ViciousSquid/Dosidicus

What it solves

Dosidicus is a transparent cognitive sandbox that solves the "black box" problem of modern AI. Instead of hiding intelligence in millions of opaque parameters, it provides a visible, understandable neural network where every neuron, synapse, and weight change can be traced back to a specific experience or action.

How it works

The project uses a custom simulation engine called STRINg, built from scratch in NumPy without relying on TensorFlow or PyTorch. It implements a neural network that evolves through experience using Hebbian learning and Spike-Timing-Dependent Plasticity (STDP). The brain starts with a small set of core neurons (drives and senses) and grows new neurons dynamically when the existing structure cannot cope with new demands. It features a dual memory system and a temporal credit assignment system to link actions to consequences.

Who it’s for

It is designed for people interested in neuroscience, artificial life, and computational cognition who want to learn how simple learning rules create complex behaviors by raising a digital pet squid.

Highlights

• Full Provenance: Every synaptic change and new neuron records the exact mechanism and experience that caused it, allowing users to ask "why" a behavior emerged.
• Dynamic Growth: The network grows new neurons based on capability needs rather than scripted events.
• Visual Brain: A GUI allows users to watch neurons fire in real time and inspect the brain's structure.
• Headless Mode: Includes a CLI trainer for background simulation and training without the GUI.
• Custom Brains: Supports pre-configured neural network templates for specific behaviors.