Induction Labs has released a research result that challenges a long-standing assumption in artificial intelligence: that machines must be trained with carefully labeled examples of each action they are expected to perform. The company introduced Photon-1, the first model in what it calls a new class of foundation architecture known as “imagination models,” designed to learn from internet-scale video without seeing action labels during pretraining.

Photon-1 is a sparse 106-billion-parameter mixture-of-experts transformer with 5 billion active parameters. It was trained on about 575 million frames of computer screen recordings, which Induction Labs says is equivalent to 18 years of video sampled at one frame per second. The training data was derived from an initial index of two billion publicly available videos, narrowed to roughly two million screen recordings and then stripped of redundant frames using an internal keyframe detection model. The model was pretrained from scratch for one epoch, requiring approximately 30,000 NVIDIA H200 GPU-hours and 4.4 × 10²² FLOPs.

Induction Labs’ main claim is that Photon-1 outperforms Google’s Gemini 3.1 Flash-Lite on the company’s internal computer-use benchmarks, despite being trained with at least 30 times less compute. The company also says Photon-1 costs about three times less per million tokens to serve, reporting a weighted inference cost of $0.11 per million tokens versus $0.36 for Gemini. Those figures carry important limitations: the benchmark is internal and unreleased, so the results are not independently reproducible, and the Gemini compute estimate is Induction Labs’ own conservative projection rather than externally verified data.

The claim is notable because computer-use agents are one of the areas where AI systems still depend heavily on supervised demonstrations, tool-specific workflows, and environment-specific evaluation. If models can learn useful interface dynamics from unlabeled screen recordings, the amount of human annotation needed to build task-performing agents could be reduced, though Photon-1’s reported results still depend on a smaller labeled finetuning stage and reinforcement learning after pretraining.

We’re introducing imagination models: a new foundation model architecture that unlocks learning from internet-scale video. Our first imagination model, Photon-1, learned to use a computer by watching 18 years of screen recording video without action labels. pic.twitter.com/DMhRqL28si — Induction Labs (@induction_labs) July 23, 2026

We’re introducing imagination models: a new foundation model architecture that unlocks learning from internet-scale video. Our first imagination model, Photon-1, learned to use a computer by watching 18 years of screen recording video without action labels. pic.twitter.com/DMhRqL28si

How Imagination Models Work

Photon-1 differs from conventional approaches not only in scale, but also in architecture. Instead of predicting the next word or generating raw pixels, imagination models autoregressively predict future frames inside a learned representation space using a next-latent-token objective. Each video frame is compressed into 960 discrete tokens through finite scalar quantization and occupies only 2.2 kilobytes. Induction Labs says that is roughly 100 times smaller than existing OCR and multimodal representations, while still preserving text, layout, and state changes.

A differential latent encoder processes frames in pairs, encoding the differences between consecutive states rather than representing each frame only as absolute visual content. This compression makes it computationally practical to predict future states autoregressively at scale. During pretraining, the model learns what the company calls an “implicit policy”: by forecasting what a screen will look like next, it internalizes the causal structure of computer interfaces without being told which mouse clicks or keystrokes caused the observed transitions.

Converting that observational knowledge into a working agent required a second stage. Induction Labs finetuned Photon-1 on fewer than 35,000 labeled computer-use trajectories to teach the model the correct action format. The company added special tokens that allow Photon-1 to emit keyboard and mouse commands. At inference time, the system works in two steps: it first imagines the next state that would move the task forward, then generates the action intended to reach that state.

Induction Labs then used online reinforcement learning, with real-time rollouts on Linux virtual machines across five desktop environments. Outcomes were programmatically verified, and reward signals were used to further improve the model. That evaluation setup keeps the reported work focused on software environments where states and outcomes can be instrumented, rather than on open-ended physical settings where verification and safety constraints are harder to automate.

The company also reports that Photon-1’s capabilities extend beyond the screen-recording domain it observed during pretraining. When finetuned on 20,000 tournament checkers games, Photon-1 outperformed both a vision-encoder baseline and a similarly sized language model baseline on 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 with 1.15 for the LLM baseline and 1.44 for the vision baseline.

According to Induction Labs, the model also learned human behavioral patterns from its pretraining data. It learned to prompt an in-virtual-machine ChatGPT clone, evaluate its outputs, and steer the conversation until a task was completed, resembling the way people use AI tools in practical workflows.

The Expanding Frontier of World Models

Photon-1 is being introduced as the AI industry increasingly focuses on what researchers broadly call “world models.” These systems are designed to build internal representations of how environments evolve in response to actions, rather than only predicting text or producing isolated video clips.

In 2026, development in this area has accelerated. In May, Google DeepMind released Genie 3, a real-time interactive world model that can generate persistent 3D environments at 24 frames per second from text or images, using self-learned physics instead of 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 used by robotics firms including 1X, Figure AI, and Agility to generate synthetic training data.

Fei-Fei Li’s World Labs launched Marble, a commercial system for creating editable 3D worlds from text, images, or video. 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 developments include Runway’s Gen-4.5, which the company explicitly describes 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 and no action labels.

Taken together, these efforts point to a broader architectural shift in AI research. While large language models have become effective at pattern matching in text, the next generation of AI systems is increasingly aimed at understanding causality, physics, and procedures through direct observation. Induction Labs’ method, which learns from unlabeled video through state prediction, fits that direction while addressing a key bottleneck: the need for humans to convert observations into labeled actions before training.

The next tests for the approach will be external reproducibility, broader benchmark coverage, and evidence that the same training recipe works across less controlled environments. Whether imagination models can scale beyond screen recordings to physical labor, social interaction, and complex real-world dynamics remains unresolved. For now, Photon-1 is presented as a proof of concept that machines can learn to act, at least in part, by watching.