Google’s AI can create… interactive worlds: find out more about the possibilities of Project Genie

Initially, the possibility arose to generate images from a simple text instruction. Subsequently, it became easier to create video clips simply by requesting it from an AI. At the same time, artificial intelligence models honed their ability to generate code at increasingly complex levels.

Now, AI is capable of generating worlds. At least, virtual worlds, of course 🙂

Project Genie, the latest evolution of Google’s AI

Project Genie represents a new milestone in Google DeepMind’s ambition to achieve Artificial General Intelligence. Based on the powerful Genie 3 world model, this experimental prototype is capable of simulating dynamic environments that respond in real time to user actions.

It is as if one could create an open world from a video game using pure natural language.

Project Genie is a prototype web app powered by Genie 3, Nano Banana Pro + Gemini that lets you create your own interactive worlds. I have been playing around with it a bit and it is…out of this world:)

Rolling out now for US Ultra subscribers. pic.twitter.com/rNDXn3VUF6

— Sundar Pichai (@sundarpichai) January 29, 2026

What is a “world model”?

Unlike traditional AI systems that predict the next word or pixel, a world model like Genie 3 simulates the physics and interactions within an environment. This ability allows it to predict how a scenario will evolve and how actions will affect it. While Google DeepMind has already mastered closed environments such as chess or Go, Genie 3 has been designed to navigate the diversity of the real world, ranging from robotics to the recreation of historical settings.

Three pillars for creating realities

Project Genie is a web application powered by the combined capabilities of Genie 3, Nano Banana Pro, and Gemini and relies on three main features:

• World Sketching: By using text or uploaded images, you may define your character and the landscape. Thanks to its integration with Nano Banana Pro, it is possible to preview and adjust the scene, and even select the perspective (first- or third-person) prior to entering.
• World Exploration: Once generated, the world is fully navigable, much like entering Roblox or Minecraft. As you progress, the AI generates the path in real time based on your actions and also allows you to adjust the camera while moving. Additionally (and this temporal consistency capability is more impressive than it may appear), you can retrace your steps and return along your previous path. The world does not simply vanish as you leave it behind—much like in a video game.
• World Remixing: Users may take existing worlds and reinterpret them by modifying their initial instructions through natural language, from the features of the world to the appearance of characters, or draw inspiration from a gallery of options to build new experiences. After finishing, they are able to download videos of their worlds and explorations.

A responsible path toward AGI

The relationship between advances like Project Genie and Artificial General Intelligence (AGI) is not direct, but it does suggest a path to follow. Project Genie is not AGI, nor does it aim to be, but it addresses one of the fundamental challenges that any truly general intelligence must resolve: understanding how a world operates and anticipating the consequences of its actions. Until now, most AI systems have focused on describing reality based on statistical patterns. World models, by contrast, seek to achieve something different: to simulate causality, maintain coherence over time, and respond in a consistent manner to interaction.

Google DeepMind has spent years exploring this approach, as training AI directly in the physical world is costly, slow, and, in many cases, unfeasible. Simulated environments provide a safe and scalable space to experiment, fail, and learn without real-world consequences. Project Genie fits within this strategy as a laboratory of worlds where future systems could be trained before facing real tasks, ranging from robotics to complex decision-making.

Nevertheless, it is prudent to exercise caution. The worlds generated by Project Genie are limited, short-lived, and lack the semantic richness, long-term memory, and autonomous objectives that AGI would require. The system does not understand the world as a human being would: it simulates it. However, even such a simulation represents an important step, because without an internal model of its environment, any intelligence—no matter how advanced in language or reasoning—will remain incapable of acting autonomously and coherently.

Despite its potential, Google emphasizes that it is still a research prototype. As with any emerging technology, there are limitations: visualizations may not be entirely faithful to reality, characters may display latency in control, and the generations are currently limited to 60 seconds.

Who can use Project Genie

At present, access to Project Genie is being rolled out to Google AI Ultra subscribers in the United States (aged 18 and older). The ultimate goal is ambitious: to use these environments to train future AI systems in safe and diverse scenarios, bringing us a step closer to intelligence that is capable of understanding and acting within our physical world.

Image: Google