Entertainment news Published
Google's Genie 4 AI Generates 2-Minute Worlds With Sound
Google has released Genie 4, a new world-generation AI model, as a research preview. According to Google, it can create more realistic worlds, generate audio, and let users change the experience in real time with natural language instructions. The generated experience reportedly runs about two minutes, up from roughly one minute with the previous generation, Genie 3.
Translated from Japanese with AI. Read the original (Japanese)
The facts
What Is Genie 4?
Google has released a new AI model called Genie 4 as a research preview. A research preview is an early release shown to a limited audience for research purposes before a product is broadly sold.
Genie 4 is a type of AI known as a "world model," which generates an animated virtual world itself from instructions such as text or images. According to Google, Genie 4 can create worlds that are closer to reality than before.
Sources: theverge.com, techcrunch.com
Audio and Real-Time Control
According to Google, Genie 4 can generate audio. A key feature this time is that it can produce not only visuals but also the sounds heard within the world.
It can also reportedly change the experience in real time through natural language instructions. Natural language means instructions written in everyday words rather than programming commands. When a user gives instructions in words, the content of the world being generated changes on the spot.
Sources: theverge.com, techcrunch.com
Experience Length Grows to About Two Minutes
According to Google, the experience Genie 4 generates lasts about two minutes. The previous generation, Genie 3, ran about one minute, so the generation time has roughly doubled.
The longer the time, the longer the events within the world can continue to be shown. It marks a step from short demos toward experiences with a somewhat coherent flow.
Sources: theverge.com, techcrunch.com
Initial Access Limited to Trusted Testers
Initial access to Genie 4 is said to be limited to "trusted testers." This means a stage where only trusted parties try it first, so general users cannot use it right away.
Given its position as a research preview, the focus at this point is likely on verifying performance and safety. How far access will expand remains to be seen in future announcements.
Sources: theverge.com, techcrunch.com
Analysis
What the World-Model Race Means
World models are a field where companies are investing as the next stage after AI that creates text and images. If text-generation AI "creates words," world models can be said to "create experiences with motion and space." They have a wide range of uses, including games, video production, education, and simulation.
What stands out in this announcement is that the direction of progress converges on three points: longer, more realistic, and controllable. The shift from Genie 3 to Genie 4 shows up as extended generation time, added audio, and on-the-spot changes through words. This can be seen as a move toward bringing AI-generated worlds from something you "watch" to something you "participate in."
At the same time, the question of how far generated worlds can be trusted remains. With realistic visuals and sound, the risk that people mistake them for the real thing also grows. Limiting release to trusted testers may also serve to check these points.
As for the impact on Japan, there is strong interest in how this kind of technology will enter production settings in a country with strong game and video industries. However, it is still at the research stage, and no timeline for use as an actual product or service has been given. Future points of interest will be the expansion of access, further extension of generation time, and the quality of experiences combining audio and control.
Sources: theverge.com, techcrunch.com
Putting the Announcement in Context
This announcement takes the form of showing a research-stage model to a limited audience rather than launching a finished product. This approach is thought to aim at showcasing performance while picking up unexpected uses and problems at an early stage.
What matters for readers is to look separately at "what has become possible" and "when it will be usable." Genie 4 demonstrated capabilities such as about two-minute experiences, audio, and real-time instruction changes, but the timing and conditions for general availability have not been disclosed. The capability announcement and the start of service should be seen as separate events.
Sources: theverge.com, techcrunch.com
Why World Models Are Drawing Attention
Hearing only the term "world model," it is hard to tell what is new. Until now, AI has mainly focused on creating text or drawing images. In contrast, world models create space itself with the flow of time. The difference is that they can generate an entire experience in which objects move, sounds play, and scenes change based on instructions.
What Genie 4 showed in this announcement is that this world model has grown from a "short sample" to a "slightly longer experience." The extension from about one minute to about two minutes is not just a change in numbers. It means that events can now be shown in a somewhat coherent form from beginning to end. For readers, it is easier to understand by seeing AI-generated worlds as shifting from a single frame like a still image to an experience with flow.
What the "Research Preview" Format Means
Genie 4 was released not as a product for sale but as a research preview. Moreover, the first to try it are trusted testers, a limited group judged to be trustworthy. This format can be described as a way of release that communicates the technology's potential while also showing caution.
Why not release it widely all at once? One reason is that experiences combining realistic visuals and sound may have unexpected effects on how people perceive things. If worlds that can be mistaken for reality become easy to create, it takes more effort to confirm whether information is correct. Having a limited group try it first makes it easier to find such problems before they grow.
The meaning for readers is to understand that this technology is not "about to be available to everyone." Research preview is also a term that shows there can be a large gap between a capability announcement and the actual start of use.
The Difficulty of Extending Generation Time
The extension of experience length from about one minute to about two minutes can straightforwardly be seen as progress. However, the longer the time, the longer AI must keep a consistent world going. If objects suddenly change position or audio and visuals fall out of sync along the way, the experience breaks down.
In other words, extending time is also the result of tackling a difficulty separate from visual realism. Genie 4 is said to support audio as well, but keeping both visuals and sound aligned over a long period is thought to be far harder than creating a short sample.
This point serves as a guide when watching future announcements. When there is news that generation time has been extended further, we should also look at how the world's consistency is maintained over that longer time. We are entering a stage where not just the size of the numbers but the stability of the content is questioned.
Five story prompts by genre, each with a hook and a twist.
Questions
- Can anyone use Genie 4?
- No. Initial access is limited to "trusted testers," so only a limited group can try it at this stage.
- What can Genie 4 do now?
- According to Google, it can generate more realistic worlds and also generate audio. It is said to be able to change the experience in real time through natural language instructions.
- How long can the generated experience last?
- According to Google, about two minutes. The previous generation, Genie 3, was about one minute.