Your Game Engine Isn't Dead: What Decart's Oasis Actually Proved
AI-generated games run at playable framerates. But 'no physics engine' is the wrong takeaway. Here's what neural rendering really changes for developers. In October 2024, Decart and Etched released Oasis, a Minecraft-like game with a striking claim: there is no physics engine, just a foundation model. You press WASD, and a network generates the next frame. The headline takeaway was "engines are…
In October 2024, Decart and Etched unveiled Oasis, a Minecraft-inspired game that eschews traditional physics engines in favor of neural rendering. The game's headline boast—"no physics engine, just a foundation model"—has led some to speculate that game engines are obsolete. However, the real lesson from Oasis may be more nuanced.
Oasis leverages next-frame prediction, anticipating the subsequent game frame based on player input, trained on extensive gameplay data. This approach employs Transformer-based architectures, including a Vision Transformer autoencoder and a diffusion backbone, to generate statistically realistic outcomes rather than simulate physical laws.
Initial impressions of Oasis were mixed, with early reviews criticizing its 360p resolution, sub-20 FPS performance, and issues like teleportation and unstable builds. These problems highlight the inherent limitations of neural world models that lack explicit state management. Games such as Google's Genie 3 have demonstrated that neural rendering can produce navigable, consistent environments at higher resolutions, albeit with limited long-term memory and persistence.
Oasis 2.0 and subsequent iterations have focused on real-time world transformation at higher resolutions, transitioning from a full game replacement to an interactive world model suited for specialized applications like robotics and simulation environments.
The evolution of Oasis reflects a broader trend in game development: shifting focus from replacing traditional game engines to creating neural models that efficiently simulate complex worlds where traditional authoring methods are impractical. While traditional engines handle explicit state management and deterministic gameplay, neural world models offer a more flexible but approximate approach to world representation.
They are particularly adept at generating diverse, immersive environments where correctness is valued over absolute accuracy. The future may lie in hybrid pipelines that combine the strengths of both explicit game engines and neural world models, leveraging the former for core gameplay mechanics and state management, and the latter for enhancing visual fidelity, lighting, and procedural content generation.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
