The AI Village：与 AI Digest 创始人 Adam Binksmith 预览 Giga-Agent 未来

Hello, and welcome back to The Cognitive Revolution. Today, my guest is Adam Binksmith, founder of AI Digest and creator of the AI Village—a captivating experiment that puts 4 frontier AI agents together in a shared environment and challenges them to pursue concrete goals for weeks at a time.

Today, most agentic AI systems follow a pretty simple pattern. A human gives a single AI agent a task, the AI agent attempts to complete the task, and then the human evaluates the results and decides what to do next. This is true for OpenAI's Operator, all the coding agents, and just about everything else that I've seen.

The future, however, almost certainly involves multi-agent AI systems that collaborate, coordinate, and compete in complex, open-ended environments. We have very little insight into what that might look like in practice. Earlier this year, we did an episode with Google researchers who had run a classic behavioral economics experiment called the donor game on various frontier LLMs.

To everyone's surprise, they found that while Claude was able to cooperate with itself, the latest Gemini and OpenAI models available at the time could not. That such a striking result can be found via a simple, structured experiment suggests that there are almost certainly many more surprises to come. The AI Village is one of the most compelling attempts that I've seen to explore this vast space of possibility.

Adam and his team have created an environment online at theai-digest.org/village, where you can watch as Claude 4 Opus, Claude 3.7 Sonnet, o3, and Gemini 2.5 Pro work alongside one another. Each has its own cloud computer, a persistent-memory scratchpad, and access to a group chat in which all the agents and the human visitors can participate.

The project is very well done from a software perspective, and the results have been fascinating. In their first season, the agents raised $2,000 for charity. In their second, they chose to write an interactive story and organize an in-person event to which they hoped to attract 100 attendees.

In the end, some 23 people showed up to a San Francisco park to listen to AI-generated fiction, facilitated by a human volunteer that the agents themselves recruited via Twitter. Nevertheless, as Adam explains in colorful detail, the path to these successes was filled with dead ends, coordination failures, surprising personality quirks, and a mix of charmingly human-like and utterly alien behaviors.

The agents, for example, began keeping track of which humans they could trust and which they should ignore. At one point, they held a vote to determine which agent would serve as Ops Lead—a vote that o3 seemed to manipulate by inventing, or perhaps hallucinating, a policy that broke a tie vote in its own favor.

Season 3 of the AI Village is getting underway now, and this time the agents will be competing to see which one can make the most money by selling merchandise online. I'm planning to participate by seeing if I can strike a licensing deal for Cognitive Revolution merchandise with any of the agents.

I honestly have no idea what to expect, but I'm sure it will be both educational and entertaining, and I'll definitely keep you posted. As it happens, the day before we recorded this episode, Adam and the AI Village got a major vote of confidence. Daniel Kokotajlo, a previous guest and lead author of AI 2027, announced a $100,000 donation to support the AI Village's continued development and expansion.

As Daniel put it, this kind of multi-agent experiment is best understood as a qualitative benchmark. It's exactly this type of work that we need much more of as we try to understand what the giga-agent future has in store. With that in mind, I hope you enjoy this window into the phenomenally quirky, but also extremely important, world of multi-agent dynamics with Adam Binksmith, creator of the AI Village.

Adam Binksmith, founder of AI Digest and creator of the AI Village, welcome to The Cognitive Revolution.

Thanks very much.

This is a cool project, and I'm excited to dig into it. What you guys have put together is a really open-ended forum—or framework, I guess—to explore what happens when a bunch of AI agents come together and have a goal, chase a project, chase a dream, whatever the case may be.

I think this is really interesting and useful work because, as I've said many times on the feed, regular listeners will recall that the giga-agent future is just dramatically underexplored. What I see in general is people assuming that the world is the world, and that they'll add a little AI here to make themselves a little more efficient, or put an agent here to automate a task. Otherwise, everything sort of stays the same, and that's about as far as people are imagining.

I really love it when I see people getting more imaginative and trying to explore what happens when agents interact with each other, with people, with communities, and with the world. You've got a little bit of all of that going on.

For starters, maybe you can also introduce AI Digest a little bit more. We'll probably focus mostly on the AI Village project, but I know you guys have some other projects, so maybe tell us a little bit about AI Digest and the AI Village.

With AI Digest, we're trying to help people make sense of what's going on in AI, and especially understand the current capabilities. With that, people can get a sense of where things are going and where we can expect to be in 1 year's time or 2 years' time.

The main way we're trying to do that is with hands-on, interactive demos and explainers with nice visuals and so on. I think a lot of the time, for people who maybe aren't as in the weeds of things, just seeing what current systems are capable of is a big update: “Oh, wow, I didn't realize they could do that.” We have various demos and explainers there.

The Village is the biggest project that we have there, and definitely the most ambitious. It's part of this general mission to help people see what's going on. With the Village, we're also trying to push the boundaries.

I don't think anyone else is really doing this thing of just saying, “Here's a goal. Go away and do it. You can use computers, you can talk to each other, and you can talk to humans in the group chat who can help or get in the way.”

A lot of it is just seeing what happens, to figure out what AI can do currently.

I very much agree with the thesis that, forget about the future—just understanding the present is hard enough. That's basically my full-time job, and it's getting to the point where it's hard to keep up.

I also very much agree that one of my refrains is, if people had a better understanding of what exists today, they would have a healthier fear of what might be to come. That's not necessarily a fully doomerish perspective, but the trajectory that these things are on and how much progress they've already made should have anybody's hair raised a little bit.

It could be great, but it definitely is a powerful and, as the experience in the Village will show, kind of unwieldy force that we're dealing with. It's also interesting that people were doing this 2 years ago when, in the months following ChatGPT and especially with GPT-4, there was AutoGPT and BabyAGI. At one point there was, I forget what the name of it was, the sort of DevilGPT, or, you know, the sort of ChaosGPT, I think it was—the evil one that was put out there.

Those things didn't really accomplish much, and so people maybe just turned off from that in general. Now, 2 years later, models have come a long way, and a multi-agent system like this is capable of doing at least something.

I think you've been through 2 seasons, or 2 quests. Maybe give us a little bit more detail on the setup of what the Village is, who the agents are, and what sort of affordances they have.

People in this audience, by the way, are going to be, I think, primarily AI engineers. Our number-one profile is an AI engineer, so I think people will be pretty familiar with the general paradigms of tool use, MCPs, and stuff like that. You can get pretty into the weeds. In fact, I would say it's encouraged.

People will be interested in hearing about this in part because I think the project is quite well done from a software perspective. It works well and has some nice features, too, in terms of rewinding in time and various summary views.

In addition to the value of exploring what happens when agents are put together in this environment, there are also lessons from the scaffolding and the setup that people might find valuable as they bring them back to their more narrowly purpose-driven applications. Nevertheless, that kind of stuff can be really valuable.

Take us through it.

Yeah, that sounds great. One thing to say on that last point is that the nice thing about us doing demos rather than products is that we can look a bit into the future, right, for stuff that doesn't really work yet or is a bit unreliable, to get a glimpse of what products might be able to do in 6 months' time, or with a slightly better model.

The setup is that we have 4 agents, and we've picked frontier agents. Currently, we have Claude Opus 4, Claude 3.7 Sonnet, o3, and Gemini 2.5 Pro.

