In this issue
The short version
- Genie 4 generates explorable 3D worlds in real time with native sound, and you can change weather, characters and camera by typing. It is a research preview for trusted testers.
- Nvidia is investing $3 billion in Meta's share offering and expanding work with Meta on its Muse agent, which Meta says has passed 10 million accounts.
- China's first national standard for data processing units (DPUs) gives buyers of AI computing clusters common requirements and test criteria.
- OpenAI's GPT-6.1 Terra halved mid-tier prices this week; see our Special Edition for October 26 to 31.
October 28 brought a model that builds interactive worlds rather than videos, a large chip-maker investment in a personal AI agent, and a technical standard aimed at the less visible parts of AI data centres. OpenAI's cheaper GPT-6.1 Terra also arrived this week; we cover it in our Special Edition for October 26 to 31. Each section explains what changed and how to use it.
01
Genie 4: interactive worlds with sound
What Google DeepMind announced
Google DeepMind announced Genie 4 on October 27, a world model that generates a 3D environment you can move through in real time. DeepMind says it is the first of its real-time world models to generate audio natively: ambient sounds such as wind and waves, music and spoken dialogue that match the scene. During a session you can type changes, such as new weather, a different camera angle or additional people and objects, and the world updates.
Genie 4 also works with Gemini. A Gemini model observes the scene and plans actions, which lets researchers test how an AI agent or a simulated robot handles multi-step tasks in a generated environment, such as a humanoid robot playing chess. Access is limited to a research preview for trusted testers, with Google AI Ultra subscribers and API partners expected next. DeepMind lists accurate real-time audio, legible text and consistency over long sessions as unsolved problems, and details can drift as a session runs.
What world models are good for today
World models are not video generators. A video tool gives you a fixed clip; a world model gives you a space you can act in. The near-term uses are testing and prototyping: walking a client through a rough version of a venue, trying out the mood of a game level, or training and testing agents in situations that are hard to stage in reality.
Until Genie 4 is widely available, you can prepare by writing briefs that would work for both kinds of tool: describe the setting, the time of day, the sounds you expect and what should happen when someone moves. Test those briefs in video generators such as Google Veo now, and you will be ready to compare them with interactive output later. Gemini is the place to watch for wider access. Our AI video generator guide covers the current tools.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
Nvidia's $3 billion investment in Meta and the Muse agent
What was announced
Nvidia will invest $3 billion in Meta as part of Meta's share offering and expand the two companies' work on Muse, Meta's personal AI agent. Meta plans to deploy more Nvidia hardware, including the upcoming Vera Rubin platform, and the companies will consider further collaboration on Nvidia's OpenShell and Sentry security software; no deployment of those has been committed. AMD also committed $1 billion to the offering. Meta's wider financing plan is covered in our Special Edition.
On Meta's results call, Mark Zuckerberg said Muse has passed 10 million accounts. Meta has reported strong daily use and shopping as one of the most common tasks; those engagement figures come from the company.
Trying a personal agent sensibly
Muse and similar agents are moving from answering questions to doing tasks, including shopping. That is convenient and raises the cost of mistakes. If you use Meta AI or any assistant that can buy things, set a spending rule before the first purchase: the agent proposes, you approve. Check the item, seller, price and delivery in the summary, not in the conversation.
Keep a short list of preferences the agent should always apply, such as sizes, brands you avoid and delivery address, and add to it whenever it gets something wrong. Compare the result with what you would have chosen yourself for the first few purchases. If the agent saves time without mistakes, extend its role gradually. General assistants such as ChatGPT and Claude offer similar agent features worth comparing.
03
China's first national standard for DPUs
What was published
China published and implemented its first national standard for data processing units in intelligent computing clusters. DPUs are chips that take networking, storage and security work off the main processors, so that GPUs and CPUs spend more time on AI tasks. The standard was led by the China Academy of Information and Communications Technology with China Mobile (中国移动), China Unicom (中国联通), China Telecom (中国电信), ZTE (中兴通讯), Inspur (浪潮), JD (京东) and others. It sets technical requirements for DPUs used in AI computing clusters, covering their development, deployment in servers and clusters, and management, and gives buyers a common basis for choosing and testing products.
The standard addresses practical problems: vendors' different designs, performance figures that cannot be compared, poor compatibility between systems and no common way to accept a project as complete.
Why buyers of AI infrastructure should care
Most AI users never see a DPU, but anyone buying servers or private AI clusters will. Without common criteria, each vendor's benchmark measures something slightly different, and problems appear only after installation. A national standard gives you a checklist to put in requests for proposals.
If you procure AI infrastructure in China, ask vendors to state compliance with the standard and to provide test results against it. Define acceptance tests before delivery, including network throughput under load and compatibility with your existing management tools. Elsewhere, the same approach applies: insist on comparable, independently reproducible figures and written acceptance criteria. Assistants in the Data Analysis category can help you build comparison tables from vendor submissions.
What to try next
Put one change into practice
Write one world-model-ready brief and test it in a video generator. If you use a shopping-capable assistant, set a propose-then-approve rule today. If you buy AI infrastructure, add standard compliance and written acceptance tests to your next request for proposals.
Explore Video Generation, Chatbots and Data Analysis to build your shortlist. Our Gemini review covers Google's assistant.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.