In particular, they talk about the "emergent movement" behavior in their "organisms," and how surprising this is given they weren't trained or incentivized to move.
Calling this "movement" is intuitive, but misleading, since what they're describing is a field of pixels that do not move, but change color. The image produced by the NCA appears to move, because each pixel is negotiating with its neighbors according to a fixed rule set, trying to minimize conflict between them. There is no "organism" to issue top-down commands, or learn from training.
The cool part of their paper is they got an NCA to "notice" when two patterns are colliding with each other, and negotiate that kind of conflict to minimize disruption. But they didn't actually analyze at the ruleset level how / why their tweak caused "emergent movement". Instead, they speculate about what this could teach us about social interaction between organisms, and emergent self-awareness in artificial life. Which... just... no?
I'm fascinated by Neural Cellular Automata, and am considering using them in my research. Cellular automata (like the Game of Life) are pretty cool already, but NCAs allow you to work in a continuous space, and easily discover rule sets for generating whatever patterns you choose! They also make for an interesting model of a developing organism. An NCA "grows" from a single pixel into an image, with surrounding pixels recruited to be part of the "body" and negotiating their identity with their neighbors. The classic demo is still available here.
One problem with this, though, is that sometimes people take that metaphor too seriously. Take this paper, for instance. It's pretty good, actually, but talking about their work as if they're watching organisms behave is the wrong metaphor. It exaggerates their findings and makes them harder to interpret.
@timhutton Hey Tim! I built a 2026 fork of your Squirm3 that keeps the chemistry rules unchanged but modernizes the substrate: WebGPU rendering, 5000×3000 arena, water as continuum heterogeneous-Brownian regions, seedable RNG, save/load.
In a single seeded run I observed a closed a-a-a membrane loop spontaneously form around a cluster of intact protocells, then accumulate more cells over time:
1. iter 117,632 - outer loop, 1 inner protocell
2. iter 140,920 - 69-atom outer loop, 2 inner protocells
3. iter 157,068 - 78-atom outer loop, 4 inner protocells
Verified from the bond graph, not just visually, I also made an analysis script in the repo that walks the graph and runs point-in-polygon containment. The outer ring is ~3x the size of a typical single-cell membrane in the same run.
Your polymerization rule is scale-agnostic, so this is consistent with your chemistry, but I don't think meta-membrane formation appears in your 2002 or 2007 papers. Did you ever observe it in your own runs?
Three save files + the script are in the repo so it's reproducible. Load any save in the demo and you boot directly into that state.
Demo: https://davidortsac.github.io/primordium/
Saves + script: https://github.com/DavidOrtsac/primordium/tree/main/samples
Repo: https://github.com/DavidOrtsac/primordium
Big fan of your work!
#CellularAutomata #ArtificialLife #ALife #ReactionDiffusion #ArtificialChemistry