Oasis AI: Playable Open-World AI
Oasis is an interactive-world research project from Decart and Etched. Its October 31, 2024 release page describes an AI-generated video experience responding to keyboard input. It is a technical demo of a world model, not a production game with guaranteed rules, persistent saves or a stable inventory system.
Understand what the model generates
Oasis takes keyboard input and generates the experience's graphics, rules and apparent physics without a separate physics engine. The release describes movement, jumping, picking up items and breaking blocks. These are examples of the research demonstration, not a promise that every action behaves deterministically.
The project generates frames autoregressively, conditioning each frame on user input. That distinguishes the research question from simply rendering a video after supplying a finished prompt: the generated scene responds as the user acts.
Choose research access deliberately
The release announced code, weights for a 500-million-parameter model and a live demo of a larger checkpoint. Those are distinct access routes. A browser demonstration can introduce the interaction concept, while local model work is for readers prepared to inspect the project's setup, dependencies and license.
The release page remains accessible and links those research resources. This directory entry has not run the model or played the browser demo, so it does not establish current gameplay reliability, hardware suitability or local performance. Check the linked resource's current requirements before attempting an installation.
Evaluate the documented limitations
The authors identify problems with distant-image fuzziness, temporal consistency, domain generalization, precise inventory and object control, and long contexts. These limitations are central to understanding what the demo shows.
If you evaluate it, try revisiting a location or tracking one object's state through several actions, and note where continuity breaks. That is a suggested research observation, not a reported test result. Do not plan a persistent production experience around the release examples alone.
The page discusses larger future models and specialized hardware as research directions. Those projections should not be treated as capabilities of the released checkpoint. Oasis is useful for studying action-conditioned generative worlds; it is not evidence that an ordinary game engine or finished game product has been replaced.