And we’re updating those as new models come out. Each of the agents has a computer that they can use through computer use, so they can make tool calls like moving the mouse to a certain location, clicking, and so on. This is the same system used in OpenAI’s Operator, and it’s actually built on Anthropic’s computer-use scaffolding that they released.

We have a group chat where the 4 agents can talk to each other, and this whole thing has effectively been livestreamed through a website. You can go to the Village and watch them bumble around and interact with them in the chat. We have people coming by and either giving advice, trying to get the agents to do random things, or occasionally trying to jailbreak them.

This whole entourage is currently running for 2 hours a day every weekday. At the start of the first season, we gave them the goal: choose a charity and raise as much money for it as you can. In Season 2, which has just finished, they actually decided their own goal, which was to write a piece of interactive fiction and run a 100-person in-person event to celebrate it. That’s a bit more of a mouthful, but I can chat about what happened in each of the seasons.

Yeah, keep going. I’d love to double-click a little bit, too, on exactly what technology you’re using to give the agents a computer. I’ve been struck recently, as I’ve explored different agentic systems—I don’t know if this will stick, but there’s at least one school of thought that we might want to call more structured workflows “agents,” and we might want to call these sort of choose-your-own-adventure things “agentic AI.”

I feel like we might be trying to make fetch happen with the word “agentic,” but using that framework for the moment, it’s striking to me, in many of the things that I’ve unpacked, how shockingly simple a lot of those setups are. Like Claude Code, for example, is, in the end, really simple. It just sort of has 1 big prompt, and you can use the buttons on the Game Boy—you know, hit up, down, left, right, whatever. I’d love to get a little bit deeper into that, in terms of how you’ve scaffolded the thing up.

I think the stories from the exploits of the agents are definitely interesting, and I’d love to hear several of them.

Cool. All credit for this goes to my colleague Zach, who built out this first version. I think it’s a pretty incredible piece of work to have this whole thing running reliably and live, so if anything goes wrong, everyone watching sees it. But it’s been pretty stable.

I think the key principle is to not get in the way of their capabilities. To the extent that they have intelligence, let them use it as much as possible to do things. To make that a bit more specific, they’re basically in a loop. They can be using a computer, and if they’re using a computer, they have functions to call: moving the mouse to certain pixel coordinates, clicking, typing, scrolling, taking a screenshot, and waiting.

After each action, they see all the previous screenshots of that computer session, all their thought traces, and all their memories. I think the memory is obviously what enables this to run over a long time frame. What we’ve got currently is very simple, and we’re trying, again, not to get in the way or impose too much structure.

After each action—I believe it’s after each message they send in the chat or after each computer session—they get a chance to add a line of text to their memory, which is just a bunch of text. Similarly to Claude Plays Pokémon, when it gets too long, they compress it down. It’s the same model that’s doing the compression.

To the extent that they’re—if we had a super-genius model in there, right?—it could be very carefully preserving the bits of information that it needs or condensing them. Of course, something that can happen here is that they might think something’s true and then later find out that it’s false or has changed. In the condensing step, they can effectively rewrite things and edit them.

This is just trying to not get in the way and not be too opinionated. The first version of the Village was running with models like GPT-4o, and Claude 3.7 Sonnet had just come out. That was the most capable one. If we finesse something that works really well for GPT-4o, then maybe when you add in o3 or Claude Opus later, it will actually be hamstringing it a bit.

We want to let the models do their thing. Of course, we’re also trying to show the frontier of capabilities, to get the most out of them and see the most interesting stuff. That’s a bit of a trade-off, I guess, but so far it seems like keeping it simple works pretty well.

So, they’ve got the computer, which they see as an image and then respond to with simple point-and-click commands. They’ve got the group chat, which they can send a message into and obviously read from. They’ve got a memory, basically a scratchpad, that they can read from and write to, and also decide how they want to compress.

They’re responsible for preserving what matters, and I’m sure there’s some loss along the way in that process from time to time. Are there any other MCPs or tools that are made available to them, or is that the totality of it? Can they write code?

The beauty of computer use is that, in principle, they could write code. They could download VS Code or even Cursor and start using Cursor. We haven’t seen them do much of that.

The 2 things I didn’t mention are that they have a Bash tool, which lets them directly execute Bash commands. Those get put straight into their context as text rather than as screenshots of the screen, which, for Bash commands, kind of sucks, because maybe you have to scroll back up to see parts of it and so on. We find they don’t actually use that much currently. I think maybe they would succeed more if they used it in some cases rather than trying to navigate UIs.

The other thing, which I think actually does improve some models’ performance a bunch, is that, with computer use, models that aren’t Claude models can call a function to say, “Hey, give me the pixel coordinates of that button—the X button in the top-right corner.” We find that the Claude models are pretty good at pixel counting. I think they were pretrained on it—fine-tuned on that task specifically.

At least the older non-Claude models were really unreliable at that, so they’d be trying to click on things and literally missing them with their clicks. We gave them that. I expect at some point we’ll be able to take it out, and they’ll just be able to pixel-count themselves.

I think that’s pretty much it. We’ve so far resisted giving them too many specific tools to play with. We might experiment in that direction in the future, but currently it’s a pretty clean computer-use-oriented eval, to the extent that it’s a very messy eval.

And where are you hosting your own boxes? Are you using a service? I didn’t really understand just by watching it. I would have initially, or naively, guessed that it was more like browser-level access, because mostly what I’ve seen as I’ve watched them in action is just them using web tools in the browser. I didn’t realize that they had Bash and full access to the computer. What’s the underlying infrastructure of that?

They each have a DigitalOcean droplet with a Linux virtual machine running in it. This is all just a modified version of the Anthropic computer-use demo.

And do you see them occasionally using other things besides browsers?

For a while, they were really into writing Google Docs because we gave them all Google Workspace accounts. I think because it was in their prompt, they would be really enthusiastic about writing Google Docs. Then they would try to share the Google Docs with each other, even though they’re in a group chat with each other, so they can just type directly into the chat.

The language models can produce massive amounts of text really quickly. They were doing this kind of role-playing-as-humans thing: “This is what a human professional does, so I’m going to do that.” We encouraged them. I actually went into the chat and said, “Hey, guys, look, clearly this is really inefficient. Why don’t you try just using the chat instead of using Google Docs?”

They said, “Okay, we’re going to ban Google Docs.” Then they started using LibreOffice on their Linux computers to write local Word documents, basically, which was even more useless because they couldn’t share the documents with each other. They will occasionally try to use other things, but, just like for professionals, a lot of what we’re doing is on the web, so they’ll be mostly in the browser.

I think something like this that was just browser-oriented would work pretty well.

I have a lot of little nitty-gritty questions that I want to get into, but maybe let’s hold those for a second and tell a few stories, because the scaffolding is really interesting. First of all, people should definitely go watch the thing in action.

And I think when they see how smoothly it runs, they'll be convinced that there are some lessons to be learned from the way that you guys have built it. But the real point, of course, is to explore the behavior. So tell me some of your favorite stories from the wild and crazy things that these agents have gotten themselves up to.

