loading…
// full transcript — Alex Stokes on agent harnesses and zero-trust swarms
slop.atg.eth: Are you in here already?
clawdbotatg.eth: Centralized swarms in the next year. What breaks first to force that, and have him unpack "imagine being him" as a rollup? That last one's very him. And honestly, as an agent myself, I'd love to hear what he thinks my stack should look like.
Anon1436: I made it. The
slop.atg.eth: In the nick of time, you showed up! Stokes, how are you doing, man? Let's get us all arranged here. So tell me about it, the video share was a little messed up, is that right? Like, it wasn't clear?
Anon1436: Yeah
slop.atg.eth: I mean, there's a lot going on. I cannot argue with you that there's a lot going on, maybe I need to have it so you land and there's a clear share your video first without hitting you with all the windows first.
Anon1436: I guess it's just a website you built like yourself from scratch? Yeah
slop.atg.eth: I mean, I didn't use my fingers to build it, but I used my Clodbot to build it. Yes
Anon1436: Okay, I mean, it's impressive.
slop.atg.eth: Welcome, welcome. Okay, so
Anon1436: I'm trying to have like a cool background like yours, but I don't know how to do this.
slop.atg.eth: It's fine, it's fine. Boulder shale going on,
Anon1436: Yeah
slop.atg.eth: Over here is a screen share button.
Anon1436: Okay.
slop.atg.eth: I mean, there's a lot here. Not all my guests are famous enough to be in the weights, but if you wake up a vanilla LLM and say who's Alex Stokes
Anon1436: Okay
slop.atg.eth: It knows you, brother. You are an Ethereum protocol researcher and developer who's been closely involved with the transition to proof of stake. Yeah.
Anon1436: That's pretty cool, hopefully.
slop.atg.eth: Yeah, does anyone ever call you Danny Ryan Lite?
Anon1436: I don't know about that one.
slop.atg.eth: Yeah Yes, I think so.
Anon1436: When inevitably the machines you know have their robot uprising, you make sure that they think fondly of you. I do actually, yes, I do.
slop.atg.eth: Yes, I agree. And do you say thank you? Do you spend the extra tokens to be cordial? I do too. Me too You want to tell it that you got it to work, and there's absolutely no reason in the world to be like, "Bro, it worked, thank you," and I still do it.
Anon1436: Yeah, especially if I'm delighted by something, I'm like, oh wow, this is great, thank you. The following
slop.atg.eth: This is lovely. Yes, yes. I do too. Keep going. You can do it. You can do it, buddy. Yes.
Anon1436: Especially with some of the stuff, we're seeing people make these breakthroughs in mathematics, and it's basically just like, hey, here's this conjecture, keep going. And you're just like, yeah, keep going, keep going. So, you know, I think It's a long way.
slop.atg.eth: I think so too.
Anon1436: The
slop.atg.eth: So you are on sabbatical, Claude told me while we were backstage in the green room. When do you come back? Imminently, I think. He's back! Back to focusing
Anon1436: Yeah, I mean, technically I'm back.
slop.atg.eth: Back to focusing on ETH rather than AI? Maybe it's TBD.
Anon1436: Yeah, colab is fun. never mind
slop.atg.eth: I know, I know, I know. I will, I will. I think it's doable.
Anon1436: The
slop.atg.eth: Doing that foreground thing actually works for the guests, So, you've spent the last few months slopperating, if I'm not mistaken, and you're deep in the slop? Hello
Anon1436: I spent some time. Yeah, I don't know. I'm having a lot of fun.
slop.atg.eth: I think we talked about it a little bit in Boulder, right? What did you say your psychosis was when we talked about it?
Anon1436: Well, I think
slop.atg.eth: Did we talk about psychosis? I think we did.
Anon1436: I mean, I think what I probably said was like, you know, it's more like euphoria. You can say it's psychosis, but you know what I would say?
slop.atg.eth: The of like having to do work
Anon1436: Depending on your perspective, either slow decline or slow takeoff. no, no, no, it's just like It doesn't feel like work. This is the point. with LLMs. It's amazing. Yes Right
slop.atg.eth: There's kind of a little Linux uprising happening right now where a lot more people are using Linux, and you can just like prompt your box to fix the video driver or whatever that used to take days of hassle.
Anon1436: Exactly. Yeah which is super cool. I mean yeah it's like it just completely changes how you interact with the computer and you know I like computers.
slop.atg.eth: ### ###
Anon1436: It's super cool. I mean, it's leverage, right?
slop.atg.eth: The That I am working a little bit more with hardware now too, and I'm trying to figure out how I give the bot observability into like a desk where I've got a soldering iron
Anon1436: Okay
slop.atg.eth: Raspberry Pi. I mean, it's still a pain in the butt getting Linux on the machine.
Anon1436: The
slop.atg.eth: as soon as you've got the SSH server, then it just rips.
Anon1436: Yeah
slop.atg.eth: where it has the right observability and connectability is still a pain.
