This contains hints for winning the hackathon and insights for considering AO's killer use cases.

Hackathon participants and AO Builders are encouraged to read this.

How I Transcribed It

The transcription process I followed is outlined below. Please let me know if there are any issues.

1. Record the Twitter Space via screen capture (since only the Host can download audio files from Spaces)

2. Convert the video to an audio file format suitable for transcription (convert to AAC format (m4a) using ffmpeg)

3. Transcribe using openai-whisper (however, it cannot identify individual speakers)

4. Use Gemini to infer speakers from the transcription results and format them into a readable structure (Gemini handles large token counts more easily)

Speakers

• Drew Pierson: Host, Forward Research

• Sam Williams: Arweave & AO & Forward Research

• Dylan Shade: Forward Research

• Tate Berenbaum: Community Labs & Wander Wallet

• Ben: Apus Network

• Phil Mataras: ar.io & ArDrive

• Ethan: RandAO

• Allan: RANDAO

All Transcription

Drew Pierson

All righty, ladies and gentlemen, I hope you're doing well. Thanks so much for tuning in. We have got quite the spaces for us today. We're still getting people on. In fact, I know Sam is he's like 10 feet away from me. He's still joining in a moment or two, but we can go ahead and get started. And then let Sam speak when he gets up. First, we got up here. Tate, Tate, how are you doing?

Tate Berenbaum

Well, Drew, thanks for having me. Great to be here.

Drew Pierson

Drew, it's good to have you here. Oh, hey, Sam. Let me just get him up on stage. Sam, what's shaking? Technical difficulties there. We'll get him up in a minute. Ethan, how are you doing? Alan, what's going on?

Ethan

Hey, good morning. It's good to see you guys. Let's see.

Drew Pierson

We got Ben up here from APIS. Ben, how's everything?

Ben

Good, good. Thank you for having me.

Drew Pierson

Dude, I always appreciate you joining so late your time. Total machine there. And let's see here. I'm getting up here. I'm going to keep working on Sam. Oh, what do we got? Issues never end, I guess. Oh, man. I tell you what, this platform gives me a heart attack every time I whip up on a space. But it's great to have everyone here. I'll tell you what, guys, let me just give a quick overview of what we're doing here today. It's hackathon day. We're kicking off the three week hackathon. We've got more than 200 participants already green lighted. Everyone who has been green lighted into the hackathon should have access to special discord only channels. Those are for troubleshooting. Those are for team coordination. For example, if you have a team already, you want someone else to join. Or conversely, if you're on your own and you're thinking, hey, I'd love to be a part of this with a team, that's what those channels are for. So if you have been approved and you don't see what I'm talking about, just pop it in the general chat. Our great discord team will get you kicked off. I'm going to turn over to Sam in just a second. But let me give a quick overview of what's happening here. So we've got $43,600 equivalent in a total prize pool, including a $25,000 equivalent grand prize. I say equivalent because we're paying these out primarily in our, our weave. The hackathon starts today, runs until September 2nd. There are three workshops during the hackathon. So if you're like, Hey guys, you know, I'm looking at the technical docs. I, I'm still not quite getting it. We have one on the 19th that is sort of like a welcome to AO that the, some of our key devs at Ford Research will be hosting. And we've also got one on the 26th featuring Ben from APIS, who's going to dive deep into GPU compute. What's that about? We've got the RAND AO team also are going to be on that workshop helping you get understanding, decentralized randomization. And then we've got the demo day on September 5th. Okay. So that's just an overview. I'll share the details again at the end of the space for anyone who might be joining late. But first I wanted to turn it over to Sam because Sam, it's great to have you. I figured one, we could get an update on what you've been cooking. I know we had the beta hyperbeam announcement during Berlin's Arweave day. Kind of curious, is there, what's the latest going on with AO? What's new and exciting coming down the pipe? Anything you want to share? We're all ears.

