Peter Yang Tests Genie 4: Convincing Worlds, Unfinished Gameplay

I tested an early version of Google’s Genie 4 by creating five worlds: a San Francisco street attacked by a giant creature, a Tokyo ramen stand, a dragon battle, a safari and an alien attack on New York. I will also discuss what the results might mean for movies and games.

I describe a world model as predicting the next visual frame based on user input. A prompt creates an initial scene; pressing a movement key changes what the model predicts next, and the process repeats while retaining information about what has already appeared.

The features I highlight are generated sound, the ability to change a scene with a live prompt, and Gemini 4 planning multi-step events behind the scenes. Sound can include voices, engines, wind or music.

For the first world I provide prompts for the environment and character. A San Francisco street appears in about ten seconds, with a cable car and a first-person character holding coffee and a phone. I add a golden retriever and a warm glow.

I then ask for a giant monster to destroy buildings and breathe fire down the street, followed by a bridge collapse. The scene becomes chaotic, although the character continues holding the phone. The demonstration shows how I can alter the same scene as it runs.

In the second scene, a ramen chef welcomes me to a rainy Tokyo stand and speaks. I ask for ramen and for the scene to show me eating it. The chef serves two stacked bowls, but the requested eating action does not appear.

I add a sleeping cat, warm sake, cherry blossoms and a saxophone solo on the radio. Music becomes audible. The scene is evocative, but the stacked bowls are one example of an imperfect result.

For a castle battle I generate a rider on a dragon, then request a flame burst, arrows and a dive. The generated sound resembles a person imitating a dragon more than the creature I expected. I also ask for a spaceship to fight it; the ship appears and then seems to vanish.

Next I create a jeep safari with elephants, lions and antelope. The jeep does not seem to drive. When I ask a lion to chase an antelope and then request a migration, animals appear abruptly and the action is not always coherent.

For the final scene I request a first-person shooter style and a rock soundtrack as aliens attack New York. The world produces voices, a mothership strike and a helicopter. When I ask the gun to shoot and eliminate the aliens, however, the gameplay does not work as I hoped.

I do not want to present the technology as better than it is: in this test it is good at producing graphics but not yet at producing satisfying gameplay. A good game also needs a core loop, progression and mechanics beyond its appearance.

For movies, I think world models could eventually reduce the cost of visual effects by generating alternative takes of elaborate scenes from prompts. For games, they may allow people to create personalized worlds, but the shooter demo shows why graphics alone are insufficient.

I compare the technology to an early stage of language models: it can already create impressive scenes, but objects appear unexpectedly and there are still bugs. If the models improve, I can imagine people prompting short films and games; that remains a future possibility rather than what this demo already delivers.