· via Hacker News – Front Page (native)
Google launches Genie 4 world model inside revamped Project Genie for AI Ultra subscribers
Google launches Genie 4, a world model you can edit mid-session, inside a rebuilt Project Genie experience for AI Ultra subscribers that adds an AI building agent and a remixable community gallery.
Google has released Genie 4, the newest version of its AI world model, together with a rebuilt Project Genie experience for eligible Google AI Ultra subscribers. The announcement, published on Google's blog on November 18 by Pavan Kumar, Group Product Manager at Google Labs, also introduces a conversational agent for building worlds, upgrades to the older Genie 3 model, and a community gallery for exploring and remixing creations. Project Genie, first shown earlier this year as an experimental research prototype, pairs Google's world models with its Gemini models so users can build, enter and reshape interactive worlds from a browser.
Genie 4: live edits and adaptive audio
According to Google, Genie 4 is its highest-fidelity world model so far, and its headline capability is mid-session control: while navigating a generated world, users can type a new prompt to reshape the environment on the fly, shifting the weather, dropping in a character or steering the action somewhere else. Genie 4 also produces native audio in real time that adapts to what is happening on screen, from background ambience to spoken dialogue between characters, which Google says is a first for the model family.
Genie 3 sticks around with longer sessions
The previous model remains available in Project Genie's model picker, positioned for longer runs and for worlds seeded from panorama imagery. Google has extended Genie 3 sessions to three minutes per generation, up from 60 seconds, and added the ability to chain several worlds into one longer journey. New worlds can be seeded from 360-degree panorama images, and natural-language commands can trigger in-world events while a world is being explored.
A conversational agent for world-building
The new Project Genie agent is meant to move creation beyond a single text or image prompt. Users describe what they want in plain language, and the agent orchestrates Google's multimodal models to produce the opening visuals that set a world's look, along with a background score composed to fit the mood. Gemini can also be brought into an active session, where it observes the on-screen action, surfaces suggestion chips that hint at where to go next and can automatically trigger events as the session unfolds.
Curated and community worlds to remix
A new gallery collects curated worlds alongside community submissions. Google's own examples include "Living Memory," which converts a user's own photos into spaces they can walk through; "Dino Field Guide," which grounds prehistoric habitats in 360-degree panoramas; and "Rock Paper Scissors," where objects with different themes face off on a desktop while Gemini referees in real time.
The gallery also features work from creators with early access: "Villa Giallo," a mystery that unfolds room by room inside a gothic manor built with creative technologist Will, known as BLVCKL!GHT, and "The Life of Ants," made with filmmaker Anand Gandhi, which sends users through a sprawling ant colony from an ant's-eye view. Any gallery world can be remixed through the agent by adjusting its underlying structure, soundtrack or rules, and finished creations can be published back to the gallery for others to explore and modify.
Access, watermarking and open problems
The new experience is open to eligible Google AI Ultra subscribers now, while Google AI Pro members can join a waitlist. Google says every video and audio clip generated in Project Genie carries a SynthID watermark, and all model inputs and outputs go through safety screening. The company still describes Genie 3 and Genie 4 as frontier research, and points to unsolved problems including camera latency, fine-grained physical interactions, precise real-time audio and scenes that stay consistent over time, with further detail in its Genie 4 model research overview.
Why it matters
World models, systems that generate navigable and interactive environments rather than static images or video, have mostly lived in research demonstrations. Google is now shipping one as a consumer feature, gated behind its most expensive subscription tier and wrapped in creator-platform mechanics: an agent to lower the barrier to building, plus a gallery built on remixing and sharing. That positions Project Genie as a differentiator for AI Ultra and a signal of how quickly prompt-to-playable-world tools are being productised across the industry. Equally telling is what Google admits does not work yet, since latency, physics fidelity and scene consistency remain hard problems, and naming them in the launch post suggests how early the format still is. Building SynthID watermarking in from day one also indicates Google expects these generated worlds to circulate widely, and wants provenance attached before they do.
- #ai
- #generative-ai
- #world-models
- #gemini