First-Order Recoverability Collapse in Self-Referential Information Decoders: The Operating Loop of an AI System as a Driven Nonequilibrium Steady State
Résumé
What kind of physical object is an artificial-intelligence system: a machine in the sense a refrigerator is -- dissipating free energy while holding an order imposed from outside -- or a dissipative structure in the sense a convection cell is -- an ordered state that exists only under throughput and loses stability past a critical drive? We argue the answer is split, as it is for the living cell and the star -- the trained artifact is a machine, quenched and storable; the operating loop is a dissipative structure in the informational sense -- and develop the framework in which the loop's classification becomes decidable. Modeling systems that couple inference to irreversible action as finite-capacity decoders under sustained informational driving, we characterize recoverable operation by a feasibility margin, local invertibility, and a stability diagnostic that diverges as capacity saturates. Making the feedback of uncertified output onto load explicit converts this continuous transition into a first-order one at mean-field level, sharpening in the fleet limit: lucid and collapsed states coexist in a cusp-organized bistable region with closed-form spinodals, collapse pre-empts the divergence, recovery is hysteretic, and for ungatedness alpha >= 1 load reduction alone cannot restore operation; reset restores only what is archived, making certification the operative stability lever. Cascades are subcritical branching with mean-field exponent 3/2 and a cutoff set by the grounded fraction of input. An instrumented real-workload pipeline experiment exhibits the collapse just below the spinodal computed from the measured service law, the backlog-delayed hysteretic recovery, and the cascade statistics. This supplies a statistical-mechanics account of the "metastable failures" documented in large-scale distributed systems, identifying recoverable dissipation as the stability criterion.
Citer ce document
Accès au document
Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter
Voir l'article sur le site de la revueAuteur(s)
Statistiques
Consultations : 1
Téléchargements : 0