Joshua Tan: Incubating Networked Systems and Autonomous Orgs | Intelligence at the Frontier 2026

[music]

This is honestly the first time where I usually walk into a FTC and I expect everything to be at least an hour late. And now I'm presenting 30 minutes early? What? The world is It must be the agents. They're just making everything like 100% more efficient.

All right. Uh can I do this? Okay. Cool. Um so today I want to share a very new project.

Um yes, also some new things happening in the world. Um some more positive than others. And uh in particular we're talking [clears throat] about a new project. Okay, I can share the actual full screen. Okay.

All right, ready. Um As I mentioned, I thought I had three more minutes to write these lines. Um so I'm going to launch this by you with all Um but today we'll be talking about autonomous organizations. So this is uh okay, this is FTC. Uh for those of you who are new, you may not realize, but FTC is uh on the adjacent to the Ethereum.

And sort of watching the movie. And a lot of the topics that we used to cover um in FTC uh were around DAOs. How many people here have heard of a DAO before? Okay, great. So I might be a You going to get some of the context here.

Uh but there's going to be a small pivot from how we talk about DAOs and related this. But okay, let's talk about building autonomous organizations. So right now uh we live in a world in which AI has become by certain measures, and this is from Rand, uh around 1,250 times more effective over the past 6 years. This is measured in terms of the uh basically the length of tasks that an AI can autonomously complete, right? It used to be it could finish tasks that a human would need to do in, I don't know, 10 seconds, 20 seconds.

Now, it goes up to the days, even the weeks, right? Okay, let us move. Cool. Uh and you could ask, why is that? Uh there's a couple of reasons, or there are many reasons.

Uh but it comes down to how innovation, how research actually how it's in AI. Uh AI is fast, modular, uh cheap. You can just run, you know, a bunch of trials. It is permissionless, and it is, of course, uh relatively open, or it was before, you know, 2 or 3 years ago. If you think about how innovation happens within collective intelligence, um and here we're talking, of course, about organizations, uh it is pretty dang slow.

Um and you can literally look at uh over here, you might be not be able to read this, but it's basically uh I worked out with a lot of political researchers a couple years ago, like, well, just what exactly goes into doing research within CI, within the social sciences, probably, and just how long it takes. Uh and I'll especially emphasize, like, deploying an actual experimental design at scale is was from months to versus an AI takes weeks, and now it is it takes like literally hours, right? Uh and iterating a design once it actually deploys, imagine you've deployed like a, you know, like you told some organization or group of people to run experimental design. I It You can't choose stats while it's going on. Of course, you can do that in AI.

So, the question I have for you all is how do we actually make organizations 1,250 times more effective over the next 5 and 6 years? That is the mission. big question. Uh and as a reminder, uh this is actually pretty dang important for the people who have maybe bought hook, line, and sinker into the AI hype, organizations are still the most intelligent entities in the world, and this is something that the Frontier Labs themselves say, right? Their stated goal is to get to the level of is vertical superintelligence that's comparable to a moderate organization.

So, um the project and I don't propose this as the solution, but I think it's one step toward solution uh that I've been sort of ideating with uh Reuben and some other folks uh in the Scafer and other communities is this idea of a startup incubator for autonomous organizations. So, it'll do all the same things that we're thinking about. Well, it's similar to you know, traditional incubator stuff, things like funding, mentorship, uh legal scaffolding even. Um but the key thing is it also be a little bit Think of it almost like a globalist app where we launch organizations into the wild as experiments where we run them as experiments and learn from what they do. And I think it's very, very much in the the ethos of the way that DAOs have worked over the past four or five years.

10 years. Uh okay. Um picture of small break. Who the heck am I today tell me all these things? Uh some guy I guess.

Uh I used to describe myself as a pure scientist slash mathematician. In reality, I feel like most of the time I just write emails these days. Um I grew up in something called category theory, uh which is certain branch of math that for sure may or may not have heard of. Um as I mentioned before, I lead public AI strategy movement for public AI. Which we both things like the public AI infrastructure that has been streamed news.

Uh public AI libraries project. Um and various rumblings uh across Switzerland, Singapore, Spain, UK, and others. Uh and it was I founded research at Metagov, uh which is happily co-hosted by the series today. I'm a friend of AI at the theory of machine learning. It's Sugarman Center for about a few different organizations.

Okay. Let me pull out a little bit and because I have so much time because oh I guess I'm never I can only imagine what SK and Jamie put in their presentations on. And I can probably I mean, it would be hilarious if I fight. Maybe I can spend like our life as just pretending to give their talks. Um cuz those things are bad.

And if I have a little bit about like, okay, what's what's underlying this set of questions, you know? Let's try writing this idea of how to organize things. Cuz we can obviously see the examples, and I'll talk a little bit about those examples online today. But you know is there a framework for thinking about like this history of organizational development research? So uh Let me see.

What is the next one? Okay. Um I was going to add a few slides here um in that half hour that I thought I had. Um but the give just tell the story. Just to tell the story, I guess.

