01What Is OpenAI Destroyed Sora? The End of a Video Tool

OpenAI destroyed Sora on March 24th, shutting down the app and its API without warning. Sora was a tool that generated beautiful videos from text prompts. But its downloads had dropped 75% from their peak, and total revenue was only about $2.1 million.

For a company valued at hundreds of billions, that amount was basically nothing.

Internal sources said Sora's GPUs were "melting under the strain of generating video." OpenAI decided to stop burning compute on Sora and redirect the entire research team toward something much bigger. That something is called a world model.

02Key Features of the World Model Pivot

Unlike Sora, a world model actually understands how the physical world works. It grasps physics, cause and effect, and spatial relationships. You can tell it "turn left," and it generates what happens when you turn left in real time, interactively.

  • It maintains a consistent environment where objects stay where you put them.
  • It can plan ahead by simulating multiple futures internally.
  • It is more like a city simulator than a painter, allowing you to walk around and make decisions.

The same team that built Sora—led by Bill Peebles, inventor of the diffusion transformer architecture—is now leading the world model push. They have a massive head start from two years of training video models on the physical world.

03How to Set Up a World Model (No Direct Setup Yet)

World models are still very early, and there is no public setup guide. However, companies like World Labs have launched a product called Marble. Here is how you can get started with it:

  1. Go to World Labs' website and sign up for Marble.
  2. Type a text description of the 3D environment you want to generate.
  3. You can also upload an image to guide the generation.
  4. The tool creates a persistent, explorable 3D environment.
  5. Download the environment for use in Unreal Engine, Unity, or a VR headset.
  6. Explore and interact with the environment in real time.

Note: Marble has limitations after a few steps, and it costs money. It is not a finished product yet.

04How to Use World Models: Examples from the Episode

The episode showed several examples of world models in action. Google DeepMind released Genie 3 in August 2025, which generates navigatable 3D environments at 720p and 24 frames per second from a text prompt. You can explore them, walk around, and interact with objects.

It maintains visual consistency for several minutes.

World Labs' Marble lets designers generate fully editable 3D environments from one text prompt. This replaces a workflow that used to take a team of 3D artists weeks. For agencies, you can create interactive 3D product demos for brands as easily as you generate a video today.

In robotics, companies train robots by running them through real-world scenarios. World models replace that: you simulate the environment, train in the simulation, and then deploy in reality.

05Results and Comparisons from the Episode

OpenAI's Sora had 75% lower downloads from its peak and earned only $2.1 million in total revenue. Google DeepMind's Genie 3 is the most technically impressive real-time interactive world model anyone has shipped, but it holds consistency for minutes, not hours.

The World Model Benchmark from Nuri IPS in 2025 found that even the best world models follow instructions correctly only 61% of the time. Marble hits limitations after a few steps, by its own co-founder's admission.

Despite these limits, the humanoid robot market is projected to hit $38 billion by 2035. Accenture estimates that virtual twin technologies could unlock $1.29 trillion in economic value by 2030.

06Tips, Limits, and Mistakes to Avoid

World models are very early technology. Do not expect them to work perfectly yet. They follow instructions correctly only 61% of the time, so you need to check and correct outputs manually.

  • Do not assume a world model understands everything—it can lose consistency after several minutes.
  • Do not rely on a single model for critical tasks; test thoroughly in simulation first.
  • Avoid expecting instant mainstream adoption—the episode says we are likely 12 to 18 months away from world models being mainstream for creative tools.
  • Do not ignore the robotics and automation angle—this is the one most people are not watching yet.

Remember: Sora was killed because it was a drag on resources. If you build on top of a dying model, you waste money.

07Who It Is For and Why It Matters

This matters for agencies, freelancers, creators, and e-commerce people. In the near term, world models will show up first in creative tools. Marble already lets designers generate 3D environments from text.

If you do product photography, virtual staging, game-ready assets, or architectural visualization, this technology is coming for your workflow fast.

For agencies building content for brands, interactive 3D product demos will become as easy to generate as a video is today. The episode says this could be mainstream in 12 to 18 months.

On a bigger scale, world models are how AI learns to operate in the physical world. They could unlock trillions in economic value by 2030. Pay attention because the AI race is moving beyond text and images into understanding reality itself.