AI startup Decart has officially launched Oasis 3, its advanced interactive world model capable of generating photorealistic driving environments in real time. This unveiling follows the company's recent $300 million funding round, which elevated its valuation to nearly $4 billion. The new model is now accessible to developers through an API, signaling a strategic push into the physical AI sector.
A New Tool for Autonomous Development
Decart is initially targeting autonomous vehicle companies that require extensive simulation of rare and complex driving scenarios for training and testing. Oasis 3 generates physically accurate, multi-camera environments, providing a crucial tool for developing robust autonomous systems. The service is priced at $0.02 per second of generation, with custom enterprise pricing available for specific use cases.
Fostering a Developer Ecosystem
The company's larger ambition is to cultivate a developer ecosystem around its world model, mirroring OpenAI's successful strategy with language models. By providing immediate API access, Decart aims to empower a community to build novel applications on its platform. CEO Dean Leitersdorf anticipates that this open approach will lead to a wave of unexpected and innovative uses for the technology.
This strategy builds upon the foundation of Decart's existing community of over 100,000 developers using its real-time video model, Lucy. Oasis 3 represents the company's significant expansion from e-commerce and livestreaming applications into the realm of physical AI. The recent funding round included strategic investors like Toyota, Adobe, and eBay, who Leitersdorf notes are all potential customers.
The Competitive Edge in Efficiency
Oasis 3's key advantage lies in its photorealism and its capacity for infinite generation, made possible by remarkable efficiency. This is powered by the Decart Optimization Stack (DOS), a proprietary software that optimizes model performance on various hardware platforms. This vertical integration allows Decart to operate at a fraction of the cost of its competitors in the growing world model arena.
Acknowledging Current Hurdles
Despite its impressive capabilities, early testing reveals that the model is not without its limitations, which are common in this nascent field. During extended use, the thematic integrity of the generated world can degrade, and user controls can become unresponsive. These issues highlight the significant computational challenges involved in maintaining long-term consistency within a dynamically generated environment.
Another significant challenge is the model's current inability to simulate physics accurately, as evidenced by vehicles passing through one another. Leitersdorf identifies this as a major research problem the team is actively working to solve, attributing it to a data imbalance favoring normal driving over accidents. The model's auto-regressive nature, generating frame by frame, also contributes to the difficulty of maintaining physical consistency.
Future Enhancements and Vision
To address these issues, Decart is focused on improving the model's memory by researching methods to expand its context window. The goal is to enable the system to retain more information over longer periods, thereby enhancing environmental coherence. Leitersdorf believes the next version, which will allow generation from video prompts, may partially solve the consistency problem.
The launch of Oasis 3 marks a significant development in the commercial application of interactive world models, backed by substantial new capital. While the technology is still in its early stages and faces notable hurdles in consistency and physics simulation, Decart's developer-first strategy and focus on efficiency position it as a key player. The platform's evolution will be closely watched as it opens new possibilities for simulation in autonomous systems and beyond.