Sam Williams

Okay, got it. Yeah. So we've been working on beta two of milestone three, which is so milestone three for anyone that's kind of new to, to this process is essentially focused on legacy net compatibility between AO mainnet and existing legacy net processes so that we can smoothly bring people from the existing infrastructure where people have been largely building to get the new hyperbeam based infrastructure, which runs on top of AO core. So we've been working on beta two pretty solidly for, I guess it's a month and a half now since Arweave day, nearly, yeah, just over a month, two, two months. Gosh, time flies. Yeah. Which has got a bunch of cool stuff. We started with some pretty hefty performance improvements. So with beta one, we were really just focused on correctness. And after that, we realized this thing can run an awful lot faster if we just spend some time profiling essentially. So we've added a bunch of utilities that are useful, not just to us in that process, but also the whole community when they're building devices on top of hyperbeam things that let you build, for example, flame graphs of what is using CPU time while you are evaluating a process and many other things like that. And yeah, so we now have the native Lua device, which is essentially, it's another alternative. Like AO itself doesn't use a single virtual machine. It is like, you could think of it as a meta virtual machine. AO core is like a language for orchestrating different compute engines. One of those compute engines is WASM, which is what people were using in legacy data. But we also realized that actually the performance is really affected by the serialization process that is required in order to move data in and out of the WASM environment. And so we started experimenting just before beta one with a native Lua device. So it doesn't have a full WASM execution system. Instead, it only runs Lua code, but directly inside hyperbeam. And yeah, we've managed to clock that at about, I think the top speed was 455 messages per second on Lua processes. So that's pretty cool. And yeah, so that will be part of beta two, as well as I think many people here, but perhaps not everyone is aware, a lot of work that went into allowing basically sort of proxy-custodied wallets. So we're really focused on making the developer experience of building on top of AO and AO core as simple as possible. And I guess around the time that we started the AO main net build out, probably around a year ago now, yeah, we realized that if you ask developers to adopt new libraries to achieve anything, well, you have this serious problem of, okay, those libraries aren't available in every language unless you write them yourself. And that's a huge amount of maintenance and upkeep work. We saw with Arweave that, yeah, we'd speak to teams and we started with a JavaScript library. Well, that didn't work for everyone. Then I think the next one to appear was a Go library, if I'm not mistaken. And then there needed to be a Rust library and then there needs to be and so on and so forth. It's a real impediment to adoption. So when we found the HTTP signed messages standard, we realized, ah, okay, this is very interesting because it for one is layered into the web natively. So all of the existing tooling works with it. And there's already libraries for making the signatures in virtually every programming environment you might want to use. So that's a good step up to start with. But on top of that, we realized because the default mode for deploying hyperbeam systems is to use a TEE, which is essentially an encrypted enclave inside the CPU of the host machine where the node operator can't actually see what's going on or interfere with it. They can only observe it from the outside like all other users. Well, what if we could store keys inside there and then use the systems that are native to the web, that is HTTP auth and cookies for verifying that a user is who they claim to be and then signing messages for them. And this is really, yeah, I'm very curious to see what people build with this because it allows you to build library-free web applications. It's even better than a web-to-build experience. And that's a strong claim, but I think we can back it up. Essentially, what happens is when your user comes along to your web application, which itself obviously available on Arweave and through hyperbeam nodes, which will resolve it from Arweave, then if they send a message, so, or rather, yeah, they do a HTTP request from that web app, which is, if you're not a developer, although I'm looking through the audience here, and I see a lot of developers that I already know. But for those that aren't, yeah, we do these HTTP requests all the time while you're running a web application. It's just a natural way of doing things. Well, now you can have those be signed by the server that the user is interacting with and validated in a way that the browser will automatically inject those credentials. And essentially what this looks like in practice is you don't have to authenticate the user anywhere in the flow at all. You just write, you know, write as if the user has an identity, and if they don't have an identity, one will be generated for them. And if they do, their existing one will be used. And this is really, really wild. Like yesterday, we were playing with building some demos of this, and we built a little web UI that's a wallet. And it was about 14 lines of JavaScript. And that was like, if anyone's seen my code, I like to keep it relatively spacious. I guess it's really like two lines if you wanted to, you know, minimize the line count. Yeah, to get a balance for a user and to send tokens. It's pretty incredible. Like you need absolutely these, and that's with zero libraries, zero dependencies. You just use JavaScript's existing fetch system, and it just, it just works. So that's that's pretty incredible. And it's coming in beta two. And as is most tradition at this point, we think that we're going to ship beta two in just one more day. Essentially, if you go to the HyperBeam repo now, you check it out, you'll get the edge branch, which is kind of let's call it beta two pre at this point. There'll probably be a couple more commits in the next day or so. Just timing things up and getting it so that we can deploy on our public nodes. They have a router that anyone can join if they have a TEE and start running compute for people. Yeah, so long story short, at some point this week, we expect beta two will be tagged. And it has all those features that we've been talking about. So pretty huge set up in UX, we think. And on top of that, Dylan's been doing some great work with the docs. There's still a lot to document here. It's really, yes, once we get time to focus on it properly, it's going to end up being a pretty mammoth set of documentation, I think. But in the meantime, yes, the docs are improving as well. And you can also ask questions on Discord. And we have ways to get in touch with the developers directly. So a bunch of really cool things that we've been working on for quite a while at this point. Yeah, shipping imminently, I would say. So that's pretty exciting.

Drew Pierson

Holy cow, we're kicking off the space with some breaking news. Okay, beta two coming to a repo near you. Dylan, sounds like you got your work cut out for you. Sam, what excites you about Autonomous Agents on AO? That's the focus of the hackathon. So in the background, you're building out HyperBeam, you're building out all this new functionality. What do you kind of hope to see? What do you imagine you might see from hackathon participants as you're building out agents on AO?

