• Microsoft has released a demo of id Software’s 1997 classic Quake II that is entirely generated using AI.
• It uses a new technology that can generate images quick enough to simulate the feel of gameplay.
• You can play the demo right now.

It was only a matter of time before Microsoft began pushing its Copilot AI models towards videogame applications. Previously, it was only known that Microsoft was experimenting with using generative AI to give NPCs real-time responses to player input, now the company via its Xbox wing has revealed it has been testing generating gameplay in its entirety.

This weekend, Microsoft announced “real world modelling of interactive environments” which at this stage takes the form of a fully-playable version of id Software classic Quake II where every second of gameplay is AI generated.

Developers are able to do this through its Muse family of AI models for videogames, which were revealed in February. With Muse and a system that Microsoft is calling “WHAMM” or “World and Human Action MaskGIT Model,” the AI is able to generate visuals much faster than other image generators.

Since gameplay is frames of visual data represented by pixels on your screen, AI is able to generate these frames too like any other image, but WHAMM allows the speed of generation to be fast enough that it will respond to inputs in real-time, or rather with a bit of delay.

The ability to speed up the image generation is thanks to technology that can correctly predict what the next image is going to be before the player makes a move.

“We can think of this as producing a rough and ready first pass for the image. We would then re-mask some of those tokens, predict them again, re-mask, and so on,” the developers behind the technology explain.

“This iterative procedure allows us to gradually refine our image prediction. However, since we have tight constraints on the time we can take to produce an image, we are very limited in how many passes we can do through a big transformer.”

You can try a demo of the AI-generated Quake II on browser right now, by heading here. Microsoft is calling it part of the “Copilot Gaming Experiences.”

“Copilot Gaming Experiences represents a practical step forward in exploring how AI can create brand new ways to interact with games while building upon existing ones,” the company writes in a blog post.

“By generating gameplay in real time, the underlying Muse shows how classic games like Quake II can be reimagined through modern AI techniques.”

While the technology behind how this works is quite extensive and interesting, we have an issue with Microsoft claiming that anything in the demo is Quake II being “reimagined.” After playing the demo we can safely say that the AI has simply been trained on likely many, many hours of either Quake II footage or gameplay.

We would say that the product that is currently on demo is the kind of Quake II you would see in your mind after binging the game for 12 hours while having a terrible flu and then going to bed after taking loads of meds.

The artifacting is horrendous, especially of enemies in the game which appear like vaguely humanoid blobs. While Microsoft itself has admitted that there are some “limitations” to the technology in its current form, we’re not quite sure where the company takes it into the future.

As far as we understand, the system only works when an AI is trained on images of gameplay, which it is able to very quickly piece together to simulate gameplay. The only thing is… We already have a Quake II.

Can Microsoft’s WHAMM-induced Muse model create completely original gameplay? Definitely not right now. But we suppose the technology is still very early and this is just a demo of what can be accomplished. But unlike AI allowing real-world interactions between NPCs and the player, generating gameplay in its entirety doesn’t fulfill the same fantasy.

There’s also more to gameplay just than just the visuals. There’s the audio too, and by far the most important aspect which is game design. Games are not just good from the initial idea, they take time and effort to eventually feel good while playing and become fun for the player. It’s an artform, one where you can go to university for three years to learn and you still won’t master it.

Generating games with AI also eliminates the entire game-development process. Using AI to speed up development is one thing, but using AI to not have to develop games at all? That sounds a little bit to dystopian for our liking.