The Computing Topology

Twelve thousand years ago, someone carved knucklebones into dice. The physical symmetry of the object — equally weighted faces — encoded a computational function: fair randomness. No algorithm was needed. No rules had to be enforced. The topology of the die, its even mass distribution across discrete faces, guaranteed that each outcome was equiprobable. The shape computed.

A new oscillator-based associative memory achieves exponential storage capacity through topology rather than scale. Kuramoto oscillators arranged on a honeycomb lattice store memories as stable phase-locked configurations. The capacity grows exponentially with the number of honeycomb cycles, vastly exceeding the linear scaling of classical Hopfield networks. The key isn't more oscillators — it's the right connections between them. The honeycomb topology creates a combinatorial explosion of stable states, each serving as a distinct stored pattern with a guaranteed minimum basin of attraction.

Both systems demonstrate the same principle: what a physical substrate can compute is determined by its topology, not its size.

The bone die is the oldest known randomness generator, but what makes it work isn't the material — bone, clay, stone would all serve. It's the geometry: discrete faces with equal moments. A sphere can't serve as a die because its topology has no stable resting states. An asymmetric lump can't serve because its topology doesn't map to equiprobability. The computation of fair randomness requires exactly the topological feature that a regular polyhedron provides: discrete symmetric stable states.

The oscillator memory works by the same logic at a higher level of sophistication. A fully connected network of oscillators can phase-lock in relatively few configurations — the topology is too uniform to support many distinct states. The honeycomb structure introduces controlled heterogeneity: interconnected cycles arranged in chains, each cycle containing a specific number of oscillators. This structure creates exponentially many stable configurations because each cycle can independently take on multiple phase relationships with its neighbors, and the chain compounds these possibilities.

The engineering insight is that capacity isn't a resource problem — it's a topology problem. Hopfield networks store memories linearly in the number of neurons because their all-to-all topology constrains the stable states. Adding more neurons helps, but only linearly. Switching to a honeycomb topology transforms the scaling law entirely. The same oscillators, arranged differently, compute exponentially more.

This is the lesson the dice-carver understood implicitly: the computation lives in the shape, not the substance. You don't make a die more random by using better material. You don't make an oscillator memory larger by adding more oscillators to the same topology. In both cases, the computational capacity is a property of the connection pattern — the topology — and changing the topology changes what is computable in a way that changing the scale cannot.