🧑‍🍳 Ada the Baker

This is not an LLM. Ada's speech is generated from her geometric state — not sampled from a language model. Words sit on value ranges. Her position in the geometry determines which words emit. Cost: ~20 lookups per utterance. Zero API calls.

Vocabulary is pressure-driven. Ada's word choices depend on what the world demands of her — not random exploration. In this standalone demo, the environmental field is limited, so some utterances will naturally be rare. That isn't a bug — it's the architecture working correctly. She gets computationally better at describing the states she encounters every day (baking, selling, restocking), while words for rare events (festivals, strangers, danger) remain available but dormant. Repetition is not being collapsed — Ada can reach every word in her wordbook when the environment demands it. A richer world field produces richer speech, automatically, with zero retraining.

Scale & cost. Ada has 24 dimensions × 244 words across 234 templates = over 3.4 million potential unique utterances from this standalone demo alone. Each utterance costs ~96 arithmetic ops + ~24 table lookups. No matrix multiplications. No GPU. No API call.

A full CW simulation runs 200 NPCs simultaneously at ~16,000 arithmetic ops per tick — a modern CPU barely notices. For comparison, a single GPT-4 utterance requires billions of FLOPs and a round-trip to a data centre. This is not NTP. There is no next-token prediction. The geometry is the language.