DeepMind believes that the new Genie 3 model introduces a step for AGI

Google DeepMind has unveiled Genie 3, which is the latest model in the world that can be used to train artificial intelligence agents for general purposes, a capacity that the artificial intelligence laboratory says on decisive tendency on the road to “artificial general intelligence” or human -like intelligence.

“Genie 3 is the first interactive global model for general purposes in an actual time,” Shlomlie Fruce, Research Director at DeepMind said at a press conference. “It goes beyond the narrow models of the world that were present before. It is not about any specific environment. It can generate both realistic and imaginative worlds, and everything between them.”

It is still in the research inspection and not available to the public, Genie 3 depends on both of its predecessors Jenny 2 (Which can generate new agents for agents) and the latest DeepMind Veo 3 (Who is said to have a deep understanding of physics).

Through a simple text router, Genie 3 can create several minutes of 3,720 -pixel interactive environments at 24 frames per second – a big jump from 10 to 20 seconds that Genie 2 can produce. The model also contains “speed global events”, or the ability to use a router to change the world created.

Perhaps more importantly, Genie 3 simulations remain physically consistent with time because the model can remember what has been previously created – the ability that Deepmind says that researchers have not explicitly boredom in the model.

Although Genie 3 has traces on educational experiences, Ferkhter said. Games Or the initial models of creative concepts, its real opening will appear in the training agents for general purposes, which he said is necessary to reach AGI.

“We believe that global models are the key to AGI, specifically for embodied factors, as the real world’s scenarios simulating a particular challenge,” Jack Parker, a research scientist in the open Deepmind team, said during the briefing conference.

Genie 3 is supposed to be designed to solve the bottleneck. Like Veo, it does not depend on the coded physics engine; Instead, Deepmind says, the same model knows how the world works – how things move, fall and interact – by remembering what she gave birth and thinking during long horizons.

“The model is considered automatically, which means that it generates one frame at the same time,” Frozter told Techcrunch in an interview. “We must look back at what has been created before to report what will happen after that. This is a major part of architecture.”

The company says that this memory gives consistency in the genetic 3 worlds simulation, which in turn allows it to develop physics, similar to how humans understand that the glass is on the edge of the table is about to fall, or they must duck to avoid the fall.

It is worth noting that Deepmind says that the model also has the ability to push artificial intelligence agents to their borders – forcing them to learn from their own experience, similar to how people learn in the real world.

For example, DeepMind shared its test for Genie 3 with a recent version of the general expert The world’s multi -developed agent (SIMA)Directing him to follow a set of goals. In the preparation of warehouses, they asked the agent to perform tasks such as “approaching bright green garbage rolls” or “walking to the packed red spinal lever.”

“In all three cases, he is able to achieve the goal,” said Parker pregnant. “He only receives actions from the agent. So the agent takes the goal, and sees the world simulating it, then takes the actions in the world. Genie 3 mimics forward, and the fact that he is able to achieve this because Genie 3 is still fixed.”

However, Genie 3 has its limits. For example, while researchers claim that it can understand physics, the illustration showing a skiing wandering on the mountain did not reflect how the snow will move regarding the skipper.

In addition, the scope of action that the agent can take is limited. For example, speed -up global events provide a wide range of environmental interventions, but it is not necessarily done by the agent himself. It is still difficult to design complex interactions accurately between multiple independent factors in a common environment.

Genie 3 can also support just a few minutes of constant reaction, when working hours are necessary for appropriate training.

However, the model offers a convincing step forward in teaching agents to overcome interaction with inputs, allow them to plan, explore and search for uncertainty and improvement through experience and error-the type of self-embodied learning, as many say it is a key to moving towards public intelligence.

“We haven’t already had a 37 -minute step for embodied agents so far, as they can actually take new measures in the real world,” Parker said, referring to the legendary moment in the GO game, referring to the legendary moment in the GO 2016 game between a Deepmind agent from AI AI to discover Ai Beash. ”

“But now, we can enter into a new era,” he said.