NEURAL ACTIVITY 2,048 of 49,393 sampled
STATE RASTER 256 neurons × last 96 tokens
DECK A CUED
I32rank64fixed10k
15,552 updates 3,956,469 PARAMS
TEMPERATURE 0.70
TOP-P 0.90
PACE 1.00×
DECK B SILENT
Awaiting a prompt
0 TOKENS 0.0 TOK/S
Every number above is measured from the running model. The point cloud shows real
soma positions for sampled neurons and its brightness is |h|, a
continuous rate — not biological spiking. The body is a female
NeuroMechFly rig; the connectome is from a male specimen. Its movement is
illustrative animation, not a motor output of the language model.
The wiring is frozen: all 9,050,172 synaptic weights are byte-identical to
the reference model's. The neurons are not — each of the 49,393 has a
learned input gain, recurrent gain and bias (148,179 parameters), and training
moves the effective per-neuron scaling by about 51% in relative L2. That rescales
each neuron's incoming sum by one factor; it cannot change the relative strengths
or signs of individual synapses.
Trained on 10,000 TinyStories at a 16,800-update budget — 1.83 passes, so it is
not converged. It scores 2.9493 cross-entropy on a held-out population, against
3.9882 for the 52,756,661-parameter model it derives from.
Connectome: MaleCNS v1.0 (CC BY 4.0) — FlyEM / HHMI Janelia, University of Cambridge,
MRC LMB, Google Research. Architecture and tokenizer after
ngxson/fly-llm-hf .
Graph: fly-connectome-49k .
It only knows how to ramble about Lily and Tom.