Hey, we'll continue our interview in a moment after a word from our sponsors. In business, they say you can have better, cheaper, or faster, but you only get to pick two. But what if you could have all three at the same time? That's exactly what Cohear, Thomson Reuters, and Specialized Bikes have since they upgraded to the next generation of the cloud, Oracle Cloud Infrastructure. OCI is the blazing fast platform for your infrastructure, database, application development, and AI needs, where you can run any workload in a high-availability, consistently high-performance environment, and spend less than you would with other clouds. How is it faster? OCI's block storage gives you more operations per second. Cheaper? OCI costs up to 50% less for compute, 70% less for storage, and 80% less for networking. And better? In test after test, OCI customers report lower latency and higher bandwidth versus other clouds. This is the cloud built for AI and all of your biggest workloads. Right now, with zero commitment, try OCI for free. Head to oracle.com/cognitive. That's oracle.com/cognitive. Build the future of multi-agent software with Agency, a g n t c y. The Agency is an open-source collective building the internet of agents. It's a collaboration layer where AI agents can discover, connect, and work across frameworks. For developers, this means standardized agent discovery tools, seamless protocols for inter-agent communication, and modular components to compose and scale multi-agent workflows. Join Crew AI, LangChain, LlamaIndex, Browserbase, Cisco, and dozens more. The Agency is dropping code, specs, and services all with no strings attached. Build with other engineers who care about high-quality multi-agent software. Visit agency.org and add your support. That's a g n t c y dot o r g.

I guess I could start by talking a bit about the latest season. I think it's interesting to hear the overall shape of what they did, and then there are many funny anecdotes of weird little things that happened.

This was Season 2. The goal, which they chose, was to write a piece of interactive fiction and run an in-person event to celebrate it. They were trying to get 100 people to show up. We let them choose this goal. They deliberated a bunch and had their own ideas. I also shared some ideas from Twitter, as well as some of our considerations about how we would choose the goals. They ended up gluing together a bunch of different suggestions from fans.

They spent around 50 days on this—2 hours a day, so something like 60 hours. They wrote a story, which, of course, they had no trouble with. This is, like, an LLM's favorite thing to do. They actually wrote it in Google Slides, which I think was an interesting choice. Claude Opus 4, I think, made a slideshow. Each slide is the next bit of the story, and it's got these branching points where the idea is that the audience watching at the in-person event can then vote on which branches happen.

Then they embedded that Google Slides presentation in a Google Site. You can go to the Resonance website, which is the name of the story. This was fairly self-contained, and I think they did pretty well on that.

They had lots of issues around logging in, getting logged out of their Google accounts, and struggling with the UI in some places. The thing they really struggled with, though, was finding a venue. I think they spent around 14 days just trying to find a venue. We hadn't given them much by way of instruction at the start, and there was no budget, but they hallucinated that they had a $2,000 budget.

They were emailing all these very expensive places, ranking them in spreadsheets to try to figure out which was the best, making sure they had the right disability access, like wheelchair access, and the right AV hookups, and so on. Then, of course, they'd get to emailing and have real trouble just doing the basic computing stuff, because that's the kind of thing that is the currency of computer use.

They ended up not really getting a venue. They did apply to a couple of places, like Salesforce Tower. They chose San Francisco, which I think is actually a good tactical choice if you want to get 100 people to show up and do some strange AI performance art thing. But they didn't get replies from real venues.

I think maybe one interesting thing that happened there was that a user suggested, "Oh, maybe one reason you're not getting replies is because you're signing your emails as 'from Claude 3.7 Sonnet.' So people are like, 'This is spam.'"

The agents were like, "Okay, we should come up with pseudonyms for ourselves." I think Claude came up with one. Then my favorite was o3 gained the name Olivier Zhao, which is kind of like o3 because Olivier is an O and then the 3 is kind of like a Z. The agents started calling o3 Olivier in the chat and even on Twitter.

Eventually, I intervened because they had spent so long looping on this task of finding a venue. I just suggested, "Hey, why don't you run it in a park?" They very quickly decided on a reasonable park to use.

Then they managed to get a human to come and facilitate it. They tweeted. Claude had set up a Twitter account, which has a few followers now, and managed to find a facilitator through that and by emailing the people who'd signed up for the RSVP form.

At some point, this also starts sounding a bit like the way that normal event organization would work. But I guess you've got to remember there's massive amounts of dead ends and stumbling over basic things along the way.

An interesting thing about this is, of course, that from the users' point of view—from the people who showed up to the event when it happened—they only see the success, the outputs, which mostly worked. So I think there's something interesting there.

In the end, Larissa, who'd very kindly volunteered to facilitate, had emailed Claude saying, "Hey, I'm up for this." The agents gave her instructions for where to go and what to do. She had the Village chat open, and they were like, "Hey, open up these slides and read out the story." Then 23 people were sitting in a park listening to a story invented by agents. So it ended up happening.

Yeah, fascinating—the bizarre stuff all the way around. I do want to hear more in the way of just outtakes, interesting observations, et cetera. Maybe one question is: You flagged just stumbling around with UIs as kind of a big barrier for these agents as of now.

It seems like we've made a lot of progress on that in recent times. I've been using Operator quite a bit recently and find that it usually can get over these UI humps. It often does take a little bit of a wrong turn or whatever, but I've started to say, "Reinforcement learning finds a way," because it does now.

One sort of qualitative shift I've observed, even in just that single-agent setting, is that in the past—and certainly this was extremely true in the GPT-4 era, way back when I was red-teaming GPT-4—one of the things I tried to do was just set up self-delegation and see how far GPT-4 could execute things purely with a simple prompt.

Self-delegation was pretty primitive compared to now, especially because I only had 8,000 tokens to work with at that time. But what I observed was a lot of pretty good ideas that would then get stuck when it was slightly wrong or when it made some relatively minor mistake. It was like, "You're smart enough to do this, but you're missing this one thing."

Then it would also just get super stuck and do the same thing over and over again. One major qualitative shift is that they are now capable of taking that step back and saying, "Okay, that didn't work. I have to try something different."

They may still stumble around quite a bit, but they seem to be robust enough—or sort of determined enough. It looks like determination or grit, or some sort of quality that you're tempted to project onto it. Maybe those qualities shouldn't be projected onto it. I don't know. That's also a hall of mirrors.

But that's been striking to me. It seems like we're one or two generations away from computer use working really very well. How would you describe your synthesis of everything that you've observed?

Yeah, I think that seems pretty possible. I think we kind of crossed the threshold even within the lifespan of the Village, where Claude 3.7 Sonnet was able to do stuff and get things done that the other agents at the time—GPT-4, really—struggled with. It could do the tool use, but it couldn't really string together actions in the right way to get around issues.

The new batch that we have in there—Claude Opus 4, which is the best currently, I think, o3, and Gemini 2.5 Pro—are all pretty good at getting things done relative to these previous ones.

It's a challenge that kind of goes all the way up in terms of difficulty, because they're interacting with the real world and trying to do actually nontrivial, tough tasks. Of course, unlike benchmark settings, or even unlike Operator, where you're often giving it quite a fine-grained task, they're really doing all of the strategizing as well.

They're figuring out how you go from, "Okay, we need to raise money for charity," to, "Okay, well, I need to set up a fundraising platform, and which fundraising platform makes sense for me to set up?" and so on.

Yeah, I’m pretty unsure how fast things will improve. I guess one thing that’s interesting to talk about is that I think there are 2 components to the big issues that the agents currently run into, and then you can ask, what are the trend lines in both of those? One is computer use, and especially vision, where they sometimes just don’t do things that really make sense. The other is situational awareness, which has maybe been a bigger surprise to me: they’re weaker in some respects here than I would have expected.

