Based on OpenAI’s GPT-2 technology, this solution has so far only been powered by data from the game “Super Mario Bros: The Lost Levels”, released in 1986. Despite its retro look, the result is quite impressive.

Playfully named MarioGPT, this AI is still in its infancy and works on explicit written prompts. Indeed, you have to detail what you want to see in the levels that will be created.

Instructions must be precise, such as “many pipes, many enemies, few blocks, low elevation”, or conversely “few pipes, few enemies, many blocks, high elevation”, etc.

Using MarioGPT is not as easy and accessible as using ChatGPT, as you need to have some knowledge of code (Python 3.8). That’s why its creators have put a dedicated guide on the GitHub platform to help anyone interested get started at creating their own game levels.

According to its developers, MarioGPT is already capable of generating diverse and mostly playable environments, and therefore of improving the quality of the game. When trained on more complex games and based on even newer language models than GPT-2, this type of solution could one day revolutionise the entertainment industry.

It would then be hypothetically possible to enrich the content of these games and tailor them according to the players’ own wishes.