Induction Labs Photon-1 Simulates Desktops, Plays Checkers, and Models Billiard Physics From One Pretraining Run

Every agent that's learned to click buttons by watching video has needed someone to label what each click did. Induction Labs just published a model that skips that step entirely, and the result, Photon-1, is one of the stranger training recipes I've seen in a while. Instead of learning from action-labeled demonstrations, it just watches. A lot. Roughly 575 million frames, pulled from an index of 2 billion public videos and filtered down to 2 million computer screen recordings, sampled at one frame per second. That's 552 billion tokens, or about 18 years of continuous footage, fed through a single training epoch.

The trick is what Induction Labs calls an imagination model: it predicts future frames in a compressed latent space rather than generating pixels, and it never sees an action label during pretraining. The bet is that if you get good enough at predicting what a screen will look like next, you've implicitly learned what the person operating it was trying to do. They call this an implicit policy, and to make prediction cheap enough to do at scale, each frame gets squeezed into 960 tokens using finite scalar quantization — an 8-dimensional code with five possible values per dimension, landing around 2.2 KB per frame. Induction Labs claims that's over 100 times more compact than typical OCR or multimodal encodings, while still holding onto text, layout and state changes.

Photon-1 itself is a sparse 106B-parameter MoE with 5B active parameters, trained from scratch for about 30,000 H200 GPU-hours — roughly 4.4×10²² FLOPs. After pretraining, the team finetuned it on fewer than 35,000 labeled computer-use trajectories to teach it an action vocabulary, then ran online reinforcement learning across virtual machines spanning five different Linux desktop environments, with programmatic reward checking. On an internal, unreleased computer-use benchmark, Induction Labs says Photon-1 outperforms Gemini 3.1 Flash-Lite while needing about 27 times less pretraining compute and costing roughly a third as much to serve per million tokens — though that Gemini comparison rests on Induction Labs' own estimate of Gemini's size and training data, not a disclosed figure from Google.

The more convincing part, honestly, is what happens when Photon-1 leaves the desktop. It was never shown checkers or billiards during pretraining, yet after finetuning it beat both a matched vision-encoder baseline and a 20T-token LLM baseline (Ling-flash-2.0) at simulating checkers moves and predicting billiard ball physics, hitting a mean error of 0.47 against a ground-truth physics engine versus 1.15 and 1.44 for the baselines. It also picked up human habits from its training video, like using an in-VM chatbot to draft documents instead of writing them itself. That's a real signal that watching video without action labels teaches something closer to general world modeling than a narrow desktop-clicking skill.

None of this is independently verifiable yet. There's no released model, no API, no public benchmark to poke at, and the headline comparison against Gemini is Induction Labs grading its own homework against an estimate of a rival's training run. Still, the architecture idea — that action labels were never necessary, just a crutch — is the kind of claim that, if it holds up under outside scrutiny, could reshape how agent training gets done.