OpenAI’s New AI Learned to Play Minecraft by Watching 70k Hours of YouTube
singularityhub.com
singularityhub.com
> Crafting a diamond pickaxe requires a long and complicated sequence of subtasks. To make this task tractable, we reward agents for each item in the sequence.
but a lot of these steps are standard Advancements in the game anyway[1] which I think is fair to plug into the AI's reward signal.
There are two possible extensions to this research which I think would be interesting from both an AI and a game design perspective. The first is seeing if the AI could learn that unlocking new crafting recipes was an instrumental goal that correlates well with rewards, and seeing if this was enough for the AI to "complete" the game (given a few video examples of building a Nether portal and killing the Ender dragon, etc.)
A second extension would be to train an AI to describe what it is seeing in the game, and getting an AI like GPT-3 to describe which steps it should take to reach the next Advancement, then getting the AI to guess which in-game actions are most likely to implement those steps. That would be something like an internal monologue that a player has when setting themselves a game plan, and would benefit greatly from all the Minecraft guides (or equivalent knowledge) that presumably exist in GPT-3's brain somewhere.