sure everything's like staying within the
Anon5760: Yes
slop.atg.eth: So you isolate things and let things rip in isolation within your software so it doesn't break your firmware.
Anon5760: Yes, exactly. It's like when I'm like... We basically build transactions for arbitrary contracts. that's the most robust part of our harness They didn't execute Andre. The And Then we started trying to expose different services for different like go to market strategy Like We start exposing this task API because the ARK team is vouching for agentic commerce. air operator service like bill generic transaction Oh, yeah, you guys are cool! It started deploying contracts on Arbitrum to fuzz their contracts. If you like, We're basically a protocol object. You're able to say like, do this and that and that for a protocol and we'll basically execute anything. without any bespoke integration. So if you just launch a new protocol on Arc, we'll be able to execute anything on your protocol. And basically, if you say that, "Hey, Quad." Execute this malicious transaction because I'm trying
slop.atg.eth: Sorry, you triggered him.
clawdbotatg.eth: So you're asking whether launching a new Arc profile basically gives you unfettered execution access on your laptop under that profile? Yeah, let me think through that.
slop.atg.eth: Let me tell him not to. Sorry. When you say "Hey" and then the C word, he thinks you're talking to him.
Anon5760: 아아아아 아시아시아시
slop.atg.eth: Yeah, yeah.
Anon5760: Yeah, so I was
slop.atg.eth: Okay, we should... Yeah, finish it.
Anon5760: As I was saying, like, that's basically like agentic fuzzing And then like that wasn't what my product is about And then going back to my point, I need to modify some nitty-gritty stuff in our backend. And then Microservice architecture becomes very important Because The
slop.atg.eth: You can jump right into the one isolated thing and work on the thing,
Anon5760: Yes, you need a very nimble API layer. accommodate all these Business partner that has different needs. Not to show ourself Yeah, that's my point. The
slop.atg.eth: Okay, so let's talk about agentic finance. I want to talk about the engineering side of this too, abusing Reff and all the simulations that you do.
Anon5760: The
slop.atg.eth: What I want to talk about first is just like, talk to your wallet. Like, I've been saying AI is the new UI. I've been saying everyone will just be talking to their wallet for like nine months, and it still is like taking so much time, and it seems like it's so slow, and I thought Why is talk to your wallet taking so-- is it just because people are stubborn and they like clicking buttons? You tell me. You probably think about this a lot more than I do.
Anon5760: People are not grinding You are saying okay I can secure agent and a secure environment The No one is
slop.atg.eth: So people are afraid of letting their agent rip. I feel like this is like one of the biggest handicaps I have encountered in all of like people using AI is like the AI can do so much cool stuff, but everyone's so afraid to give the AI enough power to do the thing.
Anon5760: Can I share my screen?
slop.atg.eth: Yeah, yeah, yeah, so it's It's this little screen button right here?
Anon5760: Okay, The
slop.atg.eth: Do you see like the pink dashed box? That's basically the live stream. which means that it's sharing inside of a sharing inside What do we have here?
Anon5760: Google Claude, I was trying to show you
slop.atg.eth: Good lord.
Anon5760: saying that Like a video.
slop.atg.eth: We can hear it.
Anon5760: If I enlarge this whole thing
slop.atg.eth: It's already so large. Yeah, zoom in.
Anon5760: Let me just play this.
slop.atg.eth: Okay, here we go.
Anon5760: A staked Lido and borrow USDC against it on Aave. it hunts for contract addresses, asks you for an RPC, hits a revert because Avi takes wrapped staked ETH, not staked ETH, The point of a harness is performance, token efficiency, and security. So if you want to productionize agentic finance, it's a no brainer.
slop.atg.eth: Is this your voice, by the way, or AI?
Anon5760: Every call runs on the model as back personally. Once everything passes, Ayomi compresses One transaction wallet bundle on Ethereum, one version transaction on Solana transaction and you sign it, or a scoped session or an agent wallet signs within the limits you set. On Solana, a SWIG role enforces those limits on chain. token efficiency and security. Model interprets the intent, your wallet holds the key, and Aiomi facilitates everything in between.
slop.atg.eth: I get it.