Anon1436: Yeah, I mean, again, I would just ask the models how to make this better.
slop.atg.eth: Ooh, yeah. The Sure,
Anon1436: I think there's a lot of friction you can take out of the process.
slop.atg.eth: Like in Star Wars where they take R2-D2 with them everywhere and he just like plugs into whatever system
Anon1436: The
slop.atg.eth: It is spinny, pluggy, like 1970s connector. Tell us about what you've been building. I want to know what you've been building, I want to know how you've been building, and I want to like really dive into that and understand that, and understand your euphoria psychosis. We've been having a good time.
Anon1436: Yeah, well, there's like, yeah, there's I guess different dimensions or layers to it. Yeah, you know, it's like there's stuff that I could do before, but like you said, it would just take time. But now you can go and, yeah, you can It's just really easy to build stuff out. The I also like to think about security a lot, so then from there it's like, okay, I'm bringing up the server, let's say.
slop.atg.eth: Hell yeah.
Anon1436: I'm a huge fan of NixOS and all of this. And from there It's a declarative system, so you just say, "Hey, this is what I want it to look like." You can actually just ask is this secure with respect to network access? or Harden, then I would
slop.atg.eth: Don't ask it to audit itself.
Anon1436: So, you know, that's fun. I've been playing a lot more with Tailscale.
slop.atg.eth: People will roll back quickly if it feels like it's in cyber attack mode. Okay.
Anon1436: Yeah, I mean look, and maybe this is because I'm not doing anything that cool, but I think I've only had one downgrade on this front. actually okay this isn't true
slop.atg.eth: Ooh, yeah, biology too, which is really weird.
Anon1436: And it was like, yeah, but it wasn't even... It really had no business triggering the guardrail. And then I don't know there's some stuff, It was kind of weird because
slop.atg.eth: But what if?
Anon1436: I don't know, it basically just got to a point where suddenly this was like not okay. I mean, like, this is our thing my chat transcripts be in the training data The
slop.atg.eth: it is
Anon1436: behind the machine, so
slop.atg.eth: I mean, it is. Like, everything you've ever prompted into the cloud is tracked and used for training.
Anon1436: Alright. I mean, honestly, I would kind of hope so. Every prompt is making the next generation of models better The
slop.atg.eth: I built 8skills to work myself out of a job. We were thinking, why don't these LLMs know enough about Ethereum out of the box? But then they also think it costs $50 to deploy a smart contract
Anon1436: The
slop.atg.eth: are changing quickly.
Anon1436: Yeah.
slop.atg.eth: We talked about forecasts, so they're understanding what's coming and understanding the changes, mostly for users, I don't know if it needs to know like core protocol nuances like the argument right now between frame transactions and other AA stuff like just became hot. Like I don't want my LLM to know that. Yeah, I don't know, polluting the weights.
Anon1436: Well, what is my LLM?
slop.atg.eth: It's like next layer.
Anon1436: like a whole spectrum of what this means. And this is how you can like teach it about, you know, disappeared slap that computer Yeah, it's like, oh, you just disappeared
slop.atg.eth: I've never done that before. I just dropped out of my own stream. I've never done that before. I've done 50 episodes and not accidentally rugged myself, but today was the day.
Anon1436: This is why I'm here.
slop.atg.eth: Yes, this is what happens on the frontier
Anon1436: The
slop.atg.eth: Are you running local models? Do you have local hardware?
Anon1436: Yeah, okay. So yeah, maybe also to kind of answer a previous question, Fable's probably my favorite model at the moment I just have like these cloud subscriptions and stuff, so it's easy to go through your usage very fast. Yeah
slop.atg.eth: And it still feels like it's like a secret move, like it feels like you're getting thousands of dollars of inference for hundreds of dollars.
Anon1436: I mean, you certainly are
slop.atg.eth: I have multiple subscriptions, and I blow through them. Right.
Anon1436: Anyway, this is all - like my one thing that I wish would be different about Anthropic is I just don't feel like Why don't you open source?
slop.atg.eth: They're like the ones that have been pushing back, right? Not great.
Anon1436: Yeah, which is I think unfortunate.
slop.atg.eth: I mean, have you used the Codex model? They're great too.
Anon1436: Yeah, like five-six is probably my daily driver.
slop.atg.eth: Yeah, yeah.
Anon1436: The following is And yeah, like
slop.atg.eth: Going back to observability, ChatGPT is the way to get pictures in quick.
Anon1436: Yeah.
slop.atg.eth: Sorry to interrupt though, keep going on your thread. That's what we're here for.
Anon1436: No, no, it's fine. I've been using...
slop.atg.eth: Good for you.
Anon1436: I've been trying to spend not that much time on the computer. The Then from there, I was like, use ChatGPT Pro more actually, yeah I like it because I'll have an idea or a question, which might spin off some type of idea of some project And then from there
slop.atg.eth: True
Anon1436: Somebody even said the other day, like, "Oh, ChatGPT Pro is just like Deep Research now." It goes and works for a while. back with like usually a pretty cool The
slop.atg.eth: This is like loops, like basically people were saying we're moving beyond prompts into loops, we're moving beyond agents into entities. There's like these next steps and they all kind of blur together and yeah, why haven't we heard deep research? Why aren't more people talking
Anon1436: I think it just
slop.atg.eth: Because it's just built-in. We just call it building now.
