---
title: "Minimum Viable Interiority"
description: "Iulia Ionescu, Murad Khan, Alasdair Milne and Cezar Mocan accept the claim that all intelligence is collective and then ask what it would take to build an individual from scratch."
type: "reading-notes"
authors:
  - "Iulia Ionescu"
  - "Murad Khan"
  - "Alasdair Milne"
  - "Cezar Mocan"
published: "2025-05-10"
original_url: "https://coginfra.antikythera.org/"
doi: "10.1162/ANTI.5CZG"
canonical_url: "https://antikythera.wiki/work/journal/coginfra/minimum-viable-interiority"
md_url: "https://antikythera.wiki/md/work/journal/coginfra/minimum-viable-interiority"
last_updated: "2026-08-23"
site: "Antikythera Wiki"
---

# Minimum Viable Interiority

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** Iulia Ionescu, Murad Khan, Alasdair Milne, Cezar Mocan
- **Published:** 2025-05-10
- **Kind:** Studio projects · Journal article
- **DOI:** https://doi.org/10.1162/ANTI.5CZG
- **Original:** https://coginfra.antikythera.org/
- **Length:** 5.1k words
- **Part of:** [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra)
- **This page:** https://antikythera.wiki/work/journal/coginfra/minimum-viable-interiority
- **Notes generated:** 2026-08-23

## Summary

[Iulia Ionescu](https://antikythera.wiki/people/iulia-ionescu), [Murad Khan](https://antikythera.wiki/people/murad-khan), [Alasdair Milne](https://antikythera.wiki/people/alasdair-milne) and [Cezar Mocan](https://antikythera.wiki/people/cezar-mocan) accept the claim that all intelligence is collective and then ask what it would take to build an individual from scratch. Their answer runs through interiority, defined not as consciousness or qualia but as privileged access to the internal states that drive action. Taking video-game non-player characters and language-model agents as working philosophical zombies, they propose building interiority up from a pandemonium architecture, Oliver Selfridge's 1959 model of competing daemons reconciled by a selector, with layers of functional closure nested inside one another. The hypothesis is that some number λ of such layers, the minimum viable interiority constant, is what makes an agent opaque to outside observers and to its own introspection, and they sketch an experiment in DeepMind's Concordia library to test it. The slogan they arrive at inverts the usual one: the individual is greater than the sum of its collectives. The paper sits in the Mimesis of Mimesis theme of the Cognitive Infrastructures compendium.

## The argument

The authors grant Falandays and colleagues' thesis that all intelligence is collective and ask whether, after that turn, the individual still does explanatory work in accounts of agency. Their method is a counterfactual: if intelligence is collective all the way down, reconstructing a functionally singular entity from multiple components should show what the individual adds. They disclaim any theory of sentience or consciousness; the target is interiority as a feature of action between agents.

Interiority is then given a pedigree and a functionalist reframing. Thomas Duddy defends the interior–exterior duality against its Cartesian associations; Hannah Arendt's two-in-one makes thinking an internal dialogue hidden when one appears to others as one; John Macmurray makes the individual an actor rather than a thinker. The authors take encapsulation as a necessary but insufficient condition of interiority and repurpose Chalmers's p-zombie, normally an objection to functionalism, as a baseline to which architecture can be added. NPCs are the concrete p-zombies, language-model agents the more capable ones, and both are offered as experimental shells.

The architecture is Selfridge's pandemonium, read alongside Minsky and Hawkins: multiple daemons inside one agent, with a selector daemon arbitrating. The novelty is the nesting. Borrowing organizational closure from Mossio and Moreno, the authors define a functionally closed layer as one whose events are causally explainable only within its own domain. Models on one layer intra-act; closed systems inter-act; each layer is constituted by the one below and irreducible to it. Selection, modelled on Gazzaniga's left-hemisphere interpreter, externalises internal dissonance as the unified behaviour of an individual and thereby produces a self-model.

The hypothesis follows: interiority emerges at λ levels of closure, each adding opacity. The experiment scales Concordia agents from one daemon through three and seven to 2^λ − 1, holding everything else fixed, to look for behavioural differences. The conclusion distinguishes a functionally closed collective, which is an individual, from an open one, which is a swarm, and proposes metrics of decision opacity, intra-agent interaction density and self-model coherence to locate λ. Interiority is offered as an incremental step between the p-zombie and the person.

## Section by section

### 1 Introduction: From the Individual to the Collective (and Back Again)

Collective intelligence is surveyed across biology, human systems and technical systems, with particular purchase on foundation models built from the traces of vast numbers of people. The authors set up their counterfactual, define interiority, and raise the possibility that synthetic agents may become black-boxed to themselves as well as to observers.

### 2 Engineering Interiority

The philosophical ground: Duddy, Arendt, Macmurray, and the functionalist question of minimal conditions. NPCs are the p-zombie's digital elaboration, with believability research holding that a good NPC should be indistinguishable from a human-controlled character; language models are the more capable version, valuable precisely because their internal states are uncertain.

