Induction Labs Photon-1 Simulates Desktops, Plays Checkers, and Models Billiard Physics From One Pretraining Run
Most brokers that study from video have to know what motion produced every body. Induction Labs is arguing that this requirement is the bottleneck. Last week, they launched creativeness fashions, a basis mannequin structure that pretrains on uncooked video with no motion labels in any respect.
Their take a look at system is Photon-1, a sparse 106B-A5B mixture-of-experts (MoE) transformer skilled on 18 years of laptop demonstration video. On an inside laptop use benchmark, Induction Labs stories that Photon-1 beats Gemini 3.1 Flash-Lite whereas utilizing far much less pretraining compute and costing roughly 3× much less to serve.
What an creativeness mannequin really does
An creativeness mannequin predicts future frames autoregressively utilizing a next-latent-token-prediction goal. It doesn’t generate pixels throughout pretraining. Everything is modeled in a discovered illustration area.
The declare that issues is that this: predicting future states teaches the mannequin to finish duties, despite the fact that it by no means sees an motion throughout pretraining. Induction Labs calls this an implicit coverage. The mannequin learns ideas of what an individual is doing, somewhat than a label for every mouse click on.
The compression trick that makes it scale
The structure is determined by a imaginative and prescient encoder that makes use of finite scalar quantization (FSQ). Each body is compressed into 960 discrete tokens. Each token is an 8-dimensional vector. Each dimension takes one among 5 values: −1, −1/2, 0, 1/2, 1. That offers a codebook of 5⁸ attainable codes.
The ensuing encoding is about 2.2 KB per body. Induction Labs stories over 100× higher compression than present OCR and multimodal-model representations, whereas preserving textual content, structure and state adjustments.
To hit that charge, Photon-1 makes use of a differential latent encoder. It encodes video frames as pairs, so the latents describe variations between frames somewhat than body contents.
Data and pretraining compute
The corpus begins from an inside index of 2 billion publicly obtainable movies. Filtering reduces that to roughly 2 million laptop display recordings. An inside keyframe detection mannequin strips redundant frames.
The last dataset is 575 million frames, sampled at 1 body per second. That equals 552 billion tokens, or about 18 years of video. Photon-1 was pretrained from scratch for a single epoch.
Training the 106B-A5B MoE at 32K context took roughly 30,000 H200 GPU-hours, or 4.4×10²² coaching FLOPs. The analysis crew carried out coaching in PyTorch with customized fused kernels for the imaginative and prescient encoder and MoE layers, sustaining 40% end-to-end MFU. Those three figures are mutually constant: 30,000 H200-hours at 40% MFU lands nearly precisely on 4.3×10²².
From creativeness to motion
Induction Labs finetuned Photon-1 on fewer than 35,000 laptop use trajectories to show the motion and instruction format. Special laptop use tokens let the mannequin emit actions. At inference, Photon-1 predicts the subsequent body’s state first, then outputs the motion that will get there.
Online reinforcement studying follows. Rollouts run in actual time on digital machines at scale, and outcomes are verified programmatically to supply reward. The Linux VMs run 5 desktop environments (LXQt, Xfce, MATE, GNOME and Plasma), every with a Google account for login-restricted net apps and an inside ChatGPT clone with no charge limits.
The compute and value comparability
| Model | Pretraining compute | Weighted inference value / 1M tokens* |
|---|---|---|
| Gemini 3.1 Flash-Lite | 1.200 × 10²⁴ FLOPs | $0.36 |
| Photon-1 | 0.044 × 10²⁴ FLOPs | $0.11 |
*Weighted at a ten:1 input-to-output token ratio, which Induction Labs says matches its laptop use assessments.
Two caveats belong subsequent to that desk. First, the Gemini determine is Induction Labs’ personal conservative estimate, assuming 8B energetic parameters and 25T pretraining tokens. Taken at face worth the ratio is about 27×, not the 30× headline; Induction Labs states “not less than 30×” on the idea that the true Gemini quantity is probably going increased and the mannequin was probably distilled. Second, the benchmark is inside and unreleased, so the consequence just isn’t independently reproducible in the present day.
Photon-1’s personal breakeven value on Induction Labs’ {hardware} is $0.06 per 1M enter tokens and $0.60 per 1M output tokens, with no speculative decoding.
Does it generalize previous the desktop?
This is the extra fascinating take a look at, as a result of Photon-1 noticed solely laptop use video. The analysis crew finetuned it on domains absent from pretraining and in contrast in opposition to two baselines: a imaginative and prescient encoder baseline with the identical structure and dimension however no creativeness pretraining, and an LLM baseline (Ling-flash-2.0 from Inclusion AI, pretrained on 20T tokens).
On 20,000 match checkers video games from the Open Checkers Archive 2.0, Photon-1 beat each baselines on world simulation and on transfer high quality. On 10,000 synthetically generated billiard video games simulated at 5 fps, it produced a imply absolute error of 0.47 in opposition to the ground-truth physics engine, versus 1.15 for the LLM baseline and 1.44 for the imaginative and prescient encoder baseline.
Photon-1 additionally picked up human priors from the pretraining video. After RL, it discovered to make use of the in-VM ChatGPT clone to draft artifacts and reply data questions, steering the LLM the best way an individual would.
Key Takeaways
- Photon-1 learns an implicit coverage from 18 years of display recordings with zero motion labels, utilizing next-latent-token prediction.
- FSQ compresses every body to 960 tokens (~2.2 KB), a reported 100× acquire over OCR and multimodal representations.
- Trained for ~30,000 H200 GPU-hours, it beats Gemini 3.1 Flash-Lite on an inside benchmark at ~27× much less pretraining compute.
- Despite seeing solely desktop video, it beats an LLM baseline at checkers and billiard physics after finetuning.
- No weights, no API, no license — this can be a analysis consequence, not a deployable mannequin.
Check out the full technical writeup from Induction Labs and the announcement thread on X. All credit score for this analysis goes to the researchers of this mission.
The put up Induction Labs Photon-1 Simulates Desktops, Plays Checkers, and Models Billiard Physics From One Pretraining Run appeared first on MarkTechPost.
