Neural Cellular Automata

"Traditional AI works like a corporate hierarchy - every decision must travel up to the CEO (the cloud server) to get a response. The NCA stack works like an ant colony - individual ants only interact with the ants right next to them, yet through local communication, the entire colony effortlessly builds complex structures and reacts to threats instantly."

To understand NCAs, we first need to look at how biological entities grow. Consider a human body: it starts as a single egg cell. That cell splits into billions of others. There is no "master brain" cell directing where the heart goes or how the skin shapes. Instead, every single cell contains the exact same DNA blueprint. Each cell survives by looking at its immediate neighbors, exchanging chemical signals, and executing its local genetic instructions. Together, this completely decentralized system builds a complex organism.

An NCA is a computational model that mimics this biological phenomenon. It merges classical, grid-based simulations with modern deep learning.

Cellular Automata (CA)

A cellular automaton is a computational model consisting of a grid of cells. Each cell has a state (e.g., 0/1, alive/dead), and the state evolves over time based on simple rules applied to neighboring cells. Each cell can only see the state of neighboring 8 cells.

Key components:

Grid: 1D, 2D (like a chessboard), or higher dimensions
Cells: Each holds a value/state
Neighborhood: Nearby cells influencing a cell (e.g., adjacent 8 cells in 2D)
Rules: Fixed update rules applied simultaneously to all cells at each time step

A great example of CA is

Conway’s game of life

Neural Cellular Automata (NCA)

A Neural Cellular Automaton is a modern extension of cellular automata where the update rules are learned by a neural network instead of being manually defined.