Google DeepMind’s Genie 3 is the latest in its family of AI world models, able to generate fully interactive 3D environments in real time from plain text prompts. With live navigation at 720p resolution and 24 fps, Genie 3 sustains several minutes of consistent exploration, a major improvement over Genie 2’s mere 10–20 seconds. Unlike static video models, Genie 3 remembers object placements for about a minute—even when they go out of view—and supports dynamic “promptable world events” (e.g. changing weather, adding characters) to modify environments on the fly.

This richer, longer‑horizon interactivity enables both human users and AI agents to pursue goals in evolving virtual worlds—an important step toward embodied agent training and general artificial intelligence.

Despite its breakthroughs, Genie 3 comes with limitations: lower resolution and memory horizon than traditional rendering, restricted agent action space, difficulty modeling multiple independent agents or real‑world accuracy, and text rendering that only works when explicitly specified. Currently, access to Genie 3 is limited to a small group of researchers and creators in a controlled preview, while DeepMind studies the technology’s risks and potential misuse before broader release