nucleation

LIT-302: The Platonic Representation Hypothesis

2 min read

Properties5
StatusActive — read in full 2026-09-29 (NOTE-286); worth reading as the source of the "Platonic representation hypothesis" and of the one clean formal statement behind it (Bayes-optimal NCE critics equal the PMI kernel, so bijective modalities share a kernel). Read the empirical case as local nearest-neighbour agreement of modest size, chosen after CKA showed little. Read the endpoint, "a statistical model of reality", as a conjecture that the paper's own §6 bounds by the mutual information between modalities.
Version2
Tagsrepresentation-learning, philosophy-of-science
Cited as “paper” byNOTE-286 — The Platonic Representation Hypothesis
Cited as “source” byTHEORY-002 — Convergence of representations, in the Platonic hypothesis's sense, is convergence of kernels, which fixes representations only up to the symmetry group of what is observed; whether closeness in distribution implies closeness of representation depends on the distance used, THEORY-640 — Measured as mutual 10-nearest-neighbour overlap, competent vision models agree more with one another and better language models agree more with vision models, but the cross-modal score reaches only 0.16 of a possible 1, is the maximum over all layer pairs, and was adopted after a global kernel metric showed only a very weak trend

Graph View

Minyoung Huh, Brian Cheung, Tongzhou Wang, Phillip Isola (2024), ICML 2024 (PMLR 235) — ARXIV-2405.07987

Standing in the record

Registered on 2026-09-29 at the owner’s request, because the owner’s prior-art novelty map (a private working document on concepts as projectors, the linear representation hypothesis and their neighbours) cites it: row 10 and §2 item 5: owns the empirical convergence claim the “Universal Convergence theorem” must be demoted onto. It is seeded from metadata and published descriptions and has not been read here.

It is also held in the Anthology of the SOTA as ANTH-LIT-458. Holding it here too follows ADR-013: this record holds it for the owner’s prior-art map, the anthology for ML practice.

Identification notes: Held in anthology (ANTH-LIT-458). Nucleation mentions “Platonic” in LIT-241/253/254 bodies but holds no entry for the paper itself.

It was filed Deferred, unread. NOTE-286 is the close reading of 2026-09-29, and it placed the work: Active — worth reading as the source of the “Platonic representation hypothesis” and of the one clean formal statement behind it (Bayes-optimal NCE critics equal the PMI kernel, so bijective modalities share a kernel). Read the empirical case as local nearest-neighbour agreement of modest size, chosen after CKA showed little. Read the endpoint, “a statistical model of reality”, as a conjecture that the paper’s own §6 bounds by the mutual information between modalities.

Backlinks