You train a real model. Click "Join federation"
and your tab runs ~2 SGD steps each round on a JEPA world
model — same forward graph the published research uses.
Nothing fake; the loss really decreases.
2
Your contribution is graded. The server
validates each delta against held-out data BEFORE merging it.
Helpful contributions raise your reputation; junk gets
rejected and slashed. Leaderboard ranks by reputation, not
volume.
3
Cost: tab CPU. Each round is ~30 seconds of
WebGL compute. You can leave the tab open in the background;
no audio, no popups, no third-party trackers. Stop any time
with the Stop button.
A new game generation has been deployed.
AURA Federation
Real TF.js training of a JEPA world model, federated across
browser tabs. Click Join federation (REAL TF.js) in two or
more tabs and watch the val_loss curve descend.
Status
idle
Client ID
—
Current round
—
Last val_loss
—
Δ vs baseline
—
Rounds contributed
0
one-shot diagnostic: load v9, run 2 SGD steps, log loss + delta L2
Leaderboard — by reputation (improvement × decay − slash)