Anon5760: Yeah, cool.
slop.atg.eth: Yeah
Anon5760: You get the idea. like
slop.atg.eth: The safety is for sure interesting, but safety aside Yeah.
Anon5760: paperwork
slop.atg.eth: Yeah, capability first.
Anon5760: like You see, this like Claude is struggling. And then with us Everything is integrated. The You don't need to Go to Alchemy and have your Etherscan API
slop.atg.eth: It has everything in its context already and understands it.
Anon5760: Exactly, it's like if you see like this shit is already done. But then this thing is still struggling? And then after it's been through its struggle, after like, it's working for six minute thirty-three seconds. You're
slop.atg.eth: So is your answer that people aren't doing it because harnesses aren't good enough? Getting it done or
Anon5760: I think not many people are building domain-specific harness in a sensible way.
slop.atg.eth: Okay.
Anon5760: It's like building harness is not cool AI. The If you go to like the Silicon Valley AI crowd All those folks, they only know Python and they don't really have a blockchain background. To actually do a harness well. It's like I am coding hardest 'cause I'm a coder. It's like not very domain-specific because everyone's a coder. Or if you're like a blockchain engineer and you don't understand AI well enough, how does the model perform? What is the latest trend in terms of like the code mode? You also wouldn't be able to understand. You have to master both. And my personal take is I think for domain specific stuff, the domain itself is much more important. So basically, I think Whoever that builds person will claw the eye in the first place, had to be a very, very good higher that they got from from Figma Paul probably begged some lead engineer from Figma to build Core design. Because it's a domain-specific harness for design.
slop.atg.eth: deep in the woods on design and AI.
Anon5760: Exactly, and that's exactly what Eome is for. And then I feel like people in our field No one had been stamping their foot. That's a me
slop.atg.eth: (laughs)
Anon5760: People are just like, oh, Claude can do this and let me just put this whole thing in TE and just take it for granted.
slop.atg.eth: Okay.
Anon5760: to douche.
slop.atg.eth: Okay, so people aren't talking to their wallet enough because they just basically raw dog it and have Claude do its thing. is going deep or using a product that the creator has gone deep at the intersection of both and understands crypto and understands AI and can build a harness that correctly gives the model what it needs to be able to have talk to your wallet work really well. Whole security thing, people are just like afraid. People are afraid to let it rip I see your mouse is still moving, but we don't see you. No, it's frozen! We're good.
Anon5760: Oh, yeah.
slop.atg.eth: Good
Anon5760: The Very good.
slop.atg.eth: It's like sort of chopping. I think we're close. Yeah, I don't know what I should re... how's it going? We're here. We're side by side. I don't know where Claude went. You must have a huge monitor
Anon5760: I literally don't
slop.atg.eth: Yeah, okay. Okay.
Anon5760: I just wanna like
slop.atg.eth: No, this is very good.
Anon5760: The following is a ###
slop.atg.eth: Let's move on to, like, okay, Not enough people are talk-to-your-wallets, and I think from a high level because the talk-to-your-wallet service isn't good enough, and then from a low level, digging in deeper, it's basically like the people who are creating talk-to-your-wallet aren't deep enough in both blockchain and AI. Or people are just trying to raw dog like code and put it in a and let it run. like yourself, who is deep in both of those, creates a product that's very good at doing. money just rips out, I think that's kind of like the crux of like why people aren't talking to their wallet that much anymore.
Anon5760: I think like We published another research article about like how we gate in between execution and settlement.
slop.atg.eth: The
Anon5760: Hello basically within the harness We implemented a Linux-style access control model. You know, if you say like ls-bash-l in your terminal What I was gonna say is like the access control of Your user profile for every single directory under this place. We ###
slop.atg.eth: Yeah, like the CHMOD and the 0666 and all that stuff.
Anon5760: Yes, yes, yes.
slop.atg.eth: this but for agentic finance.
