Google DeepMind previews Genie 4 world model with audio

In short: Google DeepMind has introduced Genie 4, an AI system that can generate interactive virtual worlds with sound, and access will expand through Google’s paid plans.

What happened

Google DeepMind says it has built a new version of its “world model” called Genie 4. A world model is an AI that makes a virtual world on the fly as you explore it, like an improvised video game that is created frame by frame instead of being fully built ahead of time.

DeepMind says Genie 4 improves on Genie 3 in several ways. It is better at physics, meaning objects are more likely to move and crash in believable ways. DeepMind also says people in the simulations should look and move more naturally, which may reduce the creepy, “almost human” look some AI videos have.

Genie 4 also adds native audio. That means it can generate background sounds, and users can ask for specific audio in a prompt. DeepMind also says it can create spoken dialog that changes as the scene changes.

DeepMind says Genie 4 uses a Gemini model as a guide to help keep the simulation on track. Even so, DeepMind notes the worlds can still lose consistency after a couple of minutes, meaning the simulation can forget how it started.

Why it matters

If this works as described, Genie 4 could help developers test ideas quickly, and help researchers create many different training situations for robots, like a flight simulator but for almost anything. For regular people, it points to a future where interactive, AI-made scenes could show up inside creative tools and games, but early access is expected to be limited and tied to higher-priced plans.

Source: Arstechnica