Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarTNW→ original

Google Opens Genie 4 to AI Ultra Subscribers: Interactive Worlds With Sound and On-the-Fly Edits

Google has opened Genie 4, a model that builds an interactive world from a description you can walk through, to subscribers of its paid AI Ultra tier. Mid-session, you can type "let it rain" and the scene changes without a restart. For the first time, the worlds also have sound: ambient noise and character dialogue. But in long sessions, details still drift. For now, that makes it a tool for quick sketches, not for finished games.

AI-processed from TNW; edited by Hamidun News
Google Opens Genie 4 to AI Ultra Subscribers: Interactive Worlds With Sound and On-the-Fly Edits
Source: TNW. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTNW ↗
Editor-in-Chief
Genie 3 remains for longer runsup to 3 minutes instead of 60 seconds
Key metric #2AI Pro subscribers can join a waitlist
Key metric #3Everything created in Project Genie gets an invisible SynthID watermark

Imagine writing a single line of text and having a neural network build a world you can walk through, then changing the weather or adding a character on the spot. That is what Google is now offering paying users: Genie 4, a new "world model" (a neural network that generates not a video clip but a space you can move through), is available to Google AI Ultra subscribers in the Project Genie browser prototype. For anyone making games, simulations, or following the race around world models, it is a good chance to see where the technology stands right now.

What Google Actually Released

Access is open to eligible Google AI Ultra subscribers. AI Pro owners can join a waitlist. Nothing needs to be installed: Project Genie runs directly in a browser tab.

The principle is familiar. You describe a world in words, and the model turns the description into a scene. What is new in the fourth version is editing on the fly. Say you are walking through a forest and type "let it start raining" or "add a traveler nearby." The weather changes, the character appears, events unfold differently, and you do not have to start the session over.

The second addition is built-in sound. Worlds in Genie used to be silent. Now the model adds both ambient noise and character dialogue. Even a rough scene with sound feels very different from a silent one.

The Creator Agent, the Gallery, and the Future of Genie 3

Alongside the model, Google added two tools. The first is an AI agent that builds worlds from conversation: instead of hand-tuning phrasing, you discuss an idea and the agent turns the talk into a finished world. The second is a shared gallery where you can publish your worlds and remix others'. For the platform, this is a fast way to build up a library of prompts that work well.

The older Genie 3 has not gone away. It has been kept for longer runs, and a single generation now lasts up to three minutes instead of the previous 60 seconds. The choice is simple: Genie 4 offers on-the-fly editing and sound, while Genie 3 makes sense when run length matters more.

Limitations and Context

Google itself calls both models research systems. In long sessions, world details can "drift": objects change and the image loses coherence. Readable text inside a scene and complex physics are still handled poorly. Everything created in Project Genie receives SynthID, Google's invisible watermark for AI-generated content.

Large game studios approach technologies like this with caution. The source material mentions Take-Two as an example, but gives no details.

Serious money is flowing into the area. In September, AMD agreed to acquire World Labs, the company led by Fei-Fei Li. Google DeepMind sees world models as a way to train robots and AI agents before they operate in the real world: many scenarios can be run in virtual space without risking hardware. A competitor, the Odyssey lab, claims its model can already control robots and cars, but the material contains no independent verification.

What Practitioners Should Do

If you make games, use Genie as a quick sketch of atmosphere and mechanics: on-the-fly editing is handy for trying out variations. If you work on agents or robotics, watch whether world models become a source of training environments. If you need a finished production tool, you will have to wait: drifting details and weak physics do not yet allow you to rely on the system seriously. Still, the fact that this technology already runs in a browser for paying users shows how quickly the path from demo to product is moving.

Story

Key Questions Answered

What Google Actually Released?

Access is open to eligible Google AI Ultra subscribers. AI Pro owners can join a waitlist. Nothing needs to be installed: Project Genie runs directly in a browser tab. The principle is familiar. You describe a world in words, and the model turns the description into a scene. What is new in the fourth version is editing on the fly.

What are the key details on The Creator Agent, the Gallery, and the Future of Genie 3?

Alongside the model, Google added two tools. The first is an AI agent that builds worlds from conversation: instead of hand-tuning phrasing, you discuss an idea and the agent turns the talk into a finished world. The second is a shared gallery where you can publish your worlds and remix others'. For the platform, this is a fast way to build up a library of prompts that work well. The older Genie 3 has not gone away.

What are the key details on Limitations and Context?

Google itself calls both models research systems. In long sessions, world details can "drift": objects change and the image loses coherence. Readable text inside a scene and complex physics are still handled poorly. Everything created in Project Genie receives SynthID, Google's invisible watermark for AI-generated content. Large game studios approach technologies like this with caution. The source material mentions Take-Two as an example, but gives no details. Serious money is flowing into the area.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…