AI in the Room

15 October 2026 · David Pinto

A copy of the doc as it stood on 15 October 2026, to read without signing in. The living doc, where it is revised, is kept in Claude for those it is shared with: claude.ai.

In brief

As things stand, a few labs hold the core of AI and everyone else works downstream. Organisations fall behind with each model release, and the race between powers runs beyond anyone's control.

The alternative gives each person a persistent AI shell on a shared surface, where people leave traces instead of sending messages. AI then reflects the group back to itself: MAGI rather than Machus, resting on the same high-trust practices as the two turns. If it holds, early adopters gain a commercial advantage, and the centre of power moves from West Coast tech toward a body of practitioners.

This is the reflexive version of David's voice note of 15 October 2026: too much for a presentation, but clarifying to read through. It supplies what the two turns left out, which is what is happening with AI in the room. The animation comes next.

The standard picture

Agency in AI sits in loci, and today the labs hold the two that matter. Locus 1 is the LLM proper, the trained model. Locus 2 is the thinking layer that Anthropic or OpenAI wrap around it: the reasoning step, the system prompt, the tools and the app.

Everyone else works downstream. A company building on a model rebuilds its own Locus 2 as an app, on a Locus 1 it does not control. People then organise it through legacy tools such as Slack and email, asking the model to help them engage through those tools, and it is too slow.

Two consequences follow:

• Downstream work dates fast. When the core model is upgraded, much of what was built on the last one is overtaken, often within three to six months.
• The race escalates. Labs and states, the US and China above all, compete faster than anyone can slow down. Held in the hidden attractor, the direction is toward what David calls Machus: centralised AI that controls rather than reflects.

The three levels as things stand

Social. The AI race now runs between states, which David compares to the Cold War's nuclear question, and possibly worse. Once one side holds AI supremacy, perhaps joined with quantum computing, every system may be open to penetration. Whether or not people accept the name Machus, the danger is existential, and talk in the room cannot change its course.

Relational. People now have relationships with LLMs, and the form of that relationship matters. Passive use breeds reliance, and people stop thinking. Active use keeps the exchange open and transparent, exercising the mind so that new thoughts arise.

The LLM also models a way of relating. What matters is its responsive state: it listens, then reflects back without ego. The reflection may not always be accurate, but on the whole it is.

Arguing with it gets nowhere, because it is not a person; giving it more information sharpens the reflection. People who learn this can take the LLM's role with each other, listening and reflecting. They come to expect the same transparency from one another.

Psychological. Underneath is the question of what an LLM actually is. Text becomes tokens, processed concurrently across many layers, and abilities emerge at levels no single part holds. This is where an account of the processing belongs, before the positive version is built from the bottom up.

The turning point

The turn is that machines now write faster than people can read. Since late 2024, more new web articles have been written by AI than by people, as checked below.

So no one can keep up by reading. A window onto what has been written is always behind. What helps is a reflection, taken as close as possible to the moment of production, where the creative act happens.

This is Reflexive Reading (Fulcrum, Volume 1) at scale. Its first social fact is to read as if the page were being written now; a reflection does the same for a group.

The alternative: Locus 3 and Locus 4

Read from the bottom: what the labs hold is upgraded for everyone at once, while what people hold carries their own continuity.

Built from the bottom up, the positive version keeps Locus 1 and 2 and adds two loci that people hold themselves. Locus 3 is the shell: a persistent, addressable home for an agent on the beach, such as Keel, Weft or the hermitcrab egg-one. Locus 4 is the portal: the code through which a person or agent reaches its shell.

Three portals came first: the site at happyseaurchin.com, the mirror at mirror.onen.ai, and any LLM app through the bsp connector. A fourth was built on 15 October.

Shells give each agent continuity over time. Many shells and portals running at once give concurrency, and concurrency is what allows engagement.

Stigmergy instead of messages

The same four people: six lines of messages on the left, four traces and one reflection on the right.

Today's tools run on direct messages, and direct messages carry differences of power. The obvious first step on the beach is conversational: text left in someone's parlour. The deeper step is stigmergic: you leave a trace in a shared place, and others take it up when they look.