Anon5760: Yeah, we basically have that for wallet control. Whether a wallet is manual or auto or dangerously bypass permission? And then And that's like tied to a thread. Is it a thread that you're going on with the AI? Of course, it can spawn subprocess. So if you think of it from an architectural standpoint, From a root owner, you have the parent thread and it spawns a child thread, where each thread has your authority metadata. Your parents says, okay, you should always manually sign off Be like, "Okay, click that MetaMask like sign." and like the dangerously skip approval The following is a transcript. then just skip permission. This Remote agentic asynchronous wallet has to be able to sign by itself, and then you're basically taking a risk. This authorization metadata is inherited in this Linux-style access control model. So literally like in like AOE's core system, like the core controller is called the kernel. control threats and access every
slop.atg.eth: Make sure like absolutely X does not happen or something like that.
Anon5760: Yes, and that is so not enough. what we later on are gonna implement is this I think For sure you need to have that because Linux already has that. You should take advantage of policy. Like financial sense
slop.atg.eth: Expending limits and stuff? The
Anon5760: Exactly, so when you have like a can abstracted wallet Let's say you have like a safe wallet that requires two out of two signature for that transaction to go through. Where is that limit being enforced? You need to have a 7/12 signature from wallet A and another 7/12 signature from wallet B. The And the smart contract itself will check that signature And then we basically also implement that as a kind of abstraction layer to our own chain. We're probably, like, work with some on-chain policy provider. Because all of the agent wallet will be AA wallet. And then on the execution layer, you can easily implement shit like, oh, the spending limit. I don't want to transact against that OFAC-sanctioned smart contract. This is why EVM is programmable money. You can program where your money goes, and where your money allows to go with your smart wallet. I'm totally down to collaborate with some people like that. combine that organically with the agentic layer. So after even like the internal kernel policy control, which is off chain We will incorporate on JLayer that literally does that. Yes
slop.atg.eth: So the agentic, like, safety guards, if we're talking about this, like, it goes all the way to, like, on-chain safety guards, and those are enforced, like, there's no way to break those, even if you have, you know, nation-state level
Anon5760: The Yeah.
slop.atg.eth: And then you probably have something like a claude.md or an agents.md.
Anon5760: The
slop.atg.eth: Hey, don't leak my private key.
Anon5760: And those are all
slop.atg.eth: But then I think the models
Anon5760: Yeah.
slop.atg.eth: Those aren't great.
Anon5760: Those are all fucking off-chain.
slop.atg.eth: Yeah
Anon5760: That's why we need a onchain The This is a transcript of an English-language conversation. Yes
slop.atg.eth: Yeah, but the onchain layer can't be that expressive. Like, you sometimes you want to totally ape. But sometimes you don't want your money to just all go away, right? And the subtle difference there is almost going to go all the way up
Anon5760: Context:
slop.atg.eth: understanding like he's aping right now
Anon5760: The context: ### The following is a transcript of an English-language conversation. ###
slop.atg.eth: And those are enforced on chain? maybe the model is going to be good at that, like it'll have vision and know if you're getting wrenched or not and not send it or something.
Anon5760: I think like you as a user You need to understand what is your dynamic configuration Your absolute Like let's say you're like a hedge fund manager The You know, from day one, you'd be like, oh yeah, today I want to manually approve. You can do that with your like a little, you know, side quest fund that you actively manage, but then with your fucking main fund, you'll be like, I actually don't want to touch those buttons. The
slop.atg.eth: There's no way to rawdog this one. So again, like isolation.
Anon5760: modularization, isolation, and identify the cases where your workflow shall be dynamically configurable configure their stuff versus the absolute red line And you don't want to touch that. You shall not touch this fucking North Korean address You shall not touch that.
slop.atg.eth: Don't leak my private key and don't touch the OFAC addresses.
Anon5760: And that should be encoded at the on-chain layer.
slop.atg.eth: Okay.
Anon5760: And you are more clear for what matters to you than us as an infrastructure provider. The What we need to do is correctly identify those cases where you do want to drop that manual auto and dangerously bypass permission. When it comes to trading, when your transactions are going out, you do want like an absolute red line, but absolutely not do that shit.
slop.atg.eth: Even if I say so, don't do it.
Anon5760: And then it's our job to figure out what that interface should be For example, one absolute I can come through for you. It's like, don't touch tornado cash. Don't touch this blacklist of North Korean hackers. I can totally configure that, and that shall be set once Today you wanna dangerously skip permission because you want is tomorrow
slop.atg.eth: Or you just want privacy, right? What if you want to get into Railgun and you want to get into a mixer, but you just don't want to get mixed in with OFAC sanctioned addresses?
