Genesis Deep Dive

The same nine lessons, taught from the application

Project Genesis teaches a language model at the command line. This is the identical course with InferenceGenesis on screen — the corpus, the alphabet, the split, the architecture, a live training run and a page of Shakespeare it wrote itself. Every close-up is a real recording of the real program, and every number on screen is one it produced.

9 lessons 13 minutes 1,115,394 characters of Shakespeare 3,225,665 numbers trained

Pick any lesson below. It opens here, over the page.

The nine lessons

How these were made

Each lesson was filmed against the running program rather than assembled from stills. The interface at full size is unreadable on video, so the camera travels to a named region of it, holds while the narration explains that region, and pulls back. Those regions are named in the application's own markup, and the length of every hold is computed from the measured narration of the beat it backs — so nothing is cut off mid-sentence and nothing freezes waiting for the narrator.

Lesson eight is a real training run: twelve hundred steps, on the machine doing the filming, with the loss curve drawing as it happens. InferenceGenesis is a prototype the Webspinner Foundation is releasing later this year.