AI learns to play Minecraft after watching 70 thousand hours of video on the Internet: how did it do it?

The AI of OpenAI has learned to play Minecraft after training with 70 thousand hours of video on the Internet. The company indicated that artificial intelligence is capable of doing all the activities of the video game and that it even performs very well.

“With fine tuning, our model can learn to make diamond tools, a task that normally takes competent humans more than 20 minutes (24,000 actions). Our model uses the native human interface of keystrokes and mouse movements, which makes it quite general and represents a step towards general agents using computers.”, assured OpenAI in the article.

“The Internet contains a huge amount of available and public videos, from which we can learn. You can watch a person make a beautiful presentation, a digital artist draw a beautiful sunset, and a Minecraft player build an intricate house. Nevertheless, these videos only provide a record of what happened, but not exactly how it was accomplished, i.e. [la IA] you won’t know the exact sequence of mouse movements and keys pressed”, he added.

LOOK: Parti, the new Google model that generates photorealistic images from text

That is, the AI ​​could not learn which keys to use to perform the actions, so the developers had to pre-train it. “In order to utilize the vast amount of untagged video data available on the Internet, we present a novel, but simple, semi-supervised imitation learning method: Video PreTraining (VPT). We start by collecting a small data set from contractors where we record not only their video, but also the actions they performed, which in our case are keystrokes and mouse movements”, he indicated.

“With this data we train an inverse dynamics model (IDM), which predicts the action being taken at each step of the video. Importantly, the IDM can use past and future information to guess the action at each step. This task is much easier and thus requires much less data than the behavioral cloning task of predicting actions from past video frames alone, which requires inferring what the person wants to do and how to achieve it. Later we can use the trained IDM to tag a much larger dataset of online videos and learn to act through behavioral cloning”, added OpenAI.

LOOK: We tested DALL-E Mini, the AI ​​that “creates art”, and this is the curious result that it gave us when faced with phrases about Peru

This model generated that AI can perform the same actions as humans. “Trained with 70,000 hours of online video tagged by IDM, our behavioral cloning model (the “basic VPT model”) performs tasks in Minecraft that are almost impossible to accomplish with reinforcement learning from scratch. Learn how to cut down trees to collect logs, turn those logs into planks, and then turn those planks into a workbench; this sequence takes a human skilled in Minecraft approximately 50 seconds or 1,000 consecutive game actions”, he assured.

Finally, in its conclusions, the company believes that this test can be a step for AI to learn to use computers on their own. “While we only experienced Minecraft, the game is very open and the native human interface (mouse and keyboard) is very generic, so we believe that our results bode well for other similar domains, for example, computer use”, he concluded.

We trained a neural network to competently play Minecraft by pre-training on a large unlabeled video dataset of human Minecraft play and a small amount of labeled contractor data. https://t.co/a2pyBqvLvg pic.twitter.com/XbqtwQSTwU

—OpenAI (@OpenAI) June 23, 2022

Source: Elcomercio

James

I have worked as a journalist for over 10 years and have written for various news outlets. I currently work as an author at 24 News Recorder, mostly covering entertainment news. I have a keen interest in the industry and enjoy writing about the latest news and gossip. I am also a member of the National Association of Journalists.