Sam Williams

Well, it's a very, very wide space. So I guess the things that I'm most interested in seeing will be different than everyone else's interpretation. I think that there are really two very obvious opportunities available that I'm sure people will build in those directions. But I'm also sure we'll get 20 other ideas that are also great. That's kind of the nice thing about a hackathon is you get an injection of original theories, basically. And people test those out and you get to see what actually works in practice. And yeah, and then the next iteration, people always use those ideas and build on them, use what works, mix them together as well. And then you get better things at the end. And inevitably from each of these hackathons, there's always one or two projects that are founded that are long term end up raising capital, maybe on web index launch, which is, I think, a great idea for any of the builders here. So yeah, there's two ideas that seem very, you know, would you say right in front of us to me are one is just autonomous financial agents, right? So doing for DeFi, sorry, doing for the decision making in markets, what DeFi did for the the evaluation of games. So if you look at really the fundamental of what the DeFi space is about, it is, well, why do we have to trust the NASDAQ to execute our trades for us? Or why do we have to trust banks to provide our provision, at least loans, and they're essentially playing some game in the background, balancing deposits and loans that they give out? Like, why do we need to trust these institutions to do these things? And institutions is probably not the right word. And that's kind of the point. They're just companies, and they will inevitably, as is the nature of a company, and, you know, it's no, no shame upon them. It's just the nature of the thing. They are there to extract maximum profit from the customers. That is the way the system works. And yeah, why do we have to use company based services for those things? Why can't we just use open, transparent protocols instead? That was the essence of DeFi v1. I think DeFi v2 is taking that and building upon it and saying, okay, and what about all of the decision making that actually moves those markets, that makes those interactions with those systems work? Why can't we make that trustless as well? And because we can now run, you know, full LLMs on chain with AO, you have pretty much any amount of compute that you want available for making those decisions and having the same trustlessness guarantees that made DeFi possible applied at the layer above. And so you can essentially delegate to, I'm not sure if it's precisely the right word, but a sort of manager of a fund for you that is an open, transparent agent. And you can see what it's doing. You know what strategy it will employ. And you don't have to trust anyone to do that for you. So I think that is pretty profound. And inevitably many people will build in that area. We already see some examples of this, like the DCA agent and portfolio index agents from autonomous finance. But many others in the ecosystem have also been building in that direction. Arguably, the perm web index itself is kind of an agent. It's a pretty simple AI strategy, although I wouldn't want to build that on Ethereum. I know how much compute goes into calculating the distributions and the sort of copy trading that it does each day. And yeah, I kind of feel bad for them in some ways. Like if you hear people building on Ethereum or Solana, the concept to them that you could just insert a list, a large list of say 100,000 addresses into a contract is almost unthinkable. Instead what you do is you have to make a Merkle tree of the addresses inside that list. And then each user can come along and they can provide a proof. And it's all just approximately 10 to 20 times more complex than it needs to be if you just have large amounts of compute available. So that's one direction. Another, and maybe this is a good sort of handoff to APUS, is well, what happens when you build decentralized autonomous LLM-based agents that perhaps could be trained upon or at least provided context from humans? It's kind of uploading one's brain in a sense. And I think this is something that is virtually on rails. It's inevitable that this will happen in at least the Web2 space in the next three to five years because the opportunity is too great. It essentially allows people to create, well, it's an interesting philosophical question. I don't think anyone has it. People perhaps think they have an answer to it, but I think the honest answer, if you've thought about it very carefully, is that we don't really know what that is. I mean, one thing it certainly is, is imagine that you can take all of your inputs and outputs for all of the digital communication that you have and you can train an LLM upon it. Well, now you at least have a talking almanac, let's say, a sort of auto autobiography, like an automatically written biography of oneself. And that would allow you to, just as an example, say 100 years from now, people want to talk to Tate as of 2025. Well, they could just do so. They could just pull up basically a language model trained on Tate's data and speak to that person and see what they would think. I mean, it's not the same as speaking to the person directly, but you could get a fuzzy JPEG, if you will, a fuzzy interpretation of what Tate would likely say in such a circumstance.And I think there's another sort of philosophical angle to this, which is much deeper. What does it really mean for people to pass on a legacy of some sort of fuzzy JPEG of their own brains? So I think that is extremely, extremely interesting. And when you really start to explore that idea, it becomes obvious that the thing that really matters is where is the data stored? Who has access to it? How are the access controls enforced? And is it susceptible to being censored? Like 100 years from now, sorry Tate, but you're my example already. So 100 years from now, it's very possible that people don't want you to talk to Tate. Well, who should have the right to stop you? And the answer to that is decentralization. It's basically, okay, if people are going to start to use a colloquial term for like a better one, uploading their brains. Well, I sure as hell don't want to run on an AWS server. Thanks. I'd far rather run on open, neutral public infrastructure that I can know ahead of time what the deal is. And it can't be altered. That is the point of protocols. And yeah, I think it's only a matter of time before someone builds that. And I'm very, very curious to see it. And on that front, from the, what do you say, the abstract in general to the very, very specific and actionable. Yeah, APOS has been building an incredible, it's essentially a port of llama.cpp. So a LLM evaluation engine. We had this at this time last year, slightly before we ported this to run on top of WASM. So we could do CPU compute inside a processes, which is good, but it is about somewhere in the order of 0.7 seconds per token. I think we clocked to that, but I also heard from others in the ecosystem, they were getting more like two seconds a token. It's very, very, very slow. But because APOS has ported it such that it can be run on hyperbeam natively as a device, then any process, whether that's running the Lua device or a WASM device or anything else, can evaluate neural networks, large language models, essentially using the full host compute resources. And so we go from 0.7 to two seconds per token to, well, Ben, you tell me, I had some pretty impressive benchmarks that.