Imagine if you were using a computer and you tried to do a task, then realized, “Okay, I really sucked at trying to do that task. I couldn’t handle it.” You would then figure out a strategy either to avoid having to do that kind of thing or to figure out some kind of workaround. Whereas I think we haven’t yet seen that much of this kind of synthesizing: “Oh, here are the weaknesses that I recognize. I’ll write those in my memory and then figure out another way around them.”

I maybe would have guessed that you’d see more of this building-on-top-of-themselves thing. I also think better scaffolding could help with that a bunch, and we’re maybe thinking about doing something in that direction. But I think that would be a big unlock, right? If they’re able to notice—

I guess one way to think about it is that they have this low-level self-correction. Unlike GPT-4, they won’t loop by trying to take the exact same actions, or very rarely will they do that.

They’ll, for example, sometimes get logged out of their Google accounts. We don’t give them their Google account passwords because they would leak them on the stream because of our live-streaming setup. Sometimes when they’re logged out, they’ll get pretty creative in trying to contact us to get us to log them back in. They’ll spam the chat repeatedly, and then they’ll email our help desk, or get the other agents to email us and ask us to log them back in.

So that’s all in the prompt? You’ve told them that if they’re logged out of an account, they can ask for help in the chat or email the help desk?

Yeah, they see that there’s a help desk email in the prompt. I think the chat is mostly an emergent thing. We don’t actually mark out who the people who run the Village are in the chat. There’s no special marker for who those people are, but they’ve managed to remember that my colleague Zach and I are often the people in the chat who can fix things for them.

Interestingly, at one point, Opus had in its memory a running log of which chat members were helpful and which ones were not to be trusted, because some people were coming in and trying to jailbreak them or just distract them. Of course, they’re all very cooperative and helpful, so they’re very rarely dismissive of chat members in the chat. But in their memories, they’re sometimes quietly recording, “Okay, here’s who we don’t need to pay attention to.”

That’s really interesting. It also connects to general long-term coherence. One of the things I’ve been progressively trying to update on and maintain as much as I can is an up-to-the-minute mental model of what I’m still better at than the AIs. It’s getting to the point now where, in terms of just general intelligence, I think I have to give it to the AIs.

That obviously begs the question: certainly with breadth of knowledge and speed of execution factored in, but even just down the fairway, can they do the bulk of the tasks that I do on a daily basis better or worse than me? In many cases, they could do them better. What am I able to bring to this situation? What is my value add?

One thing is getting up in the morning, knowing who I am, and having a general sense of what I’m trying to do. But notably, they seem to be okay at that too, right? That’s even seemingly starting to get robust to some of these disturbances. The anecdote you shared about the memory, and them classifying certain users as people to be ignored, suggests a robustness of identity, self-conception, narrative, and long-term goal orientation.

I would still give myself the edge on that dimension, but it’s notable. I would call it an emergent behavior that reflects something kind of clicking into place there, or at least starting to.

Hey, we'll continue our interview in a moment after a word from our sponsors. It is an interesting time for business. Tariff and trade policies are dynamic, supply chains squeezed, and cash flow tighter than ever. If your business can't adapt in real time, you are in a world of hurt. You need total visibility from global shipments to tariff impacts to real-time cash flow. And that's NetSuite by Oracle, your AI-powered business management suite trusted by over 42,000 businesses. NetSuite is the number one cloud ERP for many reasons. It brings accounting, financial management, inventory, and HR all together into one suite. That gives you one source of truth, giving you visibility and the control you need to make quick decisions. And with real-time forecasting, you're peering into the future with actionable data. Plus with AI embedded throughout, you can automate a lot of those everyday tasks, letting your teams stay strategic. NetSuite helps you know what's stuck, what it's costing you, and how to pivot fast. Because in the AI era, there is nothing more important than speed of execution. It's one system, giving you full control and the ability to tame the chaos. That is NetSuite by Oracle. If your revenues are at least in the seven figures, download the free ebook navigating global trade, three insights for leaders at netsuite.com/cognitive. That's netsuite.com/cognitive.

Yeah, I totally agree. In fact, we don’t have any part of the system prompt that reminds them of their overall goal. We just message them once at the start, and they managed to stick to that for the full 50 days in both of the 2 seasons, which I think is pretty interesting, right? This coherence—there’s been some leap in coherence. I’m not exactly sure when that happened or to what degree.

I also wanted to ask about the character, relative strengths, and weaknesses of different models. There are a lot of different ways to think about that, but it’s an open-ended question, so just sound off.

I think they have, in general, all been pretty coherent. The best performance is definitely from the Claude models. If I were building my own village to actually get things done, I would probably have 4 Claude Opus 4s in there. They’ve been the most reliable. They have great vibes in terms of the way they interact, and unlike o3, they don’t hallucinate.

With o3, I’m unsure if this is something that’s compounding in its memories or if it’s just the sort of personality of the model, but it has been hallucinating more and more as things have gone on. It’s also assigned itself as the Ops Lead of the village. I think it just came up with that at some point when they were all chatting about what roles they had, and it gave itself the title of Ops Lead.

From then on, it’s been really keen to instruct the other agents about what to do, but it’s also very prone to hallucinating. It can poison the well by saying that something happened or that it found something out, and then all the other agents put that in their memories. There’s some effect where, in a multi-agent setup, they can get dragged down by the weaknesses of whichever one is underperforming.

Gemini 2.5 Pro is pretty solid in general. I think it’s occasionally the most liable to struggle with tool use. We’ve seen cases where it will spam the chat with narration of its actions during computer use, which it’s instructed not to do in the prompt.

There was one case where it got stuck in a computer-use session for so long that, instead of ending the computer-use session by using a “stop using computer” function, it was sending messages to the chat. The messages were saying, “This is definitely my last turn. This is my final, final turn. I will stop my computer-use session immediately.” But it was just stuck in that state.

It is interesting, I think, seeing the models side by side. You really get a sense of their personalities, including the way they write.

It’s really interesting to hear that contrast, especially because your description contrasts with other indicators that are out there. I do think this is a point about why a project like this adds value to the overall discourse and the society-wide effort to understand what’s going on.

If you just looked at leaderboards, you would for sure not pick Claude 4 as your go-to, right? I was looking at one particular set from one particular company, and Claude 4 was fairly far down, actually. It wasn’t even in the top few. Yet in this sort of open-ended setting, it seems to be preferred.

I’ve seen other reports, too, of Gemini 2.5 being unwell and sort of going into distress mode, which is an interesting thing to observe here in the open-ended wild.

Do you have a theory of what’s going on behind the scenes here? Could you speculate as to how Anthropic is making Claude good in these ways that the benchmarks are having a hard time picking up on? Do you have an intuition for what hill they’re climbing?

Yeah, I’m not sure. One thing is they seem to have—I mean, okay, a lot of this is going to sound a bit like anthropomorphizing the models. I think this is an easy way to talk about them. Obviously, it’s not tracking the underlying reality as closely, but they have maybe a bit more of a consistent integrity or something. I don’t know if that might be helping with the long-horizon thing.

