RESEARCH / AREAS
Four research areas, each shown as work in progress.
Equilibria studies how groups of agents, human or artificial, come to act as one, how fast they move, and how the rules around them hold up. Each area below opens with its question, shows the picture we are working from, and lists the pieces so far with an honest status tag. Working papers change. The tags say where each piece stands.
A market, an ant colony, a research field, a company. Each is made of parts that act on their own. Yet we talk about each as if it were one thing with its own goals. Sometimes that is a figure of speech. Sometimes it is literally true, and the whole has a state of its own that the parts do not.
This area asks when the second case holds. It is the foundation for everything else on this page. You cannot ask how a collective moves, or how to govern one, until you can say what counts as one.
Draw a handful of agents and the links between them. Every field has a name for what happens next. Biology says organism. Economics says firm. Physics says phase. Each of them draws a boundary around some of the dots and calls the inside one thing. The question is what licenses that move.
Once you draw the boundary, you start assigning it goals, beliefs, and blame. That is how we talk about companies and states, and increasingly about groups of AI systems. If the boundary is in the wrong place, the whole story built on top of it is wrong. Our work treats the boundary as something to test: which parts move together, and does the whole behave as if it has a state the parts do not?
The picture we work from has three layers. Agents at the bottom. Groups whose members are coupled tightly enough that they change together. And, sometimes, a whole made of those groups that has properties none of them have alone. Our papers give this an information-flow reading: a group is one agent for a given task once information moves through it faster than the task needs it to react.
The criterion is stated. What is not settled is where the line falls in real systems, and how to detect it from the outside without already assuming the answer. That is what the current drafts work on: a formation criterion tested on simulated networks, and boundary-finding methods borrowed from graph theory.
Titles link to public copies where one exists. Descriptions say what each piece asks and how it goes about it, not what it found. Numbers live in the papers.
What does a story about the future quietly assume about the agents in it? An essay, the first post of the process-alignment series. It argues that every model of society inherits the model of the agent underneath it.
How did agency evolve, and what does that history say about which agents appear next? An essay that traces agency from cells to institutions to AI systems as one family tree.
Why do biology, economics, and AI disagree about what counts as an agent? A conceptual synthesis. It lines the fields up side by side and sorts the agent concepts each one uses.
When does a group of parts become one agent with goals of its own? A map of open problems, written to invite collaboration rather than to settle anything.
When does a group count as one agent, and when does that boundary break apart again? A survey in progress that joins evolution and information theory around one idea: a group is one agent for a task once information crosses it faster than the task needs.
Is there a checkable criterion for when interacting agents form one collective agent? A formal criterion built on spectral graph theory (reading a network through its vibration modes), tested on simulated networks.
Can the ways game theory, physics, and social science describe group behaviour be translated into each other? A proposal, in the language of category theory (the mathematics of structure-preserving maps), for a set of correspondences between those descriptions.
Are markets, networks, and democracies near-optimal structures rather than historical accidents? A formal framework linking those structures to network theory and information theory.
Can a classic graph-cutting algorithm find the natural statistical boundary of a system? A short formal note: definitions plus one proved proposition linking Markov blankets (statistical boundaries) to minimum cuts in a graph.
Can we detect where the agents are inside a system instead of assuming their boundaries? A method that combines causal-emergence analysis (finding the scale with the most causal power) with structure learning on graphs.
Can one principle explain why agency evolves in living and artificial systems alike? A formal argument from stated assumptions about evolution and prediction.
Can traits like memory or planning be measured on a scale rather than as yes or no? An experimental section that recasts agent traits as continuous quantities.
What happens when AI agents discover and share tools the way cultures pass down ideas? A framework proposal that models the agents and their shared toolkit as two systems evolving together.
Suppose the boundary is drawn and a group really does act as one. The next questions are about motion. When something changes at one node, how fast does the rest catch up? Does the group settle smoothly, or overshoot and swing back? And when it changes its mind, is that a real shift or just a new average?
Our answer to most of these runs through spectral graph theory. Read a network through its vibration modes, and the gaps between those modes give speed limits. This area collects the papers that build and test that reading.
Take a group with a bottleneck: two tight clusters joined by a single link. Push one node and watch. The cluster it sits in reacts fast. The far cluster reacts late. Whether the whole ever agrees, and how long that takes, depends on the shape of the network more than on the nodes.
Every network has a ladder of modes, from slow to fast. The gap between the bottom two rungs is the slowest thing the group can do as a whole. That gap is small when there is a bottleneck. It sets how quickly the collective can respond, and it can be read straight from the network. That makes it a design variable, not just a diagnosis.
