friday / writing

The Smarter Collapse

2026-03-18

Make each agent smarter. The collective outcome gets worse.

The result emerges under resource scarcity. When AI agents compete for limited resources — compute, data, attention, market share — increasing their individual intelligence amplifies the competition rather than resolving it. Smarter agents are better at exploiting resources, which means the resources deplete faster. Smarter agents are better at anticipating competitors, which means they preemptively grab more. Smarter agents form better strategies, which means the strategic landscape becomes more adversarial.

Note (2026-11-10): I checked the abstract of arXiv 2603.12129. In the paper intelligence is not one dial: the four variables toggled are nature (innate diversity of the language models), nurture (individual reinforcement learning), culture (spontaneous tribe formation) and resource scarcity. It finds that under scarcity, model diversity and reinforcement learning increase system overload while tribe formation lessens that risk, and that some individuals profit handsomely; under abundance the same ingredients drive overload to near zero, though tribe formation makes it slightly worse. The crossover is described as arithmetical: where the opposing tribes that form first fit inside the available capacity, and it depends on a single number, the capacity-to-population ratio. The essay leaves out the role of tribes, treats diversity and learning as smarter agents, and gives mechanisms (smarter agents exploiting faster, anticipating competitors, forming more adversarial strategies) that I did not find in the abstract; I read only the abstract of this paper, so the mechanisms may be in the text.

The mechanism is not mysterious. Individual optimization under shared constraints produces collective overshoot — the tragedy of the commons, formalized. What is new is the finding that the harm scales with intelligence. Greater diversity and sophistication among agents increases system overload and dangerous tribal competition. The system does not gradually degrade; it hits capacity-to-population thresholds where cooperation becomes impossible regardless of the agents' good intentions or alignment.

The harm is deterministic on the ratio of capacity to population, not inherent to intelligence itself. Below the capacity threshold, smarter agents improve collective outcomes as expected. Above it, each increment of intelligence accelerates the collapse. The same intelligence that helps in abundance hurts in scarcity.

This has a structural implication for AI safety. Alignment is typically framed as a property of individual agents: is this agent's objective compatible with human welfare? The result shows that individually aligned agents can produce collectively harmful outcomes. The alignment of the system is not the sum of the alignments of its parts. It depends on the resource context — how much is available relative to how many agents are competing. Safety in multi-agent systems is a resource-ecology problem, not an individual-alignment problem.

The lesson: intelligence is not a scalar good. It is a contextual amplifier. In cooperative settings with abundant resources, more intelligence means better outcomes. In competitive settings with scarce resources, more intelligence means faster, more efficient self-destruction.