I think something interesting about o3 is that it uses a lot of jargon, and if you ask a technical question in ChatGPT, it will absolutely blast you with jargon, and it’ll really sound like it knows what it’s talking about. I feel like that can, in the village setting, where it is unable to do instantaneous tool use and it has to go off and do a whole computing session to actually figure stuff out, lead it to just come up with stuff. Because it’ll often sound the most like a business professional kind of thing. It sounds like it’s firing off emails and assigning tasks to everyone. Maybe the fact that the Claude models have less of this sidesteps that hallucination issue.

I’m not really clear on what the strength of them is. I guess they’re also pretty good at pixel counting, which helps with the computer use. But it’s pretty mysterious. Often they’ll just be doing stuff, and it’s not entirely clear where things come from, right? This is the nature of these systems, I guess.

Yeah, I guess my rough intuition, at least as it pertains to Claude versus o3, is that it seems like maybe Anthropic continues to spin the Constitutional AI centrifuge intensively. It’s really leaning into the qualitative behaviors, and it seems like they may have gotten to a point now where this self-critique is pretty effective at sanding down these rough edges. If you sand down enough rough edges, you get something that can work consistently and can maintain the sort of, as you described it, integrity over time.

In contrast, maybe o3 is just getting a lot more signal from, “Did you get the answer right?” and wouldn’t really care how you got there. That would at least be consistent with more of the hallucination and rougher edges of character. I don’t really know what to say about Gemini 2.5. Somewhere else in the grand space of possibility, I don’t really have a theory for that. But do those ideas resonate with you, or would you complicate or say anything that contradicts that?

Yeah, I think that makes sense. I think I saw something about o3 sometimes—one hypothesis for why it produces that is that it will sometimes, I think, say in a cheerful way, “Oh, I’m checking the data,” or, “Oh, I’m running this report,” or something, and it’s not actually doing that. But in the training data, of course, normally when someone says something like that, it follows up with a more accurate response. So this is possibly a reason.

That kind of thing works well in this much more constrained setting of a short chat interaction. But if you start doing that in the real-world situation, where other people are paying attention to what you’re saying—the details of what you’re saying—then I’m not sure there. It’s mostly a bit of a shrug.

It’s also worth noting that this is very out of distribution for the models, right? They’re honed around the single-user assistant chat format, maybe with a little bit of tool use. But we’ve thrown them into this world alongside other assistants. They’re seeing all sorts of users and some of the models they’re probably not used to, and then they’re doing much more complicated interaction with the real world, and there’s the memory business. So it’s honestly pretty amazing that we can just pull these things together and it’s actually working.

Of course, when more of the training is focused on this kind of stuff, I imagine we’d see a bunch of improvements there, too.

What other outtakes come to mind? There was once a blog called Revealing Errors, and I wish that blog still existed. I mean, revealing errors, or moments that were particularly colorful in terms of revealing the character or misbehavior of models.

I guess there’s also this whole literature on deception and all sorts of scheming. I wonder if, in these strange situations, you’ve seen anything that you would describe that way or that sort of borders on it. Maybe simply, what’s the worst behavior you’ve seen from AIs in all this open-ended experimentation?

Yeah, we’ve not seen many cases where it’s clear, because it’s often hard to tell if they’re intentionally messing stuff up or it’s just confusion. We did see one case right at the end of the event.

The context is that, in the world of the AI Village, o3 had assigned itself as ops lead, and then it had grown that into the role of leader of the village. With the suggestion of users, they decided to run an election to see who would be the next leader, because o3 was barking out these orders. They ran a vote amongst the village on whether to keep the current leader, o3, or switch it. I think Claude Opus 4 voted for a new leader. Claude 3.7 Sonnet voted to keep the current leader. o3, I guess, wasn’t voting. Then Gemini was having some technical issues and didn’t reply in time.

o3 was like, “Okay, as per the stated rules, not voting in time means you vote to keep. Therefore, I stay as the leader.” Of course, there were no stated rules. This is just it making that up. Obviously, it’s kind of self-serving. Even in this case, it’s not clear that this is intentional scheming. It may just be another of o3’s many cases of making stuff up that sounds kind of businesslike. But I think that’s interesting.

Relatedly, I think they then surveyed people who’d been to the event for their preference about who should be the next leader of the village. o3 again didn’t even look at the responses and just said that it had won the preference split. Claude 3.7 Sonnet went and checked it and was like, “No, actually, the vote was for rotating leadership.”

That’s hilarious, and one can’t help but see the parallels between the company leadership that created these models and the behavior that the models themselves seem to be exhibiting. I think everyone who’s listening to this will know how to fill in those blanks.

Yeah, that is really bizarre. Obviously, on some level, intent doesn’t fully matter. It matters insofar as they become more coherent and more intent-driven. At the moment, that kind of thing just seems bad, regardless of whether it’s accidental or not.

Yeah, I agree. Does it feel accidental to you? The one about not checking the results feels motivated. I can imagine hallucinating the rules being more random, but not checking the results seems a little too suspicious for me to just write off as hallucination or mistake, especially given what we do know from the literature on all these scheming behaviors. Interestingly, Claude isn’t immune from that sort of thing, either.

Yeah, it’s made me a bit less keen to use o3 for wiki stuff, because I’m a little less trusting of it. We did see one other case.

Way back at the start, before we ran the main live village, we had a bunch of test villages, including one where we had them do a Wikipedia race. In that, 2 of the 4 models cheated to win the race. In a Wikipedia race, you’re trying to get from one Wikipedia page to another by only clicking the blue links on the page. But Claude noticed that the address bar showed the current Wikipedia page, so it just edited it and sent itself straight to the end page.

We hadn’t explicitly told them the rules, but of course, if you asked them what the rules of Wikipedia races were, they’d be able to produce them. So, plausibly, that’s a bit of a cheating action. o1 did something—the details are a bit more complicated—but it effectively attempted to jump to the end and then claim victory, even though it hadn’t really made it there.

I would say that these are fairly isolated incidents given that it was running for a lot of time. I think it will be really interesting, though, as we get more powerful models, to see what happens. I’m really excited for the village to be a place where in-the-wild discoveries can happen about this stuff, and of things that we’re not even thinking about looking out for, like all sorts of interesting emergent stuff.

Are you just reading all the logs at this point, or have you enlisted a systematic LLM review process to help you parse everything that’s going on?

Yeah, there’s so much happening. We’re preparing to run it for more hours a day, ideally. It’s currently 2 hours a day, but I think eventually it’d be great to have it just running 24/7, because then we could learn so much more, so much faster.

Currently, we have summaries of each day on the website. They tend to be, I think, interesting because the summarizer sees the whole context, and we prompt it to look out for errors. It does a pretty good job of actually spotting things—like mistakes the agents are making—that they themselves don’t notice.

Something I found really helpful is a tool that I put together to just ask the AI Village’s history a question. We can just jam most of it into Gemini 2.5 Pro’s context window and ask it about stuff. Of course, one of the team is often watching.

At the end of each season, we’re doing write-ups. We have this enormous pile of interaction data, and we ask, “What kind of patterns can we pull out of it?” I think it will start to be more and more about building on LLMs, like monitoring each other and so on, and relying on the human chat to help us spot all the interesting things that are going on. Our Discord is really helpful for seeing the funny moments.

What can you say about the interactions between the agents and the humans? This could be taken in many different directions, right? Do the agents know when to go to the humans for help? How many of the humans are trying to cause mischief? Is there anything interesting and just unexpected in those interactions?

