Generative AI creates playable version of Doom game without code

Recreating classic computer games with AI Doom It’s playable without any computer code or graphics, and researchers on the project say it may be possible in the future to use similar AI models to create games from scratch, in the same way that we currently create text and images.

The model, called GameNGen, was created by Dani Valevski and colleagues at Google Research, who declined to be interviewed. New ScientistAccording to the research paper, the AI ​​can play for up to 20 seconds while retaining all the features of the original, including score, ammo levels, map layout, etc. The player can attack enemies, open doors, and interact with the environment as normal.

After this period, the model’s memory begins to run out and the illusion begins to break down.

original Doom It was released in 1993, and has since become a popular subject for computer science projects, including attempts to make it run on specialized and limited hardware like toasters, treadmills, and espresso machines.

But in all of these cases, the hardware is simply running the original game’s code. What GameNGen does is fundamentally different: a type of AI called a neural network learns by observation how to recreate a game, without ever seeing the game’s code.

The researchers first created an AI model that learned how to interact with Doom in the same way a human would, and then tasked it with playing the game over and over again, while a second AI model based on a Stable Diffusion image generator learned how the game state changes with hundreds of millions of inputs.

The second model was essentially a copy of the game, with all of the knowledge, rules, and instructions from the original code encoded into a mysterious network of artificial neurons within its own architecture. In tests, human players were slightly better than chance at distinguishing between short clips of the game and clips of the AI ​​simulation.

In their paper, GameNGen’s developers claim that this is a proof of concept that games can be created by neural networks rather than lines of code. They suggest that games can be generated from text descriptions and concept art, making them cheaper to produce than using human programmers.

Andrew Rogojski of the University of Surrey in the UK says the idea of ​​having neural networks hallucinate game environments and human interactions is an interesting advance, but it’s not a replacement for human game designers.

“I don’t think it’s the end of game studios. I think what game studios have is the imagination and the skill – the ability to actually create these worlds, understand gameplay, understand engagement, understand how to draw us into a story – that’s not just the nuts and bolts or bits and bytes,” he says. “There’s something very human about creating compelling experiences that we humans enjoy, and that’s going to come primarily from other humans at the moment, and for the foreseeable future.”