The AI Path

The goal: stop treating LLMs as magic boxes. Learn how inference actually runs and what it costs, then how to design products around those costs, then build one yourself. Three books, in that order: mechanics → design → code.

Part 1 · The mechanics — AI Systems · First Principles

1. Foundations — what an LLM actually costs in latency, memory, and dollars, and why efficiency is the whole game.
2. The transformer, mechanically — attention, prefill vs decode, and why the KV cache exists.
3. Inference efficiency — batching, PagedAttention/vLLM, FlashAttention, speculative decoding — the levers that make it cheap.
4. Application layer + the Claude-API capstone — prompt caching, routing, RAG, and a real before/after cost dashboard.

Part 2 · The design — System Design · First Principles

5. Building With LLMs — how LLM-backed features change system design: RAG, vector DBs, the new latency/cost/non-determinism budget, streaming UX, and guardrails.
6. The data-intensive frontier — vector search (HNSW/IVF), streaming, and the lakehouse that feed AI features.

Part 3 · The build — Rust · Project-First

7. Project 4 · askr, an AI CLI — build a real AI command-line tool on the Claude API: streaming, a tiny RAG (embeddings + vector search), and production hardening. The mechanics and design, made concrete in code.

Where it leaves you

You’ll understand an LLM call from the GPU up — what it costs, how to make it cheaper, how to design around it, and how to ship one. The opposite of cargo-culting prompts.