A 32M-parameter discrete diffusion transformer that generates monophonic melodies, conditioned through cross-attention on gestures tracked from a webcam. Trained on more than 10 million four-measure excerpts.
Heads-up no-limit hold'em against an AI that starts from an equilibrium strategy learned by self-play, models its opponent's habits as it plays, and uses a reinforcement-learned policy to exploit them. Runs entirely in the browser.
