Minecraft Could Make AI Perfect In a Fascinating Experiment

By Jason Collins | Updated


The development of Artificial Intelligence implies the advancements in neural networks, which are powerful tools for modeling complex patterns in data, thus enabling machines to learn from experiences and perform tasks that would be difficult or near impossible to program explicitly. And it’s no secret that video games have been used to develop AI and neural networks. Now, a team of researchers at the University of the Witwatersrand in South Africa is using Minecraft to measure the general intelligence of AI models.

Minecraft Is Being Used To Teach AI


This fascinating experience aims to measure an AI’s ability to ignore unimportant details while solving a complex problem with multiple steps. And while current AI models are great at performing the majority of tasks they’re given, they struggle to ignore information, even when explicitly told to do so. If you tell a generative AI to create a picture of a room without the elephant, current models will comply and output a picture of a room with the elephant in it. They’re not great at ignoring information, so they’re given Minecraft to play.

Teaching AI What’s Important

The AI would have to solve Minecraft’s pre-set in-game problems by analyzing the landscape and extraneous objects and deciding which details are necessary for task completion and which should be ignored. Though it sounds relatively simple to a human, this is a major problem of nearly all current AI models. When given a task and a set of information that might or might not be necessary for task completion, the AI will try to use all the information at its disposal. However, some of that information is irrelevant or useless, which might lead to AI hallucinations.

Minecraft Is So Big It’s Confusing

How does it connect to Minecraft? Well, Minecraft is a sandbox game, and as the term implies, you’re only limited by your imagination, as there aren’t any set goals, and the game can be played as you see fit. This abundance of choice often confuses people, let alone machines that operate on a yes-and-no basis, with no “maybe” in the middle. The new experiment is a benchmark test for current AI models, and its results could pave the way for breakthroughs in artificial general intelligence—the artificial intelligence we’ve all seen in movies.

Advanced AI Can’t Beat Medium Difficulty Minecraft Challenges

The experiment consists of 15 construction problems, each with an easy, medium, and hard setting, which total 45 tasks. To complete each task, the AI must take certain steps towards completion, which also involves planning. Sadly, even the most advanced models currently developed have been able to get past the medium settings stage. This isn’t the first time researchers have used video games to benchmark or train an AI model. OpenAI also used Minecraft to train its GPT model.

Minecraft Isn’t The First Video Game To Help AI


Perhaps the most famous example of AI playing games is the work done by DeepMind, whose AI system, Deep Q-Network, played through various Atari 2600 video games, perfecting the gameplay to a level that’s beyond human capabilities. Another example is AlphaStar, an AI that achieved a grandmaster level in StarCraft, playing against human players. And now, we have AI models playing Minecraft and learning how to solve issues without receiving specific instructions.  

Source: Arxiv