INTERACTIVE NEURAL LABS
Omnicide.
Make learning visible. Train a small network, inspect its state, and see what changes one update at a time.
ExperimentalOmnicide pairs a small TypeScript neural library with interactive labs. Addition and multiplication experiments expose the network, training controls, loss curves, editable prediction inputs, and the response surface produced by the current model.
The wider collection explores trainable memory, persistent language state, readable object representations, learning speed, and shared neural structures. Some experiments run in the browser; others use local Python models.
The aim is to make a model's behavior easier to inspect and question. Learned approximations, experimental assumptions, and negative results are kept visible alongside the demonstrations.
A closer look
Watch the update
Start, pause, or step through training while inspecting actual parameters and loss.
See the surface
Interactive 3D plots show how two inputs map to the network's prediction.
Explore the state
Dedicated labs investigate memory, object representations, and reusable structure.