### 3 Building NPC-Zombies with Pandemonium Architecture

A short history of NPC design, from rule-based systems through behaviour trees to The Sims' utility-based needs and the planning architectures ACT-R and SOAR, whose common thread is concurrent processes competing for priority in one decision-maker. Selfridge's tiers of feature, cognitive and decision daemons are adapted so that several models coexist inside one agent under a selector daemon.

### 4 Intra-Agent Intra-Action and Functional Closure

The theoretical core. Mossio and Moreno's organization of constraints is extended from physical and biological to psychological self-maintenance. A functionally closed system is irreducible to a genealogy of causes; each closure explains phenomena local to its layer. Intra-agent and inter-agent interaction are distinguished, and the diagrams, not reproduced in the text, show planes of daemons whose intra-action yields emergent components on the layer above until a final layer presents the whole.

### 4.1 Selection

The selector daemon arbitrates among the dynamics of each closure and, in doing so, re-presents internal dissonance as unified action. Following Gazzaniga and LeDoux on the left hemisphere as interpreter, the authors claim that selection interiorises the individual by generating an emergent self-model.

### 4.2 Minimum Viable Interiority Constant

Interiority is defined as emergent, partly visible in behavioural irregularities and otherwise hidden by closure. A system is hypothesised to need λ levels of closure to achieve it, each level adding opacity to observation and introspection. The section states the inverted slogan: the individual is greater than the sum of its collectives.

### 4.3 Pandemonium Architecture Versus Neural Networks

Transformers are also hierarchical, but pandemonium's daemons have designed roles and are therefore interpretable, which the authors prefer for exploratory work on interiority. They leave open integrating the two.

### 5 Experiment Design

The Concordia proposal. A baseline of N single-daemon agents; a two-level version with two input daemons and a selector, three daemons per agent; a three-level version with seven; and in general N × (2^λ − 1) daemons, which the authors concede becomes theoretical as λ grows. The aim is to observe and quantify behavioural differences across levels of closure.

### 6 Conclusions

A compatibilist position: the collective composes the individual without undermining it. The closed-versus-open distinction separates individuals from swarms and may reclassify some organisations as individuals. A final paragraph speculates, via Metzinger and Dennett, that hard-problem qualities might appear at higher orders of the same complexity.

## Key concepts

- **[Minimum Viable Interiority](https://antikythera.wiki/terms/minimum-viable-interiority)** — the least architecture that gives an agent privileged access to the internal states driving its action; hypothesised to require λ nested levels of functional closure.
- **Interiority** — privileged access to internal states that drive action, distinguished from consciousness, sentience and qualia.
- **[Pandemonium Architecture](https://antikythera.wiki/terms/pandemonium-architecture)** — Selfridge's model of parallel daemons whose outputs a decision daemon reconciles; here nested as layers inside a single agent.
- **[Functional Closure](https://antikythera.wiki/terms/functional-closure)** — a layer whose events are causally explainable only within its own domain, adapted from Mossio and Moreno's organizational closure.
- **Selector daemon** — the arbiter that turns internal dissonance into unified action and, the authors claim, produces a self-model.
- **NPC-zombie** — a non-player character or language-model agent treated as an experimental p-zombie to which interiority can be added.
- **[Functionalism](https://antikythera.wiki/terms/functionalism)** — the stance that interiority is to be specified by its functional conditions, not its phenomenal ones.

## Connections

- [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) — the compendium's Mimesis of Mimesis framing asks whether machine representation is mimesis of mimesis or something else; this paper answers by engineering the representing subject from parts.
- [Generative Topolinguistics](https://antikythera.wiki/work/journal/coginfra/generative-topolinguistics) — the sibling paper in the same theme, working on representation from the language side rather than the agent side.
- [Cognition With and Beyond the Brain](https://antikythera.wiki/work/journal/coginfra/cognition-with-and-beyond-the-brain) — tension. That paper extends cognition outward into assemblages; this one asks what encloses it.
- [Organoid Array Computing](https://antikythera.wiki/work/journal/coginfra/organoid-array-computing) — a parallel question about whether networks of units yield an individual, posed in neural tissue rather than software.
- [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence) — agreement on Gazzaniga's interpreter and on cortex as many competing models; Agüera y Arcas is more willing to locate the self in the collective than this paper is.
- [Modes of Cognition](https://antikythera.wiki/work/journal/modesofcognition) — the autopoietic lineage from which organizational closure descends.


## Terms used

- [Functional Closure](https://antikythera.wiki/terms/functional-closure)
- [Minimum Viable Interiority](https://antikythera.wiki/terms/minimum-viable-interiority)
- [Pandemonium Architecture](https://antikythera.wiki/terms/pandemonium-architecture)