A crowned crane is walking across a snowy landscape. A person types an instruction for it to approach a white piano.

A keyboard appears in the snow. The bird walks towards it and lowers its beak to the keys. Another instruction tells it to spread its wings. The bird does. Watch Google’s demo

This is a demonstration of Genie 4, Google DeepMind’s new model for creating interactive worlds. The person can move the camera and send new instructions while the scene is running.

There’s also a small box on the screen showing “reasoning steps.” Google says another AI model, Gemini, helps guide what happens in the world while Genie generates it.

I am Alex, welcome to ShortCu8 by Innov8.

Lets Dive Deep 🐰

Think of the demo as two jobs working together: Genie generates the scene and its sound. Gemini helps plan and track what happens inside it. How the models work together

Google’s previous model, Genie 3, could already create a world you could move through. It also let users change things, including the weather, with text. The ability to steer a world did not begin with this release. Genie 3 announcement

Genie 4 adds sound generated as part of the scene. Google says that includes ambient noise, music and dialogue that can respond to events. It also says the new model improves the appearance and movement of people and objects. Genie 4 announcement

Google describes using Gemini 3.8 Flash as a guiding agent: it watches the scene, breaks a goal into steps and checks the incoming frames as Genie produces them. If the scene changes or the user intervenes, Gemini adjusts the plan. Google calls this a step towards a hybrid world model. Google’s explanation

In the crane demo, the camera moves closer as the bird and keyboard occupy the same scene. The person directing it can then add another action without starting with a blank page.

Google says its trusted testers are exploring Genie 4 for film previsualisation: trying out a scene before making the finished version. They’re experimenting with camera angles, lighting, dialogue and action. Google’s tester notes

For that to become useful, a creator needs to be able to change one part of a scene without losing the rest. It’s hard to compare two camera angles if the person or the room changes between them.

In a separate experiment, Google paired Genie with a different model, Gemini 4 Argon, to simulate Apptronik’s Apollo humanoid playing checkers. Argon breaks the task into steps and passes them to Genie as text while the scene is rendered. The clip doesn’t show a physical robot learning to play. Google’s robotics example

The ambition is to give AI systems more places to practise and test actions. A world could present an unfamiliar room, an unexpected object or a change in weather without requiring a team to build each situation physically.

That would only be useful for real machines if the lessons carried over. A convincing picture of a robot moving is not evidence that a physical robot can repeat the move, or that the simulation got the physics right.

Back to the crane. If it walks away from the keyboard and returns, the keyboard needs to stay where it was. If the camera turns, the landscape should still make sense from the new angle.

In a conventional game, the software can keep a record of where an object was left. A generated scene has to keep its next images consistent with what came before, even when the camera moves and the user adds new instructions.

Earlier this year, The Verge’s Jay Peters tried Project Genie with the older Genie 3. In one world, a rolling ball left paint on the snow, but some marks disappeared. In another, part of a racetrack turned into grass while he was driving. Those were his observations of Genie 3, not a test of Genie 4. The Verge’s hands-on report

Google says Genie 4 can still lose that consistency. Details and objects may shift as a scene runs longer or after more complex changes. Its model page also says interactions currently last a few minutes, and audio accuracy remains a work in progress. Genie 4’s limitations

The crane clip is Google’s own demonstration. Google says its recordings are illustrative and the sequences have been shortened, so a demo alone cannot tell us how often an interaction will work. Google’s demo disclaimer

Genie 4 is currently a research preview for trusted testers. Google says it is coming to AI Ultra subscribers through Project Genie and to early-access partners using the Gemini API and AI Studio. It has not given a release date. Access and waitlist

Google has shown the keyboard appearing in the snow. I want to see it stay put when a creator turns away, tries something else, and comes back for another take.

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.

Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.

Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.

Try Notetaker Free

#AD2

Shoppers type exactly what they want and get “0 results found” back. They don’t complain. They just vanish to a competitor.

Gladly understands what shoppers mean, typos and all, and turns dead-end searches into product recommendations.

See how Gladly brings search back to life

Turn Dead Ends Into Sales