A new AI-generated version of Minecraft from Decart and Etched offers a practical glimpse at real-time interactive video generation. The demo lets a player move, cut down a tree and place a dirt block, but it can also change the world unexpectedly because the experience is generated by AI rather than built from conventional game code.
What Decart and Etched Built
The project is a proof of concept from Decart and Etched. Instead of writing a traditional version of Minecraft line by line, the companies trained a model called Oasis using Minecraft gameplay clips and matching keyboard inputs.
The system produces the game view in real time through a technique known as next-frame prediction. In plain terms, the model predicts what the next moment on screen should look like based on what the player has just done.
That approach makes the demo feel familiar at first. A player can move forward, interact with the world and place objects. But the same approach also explains why the experience can break from the logic of Minecraft. If a player turns away and looks back, a dirt block may no longer be where it should be, or the scene may have become a different environment.
Microsoft, owner of Minecraft, clarified after publication of the source article that this AI-generated version of Minecraft is not officially sanctioned.
Why This Matters For Real-Time Video Generation
The larger point is not simply that AI can imitate part of a popular game. Decart and Etched are using Minecraft as a demonstration of a broader idea: interactive worlds might one day be generated on the fly rather than fully coded in advance.
Dean Leitersdorf, cofounder and CEO of Decart, described the ambition as turning the screen into a portal to an imaginary world that can be changed immediately. In that vision, the player would not only press keys inside a fixed environment. The player could ask the model to alter the world itself.
The source article gives examples such as adding a flying unicorn, changing the setting into medieval ages or turning it into Star Wars. Those examples point to a major shift in how interactive media could work if the technology improves. Instead of building every asset and behavior ahead of time, creators could use AI to generate scenes and variations while the user is already inside the experience.
For now, the Minecraft demo is a limited version of that idea. Its value is that it shows a path: a model can learn environments, controls and basic game physics from gameplay data and user actions alone.
The Demo Still Has Clear Limits
Decart and Etched acknowledge that the current experience is rough. The resolution is low, sessions last only minutes at a time and hallucinations are common enough to be part of the story.
Those problems matter because interactive video generation has a higher bar than ordinary video generation. A generated clip can be impressive even if it only needs to hold together for a short sequence. A game-like system has to respond to player input, preserve the state of the world and remain coherent while the user changes direction or takes unexpected actions.
The Minecraft demo shows both sides of that challenge. It can react to keyboard input in real time, which is the promising part. But it can also forget or transform what just happened, which makes the experience less stable than the real game.
- Low resolution: the current visual output is not yet high fidelity.
- Short sessions: the demo can only be played for minutes at a time.
- Hallucinations: the generated world may change in ways that do not match Minecraft.
- Unofficial status: Microsoft clarified that the demo is not officially sanctioned.
Hardware Is The Next Big Test
A major constraint is hardware. Decart and Etched used Nvidia cards for the current demo, but Etched is developing a new card called Sohu. The company claims Sohu will improve performance by a factor of 10.
If that claim holds, the companies say it would reduce the cost and energy required for real-time interactive video. They also say it could support a better version of the demo, with longer play, fewer hallucinations, higher resolution and more simultaneous players.
Siddharth Garg, a professor of electrical and computer engineering at NYU Tandon who is not associated with Etched or Decart, said custom AI chips could deliver meaningful performance and energy efficiency gains. But he also expressed skepticism about a factor of 10 improvement coming from smarter or more specialized design alone.
Etched says its advantage comes from designing cards specifically for AI development. According to the company, the chip uses a single core to handle complicated mathematical operations more efficiently, and it focuses on inference rather than training.
Robert Wachen, cofounder and COO of Etched, said the company is building something more specialized than chips currently on the market. The company plans to run projects on the new card next year. Until the chip is deployed or its capabilities are verified, the performance claims remain unsubstantiated.
What Comes After The Proof Of Concept
The companies see the Minecraft demo as an early step toward more than games. If the efficiency gains are close to what Etched claims, they believe the same approach could help generate real-time virtual doctors or tutors.
That ambition depends on two things happening at once. The model must become more coherent, so generated environments and interactions remain stable for longer. The hardware must also become powerful and efficient enough to make real-time generation practical at higher quality.
For now, the demo is best understood as a preview of a possible direction for AI-generated games and real-time video generation. It is not a finished replacement for Minecraft, and it is not an officially sanctioned Minecraft product. But it shows why companies are looking beyond static generated clips toward systems that can respond, adapt and create while a user is actively engaged.