A four-minute cutscene from a late-90s shooter that never existed. Scientists at a secret research lab walk you through how an AI video generator works: five seconds of video is 331,776,000 numbers, an autoencoder squeezes that to about 7 million, space-time patches turn it into 108,000 tokens, and a diffusion transformer trained on noised clips, xt = (1 − t)·x0 + t·ε, walks pure noise back to a video in about fifty steps. The credits name the real research.