AI.Planet is the kind of old software project that makes modern AI branding look oddly narrow. Its goal was not a chatbot or an API. It was a whole artificial planet where weather, water, plants, animals and simple learning systems could interact inside a 3D environment.
The project appeared on SourceForge in the early 2000s and described itself as a virtual world for artificial intelligence. Users could modify terrain, add water, suns and moons, change the atmosphere, populate the world with plants and creatures, and then watch what happened.
The software was primarily written in Delphi using GLScene. Its old requirements page is a period piece by itself: Pentium III 800 MHz or faster, 256 MB of RAM, an OpenGL graphics card around the GeForce2 generation, and Windows 98, 2000 or XP.
The creatures actually had things to do
The changelog documents a surprisingly ambitious simulation. Versions added clouds, mountains, water, physics, creatures, tides from moons, wind from temperature changes, icebergs, snow, ducks, tigers, dolphins, fish mating and predator-prey relationships.
One release introduced turtles that remembered what happened when they attempted to eat different objects. The project reported individual turtles developing different food preferences based on their experiences. One preferred oranges, another fish, another crabs. Apparently even fake turtles can become picky.
The world could be explored from space, on foot, by following creatures or through a controllable robot. Later builds added visual object trackers and tools for inspecting the population.
Open source planetary tinkering
AI.Planet was released under the GPL, and the site’s tone is gloriously from the SourceForge era: if the fake planet does not do what you want, you have the source code. Add it yourself.
The project reached version 1.1.x after a long sequence of small releases. The changelog shows how software like this often evolved at the time: somebody adds a duck, somebody else adds a tiger, then someone realizes the icebergs are multiplying too quickly and fixes that before moving on to fire trees.
It is easy to look at a 2003 graphics engine and see obsolete technology. The more interesting view is that AI.Planet treated artificial intelligence as behavior embedded in a world. Learning meant an organism sensing, remembering and acting inside an environment rather than producing text in a browser window.
That makes the site useful again in 2026, even if the executable itself belongs to another hardware era. It preserves a different imagination of what home-computer AI experimentation could look like.
Explore the surviving AI.Planet site and its wonderfully detailed release history.