The Swarm that Builds the Swarm - Trending on GitHub! I accidentally created the Borg Collective... | Artificial Intelligence Masterclass

SPEAKER_1 (0:01)

This episode is brought to you by LifeLock. Between two-factor authentication, strong passwords, and a VPN, you try to be in control of how your info is protected. But many other places also have it, and they might not be as careful. That's why LifeLock monitors hundreds of millions of data points a second for threats. If your identity is stolen, they'll fix it, guaranteed, or your money back. Save up to 40% your first year. Visit lifelock.com/podcast for 40% off, terms apply.

David Shapiro (0:30)

So I think I might have accidentally created the Borg Collective, and I'm not sure that it's a bad thing. So for some context, if you're just tuning in, I've been working on a project called the OpenAI Autonomous Agent Swarm, or HAAS, for Hierarchical Autonomous Agent Swarm. So basically, with the release of OpenAI's agents endpoint, or assistants endpoint, we realized it would be really easy to spin up an arbitrary number of agents. And so I created an open source project, and I've been using daily YouTube videos as more or less kind of a scrum stand-up. But then the comments started getting really cool. Someone said, the swarm is pleased by your videos. And then someone said, is that what the subscribers are being called nowadays, the swarm? And I said, the Borg Collective approves. And someone says, thank you, Locutus. There's a lot more comments like this. But yeah, so hello, Collective. Good morning. We are hard at work, so let me give you some updates in terms of what we're doing and what we've achieved. But first and foremost, the most important thing is, I set out a bounty. So the bounties are put under issues, but I set out a bounty for someone to create a new ChatGPT that specifically reads the discussions here so that you can talk with a chatbot. And someone delivered. So thank you, team. The Collective approves. So here's the Haas Board concierge, which you can get to it right here. So I put it right at the top since it's super useful. And I just said, like, what's the latest set of tasks I can do? So it gave it like literally just as like, okay, cool, this is what we can talk about. And I'll ask, like, what are the latest PRs? It's slow, but I think it'll get faster. So anyways, let's see what that says after a little bit. So we'll come back. But point being is that it's moving fast and there's a lot going on, but we have a tool that acts as an information concierge to the actual discussions. Okay, great. We'll delve into the discussions in just a second, but what I want to start with is the issues. So this system has worked really well, where I'll just put out a bounty. And I've closed a bunch of bounties that I realized were a little too ambitious, but it's starting to really work. And so the first bounty of the day is Proposed Milestones Roadmap and General Principles. So basically, I started with, okay, what is it that we're trying to build? So we're trying to build a mission-driven, hierarchical, autonomous agent swarm. It needs to be self-organizing, and we're realizing that's the hardest part. So I've got a few other bounties that are around that. And then ideally, in the long run, as a stretch goal, it should be self-improving, which there's been a lot of conversation around that. But we need milestones and a roadmap. So we need to get more organized around the general principles that we're discovering, so on and so forth. And to do that, I created a wiki. And I actually allowed for, like, this is Wikipedia style. Anyone can create and edit. But I have a general principle. Wait, it looks like it didn't save the... Okay, I'll update this. I had a home page that said that we need to abide by some general principles, so I'll fix that. But anyways, I've documented the ChatGPT agents. So there's the Haas Board Concierge, and then there's the first one that I built. Then, oh, here's the home page. This is supposed to be the home page. General rules and principles don't modify this page. Less is more and document hygiene. And then finally, I put a placeholder for the roadmap. So that's how we're going to organize. So this is going to be... This is very similar to like the standard operating procedures that we should attach to all agents. So basically, all agents in the autonomous swarm should have a standard set of documents that they can follow and abide by. So that is how you can ensure that the behavior of all the agents is at least somewhat aligned and so that they have some relative information. So we are the agent swarm and we are building the agent swarm. Great. It's swarm inception, agent inception. So that's the first bounty. Another bounty is I'd like for the community to look at the recent pull requests and the current code base and the open pull requests and comment on them. So there's a few things that I was not quite so happy about with the current ones. Sometimes I want to, so let me just say I want to integrate all this code, but we need to make sure that we're adhering by best practices and keeping the repo organized. So that's first and foremost, I did merge the pull requests that are good. So just test the code base. Let's make sure that it's working and share your results. That's what the discussions tab is for. People are using that to great effect. And then let's see, the next big thing is we realized that the agent communication is actually the hardest part. And I also had a meeting with the Ace Framework team, which I'll talk about in just a second. But basically, like, we can take inspiration from chat dev and autogen, but we really need to establish, like, the best practices and general principles for establishing agent-to-agent communication. And also one thing to keep in mind is that we are going to need to have multiple layers of communication, which, actually, let me just go ahead and show you that in the wiki. So in the roadmap, I had, you know, my idea is like, based on some of the discussions, so I'm reading all the discussions, by the way. What's kind of emerging is that it looks like we need kind of a global communication space for all agents to talk, like, on broadcast, so that'll be like an internal Twitter. Which, you know, honestly, Pinecone would be good for that, because then, like, all agents can search everything that is put on the global space, and it's super fast, and it's got metadata. So that's basically an internal Twitter, and it's a third-party API, which also means that it's arbitrarily accessible. But people are also talking about, we need to document topics, and what needs to be organized. So I think an internal Reddit kind of style thing would be really good for basically accumulating knowledge around specific topics. But again, that can be a third-party API. So knowledge accumulation, topic management, and that sort of thing. But let's use the Reddit. And then we also need a real-time set of chat channels, such as Discord, which could also be an external third-party kind of thing. But yeah, so I think these three styles of communication should all be used eventually. But obviously we're very early. So hence this bounty, which is talking about multi-agent. We need to start with the basics, because we as humans know how to use Twitter, Reddit, and Discord. But now we need to make sure that we can build agents that can use those communication platforms as well. And I think between the three of those, there's enough differences, like structural differences in terms of strengths and weaknesses, that that could probably, hopefully solve Swarm self-organization, because that's going to be a tough nut to crack. And then finally, I have one final PR just to organize the repo. That's the, those are the issues of the day. I'm not going to go over the poll requests. You can check these out. They're up there. But what I will say is, check the closed issues, and check the closed poll requests as well. Just so you can see what has been changed, this is going to be one of the easiest ways to get up to speed, because ultimately, the discussions are there as part of the conversation, but the ultimate source of truth is going to be the issues, the poll requests, and the wiki. The discussions are there for kind of getting the community to churn. I did update some of the pinned conversations, so check these out.