This pen's ink stretches backwards into the past and forwards into possible futures. The two sides make a strange loop: the future ink influences how you draw, which in turn becomes the new "past" ink influencing further future ink.

Put another way: this is a realtime implementation of SketchRNN which predicts future strokes while you draw.

Play with it here! (You'll need to use Chrome for now.)

Let it guide your hand, or not, while you draw. Enjoy the gentle interplay between your volition and the machine's.

[I'll be writing more about the background of this project before talking about it broadly; a brief sketch follows:]

I'm excited about applications of machine learning to human augmentation—as opposed to automating tedious work or solving number-crunching problems. I believe effective media for human augmentation require feedback loops tight enough to work at the speed of thought. All that is to say: I'm interested in what happens when we take complex operations and make them interactive in real-time.

I'm grateful to David Ha, Douglas Eck, and their team for their great work with SketchRNN. The paper's wonderful to read; the models from Quick, Draw! are a great resource to the community; the demo implementations were a very helpful reference.

I'm also grateful to Michael Nielsen, Robert Ochshorn, and M Eifler for useful conversations about this project.

Thanks also to the OpenAI Hackathon for providing a nice venue for polishing the last bits of this project up.