Genie 3
World models · 2025
The idea
Section titled “The idea”The JEPA and Dreamer pages predict a state so a planner or policy can act. Genie asks a different question: can the world model be the environment - a place you walk around in, generated frame by frame?
Genie 3 (Parker-Holder, Fruchter et al., DeepMind, 5 Aug 2025) is the first in that line DeepMind calls real-time. Give it a text prompt; navigate the resulting world at 24 frames per second, 720p, with consistency DeepMind reports for a few minutes. That is the jump from Genie 2 (short, not real-time) and from Genie 1 (2024: playable 2D worlds, latent actions learned from unlabeled video).
Why minutes matter: a video model that forgets what was behind you the moment you turn is a clip generator. One that keeps the trees on the left of the building when they leave and re-enter the frame is starting to behave like a place. DeepMind calls that consistency emergent - not an explicit NeRF or Gaussian splat you handed the model, but autoregressive frames that look back at the growing trajectory. Looking away and returning after about a minute is the memory window they advertise.
Promptable world events sit beside navigation: change the weather, add an object, mid-session, by text. Useful as counterfactuals for agents; DeepMind is explicit that these interventions are not always actions the agent can take.
They also ran a SIMA agent inside Genie-generated worlds. The world model does not know the agent’s goal; it only simulates the next frame from the actions. That is the intended use: unlimited-looking curricula for embodied agents - with the honesty that this is still a research preview.
The misconception to name
Section titled “The misconception to name”It is tempting to treat a 24 fps demo as “we have a physics engine now.” The blog’s own limitations list is the corrective:
- a few minutes of continuous interaction, not hours
- limited action space for the agent itself
- multi-agent interaction is unsolved
- real-world locations are not geographically accurate
- text in the scene is unreliable unless you put it in the prompt
Learned physics here is regularity in video, not Newton’s laws. Judge it as a steppable gym and media tool, the way Three programs named this column - not as a replacement for MuJoCo.
Verdict
Section titled “Verdict”The story, animated
Section titled “The story, animated”Turn away, turn back - the same trees, the same building, still there. Then the honest limits.
Where it sits
Section titled “Where it sits”Worth having first: Three programs · Dreamer 4
Unlocks: NVIDIA Cosmos - the open-weights platform aimed at Physical AI data, not playable tourism.
Source-checked against Parker-Holder & Fruchter, “Genie 3: A new frontier for world models,” Google DeepMind, 5 Aug 2025 (capabilities and limitations list).