Think about this: in a cellular automaton, all the cells follow one rule and perform certain executions. In our neural cellular automaton model, a neural network tries thousands of times to figure out the rule to generate a particular pattern. The cells don't know about the pattern - they only care about their own states and their neightbours', however, together they output sophisticated visuals that make sense for viewers like us, who look at it from a wider perspective.

Using cellular automata to simulate the complex social dynamics is not a new thing. A Plea For Cellular Automata Based Modelling is an example that demonstrates low-dimensional (and especially two-dimensional) CA are a promising modelling approach for understanding social dynamics. People have been working on this topic for decades, solved lots of social issues with CA modelling.

With the introduction of neural networks, the CA aqcuire the ability to generate given pattern stably and even transform from one to another. And in the Chinese language, each word has its own meaning, and a connection of a few characters can form phrases that make very good sense to people. What interesting features can these bring to the CA? If we can directly use language to represent the states of the cells, will people pay more attention to the details (the states of each cell), or the greater picture?