Google's Genie 4 builds worlds you can steer⁠.

Google DeepMind has announced Genie 4, the newest version of its world model, which generates an interactive scene that a user can move through in real time¹. Google released it as a research preview on Tuesday³. It is open first to a group of invited testers, with Google AI Ultra subscribers and early access users of the Gemini API and Google AI Studio to follow¹.

The model makes sound as well as pictures, including music, background noise and dialogue, generated together with the world¹. You can type an instruction mid-scene to change the weather, move the camera or add a person or an object, and Gemini adjusts the request to fit what is already on screen¹. Google says Gemini also works behind the scenes to plan the scene and keep it coherent¹. The Verge reports that Genie 4 uses Gemini 3.8 Flash for this².

A scene can now hold together for about two minutes, up from about one minute for Genie 3². TechCrunch reports that Genie 4 also offers better physics and more humanlike characters than its predecessor³. Google says all video and audio it generates will carry SynthID, its hidden watermark for AI-made content¹.

Google is pitching it for pre-visualisation in film, prototyping game ideas and building training scenarios for robots². It says enterprise partners are testing it to generate scenarios quickly, and that game developers are using it to go from an idea to an explorable prototype before writing any engine code¹. A Google DeepMind engineer told reporters it is still early days and that the finished product is not yet clear³.

Google is plain about the limits. Details and object permanence can drift as a simulation runs longer, and accurate audio, legible text and long-term consistency are still being worked on¹. The Verge notes that it has not been able to try the model itself². The release comes as the field consolidates: TechCrunch reports that World Labs is being bought by AMD and that Amazon is in talks to buy Decart³.

A preview, not a product yet

The sources give no British angle, no UK release date and no price. For most firms in Leeds or Bristol this is one to watch rather than one to buy. The practical uses on offer today are narrow: storyboarding, a quick prototype of a game or a training scene, and a test bed for robots. Anyone in product design, training or marketing who wants a feel for it will need a Google AI Ultra subscription or a place on the waitlist, and should expect it to wobble after a couple of minutes.

The more useful point is the direction. Google is joining its language model to a picture engine so that one plans and the other draws. If that holds, the cost of mocking up a space, a process or a scenario before you build it should fall. Until it leaves preview, treat the demonstrations as Google's best examples, not as what you would get on a Tuesday afternoon.

Also today

• AI bots sent three quarters of replies to a UK vapes consultation

  The Department of Health and Social Care found that 90,835 of 118,756 responses to its 2023 consultation on disposable vapes were sent by bots, and excluded them⁴.

• Australia moves to ban chatbots that seek emotional intimacy

  Australia's digital duty of care bill, which would ban AI chatbots that seek to create emotional intimacy, has passed the House of Representatives and now goes to the Senate⁵.

• Most firms have cut back AI plans because of token costs

  Cloudera research found that 89 percent of organisations scaled back or redesigned AI plans this year because of token costs, and 45 percent moved workloads as a result⁶.

• UK trade body asks the budget for a sovereign AI plan

  UKAI, which represents more than 270 British AI businesses, wants next week's budget to merge the government's separate AI programmes into one sovereign AI action plan⁷.

• Chinese start-up shows a four metre robot built for heavy lifting

  Beijing-based NeodraDynamics says its 1.2 tonne Neo1 robot can carry 500 kilogram loads, though a display model at a trade fair showed only limited movement⁸.