Recently, Oasis was highlighted on social media as the world's first "AI video game" capable of responding to complex user input in real-time. If you're interested, a scaled-down local version's code is available on GitHub. More detail and background can be found in the project's write-up, which discusses its potential and various limitations. We believe the creators emphasize the ability to handle complex user input—such as mouse movement and inventory management—as a key distinction from AI-generated DOOM. The latter was a concept that showed AI image generators could (to some extent) function as real-time game engines. Image generators essentially act as prediction machines.

By giving a trained model a short history of recent events plus user input as context, it can quickly generate a usable prediction of what should happen next, allowing for interactivity. This loop can produce impressive clips for social media. It's an intriguing idea, and we commend the creativity of adapting an image generator for this purpose, but the limitations are evident. If you pause or explore dark or repetitive areas, the system falters, leading to what can only be described as a "dreamily broken" state.