Generative AI has come a long way. In the past few years, we have learned to generate images, poems, videos, and even entire apps and websites. However, the next wave could be the most impressive. Google DeepMind recently released its latest tool, Genie 3. Google describes this as a "world model." What that means more specifically is a tool that can generate entire digital worlds from a single prompt.
Imagine being able to create an entire explorable world from a single prompt. This could change the world of gaming, education, movies, and more.
While this isn’t the first version of the tool — hence the 3 in the name — it is the first time the tool has reached a level where it could have genuine use cases out in the real world.
So what exactly is this tool, and what does it mean for the future of AI?
What is Genie 3?
Genie, or its less catchy name, Generative Interactive Environments, is a world model tool that lets a user generate and then explore virtual worlds.
It is trained on internet videos, and, like other generative AI tools, simply requires a worded prompt to get started.
The first version of Genie was limited. It could generate worlds, and allowed for frame-by-frame interaction from the user. However, it was primarily operating in 2D game-like environments and didn’t offer anywhere near the quality that is available now.
Then, Deepmind came out with Genie 2. It scaled the tool to offer immersive 3D environments, improved control and a noticeable improvement in real-world physics and graphics.
Now, Genie 3 has gone a step further. Worlds are generated at 720p resolution and play in 24 frames per second and a huge amount of work has gone into making the explorable worlds even more interactive.
What is important to note about the advancement through these models is that going from Genie 1 to 3 only took about a year and a half. In other words, the next version of Genie is unlikely to actually take a very long time.
What can Genie 3 do?
Once you give Genie 3 a prompt, what will it actually generate? Effectively, Genie is creating a "world." You can control it with a keyboard or a touch screen, and the world can stay coherent for a few minutes. It even remembers off-screen objects that you’ve moved away with for up to a minute, with any changes you’ve made to them staying intact.
These worlds generate on the fly. That means that, in theory, they are infinitely explorable, with new parts loading as you move around them. However, the limit is how long the model can keep things consistent and in memory.
The details will begin to drift and fall apart after a few minutes. In other words, this would better apply to small regions like a house, instead of entire explorable worlds and cities.
An important feature that DeepMind has explored with Genie 3 is how editable these worlds are. You can trigger world events mid-play, adding objects or changing weather, and the model should, in theory, keep up with the change.
DeepMind has indicated that one of the key differences in this model is its understanding of real world physics. It can generate vibrant eco-systems and can replicate animal behavior and intricate plant life.
There are still plenty of limitations here. Genie 3 can’t always simulate real-world locations with absolute accuracy. It also struggles with creating text in these worlds and accurately recreating more complicated events in the world continues to be a struggle.
However, seeing the progress through the Genie models shows that these challenges are likely temporary, with improvements happening quickly.
What is the future here?
Currently, Genie 3 is only available to certain developers for testing. However, DeepMind sees multiple uses for the technology.
One of the more obvious uses is in game generation, since Genie 3 is able to quickly and effectively generate large, detailed worlds that are interactive. However, this could be used in other areas too.
DeepMind explains how it could be used to train robots on factory floors, something that is becoming more and more common. This kind of technology could create environments for them in which they could be trained effectively.
Outside of these areas, Genie 3 could make training programs more interactive on a more affordable level. Or a much more simple use, allowing people to explore the world virtually.
This technology could generate historical landmarks that are no longer around, or recreate cities historically for different time periods. Currently, the technology isn’t capable enough for these tasks, but it does mark a huge jump towards them.
More from Tom's Guide
- I tested ChatGPT-5 vs Grok 4 with 9 prompts — and there's a clear winner
- Apple plans home robot invasion with lifelike Siri that 'injects itself into conversations' — and there's a lot more devices
- 5 AI tools students need to excel in 2025 — without cheating
You must confirm your public display name before commenting
Please logout and then login again, you will then be prompted to enter your display name.
