Google Deepmind has announced Genie 3, the most recent most recent model of the AI granted to give the illegal stone on the road for public intelligence, or intelligence.
“Genie 3 is the first interactive model model in the world,” Fruit’s Shortter, research director in Deepmind, said when the press. “You extend the previous narrow model. It is not specifically for a particular environment. This can produce the world-eight and imaginary and illusion between.”
Genie 3, which is still in previewing research and not available in general, build before Genie 2 – who can produce new environments to agents – and the latest-video-generation model I see 3 – who showed the understanding of the deep physics.
With simple text prompt, Genie 3 can returns for many minutes – from 10 to 20 seconds on the Genie, interactive, interactive, 3D environment in 24 frames per second with 720p resolution. The model also displays the “event of the world that is not working,” or the ability to use directions to change the world.
Maybe the most important, Genie simulation 3 remain consistent during the time because the model can remember what you previously made – the possibility underwear is not a program.
Fruchter said that when Genie 3 clearly has implications for educational experience and new generative media GAME Or prototing the creative concept, the real unlock will be real in the training agent for the common destination task, which says it is important to reach Agri.
“We think the world’s model is the key to the road to AGI, specifically for the actual agent, where the scientist is a flight in Deepmind’s open team, saying that before it is short.
TechCrunch Events
San Francisco
|
October 27, 2025
Genie 3 is designed to complete the bottleneck. Like Veo, not depending on the difficult physics engine. However, teach yourself how the world is the world – how the object, falls, and interacting – with memories of what has been made and showing the long horizon.
“The model is automatically generated, meaning returns for one frame,” the fruchter said to tell the techcrunch in a separate interview. “You have to see you again about that you make before deciding what will happen to the next time. This is the main part of the architecture.”
It is expressing the world in the world of simulation, and consistency allows man knowing that humans know that there are glasses on the edge of the table approximately to avoid falling objects.
This ability to simulate Koherent, the environment that can be physically in order to make Genie 3 more than a generative model. Being an appropriate background for a public destination agent. It is not only possible to produce that there is no end, but it may have the potential to push the agents – forcing them by adaptation, and learn from their own experience in the real world.
Currently, the various actions can take the agents can still be limited. For example, a world event that cannot allow various environmentally interventions, but it does not have to do by the agent itself. In addition, it is still difficult to emphasize complex interactions with complex interactions between various independent agents in the environment in shared environments. Genie 3 can also only support some constant interactions, when the clock should work right.
Still, Genie 3 displays the instructions to begin with the question to sign in, in order to design, to do, pursuing, to pursue it, and work on public intelligence.
“We’ve never seen 37 move for an existing agent, so he can take novel actions in the world,” said the legendary agent, called a lovely experson to find a new strategy.
“But now, we can potential USH in the new era,” he said.