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

Dosidicus electronicus

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
• A unique intersection of 1990s retro-gaming aesthetic and modern computational neuroscience.
• Build-your-own neural network - learn how an NN works by raising one as a pet

Compiled binaries for Windows, Mac and Linux: see Releases page

curl -sSL https://raw.githubusercontent.com/ViciousSquid/Dosidicus/2.6.2.0_LatestVersion/linux_setup.sh | bash

---

Manifesto | Wiki | Changelog

---

Myth & Mechanism

Dosidicus is a digital squid born with a randomly wired brain.

Feed him, stimulate neurons, watch him learn.

• He starts with 8 neurons.
• He grows new structure via neurogenesis and rewires using Hebbian learning
• He forms memories.
• He develops quirks.

Every squid is different. Every save file is a cognitive history.

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, stimulatable, understandable.

The squid serves as a digital pioneer in our quest to understand the mechanisms of thought and the evolution of autonomy in a synthetic world.

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.

---

Project 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!

---