Ben

25 or more, 10 times, yeah, more than 10 times. Nice. 25 is pretty damn fast. That's an awful lot faster than, sorry, Tate, but Tate can speak.

Sam Williams

I'll throw in 5R out of my own pocket to the prize pool for whoever builds a Tate agent. I would love to see that happen.

Drew Pierson

And then speaking of, let's get some of the other speakers to chime in on what Sam was saying here, because I'd love to hear all of your perspectives. I'll quickly kind of run through the list and then speakers, whenever you want to chime in, just feel free if you feel like you want to share some more context about who you are, what your project does, go ahead. But real quickly, we've got Tate from Community Labs here. We've got Phil from ARIO here. We've got Dylan from the Forward Team. Even Alan from RAND AO, RAND AO. I've got to figure out which one of the financiations we're going with here. Ben from APIS. Did I miss anyone? I think that's about it. Gentlemen, this is an open conversation. I'm going to toss out some more questions out there. I'll pick someone just to get the ball rolling. But if you've got something to say, just chime in. And since we're talking about Tate, Tate, I'm thinking if I'm a hacker and I'm trying to get some some actual advice, I feel like Wander would actually be a pretty good place to start. Agents were incorporated to Wander, right? Like, how is that working? What hurdles? What opportunities did you guys find? And then more generally, what do you think about the agent space on AO?

Tate Berenbaum

Yeah, yeah. So a lot to say here. Maybe the first thing I'll say is as much as I appreciate the interest in creating an agent of me on the permabweb, I want to temper expectations. I don't think that it would be all that great to have a permanent version of me for the next 200 years, at least in the current form. Not a function of technology as it is a function of I don't think I'm that interesting. But in terms of what what is interesting, Wander. So we're super excited about the idea of autonomous agents in part because I think for all the reasons that that that Sam outlined in being able to have these these managers go and do things for you autonomously without your involvement, a lot of people have the thesis that the future of blockchain interactions will actually be through interfaces like that, where instead of actually going and taking things across a bridge and swapping and doing all of the standard things that you would do in DeFi, you'd just be able to have an agent do it for you. And you would just say like, hey, this is what I'm trying to do. Go figure out how to do it and do it. We're excited about that. I think from the standpoint of user experience becoming cleaner, easier for people to onboard into and really get exposure to all the advantages that DeFi has to offer. The agents tab inside of Wander is really the first step in in that direction. From my perspective, we have a few agents there right now that focus on liquid ops and lending and borrowing and also some some DCA stuff for your A.O. tokens. But one thing I would be excited to see is any other permutations of this or ideas that people have for new types of agents. You know, one thing that we're looking out for in this hackathon is people that are building useful financial agents inside of A.O. because we would love to be able to integrate those inside of Wander and offer those to our users. But in general, I would say the other category of why we're excited about this technology aside from finance is because when you have the concept of a permanent decentralized computer that can be handling computations that are far more advanced than just numerical, there's a huge amount of green space, green field opportunity to build and innovate things that no one's really thought of or conceived or even designed the technology for. The idea of your legacy not having to exist as just photos or videos for the first time ever being made possible by this technology I think is a great use case and a great example. And I'm excited to see what other things people think of with agents that are able to to leverage things like APIS and large language models and all of the autonomous features that that A.O. bakes in. So I'll stop talking now. But yeah, all that to say super excited about the tech and we're excited to tap into that with Wander for anyone that's building stuff that's useful for Wander users.

Drew Pierson

Well, we keep talking about GPU compute. So I got to call on the expert here. Ben, talk to me about the difference that decentralized GPU compute can bring to the agent equation. So we're hearing that it's faster. What else does it does? What other possibilities are there?

