Lead story / launch

Google adds sound to its AI worlds, but they still lose context in minutes

Google DeepMind has introduced Genie 4, its first world model with native sound. Alongside ambient noise and audio specified in a prompt, the company promises characters whose speech adapts when the simulation changes. DeepMind also says the generated worlds have more realistic physics and more natural human movement, though those improvements have not been independently assessed.

Rather than playing back a pre-rendered environment, the system generates a world frame by frame as a user explores it. Gemini 3.8 Flash breaks down goals and evaluates frames to help keep the experience on course. Users can change the weather, add elements or move the camera while the simulation runs, with Gemini helping steer those updates.

The central limitation is continuity: DeepMind says Genie 4 worlds remain consistent for only a couple of minutes before losing their initial context. The company sees potential uses in robot-training environments and rapid creative prototyping, but those ambitions depend on how faithfully a generated world behaves—not simply how convincing an individual frame looks.

For now, a small group of trusted testers has access, while interested users have a waitlist rather than a release date. Google plans developer access through the Gemini API and AI Studio, with pricing still unannounced. Consumer access is planned in phases through Project Genie for Google AI Ultra subscribers, whose plans start at $100 a month. Possible access for AI Pro and AI Plus is not a committed rollout.

culture

Some California professors turn to paper exams as AI clouds homework grades

Some California instructors are moving grades toward in-person, device-free exams because homework answers no longer reliably show individual understanding. UCLA’s Thomas Maierhofer still assigns homework and labs but grades them for completion. UC Berkeley’s Richard Borcherds attributes higher online exam scores to AI, while acknowledging he cannot identify who used it. The changes distinguish AI as a learning aid from evidence of independent work.

business

Walmart says OpenAI-trained workers reported a 27% confidence gain

Walmart says associates who completed OpenAI training reported a 27% gain in confidence using AI. The retailer offers free learning and certification programs with OpenAI and Google, and received more than 1.7 million responses to its first AI experience survey. The reported gain measures how capable workers feel, not their skills or productivity, and does not establish that training caused the improvement.

research

Google says offline learning helps AI retain new knowledge

Google Research’s latest overview revisits a paper submitted in June, describing a “sleep” framework that pauses new inputs to consolidate information into model weights. The authors report near-zero forgetting in sequential translation tests and near-perfect scaling on BABILong through 10 million tokens. That result reflects information compressed into weights, not attention over one enormous prompt. The architecture and training rules remain human-defined.

Clueso turns rough screen recordings, slide decks, docs or screenshots into polished, on-brand videos in minutes.

Also worth knowing

•  Anthropic says Claude cut drug-document review time by 89%

•  OpenAI reportedly seeks to buy Google’s AI chips

•  Amazon tells engineers to judge AI by results, not usage

•  Former Meta director builds Ndoc as an AI alternative to Office

•  Microsoft takes two AI image models out of preview in Foundry

•  Daniel Wong’s campaign removes AI endorsement video over consent dispute

•  Five AI models favored branded drugs in a study, with higher projected costs

Daily tool drop

6 AI tools worth knowing today

The Internet Had a Point

Meme from the stashClaude Saved the File Somewhere

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.