An AI Was Trained To Play Minecraft With 70,000 Hours Of YouTube Videos
OpenAI, the artificial intelligence research organization founded by Elon Musk, has trained an AI to play Minecraft almost as well as humans. It only took about 70,000 hours of binging YouTube videos.
A blog post detailing the feat reveals that researchers used a technique called "Video PreTraining (VPT)" to train a neural network on how to play Minecraft. This involved gathering 2,000 hours of sample dataset from actual humans playing Minecraft to include not just the raw video, but also exact keypresses and mouse movements.
From there, the researchers trained an inverse dynamics model (IDM) to predict the future action being taken at each step in the videos. Finally, the "trained" IDM was shown 70,000 hours of Minecraft YouTube videos. Afterwards, the model was able to copy the behaviors from the videos including chopping down trees to collect logs and turning the logs into an actual crafting table.
That's not all the AI model was capable of doing. It could also perform other complex activities such as swimming, hunting, and eating. The AI could even do "pillar jumping" which involves keeping yourself in the air by repeatedly jumping and placing a block underneath yourself.
The researchers were able to create a "foundation model" based on the data given to the AI and from there fine-tune behaviors or learn new one. The AI was noted to be able to perform early game skills such as building wooden and stone tools, raiding chests, and even build shelters.
The more hours of data that are fed into the foundation model, the more able the AI was. The researchers even used "reinforcement learning" to "reward" the AI for performing especially hard challenges. The reward system was used so effectively, the model was able to
Read more on ign.com