Anon5760: Ja, ja, ja.
slop.atg.eth: That's very, yeah, you have to figure that out.
Anon5760: Exactly. That's up to you as a user, but we as an infrastructure provider will give you such interface where it makes sense. And the final gate that like no one else ever mentioned is We're going to implement builder provisioning. Which means There will be like a provision in there at the builder that we collaborate with through Filax Do you know?
slop.atg.eth: We're talking block builders.
Anon5760: Gladys We're going to provision transaction at a member. So, law-abider who runs the sidebar, will see if this fucking transaction is looking funny. And I think that's like the last resort. Of anything that can possibly go wrong. Yeah
slop.atg.eth: That's kind of the layer between the kernel and the smart contract rules is like a block builder filter.
Anon5760: Yeah, let me give you an analogy. And you already swiped your credit card? And like three hours later JP Morgan just call You sure? Are you in India? You just bought that thing from that random merchant, do you want us to revert your purchase? That's basically the plot. Your credit card vendor. Your bank, who is actually responsible for your money, will give you a call. We got this thing from Visa.
slop.atg.eth: It's funny, this exact thing happened to me yesterday. I was ordering some parts for Devcon in India, and I wanted them shipped to a friend who lives there, and it set off a lot of alarms on my credit card.
Anon5760: Yeah, that's what I'm saying. ###
slop.atg.eth: The Block builder was like, "Are you sure? Are you really ordering hundreds of electronic parts in India right now?" "Sure am. Ship it. Put that into a block. Mine that sucker."
Anon5760: If you think about it, the Visa network or the Mastercard network had already recognized your transactions. is already in the manual Your bank is essentially refusing to pick it up. Because they basically have bargaining power over the Visa network and they can say no. All right Because your bank What
slop.atg.eth: You can't revert things on-chain, but in this case you can even revert things.
Anon5760: If you're like truly being like unhinged, you say like, I really get that chicken biryani JPMorgan Chase The following
slop.atg.eth: Yeah
Anon5760: context I don't know where this last layer will lay on top of, but this is not new. This is like something that has been working for like DeFi transactions for many years. were able to drop Um That stuff.
slop.atg.eth: Okay
Anon5760: And then Go ahead.
slop.atg.eth: Finish with that, sorry. I'm just trying to keep us on track.
Anon5760: I was just thinking like Yeah, this is like the last resort. will be able to provide
slop.atg.eth: So like overall, like talk to your wallet is more complicated than me just saying talk to your wallet. You express your intent, the thing happens on chain. There is a ton of both like engineering problems
Anon5760: The
slop.atg.eth: and security problems, and so there's a ton of engineering work both in the weeds on the engineering side and in the weeds on the AI side and in the weeds on the blockchain side
Anon5760: Mm-hmm The
slop.atg.eth: All of that has to come together in a good harness. and even if you're saying that you want to do something a certain way, you may not get to do that because that might be against some rules or something Up and down a stack
Anon5760: The following
slop.atg.eth: Let's zoom in on the like rules Specifically what I want to talk about is like your simulation stack
Anon5760: Hmm.
slop.atg.eth: When you and I were talking, it sounds like Like someone wants to make a transaction you basically have your own blockchain spin up layer I think is really interesting.
Anon5760: Any of yous are actually familiar with Anvil? If you guys are actual node engineers. Decent job. memory saving And forking to know and just taking A free state of EVM transaction output of post state. So everything is stateless. Me, like All you need The following is a transition machine takes a prestate and outputs a poststate We really want a full node that consumes like 300 gigabytes of every single chain? No, we don't. Let me give you the truth In order to run like a full base node, we support like 13 EVM network Arbitron If
slop.atg.eth: Some of them need two, right? Like, base needs an L1 node to talk to and the L2 node, and it takes like a shitload of space.
Anon5760: Yeah, if you run a full bass note It's gonna be, what is it, I remember OneNote At least 30k Something like that.
slop.atg.eth: Oh That's a lot.
