One spoken command in, every step routed to the right app and skill.
A 395M-parameter local model that splits a command into steps, routes each step to the right part of an assistant and picks the skill, with a probability for every choice. It is a fine-tune of Von by Victor Hugo, and a Mac app lets you fine-tune it on your own tools.
Hear a guitar tone on a record, get the amp settings that make you sound like it.
A 3.4 MB tone encoder listens to the song and guesses seven amp knobs, then a CMA-ES search refines them against the real audio. Measured on real mic'd amps with tonebench, its benchmark.
Reads a play off the film, then predicts the call from the pre-snap look.
It finds, tracks and labels all 22 players: teams by jersey colour, the field in real yards, the line of scrimmage and the snap. From that alignment it reads the formation, strength, backfield and box, and an XGBoost model you train on your own tagged snaps predicts run or pass and the run direction.
Issues and pull requests are welcome on GitHub.
Lab