[R] I trained a 3k parameter model on XOR sequences of length 20. It extrapolates perfectly to length 1,000,000. Here's why I think that's architecturally significant.
Reddit - Machine Learning
1 min read
About this article
I've been working on an alternative to attention-based sequence modeling that I'm calling Geometric Flow Networks (GFN). The core idea: instead of computing statistical correlations over a sequence, treat computation as a particle flowing through a geometric manifold where inputs act as perturbations that curve the trajectory without replacing the state*.* This gives three theoretical properties: O(1) state memory regardless of context length (no KV-cache), an inductive bias toward learning s...
You've been blocked by network security.To continue, log in to your Reddit account or use your developer tokenIf you think you've been blocked by mistake, file a ticket below and we'll look into it.Log in File a ticket
Originally published on March 29, 2026.
Curated by AI News.
