01 / CELLULAR MEMORYTRAINED CHECKPOINT
Memory, cell by cell.
Write a fresh association. Follow a query around a ring. Discover what disappears when private memory is erased.
Explore cellular memoryIndependent experiments · Open implementation
Small rules.
Emergent
possibilities.
Exploring how local neural systems remember, communicate, and generate language. A research division of F2i.
Two models. Explicit controls. No oversized claims.
01 / LOCAL NEURAL SYSTEMSSTATE ≠ WEIGHTS
SHARED RULE · PRIVATE STATECONCEPTUAL DIAGRAM
02Independent research models
385Shared parameters in the memory model
17 bytesMaximum context of the language checkpoint
Properties of the supplied research checkpoints, not product benchmarks. Read the evidence and limitations →
The research programme
Different questions.
One local perspective.
We study what can be built from shared rules and distributed state—and which mechanisms actually explain the result.
02 / CAUSAL LANGUAGESEPARATE MODEL
A local rule for the next byte.
An attention-free model reuses a small neural update to generate template text, one byte at a time.
Explore causal languageA field guide to the idea
What if the system
grew from its rules?
Neural cellular automata replace a hand-written local update with a learned neural rule. Many cells share the same parameters, while each keeps its own state.
Understand neural cellular automata →01
Local information
Cells read a neighbourhood, not an unrestricted global view. Geometry determines what can travel where.
02
Shared rules, different states
A small neural rule can be reused across cells and updates. Shared parameters and private memory are not the same thing.
03
Evidence through intervention
Erase memory. Block communication. Check causality. An explanation should survive a controlled test.
From explanation to experiment
Don’t just read the rule.
Run it.
Inspect trained checkpoint state in your browser. Python runs through ZIPP WebAssembly; bounded compute graphs use WebGL2 or an explicitly selected CPU reference.
Enter Neural Cellular Automata Lab// an inspectable execution pathPython experiment ↓ ZIPP / WebAssembly ↓ validated float32 graphWebGL2 · local compute ↓ actual model stateYour browser. Your experiment.
Before you dive in
Curiosity, with context.
Is this a self-learning chatbot?
No. The current work contains a structured associative-memory model and a separate small causal byte model. The memory experiment changes private fast state during inference; neither browser demo updates its shared trained parameters.
Does the language model understand general language?
That has not been demonstrated. Its checkpoint was trained on original template text and has a 17-byte context. It is a controlled architecture example, not a general-purpose assistant.
Can I reproduce the results?
The source repository contains the native models, training commands, checkpoints and recorded results. The browser lab adds an inference port with numerical reference tests. Choose a reproduction path.
How does this connect to F2i?
Neural Automata is F2i’s research division. The interactive lab uses ZIPP, part of F2i’s broader software ecosystem. Explore F2i.