Product Launch3 min read

World Models Let Machines Rehearse Their Worst Days First

By , Senior AI ConsultantPublished

Google's Genie 4 builds walkable worlds with sound and talking characters, and the first use of a tool like that is rehearsal, as Waymo showed by turning Genie 3 into a driving simulator.

Genie 4 is Google DeepMind's newest world model, a system that draws a moving, explorable scene from a text description and keeps drawing it as you steer. It adds sound to the picture, including characters who speak, and its people move more naturally than before. It also lets a Gemini model act as a director that plans the steps and checks each frame as it appears. Trusted testers have it now, and Google AI Ultra subscribers in Project Genie and early partners of the Gemini API come next.

A simulation is a bad day you can order on purpose

A robot that learns only from the days it has lived will meet most rare problems for the first time on the job. A world model removes that limit, because it draws the next moment from whatever you describe. Typing that a pallet slides off a forklift gives the robot a new day, and typing it again with the lights out gives it another.

The director is what makes this useful for machines. In Google's own demonstrations, Gemini breaks a task into steps, such as a humanoid robot playing checkers or a robot arm dropping liquid into a lab plate, then feeds them to Genie one at a time and checks that each step worked. Those are research demos, and nobody has reported certifying a robot this way. Waymo's driving simulator shows the same path already working in a real product.

A shift manager choosing a picking robot could one day ask the vendor for the footage: the aisle blocked, the light failing, a box splitting open as someone steps in. Staging that for real takes staff, stock and a closed aisle. Typing it takes a sentence.

The rough draft has a history

The nearest precedent is film. In the mid-1990s, the makers of the first Mission: Impossible needed to show the studio what a fast train chase would look like, and storyboards could not show a camera moving at that speed. An artist in George Lucas's effects company, Industrial Light & Magic, spent about two months making rough computer animation of the whole scene. It had no textures and the characters had no faces, and he later called it awful. It told the story, and the studio went ahead.

A few years later the same method was a large part of making a Star Wars film, and then a whole trade of previs companies grew up around it. The rough draft became ordinary because it was cheap and let people argue about a scene before paying to shoot it.

Genie 4 moves that draft one step earlier and makes it something you stand inside. Testers are already using it that way: a filmmaker can run several takes of a scene and change the light, the camera or the dialogue between them. The same will be true for anyone who has to explain a place or an event before it exists, such as a restaurant owner who wants to show a small film crew the ad she has in mind before paying for a shoot day.

Why it is a draft and not a measurement

Details and objects drift as a scene runs, text in the scene is often hard to read, and a scene holds together for up to two minutes. That makes a world model good at asking what might go wrong and poor at saying exactly how it will go. A simulated bad day that looks right but behaves wrongly teaches the wrong lesson, and it does so convincingly.

So the rehearsal worth trusting is the one that surprised somebody. When a vendor, a studio or a supplier shows you a simulation, ask what it found that they did not expect, and what they changed because of it.

Share this

STAY INFORMED

Get AI intelligence like this delivered to your inbox.

Free forever · Unsubscribe anytime


You May Also Find Valuable