Wink Pings

Genie 4: The Highest Peak in World Models, and the Unfinished Slope

Google has released the Genie 4 world model: real-time native audio, improved visual quality and physics, and prompt-triggered event generation are now implemented, and the demo is stunning. But the official team itself listed three limitations: long generations drift, real-world locations are not accurately reproduced, and sessions last only a few minutes. The API may open before Ultra membership; the details of the built-in Gemini agent are more worth studying than the visuals.

Google has released Genie 4. The blogger Guizang gave a very direct assessment: the highest mountain in the world-model field.

This update has three substantive changes. It generates ambient sound in real time; in the video, a motorcycle runs along a coastal road, and you can hear the surrounding sounds. Image quality and physics are noticeably improved. The ability to generate events via prompts has also been implemented; during generation, prompts can be used to change the direction of events.

The demo really is stunning. But Ultra members cannot experience it yet; Guizang says he hopes to be able to try it soon.

Don't rush to conclusions. Netizen TutuGoBig reminds us that the official team itself wrote three limitations: if it runs long, the scene drifts; real locations are not accurately reproduced; and sessions are still only a few minutes. Genie 4's current strength is the visual and audio experience of short-term generation.

The more interesting information comes later.

Developer Road N mentions that the API will also open, and developers may get to try it before Ultra members. For people building products and doing research, this is far more important than member access.

Eric Zhou adds a detail that is easy to overlook because of the visuals: Genie 4 has Gemini 3.8 Flash built in as an in-loop agent, responsible for breaking goals into steps and then driving the scene. This means what it generates is not just good-looking footage, but an environment that an agent can operate. Once you understand this, you can understand why some people are shouting "Earth online."

Ding, a Physical AI practitioner, offers a more practical angle: the greatest value of world models for industry is giving robots a cheap trial-and-error ground—first hit walls in a generated world, then go onto real production lines.

As for when ordinary people can get hands-on? There is still no certain word. Netizen wcg8904 puts it bluntly: "Don't let it be another case of a godlike launch event and then half a year of waiting for real-world rollout." It's not pleasant to hear, but it hits the mark.

Genie 4 is indeed the peak. But the path to the summit is still a stretch short.

发布时间: 2026-10-28 10:28