Tabula Zero is a chess engine that learns by playing itself. Playing here is by invitation: sign in with the address you were invited with, and a one-time link comes by mail. There is no password.
This is a sign-in link for Tabula Zero. It works once, for ten minutes.
What the engine thinks
| Move | Visits | Network's prior | Value Q |
|---|
Before searching, the network proposes a probability for every legal move (its prior). The tree search then spends its simulations where prior and measured value Q together look most promising, and plays the move it visited most.
Training
Quick lesson: training on the games you played, half of every batch from them, with no new self-play and no match. Your next game plays the result, even if it got worse. One generation: the engine plays itself, trains with a tenth of every batch from your games, then plays the network it would replace, and replaces it unless it scores below 50%. Keep training repeats generations until you stop it; games you finish meanwhile join the next one. Training runs at the lowest priority, so the engine keeps its full strength in a game you play meanwhile.
| Gen | Games | Positions | Loss | Score | Elo | Promoted | Your positions |
|---|
Lessons
| # | Generations before | Steps | Loss first → last | Your share | Draws per position | Seconds |
|---|
People
Only the people listed here can sign in. Invite someone by address (they get a mail saying so) or make an invite link that works once, for one person, for seven days. Guests play on tables of their own, with a lighter search; their games are kept apart from yours and do not train the engine.
| Address | Role | Since |
|---|
Invite links
| Link | Expires |
|---|
Guests at their tables
| Guest | State | Moves | Last move |
|---|
Game archive
| # | Result | Reason | Plies | Yours |
|---|