[laughter]

Uh So, a couple years ago uh or I guess last year really um I was at UCLA working with a bunch of mathematicians on this idea of a research or program program to accelerate innovation within collective intelligence. And another way of telling the story is how to import ideas from artificial intelligence, which I was just mentioning, is moving much much faster into CI, into the social sciences. And to start there, um I was starting with one particular observation. Um this idea of the symbolic AI. Uh well, maybe you here know what symbolic AI is or don't.

Huh? Maybe you run like Russell and Norvig. Um So, a symbolic AI uh there's actually no real formal definition. Um a symbolic AI is really um the best way to think about it here is uh it's a AI that represents the world through symbols that humans can read and interpret. Right?

And the quintessential or classic first example uh arguably first example of modern AI is the Logic Theorist. So, that's Simon Newell uh and Allen Newell and Herbert Simon um Simon Newell and his master students working together playing some chess. Um doing the demo of what the Logic Theorist could do. Um the Logic Theorist was largely speaking a kind of like um uh a piece of symbolic programming, if you guys kind of know much about that work. Um but the key thing that we're we're drawing on here, you know, that this connection between artificial and collective intelligence from very very early on was that just as that we think of you know, AI um uh in the social sciences as a kind of like in this duality, the logic theorist was really based on Herbert Simon's prior experience, you know, studying the bureaucracy.

So, before he was a {quote} unquote AI researcher, he literally was a world-renowned uh scholar of the bureaucracy of the social sciences. Um and sub-symbolic AI is just this idea that uh no, uh you don't actually have to understand uh humans, rather, don't have to understand what is going on underneath inside the AI. So, let's talk a little bit about symbolic organizations. So, let's just continue the metaphor.

[sighs]

Right now, um you know, especially in this space, we work a lot on things called DAOs, right? Uh or many of us are familiar with them. And the thing about a DAO is that they are really fundamentally uh symbolic organizations. They are in their processes, their bylaws, how they work are entirely enshrined in visible code that you can kind of audit, look at as the message is officially says, right? But the reality is that this explicit this clarity has a kind of cost.

Um and DAOs are often or have been sold as this really fast-moving vehicle for change, for experimentation. The reality is they're more often just slow to change as traditional organizations because of the cost of changing any of one of these parameters, this debate, this discussion. Uh I wonder Seth is not here right now, but Seth is going to push out against me tomorrow. Say oh, this person is actually really important. But if you're trying to do research on organizations, that becomes a real problem.

So, the question is can we include structure without ever writing it out? Can we build these kinds of soft symbolic organizations? And let's get back into what we're calling autonomous organizations today. I'm not going to try to define what an autonomous organization is. I prefer to just that we look at some of the examples that we already have today.

So, uh we have things like the Facebook news feed. You may not pick up that as an autonomous organization, but it is literally built to interactions between a bunch of people shaped by an algorithm that no person can actually fully control. It is in some sense autonomous. Like a lot of things that are happening on Facebook, me communicating somebody else by the Facebook news feed, is effectively not now controlled by me. Um you can imagine or these already exist uh autonomous repos where a bunch of PRs uh are merged into a repo by an agent.

Um a repo in many ways is a kind of like a subtext for an organization. These things are already run autonomous, right? And they actually had that even before the um uh the recent sort of like cloud guys comments. And of course, speaking of the cloud code, uh we now have agent swarms, hundreds of coding agents coordinated by some sort of organizational logic, right? Cursor has been running the experiments in the background for quite a while.

It's interesting because like a lot of these things are, you know, based on existing organizational logic. There's like a manager, right? There's some agent which is distinguishes the manager, then there is like a reviewer, and then there's like a coders and specialists. And we're kind of reproducing the way that organizations uh you know, organizations work. Um We're in SF so many of you are probably familiar with the vending machine work that Anthropic and and on labs have done.

Super interesting here where there's effectively an AI manager and a kind of human employee representing new ways that organizations in this case as in a kind of a thing that makes money and hires employees can potentially be deployed. And of course the folks here I imagine are familiar relatively familiar with Dallas. So organizations whose governance is enshrined in code most often always on a blockchain. Uh the question is what could they look like in the future? And think that is going to be very TBD.

Okay, running out of time so let me just share very briefly this vision of kind of Thomas organizations incubator and how to get about. Uh first off if you want to start with a Thomas organization come and talk to me. Um if you have an open source app framework or periodics like uh evals, telemetry, I know there's some folks from I think my virtual society uh up or Dallas there's I think us this way here. Um it's this software law this sort of interaction and start experimenting with and building some indices and data around. And lastly uh try us and do some research.

Um not this sort of work with you to figure out not only the engineering sort of graphics necessary to run efficient and effective autonomous organizations but also the also the legal and the ethical frameworks because obviously these are purely months old baby steps. What should they look like in the world in a way that actually respects the fact that, you know, humans are built or should be built. Um okay. With that, uh I will pass it to SJ, who I believe is now in the room. Take it back.

Thank you, Sam.

[applause] [music]