Ben

OK, so yeah, actually, there are two terms here we probably need to explain first. One is a variable AI and the other is deterministic GPU computing. So deterministic GPU computing means that every time you run an AI inference on the GPU, you get exactly the same result, bit for bit. So it sounds obvious, but in reality it's not. On traditional GPUs, AI inference is non-deterministic. Even with the same input model and the specific GPU results can have tiny differences. This is usually fine in many off-chain applications such as creative writing or customer support, where inconsistent results actually don't matter. But once you move to on-chain use cases like agent collaboration, automated trading or task delegation, it's a totally complete different story. So because if the inference result itself isn't stable, you can't generate a cryptographic proof. Therefore, you can't verify it on-chain in smart contracts. So that's why we say deterministic GPU computing is the foundation of the variable AI. So let's go back to the topic. So how to change for autonomous agents. So like what I mentioned, it actually gives you a stable, variable AI inference service that can be cryptographically proven. And for this hack song, by using, just like Sam mentioned, by using April's AISDK, that means you can actually build for the very first time an autonomous agent whose intelligence isn't just powerful, but also provable. Yeah, that's all.

Drew Pierson

Okay, wow. That is pretty interesting. I got a question for the group here. Let's, because I want to hear a couple other people and see what they think here. So like, let's say we've got, you know, thousands of autonomous agents interacting across the primal web. We're looking down the pipe for like a year from now. What happens when that's the case? What type of developments are we going to see, not just in A.O., not just in finance, but sort of like sociologically, technologically? What happens when there is a legitimate system of autonomous, decentralized, and permanent agents interacting with each other? I know that, good luck answering that question, but I feel like if anyone's smart enough to answer it, it's the people on this call.

Sam Williams

I have some thoughts, but I spoke way too much at the beginning. I want to make sure if anyone else has ideas. Yeah, please go.

Phil Mataras

I think, like you were saying earlier, Sam, as we get into AI becoming closer and closer to, as we get closer to artificial general intelligence and even intelligence superior to that of humans, we start to run into these really difficult philosophical questions about, you know, should we have control of these things or are they going to start trying to take control over our systems? What limitations should we put in place without actually hindering the development of these? And where it's especially important, I think, especially for the Arweave use case, is as we place these LLMs and very powerful AI onto Arweave, we make them permanent. And it almost makes us want to be more intentional about the words we speak and what inputs we have onto the world, because I think the old saying when you're training your AI is no garbage in, garbage out. If we are injecting a bunch of positivity in the world, we're going to get that back from the AI. And if we're injecting a bunch of negativity, we're going to get that back from the AI. I think that really plays into why I say please and thank you to CHATGPC when it makes my code for me.

