Google DeepMind’s Genie 3 can change the state of its dynamic simulation worlds

At the beginning of December, Google DeepMind Jenny’s chest 2. The Jenny family of artificial intelligence systems is what is known as the world models. They are able to generate images as a user – either human or most likely, AI agent – moving across the world that the program simulates. The resulting video may look like a video game, but DeepMind has always put Genie 2 as a way to train other artificial intelligence systems on what they are designed to accomplish. Through the new Genie 3 model, announced by the laboratory on Tuesday, Deepmind believes it has made a better system for training artificial intelligence agents.

At first glance, the jump between Genie 2 and 3 is not as exciting as the model made by the model last year. With Genie 2, the DeepMind system is able to generate 3D worlds, and it can accurately rebuild the environment even after the user or an AI customer left it to explore other parts of the created scene. Environmental consistency was often a weakness in the previous world models. For example, Dert Oasis The system faced a problem remembering the design Minecraft The levels you will generate.

In comparison, the improvements made by Genie 3 seem more modest, but at a press conference issued by Google before the official announcement today, the SHLOMI FruCTER, DeepMind Research Director, Jack Parker, DEPMIND, argued that they represent important stones on the way towards artificial public intelligence.

So what exactly Genie 3 do? To start, shots come out in 720 pixels, instead of 360 pixels like its predecessor. It is also able to maintain a “consistent” simulation for a longer period. Genie 2 had a theoretical limit of up to 60 seconds, but in practice, the model would often begin with hallucinations early. On the contrary, DeepMind says Genie 3 is able to run for several minutes before starting to produce artifacts.

Also new in the model is the ability to call “viable global events”. Genie 2 was as interactive as the user or artificial intelligence agent was able to insert movement orders and the model would respond after he had a few moments to create the next frame. Genie 3 does this work in the actual time. Moreover, it is possible to adjust the simulation with the text calling for the guidance of Genie to change the state of the world that generates it. In DeepMind, the form was informed by inserting a flock of deer in a scene of a person skating on the mountain. Deepmind says that deer did not move in the most realistic way, but this is the deadly feature of Genie 3, says DeepMind.

As we mentioned before, the laboratory primarily imagines the form as a tool for training and evaluation of artificial intelligence agents. Deepmind says that Genie 3 can be used to teach artificial intelligence systems to address the scenarios “What if” is not covered with their pre -training. “There are a lot of things that should happen before a model is published in the real world, but we see this as a way to train models more effectively and increase its reliability,” said Fruchter.

Despite the improvements by DeepMind to Genie, the laboratory admits a lot of work to do. For example, the model cannot generate real sites in full accuracy, and it is depleted with the presentation of the text. Moreover, for Genie to be really useful, DeepMind believes that the model should be able to maintain a simulation world for hours, not minutes. However, the laboratory feels that Jenny is ready to make an influence in the real world.

“We are already at the point that you will not use (Genie) as the only training environment, but you can definitely find things that the agents do not want to do because if they behave in some settings, even if these settings are not perfect, it is still good to know that,” said Parker Holler. “You can already know where this is going. It will be increasingly beneficial with the improvement of models.”

Currently, Genie 3 is not available to the general public. However, DeepMind says it works to make the model to the additional laboratories.