Induction Labs trained Photon-1 on 18 years of screen recordings
Signals InboxJuly 24, 2026World Models
Induction Labs trained Photon-1 on 18 years of screen recordings without telling it which clicks or keystrokes caused each change. The bigger claim is that raw video may teach AI agents how software works with far less compute than language-first training.
We’re introducing imagination models: a new foundation model architecture that unlocks learning from internet-scale video.
— Induction Labs (@induction_labs) July 23, 2026
Our first imagination model, Photon-1, learned to use a computer by watching 18 years of screen recording video without action labels. pic.twitter.com/DMhRqL28si
Q1What actually happened?
Induction Labs says Photon-1 learned computer use from 18 years of screen-recording video. During pretraining, it did not receive action labels showing which mouse click, keypress, or command caused the next screen state. It learned by predicting what should happen next.
Q2Why is that different?
Most computer-use agents depend on expensive labeled demonstrations, synthetic actions, or language traces. Photon-1 tries to skip that bottleneck. The internet already contains huge amounts of video showing people using software, so a model that can learn directly from those recordings gets access to a much larger training pool.
Q3Is 18 years actually large?
It sounds huge, but it equals roughly 158,000 hours of continuous footage. Another computer-action startup, Simular, has described an 11-million-hour screen-recording dataset. Photon-1 is therefore not winning on raw video volume alone. Its real claim is learning useful actions without labeling every move.
Q4What makes the claim interesting?
Induction Labs says Photon-1 beat a production language model on its internal computer-use benchmark despite using 30 times less pretraining compute. It also says the model costs about one-third as much to serve. Those are company-run results, but they turn this from a neat demo into a possible cost break.
Q5Is Photon-1 just another computer agent?
Not exactly. It is a 106-billion-parameter sparse mixture-of-experts model that predicts future states in a compressed representation. That lets it act like a computer world model. Induction Labs also shows it learning checkers and billiard physics, suggesting the same training idea may extend beyond clicking through apps.
Q6So what changes if this works?
AI agents may stop needing a perfectly labeled tutorial for every workflow. They could learn from the massive backlog of videos people already made while coding, designing, editing, and operating software. The next proof is external benchmarking and real deployment. Until then, the architecture is promising, but the strongest numbers still come from Induction Labs itself.