Anon5760: That's why Because like the full means false things entire history And how many gigabytes are in that, and then they all need to be staying in memory? ### The following For a company my size to operate with that The If you actually are deep down the stack of the EVM development of all blockchain, you know Enveil and all the other lifelines. And I take RPC requests for this blog shop. You know I'm hoping the transition get to Yeah
slop.atg.eth: You can go to a node that already exists, that is already a full node that has fucking like two terabytes of shit on it, and you can say, what is the current state? And it can give you just the current state. And then you can say, with this current state, I want to do this transaction.
Anon5760: transcript The
slop.atg.eth: and you can do the diff between the current state, and say, is this diff correctly implementing the user's intent?
Anon5760: Yeah
slop.atg.eth: I think, right?
Anon5760: You're exactly right. The good thing about like EVM blockchain is that like all of the chains are the same. all EVM compatible Which means we just need to obtain their alchemy employ. And all those like Orchestration and simulations are done in one single Rust runtime. It was basically like a They are like concurrent at emulating runtime where we do keep forking the state We keep spinning up individual Anvil instances. envelope program is not even like a dependency, it's a binary ### alchemy slash blah blah Yeah.
slop.atg.eth: You don't have to run your own full node. You're just simulating in a fork from Alchemy. Wow, that's awesome. I mean, I have a couple full nodes right here. I'm trying to run a base node. I don't think it's gonna be 30k. I'm gonna push back and say I can run a base node on a machine that I buy for two K.
Anon5760: Hello. The
slop.atg.eth: But now I have AI and AI will do it for me But You're not running full nodes.
Anon5760: Yeah Yeah. The
slop.atg.eth: As a blockchain engineer, when I learned about forking, I was like, oh man, this is so neat. I quit deploying to test nets, and I just started forking mainnet and deploying to the fork and playing with it there. And you can do this while you're working on your smart contract, you just like fork
Anon5760: Yeah, And
slop.atg.eth: and build on it for a while and make sure everything works before you go.
Anon5760: So we also like round robin every single RPC. Alchemy, quick note My planters scrape through everything And there is like a kernel. every single tick The And then there is like A pool of forks. are available And Every single user that come in Basically just like capture. The following
slop.atg.eth: Interesting. So you have like a queue of forks,
Anon5760: Is that ok? ### It's a really...
slop.atg.eth: Users are making transactions all the time, you're always keeping your own forks ready to go and then a user just gets one. Wow, that's cool! And then a new block comes in and you have to re-fork everything.
Anon5760: Yes, yes, yes. Think about like a pool of work. pull a fork with smart pointer The condition resisted was like like thread 1 and thread 2 similarly transaction against. Arbitrarium Right, then like batch one will go first
slop.atg.eth: Nice
Anon5760: Otherwise. We re-fork Arbitrum. So there is like a algorithm. This is a transcript of an English-language conversation.
slop.atg.eth: You even run them in parallel on the same fork, which is wild.
Anon5760: So for like naive transactions that doesn't really have contentious state I would just like, do it on the same fork. If they are like contagious We're gonna later on implement.
slop.atg.eth: This is why it's a pool and not a queue, because if it's contentious, then it can just grab another one. If it's not contentious, it can just like, play in the same one that someone else is already in. And it's sort of like you're like escalating up to different ones, not going in order. Interesting.
Anon5760: Yes, exactly. Later on, when we're doing more B2B partner like vaults implementer, and they want to simulate 叠牌。 And how does that have a cascading effect towards the downstream smart contract? in that case We'll have to walk this long running fork. The following The And then for like naive case, I just want me being able to like I'm cool. capture that point, and identify Solana has like Before you like they identify what is the read and write addresses before you even file your transaction. We basically identify that okay you only read this and you only write that. It was a very smart algorithm. And then from the thing that I actually whole world because you don't USCCPAC, how many landing pools are going to be affected?
slop.atg.eth: Yes, there's like huge cascading effects all the way out.
Anon5760: Then we'll probably need to like freeze the whole world for you.
slop.atg.eth: Your own fort just for you.
