AI startup Decart unveiled Oasis 3, its latest interactive world model capable of generating photorealistic driving environments in real-time. The model is currently accessible via API, targeting autonomous vehicle companies needing to simulate rare driving scenarios at scale, with plans to expand into robotics and other physical AI applications. Dean Leitersdorf, co-founder and CEO, stated, "It’s going to be the first usable world model that people can actually program on top of."
Decart boasts a community of over 100,000 developers, many building products on its real-time video model Lucy, primarily in e-commerce and livestreaming. Oasis 3 builds on this foundation, marking the company's push into physical AI. Access is priced at $0.02 per second, with enterprise pricing varying by use case.
Oasis 3’s edge lies in the photorealism and infinite generation capability of its models, thanks to efficiency innovations powered by the company's main product: the DOS (Decart Optimization Stack) software, which enables models to run efficiently on Nvidia, Amazon, and Google hardware at a fraction of competitors' costs.
The model generates physically accurate, multi-camera environments for training and testing, allowing developers to create scenarios infinitely—ideal for autonomous vehicle developers testing edge cases. While Oasis 3 produces the most photorealistic environments I've encountered compared to other models like Google’s Genie 3, the model significantly degrades with prolonged use.
In testing, I found it could create a strong initial scene matching the prompt, but thematic integrity rapidly deteriorated as I navigated the world. The model also struggles with physics simulation, as cars can pass through one another. Leitersdorf acknowledges this as a major research challenge, attributing it to the disparity in data availability between good driving and accidents.
To maintain consistency, the Decart team is working on extending the model’s memory length. "Every frame we generate is roughly 8,000 tokens," he explained. "Generating this at tens of frames per second means hundreds of thousands of tokens per second. We’re researching how to do longer context to store millions more tokens."
Leitersdorf believes the consistency issue may be partially addressed in the model's next version, which will enable users to generate worlds based on videos rather than images. Although current limitations exist, he is focused on the potential when developers access the technology, reminiscent of the early LLM days when OpenAI invented the API for models.
Blogger's Review: Decart's Oasis 3 shows tremendous potential in realistic driving simulations, yet the limitations in physical simulation need addressing. With the rise of a developer community, the future applications are promising and could accelerate advancements in this technology.