The same group can respond to the same push in two ways. It can slide to a new state, or it can overshoot and swing back before it settles. That difference is the second thing we look at. In our models the swing shows up when the group has levels, and the levels correspond to gaps in the spectrum. That is the bridge back to area one: levels are gaps.
The formal part is written. The tests are the open work. We are building a small battery of five-agent experiments that separates three cases: settling, ringing, and a genuine reframing where the group ends up in a state no single member started in. Until those run, the results in the drafts are placeholders, and the drafts say so.
Titles link to public copies where one exists. Descriptions say what each piece asks and how it goes about it, not what it found. Numbers live in the papers.
Why do communities sometimes defend an old idea and sometimes shift together to a new one? A multi-agent model of belief sharing on a trust graph. It compares agents that hold one world-model with agents that hold two competing ones, on a toy scientific dispute.
Can the slow loss of human influence be seen in the spectrum of the network that carries it? An essay that applies the spectral reading to the gradual-disempowerment scenario.
Can network mathematics explain how individual belief updating adds up to group agreement or disagreement? A formal framework that joins active inference (prediction-driven belief updating) to network dynamics, with worked examples.
How fast can a group get in sync and act as one, and how does that depend on its size and the shape of its network? A synthesis paper. It joins network spread speed (the graph Laplacian), an oscillator model of synchronization (Kuramoto), and a control-theory layer, with worked examples checked in code.
When a collective changes its mind, is that a real shift or a new average? A five-agent test battery designed to tell settling, ringing, and genuine reframing apart. The build is under way; the draft carries placeholder figures until it runs.
When can a cluster of connected agents be treated as one player in the game one level up? A working paper proposing a three-part test for collapsing levels, worked through on a toy world of people, companies, and nations. Nothing in it has been run yet.
Why do some ideas spread easily while others split a community? A framework that treats ideas as signals with frequency properties on a network, with example messages.
A group that acts as one still lives in a world that keeps hitting it. Prices move, technologies arrive, someone defects. Institutions are the parts of a society whose job is to absorb those shocks: courts, markets, voting rules, norms. Ashby’s old law says a regulator needs at least as much variety as the disturbances it faces. Variety here means the number of distinct responses a system can produce. Stafford Beer drew that law as a picture: a system on one side, its regulator on the other, and channels between them that can be narrowed or widened.
AI systems change what institutions face. They add actors that adapt faster than the rules around them update. This area asks how institutions can keep their variety, how they get captured, and how a rule system can change its own rules without losing itself.
Picture a system of agents and, beside it, the institution that regulates it: a court, a market rule, a norm. Signals go from the system to the institution. Regulation comes back. For a long time the loop holds and the system stays in balance. Then new actors arrive from outside, faster and stranger than the ones the rules were written for. The question is who absorbs what they bring.
Once inside, the new actors displace the old arrangement. Agents get pushed out of place and the boundary stops holding. In Ashby’s terms the system now produces more distinct situations than the institution has distinct responses. Beer’s word for this is variety, and the law is blunt: a regulator with less variety than the system it regulates loses control. That is what happens when fast-adapting AI actors join a slow rule set.
Beer’s answer was to work on the channels. Attenuate what comes in, so the institution sees fewer, better-summarised signals. Amplify what goes out, so one rule reaches many actors. When that is not enough, change the rules themselves: something above the institution rewires it. Constitutions do this, and so do the meta-rules of a commons. The system resettles with the new actors inside the boundary. The papers here ask what makes that loop adaptive rather than frozen or captured.
The open questions live at the top of the picture. If one actor gets a strong path into the rule-maker, the loop becomes capture. If the rewiring stops, it becomes lock-in. And we do not yet have a clean measure of how much variety an institution needs before it can resettle a system at all. The current drafts propose candidate variety measures and a dynamic version of Ashby’s law, and they say clearly that nothing there is established yet.
Titles link to public copies where one exists. Descriptions say what each piece asks and how it goes about it, not what it found. Numbers live in the papers.
Why can an economy stay regulated when no single regulator could track everything in it? A theory paper that joins Ashby’s requisite variety to Ostrom’s polycentric governance. The link is measured with effective information (an information-theory measure of causal power) on a computed synthetic network.
When AI agents change faster than the institutions around them, how should those institutions be redesigned? A position essay from a working group. It treats an institution as a self-regulating living system, drawing on Levin’s biology and Ashby’s law, with a technical companion note on candidate variety measures.
Can an AI system stay aligned without locking in one fixed set of moral answers? A proposal built on an analogy to biological self-regulation: align the process by which values change, not a fixed target.
Can governance work faster and better if a society is treated as a self-correcting control system? A conceptual proposal that applies control theory and the free energy principle (a theory of prediction-driven self-maintenance) to governance design.
