Oasis AI Review — Real-Time Generative Game Engine

Oasis AI, developed by Decart, is the world's first fully playable real-time generative game engine. Built entirely using diffusion models instead of traditional software rendering, it hallucinates a Minecraft-like 3D world frame-by-frame based on player inputs. While it exhibits visual hallucinations and memory limitations, it is a revolutionary proof of concept for the future of interactive neural simulation.

The Dawn of the Neural Game Engine

For decades, rendering a 3D environment meant passing polygon meshes, textures, and lighting data through a graphics card using rendering pipelines like DirectX or Vulkan.

Oasis AI completely bypasses this standard. Developed by AI startup Decart in partnership with Etched, Oasis does not run a single line of game code. There is no underlying game logic, no physics engine, and no spatial coordinate system.

Instead, a massive video diffusion transformer model predicts and outputs the next video frame based on the previous frame and your keyboard and mouse controls.

We tested Oasis AI to evaluate its responsiveness, visual consistency, and overall gameplay feel to see if generative models can truly compete with traditional game rendering.

Hands-On Experience: Navigating the Neural Minecraft

When you launch Oasis, you are dropped into a 3D, block-based world that looks and feels exactly like Minecraft. You can move around using the W, A, S, and D keys, look in all directions with the mouse, jump, and interact with the environment.

The model generates the world at approximately 20 frames per second. When you press "W" to walk forward, the neural network predicts how the pixels of the floor, sky, and distant hills should warp to simulate forward movement.

+------------------+     +-------------------+     +------------------+
| Current Frame    | --> | Oasis Transformer | --> | Next Frame       |
| (Pixels + Input) |     | (Diffusion Model) |     | (Rendered at 20p)|
+------------------+     +-------------------+     +------------------+

You can even "break" blocks or place them. When you click, the model understands the visual action of breaking blocks, rendering consecutive frames where the target block cracks and disappears, leaving a hole that reveals the background behind it.

However, the controls suffer from a slight floaty feeling. Because the model must infer the camera angle from mouse movements and then render the frame, there is a minor input delay that makes quick aiming difficult.

The Challenge of Persistence and "Memory Drift"

The most significant technical hurdle for Oasis AI is the lack of a permanent world database. The model has no memory of the world outside of its immediate attention window.

This leads to a phenomenon known as "memory drift" or visual hallucination:

• Object Mutation: If you look at a tree, turn 180 degrees away, walk for a few seconds, and then turn back, the tree may have turned into a stone block or disappeared entirely.
• Layout Shifts: The world morphs as you move through it. Exploring a cave and trying to return to the surface using the same path is nearly impossible, as the tunnel behind you will have naturally closed or changed shape.
• Structural Decay: If you build a house, the model will struggle to keep the dimensions consistent. Looking away and looking back will often cause the walls or doors to shift.

Essentially, you are playing inside a controlled dream. The world exists only as long as you are looking at it, which limits long-term playability.

Infrastructure and Cloud Compute Demands

Running a diffusion model at 20+ frames per second with sub-100ms latency requires astronomical amounts of computational power.

Decart achieved this by partnering with Etched, utilizing specialized transformer accelerators designed to run attention-based architectures with maximum efficiency.

Because of the massive cloud compute costs, the publicly available demo runs at a low resolution and is heavily compressed, giving the game a soft, slightly pixelated look.

Pros & Cons

Pros

• First playable, real-time 3D game engine generated entirely by a neural network.
• Physics, lighting, and environmental interactions are learned directly from video.
• Zero traditional game design assets (meshes, textures, or code) are required.
• Highly responsive keyboard and mouse controls for a generative system.

Cons

• Extreme memory drift causes the world to constantly morph when looking away.
• Low resolution and visual compression artifacts reduce clarity.
• Very expensive cloud infrastructure makes local or offline play impossible for now.

Conclusion: A Glimpse into the Future

Oasis AI is a technical triumph. While it is too unstable and low-resolution to replace traditional video games today, it demonstrates that deep learning models can simulate complex, interactive 3D worlds on the fly.

As hardware accelerators improve and diffusion models become more efficient, we may see a future where game environments are generated in real-time, responding dynamically to player inputs in ways that pre-programmed code could never anticipate.

Frequently Asked Questions

Is Oasis running a modified version of Minecraft?

No, Oasis does not run the Minecraft game engine. It is a neural network that has watched millions of hours of Minecraft gameplay and has learned to generate similar visuals and behaviors in response to controller inputs.

Can I build permanent structures in Oasis?

No, because the model lacks a permanent database of the world's coordinates. Any structure you build will begin to morph or disappear once you look away or travel too far from the area.

How is the input latency managed?

Oasis uses highly optimized inference pipelines running on custom hardware accelerators. This reduces the time it takes for the model to process your mouse movement and generate the next corresponding frame to under 100 milliseconds.