myceliumA living pattern laboratory
Preparing Python

Experiment 01 / Emergence

From one
cell, form.

A single seed. A shared local rule. Watch a pattern assemble itself, one update at a time.

Specimen & training
Choose a specimen01—03
Preparing the specimen collection…

Let the seed grow, then try removing part of it. The target is a reference, never an input to the live rule.

Live culture32 × 32 cells · 16 channels
LOCAL RULE / NCA48 → 64 → 16
Training target—
032-cell simulation domain32
Preparing a place to grow.

Loading the local simulation engine…

Grow a pattern, then drag to erase.

Loading
Simulation step0000
Occupied cells
—
of 1,024
Mask overlap IoU
—
1 = target mask
Empty-relative error
—
1 = empty baseline
Balanced error
—
lower is closer

A trace of growth.

A

Reconstruction error in this live run. Lower values are closer to the known target.

A trace will appear as the culture grows.

New run. No interventions yet.

No measurements yet.

Empty-relative error · 1 = emptySimulation step →

Your experiment / Kept local

Keep a record.
Follow the evidence.

Save the complete state, random stream, and intervention history. Reopen it to continue the same experiment in this runtime.

Continue this experiment

A data-only snapshot, including both branches.

Take the measurements

Export the retained live measurements.

Model files & runtime

Models are supported data-only Mycelium files. A new image requires local training. Recreating the runtime discards unsaved state. Parity checks a frozen PyTorch reference; the runtime benchmark uses an independent clone.

No browser diagnostics run yet.

The evidence / Recorded native trials

Learning to grow.
Learning to remain.

These are saved evaluations, separate from your live experiment. The targets were known during training. Different rollout seeds are not independent training replications.

Loading recorded results…
Download raw summary ↓

Field guide / A learned local rule

The whole begins
with a neighbor.

Every cell carries sixteen numbers. It sees only its 3 × 3 neighborhood. The same tiny neural network updates every cell; together, they can grow a recognizable form.

16values per cell
48local features
64hidden units
16proposed updates
01 / Growth

Begin.

Train from one seed. Learn a local rule that produces a visible pattern.

02 / Persistence

Remain.

Continue from a pool of previous states. Learn to maintain the pattern over time.

03 / Regeneration

Return.

Damage pooled states during training. Learn a rule that may reconstruct missing tissue.

A rule, not a picture to copy.

During inference there are no coordinates, no target image, and no global view. The reference and live measurements help you inspect the result; they do not steer the rule. Erase every living cell and the culture stays extinct until you explicitly plant a seed.

All learning takes place locally in Python. Your browser runs the same NumPy transition in a persistent Pyodide worker. No inference service or API is involved.

Exact browser contract

32 × 32 cells, 16 float32 channels, identity and normalized Sobel perception, shared 48 → 64 → 16 update network. Hidden displays use a fixed scale. Pyodide 0.27.7, Python 3.12.7, NumPy 2.0.2. Exact replay is scoped to the pinned runtime; floating-point backend differences can grow over long rollouts.

Research foundation: Growing Neural Cellular Automata, Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, and Michael Levin (2020). Mycelium is an independent implementation and local training case study. It makes no claim of biological realism or guaranteed repair.