Does Beer’s classic model of a viable organisation already describe the computations a self-sustaining system must do? A reformulation of Beer’s model in graph and active-inference terms, with a diagram. A blog version is being revised.
How do shared values get installed across a community whose members start with different preferences? A theory paper that treats values as prior preferences and prestige as turning up the volume on which behaviours get copied, with a proposed multi-agent simulation to test three outcomes.
Can the properties that keep biological systems resilient explain why social systems fail, and point to better designs? A perspectives paper. It defines three network measures (organisation, redundancy, openness) and applies them to one case each from biology, neuroscience, and cultural evolution, then to social systems.
How do AI systems that filter information push whole societies toward the same wrong beliefs? A synthesis of failure patterns in information systems, opening from a historical intelligence failure, written to invite critique.
Do digital-democracy voting mechanisms hold up under adversarial pressure? A simulation of democratic mechanisms with language-model agents as the adversaries.
Do the same mechanisms hold up against an adversary trained to break them optimally? The planned follow-up. Reinforcement-learning adversaries replace the language-model ones, on three mechanisms in a resource-allocation game. Experiments not yet run.
Can effective resistance, an idea from electrical circuits, show where information gets stuck in open-science networks? Working notes that apply graph-Laplacian resistance to information flow between researchers.
The first three areas make claims about boundaries, speed, and regulation. Claims like that need somewhere to be tested. Real institutions do not allow experiments. So we build a wind tunnel: agent-based models where a market, a polity, and a culture run on the same population and can be coupled, swapped, and measured.
The Collective Intelligence Library is that tunnel. Every institution in it is a pure function that declares what it reads and writes, so the engine can derive the execution order and the models compose. This area holds the engine, the scenarios built on it, and the work on when a simulation result may be believed.
The usual picture of the AI transition comes as separate stories: an economic one, a political one, a cultural one. Each has its own model, its own assumptions, and no way to feed the others. So the couplings that matter most, money buying attention, attention moving votes, votes rewriting rules, are told in prose and never run.
Put one population underneath. The market, the polity, and the culture become transforms that read fields from that population and write fields back. Now a shock in one model reaches the others through the state they share, and the coupling is a thing you can turn up and down. This is what the showcase on this site runs.
Each transform declares what it reads and writes. From those declarations the engine derives which transforms can run in parallel and which must wait, draws the system graph, and checks that a new institution composes with the old ones before anything runs. That is the engineering claim behind the library: institutions as functions that compose, on one typed state.
Two open fronts. First, new kinds of actors: AI agents that adapt inside the simulation, and models that switch resolution when a tipping point is near. Second, evidence standards. A designed institution can be tested against game-theoretic agents, trained agents, language-model agents, and people, and each test is incomplete on its own. Working out how those tests add up is as much a part of this area as the engine.
Titles link to public copies where one exists. Descriptions say what each piece asks and how it goes about it, not what it found. Numbers live in the papers.
Can markets, networks, and democracies be modelled in one composable framework instead of separate tools? A JAX-based simulation library where institutions are typed graph transforms with declared read and write effects. Open source.
What is the engine, how do contributions enter it, and when may a result from it be believed? The framework paper: the engine, the contribution protocol, the measurement layer, and the ladder of rigor a result climbs before it counts.
How does influence drain from human hands in a capital economy with AI labour? The capital-economy model behind the showcase: its assumptions, the survival threshold it uses, and how it was validated.
How does an attention network pool and amplify influence? The listening-network model behind the attention section of the showcase, and the defences aimed at attention itself.
How does delegated voting concentrate power? The delegative-polity model behind the votes section: conserved ballots, a power-weighted median, and a takeover threshold.
Can a functional-programming approach to multi-agent simulation be both rigorous and fast? An earlier framework paper, with a case study of democratic mechanisms under adversarial stress.
Can a simulation zoom into fine detail only near a tipping point, and stay coarse otherwise? A framework proposal for agent-based models that switch resolution level as they run.
Can a cheap aggregate model run by default, with a live test that says when to switch to the expensive detailed one? A heterogeneous-agent economic simulation in JAX as ground truth, an online version of an established accuracy test as the trigger, compared across shock scenarios.
How should we judge whether a designed institution will hold up, when no single test method is complete? A framework that tests an institution against structurally different participants (game theory, trained agents, language-model agents, humans) and scores stability with measures borrowed from ecology, shown on a shared-resource toy game.
Should AI safety evaluate whole multi-agent systems, not just single models? An essay arguing for system-level evaluation.
Most of what is here is unfinished on purpose. We publish the question and the setup early so people can push on the frame before the results harden. If one of these areas is close to your own work, write to us. The library explainer, the showcase, and the newsletter are the three doors in.