TL;DR: We're Induction Labs, and today we're announcing research we've been working on: imagination models, a foundation model architecture that unlocks scalable learning from internet video. Our first model, Photon-1, beats a production LLM on internal computer use benchmarks with 30x less training compute, at 3x lower serving cost. [Main Article]

Internet video contains millions of hours of people using computers, doing skilled work, and interacting with the world.

Imagination models can scalably learn from this video to get better at completing tasks and understanding the world.

We tested the architecture with Photon-1, a 106B-A5B MoE transformer pretrained on 18 years of computer screen recordings.

Some Results:

• Outperforms a production LLM (Gemini 3.1 Flash-Lite) on our internal computer use benchmarks, with 30x less pretraining compute and 3x cheaper inference.

• Simulates full desktop sessions from a single screenshot: it can imagine VS Code, Gmail, and ChatGPT frame by frame
• Generalizes beyond computers: finetuned, it learns checkers and simulates billiard physics better than LLM baselines.
• Picks up human behavior from video, like reprompting ChatGPT until it gets what it wants.
• Demos here

We see imagination models as a path to intelligence that learns by observing the world directly, without a human first translating it into text.

Our Ask

We’re working to scale this method to more kinds of video and looking for exceptional people to work with us! If you’re an exceptional thinker, engineer, or researcher (or know anyone who would be a great fit) - shoot us a message at team@inductionlabs.com