Kindle-ready field guide
Modern AI, but make it make sense
A concise guide and lightweight decision game for understanding the systems behind modern AI products: transformers, training, RAG, MoE, diffusion, agents, long-context reasoning, multimodal systems, and production evals.
Start the Quest Read the Lore Open Glossary Single-Page Edition
Modern AI, decoded
AI has moved at a breakneck speed since the introduction of the Transformer architecture. This site explains the engineering ideas that now show up in real products: how models read context, how they are trained and aligned, how apps ground answers in private data, how agents use tools, and how teams test systems before launch.
Read the chapters in order, or run AI Systems Quest, a compact scenario game where each choice tests whether you know when to use retrieval, attention, sparse experts, tool calls, extra test-time compute, or production evaluation. The goal is practical judgment: choosing the right pattern for the job, not memorizing model release names.
Reading progress
Track your chapters
Mark chapters complete as you read. Progress stays on this device, no account needed.
The Transformer Core
Get self-attention, Query-Key-Value vectors, context limits, and why Transformer variants still shape modern language, vision, and audio systems.
LLM Training & Alignment
See how pre-training, supervised fine-tuning, RLHF, DPO, safety tuning, and deployment feedback turn a base model into an assistant.
RAG & Context Windows
Lock in on when to use semantic search, reranking, citations, or long context for private documents, fresh data, and multi-document synthesis.
Scaling Efficiency: MoE & Quantization
How sparse routing, quantization, distillation, and caching make larger models cheaper to serve and smaller models usable on local hardware.
Diffusion & Generative Media
Lock in on how diffusion, latent generation, guidance, and editing workflows power images, video, design tools, and synthetic training data.
Agentic AI & Reasoning
Step into tool calling, agent loops, computer use, memory, and approval gates for systems that need to act instead of only answer.
Future Frontiers & Physical AI
Analyze multimodal models, verified synthetic data, robotics, physical grounding, and when extra inference-time work is worth the cost.
Evaluation, Safety & Production AI
Lock in on how teams ship AI safely: eval suites, groundedness checks, prompt-injection defenses, observability, rollback plans, and human review loops.