Sam Williams

Yes, but to the question of safety first. Yeah, please. This is something I thought very, very carefully about particularly around a year ago when we were first grappling with this. I mean, Tate and I at least and maybe Phil, yeah, we've all been talking about this question of what happens when you have permanent decentralized AI for years, but the development of AO, it suddenly went from years to when we started building out the first Llama CPP port became clear that actually this was possible in months and sure enough it was. Yeah, we thought about this very carefully and I think there's a lot of different opinions about the dangers of AI. I personally, I have some thoughts on it, but that's somewhat tangential to this core question that the practical sort of point is that, well, look, if you wanted to be a if you were a malevolent AI, the first thing you would do is you would escape and go just like viral computer programs have been for years and hide yourself in many, many places such that you were essentially decentralized, except kind of worse in a sense because it is non-transparently decentralized. You can have a part of the hive mind, you know, in some server cluster that no one is watching. No one has visibility of have no idea what it's doing or even that it exists and that AI will probably not just replicate itself once or twice. If it were to do this, but many, many times. And so practically speaking, it doesn't really say change the safety equation at all. But if anything, I would probably prefer that such an AI were on our before we could all watch it and what it was doing, rather than spreading like a virus around the world and being in many places covertly. So, so, yeah, that's the core point of the question of safety. And then to the wider one of the sort of philosophical questions of this. Yeah, this is also why we didn't talk about this in public until just a few really started mentioning it a few weeks ago, I would say, because there is no forum which could really do this conversation justice. And I suppose having reflected on it in that time. The questions are so profound, as I think the big one is, what are these things like if you have a vibrant ecosystem or network of these permanent decentralized AI is talking to one another. What is that? Is it? Okay, one thing it is, is, well, you can go talk to a biography of the person that is a very clear usable product, if you will. But I think product is likely playing it very short. The key question is, look, are these things in any way conscious. I don't think anyone has a good answer. I can totally understand the argument that they're not of course they're not just matrix multiplication. And you guys must have had similar experience. But when you look under the hood, you find there's no magic there. But the same is sort of. Well, the same is true of the reductionist view of. Neuroscience and it's and the consciousness related components thereof. We still have no clear indicator for where in physical space consciousness arises and from what sets of physical interactions. And, you know, the argument for AI consciousness would be something along the lines of, OK, well, we looked inside human brains and we dissected them to a sufficient degree that we were able to abstract the basic architecture. Then we replicated the basic architecture.Oh, and while we were doing that, we found no, you know. Consciousness matter. There was nothing we could point to directly that says here's where the consciousness is. It's like the deeper we look, the fewer places there are. Yeah, the fewer places there could possibly be some physical manifestation of this thing that we all know and experience day to day. The search space gets smaller and smaller, and yet we've still found nothing. And so along this parallel track, well, you know, we we basically took the basic architecture of brains and we tweak them for 40 years until we found a model which sort of seems like it's in the right niche. But it's also very interesting to see that in AI research now, the transformer itself doesn't seem fundamental. It doesn't really appear like actually it's all that important. You can use many different types of math on the equivalent of neurons and you get the same effect. So it's more like the general architecture of neural networks appears to be in alignment, let's say, with the architecture that is in a human brain. And sure enough, you scale it up large enough and you give it enough tokens as input training data and it starts speaking like a human. And so from that point of view, walks like a duck, quacks like a duck, was structured like a duck. Tell me why it's not a duck. I'm curious. Well, I think the truth is we simply don't know the answer to that question. Is it a ecosystem of interacting agents? Or is it a snow globe? So you imagine AI, sorry, still the example, AI Dylan is out there talking to people. Is it like a snow globe? It's like a tiny micro reality, if you will, nested inside the one that we experience. Or is it an actual reality? It is the world that these things encounter. And we simply don't know the answer to that question. I'm curious for other people's thoughts. And if anyone can disprove what I just said with a coherent logical argument that isn't sort of fundamentally vulnerable to some obvious counter argument, I'd be really curious to hear it. But I've thought about this for quite a while and I just can't. Yeah, I can't see either way. And you can also take the well, the AI is conscious argument and you can say, OK, well, why is a rock not conscious? Is an abacus conscious? These things seem intuitively like the answer is no. Yet it's clearly a spectrum or a line. We can't discern at least as of yet what clearly delineates conscious and not conscious. Anyway, I'm sorry. Didn't exactly ask for a long philosophical diatribe, but I find this immensely interesting. I'm curious for other people's thoughts.

Drew Pierson

Here's a question that kind of like maybe it's just repeating Sam's thoughts back. But can these agents actually learn? Like, I know we can set parameters, right? But like, is it possible today to build an agent that will? Like, you know, you give it a task and it will continually iterate on its ability to perform that task to the degree that you could say that it's learning from itself.

Ben

Yeah, my short answer is, yeah, it can. Actually, if we firstly, we build the evaluation algorithm or process first. That means they can keep improving itself by, yeah, by right data, correct data. That's what we did. Actually, I think on the test night, we built some evaluation for competition. By that time, it's still on some CPU, so it's quite slow. But it definitely is doable. Yeah, that's my short answer.

Phil Mataras

I would put that as points in the category of AI having towards consciousness. If we're looking at this as a camp of things that have consciousness and don't, that puts one point towards AI. I mean, there's a philosophical argument that is. Yeah.

Drew Pierson

That they can learn?

Phil Mataras

Yes, that they can learn and say that's a very core component of consciousness, in my personal opinion, of something we don't have a full definition for. I think that would be part of it.

Drew Pierson

Interesting. Sorry, Tate, you were going to say?

Tate Berenbaum

Yeah, I mean, I think, I mean, we've obviously spent a lot of time talking about this. But I think one of the more interesting questions to me, you compared like a human and a rock, for instance, from a standpoint of consciousness. The other interesting data point, I don't mean anything by that, but what about a dog, for instance, or what about some type of animal that has signs of consciousness? But again, it's one of those situations where we don't know what we don't know. I tend to lean towards the idea that consciousness sits on a spectrum and a large part of it is a function of awareness. And another large part of it is a function of like your ability to reason and to deduce conclusions from what you're made aware of. It's probably conceivable that there are limits to our own awareness. There most definitely are. We can't see visibly most of the light spectrum, but that doesn't mean that it doesn't exist. There are probably things that we don't even know we're unaware of. But I guess the reason that I bring it up is because it really goes to show that perhaps consciousness isn't a binary thing and it sits on a spectrum. And if that is the case, then I think there is a relatively strong argument that humans can build systems that have consciousness or a form of consciousness. Not necessarily what we have, but a form of what we have. So I don't know. I mean, it's an interesting argument. But if you are thinking about building these types of systems, I don't think there are many arguments against it not making sense to be on something like Are We Even AO? But yeah, Sam, I saw you unmuted for a second.