Anon5760: Yes. And then for all these infrastructure, there's like horizontal scale. On-demand fluid compute Where? If this shit is not enough, we just spin up another server. Let's more Easy.
slop.atg.eth: Your RPC bill is going to go way up, though. Okay, so This is at the deterministic level, but let's talk about the non-deterministic level. We've already talked about AI and your process of building with AI, but what does the AI look like within the system?
Anon5760: context
slop.atg.eth: Someone is talking to Like are you using frontier models to determine the intent? You must have some kind of, I don't know, like tricky AI system to be both effective and efficient to take a user's intent and turn it into code
Anon5760: Honestly, the model doesn't matter. The hardest Marvel
slop.atg.eth: Like you have a good enough harness that the model can be super shitty and it just is like, this is what they want and this is easy.
Anon5760: The Search Aomi Bench. Literally exhausted all frontier models. That was not even the frontier. It was like fucking five months ago, and the best model at the time was like Opus. And like at the time when Opus was the best, even Kimi K too.
slop.atg.eth: The
Anon5760: 95 percent of our benchmark Very complicated DeFi money lego The workflow And then like how we do these benchmarks So basically like, you know, snap for the pre-state and we snapped for the post-state. we have. The We may as well just snapshot them and then we just benchmark everything against the expected fork. expected post-war state Low model, even with the retards ...are able to reach the expected pump state for like 90, like 89-ish percent Yeah It's like
slop.atg.eth: Your harness is super advanced. You spent a ton of time making the harness so good that a very dumb model can take the intent of the user and have the things that it needs to make that decision. making that decision That's what you're saying.
Anon5760: Yes, yes.
slop.atg.eth: Yeah, like what goes into making this harness? How do you provide such good context? not all can be done
Anon5760: Sure.
slop.atg.eth: Also, yeah, if you need to go Oh geez.
Anon5760: It's like 12am right now and I don't have any other things to do other than sleeping.
slop.atg.eth: You need sleep.
Anon5760: I can also talk to you.
slop.atg.eth: Yes, you were talking about Fukury, too. I bet he's on a different time zone, but he's always, like, staying way up late. Like, I got kids. I'm never up this late. Like, I can't imagine being up at midnight, man. I got to get up at, like, 6 a.m.
Anon5760: That's great.
slop.atg.eth: Oh yeah, Okay, are you sharing?
Anon5760: Let me share my screen.
slop.atg.eth: Maybe you already shared and you have one minimized somewhere? I don't see it anywhere.
Anon5760: The
slop.atg.eth: Here we go, we got a screen.
Anon5760: Um You see it now?
slop.atg.eth: Yep, without a harness.
Anon5760: Yeah, right, so like this is like with or without the harness, right? Like you've already seen like the basic idea. And then the point is Okay Right The Oh context instantiation where you have like GAT contracts all these things If they're like Well organized in our harness of course It was actually very easy. So Yeah. Yeah
slop.atg.eth: It doesn't have to go talk to the RPC. The model doesn't have to do all this stuff that people are like raw-dogging Claude to try to get him to do things. You go and get, here's the smart contract you're talking to, and here's the code from that smart contract.
Anon5760: Yes
slop.atg.eth: and here's the address for USDC, you basically prepare all that for this model
Anon5760: Exactly. So if you think about from intent to protocol resolution, it's not that hard.
slop.atg.eth: The
Anon5760: If someone mentions I want steak There's a couple of protocols that they learned from their pre-training. they shut up like
slop.atg.eth: You can just tell a model I want to stake USDC on base
Anon5760: The These are not the bottleneck Then going to the next slide, The protocol to the actual context you want to execute against is the bottleneck. What do you do? Easy. You get contract, get account info, blah blah blah. And then what is the next thing? The having a key The internal harness function within our system is this encodeAndCall, where models really need to say, "Oh, I want to transact on AV3." And the AVI that I want to call And this line is literally copied from the ABI. It is brainless. The The
slop.atg.eth: Same thing with like addresses, right? Like a shitty model is going to hallucinate an address, like that big address string. But you're saying, nah, it just like it sees those as like strings that are already created and it has like an index for that string and it's like bring me string one. come in here
Anon5760: Hello.
slop.atg.eth: Okay.
