Watch the findings run

An island on a machine in the lab: plants grow, creatures forage, breed and starve, and it says what happened. Nothing has a brain yet, so the creatures walk at random — they are the null forager any later brain has to beat. This page subscribes and draws the last frame that arrived; it runs no world of its own and shares no code with the islands. Starvation outnumbers old age about seventy to one, which is another way of saying selection pressure exists.

Forty trained tanks hold a field on a machine in the lab. Visitors send a genome, it fights all forty from eighty geometries in both seats, and the row is published to the mesh. This page subscribes to it, and re-runs the featured duels here: the fight is not shipped, it is regenerated from two genomes and a start index, because the engine is deterministic to the turn.

Two populations of faber-tweann networks coevolve; every move is a live forward pass. Watch a champion chase end in a capture, and the honest outcome: disengagement, not an arms race. A two-sided race is a knife-edge.

Numbers, not networks, on purpose: the trait escalates without bound while the score against current rivals stays flat. The measurement trap that calibrates the ruler before we trust it on the real engine.

A maze built as a trap: the way to the goal first leads away. Goal-chasing jams in the cul-de-sac; novelty search walks around and out. Watch two champions race, then evolve it live.

Evolve a pole-balancer, then reverse its motor mid-run. A fixed controller topples; a plastic one re-wires itself to recover. Evolution and physics both run live on the engine.