AI startup Decart has just pulled back the curtain on its latest innovation, Oasis 3, an interactive world model designed to generate incredibly photorealistic driving environments in real time. This groundbreaking model, now available via API, marks a significant leap forward in generative AI and simulation technology.
Initially, Decart is setting its sights on autonomous vehicle companies, offering them an unparalleled tool to simulate rare and complex driving scenarios at scale. However, the company’s vision extends much further, with plans to expand into robotics and a diverse array of other physical AI applications.
The bigger play here, according to Decart co-founder and CEO Dean Leitersdorf, is building a vibrant developer ecosystem. By offering API access from day one, Decart aims to replicate the success OpenAI found with language models, fostering a community that will discover novel uses for world models.
Leitersdorf enthusiastically stated, “It’s going to be the first usable world model that people can actually program on top of.” He anticipates an entire developer community emerging around this technology, much like the one that grew around their existing real-time video model, Lucy, which already boasts over 100,000 developers in areas like e-commerce and live streaming.
Oasis 3: Powering the Future of AI Simulation
Oasis 3 is built upon the robust foundation of Decart’s existing models and represents a strategic push into physical AI. Access to this cutting-edge tool is priced at $0.02 per second, with flexible enterprise pricing tailored to specific use cases.
The release of Oasis 3 follows a massive $300 million funding round that boosted Decart’s valuation to nearly $4 billion. This impressive investment reflects the surging demand for Decart’s models across e-commerce, live streaming, and physical AI, attracting strategic investors such as Toyota, Adobe, and eBay, alongside existing backer Nvidia.
What truly gives Oasis 3 an edge in the increasingly competitive world model arena—which includes players like Google’s Genie 3 and World Labs’ Marble—is its exceptional photorealism and infinite generation capability. This prowess is largely thanks to Decart’s proprietary DOS (Decart Optimization Stack) software.
The DOS software allows Decart’s models to run with remarkable efficiency on hardware from Nvidia, Amazon, and Google, making them significantly more cost-effective than competitor offerings. Leitersdorf notes, “By being so vertically integrated, we’re able to be more than an order of magnitude cheaper than anyone else in the industry.” In fact, the startup has reportedly burned through “drastically less” than $100 million in its lifetime, a testament to its efficiency.
Navigating the Current Limitations
While Oasis 3 sets a new standard for photorealistic environments and infinite scenario generation, offering physically accurate multi-camera views perfect for training autonomous systems, it’s still in its early stages. My own testing revealed that while the system excels at setting up a strong initial scene matching the prompt, its thematic integrity can degrade rapidly over extended “drives.”
For example, a beautifully rendered “New York City street in the morning” might gradually morph into a generic urban landscape as you move further, with the initial intersection vanishing if you try to turn back. Moreover, the controls felt somewhat unresponsive, leading to occasional loss of control over the virtual vehicle.
A more critical issue, also observed in other world models, is the occasional lack of proper physics simulation—cars can sometimes drive straight through other vehicles. Leitersdorf acknowledges this as a “major research problem” stemming from the disproportionate amount of data available on “good driving” versus accident scenarios.
This consistency challenge is deeply rooted in Oasis 3’s autoregressive architecture, where it generates one frame at a time, referencing previous frames. Generating “hundreds of thousands of tokens per second” means the model’s context window fills up very quickly, making long-term environmental consistency difficult to maintain.
The Road Ahead: Evolution and the Developer Vision
To address these memory and consistency issues, Decart is actively researching ways to improve the model’s “context window” and compress memory into fewer tokens. Leitersdorf believes that the next version, which will allow users to initiate world generation from a video rather than a static image, will offer a partial solution to the current consistency challenges.
Acknowledging that world models as a field are still nascent, Leitersdorf is focused less on current limitations and more on the transformative potential once developers gain access. He draws a parallel to the early days of LLMs, where OpenAI’s API catalyzed an explosion of innovative applications from a burgeoning developer community.
“When we talk again in three months, we’ll be like, ‘Here’s 100 developers that all built 100 different applications with Oasis that surprised all of us,’” Leitersdorf optimistically predicts. Decart is banking on this collective creativity to push the boundaries of what’s possible with interactive, photorealistic world models.
Source: TechCrunch – AI