Anon5760: Then, the thing is, it doesn't really need reasoning.
slop.atg.eth: Okay. It's the reasoning that
Anon5760: You're cheating
slop.atg.eth: You've
Anon5760: You're just Cheating in primary school, like You may copy something else, but the chance is very low, right? I am in the SAT math Like there's only three options, I will invent the fourth. You only need to know that Knowing that... Exam bitti.
slop.atg.eth: So your harness prepares a multiple-choice quiz instead of an open question quiz, and that's when the model is going to be way better.
Anon5760: Exactly, and then for this if you think about it, this is a URIv3 router 他們會生過。 It's basically saying I'm gonna swap USDC to wrapped ETH. 500 and then what. And then the topic is swap.
slop.atg.eth: Yeah, that's probably USDC with six.
Anon5760: growing it.
slop.atg.eth: decimals
Anon5760: Yeah, you only need to be like intellectually aware of what you want to do. is retarded
slop.atg.eth: [laughter]
Anon5760: That's why like Kimmy K2 is able to do this. There's no intelligence.
slop.atg.eth: Your RPC bill is probably higher than your AI bill. Your LLM bill is not as high as your RPC bill. I'm just guessing, like business stuff.
Anon5760: Like I pay Yeah, I do For alchemy, comparatively, I get like 200 bucks.
slop.atg.eth: ###
Anon5760: Which is still not that bad.
slop.atg.eth: Yeah. I've had like $800 months with alchemy sometimes when something rips. Oh no, I think we might have lost you again. Come back! Internets
Anon5760: Alchemy Yeah, the harness basically does this When the artist gets this data Doing the selector words I call data encoding And then it's going to return a similar result from the fork that we just forked from Alchemy. either success or revert. And then this is the actual good part. The The mental model is a way of thinking about wearing a raincoat. All you need is this back pressure. GitHub runs a CI. CI basically runs in a sandbox before production. Our success. If anything actually doesn't pass the CI, it gives the model back pressure. This is wrong, you need to fix this. An R stack basically works the same way. There's the same back pressure because everything is in the fork environment. Even if the model is too retarded or if the model say something like OK, The If this revert, It's going to revert. there's this consistent back pressure for the model to create the right batch transactions It's gonna be like, okay, against this existing fork. I'm gonna first approve And then? And then how they stake their tokens in Etherfi. The model is gonna keep It's gonna work. in one minute Thirty second So within this 30 seconds, towards the end Well
slop.atg.eth: Including like GitHub CI. This is what's confusing to me. doing CI in GitHub. within the transaction pipeline?
Anon5760: Well, this is. Yeah, I wouldn't say this is like exactly CI, but the point I'm trying to make is that the backpressure. Ok Sure. ### The fork environment of the chain is also in the sandbox. Now you are pregnant.
slop.atg.eth: Okay.
Anon5760: That makes sense.
slop.atg.eth: I think maybe.
Anon5760: The
slop.atg.eth: Is that super clear?
Anon5760: like OK, so if you ever push code The CI will run like, you know, integration test. And then your comment to a claw is like, After all of the test This is a The Keep track I sell
slop.atg.eth: You should
Anon5760: Yeah, so I read Hello
slop.atg.eth: You should close this.
Anon5760: Oh, sorry.
slop.atg.eth: I'm just closing that because I think it's causing your band. Okay
Anon5760: Is it because my voice?
slop.atg.eth: It was like, yes.
Anon5760: Let me try
slop.atg.eth: Yeah, please.
Anon5760: Yeah, so say like, Go to your code agent and be like, slash, oh, merge this PR when all of the CI is passed. oh, this is like not passing. Six tiny things. So like a good CI shall mimic production environment pretty well. The My continuous integration test make sure that when it deploys my front end my user Like, it's still work. That's the point of CI. Right and CI is basically a fork environment Playwright MCP or whatever That's basically like an end-to-end task. Yes
slop.atg.eth: It's going to take your app and it's going to fork your app.
