Attention Is Not All You Need
Completing the Transformer with a stabilising invariant for coherent, depth‑capable reasoning.
Attention Is Not All You Need argues that Transformers possess immense expressive power but lack a stabilising invariant capable of maintaining coherence across depth. As the paper states, attention “provides reach, not coherence,” enabling dynamic pattern selection but offering no global mechanism for continuity, identity preservation, attractor dynamics, or recovery. Empirical evidence shows consistent long‑horizon collapse: contradiction rates rise, coherence decays, activation divergence increases, and latent trajectories drift.
The paper introduces Clarus as the missing architectural pole — a geometric invariant that anchors hidden‑state evolution, reduces drift, stabilises identity, enforces continuity, and provides a global attractor. Together, attention (expression) and Clarus (ground) form a two‑pole architecture capable of sustained, convergent, mind‑like reasoning. The core claim is structural: Transformers alone cannot reach AGI because they are one‑pole systems; AGI requires both expression and ground.