AI companies Etched and Decart have built "Oasis," a fully AI-generated video game that runs in real time.
The game lets players move, jump, collect items, and break blocks in an environment similar to Minecraft. According to Decart, their system processes player inputs instantly to create gameplay elements, including physics, rules, and visuals.
The system's speed stands out compared to other AI video generators. While models like OpenAI's Sora or Runway Gen-3 need 10–20 seconds to create one second of video, Oasis generates 20 frames per second with "zero latency"—a new frame every 0.04 seconds.
Decart says Oasis uses custom AI architecture based on Transformer technology. Unlike other models, it can create images instantly and respond to player actions. The system combines Vision Transformer technology with a diffusion model for image processing, letting players control the game with standard WASD keys and mouse input.
One challenge in generating a continuous game world with AI is maintaining consistent, high-quality visuals over time. Small errors can accumulate with each new frame, leading to a blurry or distorted image.
By comparison, Google DeepMind recently created an AI system that can simulate Doom, but only for three-second intervals. Oasis still has limitations, including blurry distant objects, inconsistent object rendering, and inventory control issues.
To improve this, the Oasis developers used a technique called "dynamic noising". Essentially, they intentionally add some randomness or "noise" to the image data in the early stages of generating each frame.
This helps prevent errors from building up and becoming noticeable. Then, as the AI model refines the image, this artificial noise is gradually removed. This allows the model to fill in sharp details while maintaining consistency with previous frames. The key is that the AI was exposed to this type of noise during its training process. It has learned how to deal with the noisy data and still produce a clean final result.
The road ahead
Decart believes larger models and datasets will fix many current problems. Their partner Etched is developing a specialized AI chip called Sohu, designed to handle larger models with up to 100 billion parameters in 4K resolution.
Currently, Oasis runs on H100 GPUs at 360p resolution. Decart says operating costs are "much cheaper than what a user pays for an average hour of a Steam Game," though this comparison isn't entirely clear since Steam games typically have one-time purchase prices rather than hourly rates. Decart seems to be estimating the Steam cost per hour based on average playtime and game prices.
Decart and Etched have released the code and weights for a 500-million-parameter model on GitHub that can run locally on GitHub, along with a playable demo of a larger version.
Venture capital firm Sequoia Capital, which led Decart's seed funding round, sees major potential in the project. "Today’s launch of Oasis, powered not by a gaming engine but by a single AI model trained on videos, marks a milestone in inference," Sequoia stated.