DeepMind thinks its new Genie 3 world model presents a stepping stone toward AGI

Google Dipmind Jenny 3 has revealed that its latest foundation World Model that can be used for training for general-purpose AI agents, says the AI Lab that creates an important step in the “artificial general intelligence” or human-national detective path.

“Jenny 3 is the first real-time interactive General-Perpus World Model,” said Dipmind’s research director Slommy Fruchter during a press briefing. “It was out of the narrow world models that existed it before it is not specified for a particular environment. It is a photo-realistic and fictional world and both in between.”

Still is still pre -exemplared and universally available, making both the predecessors of Jenny 3 Jenny 2 (Which may create new environment for agents) and the latest video of the Deepmind is the generation model I see 3 (Which is known to have a deep idea of physics).

With a simple text prompt, Jenny 3 can create multiple minutes interactive 3D environment in 720p resolution per second – a significant jump of 10 to 20 seconds can produce Jenny 2. The model also has the ability to use “significant world events” or prompt to change the generated world.

Perhaps most importantly, the simulations of Jenny 3 are physically consistent over time because the model may think what it was made earlier – a power that says dipmind that his researchers clearly did not program the model.

Fruuster said that Jenny 3’s involved in the educational experience of 3, Gaming Or by prototyping creative ideas, its original unlock will be revealed in training agents for general-principal work, which he said that the AGI is necessary to reach.

“We think that the world models are the key to the AGI, especially for the statue agents, where the real world situation is particularly challenging,” said a research scientist at the Jack Parker-Holder, Dipmind’s Open-endness team, during the briefing.

Jenny 3 is designed to solve this barrier. Like VEO it does not depend on the engine of hard-coded physics; Instead, Dipmind says the model teaches itself how the world works – how the objects move, fall and interact – it remembers what it has generated on the horizons of long time and arguing.

“The model is auto-regressive, which means it produces a frame at once,” Fruuchter told TechCrunch in an interview. “Before deciding what will happen then, it is important to look back that is the main part of the architecture.”

The agency says that this memory lends for the continuity of Jenny 3’s simulated World, which allows it to develop a perception of physics, as people realize that a glass shake is about to fall on the edge of a table, or ducks to avoid their falling objects.

Significantly, Dipmind says the model has the potential to push AI agents into their limitations – forcing them to learn from their own experiences, which people are similar to how the real world learns.

For example, Dipmind shared the test of Jenny 3 with the recent version of its generalist Scaleable Instructions Multi -World Agent (Sima)It is instructed to follow a set of goals. In a warehouse setting, they asked the agent to perform tasks like “Go to a bright green trash compactor” or “walk on the packed Red Foreclift”.

“Sima agent is able to achieve the goal in all three cases,” said Parker Holder. “It simply accepts the verbs from the agent so so so

It was said that Jenny 3 has limitations. For example, researchers have claimed that physics understands that a demo that barrels a mountain does not reflect on how to snow with Skyar.

In addition, the activities that can take an agent are limited. For example, instant world events allows for a wide range of environmental interventions but these agents are not edited by itself. And the complex interaction between multiple distinct agents in the divided environment is still difficult to correctly model.

Jenny 3 can only support the continuous interaction of a few minutes, while the hour is required for proper training.

Nevertheless, the model is a self-driven, figurative type that many people say that many are the key to moving towards ordinary intelligence, to overcome the input response, to explore, seek uncertainty, and to improve through trials and errors.

“We still had no 37 moments for embodied agents, where they could take fancy steps in the world champions Lee Sedol, the AI agent of the Parker-Holder Deepmind, Li Sadol,”

“But now, we can probably start a new era,” he said.