Google DeepMind has revealed Genie 3, its newest basis world mannequin that can be utilized to coach general-purpose AI brokers, a functionality that the AI lab says makes for a vital stepping stone on the trail to “artificial general intelligence,” or human-like intelligence.
“Genie 3 is the first real-time interactive general purpose world model,” Shlomi Fruchter, a analysis director at DeepMind, mentioned throughout a press briefing. “It goes beyond narrow world models that existed before. It’s not specific to any particular environment. It can generate both photo-realistic and imaginary worlds, and everything in between.”
Nonetheless in analysis preview and never publicly obtainable, Genie 3 builds on each its predecessor Genie 2 (which may generate new environments for brokers) and DeepMind’s newest video era mannequin Veo 3 (which is claimed to have a deep understanding of physics).
With a easy textual content immediate, Genie 3 can generate a number of minutes of interactive 3D environments at 720p decision at 24 frames per second — a big soar from the ten to twenty seconds Genie 2 may produce. The mannequin additionally options “promptable world events,” or the flexibility to make use of a immediate to vary the generated world.
Maybe most significantly, Genie 3’s simulations keep bodily constant over time as a result of the mannequin can bear in mind what it beforehand generated — a functionality that DeepMind says its researchers didn’t explicitly program into the mannequin.
Fruchter mentioned that whereas Genie 3 has implications for academic experiences, gaming or prototyping artistic ideas, its actual unlock will manifest in coaching brokers for common objective duties, which he mentioned is crucial to reaching AGI.
“We think world models are key on the path to AGI, specifically for embodied agents, where simulating real world scenarios is particularly challenging,”Jack Parker-Holder, a analysis scientist on DeepMind’s open-endedness workforce, mentioned in the course of the briefing.
Techcrunch occasion
San Francisco
|
October 27-29, 2025
Genie 3 is supposedly designed to resolve that bottleneck. Like Veo, it doesn’t depend on a hard-coded physics engine; as a substitute, DeepMind says, the mannequin teaches itself how the world works – how objects transfer, fall, and work together – by remembering what it has generated and reasoning over very long time horizons.
“The model is auto-regressive, meaning it generates one frame at a time,” Fruchter informed TechCrunch in an interview. “It has to look back at what was generated before to decide what’s going to happen next. That’s a key part of the architecture.”
That reminiscence, the corporate says, lends to consistency in Genie 3’s simulated worlds, which in flip permits it to develop a grasp of physics, just like how people perceive {that a} glass teetering on the sting of a desk is about to fall, or that they need to duck to keep away from a falling object.
Notably, DeepMind says the mannequin additionally has the potential to push AI brokers to their limits — forcing them to be taught from their very own expertise, just like how people be taught in the actual world.
For example, DeepMind shared its take a look at of Genie 3 with a latest model of its generalist Scalable Instructable Multiworld Agent (SIMA), instructing it to pursue a set of objectives. In a warehouse setting, they requested the agent to carry out duties like “approach the bright green trash compactor” or “walk to the packed red forklift.”
“In all three cases, the SIMA agent is able to achieve the goal,” Parker-Holder mentioned. “It just receives the actions from the agent. So the agent takes the goal, sees the world simulated around it, and then takes the actions in the world. Genie 3 simulates forward, and the fact that it’s able to achieve it is because Genie 3 remains consistent.”
That mentioned, Genie 3 has its limitations. For instance, whereas the researchers declare it could actually perceive physics, the demo displaying a skier barreling down a mountain didn’t replicate how snow would transfer in relation to the skier.
Moreover, the vary of actions an agent can take is proscribed. For instance, the prompt-able world occasions permit for a variety of environmental interventions, however they’re not essentially carried out by the agent itself. And it’s nonetheless troublesome to precisely mannequin advanced interactions between a number of unbiased brokers in a shared atmosphere.
Genie 3 can even solely assist a couple of minutes of steady interplay, when hours could be needed for correct coaching.
Nonetheless, the mannequin presents a compelling step ahead in instructing brokers to transcend reacting to inputs, letting them doubtlessly plan, discover, search out uncertainty, and enhance by trial and error – the type of self-driven, embodied studying that many say is essential to transferring in direction of common intelligence.
“We haven’t really had a Move 37 moment for embodied agents yet, where they can actually take novel actions in the real world,” Parker-Holder mentioned, referring to the legendary second within the 2016 recreation of Go between DeepMind’s AI agent AlphaGo and world champion Lee Sedol, by which Alpha Go performed an unconventional and good transfer that grew to become symbolic of AI’s skill to find new methods past human understanding.
“But now, we can potentially usher in a new era,” he mentioned.
