Researchers have “unleashed” two AIs in Minecraft… and the results are not good

With a community of over 100 million active players, Minecraft Still an extraordinary game today. A team of researchers from the University of the Witwatersrand (South Africa) wondered whether AI would be able to cope adequately in this type of sandbox game. So the researchers developed the MinePlanner model (a model developed specifically for Minecraft) and two AI units powered by machine learning (Machine learning), then released the two entities into the Minecraft game.

The two AI entities had two objectives to accomplish in the virtual world of Minecraft, namely to build 45 structures according to three levels of complexity: easy, intermediate, and difficult. Easy for the average player, but extremely complex for AI given the enormous possibilities of the game, knowing that AI comes into Minecraft with minimal prior knowledge. Everything must be analyzed on site, after each trial and error allows the AI ​​to move a little further towards the task at hand (this is the principle of machine learning).

Ultimately, the AIs failed to fulfill the objective assigned to them. One of the AIs completed 17 constructs, none of which were difficult, while the other AI performed with 6 constructs, of which 5 were easy and one was intermediate. Surprisingly, only this last AI failed to the point of no longer being able to “think” because its memory was so saturated with data. In short, a truly intelligent enabling AI “from scratch” (or almost), isn’t really happening right away…