Oasis has no game engine. It generates its environment, physics, and gameplay elements as you play.
The first playable AI model that generates “open-world” games, called Oasis, is here. It was created by an Israeli AI company, Decart, in partnership with a hardware company called Etched.
Decart recently launched Oasis after securing $21 million in funding, backed by big investors like Sequoia Capital and Oren Zeev.
Unlike traditional video games, which rely on pre-designed assets and game engines, Oasis generates its environment, physics, and gameplay elements as you play. The AI generates video frame-by-frame from keyboard and mouse inputs.
“Within a decade, we believe a majority of internet content will be AI-generated.” — Etched
It’s a bold claim from its creators.
But when you think about it, 70% of today’s internet traffic is video. With this massive amount of data, AI video models could start to scale and learn to represent entire physical worlds and video games.
Oasis promises to change interactive gaming forever.
The game renders in real-time, with each frame generated in just 0.04 seconds — way faster than other AI models that can take up to 20 seconds just to make one second of video.
The core of Sohu’s technical innovation lies in its transformer-based architecture, enhanced with a Variational Autoencoder (VAE) and an axial, causal spatiotemporal attention mechanism. These features allow Oasis to efficiently handle spatial and temporal data, enabling seamless, continuous gameplay.
This setup maintains high performance on extended sequences and ensures that every user action can dynamically influence the game environment in real time.
Sohu’s infrastructure tackles the usual bottlenecks of video models by parallelizing computations for large batch sizes, allowing for smooth performance even at 4K resolutions. Its efficiency dramatically reduces latency, enabling real-time video at scale and providing a feasible, low-cost solution for global users.
Right now, the publicly accessible demo of the game can only be played in 360p resolution. Sohu can serve next-gen 100B+ parameter models in 4K video, scaling to >10x more users than H100s.
More about the technical details of Oasis here.
Head over to the Oasis website, and you’ll be greeted with this dashboard.
No need for an account, login, or credit card — just click Start and follow the instructions on the screen.
“You’re about to enter a first-of-its-kind video model, a game engine trained by millions of gameplay hours. Every step you take will shape the environment around you in real time.”
On the map selection screen, you’ll find four available maps. You can even upload a screenshot to Oasis as a starting point for your game.
Hover over the map you want, click “Begin Your Journey,” and start playing.
You can move around, jump, pick up items, break blocks, and more. The game’s physics, rules, and graphics are all generated by AI in real time. To exit, press the escape key to open the menu.
Here are some example gameplays.
Now let’s talk about the feature that lets you upload your own images and turn them into playable environments. This opens up a world of possibilities for creating unique, personalized gaming experiences.
The system processes user-provided images and seamlessly integrates them into the game’s physics engine and interaction systems, creating personalized gameplay environments that maintain all the interactive capabilities of the base system.
You can watch an example gameplay here:
You can share your custom worlds and creations through Decart’s Discord server in the contest-1 channel.
Sure, AI is slowly revolutionizing the video game industry, from how games are developed to how they are experienced and played.
But it’s still in its early stages.
The tech behind AI-generated video games is still premature, and it’s unlikely to replace traditional games anytime soon.
However, game studios shouldn’t ignore the potential here.
Now that a prototype is available to the public, more people will be paying attention, and companies might invest in research to be the first to launch a full-scale product.
In fact, China has already announced a research paper on GameGen-X, a diffusion transformer model specifically designed for both generating and interactively controlling open-world game videos.
Here’s an example:
Environmental Basics: Widen the path in front of the main character as they walk forward. Main Character: Move steadily along the path, decreasing distance to distant buildings. Environmental Changes: Enhance visibility and detail of approaching village structures over time. Sky/Lighting: Maintain clear skies and consistent daylight throughout. aesthetic score: 5.47, motion score: 15.37, camera motion: pan_right, perspective: third, shot size: full.
GameGen-X is set to be open-sourced, giving the scientific community access to game data and sparking collaboration across fields.
You can follow their GitHub page for updates.
Its imminent open-source release will offer the scientific community unprecedented access to a broad spectrum of video game data, fostering innovation and collaboration across multiple disciplines.
Of course, not everyone is not gonna like it.
The quality is not on par with the likes of Black Myth: Wukong, animations are still clunky, and frame rates sometimes drop badly. Some users even say it’s “unplayable” due to glitches, like objects popping up out of nowhere or areas being unreachable because of rendering issues.
I know, the initial version sucks.
But remember, early versions of any tech often get mixed reviews.
Midjourney’s first version in 2022 wasn’t that impressive either, but it’s improved a lot. MJ is one of the most loved and is considered to be the king of all AI image generators.
X user Nick St. Pierre shared these two examples that show how much image generators have improved in the span of two years.
Do you see the point here?
The first version of any technology doesn’t define its future.
Dr. Károly Zsolnai-Fehér, creator of the “Two Minute Papers” YouTube channel, says it best:
“Do not judge a paper by its current state; judge it by where it will be two or three papers down the line.” — Dr. Zsolnai-Fehér
This perspective highlights the iterative nature of technology or scientific discovery, where initial findings are often foundational steps leading to more refined and impactful results in future publications/iterations.
Over the past couple of years, I’ve seen the idea of AI-generated video games come up many times, but it’s always been brushed aside as too expensive and impractical.
The best models out there barely reach one frame per second and can cost nearly $1 per minute just to run. It felt like one of those “maybe someday” technologies.
But here we are with Oasis, and thanks to new hardware like Sohu, we’re finally able to experience a (barely) playable AI-generated game that is much cheaper. This is bringing us closer to a future where traditional game engines might start to feel outdated, and who knows, even game developers might find themselves rethinking their roles.
If you’ve tried the demo and felt it was disappointing, I get it. The graphics aren’t on the level of big studio games, the animations still feel awkward, and the FPS drops can be a pain.
But AI is evolving fast, and if you’ve seen how quickly AI image generators and video generators improved, you know these things can change quickly. So yeah, it’s not replacing traditional games soon, but it’s definitely moving the needle.
‍
Software engineer, writer, solopreneur