Harder, Better, Faster, Stronger, Real-time! We are excited to reveal Genie 3, our most capable real-time foundational world model. Fantastic cross-team effort led by @jparkerholder and @shlomifruchter. Below some interactive worlds and capabilities that were highlights for me 🌎👇
Genie 3 generates minutes of consistent worlds in real-time at 720p! https://t.co/NEQqKLaXFs
In contrast to Genie 2, Genie 3 can generate real-world environments. I love how it adapts to different lighting conditions in this one. https://t.co/5XCmk9oQzX
Memory works really well! https://t.co/eTXOIKpzRL
Actions are persistent in the world. https://t.co/Wk1mrp2LyT
Major innovation over Genie 2: text events. "Another skier appears wearing a Genie 3 tshirt" https://t.co/5UmKQC8aRw
Let's explore Knosses. All the content, for example when turning 180° or walking through a door, is generated in real-time end-to-end on the fly. https://t.co/q8BCPy6260
When we started working on Genie 1 three years ago, the goal was always to one day generate an endless variety of environments for training embodied AGI. I believe the capabilities of our world models are getting there now. Next step: agent self-improvement loops! https://t.co/vPuvViuOF4
Here is @GoogleDeepMind's Sima agent following different instructions inside a world generated by Genie 3. https://t.co/prGpc0Q5se
@GoogleDeepMind More details and examples at https://t.co/JmAS3r0bjT
