Genie 3 by Google DeepMind  – AI News – #3 August 2025

Imagine you could describe any world you can think of – from the crowded canals of Venice to a surreal realm with flying mountains – and then… simply step into it. Move around, explore it, and even change its rules on the fly. Sounds like a promise from the borderlands of video games and science fiction? Google DeepMind has just opened the door to such a future by presenting Genie 3 – an AI model that generates interactive, playable worlds based on a simple text description. And not as a static image or video, but as a dynamic simulation that responds to our actions in real time.

This is the moment when generative artificial intelligence stops being just a passive content creator and becomes the architect of entire experiences. Let’s take a closer look at what this new “genie” can do and why it might be one of the most important AI launches of the year.

In the simplest terms, Genie 3 is a so-called world model. This is not just another video generator like Veo or Sora. Its primary goal is not to create a perfect cinematic clip, but to generate a coherent environment in which the user can move. Think of it as a game engine being created live, right before your eyes, based on a few sentences of prompt.

Google states that Genie 3 can generate dynamic worlds at 720p resolution at 24 frames per second while maintaining consistency for several minutes of interaction. This means smooth movement through the created space, which “remembers” what is where, even if it briefly disappears from view.

Genie 3 didn’t come out of nowhere. It is the culmination of more than a decade of research by Google DeepMind on simulated environments. They started by training AI agents to master strategic games and later expanded to developing virtual worlds for robotics and open-ended machine learning research.

World models like Genie are seen as a key step towards creating artificial general intelligence (AGI). Why? Because they give AI agents an almost infinite curriculum. Instead of being limited to real-world data, AI can learn through millions of diverse, simulated scenarios, testing the consequences of its actions in a safe environment. Genie 3 is a direct successor to Genie 1 and Genie 2 models but introduces a fundamental novelty: real-time interaction combined with a much higher level of realism and consistency.

Examples published by Google best demonstrate Genie 3’s versatility. This is not a tool limited to one style or theme. It is a true chameleon.

One of the most impressive aspects is the model’s ability to simulate basic laws of physics and natural phenomena. In demonstrations, we see the perspective of a wheeled rover overcoming volcanic terrain, where tires dig into blackened earth while smoke rises in the distance and lava flows. Another example is riding a jet ski during a light festival or walking along the Florida coast during a hurricane, where massive waves flood the road and palm trees bend in the wind. Water, lighting, and environmental interactions look surprisingly natural.

Genie 3 can create not only still nature but entire vibrant ecosystems. From running along the edge of a glacial lake, passing wild animals along the way, to diving into deep ocean waters among schools of jellyfish, to a precisely designed Japanese zen garden. The model understands how individual elements – vegetation, animals, water, light – should interact to form a believable whole.

This is where Genie 3 really shows its true claw. The model is not limited to realism. Want to become a lizard in a world made of origami? Here you go. Or maybe you prefer flying as a firefly through a magical forest with treehouse homes? No problem. One of the most extraordinary examples is a landscape of Irish hills suddenly tearing apart, with fragments floating into the sky to form surreal, brutalist architecture with waterfalls cascading from suspended lakes. This proves that the only limitation is imagination.

Want to see how the palace of Knossos in Crete looked in its heyday? Or take a ride on a water tram through the canals of Venice, observing ancient buildings and other boats? Genie 3 allows crossing geographical and temporal boundaries, offering unique opportunities to explore historical places and distant corners of the world.

Achieving such a high level of control and real-time interactivity required significant technical breakthroughs.

Generating a coherent environment frame by frame is harder than generating a finished video. In videos, errors are static. In interactive simulations, inaccuracies can accumulate, leading to the breakdown of the illusion. Genie 3 generates each new frame considering the entire trajectory of the user’s movement so far. If after a minute you return to the same place, the model must “remember” how it looked. Interestingly, this consistency is an emergent ability – it does not come from creating an explicit 3D representation (like NeRF or Gaussian Splatting), but from the generation process itself. This makes the worlds much more dynamic and rich.

This is one of the most interesting features. Besides moving around the world, Genie 3 allows modifying it with text commands. This is called “promptable world events”. For example, you can change the weather, add new objects or characters. This feature greatly expands exploration possibilities and is invaluable for training AI agents, allowing to test “what if…” scenarios.

To test the usefulness of its worlds, Google “released” one of its AI agents – SIMA – into them. The agent received specific goals (e.g., “approach the red tree”) and independently sent navigation commands to Genie 3 to achieve them. Thanks to the consistency of generated environments, SIMA could perform longer and more complex action sequences. Such simulations are meant to accelerate the development of agents that in the future will be able to operate not only in virtual but also in the real world.

Despite its enormous potential, Genie 3 has its limitations. The range of actions an agent can perform is still limited. The model struggles with simulating interactions between many independent agents, and generated locations do not have perfect geographical accuracy. Rendering readable text remains a challenge. Currently, interaction is possible for several minutes, not many hours.

Google also emphasizes its commitment to responsible development. Open, interactive models bring new safety challenges. Therefore, Genie 3 is currently available only within a limited research program for a select group of scientists and creators. This approach allows collecting feedback and better understanding potential risks.

Genie 3 is a milestone. It marks the moment when world models start leaving research labs and knocking on the doors of creators, educators, and engineers. The potential applications are vast: from revolutionary tools for creating games and films, through simulators for training surgeons or pilots, to advanced platforms for testing autonomous vehicles and robots.

We are witnessing the birth of a new form of media – interactive media created on demand. There is still a long way to go before each of us can create our own photorealistic and fully interactive world for many hours of fun. But the genie has been let out of the bottle and cannot be put back. And what it has shown us is just the beginning – if you want to follow its progress with us, sign up for the Delante newsletter!

Information source about Genie 3: https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/