Can a Fly Learn the World with Causal‑JEPA?
We built a 4,096-neuron fly brain from connectome statistics, put it in a virtual dish with wind, an odour plume and a hunting spider, and asked whether a self-supervised world model could learn how that world works — and reuse it in worlds it had never seen. It can, but only once something splits the brain's activity into entities. Finding that split without labels is still open.
Read the post