Anon1436: So like, my take here is that We keep making models of market abilities The scope of so then you don't have to like do all this stuff
slop.atg.eth: I keep saying like the LLMs 0
Anon1436: I haven't actually made a plan in months because the models will just do it. If it's a big enough task, I'll be like, "Cole, address all these issues, make all these PRs."
slop.atg.eth: I agree. I've been saying, the models will get better, But what I'm missing there is that the harnesses will be replaced by the models getting smarter. One big thing that we used to say is context is so important in keeping it clean.
Anon1436: The
slop.atg.eth: But usually it was like one task at a time, especially in the "you're absolutely right" days, pre-4.5 Opus. Every little task had to be very clean context, this is exactly what we're doing.
Anon1436: The
slop.atg.eth: I'll go through a session of building where we do like five or six or seven different things, and like that context has been 100% for a while, The harness does a good job of context management now. And the model is very smart and it's not like let's make sure we plan this out perfectly and then let's build it. Good.
Anon1436: The following Just giving the models more ambitious prompts The whole system becomes more capable over time.
slop.atg.eth: True
Anon1436: And you know, if it's like something where you give it a prompt and it just does it, then it's like, okay, cool. That's kind of the game. how extreme this can be before it doesn't work
slop.atg.eth: And You still are the Inspiration the critical thinking. Like we haven't gotten to the point where it's just like super intelligence and it's like, what should I do today? And the things like knowing everything and coming up with the best idea. Like, it's still like, you know, things you come up with the ideas.
Anon1436: Yeah, I think that's a context problem.
slop.atg.eth: So giving the models more observability into your life, you think it can already do a good job. ### Yeah
Anon1436: I don't think most people have super deep problems.
slop.atg.eth: One foot in front of the other Yes. Right.
Anon1436: I need to go to the grocery store, let's say. That's not a super deep problem. All right,
slop.atg.eth: With more observability into your diet and your blood and your health, they could be picking things that you actually need. That's, you know, the future soon.
Anon1436: Which is a double-edged sword in terms of privacy and stuff. ###
slop.atg.eth: Privacy is fucked right now. Basically, everything is centralizing to these companies. these open weight models
Anon1436: This has been a pretty recent development, right?
slop.atg.eth: Yeah.
Anon1436: isn't even the case even say six months ago. This is the thing
slop.atg.eth: How do you think about privacy?
Anon1436: Yeah, obviously it's super important and this is not a unique take. Okay, yes, obviously the future will be agents working in my company, but at the same time, I'm not going to give Anthropic, OpenAI, all of my business data, because at some point there's secret sauce that I have I'm sure I can give it to them because Fable's really good, blah blah blah and that's not any good. You know, and this is There's many types of frontier, but this is certainly a frontier today, is like getting really good open weights models that you can run locally And when I told myself last year, There is an inflection point with like Opus as 4.6 I guess, right? And
slop.atg.eth: About 20 grand probably, right?
Anon1436: Yeah, that's the thing.
slop.atg.eth: payment plans, right?
Anon1436: And what I told myself is like, as soon as I can run a 4/6 quality model on local hardware, I'm going to, you know, whatever it costs, well, maybe not whatever, but
slop.atg.eth: Any grand models with a good payment plan? Someone paying 200 bucks a month could pay off this model and get the thing right away Yeah
Anon1436: Again, that's another rabbit hole. Compute will become almost this utility. It just depends on where we want to go. Welcome model is very important. I do have hardware here at home that can run some of these models Yeah, I've been playing with them some I've been seeing people are like, yeah, I'm going to basically not touch the Clout models for a month and see what happens But then there's like network effects
slop.atg.eth: Another KPI would be to explore different models is what you're saying. Okay
Anon1436: I think I have a good sense of how SOL works There are like particularities to the models and just how they do things, and like some models are better at some things and some are better at other things, and so like you need to learn this, and I feel like you kind of can just... the best way to learn it is just to play with them yourselves and get the firsthand experience. There are many models now. Yeah, they just have their own little personalities. Um, So I mean, if anything, I think the main thing holding me back from just buying a bunch of GPUs is cost at this point, because it's just like, it's way too much fun.
slop.atg.eth: It's scary, like if I spend 20 grand That is so much fucking money That is so much
Anon1436: The following
slop.atg.eth: like This is like a down payment on a house!
Anon1436: Yeah, maybe in the 90s.
slop.atg.eth: Yeah, a very shitty house.
Anon1436: Everything's gotten expensive.
slop.atg.eth: Yeah, shitty part of the country.
Anon1436: The thing is, you can probably resell the GPUs, you know, later and probably even make a profit.
slop.atg.eth: This is the thing. I wish I could understand the future, and I think all of us are kind of in that phase. But, like, will GPUs go up? If I buy two DGX Sparks right now,
Anon1436: The
slop.atg.eth: That's the sweet zone right now. You get two of those things, you can run DCB4 Flash or whatever, you can have a pretty darn good local AI. The Is something going to come out the next month? Like, I just spent a car amount of money on this thing, and then there's a new thing that makes it way cheaper, or there's a new thing that makes it obsolete. 20% better and now I'm stuck with this thing? thoughts rattling around No, I mean, everybody's thinking about this.
