Definition
Plain language
The conjecture that as AI models grow more capable, they end up representing the world in roughly the same way.
As stated in the literature
The hypothesis that increasing-capability foundation models converge on similar internal representations of the world, motivating cross-model interfacing, distillation, and activation-level coupling.
Why it matters: If models do converge on shared representations, that opens the door to distillation, model stitching, and cross-model interpretability.
For example, two different large models, trained separately, might both encode 'is an animal' along nearly the same internal direction.
Heard on the show
“And that lands on the so-called Platonic Representation Hypothesis — the idea that models converge on similar internal representations as they get capable.”Episode 040 — Two Frozen Models Learn to Whisper: Coupling Through Hidden States