Editorial illustration for Induction Labs' Photon-1 Model Encodes Video Frames at 2.2 KB
Photon-1 Compresses Video to 2.2 KB Per Frame
Induction Labs' Photon-1 Model Encodes Video Frames at 2.2 KB
Induction Labs put a number on video compression last week that's worth sitting with: 2.2 kilobytes per frame. That's the footprint of Photon-1, a 106-billion-parameter mixture-of-experts model the company trained on 18 years of screen-recording footage, all of it unlabeled. No action tags, no mouse-click annotations, nothing telling the model what caused what. Just raw video.
That's the bet behind what Induction Labs calls an "imagination model." Instead of learning to map actions to outcomes, Photon-1 learns to predict what a screen will look like next, frame after frame, in a compressed latent space rather than pixels. The company says this alone is enough to produce something that behaves like a policy: the model appears to pick up what a person is trying to do just from watching sequences of frames, without ever being told which click or keystroke did it.
On an internal computer-use benchmark, Induction Labs reports that Photon-1 outperforms Gemini 3.1 Flash-Lite, while needing less pretraining compute and costing about a third as much to run. The architecture behind that result starts with how the model squeezes a frame down in the first place.
The claim that matters is this: predicting future states teaches the model to complete tasks, even though it never sees an action during pretraining. Induction Labs calls this an implicit policy.
Why this matters
The pitch here is that action labels, long treated as a prerequisite for video-based agents, might just be a crutch. If Photon-1 can pick up billiard physics and desktop navigation from raw footage alone, that changes the calculus for anyone building computer-use agents: the bottleneck shifts from "how do we label enough demonstrations" to "how do we get enough unlabeled video," and there's a lot more of the latter sitting around. The 2.2 KB-per-frame compression claim is the more immediately useful number for engineers, since anything approaching 100x over OCR-based pipelines would make storing and streaming long video context for agent training dramatically cheaper, if the comparison holds up outside Induction Labs' own benchmarks.
We'd want to see this tested on tasks further from checkers and billiards before trusting the "pretraining without labels" framing generally. A 106B-parameter MoE trained on 18 years of screen recordings is also a reminder that "no action labels" doesn't mean "no expensive data pipeline." Worth watching whether independent teams can reproduce the compression numbers, and whether this architecture generalizes past desktop-style demonstrations into messier real-world video.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives OpenAI Flagged GPT-5 as High-Risk After Users Got Poison Recipes
From Archives Grok Build CLI Excels at Greenfield Coding, Testing Reveals
Common Questions Answered
What is the compression rate achieved by Induction Labs' Photon-1 model?
Photon-1 encodes video frames at 2.2 kilobytes per frame, which represents a significant compression achievement for video data. This footprint was achieved by the 106-billion-parameter mixture-of-experts model trained on 18 years of unlabeled screen-recording footage.
How was Photon-1 trained differently from traditional video-based AI models?
Photon-1 was trained on raw, unlabeled video footage without action tags, mouse-click annotations, or any information about what caused what to happen. This approach contrasts with traditional methods that require explicit action-outcome mappings, allowing the model to learn from pure visual data alone.
What is an implicit policy according to Induction Labs' research?
An implicit policy is Induction Labs' term for how Photon-1 learns to complete tasks by predicting future states, even though it never sees any actions during pretraining. This demonstrates that the model can develop task-completion capabilities purely through learning to forecast what will happen next in video sequences.
Why does Photon-1's training approach potentially change how computer-use agents are built?
By proving that models can learn desktop navigation and physics understanding from unlabeled video alone, Photon-1 shifts the bottleneck from needing large amounts of labeled demonstrations to simply requiring more unlabeled video footage. Since unlabeled video is far more abundant and easier to obtain than carefully annotated action data, this makes building computer-use agents more practical and scalable.
What tasks can Photon-1 perform based on its pretraining on screen recordings?
Photon-1 can simulate desktops, play checkers, and model billiard physics, all learned from a single pretraining run on unlabeled screen-recording data. These diverse capabilities demonstrate that the model has developed a general understanding of visual dynamics and spatial reasoning without explicit task-specific training.
Further Reading
- Photon-1 Learned To Use A Computer By Watching Video. Here's What That Actually Means. - B2BNN
- Controllable Generative Video Compression - arXiv
- PruneVid: Visual Token Pruning for Efficient Video Large Language Models - Project page
- Penguin-VL: Exploring the Efficiency Limits of VLM with Dense Fusion - arXiv
- Generative Compression for Video: What to Know in 2026 - Softorino
