Doom Runs Inside a Transformer Model Without Any Training

Researchers have managed to run the classic Doom game renderer inside a transformer neural network, using the Phi-3 architecture as the base model. Rather than training the network, a compiler was used to directly set the model's weights. This means the transformer executes Doom's renderer through autoregressive generation, with no machine learning training involved whatsoever. The project, published on ood.dev, demonstrates an unconventional use of transformer architecture as a programmable computational substrate.