Imagine a world you can type into existence. Not a static image or a pre-rendered video, but a living and breathing 3D scene with which you can walk around.
That is Genie 3, DeepMind’s newest breakthrough. And no, it’s not magic, although it almost feels like it.
Genie 3 takes simple one-line prompts such as “a snowy ski resort with deer” and constructs explorable virtual worlds in real-time.
We are talking about smooth 720p, 24 frames per second. Unlike the previous models, this one is equipped with memory.
So, if you set an object down, go for a stroll, and come back, the object will still be there. Such consistency is sublime for training the AI. It makes the machines aware that the world does not reset for no reason while unobserved. It involves understanding the basics of physics, persistence, cause, and effect.
Now the fun part: Genie 3 does more than just build worlds. It reacts to you.
Want to change the weather midway through a scene? Add a new character? Give life to a dog from thin air? Go on.
The scene adjusts, and there is no need to restart, simulation, sandbox, improv theatre, or listen to your script.
And here’s where the real AGI conversations begin. Genie 3 is not just visual eye candy. It’s more like a training ground for embodied agents, embodied AI that has bodies, goals, and the ability to learn through doing.
Drop an agent into any of these worlds, and it can learn to navigate, solve tasks interactively, and so on, just as we do, one giant leap towards a general intelligence capable of acting, not just chatting.
It’s not perfect, of course. Genie 3 has many limitations. You can never have a session running endlessly. Text in the scene is sometimes not very clear.
Realism often evokes a dark reminder of an early video game, which likely delayed the product’s mass release due to the engine and UI being so fresh.
And it’s still out of reach for most creators and researchers. So, you won’t realistically be able to produce your dream landscape this weekend.
But don’t let that fool you – Genie3 is laying the fundamental bricks on the road toward AGI.
By providing AI with interactive, memory-rich worlds that change in real-time, it offers the AI one thing it’s never truly had: a place to grow up, a place to get clumsy, stumble upon wonders, and explore interesting things, to learn from.
So yes, of course, the genie is out of the bottle-and this one doesn’t grant you wishes, it builds worlds.
