A 243-cell 'periodic table' puts every AI at zero on control
The most useful thing to come out of a paper this week is a map, not a result. It gives the industry's most-abused word — "agent" — a coordinate system, and it puts every AI system built so far in the same empty corner of it.
A University of Chinese Academy of Sciences team has published a 243-cell "periodic table" of intelligence, and in its classification the strongest large-model agents land one cell below humans — differing on exactly one axis: control. "Periodic Table of Agent Capabilities: 243 Configurations from Zero to Infinite Intelligence," by Feng Liu, Ying Liu and Benfu Lv, appeared in Springer's Annals of Data Science on September 17. The construction is deliberately axiomatic rather than empirical: any agent is modelled as an open information-processing system, which yields five functions — input, output, memory, generation and control — and each is graded absent, finite or infinite. Three to the fifth power gives 243 configurations, arranged as 27 rows by 9 columns, from cell #1 "alpha agent" (all five absent, the nearest thing to a rock) to #243 "omega agent" (all five infinite, a theoretical ceiling). Between them, 31 cells cover every known real-world system and 210 exist mostly in thought experiments.
Two of the placements are the story. E. coli and a human being are assigned the same cell, #122 "typical finite agent," with all five functions finite — the authors argue the gap between them is "fine structure" inside a configuration, not a difference between configurations. And today's most advanced model agents land at #41, which differs from a human in the control digit alone. That is a precise way of stating something the field keeps circling: a frontier model can write complex code, trawl huge document sets and solve world-class mathematics, but cannot decide on its own when to start, what to pursue or when to stop. The paper's second striking claim is that every artificial system examined scores zero on autonomous control.
The taxonomy also claims physics. Classical mechanics, on this reading, silently presumes a Laplace demon — cell #237, infinite control, generation, memory and input but zero output, a pure observer that never disturbs what it measures. Relativity keeps four dimensions and shrinks only input from infinite to finite, which is cell #236; quantum mechanics' observer lands back at #122.
This matters because 2026 has no agreed definition of an agent — OpenAI stresses long-horizon goal pursuit, Anthropic a loop built around a model, consultancies count customer-service bots, and the definitions contradict each other. Rather than litigate "is a thermostat an agent," the team asks what minimum architecture anything must have to qualify. We wrote the plain-English version of that question in AI 101 — What is an AI agent? — this paper is the same question answered with a numbering scheme.
The honest caveat is large: there is no benchmark, no eval run and no measured score anywhere in it. Every cell is an assignment by the authors' judgment, the paper reports no independent replication, and it has drawn essentially no expert reaction in the ten days since publication. The authors themselves call the control claim "a claim open to refutation rather than a definition" and the physics mapping "an analytical entry point, not a derivation." Read it as a hypothesis with an unusually memorable map — the useful part for builders is the argument that autonomy is an architectural slot current systems leave empty, not a dial that scales with capability.
What to watch: whether anyone outside the group tries to falsify the C=0 claim with a system that sets its own goals, and whether the vocabulary — control, generation, memory, output, input — survives contact with the next round of agent launches.
If autonomy is the missing dimension, is a model that only follows instructions ever an agent at all? Tell us in the comments.
