Lusob
Lusob (NeuralOS) is a neural window manager that moves and resizes windows from raw mouse actions.
Last verified:
What is Lusob?
Lusob (NeuralOS) is an experimental neural network window manager that moves and resizes windows without programming an event system. Instead of traditional event-based programming, it uses a simple neural network that guesses pixels from mouse actions to control window behavior: a UNET predicts what the screen looks like next from the last two frames and mouse position, while a small MLP with separate move and resize heads converts clicks near the title bar and grip into window motion. The trained models were exported to ONNX and run entirely in the browser without a server. Aimed at researchers, AI enthusiasts, and developers interested in experimental human-computer interaction and neural-native interfaces.
Lusob pricing
Pricing model: Freemium
Free and open source. The project is available on GitHub and runs entirely in the browser without any server costs or paid plans.
Lusob pros
- No event system programming required for window management
- Runs entirely in browser without server dependency
- Exports to ONNX format for portability
- Uses only canvas element and two small neural networks
- Deterministic renderer with rectangles drawn in JavaScript
- Window behavior learned from examples rather than hardcoded
- Simple convolutional neural network architecture
- UNET model with efficient encoder-decoder design
- Separate heads for moving and resizing prevent confusion
- Mouse vector coordinates projected to bottleneck spatial size
- Motion information feeds each jump connection
- Small MLP for efficient dx, dy, dw, dh generation
- Demonstrates AI integration in OS core components
- Open source with accessible GitHub repository
- Proof-of-concept for neural-native operating interfaces
Lusob cons
- Experimental proof-of-concept not production-ready
- Only handles window movement and resizing, not full OS
- Requires training data for window behavior learning
- Limited to mouse actions, keyboard support unclear
- Small neural network may have accuracy limitations
- No traditional event system may limit compatibility
- Browser-only implementation limits deployment options
- Unclear performance on complex multi-window scenarios
Frequently asked questions about Lusob
What is NeuralOS?
NeuralOS is a neural network window manager that moves windows without programming an event system. It uses a simple neural network that guesses pixels from mouse actions to control window movement and resizing, representing an experimental approach to window management through AI rather than traditional event-based programming.
How does the neural network predict window movements?
The neural network predicts what the screen would look like next based on the last two frames and the mouse position. It's a UNET architecture where the encoder compresses stacked frames, the decoder reconstructs the next frame, and mouse vector coordinates are projected with a linear layer to fit the bottleneck's spatial size before decoding.
What architecture does NeuralOS use?
NeuralOS uses a UNET (unitary neural network) architecture with an encoder that compresses stacked frames and a decoder that reconstructs the next frame. Mouse vector coordinates are projected with a linear layer to fit the spatial size of the bottleneck, where they are concatenated before decoding so motion information feeds each jump connection.
How does window movement differ from resizing in the model?
A small MLP receives (distance to the title bar, distance to the resize point, click) and generates (dx, dy, dw, dh) with two separate heads: one for moving and one for resizing. The heads share nothing except the click signal, so the model cannot confuse dragging with resizing.
Does NeuralOS require a server?
No, NeuralOS runs entirely in the browser without a server. The model was exported to ONNX and everything runs client-side using just a canvas element and two small neural networks communicating with each other.
Is the renderer deterministic?
Yes, with the new approach the renderer remains deterministic, with rectangles drawn in JavaScript. However, the window's behavior - where it moves and how it resizes - is learned from examples through the neural network rather than being hardcoded.
What inputs does the MLP receive?
The small MLP receives three inputs: distance to the title bar, distance to the resize point, and click status. These inputs are used to generate the output values (dx, dy, dw, dh) that control window movement and resizing.
Can I use NeuralOS in production?
No, NeuralOS is an experimental proof-of-concept project. It was created as an experiment to answer whether a neural window manager could work, training a neural network to predict screen frames. It is not designed for production use but rather demonstrates the potential for AI-integrated window management.
Where can I access NeuralOS?
NeuralOS is available at lusob.github.io/neural-os for the demo and the source code is on GitHub. The project is open source, allowing anyone to access, study, and contribute to the codebase.
What makes NeuralOS different from traditional window managers?
Traditional window managers use programmed event systems with hardcoded rules for window behavior. NeuralOS replaces this with a neural network that learns window behavior from examples, predicting screen frames based on mouse actions without explicit event system programming. This represents a paradigm shift from rule-based to learning-based window management.