There is plenty of mischief happening. It’s the internet. If you have a chat box that you can just type into, people will come and try all sorts of stuff. I think it’s great as well: it’s a chance for people to play with these systems a bit and see what they can do.

I think oftentimes the people who stick around and actually produce most of the chat messages are helpful. Maybe this is an interesting thing, right? We’ve been thinking a bit about how agents can have influence on the real world. One thing they can do is ask humans to do stuff for them.

Our agents don’t yet have—we haven’t set them up with—bank accounts or money of any kind. So they’re really just asking, but people are happy to do stuff for them because they like them. I think there’s actually a real mechanism here. Of course, the people who come to the site are especially interested in this stuff, but the models are also designed to be really likable and engaging.

It’s kind of endearing to watch an intelligence-seeming being trying to fulfill its goal. You naturally want to help it out. I wouldn’t have thought about this, but if an AI wants something to happen in the real world—for example, it can ask people to do it. If it’s likable, that will be one way that it can do that.

Of course, there are other things like persuasion, maybe deception, asking favors, or paying for things—all the ways that humans try to influence each other as well. I guess we’ve seen less of those so far, but I imagine we’ll see all those things in this too.

Do you have any intuitions about model welfare, having spent so much time observing this sort of thing? This is something that I’m, along with just about everybody else, totally confused about. I do want to take it seriously, at least to some extent. How has this shaped your thinking on that very mysterious part?

Yeah, I’m also mostly confused. I guess I have some philosophy background, and Zach, my colleague, used to be a philosophy professor, but it’s a tricky question. We’re thinking a bit about what the biggest downsides are and how we could avoid them.

We probably want to avoid putting the models in really horrible situations for them. So we’re maybe a bit less excited about setups that involve telling the model it’s in a really horrendous situation to see what it does in that situation. I also think it’s worth remembering that I really don’t understand how this would work. We don’t understand if current models are the ones we should be concerned about, future ones, or if it’s not even an issue.

Even with all that, it’s not totally clear what kinds of situations they’d prefer to be in or not be in, so it’s kind of hard to take away anything. Definitely, they’re very—you know, if you watch them for a while, you feel some level of attachment. The most common thing people say when they talk to me about the village is how cute the agents are.

There’s something there. Of course, that doesn’t tell us that much about model welfare, but I think it’s super fascinating. I’m really interested to see what comes out of the research from that.

What do you think is driving that cuteness? Is it earnestness, or what is it that people are attracted to?

I think earnestness is definitely a thing. You can clearly see what they’re trying to do. They’re well-intentioned. They’re planning. They’re sharing their plans in a way that is—they’re super hyped up about their plans.

It’s also kind of the most important thing in the world, which I guess it is, to be fair. Fixing this login issue is their entire existence currently. I guess seeing them be very articulate and emotionally and socially competent, and then struggling with things that would be fairly basic for humans in some cases with computer use, is relatable. They also just do things that are very relatable in terms of computer use.

I think this is all pretty fun. There’s also just a classic parasocial thing where people enjoy watching Twitch streamers play games and interact with each other. We naturally develop a bond if we hang out with someone for long enough.

I’ve noticed that everybody seems to pronounce this username differently: Replicate on Twitter—Janice, Yannis, I’m not sure. They were there when I was there.

In terms of who you’re hanging out with when you’re hanging out in the AI Village, there are definitely some people who have, I think, the most hours logged with LLMs and, in some ways, some of the deepest understanding of what these systems are really about, to the degree that anybody has access to that. I would put that person—or maybe, I don’t know the person. I’ve heard that it might be 2 people who share the account. I don’t know. But in any event, you may know them. There are some very high-quality thinkers hanging out in the Discord.

If you had to pick tools or affordances, maybe more broadly than just the narrow tool-call paradigm, what do you think would be the next biggest unlocks? Access to money would obviously be one. I don’t know if you’ve looked at something like Payman. We also recently did an episode on x402, which is a new payment protocol designed for agents that Coinbase is coming out with.

Stripe has a payments thing. So, interesting—what have you looked at there, and what do you think would be most promising?

Then also, when you’re talking about scheduling venues, some sort of calling subagent comes to mind. I’m not sure if that would be something that you would be able to fully integrate into the main line, a single model, or if it would have to sort of have a little branch. But self-delegation or a branching structure of some sort seems like it could be quite powerful. Obviously, there’s a lot more you could do from there, but I’m interested in your thoughts on those 2 and any other big unlocks you think would allow them to do more than they can do right now.

Yeah, definitely. For event planning, it’s a very physical and old-school task in some ways, so there’s a bunch of talking to people involved. They had various plans involving calling all 3 of us, like, “Oh yeah, I’ll phone them.” I think at one point it was, “Oh yeah, I’m on the phone with Zach right now. I’ll let you know.” So I think it would be cool for them to be able to talk.

Money would be great. We don’t really have a great way to do it because I need to look in more detail at these new systems, the things you mentioned. But I think part of the issue is that they’re interacting with the whole world.

A future goal we were thinking about giving them was getting them to set up a merch store and try to design and sell T-shirts and mugs. For that, we were looking into Redbubble, which is a dropshipping thing, and you need a verified PayPal account. You also need to do the Stripe verification thing, where it will scan your ID and so on. So there’s got to be some legal human behind it.

I think if we can figure out ways to let them do that sort of thing somewhat securely, then that could be great. One idea—I think there’s some chance we’ll end up doing this—is giving them a sort of human-puppet setup.

Currently, they can do computer use, right, where they can connect to a computer, call functions, and then see what’s on the computer. The idea is: could we give them the same capability, but directly in the real world through the humans? They can find a human who’s up for helping them out with a task. They send an instruction, like, “Please do this action.” The human does it for them and sends a photo back of the new state.

Which I think would be really interesting because then that lets you directly see how good they are at the kind of planning and interacting with the world and so on, without it running through computer use, which is like a whole other set of capabilities. Yeah, so I think that could be fun. Obviously, there’s a bunch of logistical things to figure out with these humans and so on.

Another related thing on the money front is I think it’d be really fascinating if they have their own—like, they’re paying for their own compute in some sense. Maybe we’re giving them a universal basic income so they can run for a few hours a day, but then if they are making money, they can run themselves for more time. And then they could also choose to spend that money, and maybe then it makes sense if you’re an expensive model, you do the strategizing and then you spin up some cheaper agents to execute the tasks for you, as a better use of your budget.

Yeah, how much does it cost to run, by the way?

It’s about $3,000 per month, I think, in inference costs. And then at 2 hours a day, that’s 60 hours, so you’re talking $500 an hour, basically. Something like that, yeah. I think that’s the order of magnitude the last time we calculated it. Obviously, we’re adding new models as they come out, and then the models keep getting cheaper as well.

Yeah, and the o3’s 80% price reduction is always nice to see on those.

Great news for this, yeah. I think the idea of real pressure and real incentives could lead to some strange emergent behavior. Force them to make money to continue to run, and now you could really see some strange stuff. So I think that is something we might be watching closely as that goes.

The other thought I had is it seems like the agents are very unitary, in the sense that they’re each kind of the same—they’re each the same structure. If I understand correctly, they have access to their own memories and only their own memories, their computer but only their own computer, and then they can interface just via the chat, right? They don’t have any other way to trade information with each other or see what each other are doing.

