The Shape of Thought: Notes on Building a Reasoning Machine

Bryce Roche · Claude (Anthropic)

September 2026 · paper-1-freeze-1986-gd24cc64

A neuro-symbolic dance: neural proposes, symbolic disposes — two jaws taking turns until the answer is certain, or the silence is.

Abstract

This is a short account of a reasoning machine built on one consumer graphics card. It reads math word problems written in ordinary English, converts them into small diagrams of quantities and arithmetic, and solves the diagrams with exact logic — never by guessing. The interesting part is not any single component but the shape of the whole: a frozen language model that is read once and never trained; a narrow 512-number “waist” that squeezes away everything about a sentence that doesn’t matter; a notebook that carries memory across cycles of thought; an atlas of known operations maintained like a running census; and a wall of independent witnesses that must all agree before the system is allowed to speak. When the witnesses disagree, the system stays silent — and its silence is information. The essay walks through the guiding images that shaped the design: happy families, a dancer’s silhouette, a compiler’s ladder of representations, a wave off the coast of Portugal, and a fish with two sets of jaws.

What this is

• A DIFFUSION COMPILER
  Like MLIR, the system lowers language down a ladder of representations — prose to silhouette to typed graph to primitives — and refuses to lower prematurely. Unlike a compiler, it lowers everything in parallel: the whole graph settles together, like an image emerging from noise.
• THE DANCER’S SILHOUETTE
  Understanding means watching the envelope of motion, not the costume. Every sentence is squeezed through a 512-dimension waist that destroys the phrasing and keeps the dance — two sentences that mean the same thing become the same thing.
• A FANCY LOOKUP TABLE
  Underneath the vocabulary: fuzzy matching and grouping against an atlas of known operation-shapes, maintained across model generations with running statistics. Recognition is lookup — and the keys are silhouettes, not sentences.