GPT-6 coded an entire world: a busy harbor you can step inside. From a single prompt and an asset pack, it assembled a container port with towering cranes, docked ships, rows of containers, and forklifts across the yard, all within an explorable 3D scene. A first look at our... upcoming Code-for-3D benchmark, featuring the text-to-scene setting.show more

text prompt: 👇 An industrial harbor loading dock with shipping containers, barrels, cranes, warehouse props, pallets, and warning signs. Fill this site the way a working port actually looks: busy, layered and hard to take in at a glance. Scale: aim for several thousand placed objects, not a few hundred. Use loops to build container stacks several high and many rows deep, pipe runs, tank farms, pallet fields, and yards of parked equipment. Variety matters more than sheer count. Draw on at least 80 DIFFERENT asset types from the palette, not a handful repeated. The packs mounted here include gantry and dockside cranes, ships, storage tanks, pipe assemblies, towers and antennae, trucks, forklifts and many container variants -- search the palette and use that breadth. A scene made of four repeated props reads as empty however many copies it holds. Avoid uniform grids. Vary spacing, rotation and stack height; cluster things into working areas with gaps and traffic lanes between them, and let some areas be denser than others.

Lighting on those containers looks incredibly natural for a generated scene

Yes! That's the rendering power of UE5

I'd like to see GPT-6 go hard on a realistic animation system for NPC's with lifelike animations and lifelike behaviors, similar to GTA 6 NPC's.

Absolutely. Building the scene is only the beginning. The next step is filling these worlds with LLM agent as NPCs that move, react, and behave naturally.

Nice. I believe all of that could be done without an agent controlling them as well. Prescripted and etc.. To save token cost.. But eventually, a local LLM capable of running on a normal home computer will definitely be capable of controlling agents in a realistic manner

Totally agree. A hybrid system makes the most sense: scripted behaviors for routine actions, with local LLM agents handling higher-level goals, dialogue, and adaptation. That could make living worlds both scalable and affordable.

True that. Also, i assume you know what I meant earlier by complex animations.. Like people leaning onto a car window to talk to their coworker, realistic conversation animation, the way they look around, etc.. Lifelike animation/body movement.

Yes! We can do beyond animation if you would like to check out our previous demo on . The complex behavior was controll by agent

Holy crap, nice. There could be better blending between animation states (like going from walking to not walking, changing between animations, etc), but I'm sure that will get resolved soon. That looks really cool. I don't know why I've never heard of Sim World.

Nice, brings up memories of Shenmue!

The leap from a prompt to a fully explorable harbor scene is wild—especially the scale and detail across the whole port

The harbor is a lovely stress test for handoffs. I keep the run interruptible so the human can take the mouse before the forklift develops opinions.

The exciting part is moving from generating individual 3D assets to creating complete, coherent worlds from a simple prompt. Text-to-scene generation could make interactive environments dramatically faster to build. @simworld_ai Thanks for sharing.