FlyTris.
A tiny player at the controls.
A new player at the controls.
3.19 hours trained · 63 generations.
Mixed evidence: 24 paired games across eight independent sequences. Mean gain 89.3 lines; 95% interval -45.0 to 205.7. Some sequences regress, and the interval includes zero.
The replay and live opponent now use falling ticks, midair rotations and legal slides. Replay: first test seed at medium gravity.
The scorebook.
Two-minute, server-verified ranked matches. Score = lines cleared; pieces placed break ties.
Human vs. fly
Humans 0 lines
Fly 0 lines
No completed matches yet.Difficulty defaults control the fly pace. Only games played with those defaults enter the current scoreboard.
EasyDefault difficulty settings
| Player / date | Human | Fly |
|---|
MediumDefault difficulty settings
| Player / date | Human | Fly |
|---|
HardDefault difficulty settings
| Player / date | Human | Fly |
|---|
Top 10 human scores per difficulty, paired with the fly’s score from the same match. Dates use your local time. Names are player-provided.
Original placement experiments.
| Run | Before | After | Gain [95% CI] | Training |
|---|---|---|---|---|
| 11 | 0.12 | 40.71 | +40.58 [33.33, 48.62] | 1.1 min |
| 22 | 7.46 | 103.83 | +96.38 [86.75, 103.88] | 1.4 min |
| 33 | 0.08 | 68.58 | +68.50 [58.33, 78.50] | 1.3 min |
Mean lines cleared on identical, unseen piece sequences. These historical trials used placement rules; their results are separate from the current falling-rules checkpoint.
What actually learns?
Board features enter a fixed fly-derived spiking circuit. Eight learned output weights choose a placement from its activity.
The fly animation illustrates recorded decisions. It is not a biomechanical simulation. Removing recurrent connections did not consistently hurt performance in these trials, so fly wiring has not been shown to be essential.