This article explains how 3D graphics projection works, starting with the simple principle of dividing x and y coordinates by depth (z) to create perspective. It then introduces the perspective projection matrix, a more practical tool used in modern graphics that incorporates camera parameters like field of view, aspect ratio, and clipping planes to transform 3D world coordinates into 2D screen space.
Recurrent Looped Transformer (RLT) combines a causal encoder with a recurrent decoder that grows temporal depth with each token, traversing tL_D decoder blocks after t tokens while maintaining fixed per-token computation. The architecture integrates global encoder memory with layerwise sliding-window attention caches in the decoder, enabling co-design with hardware and RL algorithms through shared state transitions across pretraining, fine-tuning, and sampling.