A colleague recently brought my attention to Google’s GameNGen experiment and their paper. In the experiment they demonstrate generative AI creating a playable version of Doom without any code. Google’s experiment is able to render an interactive version of Doom, in real-time, using a neural network that understands how the game behaves, having been shown how the original game works. Apparently the playable section of the game only lasts a few seconds before it starts to fail but even in those few seconds it is mind-bending to see the machine create monsters, adjust the score, deplete ammunition, and produce the maze. A neural network has simply watched the game being played and is then able to re-create the game screens in response to user actions.

There are some flaws in the rendering but it’s really clever stuff … and thought-provoking.

Is this the future? One where there is no “code”, just extremely intelligent entities that will show us what we expect to see based on an understanding of the world what we are doing in it? I think it is. Let’s face it – code (Assembler, Fortran, C, Java, whatever) is how humans instruct computers how to behave but GameNGen shows us that powerful machines won’t need code to do the same – they will understand how the world works having observed enough of it so that they can then infer behaviour from that knowledge.