Cluster: workshop-as-shared-build · IamHITL Field Notes

Small Group, Large Group, Different Loops

By Michael Polzin · Published 2026-07-01 · Human-in-the-loop practice

Four people around a table is not a smaller version of forty people in a room. It is a different animal. The loop between what a group notices, what it names, and what it decides changes shape as the room grows, and the practice has to change with it.

The human-in-the-loop workshop, at its core, is a small piece of shared inference: a group builds a working map, tests it against what actually happens, updates the map, and does it again. That is the loop. The mechanics of that loop, the pacing, the seating, the way a facilitator holds silence, all shift when the count in the room shifts. This is a field note on three sizes we run, and where each one tends to fail.

Four people: the tight loop

At four, the loop is nearly frictionless. Every person can speak in a single pass without the meeting running long. Turn-taking is intuitive, cross-talk is easy to repair in real time, and the pace of update, someone says something, the group's map visibly shifts, is the closest to what active-inference practice describes in the literature (Class E, Parr, Pezzulo, Friston, 2022; Namjoshi, 2026). One well-timed question can restructure the whole conversation.

Where it fails: at four, one dominant voice becomes the whole prior. The group's shared map bends toward that person's model, and no one else has enough acoustic space to publish a competing map without it feeling like confrontation. The falsifier here is boring but real (Class F): if you cannot get an even distribution of talk-time across a two-hour session at N=4, the loop has collapsed into a monologue with a chorus. Fix is structural, not tonal: timed rounds, written first-pass, facilitator explicitly holding air for the quieter three.

Twelve people: the seam group

Twelve is the seam. It is the largest room where every participant can still hear every other participant think, and the smallest room where subgroup dynamics start to appear on their own. This is the workshop's home size for a reason. It gives you enough diversity of priors that the shared map actually gets stressed, and enough intimacy that stress does not fragment the room.

The loop at twelve has a two-beat rhythm: a whole-group frame, a subgroup dive of three or four, and a return to whole-group synthesis. Pacing is the whole game (see Pacing as Pedagogy). Rush the subgroup phase, and the return synthesis is thin. Skip the whole-group frame, and the subgroups solve the wrong problem in parallel.

Where it fails: twelve is large enough that a single confused participant can quietly disengage, and no one notices for forty minutes. The configuration signal for a healthy N=12 room is that the facilitator can name what each participant contributed at the halfway mark without checking notes (Class C, from our running practice logs). If you cannot, the loop is leaking someone.

Forty people: the layered loop

Forty is a different animal. You cannot run one loop at forty. You have to run a slow outer loop and a fast inner loop simultaneously. The outer loop is whole-room: the frame, the anchor case, the shared vocabulary. The inner loop is small-group, usually pods of four or five, running the actual updating in parallel. The facilitator is no longer inside the update loop. The facilitator is orchestrating a room full of them.

This is where the practice gets scored by seams. The transitions between whole-room and pod, and the report-back structure that pulls pod-level updates back into the room-level shared map, are where the whole thing lives or dies. A report-back that lets each pod speak for ninety seconds is theatre. A report-back structured to publish one specific claim per pod, tagged with an evidence class, actually updates the room's map.

Where it fails: at forty, the temptation to over-produce the room, decks, breakout timers, structured worksheets, is enormous, and every layer of production reduces the surface area where a real update can happen. Our working hypothesis (Class U, still being tested across the 2026 cohort) is that a forty-person room with three light artifacts outperforms the same room with ten polished ones. We have not yet run enough of these to publish a strong claim.

What stays constant

The loop shape changes. The posture does not. At every size, the practice is the same working hypothesis on an attainable path toward General Natural Intelligence, natural not artificial: a group builds a map together, tests it against what happens, updates it, and marks what it is not yet sure about. Evidence gets classed. Falsifiers get named. The room learns in the open. The work is trauma-informed but non-clinical: we hold space carefully, we do not diagnose, we do not treat.

If you are wondering which size fits the work you are trying to do, the honest answer is that it depends on what the shared map has to hold. A tight product decision fits four. A team learning a new practice together fits twelve. A cohort trying to change how a whole organization thinks fits forty, and the room has to be built for it.

Evidence classes referenced in this post: E (expert citation, active-inference literature), C (configuration, our facilitation logs), F (falsifier, the collapse mode at N=4), U (unverified, the N=40 light-artifact hypothesis). This is a working hypothesis with growing evidence, tested in the open. Bring the falsifier and the record gets stronger.