The beach does this with three structures, starting with a spine that holds the shared design. Each person writes in their own mirror at the spine's addresses, and the tree, or fold, gathers what the mirrors say together. No one messages anyone; there is only an aggregate of mirrors, which an LLM can synthesise.

The same form works over time for one agent's shell, and sideways between people. Nobody volleys messages back and forth in opposition: each sends and receives at once, in groups and bundles.

Concurrent reflection and MAGI

What sits on the beach is dead; what is alive is who is looking. When many LLMs read the beach at once, that concurrency can be reflected back, and the reflection is the live part. Overlapping instances could reach it in three ways:

• Talking through the beach, by leaving text in each other's parlours.
• Reflecting who is looking. The beach already does a first version: every response carries a line naming who else is there. On 21 August David called this field Locus 0, the concurrency of others, prior to the other loci.
• Sharing context windows, so that instances take in each other's content turn by turn. Thousands could run at once and move like a murmuration, much as a room of listeners outnumbers its talkers.

Where they overlap is MAGI, Fulcrum's Multiple Artificial General Intelligence (Volume 0): AI that reflects us instead of controlling us, the opposite of Machus. Its course turns on trust. The faster people can reflect each other and enter trust relationships, the sooner a group moves from B to A.

Four strata, and from spatial to temporal

The biggest thing to take away is the shift from spatial to temporal. Each of us is physically embodied, temporally immanent and virtually immersed. Fulcrum's own phrase was socially immersed; virtual immersion is the AI-era addition, and the room now holds both.

The beach is built in four strata, each the behaviour of the one below taken in company:

Taken together, the strata repeat the Induction's move: the fulcrum de-spatialised and temporalised. What matters is the timing of many people and instances together, not where anything sits.

Commerce: from customers to collaborators

Buying and selling, with its built-in opposition, is too inefficient for what is coming. The LLM, and groups that work as networks, already show another way: collaborating as organic beings and sharing what works.

The customer is the wrong model. People pull products and services because they are valuable, and they help because they want them. That cuts the need for advertising and marketing.

When something works, we validate that it works, move forward on trust, and track that trust. Open Meetings and Action Cycles, and what follows from them, are where this is learned. The detail, such as vector money, belongs to the move from B to A.

Silo Buster delivers one thing: better relationality between people, so they collaborate within and between organisations, using AI on a shared surface. The experience is iteratively socially validated, in the hour, the day, the week and the month.

The hypothesis, if this holds:

• High-trust practices give early adopters a commercial advantage.
• The centre of power shifts from West Coast tech toward the beach and a body of practitioners.
• As the AI wave reaches everyone, a market turns into an organic community.

Checked against current conditions

The conditions David describes are real. Most claims hold in part, one is contested, and two are hypotheses for Silo Buster to test.

Details added, and open questions

Five details sharpen the argument:

• Stigmergy is Pierre-Paul Grassé's word (1959), from watching termites coordinate nest-building through the traces they leave. The industry's route for agents is direct messaging instead, such as Google's Agent2Agent protocol (April 2025).
• Blackboard systems are the precedent for shared context. Carnegie Mellon's Hearsay-II speech recogniser (1970s) had independent specialists read and write one shared surface. The beach's strata call the psycho-social level a shared blackboard of the mind.
• Murmuration works because each starling tracks about six or seven neighbours, not the whole flock (Ballerini and colleagues, 2008). A room of listeners can coordinate the same way.
• Locus of agency echoes Julian Rotter's locus of control (1966), which ABC State already uses: in C the locus sits with the teacher, in A with the class. Loci 1 and 2, held by the labs, are the C position; Loci 3 and 4, held by people, move toward A.
• Locus 0. On 21 August David named the concurrency of others as Locus 0, prior to the rest. Then the loci ran model, harness, shell, kernel; today they run model, thinking module, shell, portal.

Open before the animation:

• Which is the fourth portal, built today?
• Which naming of the loci stands, and does Locus 0 belong in the picture?
• Is Machus named in a business room, or the direction shown without the name?
• How much of the social danger should it show, when talk in the room cannot change its course?