Hyper-convergence

Bernac, Farghly and Zhang's name for what happens under the intersection protocol: when two systems are optimized to agree, the space of outputs shrinks toward "the intersection of their supports." Their instances are CLIP-guided diffusion (image and text embeddings forced into one space), RLHF (a language model agreeing with a reward model of average satisfaction — mode collapse, preference drift) and GAN generator/discriminator collapse. The aesthetic symptom is the "canny valley": derivative, eerily familiar content. The diagnosis is systemic, not essentialist — the blandness is a property of the interaction protocol, not of machines. Cf. Agüera y Arcas's "monoculture is not scaling; it's collapse."