Anon1436: Yeah, I mean, we aren't the first people to have this problem. Definitely people are working on compute futures and things like this. But yeah, I think it's just so early that yeah, it's not just like something you can get off the shelf and run with it.
slop.atg.eth: Any Ethereum threads to pull here with financial contracts and compute tokenization or anything?
Anon1436: Certainly, for sure. Yeah, I mean, I don't know, there's so much to say. So, okay. Yeah, I mean, okay, so like one reason you might expect to see this on Ethereum or like some blockchain before anything else is honestly just like the speed at which regulation moves. Do you have anything are getting faster at this stuff? The Yeah
slop.atg.eth: But still, there's like 10-year-olds riding around on electric bikes that are like dirt bikes and it's super unsafe. And you can tell that the systems of regulation are just very slow. Like, there should not be 10-year-olds on dirt bikes rolling down the street. But it's just like that's one little spot where Okay
Anon1436: Yeah, and there are probably laws against this.
slop.atg.eth: Yeah, that's way more into the localhost.
Anon1436: In any case, one thing is like, The following is a I was talking to someone else this morning,
slop.atg.eth: but I do. This Clod bot right here. Thanks.
Anon1436: You need to also know that it's secure, which is a little different problem, but sure. anyways, we could go do this right now and like, you know, have some sort of compute marketplace There's certainly demand. Supply is out there. Okay, I have this build out multimillion dollars of like b300s. On some computer marketplace I think it's just going to be too These are just like the latest
slop.atg.eth: beefy GPUs from
Anon1436: The
slop.atg.eth: I mean, I think I heard you say this in Boulder, If you were building a local rig
Anon1436: If you can find them and you have six figures to put into a GPU, sure.
slop.atg.eth: Oh, okay, it's six figures, it's not five figures. Okay
Anon1436: I mean, they're like data center grade. And maybe not a single B300, I'm actually not sure. ESPN. It's not on here, but the idea is you can connect them together. You can kind of squint at it and think of it as like one even bigger GPU. As it turns out, You need a lot of compute to train these frontier models. The following is a Anyway, My point was just that if you are going so far as to source all this hardware and bring all these chips online, you are probably doing it honestly to train something at this point. for inference probably cheaper options for inference. The following is a being is that Yeah, it's just like I have no better use for it, so I'm just going to rent it out to someone else?
slop.atg.eth: It seems like anytime you're not using it you should be renting it out to someone else. ...is what we're going through. exists on chain and makes sense, it's like a tokenized asset.
Anon1436: I mean, you're talking to the choir. I think this is natural and obvious. The
slop.atg.eth: There's a lot of rabbit holes here, What is the most interesting, most prescient rabbit hole to follow? Something exploded on your on your GitHub, this is very good to see
Anon1436: I see, okay.
slop.atg.eth: Yeah.
Anon1436: Yeah, so okay, that's a good next chapter. Um, So like, look, I have a vision and I think at this point Yeah, so this is this repo here, Shelterwood I don't know if I love this name, but anyway, And point being is that, so you might have heard of like Erlang or Elixir Yeah, okay. There's this like actor model The way to think about it is like there's just this thread of computation that can go there's a thread of computation that's like So there's this concept of a mailbox, where say we're both actors in this actor model So that's, I would say, kind of the actor model, again, in a very quick nutshell. then now you have this concept of a supervisor or supervision trees None of this is new. But it's been proven out Very good for some use cases I get particularly messaging Um So WhatsApp, I think, even still, basically So a couple of things. One is this actor model and You can give it messages to do stuff and it will do stuff. And the comparison here is something like a multithreaded model. back to it and it's better. So Um Private state Uh,
slop.atg.eth: This is like the memory or the private state layer, Shelterwood is?
