Modern AI, but make it make sense
A crisp, high-contrast guide plus a quick decision game for getting the AI stack: transformers, RAG, MoE, diffusion, agents, reasoning, and physical AI.
Modern AI, decoded
AI has been moving absurdly fast since the Transformer architecture dropped. This site breaks down the big engineering ideas without math-heavy clutter, updated for the 2026 shift toward reasoning models, agentic workflows, long-context systems, efficient open-weight models, and physical AI.
Read the chapters in order, or run AI Systems Quest, a compact scenario game where every choice checks whether you know when to use retrieval, attention, sparse experts, tool calls, or extra test-time compute. Both paths are Kindle and e-ink friendly.
The Transformer Core
Get self-attention, the Query-Key-Value mechanism, and how it replaced recurrent networks to become the foundation of generative AI.
LLM Training & Alignment
Lock in on pre-training, fine-tuning, and alignment mechanisms like RLHF and DPO that make AI models useful, safer, and controllable.
RAG & Context Windows
See how models pull fresh data with vector databases, semantic search, and ultra-long context windows.
Scaling Efficiency: MoE & Quantization
How Mixture of Experts (MoE) keeps models fast by activating only the experts it needs, while quantization squeezes parameters onto consumer hardware.
Diffusion & Generative Media
See how diffusion models cook up images and videos by stripping away random noise, and how latent diffusion makes the loop faster.
Agentic AI & Reasoning
Get the agent loop: reasoning, tool use (function calling), and System 2 patterns where models think before they answer.
Future Frontiers & Physical AI
Vibe-check native multimodality, synthetic data limits, and the next physical frontiers for AI.