THE NOVEMBER EXPERIMENTS
Less theory.
More “let me try.”
Watch a story take shape, test a tiny summarizer, and explore what happens when you change a model’s weights.
Two live NCM demosBefore & after recordingsRuns on AECH
LIVE · TINYSTORIES
A story in twelve layers.
The model starts with coarse codes and refines them into a story. Change the seed to explore another result. This model generates stories without a prompt.
Ready when you are.
Your generated story will appear here.
See the 12 decoded levels
LIVE · XSUM
A short summary. Nine levels.
Paste a short article or try the fictional library example. This research model can invent facts—compare its summary with the source.
Up to 512 input tokens; up to 48 output tokens.
Your summary will appear here.
See the exact input excerpt
See the 9 decoded levels
RECORDED · WEIGHT-SPACE PROGRAMMING
Same model. A small weight edit.
Explore actual local runs of Qwen3.5-0.8B before and after the published patch. Two matrices change; 130,188 stored values differ. Both sides use the same prompt, seed, and sampling settings.
These are recorded Apple GPU runs, not live inference or a full benchmark reproduction. We include mixed outcomes, not just the featured example.
After · published edit
Model output
“Token limit” means the run stopped at its budget. It is not proof of a reasoning loop. Generated reasoning may contain errors.