DeepMind has once again appeared to push the frontier of AI with the debut of Genie 3. Unlike traditional LLMs, Genie 3 doesn’t just respond to text. It builds interactive 3D worlds on the fly, capable of evolving in real time with prompt-driven interventions.

Here’s what sets Genie 3 apart:

• Generates 720p 3D environments at 24fps in real time from pure text prompts
• Maintains scene memory objects remembered, interactions preserved, enabling storytelling and simulation continuity
• Introduces “promptable world events”, where users or AI agents can alter the simulation on the fly: change the weather, add characters, create dynamic physics scenarios all without breaking the simulation loop

Why this matters‍

Genie 3 is the most advanced world model ever built, laying the foundation for embodied AI agents that don’t just answer questions, but live inside rich, interactive environments.

It’s a major step toward Artificial General Intelligence (AGI), offering a testbed for agents to learn, act, and adapt in sandboxed simulations resembling the real world. Currently offered as a limited research preview to select partners and universities, Genie 3 positions DeepMind (and by extension Google) as a leader in next-generation simulation and AGI infrastructure.

The GoML POV

DeepMind's Genie 3 represents a leap forward in the development of "world models" and, more broadly, a critical step towards Artificial General Intelligence (AGI). By creating real-time, interactive 3D environments with a consistent visual memory and the ability to generate "on-the-fly" events, DeepMind is moving beyond static video generation and into the realm of dynamic, playable simulations.

This technology's most profound impact is its potential to serve as a training ground for embodied AI agents. Training robots and autonomous systems in the physical world is costly, slow, and dangerous. Genie 3 provides a boundless, safe, and dynamic virtual sandbox where these agents can learn, explore, and reason about cause and effect in a realistic but controlled environment. The ability to dynamically prompt events, like a sudden rainstorm or the introduction of a new object, allows for the creation of an infinite curriculum of challenges.

However, it is currently just a research preview. It remains to be seen how it performs it becomes a general purpose model accessible to builders and designers.