A 129-parameter 2D classifier trained by federated Adam across whatever browser tabs you open. Workers compute ∇L on local batches in JS; the Cloudflare Durable Object averages and steps. Switch the task from the dropdown to retrain the same architecture on a new objective. Alternate protocols / models on this substrate: float tournament · ternary tournament · char-LM tournament · all-demos dashboard
Decision boundary, live.
Each pixel = the swarm's current θ run through the forward pass at that (x, y).
Color is the predicted probability of class 1. The thin black curve is the
ground truth: sin(2x) > y. The model should bend into that wavy
shape as the swarm converges.
x, y ∈ [−2, 2] · 320×320 grid · forward pass at every pixel each round
Open this page in more tabs (or send the URL to a friend) to add workers.
The coordinator (a Cloudflare Durable Object) averages every 2+ gradients
it gets and steps θ. Loss should plunge and the boundary should bend into the
wave within a few dozen rounds.