GitHub teaches a plush squirrel to throw itself, scoring 86% live

On the GitHub Universe stage, Martin Woodward stood behind the podium. On the screen was an app called Action Trainer, with a status bar reading Ready and three tabs below—Walk, Throw, Drop—while several progress bars each showed percentages for different actions.

His demo subject was a plush squirrel. He used a machine learning model to drive it through actions, training it live and checking the results live. At the moment the squirrel was thrown, the number on the screen stopped at 86%.

This 86% is the model's confidence in the current action. One commenter said in a reply that 86% was more accurate than most answers they could guess.

The interesting part of this demo isn't the technology itself. The Action Trainer interface is plain: a few tabs and progress bars, like some kind of experimental tool. Woodward brought it onto the main stage and spent a few minutes making a plush toy move. The audience watched for entertainment, but underneath ran a complete action-training workflow: choose an action, provide data, view the model output, and adjust parameters.

GitHub Universe has segments like this every year. They take something abstract from developer tools and turn it into a scene people can watch, laugh at, and comment on. The plush squirrel is a good prop: it pulls the term "machine learning model" off the slides and lands it on something furry that can be thrown.

As for what the squirrel itself thinks, it has no say. It is only responsible for providing 86% confidence and a round of laughter from the audience.