Scaling
How fast does your computer learn?
A 30-second speed test made of real self-play training: little AI agents learning hide and seek and heists by PPO, on every core of your machine. Nothing to install. Then see where your machine ranks.
0
agent-steps / second
Takes about 30 seconds. Keep this tab in front while it runs.
Inside your CPU
Throughput by teststeps / s
Your runs here
Under the hood
| Metric | Value |
|---|
The leaderboard.
Each machine's best run. Press GO, then put yours up.
Sends this name, your score, core count, chip and browser, and the country your connection is in. Nothing else.
| # | Name | Steps/s | Machine | When |
|---|
Not a synthetic loop. Actual learning.
Every number here comes from agents getting better at a game, the same code that trains them when you press Train.
01 / SELF-PLAY
Two teams teach each other
Hiders and seekers, thieves and guards. Each team is a small neural net; every game they play becomes training data for both.
02 / EVERY CORE
One training run per core
Each core plays a batch of games, then runs PPO on them: three passes of forward and backward through policy and value nets, with Adam.
03 / PRIVATE
Nothing leaves unless you say
No account, no GPU tricks, and only what you choose to put on the leaderboard is sent. Plain JavaScript in Web Workers, so phones, laptops and workstations all run the same thing.
Now watch them learn.
The benchmark is borrowed from two games. Press Train and see the strategies appear.