Inspiration

  • the future of computing must mix centralization and distribution
  • there's no good multiplayer world model
  • the testbed of ai are video games

What it does

  • multiplayer world model: two players play, interact, and modify one shared world
  • distributed inference: optimized latent state generation on b200, optimized decoding on mac

How we built it

  • overnight post-trained oasis single player Minecraft model with gameplay from two Minecraft agents
  • flow megakernel optimized for b200, metal kernel optimizations

Challenges we ran into

  • not getting high MFU initially during training
  • wifi was slow for sending latent packets (using hotspot instead)
  • models were slow, especially on apple MLX

Accomplishments that we're proud of

  • 10x throughput on b200
  • intelligence design

Built With

  • ai
  • b200
  • codex
  • llm
Share this project:

Updates