After the bubble

GitHub ·

Hand-written GGUFs as state machines: no training, every weight placed by hand. A standard ollama server runs them like any other model.

Snake — the classic game, food and deaths computed in the weights 22 MB
Wolf — a first person raycaster 191 MB
Life — Conway's game of life on a 12×12 torus 64 MB
kv — a key-value store whose database is the context 321 MB
base64 — text in, base64 out, encoded by the weights 18 MB
www — a web server made of weights, serving the previous version of this site <1 MB

What is this?

These are machines whose engine is a GGUF file, served by a standard ollama server. The GGUFs have been hand crafted to behave deterministically: the constraint was to build something that can interact with the frontend only through the ollama runtime.

Attention heads read the current state from the conversation (ollama is stateless), MLPs implement the rules, and greedy decoding over saturated logits makes the whole system deterministic. It all started from a minuscule GGUF that always returned the same sentence; the only possible future was, of course, Doom on ollama.

All the frontend does is send inputs and draw what it receives: the state itself is always computed by the model from the whole conversation. Each page has an info button explaining what its GGUF does.

See how it is done on GitHub.