Induction Labs unveiled a research result that challenges one of the longest-standing assumptions in AI: that teaching machines to act requires meticulously labeled examples of every action they should take. The company introduced Photon-1, the first of what it calls “imagination models”—a new class of foundation architecture designed to learn from internet-scale video without ever seeing an action label during pretraining.
Photon-1 is a sparse 106-billion-parameter mixture-of-experts transformer with 5 billion active parameters, trained on approximately 575 million frames of computer screen recordings—equivalent to 18 years of video sampled at one frame per second. The dataset was distilled from an initial index of two billion publicly available videos, filtered down to roughly two million screen recordings and stripped of redundant frames by an internal keyframe detection model. The model was pretrained from scratch for a single epoch, requiring approximately 30,000 NVIDIA H200 GPU-hours and 4.4 × 10²² FLOPs.
The central claim is striking. According to Induction Labs, Photon-1 outperforms Google’s Gemini 3.1 Flash-Lite on internal computer-use benchmarks despite having been trained on at least 30 times less compute, while costing roughly three times less per million tokens to serve. The company reports a weighted inference cost of $0.11 per million tokens compared to Gemini’s $0.36. These figures, however, come with important caveats: the benchmark is internal and unreleased, meaning the results are not independently reproducible, and the Gemini compute estimate is Induction Labs’ own conservative projection rather than verified data.
What distinguishes Photon-1 from conventional approaches is not merely scale but architecture. Rather than predicting the next word or generating raw pixels, imagination models predict future frames autoregressively in a learned representation space using a next-latent-token objective. Each video frame is compressed into 960 discrete tokens via finite scalar quantization, occupying just 2.2 kilobytes—roughly 100 times smaller than existing OCR and multimodal representations—while preserving text, layout, and state changes. A differential latent encoder processes frames as pairs, encoding differences between consecutive states rather than absolute frame contents.
This compression makes autoregressive prediction of future states computationally practical at scale. During pretraining, the model learns what the company terms an “implicit policy”: by predicting what a screen will look like next, it internalizes the causal structure of computer interfaces without anyone telling it which mouse clicks or keystrokes produced the transitions.
Turning this observational knowledge into a functioning agent required a second stage. Induction Labs finetuned Photon-1 on fewer than 35,000 labeled computer-use trajectories to teach it the correct action format, adding special tokens that allow the model to emit keyboard and mouse commands. At inference, the system operates in two steps: it first imagines the next state that would advance the task, then generates the action intended to reach that state. Online reinforcement learning followed, with real-time rollouts on Linux virtual machines across five desktop environments, programmatically verified outcomes, and reward signals driving further improvement.
Perhaps more surprisingly, the model’s capabilities appear to extend beyond the domain it observed. When finetuned on 20,000 tournament checkers games, Photon-1 outperformed both a vision-encoder baseline and a similarly sized language model baseline on both world simulation and move quality. On 10,000 synthetically generated billiard games, it achieved a mean absolute error of 0.47 in ball-position prediction, compared to 1.15 for the LLM baseline and 1.44 for the vision baseline. The model also picked up human behavioral patterns from its pretraining data, learning to prompt an in-virtual-machine ChatGPT clone, check its outputs, and steer the conversation until the task was complete—mimicking the way people actually use AI tools.
Photon-1 arrives at a moment when the entire AI industry is pivoting toward what researchers broadly call “world models”—systems that build internal representations of how environments evolve in response to actions, rather than merely predicting text or generating isolated video clips.
In 2026, this field has accelerated dramatically. In May, Google DeepMind released Genie 3, a real-time interactive world model capable of generating persistent 3D environments at 24 frames per second from text or images, with self-learned physics rather than hard-coded rules. NVIDIA’s Cosmos platform, which offers open-weight world foundation models trained on 20 million hours of real-world data, has surpassed two million downloads and is being adopted by robotics firms including 1X, Figure AI, and Agility for synthetic training data generation. Fei-Fei Li’s World Labs launched Marble, a commercially available system for creating editable 3D worlds from text, images, or video. Meanwhile, Yann LeCun’s AMI Labs—reportedly valued at €3 billion before releasing a product—raised €500 million to pursue JEPA-style architectures that learn abstract representations by predicting in latent space rather than pixels.
Other notable developments include Runway’s Gen-4.5, which the company explicitly frames as a “world model” with realistic physics, and DreamZero, a 14-billion-parameter world action model that demonstrated strong cross-embodiment transfer from human video to robot control using only visual information without action labels.
The convergence of these efforts suggests a broader architectural shift. Where large language models mastered pattern matching in text, the next generation of AI aims to understand causality, physics, and procedure by observing the world directly. Induction Labs’ approach—learning from unlabeled video through state prediction—aligns with this trajectory while removing a critical bottleneck: the need for humans to translate every observation into labeled actions first. Whether imagination models can scale beyond screen recordings to physical labor, social interaction, and complex real-world dynamics remains an open question. For now, Photon-1 stands as a compelling proof of concept that machines can learn to act, in part, simply by watching.
The post At 30× Less Compute, Induction Labs’ ‘Imagination Model’ Outperforms Google By Watching appeared first on Metaverse Post.
