Experts and researchers at OpenAI have developed a neural network that is capable of playing Minecraft just as well as most humans. This sounds like a story that doesn’t have much of an impact on the real world, but it actually does. You see, the importance of this development doesn’t lie in the fact that AI can now play video games too, but rather in how it learned to play Minecraft.
Training the neural network to play the popular block-based sandbox game involved 70,000 hours of in-game footage, in addition to a limited database of specially curated videos depicting certain in-game tasks being performed.
The stroke of genius here is that the Minecraft AI Education Edition package that the neural network was presented with didn’t just contain videos of Minecraft gameplay – it also contained a log of keystrokes and mouse movements that create a roadmap of the actions necessary to perform certain tasks. This process is known as Video PreTraining (VPT).
For the longest time, we’ve been willing and able to show AI what we want it to do, but we’ve been far less forthcoming with how it was supposed to achieve that objective. It’s impressive when AI figures these problems out for itself, and it has, in the past, actually informed human research teams on how better to tackle certain problems. However, we can’t expect AI to behave in a manner we expect and understand if we don’t teach it to behave as we do.
After a brief period of fine-tuning, the neural network was able to play Minecraft just as well as most human players – and better than some. In addition to simple tasks such as punching trees to harvest wood, the AI demonstrated that it could successfully hunt animals and consume their meat. One of the most impressive discoveries was that the AI had mastered the pillar jump – a move that requires a player to place a block below themselves while jumping to elevate themselves without having to build a staircase.
This experiment, while conducted in Minecraft, has massive implications for the future of robotics and artificial intelligence. As it turns out, the key to creating AI that can perform more complicated tasks doesn’t just lie in instructing it to perform them, but rather showing it keystroke for keystroke how the objective is achieved.
I bet you didn’t know beforehand that AI could play Minecraft. In the spirit of things you may have not known about the game, here’s 15 cool things you might not know you can do in Minecraft!