Deepmind thinks his new World World Model presents a stone at the agi

Google Deepman revealed genius 3, its last model of the world that can be used to form general general agents, an ability to the road to “artificial intelligence.” Artificial intelligence.

“Genie 3 is the first model of general,” Freuchter Shlomi, a depth director at the bottom-in-the-point, he said during a picture. “Alone the narrow world models that exist first. It is not specific to any particular environment. Can generate the two pics of photos, and everything.”

Still in search preview and not available, genie 3 build on its predecessor Genien 2 (that can generate new environment for agents) and the last generation model of the depth I see 3 (who says to have a deep understanding of physics).

Image credits:Google Deepmind

With a simple prompt, Genie 3 can generate many minutes of the interactive 3D environments to 24 frames per second generation from 10 to 20 seconds. The model character also “World Events Promise,” OA ability to use a readiness to change the generated world.

Perhaps Simestian similes, genie responses to time because the model can remember, what you have usually generated – a capability meant a patterned pattern.

Fruchter said that as genie 3 has implications for educational experiences, Gaming o Creative concepts, its real unlock to handle in training agents for general purpose activities, which he said it is essential to reach the agi.

“Let’s think the world models are key on AGI, specifically for the agents of incantents, where similar scenarios is particularly served, a research scenery in the race.

Techcrunch event

San Francesco
| 0.
October 27-29, 2025

Image credits:Google Deepmind

Genie 3 is supposedly designed to solve the bottleneck. As veho, do not basically in a difficult physical engine; Instead, deep the music says, you’re teaching, how to get worldwhat, as moving, and interacts, and interacts of what has generated and reasoned and reasonable and reasoning long.

“The model is automi, which means genies a square at a time,” fruchter said to technochcrough in an interview. “You have to look at what was generated before you decide what happens following. Here’s a key part of architecture.”

This memory, the company says, consistency in geies, which will return to developing that a glass of a table, or when they built to avoid an object that goes down.

Nabbed, ProfundMind says the pattern to push the agents to their limits – forces to learn from their own experience, similar to man.

As an example, the deep shared their trial of genes 3 with a recent version of their generalist Utterable utteribly agent climb (sina)instruct pursuing a set of goals. In a warehouse environment, they asked the agent to do the functions as “approach greenish-green” or “walk in the foklift red”.

“In all three cases, the sopic manner is able to accomplish the sake of the agent. So see the world, and then take the action of the show, and the fact that it is knowing-genue 3

Image credits:Google Deepmind

Say said, genue 3 has its limitations. For example, while researchers can figure physical, demo showing a barrier barrier give a mountain dont reflect how much it will be related to the skier.

Also, the range of actions an agent can take is limited. For example, the prompt worlds adversable allows a wide range of environmental interventions but are not necessary for the agentsia. And it is still difficult to accurately model complex interactions between independent agents in a shared environment.

Genie 3 can also sustain a few minutes of continuous interaction, when the hours would be required for proper training.

Ended, presents a pattern that approaches the teaching teaching of teaching, letoming to the crossroad and error – the kind of general intelligence.

“We don’t really have a bite 37 moment for incrediable infars, where they can actually take the real moment in the world of the world in the world

“But now, all of you can carry it in a new he was,” he said.

Source link