Particle.news

Download on the App Store

Google DeepMind Debuts Genie 3 With One-Minute Memory for Interactive 3D Worlds

The private preview delivers dynamic 720p environments at 24fps that researchers can use for AI agent training.

Overview

  • Genie 3 autonomously extended its visual memory from around ten seconds to roughly one minute without developer intervention.
  • The model generates interactive, coherent 3D worlds from text prompts that persist for the full memory duration.
  • Access is limited to invited researchers and creators who are supplying feedback to guide further improvements.
  • Google DeepMind positions Genie 3 as a foundational tool for training AI agents on the path toward artificial general intelligence.
  • Technical constraints remain, with restricted agent action spaces and capped interaction durations in the generated simulations.