Research

Hilbert

A neuro-symbolic transformer that reasons across five systems of logic and learns from each input without retraining.

Status: research, not yet released

N5 · sortal N4 · temporal (LTL) N3 · modal N2 · relational (GAT) N1 · first-order strangeloop fastweights

What it is

Research into neural architectures that carry logic inside them rather than bolting a reasoner onto the side. The founding observation: a transformer’s attention layer is mathematically an adaptive filter with feedback — which means it can be made to learn from each input as it arrives.

What is built

Five levels, each with a different logic: first-order; relational with graph attention; modal with a multi-world attention matrix; temporal with LTL and dual causal/bidirectional attention; sortal with ontology-biased attention. Fast weights and episodic memory connect them; a strange loop after Hofstadter feeds the top level’s error signal back down to refine the lower four at inference time; a genetic layer evolves populations of fast-weight configurations against gradient descent. The same commitments drive Thinker, the reasoning layer inside Turing.

How to get it

There is no public release. Hilbert is active research; results will be published when they are ready.