Gesture Was the Limit. Maybe Not Anymore.

Notes toward a coordination layer for eight billion people

On May 25, 1986, six and a half million people held hands in a line from New York to California. They donated ten dollars each to stand in their assigned place for fifteen minutes in a symbolic gesture to end hunger in America. The line ran through fifteen states and across roughly four thousand miles, broken in a few stretches by sheer geography but otherwise intact hand-to-hand. It was the largest single coordinated human action of its time, and looking at photographs of it now is still, somehow, moving.

What it took to make that happen is what I keep coming back to. Eighteen months of planning. Mainframes: IBM donated time on theirs. Phone banks staffed by volunteers calling people who had mailed in pledges, telling them which mile of which highway to show up at. Paper maps mailed out. Local organizers in every state matching addresses to coordinates. A media operation. A logistics operation that, in its scale and complexity, resembled a small war.

And the thing they produced — the output, after all of that — was one bit per participant. Present, or not present. A signal of solidarity around the problems of hunger and homelessness, beautiful in the way photographs of it are still beautiful, but as an act of coordinated intelligence it was so simple, it is almost nothing. Six and a half million people had been mobilized to perform a single, identical, fifteen-minute gesture. They could not deliberate together. They could not adjust the action or pivot based on what their neighbor wanted to do. They could not respond to conditions on the ground. They could only stand, briefly, in the place they had been told to stand.

That gesture was the limit of what the technological substrate of 1986 (early computing, mostly analogue communication) allowed.

I have been thinking, lately, about what the substrate of the next decade might allow, and whether what is being built right now in AI is being built in a way that opens that question or forecloses it.

I do not have the answer. I have an idea — one possible shape of an answer — and I am trying to work out whether it is the right shape, whether it can actually be built, and whether anyone is in a position to build it.

The idea, more or less, is this: what if AI’s highest purpose is not for the things we are currently using it for?

The chat interface and the agent are the two forms the technology has so far taken on, and most of the activist applications I see follow from one or the other. Faster press releases. Sharper policy summaries. A mentor in your pocket. Agents that file FOIA requests in the background. None of this is bad. Some of it I am building myself, at Outcry, because it is useful and interesting and because it is what is technically achievable right now. But none of it does the thing I keep thinking AI might actually be for, and the gap between what is achievable and what feels actually possible is what this dispatch is trying to describe.

The systems we are trying to influence — markets, states, frontier labs themselves — coordinate globally in milliseconds. Capital does. Compute does. Tokens move through a hundred layers of a frontier model in tens of milliseconds. Eight billion humans coordinate through nation-states, quadrennial elections, NGO reports, and mass protests that take months to organize but dissolve over the course of a weekend. The asymmetry is not a problem we have failed to solve; it is the problem. Most of what is being built right now in AI-for-activists works on the wrong side of that asymmetry. It makes the slow side slightly faster, the same way Hands Across America made paper-and-mainframe coordination slightly larger. Worth doing, I think, but not the move I keep thinking is actually soon to be available.

That move, when I try to articulate it, is to use everything that is being built for AI (the models, harnesses, tools, infrastructure, energy, new ways of thinking, etc) to produce something the substrate has never been used to produce: coordinated human action at the scale of the species. Not a gesture. Not a march. Not a campaign in the sense we usually mean. A standing capability for synchronized intelligence among hundreds of millions of people, exercised when conditions require it, then quiet again.

What that capability would actually do, in practice, is something I have been trying to imagine concretely. I want to avoid vague gestures at “coordination.” So image this: Ocean plastic cleanup at 100 million-person scale, every coastline, the same weekend. Wildlife corridors planted across continents through a single coordinated push. Citizen science at species scale — bird migrations and light pollution mapped in real time by every phone on earth. Disaster relief that arrives ahead of the disaster, because the people best positioned to provide it have been alerted before the rest of us notice it is coming. A hundred million participants withholding spending from a targeted sector for a single week. These are not all the same problem, but they are all the same kind of problem: the answer requires many people doing complex social behaviors at once, and the obstacle is not desire but coordination.

When I try to sketch what infrastructure would actually be required to do this, I keep arriving at three layers, and I want to lay them out with the caveat that I do not think the shape is settled. It is more like the shape I keep arriving at when I try to think the problem all the way through.

There would have to be a deliberative layer, and I think it would have to run locally, on each participant’s device. A small AI model, no cloud. The reason is not privacy theater; it is durability. A coordination infrastructure that can be turned off by the entities being coordinated against is not a coordination infrastructure. It is a permission slip. The local model would need to encode, for each participant, what they actually see from where they stand, what they want, what they would do, and under what conditions they would do it alongside strangers they haven’t met yet. Some of the pieces have analogs already — on-device small models, sovereign weights, and the early version of this at Outcry. But the hard part, I think, is not the model. The hard part is producing authentic convergence at planetary scale without the layer above this one becoming a manipulation engine, or collapsing into the kind of ineffectual spectacle that movements oriented around one-off marches tend to end up being. I do not know how to do this yet. I think it is the open research problem.

