EPISTEMIC MONOCULTURE AND CORRELATED FAILURE: A MIMO DIVERSITY GAIN FORMULATION FOR MULTI-AGENT ADVISORY SYSTEMS

Анотація

Executive advisory systems built on a single large language model behave, in information-theoretic terms, as one noisy channel between the true state of the environment and the resulting decision. For a Markov chain X → Y → Z, the data processing inequality shows that no downstream reasoning can recover information that the channel has already discarded. Stacking a second instance of the same foundation model on top of the first does not add a channel; it duplicates the one already there. This is the structural content of algorithmic monoculture, where convergence on a single more-accurate algorithm can still reduce aggregate welfare through correlated failure.