Living learning guides · preview
Learn
Grounded in the models in the ModelDNA atlas. Each module is reviewed for accuracy before it is marked publish-ready.
New to the jargon? Terms like SwiGLU are clickable in every module — or browse the full glossary.
Model Anatomy
What we all mean by one word — 'model' — from the dense baseline through the efficiency techniques in the most modern models we've tested.
- 00Orientation: what a model is, and the five efficienciesfoundation
- 01The dense baseline (and its atoms)foundation
- 02Position: RoPE & long contextfoundation
- 03Attention & the KV-cache memory wall (GQA/SWA)attention spine
- 04Latent attention (MLA)attention spine
- 05Beyond softmax: linear attention & SSMattention spine
- 06Frontier attention sparsity (DSA)attention spine
- 07Parameter sparsity: Mixture-of-Expertsaxis: sparsity
- 08Decoding acceleration (MTP)axis: decoding
- 09Precision: bits as a structural choiceaxis: precision
- 10Topology: how the pieces are stitched togethersynthesis
- 11Putting it together: the technique stacksynthesis
- 12What structure does not tell youbridge
Model Classes
New kinds of model that do not fit the chat-assistant shape — what each one is for, how it works, and how to tell a good one from a confident one.
An in-progress preview. Model Anatomy has more rungs coming; Model Classes will add the other shapes we keep finding in the atlas — multimodal encoders, diffusion language models, embodied and vision-language-action models, and routed-adapter models whose behavior changes per request.