Research & Papers

Versor: A Geometric Sequence Architecture

This radical new architecture could make Transformers obsolete for scientific AI.

Deep Dive

A new AI architecture called Versor uses Conformal Geometric Algebra (CGA) to replace core Transformer operations, achieving state-of-the-art results with dramatically fewer parameters. It uses 200x fewer parameters than Transformers, achieves 99.3% accuracy on topology tasks (vs. 50.4% for ViT), and offers O(L) linear complexity. It consistently outperforms Transformers and Graph Networks on chaotic N-body dynamics, multimodal benchmarks, and shows superior out-of-distribution generalization where Transformers fail catastrophically.

Why It Matters

It provides a scalable, interpretable foundation for geometrically-aware scientific modeling that could revolutionize physics and biology AI.