DeepMind announces Genie 4, a world model that generates scene audio and dialogue in real time

I still think world models become the simulation engine for AAA games. The game binary never ships, it runs on their cloud and streams to you, so Stadia plus world models is just a new DRM layer.

I cancelled Stadia because streaming games over the internet sucks: input lag and ugly graphics. AI-generated games are worse, no skill or taste, and regenerating a scene and shipping it over the pipe is cost prohibitive.

Game studios still can't put regular LLMs in games without catastrophic edge cases. I think world models stay a fancy tech demo until that changes.

Why would I want a non-deterministic engine when I can have a deterministic one distributed to consumer hardware? Write the code once and ship it.

I don't see anyone playing that trash, and the economics don't work. Running a world model for every player costs way more than they'll make from microtransactions.

I think it's cheaper to write the game code once and host a download than to run hundreds of thousands of GPUs simulating a unique world for millions of players.

Genie 4 looks worse than Genie 3 for player control to me. It doesn't seem to respond to inputs; it looks like AI videos with control overlays pasted on top.

I don't get who these experiences are for. Google could hire good game developers for a couple million and get a real product, but instead it has PMs who don't play games designing for Bart Simpson.

I think the target audience is people making commercial videos, not games.

That was already a problem with Genie 3, and now they've gone all in on video generation over controllability. The original Genie had a purpose: agents learning about the world. Most of the people behind that, Jack Parker-Holder, Tim Rocktaschel, Demis, have left DeepMind.

I want to know if this exports 3D assets or an Unreal, Godot, or Unity project, or if it's stuck in Google's runtime.

I thought the stained glass landscape clip was cool. I wonder if that effect could work in a real-time game engine.

I read the prompt: 'A REAL, ultra-photorealistic, high-end 35mm cinematic first-person perspective,' and it still looks like a video game rendering. Google's prompting is doing a lot of work there.

I noticed the marble bust dropped on a watermelon should have toppled over at that point if the physics were modeled.

They were demonstrating steering, not physics.

This is the hole in the whole world model idea for me. We keep building simulations and training models on them, but the model that actually learns the world can simulate it internally and beat the simulation. The simulation ends up beside the point.

I think real-world things don't behave as expected, and a robot could use this as a mental model corrected by actual events. We all have a world model in our heads.

I thought the 'do you really think we're going to make it?' scene was out of sync, and it shows the problem with AI video capturing humanity. The painting-to-interactive scene is astonishing though.

I think world models capture physical complexity and teach AI hidden laws, which helps robots and autonomous vehicles. AMD buying World Labs is a move against Nvidia here, and Google might see this as a gateway to renting out autonomous vehicles and humanoids.

That all sounds right to me, but they also use this for generating videos and 'unlocking new interactive experiences.' Those goals compete with the robot and vehicle story.