OpenAI’s latest AI learned to play Minecraft by watching 40,000 hours of YouTube.
The company, which specialises in artificial intelligence, has developed a video pre-training (VPT) algorithm that allows its AI to learn through viewing.
Unlike previous attempts at creating Minecraft algorithms, OpenAI’s plays in the same environment as humans using the equivalent of mouse and keyboard actions, making this a huge leap in AI technology.
Out of the box, the AI learned actions such as chopping trees, making planks, and building crafting tables, as well as swimming, hunting, cooking and more.
With some tweaking, the AI eventually developed its productivity and was able to learn to build a diamond pickaxe – a skill that would take a human player around 20 minutes and 24,000 actions to accomplish.
Raw footage isn’t enough for the AI to learn, however. That’s because video alone shows what can be done, but not how to do it.
As such, contractors recorded and labelled basic Minecraft skills to create another algorithm that accurately annotated those 70,000 hours of YouTube footage with keyboard and mouse input prompts.
This allowed the AI to learn by watching human actions, rather than testing random actions, and predict actions taken using past and future information.
While the team only tested the AI with Minecraft, they believe the open-ended nature of the game and the generic human interface (mouse and keyboard) mean the AI could be used for general computer usage too.
You can learn more about OpenAI’s latest work in this blog post.