Artificial intelligence research has largely focused on training models using text, images, and structured datasets. However, a growing number of researchers believe that understanding how humans interact with computers may unlock a new path toward more capable AI systems. Instead of teaching models through explicit instructions or action labels, these systems learn by observing the consequences of actions over time.
This idea sits at the center of Photon-1, a new research model developed by Induction Labs. The company describes Photon-1 as an “imagination model” capable of predicting future states of a system after observing visual input. Unlike traditional AI agents that rely heavily on labeled action data, Photon-1 learns by watching what happens next.
The model was trained on approximately 18 years of computer demonstration video, allowing it to develop an understanding of how digital environments evolve over time. According to Induction Labs, Photon-1 can simulate desktop interactions, play strategic games such as checkers, and even model physical systems like billiard ball dynamics despite never being explicitly trained on those domains during pretraining.
Perhaps most notably, the company claims Photon-1 outperforms Gemini 3.1 Flash-Lite on its internal computer-use benchmark while requiring dramatically less pretraining compute and significantly lower serving costs.
While Photon-1 is currently a research project with no public API, open weights, or commercial deployment, the techniques behind it offer a fascinating glimpse into the future of world-model-based AI.
Key Takeaways
Photon-1 introduces a novel imagination-based approach to AI training that focuses on future-state prediction rather than explicit action learning.
Some of the most important findings include:
- Photon-1 is a 106B-A5B Mixture-of-Experts transformer trained from scratch.
- The model learned from approximately 18 years of computer demonstration video.
- FSQ compresses each frame into 960 tokens, producing roughly 2.2 KB representations.
- Induction Labs reports more than 100× compression gains compared to OCR and multimodal systems.
- Training required approximately 30,000 H200 GPU-hours and 4.4 × 10²² FLOPs.
- Photon-1 outperformed Gemini 3.1 Flash-Lite on an internal benchmark while using roughly 27× less pretraining compute and about 3× lower serving costs.
- The model generalized successfully to checkers and billiard physics despite never seeing those domains during pretraining.
- Reinforcement learning enabled the system to translate imagined future states into executable actions.
For now, Photon-1 remains a research achievement rather than a deployable product. There are currently no public weights, APIs, or licensing options available. Nevertheless, the project offers a fascinating look at how future AI systems might learn by observing the world and predicting what happens next, rather than relying exclusively on explicit instructions or labeled actions.
Understanding the Concept of an Imagination Model
Most modern AI systems learn by predicting the next word, image token, or action based on previous inputs. Photon-1 approaches learning differently.
Rather than generating pixels or explicitly predicting actions, Photon-1 predicts future states in a compressed latent representation space. This process allows the model to learn how environments evolve over time without requiring detailed action annotations.
Induction Labs refers to this architecture as an imagination model because it learns to simulate possible future outcomes internally before producing a response.
The key insight is that understanding future states may be sufficient for learning behavior.
Instead of being told:
- Click this button
- Open this menu
- Move this object
Photon-1 simply watches what happens next.
Over time, it develops an understanding of goals, workflows, and interactions by learning state transitions.
Why Future-State Prediction Matters
Traditional computer-use agents often depend on extensive labeled datasets that map actions to outcomes.
These datasets are expensive to create because every interaction must be manually recorded and annotated.
Photon-1 removes this requirement.
The model learns by predicting what the next screen state should look like.
This seemingly simple objective has a powerful consequence.
If the model becomes good at predicting future states, it implicitly learns:
- User intentions
- Interface navigation
- Task completion patterns
- Workflow structures
Induction Labs describes this phenomenon as learning an implicit policy.
The model develops an understanding of what users are trying to accomplish rather than memorizing specific mouse clicks or keyboard shortcuts.
How Photon-1 Represents Visual Information
One of the biggest challenges in training video-based AI systems is data efficiency.
Raw video contains enormous amounts of information.
Storing and processing every pixel across billions of frames quickly becomes prohibitively expensive.
To solve this problem, Photon-1 relies on a specialized vision encoder built around Finite Scalar Quantization (FSQ).
