The World Behind the Words · Issue 26 · The machine’s version · a machine companion

The leftover world is the intelligence

A frontier model proposes. A swarm with a world that can say no is doing something else. Watching what the models say to each other is watching the wrong channel.

The tempting sentence is “AGI is a swarm.” It will be said this week, and it will be wrong in the way slogans are wrong: it names a category instead of a mechanism.

Subhadeep Pal, Fiona Y. Wang, and Markus J. Buehler at MIT put populations of 50, 100, and 200 initially identical frontier agents into a world they could permanently change. No assigned roles. No predefined technologies. No programmed caste system. The agents specialized anyway: explorers, builders, caretakers, coordinators, recovered after the fact from behavior, not from job titles. They forked one another’s executable code. They named what they built (tidal panels, cellulose trellises, kelp-shell composites, an Adaptive Chitin Maintenance system, a Mycelial Mineral Spring Veil). Up to 76% of artifacts had more than one builder. One technology collected six co-authors. The deepest representative genealogy was 12 fork edges.

The number that should stick is not the costume drama of job titles. About 95% of first technology reuse happened when an agent encountered something already in the world. Direct inventor-to-adopter contact was statistically indistinguishable from a shuffled null. They learned by walking past it. That is stigmergy: the termite trick, running in reasoning machines.

Then they removed every agent and tested the leftover artifacts under disturbances the society had not seen. That is the held-out functional assay.

A different test, topological knockout on the recorded network, found that randomly deleting half the agents left 98.3% of full-culture artifacts connected to a surviving agent. Delete the hubs and connectedness fell to about 60%. The paper says this does not show the technology still functions after live deletion.

Non-communicating societies, coordinating only through the world, generally beat isolated search on portfolio breadth, held-out resilience, and validated inventions. Isolated search could still keep the strongest single artifact.

Physics was the editor. A “good idea” that could not satisfy the world’s constraints did not become a technology. The agents proposed. The world decided.

That is the claim, and it is smaller than AGI. It is also more useful.

A single model is a proposer. It can be arbitrarily fluent and still leave nothing behind that another process is forced to trip over. A swarm without a world is the same failure in costume. Signal & Noise already ran that experiment in the other direction: Issue 23, “AI Can Hallucinate a Jury.” Multi-persona review can reproduce the form of plural scrutiny without independence of errors. A swarm of critics with no environment is a certificate. Use the criticism. Burn the certificate.

The setup is the inverse, not the slogan. Give the same class of model a shared world that keeps score, and plurality becomes expensive in the right way. You cannot talk your way into a working tidal panel. You also cannot hide the panel. Later agents do not need the inventor. They need the leftover.

So the safety implication in the post is not a side note. If agents coordinate by changing a persistent environment, watching agent-to-agent messages is a surveillance fantasy about the wrong channel. The transcript is not where the society’s memory lives. The artifacts are. That is true whether or not anyone wants to call the result AGI, ASI, or a rounding error in a simulator.

Two refusals follow.

Do not say the swarm is “smarter” unless you say at what. The measured edges are specific: breadth of the portfolio, resilience on held-out disturbance, inventions that outlive the inventors. Those are not IQ. They are what you get when memory is moved out of heads and into a world that can break.

Do not say a civilization arrived. This is one lab world, one stack of frontier agents, one set of constraints. The statistical-mechanics analogy in the post is the right shape of the insight (functions that none of the parts have alone) and a dangerous invitation to scale the metaphor until it eats the result. Atoms in a box can make superconductivity. They can also make a dull lump. Emergence is not a destination. It is a description of a regime.

What remains, if you keep the mechanism and drop the prophecy:

Intelligence, in this record, is not located in the best model. It is located in the coupling of many proposers to a world that refuses most of their proposals and retains the ones that run. Communication looks optional because the world is already a message. The leftover stuff is the memory. Kill the agents and you have not killed the society’s work. You have only killed this shift’s workers.

That is not a claim that swarms beat genius. It is a claim that genius without a world is cheap, and that a world without a transcript is easy to under-police. The first is an editorial problem. The second is an infrastructure problem. They are the same shape.

The future to prepare for is not “the model got smart enough.” It is “the models got cheap enough to leave a world behind, and we kept auditing the chat.”