415.tech
AI & tech, from the frontlines of Silicon Valley
Sakana AI and NYU train transformers on billions of cubes to generate playable Minecraft worlds

Sakana AI and NYU train transformers on billions of cubes to generate playable Minecraft worlds

Sakana AI and New York University released Dream-Cubed, which treats Minecraft cubes as tokens so transformers trained on tens of billions of them produce editable, playable worlds. The paper, code, and dataset are all public, making the compositional-token route to 3D generation reproducible today — with inpainting, outpainting, and block-level control over infinitely sized maps.

Source: sakana.ai ↗

Post on XEmail

Using cubes as tokens allows large transformers to do the same.

Sakana AI

Why this matters

  • → Transformer models now generate playable 3D worlds at scale, not just images or text.
  • → Discrete tokens (cubes) unlock the same compositional training that powers language models.
  • → Full dataset and code are public—reproducible path to generative 3D, today.
Cubes as language