Reality and Practice: A Relational Reading of the Platonic Representation Hypothesis

The Platonic Representation Hypothesis (PRH) of Huh et al. (2024) argues that large neural representations trained for different objectives on different data are converging on a shared statistical model of reality, framed in the language of Plato’s cave. While we agree with the empirical observations of convergence as well as the formalisation using kernel-alignment, we argue that the philosophical framing does not follow from the formalism. The mathematical apparatus PRH employs is purely relational: representations are compared by the kernels they induce, and the fixed point identified by PRH’s contrastive analysis is a co-occurrence statistic of the data-generating distribution. The Platonic ontology is an addition on top of this apparatus, not a requirement of it. We develop an alternative reading drawn from Wittgenstein’s later work, in which categories are useful compressions held together by family resemblance, vision and language are coupled through human practice under physical constraint, and the convergent geometry is a snapshot in time of a non-stationary practice. On this reading, PRH describes convergence toward a relational system of use, not convergence toward an ideal object.