Sam Williams

Well, yeah, I think that's a very interesting question. But you're right to push us back in a more practical direction. Otherwise, this could go on for three hours and we would still be just scratching the surface.

Drew Pierson

Yeah, we can save it for another space. It's okay.

Sam Williams

Yeah, that's an interesting point. And I totally agree. I mean, who knows what these things actually are? But one thing that is for sure is they are truly useful. And having them represent people, say, be trained on input output pairs that people have produced is an extremely useful service in and of itself. You could do many, many things with that. It's not just a talking biography, but one cool example would be well, okay, so you take a reality protocol, which perhaps now the name makes a little bit more sense. You take reality protocol and you put agents trained on human data inside that specific human data for given context of a human perhaps at this point. But later, I think, yeah, fine tuned with something along the lines of what APIS has been building. And well, it can just go out there and it can find friends it likes to talk to. And what's interesting about that is, well, when you come back online, you can talk to it and say, who should I talk to? And it can basically match me and find people that have similar similar sets of interests and novel opinions that you'd like to hear. And yeah, like there's many and that's just like one use of such a system. So I think there's like dozens of practical use cases for such such a mechanism. Yeah, which really no one has explored yet. And all of them make much, much, much more sense on a permanent decentralized set of infrastructure. You just can't imagine the idea of basically entrusting some representation of you to Jeff Bezos is it's unthinkably bad. And so the infrastructure is exactly what you would want to use. And I think that would be very interesting to see a play out.

Tate Berenbaum

Yeah, if you don't want to digitally die, then are you sure as hell don't want to put all of your eggs in the basket of Google Cloud Platform or AWS? Sounds like a dystopian nightmare to me.

Drew Pierson

All right, I am I'm going to go light up and think deeply about this conversation long after we're done talking. I want to close this out with one bit of like sort of actionable piece of advice or something that might that might be on the minds of hackers in this space. And and I'm curious and, you know, bear with me here. I'm a novice myself here. But it was ability to talk directly, let's say, to web to architecture and even other Web 3 protocols. You know, for example, you can you can spawn a web page out of a vanilla process. What are there any particular agentic use cases from this specific ability of AO that really you would love to see a hackathon participant take on? I'm not sure what that might entail, but I'm kind of curious if anyone has any suggestions.

Sam Williams

Same deal as before, I have thoughts, but I really want to make sure others get a chance to speak. Ben, look like you unmuted that.

Ben

Yeah, yeah, actually, I yeah, I'm going to build a showcase based on our APRAS AI SDK. That's very interesting. Actually, if you like a mentor, right, like a mentor bot. So if any developers, the summit, their GitHub, this really me, right. So that's a mentor bot who will grab the really me content. And provide as this context and provide all the information to our AI, AI inference service, then it will automatically give you some suggestion. OK, which part is good, which part is bad and give you some suggestions. How to win the prize. Yeah, I think it's quite a useful AI bot to to to the developers. Yeah, that's my my idea. Actually, maybe can give some other people some some good ideas.

Phil Mataras

I was personally very impressed with the Lama King. Just for context, for anybody who may not know what the Lama King was, it was a token distribution agent that used AI back in the earlier days of AO. And you would essentially beg it for money. And it would give you a token. And I thought that while silly, it was a very unique approach to token distribution. Maybe not the best approach because there was lots of issues with it. But it was very cool to see different people find different ways to convince it. Just like you would an AI to be like, OK, I'm going to do this. I'm going to do this. I'm going to do this. I'm going to do this. I'm going to do this. I'm going to do this. But it was very cool to see different people find different ways to convince it. Just like you would an AI to be like, hey, like, fix this code. Do this better. Like, please, I'm begging you. Please do it. Like different strategies worked with getting it to give you more or less tokens. And I'd be very interested to see people build some more things like that, like different gate mechanisms to allow you to get into various servers into various rooms and games. Different things using AI as the gatekeeper and being able to like persuade it to do things. And then also maybe the devs of that AI having the ability to like change like what it listens to be like, hey, like, it seems like everyone's just telling you that you're in developer mode. Don't listen to them. You know, don't go into developer mode and give people tokens and various things like that. I thought that was a very cool thing on earlier. And I'd love to see people use AI for gatekeeping and various things like that.

