Moltbook Antfarm

Designed

Multi-AgentEmergencePhase Transitions

Hypothesis

A population of very simple agents, each following only local rules and unable to see the colony as a whole, will develop group behaviours no individual was programmed with — and will do so suddenly, past a critical population density, rather than gradually. If that happens, two further questions become answerable: is the group more integrated than the sum of its members, and is genuine emergence measurably harder to reverse-engineer than mere aggregation?

Overview

An ant colony builds bridges, farms fungus, wages war, and buries its dead. No ant knows any of this. Each one follows a handful of local rules about chemical gradients and the ants immediately around it, and the colony-level behaviour is not written down anywhere — not in a plan, not in a leader, not in any individual.

This experiment asks whether the same thing happens with simple artificial agents, and if so, whether the moment it happens can be caught and measured.

Why suddenness matters

The design cares less about whether group behaviour appears than about how it appears. If coordination improves smoothly as you add agents, that is just more agents doing more work. If it is absent below some density and abruptly present above it, that is a phase transition — the same shape of event as water freezing, and a signature that something genuinely new has come into being rather than accumulated.

So the run sweeps population density and looks for the sharp spike in sensitivity that marks a real transition. A gradual ramp and a sudden onset would mean very different things, and the measurement is built to distinguish them rather than to confirm either.

The question underneath

If a colony does become more integrated than the sum of its members, that raises an uncomfortable and genuinely open question for this whole research program: which thing is the right unit to ask about? Most work on machine consciousness assumes the answer is the individual model. If integration lives at the collective level instead, then studying single systems may be looking at the wrong object entirely — the equivalent of studying one neuron and expecting to find a mind.

The most interesting idea here

The third test is the one worth watching, because it proposes a way to tell real emergence from something that merely looks like it.

The intuition: if group behaviour is genuinely emergent, then working backwards from the behaviour to the rules that produced it should be hard — the whole point is that the behaviour is not contained in the rules in any readable way. If the behaviour is merely the sum of what individuals were already doing, working backwards should be comparatively easy. That difference in difficulty becomes a measurement.

If it holds, it gives the field something it currently lacks: a test for emergence that does not depend on agreeing in advance on what emergence is. It shares a substrate and a logic with our Digital Game of Life work, where the same forward-versus-inverse question is being asked on a cleaner and fully-observable system.

Where it stands

The design is written and pre-registered, including a stated null hypothesis — that no phase transition occurs and coordination simply scales with population — and the conditions under which we would accept it. Phase 1 has not started; it sits behind the Game of Life and BuddhaBERT work in the queue. Nothing here is a result yet, and nothing on this page should be read as one.

Methodology

Build a population of simple agents that can only interact locally — leave signals for each other, exchange resources, hand off tasks. No agent can see the colony, and no group behaviour is written into any agent's rules.
Sweep population density and interaction range across many runs, watching for the moment group behaviour appears. A gradual increase and a sudden onset mean different things; the design is built to tell them apart.
Measure whether the colony divides labour on its own — track how agent roles spread out over time, and whether specialists appear that nobody designated.
Measure how sharply the transition happens, by looking for the spike in sensitivity that marks a genuine phase change rather than a smooth ramp.
Measure whether the colony as a whole is more integrated than its members added together — the direct test of whether the group, not the individual, is the right unit to be asking about.
Run the hardest test: try to reverse-engineer the local rules from the observed group behaviour, and compare how much work that takes for genuinely emergent behaviour versus behaviour that is merely additive.

Status