What it is
K2K replaces the standard token-embedding paradigm with matrix-state primitives on a constrained manifold (Frobenius norm fixed to one). State updates are matrix multiplications that stay on the manifold, which gives stable gradients and a well-defined geometry the lab can derive theorems against.
NLP stack
The lab is building a native NL stack on K2K — morphology, syntax, and semantics — rather than treating language as opaque tokens for a transformer. This makes interpretability a property of the architecture, not a post-hoc analysis.
cross-references