Induction Labs unveils Photon-1, a video 'imagination model' that learns computer use without action labels
AI video model releases
Induction Labs unveils Photon-1, a video 'imagination model' that learns computer use without action labels
Induction Labs introduces imagination models, foundation models that predict future video frames in latent space and implicitly learn actions from unlabeled video during pretraining.
Photon-1, a sparse 106B-A5B MoE transformer trained on 18 years of computer demonstration video, beats a production LLM trained with 30x more FLOPs on internal computer use benchmarks while costing 3x less to serve.
The vision encoder uses finite scalar quantization (FSQ) to compress each frame into 960 tokens of 8 dimensions with 5 possible values each, totaling 2.2KB per frame, claimed to be 100x better compression than existing OCR or multimodal representations.
A differential latent encoder encodes pairs of frames to capture differences rather than raw frame contents, supporting the high compression rate.
Despite training only on computer use video, Photon-1 generalizes to play checkers after finetuning on 20,000 tournament games and to simulate billiard ball physics, beating LLMs finetuned on the same data; the pretraining corpus was filtered from 2 billion public videos down to about 2 million computer screen recordings.
cho.sh