What a 'world model' is — and what Decart's Oasis 3 does

You've heard of AI that writes text and AI that makes images. A 'world model' does something different — and Decart's Oasis 3, launched on 10 June 2026, is a good way to understand it. Here's the plain-English version, with the jargon translated and the honest caveats included.

Self-driving cars learn by practising. They need to handle thousands of different situations — a child running into the road, black ice, a sudden lane closure in heavy rain — but you can't safely create those on real roads over and over again. And waiting for them to happen naturally could take millions of miles of real driving.

With Oasis 3, an engineer can simply type a description — 'foggy motorway at night, a lorry pulling out two lanes ahead' — and the AI instantly generates that scene, in photorealistic detail, from multiple camera angles. The self-driving car's AI can then practise that situation as many times as needed, in every variation, without putting anyone at risk. Decart prices this at roughly $0.02 per second of simulated driving time — cheap enough to run thousands of scenarios overnight that would take months to capture on real roads.

The multi-camera detail is important here. Real self-driving cars see the world through multiple cameras at once — one looking forward, others looking to the sides. Oasis 3 generates all of those views simultaneously, so the simulation matches what the car would actually see. That is harder than it sounds: each camera view needs to be consistent with the others, in real time, as the scene changes.

After self-driving cars, Decart's next target is robots — the broad field sometimes called 'physical AI', which covers any AI that controls something in the real physical world, not just on a screen. Robots face similar training challenges to self-driving cars: they need to practise picking up objects in different lighting, navigating cluttered spaces, and responding to unexpected obstacles — ideally in simulation before being sent into the real world.

Decart CEO Dean Leitersdorf has a bigger vision beyond just driving and robots, though. He wants other companies to build their own products on top of Oasis 3 — the way thousands of apps were built on top of early AI services. Think of it as renting world-building capability instead of building your own simulator from scratch. That is the bet: that world models become a platform, not just a product.

To understand where Oasis 3 fits alongside other kinds of AI:

A world model is interactive and responds to actions — that's what makes it different from a video.

The honest summary is this: Oasis 3 is a genuinely new kind of tool — one that could meaningfully speed up how quickly self-driving cars and robots get good at their jobs, by giving them unlimited practice in realistic simulated environments. The company behind it has serious backing (roughly $300 million from investors including Toyota and Nvidia). But the simulations are not perfect yet, and the gap between 'realistic-looking' and 'physically accurate' is the thing engineers will be watching closely. It is a promising start, and the kind of technology that tends to improve quickly once real teams start using it at scale.