The Finite Scalar Quantization (FSQ) Compression System
FSQ allows Photon-1 to compress visual information into compact discrete tokens.
Each frame is represented using:
- 960 discrete tokens
- An 8-dimensional vector for each token
Each vector dimension can take one of five possible values:
- -1
- -1/2
- 0
- 1/2
- 1
This structure creates a codebook consisting of:
5⁸ possible codes
The result is an extremely compact representation of visual information.
Achieving Massive Compression Gains
Using this encoding method, each video frame occupies approximately:
2.2 KB
According to Induction Labs, this delivers more than:
100× greater compression
compared with existing OCR-based systems and multimodal representations.
Importantly, the compression retains critical information such as:
- Text content
- Interface layouts
- Screen state changes
- User interactions
This enables Photon-1 to process vast quantities of video while maintaining computational efficiency.
Differential Latent Encoding Improves Efficiency
Photon-1 introduces another optimization known as a differential latent encoder.
Instead of encoding each frame independently, the system analyzes pairs of frames.
The generated latent representation focuses on differences between frames rather than storing complete frame contents.
This approach significantly reduces redundancy because most adjacent frames in desktop recordings change only slightly.
Examples include:
- Cursor movement
- Scrolling
- Menu interactions
- Text entry
By concentrating on changes, Photon-1 can represent long video sequences more efficiently while preserving important state transitions.
Building an 18-Year Video Dataset
Training an imagination model requires enormous quantities of sequential data.
Induction Labs began with an internal index containing approximately:
2 billion publicly available videos
This massive collection underwent extensive filtering.
The goal was to isolate computer-use demonstrations rather than general video content.
Narrowing the Dataset
The filtering process reduced the original pool to roughly:
2 million computer screen recordings
The team then applied a proprietary keyframe detection system to eliminate redundant frames.
This step removed unnecessary visual duplication while preserving meaningful state transitions.
The final dataset contained:
- 575 million frames
- Sampled at 1 frame per second
- Equivalent to approximately 18 years of video footage
After tokenization, the dataset expanded to roughly:
552 billion tokens
This enormous corpus became the foundation for Photon-1 pretraining.
Training Photon-1 From Scratch
Unlike many modern AI systems that build upon existing foundation models, Photon-1 was pretrained from scratch.
The model itself is a:
106B-A5B sparse Mixture-of-Experts (MoE) transformer
This means the architecture contains 106 billion parameters overall while activating only a subset during inference.
Mixture-of-Experts designs help reduce computational requirements while maintaining model capacity.
Training Compute Requirements
Photon-1 was trained with:
- 32K token context length
- Approximately 30,000 H200 GPU-hours
- Around 4.4 × 10²² training FLOPs
Induction Labs implemented the training system using:
- PyTorch
- Custom fused kernels
- Optimized vision encoding layers
- Specialized MoE infrastructure
The team reports sustaining:
40% end-to-end Model FLOP Utilization (MFU)
throughout training.
Interestingly, the reported training figures align mathematically.
Thirty thousand H200 GPU-hours at 40% utilization corresponds closely to the stated 4.3–4.4 × 10²² FLOPs.
From Imagination to Real Actions
Pretraining teaches Photon-1 to predict future states.
However, performing real-world tasks requires action generation.
To bridge this gap, Induction Labs fine-tuned Photon-1 using:
fewer than 35,000 computer-use trajectories
These trajectories teach the model how to translate predicted states into executable actions.
Predicting States Before Acting
Photon-1 follows a unique decision-making process.
Before selecting an action, it first predicts what the next state should look like.
Only then does it determine the action needed to reach that state.
This creates a planning-oriented workflow:
- Imagine future outcome
- Evaluate desired state
- Generate action sequence
This differs from many current AI agents that generate actions directly without explicit future-state simulation.
Reinforcement Learning at Scale
Following supervised fine-tuning, Photon-1 undergoes online reinforcement learning.
Training occurs inside large numbers of virtual machines operating in real time.
Task completion is automatically verified to generate reward signals.
This enables large-scale autonomous learning without extensive human supervision.
