Imagine a video game that isn't pre-programmed but dynamically generated by artificial intelligence in real-time. This is the promise of Oasis, an innovative AI model developed by Decart and Etched. Released on October 31, 2024, Oasis represents a significant leap towards creating complex, interactive virtual worlds powered entirely by AI.
Oasis is an experiential, real-time, open-world AI model. Unlike traditional video games relying on pre-scripted events and environments, Oasis generates its world dynamically based on user input. This makes every interaction unique and unpredictable.
Key features of Oasis:
Try the Oasis demo here to experience the innovative AI model.
Oasis utilizes a transformer-based architecture consisting of two main parts:
The model is trained using Diffusion Forcing, allowing it to handle noisy data and generate coherent frames even with imperfections. Oasis generates frames autoregressively, conditioning each frame on user input for real-time interactivity.
A key challenge in autoregressive models is maintaining temporal stability. To address this, Oasis employs dynamic noising, injecting noise during the initial diffusion passes to reduce error accumulation and gradually removing it to preserve high-frequency details.
Interested in the engineering behind this groundbreaking model? Delve deeper into the intricacies by exploring the Decart blog post.
To delve deeper into the technical architecture, you can also view the code and explore the model weights on Hugging Face.
Oasis achieves an impressive 20 frames per second, a significant improvement over state-of-the-art text-to-video models. This real-time performance is made possible by Decart's inference stack.
To further accelerate the model and reduce its cost at scale, Oasis is optimized for Etched's Transformer ASIC, Sohu. Sohu is designed to handle massive, next-generation models in 4K resolution, potentially serving 10x more users than current hardware.
While Oasis demonstrates exciting results, there's still room for improvement:
The developers of Oasis are focusing on scaling the model and datasets to address these limitations. They also recognize the need for breakthroughs in inferencing technology to ensure a sustainable latency and cost trade-off.
Oasis is more than just a technical demo; it's a glimpse into the future of interactive AI. It showcases the potential for AI to generate complex, dynamic worlds that respond to user input in real-time. This technology could revolutionize various fields, including gaming, education, and simulation. With future iterations, experiences could be driven entirely by text, audio, or other modalities, opening up limitless possibilities.
Oasis is optimized for Sohu, the Transformer ASIC built by Etched. Read more about Oasis and Sohu on Etched's blog.
Check out these related-articles:
