Induction Labs Trained a Computer-Use Model on 18 Years of Screen Recordings. It Never Saw a Single Click.

TL;DR

Induction Labs published a writeup on what it calls imagination models, and the first one, Photon-1, is a sparse 106B-A5B mixture-of-experts transformer that learned to operate a computer from 18 years of screen-recording video containing no action labels at all. No mouse coordinates, no click events, no keystroke traces. It cost roughly 30,000 H200 GPU-hours. Induction reports it beats Gemini 3.1 Flash-Lite on an internal computer-use benchmark at about a 27th of the pretraining compute and a third of the serving cost. The result went wide today; the caveats went wide with it.

The bottleneck it claims to remove

Every computer-use agent you can rent today was taught roughly the same way: show it a screenshot, tell it which action a human took, repeat. That pairing is the expensive part. Action-labeled trajectories have to be collected, instrumented, and cleaned, and there are maybe tens of thousands of good ones in the world, not billions.

Meanwhile there is an ocean of screen-recording video on the public internet with no labels attached whatsoever. Tutorials, streams, bug reports, walkthroughs. Induction's pitch is that the ocean is the training set, and the labels were never the point.

Photon-1 pretrains on one objective: given the frames so far, predict the next one. It never sees an action during pretraining. Induction argues that predicting the future state of a desktop forces the model to internalize what a person is doing and why, which it calls an implicit policy. Learn the consequences well enough and the causes fall out for free.

The intuition is riding shotgun. Spend a decade watching someone else drive and you never touch the wheel, but you build such a sharp model of what the car does next that the first time you sit in the driver's seat, your hands already know which way to turn.

How 18 years of video fits in a training run

The data pipeline is the unglamorous half of the result. Induction says it filtered 2 billion publicly available videos down to 2 million computer screen recordings, then sampled at 1 frame per second into a 575-million-frame dataset. That is 18 years of wall-clock video and 552 billion tokens, trained for a single epoch from scratch.

Getting 575 million frames into 552 billion tokens is the trick that makes this affordable. Instead of predicting pixels, Photon-1 predicts in a learned latent space. A vision encoder using finite scalar quantization crushes each frame to 960 discrete tokens, roughly 2.2KB, where every token is an 8-dimensional vector whose dimensions each take one of five values. That is a codebook of 5 to the 8th, and Induction claims about 100x better compression than OCR or standard multimodal representations while still preserving text, layout, and state changes.

Think of it as note-taking rather than photography. A screenshot stores every pixel of a code editor; Photon-1 stores the shorthand a stenographer would write down, which is far smaller and, for predicting what happens next, most of what actually mattered.

The whole run hit 4.4x10^22 training FLOPs at a reported 40% end-to-end model FLOPs utilization. For a frontier-adjacent claim, 30,000 H200-hours is a rounding error.

The comparison, and its asterisk

Here is the headline chart, and it is genuinely striking.

Read the asterisk carefully, because it is load-bearing. Google does not publish Gemini 3.1 Flash-Lite's training compute. Induction derives that 1.2x10^24 figure by assuming 8 billion active parameters and 25 trillion pretraining tokens and applying the standard 6ND rule. That is a reasonable back-of-envelope, and it is still an envelope.

The serving-cost side is easier to check. Induction lists Photon-1 at $0.06 per million input tokens and $0.60 per million output on its own hardware, blended to $0.11 at a 10-to-1 input-to-output ratio. Gemini 3.1 Flash-Lite's published price is $0.25 input and $1.50 output, which blends to $0.36 at the same ratio. Do the arithmetic yourself and both numbers land where Induction says they do. That part survives contact with a calculator.

What does not survive is verification of the win itself. The computer-use benchmark is internal and unreleased. There is no public leaderboard entry, no eval harness, and no third party who has reproduced it. A comparison against a competitor's estimated compute, scored on a test only one party holds, is a claim with a receipt attached to it, not a result.

It also plays checkers, which is the more interesting part

The transfer experiments are what separate this from another computer-use press release. If predicting the next screen really teaches a general world model, then the same pretrained weights should adapt to domains that have nothing to do with desktops.

Induction finetuned Photon-1 and two baselines, a vision-encoder model and an LLM, on 20,000 tournament checkers games. Photon-1 beat both on world simulation and on move quality. Then it moved to billiards, finetuning on 10,000 synthetically generated games, and measured mean absolute error of predicted ball positions against a physics engine.

A model that only ever watched people click around in Linux windows is now the best of three at guessing where a cue ball ends up. That is either a real signal that next-frame prediction learns transferable dynamics, or a sign that the baselines were poorly matched. Both readings are available, and only one of them gets tested when someone else can run the code.

Turning a watcher into a doer

Pretraining gets you a model that can imagine the next frame. It does not get you one that emits clicks. Induction closes that gap with fewer than 35,000 computer-use trajectories, which teaches the model the action format rather than the behavior, followed by online reinforcement learning with real-time rollouts on Linux VMs.

The inference loop is the part worth stealing conceptually: Photon-1 predicts the next screen state first, then emits the action that would produce it. It decides where it wants the world to go, then works backwards to the mouse. The RL environments span LXQt, Xfce, MATE, GNOME, and Plasma, which is more Linux desktops than most Linux users have opinions about.

What you cannot do with it

Nothing. Not yet, anyway. The writeup mentions no weights, no API, no license, and no download. This is a research post from a Y Combinator-backed startup, authored by David Li and Jonathan Li and dated July 23, that broke into wider circulation this week. Treat it as a direction, not a dependency.

The direction still matters if you build agents. Every computer-use product on the market today is rate-limited by how many labeled trajectories someone can afford to collect. If unlabeled video is a viable substitute, the moat around action-labeled datasets gets shallower fast, and the cost floor for a competent desktop agent drops with it. Induction says next stops are physical dynamics, skilled labor, and social interaction, all of which are also things there is a great deal of unlabeled video of.

Bookmark it, do not budget around it. The claim that deserves your attention is the architectural one, and the number that deserves your skepticism is the one from the benchmark you cannot run.

Key Takeaways

• Photon-1 is a 106B-A5B MoE transformer with a 32K context, pretrained for one epoch on 575 million frames of screen-recording video (552 billion latent tokens, 18 years at 1 fps) with no action labels.
• A finite-scalar-quantization vision encoder compresses each frame to 960 discrete tokens, about 2.2KB, which Induction claims is roughly 100x better than OCR or standard multimodal representations.
• The whole run took about 30,000 H200 GPU-hours and 4.4x10^22 FLOPs at 40% MFU, roughly a 27th of Induction's estimate for Gemini 3.1 Flash-Lite.
• The serving-cost comparison checks out ($0.11 vs $0.36 blended at 10:1), but the computer-use benchmark is internal and unreleased, and the Gemini compute figure is Induction's estimate rather than Google's disclosure.
• Transfer results are the strongest evidence: finetuned on 20,000 checkers games it beat both baselines, and on billiards it hit 0.47 MAE against 1.15 for an LLM baseline and 1.44 for a vision-encoder baseline.
• No weights, no API, no license. If unlabeled video really substitutes for action-labeled trajectories, the collection cost that gates every computer-use agent today stops being the moat.

Sources: Induction Labs, "Scaling Video Pretraining with Imagination Models", Y Combinator Launch: Induction Labs, MarkTechPost, Google, Gemini 3.1 Flash Lite, Gemini API pricing