Google DeepMind calls Genie 4 its most realistic and interactive world model yet. Point a camera at a text prompt and you get a live, physically consistent space you can walk through, where objects fall the way they should and people in frame react instead of looping. Built on Gemini Omni, the announcement pulled 2.2 million views and climbed past 16.8 thousand posts in its own trending topic within a day.

The category underneath this launch is bigger than one demo. Fei-Fei Li's World Labs raised a billion dollars for Marble, its spatial intelligence model, with NVIDIA and AMD both putting money in, and is reportedly in talks for a new round near a five billion dollar valuation. Decart took the same bet toward driving simulation specifically: Oasis 3 generates hours of photorealistic road footage through an API, backed by 300 million in new funding that puts the company near a four billion dollar mark. NVIDIA went the open route with Cosmos 3, a world foundation model it ships for robotics and autonomous vehicle perception rather than licensing as a product.

Four labs, four different bets on what a world model is actually for. Google wants a creative canvas you can generate on demand. World Labs wants a spatial substrate for robots and 3D design. Decart wants a simulator cheap enough to replace real test miles. NVIDIA wants the open layer everyone else builds a robotics stack on top of. The earlier version of Genie shipped at 250 dollars a month for early access, which tells you the compute bill for rendering a physically consistent world in real time is still the whole story here. Whoever gets the cost per generated minute down first is the one I'd bet actually ships a product instead of a demo reel.

https://x.com/GoogleDeepMind/status/2115114028338962672