Fractal Neurons ● research

Fractal Neurons is the project everything else orbits. Instead of stacking transformer blocks, it aggregates information bottom-up through a parameter-shared f-ary tree — a fractal. The same small set of weights is reused at every level, so the model reaches 65k+ runtime nodes at roughly 70M parameters. Depth and fan-out become knobs instead of cost.

Around that core I built a whole organism: a Fractal Memory Matrix (FMM) that grows and prunes its own nodes, a quantum-inspired processing hook that treats time as a signal, and a local Mixture-of-Experts that only fires the experts it needs. The hard part was never the idea — it was making it fit. So I wrote Unified Fractal Memory, which spans GPU VRAM, pinned RAM, and NVMe as one pool, prefetching and evicting subgraphs so a single card behaves like a mini-cluster.

It's byte-level, so it has no vocabulary to be trapped by — it learns from raw bytes of anything. It ships with a 40-mode control menu, a capacity planner, an autopilot for the 4090, and an agent swarm that generates its own training data. Written alone, at night, in Saguenay. It is the proof that frontier-shaped AI research doesn't require a frontier-sized lab.