A slime mould on the front page

Physarum polycephalum is a slime mould. It lives on rotting wood and dead leaves as a single huge cell with many nuclei. It has no brain and no nervous system. Computer scientists like it because it solves mazes. Put oat flakes at two exits and within a day the mould has found the shortest path between them. It does this by growing in every direction and then pulling back from everywhere that did not lead to food.

That is also what the simulation on the front page does, more or less.

How the simulation works

The model comes from a 2010 paper by Jeff Jones on Physarum transport networks. There are no cells in it, only agents. Up to 260,000 of them, each with a position and a heading, stored as four bytes in a texture so the GPU can update all of them in one pass.

Every frame, each agent does the same three things. It samples the trail field at three points in front of it, left, centre and right. It turns towards the strongest sample, or randomly if there is no clear winner. Then it moves one step and deposits a bit of trail where it lands. The trail field blurs and fades a little every frame. A path that nobody uses is gone in seconds. A path that keeps getting used gets stronger.

sense left, centre, right
if centre is strongest:  keep going
if centre is weakest:    turn at random
otherwise:               turn towards the stronger side
move, deposit

Nothing else is programmed. The veins, the loops, the fronts and the slow pulsing all come out of this loop. The five regimes in the lab are five settings of sensor distance, sensor angle, turn speed, step size, noise and decay. Transport gives thick veins between food sources. Swarm makes the agents orbit each other. Lattice looks like a crystal growing. The code is the same in all of them.

Moving the pointer adds a weak attractor. Clicking plants food, a patch of trail that does not decay, so the colony grows towards it and connects it to the rest. It runs in WebGL, pauses when it scrolls out of view or when your system asks for reduced motion, and changes ink when you switch between light and dark.

Why it is here

A Physarum colony allocates effort without anyone in charge. It spends its body on routes that lead somewhere and takes it back from routes that do not. My research asks whether language models can do something similar with their reasoning, spend more on the questions where it changes the answer and less where it does not. The slime mould is not a model of that. It is a reminder that nature solved a version of the problem with local rules and feedback, and without a bigger brain.

Also, I like watching it.