Vibeleaderboard
← All Intel
Intel / post

Project Genie moves to Genie 4 with live world editing

Source
x.com
Date
Google Labs@GoogleLabs

Create a world from one sentence, then change it while you’re standing inside 🌐 Project Genie now runs on @GoogleDeepMind’s Genie 4. Here’s what’s new: — Prompt mid-session to edit your world live — Create worlds conversationally with the help of a new agent — Explore & remix other users' worlds in the gallery Available now to eligible Google AI Ultra subscribers

Context

Project Genie, Google's experimental site for generating worlds you can walk around in, now runs on Genie 4, the newest world model from Google DeepMind, and the headline change is that you can rewrite a world while you are inside it (Google Labs). A world model here generates an environment on the fly: you move and turn the camera with the keyboard, and the model produces each moment of the scene in response. With Genie 4, typing a new prompt mid-session can change the weather, add a character or redirect the action, and for the first time the world comes with generated sound, from ambient noise to characters' dialogue (Google). The update is rolling out to eligible Google AI Ultra subscribers, and Google AI Pro members can join a waitlist.

Project Genie opened to U.S. Ultra subscribers in January using Genie 3, which turned a text prompt into an explorable environment; Google says early testers then asked for deeper control over what happens inside those worlds. DeepMind announced Genie 4 on October 27 as a research preview for trusted testers (Google DeepMind). It says Gemini models help carry out multi-step events inside a world: Gemini 3.8 Flash watches the scene as Genie 4 generates it, breaks a goal into steps, checks incoming frames to confirm each step happened and adjusts its plan when the user intervenes. The Project Genie update adds two things around the model: an that builds a world from a conversational description by coordinating Google's other models for the initial look and a background score, and a gallery where people can publish worlds and remix others'. Genie 3 stays in the model picker, now with sessions of up to 3 minutes instead of 60 seconds and its own mid-session prompts for in-world events.

Live editing is what moves Project Genie from watching a generated scene toward steering one, but everything shown so far comes from Google's own demonstrations. Google describes both Genie models as frontier research, and DeepMind says visual details and object permanence can still drift in longer sessions or after several complex edits, while physical interactions, audio accuracy, legible text and long-term consistency remain open problems. Google has not said how long a Genie 4 session can run, though it points people to Genie 3 when they want longer runs. DeepMind says simulations like Genie 4's could help build training environments for robots and AI agents, but Project Genie itself is a paid creative prototype.

Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Recommended reads
Comments

Checking sign-in…

Loading comments…