Source
Nielsen et al. (2025), LIT-253, Thm 2.2, Thm 3.1, Cor. 3.2, Thm 4.7, as read in NOTE-738; Harvey, Larsen & Williams (2023), LIT-250, Thm 1 and eq. 24, as read in NOTE-753. The hypothesis itself is Huh et al. (2024), LIT-302 (held in the anthology as ANTH-LIT-458), §4.2 and App. A, as read in NOTE-286; its measurements are stated in THEORY-640.
What was actually shown
Huh et al. (ANTH-LIT-458, §4.2) prove that, under bijective observations and an NCE objective, representations converge to the PMI kernel up to an offset — the log of the density ratio K⁺ that the contrastive losses target (see the first claim in this group). Their measurements, though, use mutual nearest neighbours (App. A), which is much weaker than kernel equality. Nielsen et al. show that for softmax models the output distribution fixes the embeddings only up to an invertible linear map, not an orthogonal one (Thm 2.2, cited from earlier work), and that ε-closeness in KL does not bound representational dissimilarity (Thm 3.1, Cor. 3.2). That counterexample needs the unembedding norm to grow without bound, so that both softmaxes saturate, and the theorem proves only that the embeddings are not related by an invertible linear map; how far apart they are is shown numerically (Table 1). Under a different distance the answer reverses: closeness in their log-likelihood variance distance bounds a representational dissimilarity, d_LLV ≤ ε ⇒ d_f,g ≤ 2Mε (Thm 4.7), and the paper’s stated conclusion is that the answer “depends critically on the choice of distributional and representational distance”. All of this concerns the final embedding and the unembedding only; the paper says earlier layers would need identifiability results that do not yet exist (§6). Harvey, Larsen & Williams supply the kernel-side metric under which kernel closeness is rotation-aligned closeness, for centred linear kernels.
Nielsen et al. cite the Platonic hypothesis only as evidence that learned representations are robust, and do not argue against it. Reading their KL result as a caution for the hypothesis is this record’s inference, and it reaches the hypothesis’s kernels, which are pooled hidden states, only by analogy.
Read together: the defensible formal content of “representations converge” is that kernels converge, and the representation is then fixed only up to the symmetry group of the observed quantity — orthogonal if the kernel is observed, general linear if only the distribution is.
What this does not say
- That the Platonic hypothesis is false.
- That trained models do, or do not, converge in kernel.
- That any one metric (Bures, CKA, mutual nearest neighbours) is the right one.
- That closeness in distribution never bounds representational dissimilarity. It fails for KL in the saturation regime and holds for d_LLV (LIT-253, Thm 4.7).
Connections
- Presupposes THEORY-004. That account shows that a representation is fixed by its centred kernel up to a translation and an orthogonal transformation. This one takes that as the premise that lets convergence of kernels count as convergence of representations up to rotation, and asks what the Platonic hypothesis can then mean.
Note of 2026-10-19: the proved results filed apart
The results of Nielsen et al. this account draws on are now THEORY-723, Active. They cover what the softmax output distribution fixes and how closeness in distribution depends on the distance, for the final embedding and unembedding. Their proofs were read in full (NOTE-738). What keeps this THEORY Proposed is its own step: that convergence in the Platonic hypothesis’s sense is convergence of kernels, reaching the hypothesis’s pooled hidden-state kernels by analogy. The promote_when says what would settle it.
Note of 2026-10-29: what holds it back
The step from Nielsen et al.’s final-layer results to the hypothesis’s kernels, which are cosine kernels of pooled hidden states maximised over layer pairs and measured by mutual nearest neighbours (NOTE-286), is by analogy. No held result bounds a distributional distance by a Bures distance between those kernels, or measures whether high mutual-nearest-neighbour alignment comes with high Bures alignment.