Google's Genie 3 World Models Start to Break Down After About a Minute

Google DeepMind is being surprisingly honest about Genie 3's limits.

At a GDC talk this week, the company acknowledged that its world model technology starts to show inconsistencies after about a minute of runtime—due to memory constraints. Months earlier, the system could only maintain coherence for a few seconds.

"Memory limits mean the world models it generates start to show inconsistencies after about a minute," according to a Gamefile report cited by GamesIndustry.biz.

Alexandre Moufarek, product lead for the Inception team at Google DeepMind, said during the talk that the team isn't thinking about replacing traditional video games. Instead, Genie 3 is aimed at helping "imagine new game experiences…you couldn't even imagine without AI" and ultimately toward the goal of artificial general intelligence—creating bespoke virtual worlds for AI agents to navigate.

"We're not at all in a stage where we can just, say, make a game with it," Moufarek said, though he added it would be interesting to put the technology in developers' hands to "play around with it."

The tech works by generating frames one at a time, functioning more like a reactive video than a traditional 3D game world.

The admission is notable given that Genie 3's reveal caused video game company stocks to tumble back in January, with investors worried about AI's potential to replace game development.