Google's Genie 3 Can't Hold Its Worlds Together for More Than a Minute

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

According to a talk at the Game Developers Conference, as reported by Gamefile and covered by GamesIndustry.biz, the company's generative AI world model starts to break down after about a minute of interaction. Months earlier, the same technology could only maintain consistency for a few seconds.

The issue is memory. Genie 3 generates its worlds one frame at a time — more like a reactive video than a traditional 3D game engine. When a user revisits a location after a minute, the model has to recall relevant information from a minute ago. That computation has to happen multiple times per second to maintain real-time interactivity at 20-24 frames per second. The system can't keep up indefinitely.

According to DeepMind's own model page, Genie 3 "can support a few minutes of continuous interaction, rather than extended hours." The limitations section also notes challenges with accurately representing real-world locations, multi-agent interactions, and text rendering.

This is notable because it's a major tech company openly acknowledging a significant product limitation — not in a defensive way, but as part of a GDC talk about the technology's direction. Alexandre Moufarek, product lead for Google's Inception team, said during the talk that DeepMind isn't trying to replace traditional video games. "We're not at all in a stage where we can just, say, make a game with it," he said. The goal is AGI and creating virtual worlds for AI agents to navigate.

The gap between the hype — Genie 3's reveal caused video game stock prices to tumble in January — and the reality is still substantial. But the transparency is a refreshing contrast to the typical product launch cycle.