random-init-cka-exceeds-trained-cka

IN premise — summaries/2026/08/24/convergence-without-understanding-2026-s4-discussion.md

Created 2026-08-24T17:10:52+00:00

Randomly initialized models exhibit CKA = 0.864 ± 0.007 while trained models show CKA = 0.612 ± 0.011 (p < 10⁻⁵), indicating a large component of observed cross-model representational convergence is architectural (causal masking, residual streams, layer normalization) rather than learned.

Summary

Untrained models with random weights actually look more alike in their internal representations than trained ones do, and the difference is statistically significant. This means that when we see two trained models producing similar internal structures, we should not read that as evidence they learned similar things from data; the similarity is mostly an artifact of the shared architecture, like the same scaffolding making empty rooms look alike.