Choose a question.
Follow it down.
Each map begins with the landscape. Open only the concepts you need, then return whenever you want the larger picture.
Inside the Transformer Block
Learn the paper diagram, then trace attention, residuals, channel mixing, depth, and loops into code.
Diagram → Tensor path → Code→NEW / MODEL ARCHITECTUREPositional Encoding Evolution
See how token order grows into spatial coordinates and shared multimodal time.
Sinusoidal → RoPE → M-RoPE → TMRoPE→AVAILABLE NOW / OPTIMIZATIONOptimizer Journey
Follow scalar memory through decoupled decay and into matrix geometry.
Adam → AdamW → Muon→AVAILABLE NOW / ADAPTATIONLow-Rank Adaptation
How constrained matrix updates become practical model adaptation.
LoRA → SingLoRA / DoRA→