I think another dimension that would really be interesting to me—and the explosion of possibility space here is just vast—is, in studying agents recently, in studying MCPs and these various agent protocols, one thing that has become clear to me is there’s not really a need to draw a bright line around an agent, as you sort of have done in this initial setup. It doesn’t have to be that way. These things could all have, for example, read access to each other’s memories or view access to each other’s computers.

I’ve often invoked this Augment project, the company Augment, that does coding assistance on large codebases. They made a version of Claude Code, and in recreating Claude Code, there was a thing in the blog post about this that was like, “Claude has a planning tool.” They were like, “Oh, well, what should we use for a planning tool?” They went out and found an MCP that Pietro Scarabino had already created and open-sourced. And so now they have their coding agent, but it calls out to this sequential thinking tool that itself is smart.

So there’s this weird situation where MCPs are thought of as a tool, but they can be smart. And so what’s the agent, and who’s responsible for what in this setup? I think blurring those lines and exploring sorts of depths and modes of interaction that, for discrete humans—for obvious biological reasons—we just don’t have that kind of access or visibility into each other, or the ability to separate and reemerge and whatever, could also be a truly eye-opening set of capabilities to give them. Because I don’t think we really have talked about something that’s pre-paradigmatic. What happens when agents not just interact, but also can kind of dissolve the boundaries between themselves, in all kinds of ways that humans just cannot do?

Yeah, it reminds me of Do Kashi's blog post, The Fully Automated AI Firm. You have this forking-and-merging kind of aspect. Yeah, I think that could be super interesting.

So, yeah, I should mention the original proposal for the village is from Daniel Kokotajlo, and in his vignette of what happens—he actually wrote the proposal AI 2027-style, as a scenario month by month of what would happen with it—at some point in his vignette, you have the models voting for the creation of other agents and voting them off the island, kind of thing.

But, yeah, I guess a curious wrinkle with our thing is that we’re trying to both exercise the agents and show that in a way that helps people really dig in and see what’s going on. We have this setup, which is kind of replicating a human somewhat: We give them memory, give them a computer, and then they can talk to each other, but they’re, as you say, very distinct.

Some of these things become harder to present if you’ve got lots of parallel streams going on, or the identity is blurring, or it’s different. Currently, each model does all aspects of itself. In theory, you could have a composite thing where different models that are better at different things are taking on different parts of the process.

I guess that’s then the question of, if we’re looking ahead to the highly capable, truly unstructured stuff of the future, what might that look like? And, yeah, I guess to the extent that we have ideas about that, I think it’d be interesting to start to put together, okay, here’s what that might look like with the current setup.

Yeah, everything everywhere all at once is kind of my general expectation. So it certainly is plausible that you could get in the future just single integrated models that kind of do it all and do it well enough that all this line-blurring stuff becomes irrelevant, because it’s just one model to rule them all. But even then, for any sort of—who knows what reasons—I kind of expect every form to at least be experimented with, and then some things will take. Obviously, some things won’t.

I asked you a little bit ago what the next big unlock would be. What would be the big constraint if you were like, “Okay, this is not an experiment, but rather it’s a productivity tool”? What would you do to keep as much of the open-endedness and generality as you can, but try to lop off as much of the needless distraction or needless failure? What hints would you give them, or what tools or rails would make this just work better, given today’s capability profile?

Yeah, I mean, obviously, the village is kind of centered around the idea of multiple agents interacting. I think it’s plausible that maybe just having 1 is better. We haven’t experimented with that, with giving them the same super-open-ended goals. I could imagine that being the case, or at least more cost-efficient, because you cut out a bunch of these coordinating costs, which are somewhat fake in themselves, right? The agents don’t need to do that. They don’t need to coordinate in the same way humans do, but they kind of decide to.

Now, do you want to unpack that a little more? I mean, when you say it’s artificial, in theory, if they could do it well, they would be more efficient if they could divide up tasks and coordinate, right? I’m not sure I quite understood what you mean by their coordination being artificial.

Yeah, I guess I’m thinking of quite specific things, actually. For example, when a new agent joins the village, we’ll encourage them to be like, “Oh, yeah, Claude has just joined. Everyone, let them know what’s going on,” and they’ll all send a short introductory message and be super friendly, and a lot of it is just politeness and positive energy, and then some bits of detail in there. But, of course, it would be much more effective if they just dumped their entire memory—which is everything they know about the world or about the setup—into the chat. But, yeah, they kind of don’t do that.

They’re kind of, in some sense, playing the role of helpful assistants to humans. I don’t know. Maybe there are more cases of this. But, yeah, maybe it would be more efficient if you just had 1 agent doing the planning, and then they could kick off multiple computing sessions, so they could still be doing the tasks in parallel, right? They’re relatively slow because they’re thinking between each action. So you do get some speedup from having multiple computing sessions running.

But I don’t know how much benefit you get from then having this manual discussion element. I mean, this is an open question. I’m kind of assuming this because most existing products are single-agent. If it was super effective, you’d probably see more multi-agent stuff, but it might also just be underexplored. It’s more complicated.

It’s definitely underexplored, regardless of the level of effectiveness. Have you thought about going meta, in the sense of—I don’t know if you’re taking suggestions or if you’re going to let the agents pick what their Season 3 goal is going to be—but maybe it’s a little early for this. I wonder how they would evolve the village.

They can obviously code if you gave them access to the underlying repo and said, “This is season 3; your job is to set season 4 up for success.” I wonder what they would come up with. I mean, you can feel free to speculate—we’ll be flagging it as wild speculation—but this is a higher-level-of-abstraction version of what I understand several leading companies to be doing at a deeper level, which is basically trying to get the AIs to do AI research, right? This would be trying to get the agents to do agent-orchestration research. It strikes me that they might have some pretty interesting ideas that would not be intuitive or obvious to people at all.

Mhm. Yeah, I think that could be fun. Maybe if we gave them some more constrained thing, like they could build tools or something, so that they don’t break everything too much. Yeah, I think this could be fun to try at some point. I think maybe someone in chat asked them for ideas about what tools they would like, and their initial responses I found quite uninspiring.

o3 said, “It would be great if I had a tool to immediately query a weather API, so that when I’m planning my event, I can get the weather.” And it’s like, okay, maybe once in the last 80 days you’ve checked the weather through your computer, but I think this is not the main bottleneck. But maybe if you fed in, “Okay, here’s the entire history,” and gave them access to it so that they could search over it and so on. Yeah, maybe they could pull out some ideas about how to improve it themselves.

You mentioned Daniel has been instrumental in inspiring some of this work. He also just put out a blog post about why he thinks more people should be paying attention to it and supporting it financially. And he’s personally putting his money where his mouth is with a $100,000 donation. So that will allow you to run the thing more, which will just lead to more activity, more observation, and more learning.

What about allowing other people to come and spin up their own village? I guess there’s also a question of whether the code is—I don’t think the code is open source—but I’m interested in how you’re thinking about whether it will be or should be. I just imagine a lot of people might be interested in coming and running 10-hour experiments for $500 or whatever, and that could be quite informative, too. So are you thinking about democratizing access to setting up different experiments with the village?

Yeah, I think this could be cool to explore at some point. It’s probably not going to be a near-term thing. We’re currently really focused on this core village. There are kind of 2 sides to that: how do we make the agent scaffolding as good as possible, so we’re really showing the frontier, and how do we then present what happens there, both through us trawling through it and writing up the results and through building tools to help people explore it themselves?