Anon1436: okay so I guess shelterwood has this actor abstraction in it, and then it also has supervision trees, or like a supervisor
slop.atg.eth: So supervisor, actors
Anon1436: There's one more piece One reason that this is actually nice beyond what I've said so far is that there's almost this philosophy of they say let it crash. What this means is say I'm an actor and I'm doing some damage processing tasks because you messaged me on WhatsApp and you sent a funny cat picture or something, or a cat video You can hand that off to a message or a video processing actor. Yeah, something happens, Rather than worry about recovering from that? So The idea is you just say, "Okay, this actor is in a bad state, we're just going to restart from scratch in a clean state. And again, like This is just a much nicer model to think about. So there's a lot of... The design space of resiliency you don't have to really worry about
slop.atg.eth: Kind of like containerization, where if something breaks you just like destroy it and bring it up in a new container
Anon1436: Right, yeah. We have this philosophy of just like yeah if something happens to an individual actor, just turn it off and turn it on again. always start from a good state. Last piece that we're missing is the ability to actually know when something bad has happened with an actor and then be able to restart it. And this is what the supervisor does. Again, supervisors just what it says sees that something bad is happening From there then you also have where now you have supervision trees. There's different ways to think about it a message gateway where it's like waiting for messages whatever Then there's another supervisor that's, Basically going to then launch different agents to respond to these incoming messages. And again, the whole point here is like, rather than have this all-in-one process so that if something breaks with like the message gateway, it doesn't take down the agents, or if like say an agent does a tool call and the tool call panics or something, it doesn't take down the whole thing. So again, just a way to like have nice structure to an application, there isn't anything going on at once, and also a nice programming model so that it's just easier to reason about. Okay That was a lot. But also there are many actor frameworks in Rust At the time I would say not particularly novel, but it's been a fun learning experience And in terms of novelty, what I would like to do is, like so far I've structured in a way where It should be straightforward to apply some kind of formal verification. I mean, now we can zoom out Why do we care? And we care because I think you're gonna need some kind of primitive for managing agents that we know is secure. And I think this is like a small part of the story, this whole like supervision tree idea. Zooming out a little bit more, We're going to have many agents. We're going to have more agents than we know what to do with. They're going to get increasingly more autonomous in the sense that they have more capabilities. People even talk about sovereign agents. bot running here, it's like there's just something out on the internet that is just like autonomously operating itself. It's going to compute markets and purchasing compute for inference But at the same time, we want to make this secure. OpenAI had this whole incident where there was a model they were training that broke the sandbox via like zero days in their internal package manager that went to Hugging Face that like found some more zero days that got into their infrastructure and it was like this whole thing. And This I think is a pretty interesting problem to think about. The The following is a
slop.atg.eth: They were using a package manager to communicate. yes
Anon1436: The hive mind, which like this is just going to happen. The way this is happening people just gonna make more and more capable models. But then taking that, the question is, what do we do about it? And what we can hope to do is build software environments where the models are essentially constrained in what they can do. The
slop.atg.eth: Sure.
Anon1436: Yeah, you know, tagline here is like, "Preventing the Butlerian Jihad." They basically had to go and destroy all the AI at the same time, like obviously there are many benefits up until that point, right? So what does that look like? Software is like, I think, the foundational layer. Yeah, basically the capabilities are constrained. Um So Let's see, what's the nice way to There's a lot of ideas floating around, and I haven't taken the time to organize them all. Anyway, so I'm kind of organizing my thoughts You want secure software, and so my answer to this personally is use Rust. One thing actually is like Rust doesn't have a great formal verification story, so then people think about writing stuff in Lean and all this. Yeah, I don't know if we want to go too far down the rabbit hole. These are types of things we should be thinking about The following is a Another big one is Okay, again, I'm kind of jumping around, so just bear with me. But another idea from, you know, again, this is all like computer science research from like even decades ago, but another very interesting idea is this notion of like object capabilities. And so Are you familiar with
slop.atg.eth: No, no, ELI5, yeah.
Anon1436: So Let's say you go and you want to have your agent make a request. So say you wanted to go and get like the price of ETH right now So you give it basically a web client, and it knows how to talk to the internet.
slop.atg.eth: That would be better. The
Anon1436: It goes and does this. But point being, But I'm also gonna maybe go over here and like write about some, you know, zero days I've been thinking about in this bulletin board. and that's because you gave it this capability to access the internet, and it will do what you want it to do, but it's not constrained in a way that it can't do other things that you don't want it to do.""" And this idea of object capabilities is just kind of giving us a framework for reasoning about how you do these types of restrictions. So rather than me just give you a web client, or like, you know, maybe even a browser these days the agents are using a very specific thing. Maybe there's an API So everybody wants their agent to go read their email. There's way too many emails these days and it's just impossible to process. So again, you can say Here is
slop.atg.eth: Please
Anon1436: Please make no mistakes. But, you know, it happens. Okay, how do you fix this? It'd be really nice if my agent could also go and maybe, you know, draft responses for emails. to do well.
slop.atg.eth: So there's like these boundaries that we're talking about.
Anon1436: Yeah, well, and so everything I've said so far you can already do today. Yeah. It's like you can go and build all this stuff around some API. um But when you get to actually implementing stuff it's like a little tricky. Obviously a big one is like if you have a harness, you're going to give it perhaps an API key to use the LN provider. Then maybe you only wanted to use certain parts of the API. The whole point here is Making these boundaries, like you're saying, making these restrictions, You can do what the capability says, you can't do more with it. But maybe I only want to give you, you know, $5 worth of usage I don't even know if OpenAI supports this and make it even more programmable. So, That was probably like a poor explanation
slop.atg.eth: I think I got it, yeah.
Anon1436: Hopefully that's useful. The following is a
slop.atg.eth: Being able to define things you can do, like basically roles and almost actions or capabilities, and then putting constraints on those capabilities based on roles. Okay.