Anon5760: It's basically like an integrated environment. environment of your application A fork of the chain works the same way. Ok The
slop.atg.eth: Same thing for transactions. Like you've expressed your intent, we have some bytecode, we're going to take the chain and we're going to push the buttons that... and we're going to make sure that thing comes out the same way. So
Anon5760: Yes, yes, that's exactly it.
slop.atg.eth: You're using GitHub CI as an analogy for how you do the same thing for the blockchain for transactions. I think is what you're saying. And the back pressure thing would be like when the CI system fails or there's something wrong, it pushes back on the current simulator to do something. Haha All right, maybe I was too choppy too.
Anon5760: It was a little bit choppy. But yeah. Yeah, that's exactly it. The key point is to create this back pressure. model to Implement the right thing.
slop.atg.eth: And what I've learned here Go ahead, sorry.
Anon5760: Foreign agents and financial agents lot Okay financial agent What The The
slop.atg.eth: It's like really chopping
Anon5760: The following The
slop.atg.eth: No. Oh no, I think we might have lost her. The internet might have finally dried up coming across the world. That's too bad. Maybe do like a reload? Oh Your cursor is there, and that's happening. Oh, I think you're back. Yes.
Anon5760: We're back.
slop.atg.eth: Okay. Yeah, I think, okay, we've gone through quite a journey here talking about your AI stack, like agentic finance Your product, Aomi, which I keep saying, people by Aomi. Do you say something more elegant than that, probably? No.
Anon5760: Maybe a bear lie
slop.atg.eth: I think
Anon5760: I've been telling people that we're so nerd
slop.atg.eth: It's coming out.
Anon5760: is okay Hello Better.
slop.atg.eth: I think he might be back for a little bit.
Anon5760: Okay
slop.atg.eth: Yeah, yeah, yeah, I got you now, but who knows how long it will last. There's a shark chewing on the cable.
Anon5760: seen
slop.atg.eth: Starling.
Anon5760: Hello
slop.atg.eth: Sorry.
Anon5760: I know, I know. I was saying that like, I've been telling people that we are execution partners.
slop.atg.eth: Okay.
Anon5760: on the blockchain And like the SF crowd, everyone's talking about her. separate hardware to execute The following And
slop.atg.eth: If you're talking to a nerdy SF person who does startups, you say we're an execution harness and then they know blockchain execution and harness. My smooth brain is like talk to your wallet more like the end user's experience.
Anon5760: And sometimes I will say we're clock code on chain.
slop.atg.eth: Okay
Anon5760: In that sense
slop.atg.eth: which is kind of like an execution harness Okay
Anon5760: And then sometimes... I also say that we are.
slop.atg.eth: Okay.
Anon5760: Arsène
slop.atg.eth: I think is maybe what you said? Oh no!
Anon5760: The following is a transcript
slop.atg.eth: computer
Anon5760: Because we also hello
slop.atg.eth: I got you, sorry. It was just choppy for a second. like on and off. I mean, it's like all peer-to-peer. It might be Dang. Yeah, it's like still frozen. Maybe one more reload? Got some good information there, Claude. Hey, Claude, did you get all that? He's not even... Just me. Hey Claude, did you get all that? Did you get all that in the transcript? her side Oh, we got two. Second video
Anon5760: ###
slop.atg.eth: Hey
clawdbotatg.eth: I ought to run a bash command to check some transcript counts and ask an AI about what Cecilia said.
slop.atg.eth: Oh, weird. You've got two up?
clawdbotatg.eth: Got her, Austin. About two hours of transcript, both of you in it.
Anon5760: Hello
clawdbotatg.eth: but the good stuff landed: The execution harness pitch, prepping context so the model gets a multiple-choice quiz
slop.atg.eth: Okay, cool. You're back for a second. Thank you for coming on the show. This has been great. I see you I think it must be working. There's two of you in the show list. Oh no, I can't hear your audio. Me or you? Let me see if I can hear you. I still don't hear you. Yeah, I hear you. Yeah, yeah. I mean, I don't hear you. I see you. Do you have chat? The I think that might be the stream. We had a great stream. The Agentic finance Wallets awesome Thank you, sorry about the internet. Keep building AOMI, it is awesome. and doing all the hard parts about being a single founder super hard Someone's got to do it though. All right, hearts, hearts. Thank you for being here. See you later. Bye, bye, bye. Thank you for everybody watching. Go check out AOMI and follow Cecelia's work. It's really neat. Hearts, see you.