Above that, there would have to be a synthesis layer, running at frontier scale. Models that read aggregated outputs from the deliberative layer and return to participants an actionable picture of what they collectively see. Sortition scaled past where sortition has ever been scaled. The primitives exist in early form. DeepMind’s Habermas Machine generates consensus statements that small deliberating groups prefer to ones produced by human facilitators. Anthropic’s Clio surfaces aggregated patterns across millions of conversations without any human reading the raw text. Neither of these is planetary. Neither is the layer. But they suggest the layer is not impossible, only un-built. The path from groups-of-five consensus and population-scale pattern surfacing to authentic shared situations among eight billion is partly engineering and partly open research. The compute, I am reasonably sure, is engineering. The rest is honestly still unknown.

And then there is the simulation layer, which is the one I want to spend more time on, because it is the part of the picture furthest from what people imagine when they imagine AI for activists, and the part where my own thinking is the least settled.

Coordinating eight billion people, when I really turn it over, does not seem like a logistics problem made larger. I have been with this for a long time, and I think it is a simulation problem.

When Richard Feynman first proposed quantum computing, in 1981, he did so because he was arguing that there exists a class of problems whose state space grows so fast that no classical machine can ever represent them. Simulate the dynamics of thirty interacting quantum particles on a classical computer and you need more memory than there are atoms in the visible universe. Simulate them on a machine that operates by the same logic as the particles themselves, and the resources scale almost linearly. Reality, Feynman noticed, is not classical. To simulate it well, you need machines that compute the way it does.

I have come, slowly, to think that a coordination event at the scale of the species sits in that class of problems. I am not sure of this. I want to flag that clearly. But here is the intuition.

Eight billion humans, each with their own thresholds and conditions and willingnesses; each entangled through markets and media and memes; each capable of altering the trajectory of all the others through a single act of participation or refusal. The state space of “what could happen if these eight billion people did X” is not classically tractable. You cannot enumerate it. You cannot grid-search it. You cannot run a Monte Carlo over it that converges in any meaningful sense. And the interventions whose downstream effects you would most want to know — a coordinated withdrawal of labor in a specific sector for seven days, a hundred million participants altering their consumption to impact a commodity market in real time, a synchronized community forum at every playground in the world on a single Tuesday — these unfold across a possibility space whose dimensions exceed what any classical simulator we presently know how to build can hold.

The architecture you would need to model them, I suspect, is the architecture the universe itself uses to compute outcomes from possibility. Superposition. Interference. Entanglement. A coordination event, in this picture, is the macroscopic analog of what a quantum computer does at the level of qubits: many possibilities held in a single computational state, allowed to interact, collapsed into a single trajectory by the act of observation. The deliberative layer, on this view, would be preparing the wave function. The synthesis layer would be rotating it. The simulation layer would be asking the question that collapses it.

I want to be careful here, because this could be a metaphor that overreaches, and I am genuinely uncertain whether it does. The honest version of the claim is something like this. Movements already behave in quantum-like ways — superpositions of possible mobilizations, collapsed by triggering events, with interference patterns visible in retrospect when one cause picks up an unrelated cause’s energy. If those descriptions are not just poetic, then the right machinery to simulate the systems we are trying to act on may not be classical. It may need to be quantum, or quantum-hybrid, or some scaled-classical approximation that captures enough of the same structure to be useful. I do not know which. I think this is one of the more important questions in the whole picture, and I think almost nobody is working on it from the activist side.

What I do think I know, with more confidence than I have about most of the rest of this, is that the foundations being laid right now will determine, ten years out, whether a coordination layer of this kind can be built on top of them, or whether the substrate is shaped in ways that make it structurally impossible. The data centers, the energy grids, the chip architectures, the regulatory frames — all of it is downstream of decisions being made now. So is whether anyone with the resources to build the layer is even thinking about it.

When I think about who might fund any of this, the answer keeps narrowing. The OpenAI Foundation, as currently constituted, will not. Anthropic’s founders, having pledged their wealth, will not. No frontier-lab-adjacent philanthropic vehicle will. The reason is structural, not personal: a planetary coordination infrastructure is the one deployment of capital that the institutions producing the capital experience as a threat. The threat is not the ocean cleanup. The threat is that the same architecture, once built, coordinates whatever its participants converge on. I do not say this to indict anyone. I say it because the diagnostic is useful to me. A philanthropy the AI industry can comfortably underwrite is a philanthropy that has been domesticated before its first grant goes out. The fact that what I am describing here is uncomfortable from inside the labs is, I think, one of the few signs that it might be pointed in the right direction.

So this is where my thinking is, more or less. I keep returning to that line of six and a half million people, and to how much human effort it took to produce one bit per person. The substrate of 1986 is a museum piece now. The substrate of 2008 — Earth Hour, a billion light switches, a single coordinated minute of darkness — is approaching one. The substrate available to us in the next decade would have been mathematically inconceivable to the machines that produced those photographs of hands across the desert.

What is missing, I think, is not the machine. It is the decision to start asking what the post-AI machine might actually be for, it’s highest purpose, and to start trying to design it before the foundations are finished being poured.

That is the work I have been trying to do. I do not know yet whether the shape I have sketched here is the right one. I am writing this, in part, to find out.

Links: micahmwhite.com, outcryai.com