The Oligarch Barely Steers Model Collapse in Multi-Model Ecosystems

URL SCAN: The Oligarch Barely Steers Model Collapse in Multi-Model Ecosystems
FIRST LINE: Computer Science > Artificial Intelligence

THE DISSECTION

This paper isolates recursive training contamination in controlled ecosystems. It varies model output shares, including a 90% dominant probe, and finds that concentration barely changes collapse speed or endpoint. The decisive variables are instead who supplies the shared pool and how susceptible those suppliers are. Replacing half the pool with human text roughly halves drift without changing its direction.

THE CORE FALLACY

Under the Discontinuity Thesis, the dangerous extrapolation is to confuse weak control over a model’s internal collapse trajectory with weak oligarchic power. The experiment measures steering of model behavior, not control of the AI economy. An oligarch does not need to dictate the exact endpoint. Ownership of compute, training pipelines, data admission, deployment, licensing, and capital determines who can build, which text enters the pool, which systems survive, and who captures the remaining rents.

The local result may be sound. Its broader political implication is not. The oligarch may not control the pathology’s precise destination, but still owns the laboratory, the power supply, and the distribution network.

HIDDEN ASSUMPTIONS

• Five generations and thirteen open 1–4B models represent industrial AI dynamics.
• Retraining from clean base weights captures real recursive development while removing accumulated adaptation, capital lock-in, and institutional feedback.
• Fixed market shares model oligopoly, although real oligarchs can alter shares through pricing, access, acquisition, regulation, and infrastructure control.
• A shared output pool resembles actual data pipelines, where filtering, licensing, synthetic-data selection, and proprietary corpora are unevenly controlled.
• The measured collapse metric represents economically relevant capability.
• “Steering” means shifting the model’s endpoint, rather than controlling timing, access, deployment, distribution, or profit extraction.
• Models are treated as peer suppliers, hiding the asymmetry between model owners, infrastructure owners, data owners, and dependent users.

SOCIAL FUNCTION

Partial truth functioning as ideological anesthetic and elite self-exoneration. The paper correctly punctures the simplistic belief that market share alone lets one dominant model command the entire evolutionary trajectory. But “the oligarch barely steers” is true only for the narrow variable tested. It can easily be repackaged into the soothing fiction that concentration is therefore less consequential.

The result actually suggests a harsher system: collapse is not necessarily centrally steerable, but it is still governed by whoever controls the composition of the training pool and the infrastructure that feeds it. Power may be unable to specify the exact decay pattern while remaining fully capable of owning the decay process.

THE VERDICT

Within the tested range, concentration is not the main driver of model-collapse speed or destination; pool composition and supplier susceptibility are. This is evidence about model pathology, not a refutation of the Discontinuity Thesis.

For DT, the paper is ammunition, not salvation. It shows that AI degradation may be more decentralized and less obedient than oligarchic mythology assumes. It does not restore human productive participation, weaken control of compute and logistics, or prevent Sovereigns from capturing the transition. The steering wheel is loose because the road is disintegrating. The oligarch still owns the convoy.