Induction Labs says its new “imagination model” learned from the equivalent of 18 years of screen recordings. The underlying technique could change how AI agents are trained—but the headline leaves out an important stage.

A new AI model reportedly learned how computers work by watching people use them.

San Francisco startup Induction Labs introduced Photon-1 this week, describing it as the first of a new class of “imagination models” trained to learn from internet-scale video. The company says Photon-1 was pretrained on the equivalent of 18 years of screen recordings without being told which mouse clicks, keystrokes or other actions produced the changes it observed.

That description makes the achievement sound almost inexplicable. In reality, Photon-1 applies a familiar AI training principle to an unusually rich source of data: instead of predicting the next word in a sentence, it predicts what a computer screen will look like next.

The result is potentially significant. It suggests that much of the video already available online could be used to teach AI systems how processes unfold, without humans first having to label every action taking place.

But there is an important qualification: Photon-1 did not become a working computer agent entirely from unlabelled video. The videos taught it how computer environments behave. A later stage involving labelled trajectories and reinforcement learning taught it how to take actions.

What does “18 years of video” mean?

The training did not involve one 18-year-long recording. Induction Labs began with an internal index of two billion publicly available videos and filtered it down to approximately two million computer screen recordings.

The company then used a keyframe-detection system to remove redundant images and sampled the remaining videos at one frame per second. That produced 575 million frames, equivalent to approximately 18 continuous years of video.

Each frame was compressed by a vision encoder into 960 discrete “latent tokens.” These tokens represented information such as text, windows, menus, layout and changes from the preceding frame without requiring the model to process every original pixel.

Photon-1 was then trained to predict the next latent representation.

Imagine a screen recording in which a cursor moves to a menu, the menu opens, an option is selected and a new window appears. The training data does not have to contain labels saying “click,” “open menu” or “select Save.” To predict the next screen correctly, the model still has to learn the relationships between those visible states.

After encountering these patterns millions of times, it begins to develop an internal representation of how computer interfaces behave.

“Our key result is that predicting future states can teach models to complete tasks despite never seeing actions during pretraining,” Induction Labs said in its technical announcement⁠.

This is what the company calls “learning by imagining.” It is not imagination in the human sense. It is the probabilistic prediction of a future state based on the states that preceded it.

Where the actions come from

The distinction between understanding screen transitions and controlling a computer is critical.

Following video pretraining, Induction Labs fine-tuned Photon-1 using fewer than 35,000 computer-use trajectories. During this stage, the company added special tokens that allowed the model to output actions in the required format.

It then used online reinforcement learning. Photon-1 was allowed to perform tasks inside virtual machines, while software checked whether it had achieved the requested outcome and supplied a reward signal.

At inference time, the system reportedly works in two stages. Photon-1 first predicts the screen state that would move it closer to completing its task. It then produces the mouse or keyboard action intended to create that state.

In simplified form, it does not begin by asking, “Which button should I press?” It asks, “What should the screen look like next?” and then selects an action intended to make that prediction real.

This means the claim that Photon-1 learned to use a computer “without action labels” is directionally true about its pretraining, but incomplete as a description of the finished agent. Action-labelled fine-tuning and reinforcement learning were still required to translate its observational knowledge into reliable control.

Why this differs from earlier video training

Researchers have previously trained AI agents using unlabelled online video. OpenAI’s 2022 Video PreTraining project, for example, trained an agent to play Minecraft by watching internet videos.

That system used a small labelled dataset to train an “inverse dynamics model,” which inferred the likely mouse and keyboard actions occurring in much larger collections of unlabelled footage. Those inferred actions were then used as training labels.

Photon-1 attempts to avoid that intermediate requirement during pretraining. It learns to predict future states directly, allowing it to extract patterns from video even when the underlying action is difficult to identify or describe.

That distinction matters beyond computer interfaces. Not everything humans learn from observation can be reduced to a convenient action label. Skilled physical work, social interaction and processes involving multiple simultaneous events may be easier to represent by predicting how the overall situation evolves.

Induction Labs says it ultimately sees the architecture as “a path to intelligence that learns by observing the world directly.”

What Photon-1 reportedly achieved

Photon-1 is a sparse mixture-of-experts transformer containing 106 billion total parameters, with approximately five billion active during processing. The 575-million-frame dataset became 552 billion latent tokens.

Induction Labs says pretraining required approximately 30,000 H200 GPU-hours, equivalent to 4.4 × 10²² training FLOPs.

After fine-tuning and reinforcement learning, Photon-1 reportedly outperformed Gemini 3.1 Flash-Lite on the company’s internal computer-use benchmark while using an estimated 30 times less pretraining compute. Induction Labs also says Photon-1 is approximately three times cheaper to serve under the usage pattern assumed in its comparison.

Those results have not yet been independently replicated. The benchmark is internal, and the comparison with Gemini relies partly on Induction Labs’ estimates of a closed model’s architecture and training requirements.

The company also tested whether Photon-1’s representations transferred beyond conventional desktop activity. After fine-tuning on 20,000 tournament checkers games, it reportedly outperformed both a vision baseline and a pretrained language-model baseline at predicting board states and selecting moves.

A second experiment fine-tuned the model on 10,000 synthetic billiards games. Photon-1 produced more accurate predictions of ball positions than the comparison models, suggesting that its video pretraining had created representations useful for learning other kinds of dynamics.

These experiments do not demonstrate a general understanding of physics or games, but they offer evidence that the model learned something broader than the visual appearance of desktop software.

The demonstrations require another clarification

The company’s videos show Photon-1 “imagining” complete computer screens. However, Photon-1 itself operates in compressed representation space rather than directly generating screenshot pixels.

To make its predictions visible, Induction Labs fine-tuned Ideogram 4 as an auxiliary image generator that converts Photon-1’s internal representations into screenshot-like images. The company says this decoder was used only for visualization and was not part of Photon-1’s training or normal inference.

The displayed screens should therefore be understood as visual interpretations of the model’s predicted states, rather than native pixel outputs from Photon-1.

Why businesses should pay attention

Most current computer-use agents rely heavily on language models that inspect screenshots, reason about instructions and choose actions one step at a time. They can be capable, but they are also brittle. Interfaces change, buttons move and the model may understand the words on a screen without fully understanding how the underlying application behaves over time.

Observational pretraining offers another route. A model exposed to millions of real workflows could learn recurring interface patterns, consequences and task sequences before receiving much task-specific instruction.

That could eventually improve agents operating across legacy enterprise software, internal administrative systems and applications without usable APIs. It could also reduce the amount of labelled interaction data needed to train agents for specialized environments.

The same possibility creates governance questions. Corporate screen recordings can contain customer information, passwords, private communications, proprietary workflows and regulated data. If observational video becomes valuable training material, companies will need policies governing its collection, retention, consent and permitted uses. Induction Labs says its corpus originated from publicly available videos, but its announcement does not provide detailed licensing or provenance information for the approximately two million recordings ultimately selected.

Photon-1 is an early company result, not yet definitive proof of a new foundation-model paradigm. Its strongest performance claims come from internal testing, and the system still needed labelled fine-tuning and reinforcement learning before it could act.

The more important development is the architecture’s proposition: video does not necessarily need to be exhaustively labelled to teach an AI how an environment changes. If that proposition holds at larger scales and across more physical settings, the internet’s enormous video archives could become something more than visual content.