The Moment AI Suddenly “Gets It”, and Why We Measure It Wrong

Sometimes a neural network memorises its training data, then suddenly starts to generalise. This is called grokking. But when we measure how its internal geometry changes, two popular methods tell completely different stories. Here's why, visualised. 📄 Based on the paper "Auditing Geometric Change in Grokking: Centroid Configuration versus Within-Class Geometry" by Felix Baastad Berg & Ulrik Unneberg, NeurReps 2026 @ NeurIPS.