Induction Labs Introduces Photon-1 Desktop and Physics Simulation AI
Induction Labs unveils Photon-1, a 106B-A5B Mixture-of-Experts model trained on 18 years of computer demonstration videos capable of simulating desktops, playing checkers, and modeling billiards.
Stock photo for illustration only, not from the actual event
- Photon-1 is a 106B-A5B MoE transformer trained exclusively on computer screen recording videos.
- Uses a differential latent encoder achieving a high compression rate of roughly 2.2 KB per frame.
- Outperforms Gemini 3.1 Flash-Lite on internal computer use benchmarks at a significantly lower serving cost.
- Demonstrates world simulation capabilities like playing checkers and modeling pool physics from screen data alone.
Induction Labs has released Photon-1, a sparse 106B-A5B mixture-of-experts (MoE) transformer trained on an extensive corpus of 18 years of computer demonstration video. In internal computer use benchmarks, the research team reports that Photon-1 outperforms Gemini 3.1 Flash-Lite while requiring far less pretraining compute and costing approximately three times less to serve to users.
At the core of the system is an imagination model that predicts future frames autoregressively through a next-latent-token-prediction objective. The architecture models everything within a learned representation space rather than generating pixels directly during pretraining. Induction Labs refers to this mechanism as an implicit policy, where the model grasps the underlying concepts of human actions without needing explicit labels for every individual mouse click.
The architecture relies on a vision encoder utilizing finite scalar quantization (FSQ). Each video frame is compressed into 960 discrete tokens represented as 8-dimensional vectors where each dimension takes one of five values: -1, -1/2, 0, 1/2, and 1, resulting in a codebook of 5 to the 8th power possible codes. This encoding yields roughly 2.2 KB per frame, which Induction Labs reports is over 100 times better compression than existing OCR and multimodal representations while perfectly preserving text, layout, and state transitions.
Stock photo for illustration only, not from the actual event
The training corpus originated from an internal index of 2 billion public videos, filtered down to about 2 million computer screen recordings. After removing redundant frames via an internal keyframe detection model, the final dataset comprises 575 million frames sampled at 1 frame per second, translating to 552 billion tokens or 18 years of video. Photon-1 was pretrained from scratch for a single epoch utilizing 30,000 H200 GPU-hours at a 32K context length, equating to 4.4 x 10 to the 22nd power training FLOPs.
"Predicting future states teaches the model to complete tasks, even though it never sees an action during pretraining."
Induction Labs
The ability of Photon-1 to generalize from screen recordings to complex physical simulations like billiards and strategic board games highlights a major breakthrough in self-supervised video modeling. By observing standard computer usage interfaces, the model inherently extracts physical rules and logical frameworks without needing specialized simulation training datasets.
When evaluated on out-of-domain tasks, Photon-1 excelled against two baselines: a vision encoder baseline without imagination pretraining and an LLM baseline using Ling-flash-2.0. Across 20,000 tournament checkers games and 10,000 simulated billiard games, Photon-1 demonstrated superior world simulation and move quality. Furthermore, the model achieves a breakeven cost on Induction Labs hardware of $0.06 per million input tokens and $0.60 per million output tokens.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment