Google Deepmind has revealed Genie 3, its latest World Foundation model that Ai Lab says it is presenting a crucial step on the way to artificial general intelligence or human intelligence.
“Genie 3 is the first real -time interactive general purpose model,” said Shlomi Fruchter, director of Deepmind, during a press briefing. “He passed the narrow models of the world that were previously. They are not specific in a specific environment. It can create both photorealistic and fantastic worlds and all intermediates.”
Genie 3, which is still previewing the research rather than the public, is based on both Genie 2’s predecessor – who can create new environments for agents – as well as the latest generation video of Deepmind Veo 3 – which shows a deep understanding of physics.
With a simple text prompt, Genie 3 can create multiple minutes – from 10 to 20 seconds in Genie 2 – from different, interactive, 3D environments in 24 frames per second with 720p resolution. The model also has “the right world events”, or the ability to use a prompt to change the world.
Perhaps most importantly, Genie 3 simulations remain naturally consistent over time, because the model is able to remember what it had previously created – an upset ability that Deepmind researchers are not explicitly planning on the model.
Fruchter said while Genie 3 clearly has an impact on educational experiences and new genetic media such as games or original creative concepts, his real unlocking will be manifested by educational agents for general purpose duties, which he said was necessary for AGI.
“We believe that the models of the world are decisive for AGI, especially for the built -in factors, where the simulation of real world scenarios is particularly challenging,” said Jack Parker, a scientist in the Deepmind Open End.
TechCrunch event
Francisco
|
27-29 October 2025
Genie 3 is designed to solve this congestion. Like VEO, it is not based on a hard -coded physics machine. Instead, it teaches the way people work – how objects move, fall and interact – we remember what it has created and reasonably created on long horizons.
“The model is automatic, which means it creates one frame every time,” Fruchter told TechCrunch in a separate interview. “He has to look back on what was created before to decide what will happen next. This is a key part of the architecture.”
This memory creates consistency in its simulated worlds and this consequence allows it to develop a kind of intuitive perception of physics, similar to how people understand that a glass that hits on the edge of a table is going to fall or that they have to go to avoid an object.
This ability to simulate a coherent, naturally reasonable environments over time makes Genie 3 much more than a genetic model. It becomes an ideal training ground for general -purpose agents. Not only can it create endless, different worlds to explore, but it also has the ability to push the factors to their limits – forcing them to adapt, fight and learn from their own experience in a way that reflects the way people learn in the real world.
Currently, the range of actions that an agent can receive is still limited. For example, appropriate global events allow a wide range of environmental interventions, but are not necessarily executed by the agent himself. Similarly, it is still difficult to accurately form complex interactions between multiple independent factors in a common environment. Genie 3 can also support only a few minutes of continuous interaction when hours were necessary for proper training.
Still, Genie 3 presents a fascinating step forward to teachers to overcome inflows so that they can design, explore, seek uncertainty and improve the test and error-the type of self-guided, integrated learning that is the key to the key.
“We didn’t really have a 37-minute move for the built-in agents still, where they can really take new actions in the real world,” said Parker-Holder, referring to the legendary moment of AI’s ability to discover new strategies beyond Human Condering.
“But now, we can probably start a new era,” he said.