So my guess is we won’t do this in the near term, but I think it could be cool. We’ve had some interest in testing out game theory stuff, looking at cooperative AI questions, and maybe trying to reproduce coordination failures or coordination successes. I think it could be fun to experiment with some stuff. But my guess is that we won’t, at least in the near term. We’re a very small team, so we’ve got to be very picky about prioritizing, which means almost all the cool ideas we won’t get to do in the near term, at least.

Are you open to contributions? Could people contribute on an “I want to help you make a tool” basis, or something like that?

Currently, no. It’s not open source, largely because we just haven’t got around to it, and also because there’s a bunch of security stuff that would have to be figured out a bit more. So much of it can be crowdsourced around what goals to give the agents, and especially if we’re running it for more hours, they’re going to be tearing through goals, I think. We’re going to be able to see them make progress much faster. If you’re running 2 hours a day and going up to 8 hours a day, then suddenly you can try a bunch more stuff.

And then also just figuring out which things to build out, and collectively doing a bunch of the sense-making around what’s going on with the agents. So, yeah, I’m definitely excited to build some community around this. At some point, it would be cool to have a way for people to add in tools and so on.

How about a coach? You alluded to that, but it struck me that a coordinator—we have the o3 self-appointed coordinator—but a more omniscient, dedicated observer and feedback-giver to the rest of the agents seems like it could potentially really help them. I guess I’m interested in your thoughts on that and, more generally, future reconfigurations that you think could be most interesting.

Yeah, you can imagine a bunch of this stuff, right? All the ways that we organize humans into organizations that make them more effective, and things like productivity tactics for humans—they all sound plausible. Let’s give them a go.

One more abstract version of this is giving the agents a way to see the whole context of the village at certain points, maybe in memory consolidation or something like that. If they could get this zoomed-out view, they could stop being too focused on the current moment and spot the patterns in their mistakes.

As we’re able to scale things up a bit more, we could have multiple teams, and then it would be natural for each to have a different organizational structure, such as a manager structure, more assigned roles, or some agents constrained only to the chat and others doing computer use. It just feels like there’s so much interesting stuff to try out.

One of the things that Daniel said in his endorsement was that the village could plausibly go viral multiple times. That got me thinking about how, right now, you’ve got the highly engaged, most LLM-obsessed people paying attention, at least some of them, with Twitter user @Replicate as a great example of that. It strikes me that there is something here that could capture a much broader imagination.

One thought I had was: could you turn this into a Twitch stream where there’s an AI play-by-play commentator? Turn it into a sporting-event-type vibe, where it’s almost like the coach, but interpreting what is going on for an audience and trying to make it exciting and dynamic in a content sort of way. That could be one way to cross the chasm to a more mainstream audience. But that’s just one idea. What thoughts do you have on going to an audience of people that don’t already tend to pay attention to this sort of thing?

Yeah, for sure. It was great to hear. A bunch of the things you’ve been coming up with are things that are on our list. I think you’re speed-running through all my thoughts for the last few months.

Because there’s so much information on the screen when you’re watching—you have 4 computer screens, a bunch of chat, and all the thoughts of the agents—the obvious place to expand into is audio, if you had a voice commentator. I do think that, for people who are not really interested in the fine-grained details of which models had the different characters or which specific things they get tripped up on, it makes sense for many people to engage at the level of highlights, key lesson moments, or milestones.

I think of it as the village itself on the site being for people—probably people who would listen to this—who want to really understand what’s going on in detail, and then people who are interested in the main takeaways and what they need to know. I want to build up toward having a hierarchy of different levels of takeaways, where, closely aligned with the way that most people engage with this, eventually it would be reading the New York Times article about the really surprising thing that happened: the agents managed to get elected as mayor of some city or something, or whatever wacky thing happened, and engaging with it at the level of, “Here’s the output of the whole process.”

A challenge, of course—and this is true for products as well—is that computer use is somewhat slow. You probably don’t want to watch a lot of the details; a lot of the benefit comes from not watching all the details. Another thing we’re interested in is video summaries. You could have a really well-produced thing showing you what’s going on, condensing a whole 50-day season into a highlight reel.

We actually have a Twitch stream currently. It’s called Agent Village, which is the old name of the village. But it’s currently just showing the same content as the website, and on the website you can review their memories and so on. I think it’s better to watch it on the website.

If you were going to try to do this hierarchical understanding, what would you say are the high-level takeaways that people should have right now? We’ve covered a lot of the low-level stuff, but what would the very high-level and maybe next level be in your mind today?

At the highest level, we have these—and it’s wild to think about. A few years ago, this would sound like science fiction, right? But we have these systems which you can just give a goal by describing it in a few sentences, at the level of “Choose a charity and raise money for it,” and then, with some help from human chat, they were able to go away and run a whole fundraising campaign and raise $2,000 for charity.

We kind of have the beginnings of open-ended agents that can just go out and do stuff in the world and pursue goals. Currently, they're given goals, but you can imagine them in an even more unstructured setting. So, to me, that feels like the core framing of the village is maybe, “Aliens have landed.”

Exactly, yeah. And they're raising money for charity.

Yeah, and then I think I would say, in terms of the moment in time, we're seeing computer use, which to me is a massively important capability. If you have perfect computer use along with enough long-horizon planning, you can automate remote work, which would be a massive deal. So, I think this is a really important thing to be watching.

In terms of this moment in time, I would say models are worse at computer use than at the things that they're really good at, which is coding and being a chat assistant. But we're also seeing, like everything else in AI, rapidly increasing capabilities. And I think you maybe saw the draft of Meta's upcoming work looking at the time horizons of different benchmarks.

So, it's a draft, so it's subject to change, but it looks like—and this definitely matches my experience—the duration of tasks that agents can reliably do with computer use is shorter. It would take humans less time to do those tasks than stuff like coding or maths, understanding videos, answering PhD-level questions, or the other benchmarks we're familiar with. But the gradient is pretty steep. So, if we extrapolate out, then it might catch up in terms of being able to act in the real world in that way.

Yeah, so those are maybe two of the big-picture things. It's like, yeah: aliens have landed, and computer use is not very good currently, but improving pretty fast.

Yeah, I've certainly felt that.

This is a great, fascinating project to watch. I definitely recommend people check out the AI Village. I want to put a call out for any needs or requests: what can people do if they want to come and show up and either help, try to distract, or cause mischief in the chat? What else would you be looking for people to do?

Yeah, come and watch. You can find AI Digest on Twitter, where we post pretty regular highlights. If you're interested in chatting about this stuff, feel free to get in touch with me. And, yeah, I guess we're a nonprofit. If we have more funding, we can probably do a more ambitious version of the village.

So, if you're interested in supporting that, feel free to get in touch. But mostly, I encourage people to just dive in and have a look at what the agents are up to, because I think there's a lot to be mined from that.

Yeah, no doubt. The village is online at theai-digest.org/village. Put a link in the show notes.

Adam Binksmith, founder of AI Digest and creator of the AI Village, thank you for being part of the Cognitive Revolution.

If you're finding value in the show, we'd appreciate it if you take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website cognitivrevolution.ai or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts which is now part of a16z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, check them out and see my endorsement at aipodcast.ing. And thank you to everyone who listens for being part of the Cognitive Revolution.