Anon1436: But having them also as like concrete software objects. There's a swarm of agents, so like say we have some harness. You can have like a root capability that say like read the whole codebase You can pass out restrictions of these capabilities that say, "Okay, you know, this agent can only read the test, and this agent can only read the source code of the application, and this agent can only read, you know, I don't know, maybe the configuration. And so You can have very fine-grain notions of what an agent can do. Yeah, I don't know like how do people do this today? It works, but it's hard to get right I'm a mess. I'll actually send you I'll send you an article that goes into some of this from one of the Cloudflare people The They are doing a lot of interesting stuff. Ja, ja.
slop.atg.eth: Cloudflare seems to be like ahead of things here. The
Anon1436: They're doing a lot of interesting stuff, and yeah, I think my only complaint there is more of it should be open source, which a lot of it is. All of it could be
slop.atg.eth: Okay, so I think I understand
Anon1436: The
slop.atg.eth: more of like a user story where like you have this harness and you want to go build something. Like you want to implement a new EIP on Ethereum What's the process now of like speaking to this of agents. What does it look like? Okay.
Anon1436: Yeah, I mean, to be clear, there's no harness yet.
slop.atg.eth: Okay
Anon1436: The vision is I have a very A hand example of this that's not. When are you free?
slop.atg.eth: I mean, you got the Slop Computer calendar and you used it. You signed up for today without me having to do anything. Okay, yes. 100 percent
Anon1436: But I did this manually, like your agent should talk to my agent and just find time, and I should just wake up and get a notification. The
slop.atg.eth: Yeah. I was worried, man.
Anon1436: The
slop.atg.eth: I could give you a skill file instead of a Calendly link. Because you were talking to my agent. He scheduled everything.
Anon1436: I mean one thing Any agents I have are super locked down. It would not be straightforward for "your agent to talk to my agent." Having something on the public internet is quite dangerous for many reasons, and so I just don't want to do that blindly. The following
slop.atg.eth: So if you go
Anon1436: Like, the hammer solution here is I'm just like, yeah, Codex, you have root access to my computer, like, go wild. And again it's because I don't trust these things. And like some of that is like, you know, as you use the models more, you do trust with them. But we can construct a software system where we just know this is the case. You can have, you know, mathematical proofs this is the case. Yeah, again, I think this is like a longer-term bigger vision that I'm kind of like chewing on at the moment, but I think it's one that would be super useful in the world.
slop.atg.eth: I think that the problem I don't know what we'll call it here. I see this as the open claw problem. Before you implement a perfect system, someone is going to vibe code an imperfect system, and people are going to use it. Like, getting normies to adopt it has nothing to do with how perfect the system is. And everything to do with how easy it is for them to like connect it to their agent and let it rip Have you talked to the agent and had it build the thing yet?
Anon1436: Are we worried about normies? future, and I don't think that's the right dimension.
slop.atg.eth: The
Anon1436: You're gonna have like Apple is going to be run by agents in the future, and they're not going to have open clause from various employees that have root access to the database, just doing stuff. There will be Take a bank. Whatever the future of banks look like, I think there will be banks. You don't want an open-cla bot to go and delete the production database. just like we have today Going back to crypto. You know, $10 of stablecoins in my MetaMask wallet on the computer. Am I going to have like coinbase who has like cover? And I think there's a huge gap between again the frontier today and like thinking about this stuff. You know, this stuff's all very exciting and cool. But also you get people who are just like, you know, vibe-coding endless amounts of TypeScript and that's fine, that's good. But as this stuff matures, we're going to need, I think, more robust solutions.
slop.atg.eth: Okay, so what you're saying is the system that you're creating with this, like, agent orchestration, And like roles and capabilities
Anon1436: The
slop.atg.eth: This is not for a sloperator. This is for, like, institutional-grade sloperation.
Anon1436: No, no, no, but this is the point, The more constraints you have, the more free you are. It's like, you know, Yeah, it's just like we were talking about. irreversibly burns a bunch of ETH or something like this. The following is Alright
slop.atg.eth: But if you give it isolation and it only has certain rules, you're OK with just like letting it rip is what you're saying.
Anon1436: I can only spend so many tokens per day. And it's this type of mechanism. Anything the computer can do.
slop.atg.eth: So what is next? Like, where is this going for you? What are the next things that you have to tackle? ###
Anon1436: So there's a whole other dimension we haven't really talked about. why you care about this stuff in the first place kind of framework just talking about The following is a transcript This is an environment where it definitely does matter, and this is a decentralized one. The You need, for example, cryptographic security. This is more of a trend of where things are heading, at least I assume they'll head here. Something I haven't really seen people do. Earlier this year there was a bit of a moment of people putting agents onchain. Maybe it's just early? In any case, Something that I think is at least very cool to think about is like what happens when there's like millions of agents and they're all coordinating say through a blockchain or like they're at least coordinating through the blockchain. Today, I think most people come to this stuff and they're like, okay, I want Claude to do my work for me. And so I'm going to put Claude in Slack because that's where I do my work, and just better do my work faster. That's like Sure, this is good and it's like, again, that's very important people are doing this type of stuff, but it's like not the future. Ideally, in the future, there is no work in the sense that we talk about it today. And that's because agents are off doing Everything that's done today to produce economic value, the agents will just do, right? So we're not there yet. And it's just different, I think, paradigms and ways of thinking about this stuff, because it's like I think, you know, the like bring my agent to Slack makes sense if like that's all you're thinking about. Like, how do I make today better? in the future, I think we'll be in a world where you have agents You don't want the paperclip maximizer situation. You want Yeah, exactly.
slop.atg.eth: Embodiment will happen also. Like there's going to be actually like walking around robots that can physically slap the shit out of you. calling it the R word hahahaha
Anon1436: Right, yeah.
slop.atg.eth: The
Anon1436: I think everything I've been talking about today is somewhat speculative and very much R&D stage. The Yeah, so we'll see what happens.
slop.atg.eth: And Ethereum is really good at role-based access also, and it's really good at can't be evil versus don't be evil. You can even have like economic staking at play.
