"What if a Tamagotchi had a neural network and could learn stuff?" - Gigazine , Hackaday

Dosidicus electronicus

Learn neuroscience by raising a neural network as a pet

A transparent cognitive sandbox disguised as a digital pet squid with a neural network you can see thinking

• Part educational neuro tool, part sim game, part fever dream
• Combining 1990s virtual pet nostalgia with modern computational neuroscience.

Download for Windows, macOS, Linux and Android

git clone https://github.com/ViciousSquid/Dosidicus.git
cd Dosidicus
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate (Windows)
pip install -r requirements.txt
python main.py

---

Manifesto | Changelog | Wiki (53 pages)

---

Why this exists

What if you could understand every neuron inside a learning creature?

The project is designed to make artificial cognition visible.

Instead of hiding intelligence inside millions of parameters, Dosidicus starts with just eight neurons. Every connection can be inspected. Every activation can be visualised. Every learned behaviour can be traced back to experience.

As the squid lives, its brain rewires itself through Hebbian learning, strengthens useful pathways using STDP, and grows entirely new neurons through neurogenesis.

No two brains ever develop the same way.

Every save file becomes a permanent cognitive history.

As the caretaker you will

• Feed, clean and care for your squid.
• Introduce it to new experiences.
• Watch neurons fire in real time.
• Watch memories form and influence future behaviour.
• Observe fears, habits and preferences emerge.
• Raise a brain unlike anyone else's.

Under the hood runs STRINg simulation engine:

• Built from scratch in NumPy
• No TensorFlow. No PyTorch. No NEAT.
• Fully visible neuron activations
• Structural growth over time
• Dual memory system
• Headless training mode

Most AI is a black box: Dosidicus lets you see the mind forming - every neuron is visible & understandable.

The squid isn't driven by scripted behaviours — it develops through experience. By watching its brain change over time, you can explore how simple learning rules give rise to increasingly complex behaviour.

Want the full conceptual philosophy behind Dosidicus? Read the Cognitive Sandbox Manifesto

---

Share Your Squid

No two squids are wired the same.

• Early interactions permanently alter their structure (good or bad!).
• Tiny differences amplify.
• Habits form. Fears emerge. Personalities drift.

Your squid's brain is a cognitive history - shaped by you.

So share it.

• Export save files and let others explore your squid's neural structure.

• Post screenshots of strange activation patterns and unexpected growth.

• Show bizarre learned behaviors (Why is yours afraid of poop?)

• Compare cognitive histories and trace how experience shaped structure.

• Did yours grow 40 neurons?

• Did it develop a persistent avoidance loop?

• Did you accidentally create a neurotic reward spiral?

Every squid is an experiment.

---

Docker

Two targets are provided: headless (CLI trainer) and gui (PyQt5 app with X11).

Headless (recommended for containers):

docker build -t dosidicus:headless --target headless .
docker run --rm -v ${PWD}/headless_output:/app/output dosidicus:headless --ticks 10000 --output /app/output/trained_brain.json

GUI (Linux host with X11 or WSLg):

docker build -t dosidicus:gui --target gui .
docker run --rm \
  -e DISPLAY=$DISPLAY \
  -e QT_X11_NO_MITSHM=1 \
  -v /tmp/.X11-unix:/tmp/.X11-unix:rw \
  -v ${PWD}/saves:/app/saves \
  -v ${PWD}/logs:/app/logs \
  dosidicus:gui

Compose:

docker compose up --build
docker compose --profile gui up --build

WSLg note: If the GUI fails to start with a Qt platform plugin error, try:

export QT_QPA_PLATFORM=wayland
docker compose --profile gui up --build

Note: On Windows without WSLg, you will need an X server and a valid DISPLAY value to run the GUI container.

Note: Attempting to build the Docker container on Windows ARM64 will fail because there is no pyqt5 wheel [32] - Use the prebuilt binary from releases instead

Troubleshooting (quick):

• If DISPLAY is empty in WSL: WSLg is not active. Use WSLg or run an X server on Windows.
• If Docker errors mention docker_engine/pipe not found: start Docker Desktop and ensure WSL integration is enabled.
• If GUI still exits with Qt plugin errors: rebuild the image (docker compose --profile gui build --no-cache) and retry.

---

Technical Overview

• 41,636 lines, one developer, 28 months, GPL 2.0 license

• Dependencies:

  • Python ^3.9
  • PyQt5 ^5.15 (GUI framework)
  • numpy ^1.21 (neural network computations)
  • OPTIONAL onnxruntime or onnxruntime-directml (more info)

• Core Structure: Modular codebase in src/ including brain designer, decision engine, learning algorithms, personality traits, memory management, UI components, and interaction systems. Entry point via main.py.

Key Project Components

• Plugin System: Extensible architecture with built-in plugins for achievements (tracking milestones) and multiplayer (networked interactions).
• Save System: Persistent saves in saves/ for pet states, autosaves, and achievement logs.
• Headless Mode: Standalone training and simulation in headless/ for GUI-less operation, ideal for background training or server environments (experimental)
• Custom Brains: Library of pre-configured neural networks in custom_brains/ (e.g., "Plant-Seeker", "Insomniac") for quick behavior setup.
• Memory Management: Dual memory system (_memory/) with long-term and short-term storage for learning persistence.
• Examples and Tools: Example squids, configuration files (config.ini), and version tracking.

---

A year ago I got a tattoo of this project to celebrate its first development milestone!

---