DeepMind's Genie 3 builds playable worlds on the fly

Imagine giving an AI a one-line text prompt, and instead of spitting out a static video, it instantly generates an interactive 3D world you can walk through in real time.

Google DeepMind just unveiled Genie 3, a general-purpose foundation world model capable of spinning up playable environments on the fly. Users can explore these worlds at 720p resolution and 24 frames per second, with physical consistency holding up for multiple minutes.

Stepping inside the frame

Compared to Genie 2, which could only hold a scene together for 10 to 20 seconds, Genie 3 is a massive leap forward. It features an internal visual memory extending back at least a minute, meaning if you turn around and revisit an area you passed earlier, the scenery remains consistent rather than hallucinating into something else entirely.

Under the hood, Genie 3 does not rely on a hand-coded game engine, NeRFs, or 3D meshes. DeepMind research director Shlomi Fruchter explained that the model is fully auto-regressive, predicting each frame based on past trajectories and real-time user inputs. DeepMind researchers claim the model effectively taught itself intuitive physics through scale and memory.

The system also supports promptable world events. While wandering around, you can type a command mid-session to alter the weather or inject a new obstacle. DeepMind researcher Jakob Bauer even showcased taking a pre-rendered drone video generated by Veo 3 and seizing control mid-flight to explore the scene interactively.

The ultimate sandbox for embodied AI

While the gaming and creative potential is obvious, DeepMind is chasing a much bigger prize: training embodied agents for the path toward AGI.

To test this, DeepMind plugged in a recent build of SIMA, its generalist agent for 3D environments. In a Genie 3-generated warehouse, SIMA received instructions like approaching a trash compactor or walking toward a forklift. Without knowing the agent's internal goals, Genie 3 reliably simulated the physics of the environment in response to SIMA's actions, allowing it to complete sequential goals.

Still, Genie 3 remains a limited research preview with obvious constraints. TechCrunch pointed out that a downhill skiing demo failed to accurately simulate snow displacement. Furthermore, the action space for agents is still narrow, multi-agent interactions remain unsolved, text rendering is messy, and a runtime of a few minutes falls drastically short of the hours required for serious agent training.

Why it matters

We think Genie 3 marks a pivotal shift in generative AI, transforming world models from passive video playback into interactive, responsive simulation engines. Generating dynamic 3D worlds at 24 frames per second without hardcoded physics proves that neural networks can learn functional spatial awareness entirely through visual autoregression.

For the robotics and embodied AI sector, this could eventually shatter the physical data bottleneck. Collecting edge-case physical data in real facilities is slow, dangerous, and expensive. If foundation world models can generate infinite counterfactual training scenarios on demand, embodied agents might finally get their own AlphaGo-style self-play training loop.

That said, it is not something everyday creators or developers can build on today. It is locked behind a tight research preview, and sustaining coherence for only a few minutes means commercial game engines are not getting replaced anytime soon. But as a technical milestone, it proves that the bridge between generative media and physical reasoning is becoming very real.