I built marshmallow castles in Google’s new AI world generator

Chocolate rivers flowing through marshmallow castles in the clouds now spring to life with just a text prompt, as Google DeepMind opens Project Genie to AI Ultra subscribers across the US starting Thursday. The experimental tool transforms written descriptions or photos into playable game worlds within seconds, powered by the company's Genie 3 world model and image generator Nano Banana Pro.

Users get 60 seconds to explore their creations due to massive computing demands, with each session requiring dedicated processing power. The technology excels at whimsical, artistic worlds but struggles with photorealism, often producing sterile, video game-like environments instead of lifelike settings.

"I think it's exciting to be in a place where we can have more people access it and give us feedback," says Shlomi Fruchter, DeepMind's research director, acknowledging the tool's inconsistencies as characters sometimes walk through walls or navigation goes haywire.

This release intensifies the world model race, with competitors like Fei-Fei Li's World Labs and Runway launching similar products. DeepMind views entertainment as a stepping stone toward training robots in simulation and eventually achieving artificial general intelligence.