Google wants AI videos to stop changing between scenes
A ten-minute film shows connected scenes, but it tests just one of Google's four research systems.
By Saeed Ezzati7 min read
The audio edition
Listen to this newsletter
0:003:29
Read transcript
Google is testing four ways to keep an AI-generated story from losing track of its own characters, places and details. The challenge shows up when a video runs longer than a clip: a person’s clothes may change between shots, or a room may look different when the story returns to it. And a mistake in an early asset can travel through later production. These systems coordinate existing models, including Gemini and Veo; they are not a new video generator. One acts like a co-director. Given a person’s creative brief, it plans a story and visual style, directs agents to make storyboards, shots and audio, then asks a model to assess the assembled cut and guide another production pass. A second system, CANVAS, keeps references for characters, places and objects, so a story can retrieve them after moving elsewhere. A third, A-squared-R-D, generates video in connected segments. It saves context from earlier segments, then can either move into a new story beat or anchor a scene to people and places already shown. Google’s ten-minute film demonstrates that segment-based approach. It does not test all four systems working together. The fourth system, V-Q-Q-A, asks targeted visual questions about a generated video. Answers guide a revised prompt and another generation; it does not directly repair faulty frames. Google reports gains in consistency and character persistence, but the supplied account gives no figures. These are research systems, and selected examples do not establish how reliably they can hold together other long stories. That gap between a promising demonstration and dependable use matters even more when the stakes are personal. MIT researchers built a lightweight tool that estimates suicide risk in crisis-text conversations and highlights words contributing to its estimate. They tested it on about 16,000 de-identified exchanges from Crisis Text Line, using counselor-assessed risk categories and a clinician-reviewed list tied to 49 risk factors. The model distinguished those categories, including imminent risk. That is not evidence it can predict suicide attempts or improve interventions. A word match can miss context, too. The researchers are sharing the software and word list, but say clinical use requires further validation; human judgment remains critical. The same question of what works outside a test setting follows Feather Robotics, which has begun selling configurable robots in small quantities. The company says revenue has passed one million dollars, and prices its platform at roughly $30,000. Customers can choose AI from outside providers and adapt the hardware for particular jobs. Feather says its robots are cooking in Japanese restaurants and cleaning labs, but it has not named customers or disclosed how many machines are deployed. Sales show early demand—not reliable performance across those jobs. Customers still need to select or build the AI and application that make a robot useful on site. And Meta’s Muse puts the same proof question in front of an agent that can act across services. Meta has priced its Power plan at $20 a month and Maximum at $100, without explaining what separates them. Mark Zuckerberg also described a possible fee on purchases Muse completes; that is a future plan, not current revenue. New connections with Walmart, Best Buy, Gap and Spotify could give Muse more to do, but users must choose to link accounts. Meta says it uses secure virtual machines now and plans more privacy protections; planned features are not yet available. Across video, crisis tools, robots and agents, the useful test is not how capable a system looks in a showcase. Watch whether it performs reliably in real settings—and earns the trust and access that wider use requires.
