Have you heard of the “can it run Doom?” meme? It’s when inventors and engineers of all stripes try to get the classic 1993 first-person demon-slaying video game running on anything and everything. Doom playing on a microwave. Doom playing on a vape rig. Doom playing on a pregnancy test. Doom playing on bacteria.
Well, researchers at Google and Tel Aviv University figured out a way to get Doom running in a new way. Their GameNGen creates a playable version of Doom entirely through generative AI.
Videos by VICE
Traditional video game engines rely on pre-coded logic and graphics. GameNGen uses neural networks to recreate Doom‘s gameplay without that. First, it trains an algorithmic RL-agent, or Reinforcement Learning agent, to play the real Doom. Then, it uses diffusion models like the ones in image generators like Stable Diffusion to generate the gameplay in real-time based on the RL-agent’s past actions.
This way, a human can boot up the GameNGen Doom, and the game will predict frames based on previous frames. The AI model can manage game states like health, enemy interactions, and ammunition count all in the neural network.
There are plenty of limitations, though. It’s prone to glitches and various inconsistencies. And it’s all a lot for the system’s memory to compute, so it can only render 20 seconds of interactive gameplay before it craps out. But it does it while simulating the game’s various elements with startling accuracy.
As far as implications for this advancement go, small game development teams could one day build entire complex games based solely on a few pieces of concept art or even just simple text descriptions. The creators theorize it can also be applied to autonomous vehicles and the further development of smart cities.
