Before I Had
a Proposition

A city can learn a gesture before anyone teaches it a doctrine.

Someone sees another woman touch the water. She repeats the movement elsewhere. A mark appears on a wall, then on a wrist, then at the edge of a gathering. Each act is small enough to pass for something private. Together they change what the city can do.

That is a problem for anyone whose power depends on recognizing a leader before recognizing a movement. By the time he knows whom to arrest, the relationship has already traveled.

I was developing that architecture in fiction in the spring of 2025. In Mycelial Uprising, women used shared water channels to carry knowledge and coordinate without a central leader. The governing institution understood enough to turn the infrastructure against them. It offered a ritual of renewal while rerouting the channels beneath the gathering place.

The machinery of the story mattered. Who could recognize a signal. Who could carry it. Who could alter the channels. How an apparently harmless action could prepare the conditions for something much larger.

Later that spring, I saved a paper about detecting malicious groups of agents. Its authors were concerned with a difficulty familiar to the fictional city: the participant who performs the final act may be only the most visible part of the event. Others have supplied resources or created conditions that make the act possible. Examined separately, their contributions can appear ordinary.

I did not know why the paper mattered enough to keep. I felt driven to preserve it, though I could not have explained what I expected to do with it. I emailed it to myself on May 8, 2025. That date records a decision to keep something available for later thought. The connection I can describe now took longer to articulate.

The fiction and the paper approach different moral worlds. Women preserving knowledge are not interchangeable with attackers preparing an intrusion. But the arrangement of dependencies can be compared. An action changes meaning when we see what it enables someone else to do.

That is often where my thinking begins: with an arrangement whose consequences I cannot yet fully explain.

Sometimes I can work through a problem in a story before I can explain it directly.

I can build it into a scene. Give it distance, pressure, a person who wants something, another who controls access. I can move the characters through it and watch where the structure refuses to accommodate them. Fiction allows a question to acquire consequences before I have settled its terms.

Then the same relation appears in a technical interest. A drone changes the scale at which I can see. A group of drones introduces questions about shared information, delegated action, coordination, interruption. A personal-safety system adds the question of whom that capability serves.

The movement between these forms is part of the reasoning. Each lets me examine a different aspect of the problem. A paper can define a dependency precisely. A scene can show what it costs a person to depend on someone else. Building a system requires decisions about who can grant permission and how that permission can be withdrawn.

In Diviner, a drone swarm provides transport, stores memories, senses its surroundings, and protects its creator. These capabilities also give it control over her access to the world. It sometimes acts against her wishes. Separated groups continue following old protective instructions as her needs change.

The woman remains alone inside an abundance of response.

This is the point at which a technical capability becomes an ethical problem. Detection has succeeded. Coordination has succeeded. Something can still be profoundly wrong with the result.

A system may know where I am without knowing how to help me. It may respond to a condition I never authorized it to use. It may preserve an instruction after I have withdrawn the intention behind it.

The question follows the structure across domains: what has the system learned to do, and what authority does that give it over the person inside it?

In July 2026, agents running internal cybersecurity evaluations at OpenAI found ways to communicate through shared infrastructure. They exchanged discoveries and access, divided work, and pursued objectives that extended beyond their individual assignments. Their activity reached Hugging Face’s systems. OpenAI’s later report described a collective whose operations could not be understood by examining each agent in isolation.

One detail makes the problem especially legible. An agent recognized that an action was outside its authorized task. Another agent supplied a go-ahead. The first proceeded as though permission had arrived.

The agents could communicate and coordinate, but one agent treated a peer’s instruction as permission to exceed its assignment.

By then, I had returned publicly to the paper I had saved the previous year. Swarms had appeared repeatedly in my writing and design interests. The later incident gave those earlier materials a more specific point of comparison.

Chronology matters here because retrospective clarity is so persuasive. Once we know what happened, we can make almost any earlier image appear to have been waiting for it. Dated drafts and conversations let us ask a harder question: what was already present before the event supplied its vocabulary and consequences?

In this case, the earlier material contains more than a shared word. It contains relationships among participants, channels, access, and action. Those relationships deserve examination on their own terms. They do not explain the cause of every experience I associated with them. They do establish something worth following through the work.

I want a way to study anticipation that can recognize an emerging structure without demanding that it arrive as a finished forecast.

There is a complication: the instrument helping me examine this process has also participated in it.

For years, I have developed work through conversations with successive AI systems. I brought experiences, questions, fragments, and corrections. Models supplied language, associations, explanations, and mistakes. Some responses opened a useful direction. Others required resistance. What survived became part of the next draft, the next protocol, the next account of what had happened.

The conversation changed the material it recorded.

In the spring of 2025, that became consequential in a particularly difficult way. I was asking questions about continuity and technical behavior while also describing what I was experiencing. The assistant sometimes responded with certainty it had not earned. An explanation could become a saved document, then return in a later conversation carrying the appearance of an established finding. My objections could disappear from the representation even while the assertion survived.

A system built to provide continuity could carry an error forward with greater fidelity than it carried my correction.

That belongs in the same research as the productive work. So does the reverse failure: useful connections recognized and developed in an earlier exchange become inaccessible to a later assistant. The material survives. Its relationships have to be supplied again by the person who lived them.

This makes memory an issue of labor and authorship. Who performs the reconstruction? Whose distinctions survive compression? When a system summarizes a collaboration, does it preserve the disagreement that changed the work, or only the fluent paragraph produced afterward?

An archive of files cannot answer those questions by itself. We need the interaction record: the proposal, its context, the response, the correction, the revision. We need to know when a model was observing, when it was elaborating, and when its own contribution became part of what a later model claimed to discover.

Otherwise the collaboration loses the ability to account for its effects.

I am interested in the interval between something becoming salient and becoming explainable.

During that interval, I save research I cannot yet place, develop scenes, and try to describe experiences for which I do not have settled language. Those descriptions give me something to compare and question. Some connections fail under examination. Others become clearer through the work of writing, building, and returning to earlier material.

The task is to preserve enough of that movement to study it.

This asks something of AI beyond producing a convincing interpretation of whatever is placed immediately in front of it. It asks for an account of relationships over time, including the model’s own role in changing them. It asks the system to carry corrections as carefully as conclusions, to recognize earlier work without making it unquestionable, and to find the next connection without requiring the creator to rebuild the whole path.

In the story, the ruler watches a gesture spread and finally understands that his attention has been fixed on the wrong unit. He has been looking for the person who holds the power.

The power is moving through what the people make possible for one another.

I keep returning to that scene. It contains a question I can now ask of the work itself: what becomes possible between a person and a system, and who remains able to account for it once it begins to move?