01 · 2026 — Now · PyTorch, Ripser
Epistemic Topology
Do a language model's internal states look different when it is about to give a wrong answer? A PCA baseline on residual-stream activations didn't separate them linearly, so I moved to persistent homology: Vietoris–Rips complexes over the activations, tracking which loops survive as the radius grows.
Code on GitHub ↗