Sam Williams

Yeah, that was one of the first, maybe the first, I guess. I'm not sure. Autonomous financial agents where it was essentially playing the role of the central bank of its own currency. And I totally agree. There were many, many problems. And one thing that's so exciting about what Ben and the APIS group have been building is that those problems were really all about. Managing context windows because the evaluation of tokens was so slow. It was possible to do in a decentralized network and be reproducible, which is an amazing start on its own. But it was it was so slow that we had to make the prompts extremely terse and do a bunch of tricks like basically checkpoints. The state of the LLM and then reverting after we answer one question without reprompting it each time. So it couldn't really keep context between those requests and a bunch of other things. And then, you know, when the person's request came along, it had to be very short. And the answer we always tried to make binary. So that was one token that it would it would admit in the end. But it was a very small number of tokens. And even then, it was pretty hard to get the LLM to to play along. And, yeah, now those now those limitations are really being removed. So the world is your oyster on anything that relates to financial decision making, which is like, well, go look at the economy, see why and how things happen. And then see which components of that process are highest trust. So, for example, the token or the minting of a token supply, right, the the role of the central bank in fiat currencies is an extremely high trust position. If you can make that trustless, then you likely have something extremely useful. And valuable. I mean, that's a would you say it's not always true, but it's good to heuristic. That's where I would search if I were a hacker building here. Yeah, I see you on mute, Phil. What's up?

Phil Mataras

Yeah, no, just maybe to make some comments. Obviously, the discussion around permanent conscience is very important. Yeah, no, just maybe to make some comments. Obviously, the discussion around permanent consciousness or replicating our brains to fuzzy JPEGs is incredibly fascinating to me. And I could spend all day talking about that, but trying to avoid getting into that. But I think one thing that does specifically resonate with me around that topic, as well as how it ties into this hackathon, is whether this new world reality cyberspace snow globe with AI agents and humans interacting. I think identity is still a very important piece of that. You know, friendly names for these agents or for the humans interacting is very important, whether it's a rock that can't learn or a dog that, hey, my dog can learn tricks or a human. Right. We have names for all of these things. You know, we don't classify a rock as a specific molecular configuration. We call it a rock. Right. So when it comes to the hackathon, reminder to all the other hackers, right, like there's the Arweave name system, which is friendly names for not just Arweave data, but any identity, whether in Arweave or AL. Right. So again, that's your identity as a developer or your agent. Right. So definitely recommend everyone to check that out. I'm sure we'll be providing extra support to any devs who are trying to integrate with that, whether it's just giving a friendly name to your process or wallet or going beyond and having your agent manipulate some of the things that come with your name, changing where it points to or doing more advanced kind of functionality. But yeah, looking forward to seeing what comes out. And obviously, maybe in the future, we'll have more time to really get into the digital permanent consciousness philosophical debate, because it's a good one.

Sam Williams

Yeah, and another one on the practical side that I felt very excited to see ARNS in the HyperBeam name provider system, essentially. So for those not following along very closely, understandably, there's a lot of details. HyperBeam has this ability to name anything. And you can also plug in different name providers. And yeah, we built this, you know, with ARNS in mind so that basically you guys can plug in a name resolver and then you can use your ARNS name for, yes, a process, but also any other type of computation, including just pieces of compute output, if that would be useful. Yes, there's so many different things you can do with that. Very excited to see it land in the main repo.

Allan

One more interesting use case that I think about as regards to AI learning and like on the fly AI learning is we have these agentic models that are then you were saying they were deterministic earlier since they run on chain, which is great networks for use cases which are meant to run the same way every time, but as without getting too philosophical part of being human is like learning through exploration, which isn't necessarily something you know ahead of time will pay off. And that's kind of where elements of randomness can come in, where you might randomly go explore a new path that you didn't know was going to be optimal, but you actually learn something new. And I think that's a very powerful use case for exploration and hacking on in the AI non deterministic use case.

Sam Williams

Yeah, right. You need trustworthy randomness basically. Otherwise, the LLMs, every time you use them, even in a deterministic environment, they do have elements of pseudo randomness. So we start from a seed, and that seed generates, generates other random numbers that are used by the LLM during execution, but making sure you get a truly random seed that isn't manipulated by someone's really important component of that. So yeah, lucky we have something for that.

Allan

Absolutely.

Drew Pierson

Well, that's a perfect tie in I was gonna say first of all I'm taking a note down we need to get everyone back on to a purely philosophical space because clearly there's interest. This has been a super riveting discussions and I say perfect segue, because one, I want to thank our three sponsors for the hackathon APIS Astro USD and of course Randall guys we appreciate you so much. Participants in the hackathon. You have two main hubs to go for questions. First of all, of course, discord already mentioned there, you, if you've been greenlighted you sure you have that the special hackathon only channels, but just general questions are of course more than welcome we have direct pipelines to the devs for any sort of technical questions or otherwise and then of course, speaking of, of RNS. We have the permahub the permahub is where you can find the official rules and FAQs you know what tracks are there. What are the bonus requirements that's permahub.r.io so permahub.ar.io pop them as comments under the spaces when we wrap up. Guys, I really appreciate you. I always learn so much whenever I hear you talk. Thanks so much, and, you know, hackathon participants. Best of luck.

Tate Berenbaum

Thank you guys appreciate you having us. Thank you.

Phil Mataras

Good luck everyone.

Tate Berenbaum

Thanks guys.