This is really cool! I tried to do a similar experiment myself to get Bing chat to play Bitburner (not minecraft) a while back, but I didn't have the sort of organized prompts or iterative prompting that they set up in that paper, and with the response limit per conversation for the Bing chat, it did not go well.Blitz4 wrote: ↑Sat Jan 06, 2024 8:20 am...
Allow me to show you one extreme example to that theory. Could be the future, who knows. So nVidia, they wrote a paper for an AI that would beat Minecraft, but more than that. It wrote instructions to perform the tasks in the game that are required to beat Minecraft as fast a possible. It improved those instructions through a variety of ways, primarily by learning & testing over many iterations. Cool right. It gets better. This was crated using ChatGPT4. Doing so allows the AI to receive instructions from the player. The AI can be playing as if they were the player at every moment to decide what's the most efficient path to the end goal (beating the game), but not share what they're thinking to the player. (I just said AI is thinking, oh boy). What's cool is if the player needs help at any time, they can say, "What's the best thing for me todo right now?" Or "What's the fastest way to build a boat from where I'm at?"
Video covering it in good detail "Minecraft AI - NVIDIA uses GPT-4 to create a SELF-IMPROVING autonomous agent.":
https://www.youtube.com/watch?v=7yI4yfYftfM
...
What's really interesting to me is how the skill library in their agent is so similar to blueprints in Factorio. Instead of prompting it for every action, they prompt it to create skills, iterate on them until they work properly, then store them to be used at later times. As it progresses, it abstracts groups of smaller tasks into one task, allowing it to manage larger tasks without getting bogged down with those smaller tasks.