Neural Cellular Automata Sandbox
A neural cellular automaton gives every grid cell a small state vector (color, an alive channel, a few hidden channels) and updates all cells with the same tiny network: each cell perceives itself and its neighbours through identity, Sobel-x and Sobel-y filters, feeds that through a small dense layer, and adds the result to its state — stochastically, so only a random fraction of cells fires each step. This sandbox is honest about what it is: the alive channel follows a hand-designed local growth rule, so every pattern reliably grows from the seed and heals when you click to erase a region, while the color and hidden channels evolve under randomly initialised network weights. Randomize the weight seed to explore the space of dynamics — some settle into stable textures, some churn forever, some saturate — with live stability metrics, per-channel views, and sliders for update rate, fire rate and channel count.
Runs 100% in your browser — models are trained and computed locally on your device.
Read the full guide to this tool
Notes
- The perception stage is a fixed depthwise convolution (identity + two Sobel kernels per channel); only the dense layer after it differs between weight seeds — the same architecture Mordvintsev et al. trained in "Growing Neural Cellular Automata" (2020).
- In the original work the weights are *trained* with backpropagation-through-time to grow a specific image and regrow it after damage. Training is far too slow for a browser page, so here the life channel is hand-designed (logistic neighbour growth) and the rest is an explorable random network — a sandbox of NCA dynamics, not a trained organism.
- Stochastic updating (the fire rate) stands in for asynchronous biological cells: because no global clock synchronises updates, any pattern that persists must be robust to timing noise.
- Alive masking means a cell with no living neighbour is wiped to zero — growth can only advance along a living frontier, which is exactly why erased regions heal from their edges inward.
- Runs 100% in your browser — models are trained and computed locally on your device.
Related tools
Diffusion Model Toy
Train a tiny denoising diffusion model and watch it turn pure noise back into a 2D shape.
GAN Trainer (2D)
Pit a tiny generator against a discriminator on 2D toy data and watch mode collapse happen live.
Markov Chain Text Generator
Train an order-k Markov chain on any text and generate new text one token at a time.
AI Learns to Drive (Neuroevolution)
Watch a population of neural-network cars evolve from crashing into corners to lapping a race track.
Backpropagation Step-by-Step
Walk through one forward and backward pass of a tiny neural network, number by number.
Decision Tree Builder
Grow a CART decision tree split by split and watch the partition map and tree diagram build together.