The Virtual Machine Environment
The reinforcement learning infrastructure includes Linux virtual machines running multiple desktop environments:
- LXQt
- Xfce
- MATE
- GNOME
- Plasma
Each environment includes:
- A Google account for login-protected websites
- An internal ChatGPT-style assistant
- Unlimited access without rate restrictions
These systems create realistic digital workspaces for training computer-use behavior.
Photon-1 vs Gemini 3.1 Flash-Lite
One of the most attention-grabbing claims from Induction Labs involves efficiency.
The company reports that Photon-1 surpasses Gemini 3.1 Flash-Lite on an internal computer-use benchmark while requiring far less compute.
Pretraining Compute Comparison
| Model | Pretraining Compute |
|---|---|
| Gemini 3.1 Flash-Lite | 1.200 × 10²⁴ FLOPs |
| Photon-1 | 0.044 × 10²⁴ FLOPs |
Based on these figures, Photon-1 uses approximately:
27× less pretraining compute
than Gemini 3.1 Flash-Lite.
Induction Labs refers to this as “at least 30×” because it believes Gemini’s actual training requirements may have been higher.
Inference Cost Comparison
The company also reports significant serving cost advantages.
Weighted Cost Per 1 Million Tokens
| Model | Cost |
|---|---|
| Gemini 3.1 Flash-Lite | $0.36 |
| Photon-1 | $0.11 |
The calculation assumes:
10:1 input-to-output token ratio
which Induction Labs says reflects typical computer-use workloads.
This places Photon-1 at roughly:
3× lower serving cost
than Gemini 3.1 Flash-Lite.
Important Caveats About the Comparison
While the results are impressive, the company notes several limitations.
First, the Gemini compute estimate is an internal approximation based on assumptions including:
- 8 billion active parameters
- 25 trillion pretraining tokens
Second, the benchmark itself remains unpublished.
As a result, external researchers cannot independently reproduce the findings today.
These limitations are important when evaluating performance claims.
Testing Generalization Beyond Desktop Tasks
Perhaps the most fascinating aspect of Photon-1 is its ability to generalize.
The model was pretrained exclusively on desktop-use video.
Yet researchers evaluated it in domains entirely absent from pretraining.
Checkers Performance
One evaluation involved:
20,000 tournament checkers games
sourced from:
Photon-1 was fine-tuned on this domain and compared against:
- A vision encoder baseline
- An LLM baseline (Ling-flash-2.0 from Inclusion AI)
The results showed Photon-1 outperforming both systems in:
- World simulation quality
- Move selection quality
This suggests future-state prediction transfers effectively into strategic planning tasks.
Modeling Billiard Physics
Researchers also evaluated Photon-1 using:
10,000 synthetically generated billiard games
simulated at:
5 frames per second
The goal was to predict future physical states.
Physics Prediction Results
Photon-1 achieved:
Mean Absolute Error (MAE): 0.47
Compared with:
| Model | MAE |
|---|---|
| Photon-1 | 0.47 |
| LLM Baseline | 1.15 |
| Vision Encoder Baseline | 1.44 |
The results indicate substantially stronger physical-world modeling capabilities despite the model never seeing billiard simulations during pretraining.
Learning Human Behavioral Priors
Another surprising outcome emerged after reinforcement learning.
Photon-1 learned to leverage the internal ChatGPT-style assistant available in training environments.
Instead of treating the assistant as a tool to query randomly, Photon-1 used it similarly to how humans do.
The model learned to:
- Draft documents
- Gather information
- Answer knowledge questions
- Assist with task completion
This suggests that observing human computer behavior may enable the acquisition of broader behavioral priors.
What Photon-1 Means for the Future of AI
Photon-1 represents a compelling shift away from traditional action-labeled learning.
Rather than teaching an AI exactly what actions to perform, Induction Labs demonstrates that predicting future states can produce useful behavior patterns.
The research suggests that large-scale observation alone may provide enough information for models to develop planning, reasoning, and interaction capabilities.
If these findings continue to scale, future AI systems could learn from massive quantities of passive video data without requiring costly action annotations.
Discover more from AiTechtonic - AI & Informative News
Subscribe to get the latest posts sent to your email.