If there is a reason why Super Mario Bros. is remembered (and played) almost forty years after its release, it is none other than its wonderful level design. Japanese creative Shigeru Miyamoto demonstrated unprecedented creativity in creating stages filled with pipes, obstacles, and enemies for the player to overcome. That good work when it comes to designing fun levels has been one of Nintendo’s hallmarks ever since, and it’s the same thing that MarioGPT is now trying to emulate, an artificial intelligence capable of generating levels like there was no tomorrow.
Read also Albert Garcia
Using OpenAI’s popular GPT-2 tool, a team of researchers at the University of Copenhagen have trained an artificial intelligence model that can create classic Super Mario Bros. levels in seconds using only simple natural language commands.
In addition to publishing the corresponding paper and sharing a GitHub page where they explain the entire process more schematically, the team led by researcher Shyam Sudhakaran has also enabled a demo that everyone can use.
How to create levels with MarioGPT?
As in other popular tools such as ChatGPT or DALL-E 2, if you want to create a level, you only need to enter a text entry (‘prompt’) that the AI ββcan interpret. In this case, it is recommended to use a not too complicated phrase that refers to the number of pipes, enemies and blocks that we want the level to have.
To make things a bit easier, the aforementioned demo allows you to choose between different preset variables to set up a level without too much complication. For now, the levels are quite basic β not at all reaching the brutal creations that humans are capable of with the Mario Maker editor β but the majority (88%, according to its researchers) are functional and playable phases.
Currently, this experiment is only capable of generating levels with the same enemy. This is due to the way this model interprets the levels, which is by converting the different elements of the stage and characters into ASCII alphanumeric characters. All enemies are defined with the same letter “E” and, therefore, does not differentiate between the different types.
It is also interesting to note that MarioGPT has been based on the GPT-2 model and not its most recent and complete version GPT-3. As its creators have explained, the reason for this decision is that it is an easier system to train. In fact, the dataset used to teach the machine to create levels consisted solely of dumping levels from Super Mario Bros. and its Japanese sequel Super Mario Bros. 2.
It is not the first time that AI and Super Mario have crossed paths. For a few years now, researchers in this field have been using the Nintendo classic to teach artificial intelligence to learn to play autonomously through self-learning algorithms.
read also
Video game
Albert Garcia
Albert Garcia
Albert Garcia