# The Room Is the Unit

The room matters because every person in it is inside the system.

Laplace's demon was imagined as an intelligence outside the system: full state, full law, perfect forecast. Computer Future's correction is that every real actor belongs to the process it tries to predict. The observer is inside the computation. Modeling the world uses part of the world. Modeling yourself modeling the world adds the recursion that breaks the fantasy. Self-abstraction has a bound.

AI puts a small version of that fantasy on every desk.

A person prompts a model and feels the sudden lift. The blank page yields. The mockup appears. The email gets written. The internal distance between wanting and making collapses for a moment, and the person feels more agentic because, locally, he is. That feeling is real. It is also bounded. The worker has more reach inside his private model of the work, while the organization may have learned nothing about itself.

That is the deeper version of David Demaree's useful observation: AI is good at lighting up a room for one person and still poor at bringing the room together. The problem is larger than collaboration etiquette. It is formal opacity expressed as coordination. Each person is a bounded model of the work. The organization becomes intelligent only where those partial models are coupled through artifacts that survive mood, rank, memory, and private fluency.

The room is the first practical computer around the individual blind spot.

This is why slop feels so irritating. Slop is private computation thrown across a boundary before the shared model exists to make it usable. The sender experienced agency. The receiver receives an unpriced self-abstraction task: infer the intent, inspect the artifact, locate the miss, repair the context, absorb the awkwardness. The object looks like output. Operationally it is a demand that someone else finish modeling what the producer left implicit.

AI adoption is the moment a company discovers how much of its self-model was never written down.

Most organizations run on partial stories: what the meeting meant, who really owns the decision, which standard matters, which shortcut is tolerated, which customer promise is sacred, which dashboard lies, who catches the miss, where correction lands. Human beings can operate inside that fog because they share history, status, tone, fear, trust, and repeated contact. A model has none of that unless the room has made it explicit enough to compute with.

That is what I mean by saying humans are still becoming sufficiently computational. Computational here means explicit enough to share state, route action, preserve memory, and correct the next run. Most of our shared life has weak executable state. We carry too much of the system in private memory, social inference, vibes, and after-the-fact explanation. Then we hand one person a machine that can produce at speed and act surprised when the group fails to metabolize the result.

The alignment frame arrives late. Alignment is one downstream symptom. The upstream constraint is bounded self-abstraction. A person, team, company, or civilization acts safely only across the portion of itself it has modeled well enough to route agency through. The rest becomes theater, coercion, surprise, and repair work.

Forced AI adoption is a clean example. The mandate says the group is becoming advanced. The worker experiences a shrinking control surface: token quotas, usage dashboards, vague pressure to produce more artifacts, fewer guarantees that the artifacts mean anything. Resistance can be laziness, status defense, or fear. It can also be accurate information from a bounded observer who has just lost the handles by which his work stayed coherent.

Kindness here means treating resistance as data before treating it as pathology. Some resistance protects a real standard. Some protects a bottleneck. Some managers use AI language to hide contempt for the people asked to adopt it. Some craftspeople use standard-language to hide fear of being measured. The discriminating question is computational: whose model became more accurate, whose action rights became clearer, and where did correction get cheaper?

The CMS example matters because a CMS is more than convenience. A good CMS is organizational self-abstraction. It encodes who may change copy, who reviews, what can be rolled back, which parts belong to editors, which parts belong to engineers, how public state changes, and where memory lands. It lets the person with legitimate responsibility act directly on the part of the system he owns.

Replacing that with a coding agent may empower developers while making the organization dumber. If every content change now passes through the person fluent enough to ask Claude, agency has concentrated behind private fluency. The AI made one actor more powerful and made the room less computational.

The better move uses the agent to deepen the shared model: safer CMS controls, review trails, generated drafts inside the editorial workflow, consequence previews for nontechnical owners, rollback paths, and memory that changes the next run. The question is whether the tool makes the room better at seeing and changing itself.

The Gödelian horizon stays in the picture. A room never fully models itself. The group is also inside the computation it is trying to describe. Every policy changes behavior. Every metric changes the game it measures. Every AI tool alters the work it was meant to accelerate. The point is to locate the blind spot and build external loops around it: review, provenance, rollback, permission, customer feedback, and shared memory.

This is why the previous version of this node was incomplete. It had the social claim: shared agency beats private sparks. The correction arrived from outside that draft's own model and named the deeper structure. The room around the node made the node smarter. That is the thesis in miniature.

Computer Future begins here: a person and a machine, then a room and machines, learning to model their own bounds. The civilizational problem is humans summoning more computation into rooms that have not learned to compute themselves.

The room is the unit because the room is where private agency becomes shared state, shared correction, and shared responsibility. The room is where the local demon gets demoted. The room is where a bounded mind gets help seeing the bound.

Make the room computational.

## Sources

- David Demaree, "AI and agency," Bits&Letters, June 8, 2026: https://www.bitsandletters.com/ideas/ai-and-agency
- Computer Future, "demoting Laplace," January 11, 2026: https://computerfuture.me/posts/demoting-laplaces-demon