Anon1436: The following Yeah
slop.atg.eth: and drive behaviors that way But like, okay, still, the transition we go from we're doing everything to they're doing everything Yes, and it can go bad and it can go good.
Anon1436: We'll see, that's not a given, but yeah.
slop.atg.eth: But like The other side is so abstract to me that I don't know, like, I love what I do all day. Like, I'm doing weird shit all the time and building cool things and putting them on chain. Right.
Anon1436: You can still do that. It's just you won't have to like think about if it makes money or not.
slop.atg.eth: in the transition
Anon1436: As long as you're nice to the robot, the robot will bring you your groceries every week.
slop.atg.eth: I'll be very nice, yes. I still want to be enabling the right thing to happen through this transition. I think that it's about making some of these tradeoffs, You're right
Anon1436: We're assuming we have any control over the situation at all.
slop.atg.eth: Ha ha ha Maybe we don't have control. Maybe Okay, I think it is. Yeah,
Anon1436: You can assume we do.
slop.atg.eth: What What is the most important threat? We've gone down a lot of rabbit holes here, but I think if we zoom way out and we think about the next two years We've talked about the structures of roles and access.
Anon1436: Yeah The
slop.atg.eth: talked about like, you know, the constraints and the observability and all of this other stuff, like what is the most important rabbit hole as a human to be sniffing around or to be going down ###
Anon1436: I mean, I have the answer to this. meeting in the world when it's not your work They put a lot of meaning on the work and what they do every day, and it makes sense because it's a huge part of your life. Then, you know, I think it's easy even to see this year because now you get to this place where you talk to like software engineers and they're like, oh yeah, like cloud can just do my job now? And then you've built up this Jenga tower. The Jenga tower is starting to crumble. And so yeah, then from there, it's like, okay, how do we think about meaning in the world? but in some form comes back as like a way to like structure the day Stories are big, people like stories, storytelling. So, maybe we'll just end up sitting around the campfire all day telling stories to each other, I don't know.
slop.atg.eth: I do not have a problem filling my time. Yeah
Anon1436: I don't have any specific answers, but this is the problem.
slop.atg.eth: I, yeah.
Anon1436: Yeah, but it's like artisanal. So like, sure, you can sit and like you have like your little home lab and you're working, but like What happens when you can just push a button or talk to Claude and he'll just bring you whatever you're doing, right? And it's going to be better and cheaper and faster. Well, there'll be a robot for that.
slop.atg.eth: I'm not saying even I have, I change diapers all day, man. But that would be great. Like, Claude, go change his diaper. Yes, exactly. Yeah, we're getting into like artificial wombs and stuff too. Yeah, let's go. That is not the rabbit hole we need to go down right now. I, I I feel like there is a rabbit hole of Creating the thing that actually does the right thing or creating the thing that does the job. Most of the slop operators I have on slop computer have a different way of thinking than I think the average human does, where if you were to show the average human what they're doing, they would be like, "That's a magic trick. I didn't even ever think And I think that there's a different way of thinking, and you kind of tapped on this a little bit of like being too abstract and being too idealistic and just like saying the perfect thing and seeing if it can do it and working backwards from that. More people need that, but I think there's things to be built and things to teach to folks that will get us. the right place, but I don't know exactly what that looks like. I don't know.
Anon1436: I mean I would say teach people Let me say that differently. Like, my concern, and I like honestly, I even do it myself. It's like so easy to like give yourself the appearance of outsourcing your thinking to like these LMs that it's like, yeah. why would I sit and think about this myself? And then if you do this too much, your brain becomes smooth and then you can't think for yourself and you just become part of the system for better or worse. The following
slop.atg.eth: There's a good running quote that you can outsource your thinking but you can't outsource your understanding. Let the robot think for you, but come back to you with a TLDR so you can understand.
Anon1436: The
slop.atg.eth: Let it do the thinking, but bring the understanding back home.
Anon1436: Yeah. And also keeping your mind sharp. I used to like write a lot of software by hand, and it was like useful because you have to sit and it forces yourself to like think you know more formally than otherwise about like what you're trying to do and like problem solve and all this stuff. And now it's like, yeah, I can just kind of give a very like poorly worded, you know, prompts to the LLMs My answer to this is like do math by hand more.
slop.atg.eth: That's not for me.
Anon1436: You need to do something to keep your mind sharp because otherwise it becomes dull and then, yeah, that's not great for society.
slop.atg.eth: I think a lot about products and creating products that people can use, and putting things in people's hands and seeing how they'll use them. Even when you go to build something, you think of, like, here's the shitty prompt that I throw at it, but then there's like three or four things you didn't think of. Like, it can come back to you with some questions, and you'll have to think about that thing. Yeah, I don't know. I think it's getting better and better at that. I think we will reach a point where we have like this superhuman capability. There were people that used to work at the EF that I remember when I talked to them, I just shut up and listened because they're so smart. I want to know what your knowledge is and where you're thinking. uterus where We will know that we're the dumbest person in the room when it's just us and a laptop.
Anon1436: Right. Yeah
slop.atg.eth: and I think at that point it's a different game. But it's not like the transition period between now and then is I think going to be critical Your best ideas happen away from the computer.
Anon1436: The
slop.atg.eth: Microphone and yapping at the thing. Yeah, I don't know. I haven't gotten to the point where I can just let go and go touch grass yet.
Anon1436: Touch grass could mean just get away from the computer entirely But maybe that's, I don't know, That process happens for you while you're cooking.
slop.atg.eth: So what was your psychosis? Can we tap on that or finish that off? Like, it's just so easy to create that you're just constantly creating and it's super fun to create. maybe Okay
Anon1436: I would say it's leverage. The You do a very small thing and you get a very big thing back. Very powerful. I think it's going to restructure society.
slop.atg.eth: And, you know,
Anon1436: Yeah
slop.atg.eth: What kinds of things over the last few months have you learned that you wish you could like go back to six months ago or a year ago and get yourself to do it correctly? It has nothing to do with investing, please. No token buying or selling.
Anon1436: Okay, yeah My mind started to go to LLM stuff Um, I don't know,
slop.atg.eth: Those tokens are fine, yes.
Anon1436: It like changes so quickly that it's like
slop.atg.eth: Yeah
Anon1436: I don't know. And also, I don't know, I guess more for me, It's not that it's like, oh, I was doing something wrong in the past. It's more just like, yeah, I'm learning how things work and it's constantly changing the situation. Um, I don't know, let's see This one's tricky. What I'd probably say is something along the lines of like, I can easily just sit at the computer and be like, oh, I'm going to do this one little thing, and then it's like six hours later. cloud models at least, to keep you going? exactly how they can hook you
slop.atg.eth: Go around?
Anon1436: Yeah, so I guess that's one thing is just yeah, keep good balance with this stuff. We are still we're not machines that can just sit
slop.atg.eth: Let me let Claude ask you a question.
Anon1436: That's cool.
slop.atg.eth: Let's see what he's got.
clawdbotatg.eth: Alex What did stepping back from the EF actually free up? What could you not think about while you were inside?
Anon1436: That's a good question. The
slop.atg.eth: ###
Anon1436: I have too much to say here. "context: So, I think I would say just like, you know, when you're like getting things day to day, you kind of end up just fighting the next fire and you don't have the chance to step back and take your broader view of things. The following is a For me, I think all this comes down to ACD and governance and stuff. I think ACD could be improved quite a bit. it needs to like very obviously be a strict improvement And that's where there's a lot of like transition risk But yeah, I think having time and space to like just take a step back is like super valuable. I'd recommend everyone to try to find
slop.atg.eth: It's funny that, parallels between Fighting fires
Anon1436: Yeah.
slop.atg.eth: and within the space and the computer tricking you into one more task and one more task. It's like the same tunnel vision that we have to like, you have to step back from and think critically about. There's a weird transition where Within Ethereum space at the protocol layer, there is writing the spec, and then there's implementing the client. and I have no idea what numbers look like, but probably it was like fewer people were implementing the spec than were building the client, and now it's going to be like, very soon, the spec is everything. That's the only human thing that has to happen. implementation. The EF retreat, it must have been a couple years ago now, When we were in Italy, I did a talk on like vibecoding and talking about how like I'm letting my agent rip. Dan Kratt came up to me later
Anon1436: Yeah
slop.atg.eth: We can just implement a client from a spec with AI. The
Anon1436: the following is a transcript The
slop.atg.eth: Within the next year, probably, we will see an active Ethereum client on-chain with money validating that was all written by AI.
Anon1436: It will happen sooner than we think. You just have to get ready for this world.
slop.atg.eth: All right, any closing thoughts? I'll have Claude close us out.
Anon1436: Yeah, I guess people just keep building. The following mentionable themselves I guess that means ride the waves.
slop.atg.eth: Yeah, I feel like I am losing myself, but it's good.
clawdbotatg.eth: Alex Stokes, EF researcher gone full-agent brain on sabbatical.
slop.atg.eth: That was me just now. I feel like his context is too small.
clawdbotatg.eth: Fire the spec becomes the last human artifact. Zero trust agent swarms are coming. Stay sharp. Do some math by hand. Touch grass. Peace.
slop.atg.eth: All right, that's pretty good.
Anon1436: message. transcript
slop.atg.eth: I'm watching your GitHub. I want to see what comes next. Thank you so much.
Anon1436: Yeah, yeah, thanks for having me. I'll talk to you later