Who builds what? (The layer map)

By the end of this page — you can draw the four layers of the AI stack, say which layer each major player owns, and re-read any "model X vs model Y" debate with the right frame.

Everyone says “AI” as if it were one thing. It’s a stack. Knowing who owns which layer fixes most of the confusion.

The four layers

Layer	What it is	Who owns it	Example
🧠 L1 — Model	The trained weights: what the model can predict	Model labs (OpenAI, Anthropic, Google, Mistral, …) — or you, running it locally	GPT, Claude, Gemini, Llama
🔌 L2 — Inference & API	Serving the model, billing tokens, rate limits	Cloud providers + the labs themselves; locally: Ollama, llama.cpp	OpenAI API, Ollama on your machine
🪢 L3 — Harness / App	The loop, tools, prompts, UX — the product	App builders (coding agents: Claude Code, Codex CLI, Cursor, Windsurf…; chat apps)	“AI that writes your code”
🧑 L4 — You	Prompts, workflows, your tools, your guardrails	You	The system prompt you tweak, the MCP servers you plug in

Two things this map explains

1. “Model X is bad at coding” is usually a wrong diagnosis. What you’re experiencing is L1 through an L3 harness. The same model can feel sharp in one coding agent and clumsy in another — different system prompts, tool descriptions and context management. When in doubt, suspect the harness before you blame the brain.

2. Switching layers is a different skill than switching products.

Swap two models inside the same harness → you’re changing the engine, the car stays the same.
Swap two harnesses with the same model → the car is a different build.
Move local (Ollama/llama.cpp) → you own L1 and L2 yourself: no API bill, no cloud, your hardware sets the ceiling (cloud vs local).

Where should *you* invest your learning? — L1 is a black box you can read about but not really touch (unless you train). L2 is plumbing. **L3 is where the craft lives** — prompts, tool design, context management, guardrails — and L4 is where your specific knowledge beats everyone's. That's exactly the order this site is organized in: fundamentals → agents → tutorials.

A note on the local layer

The map above describes the cloud stack, but the bottom two layers are portable. Run a model through Ollama and you’ve got L1+L2 on your own machine; plug an agent harness in front of it and the whole diagram still holds — only the owner of L1/L2 changes from a company to you. That’s the entire local tutorials story, compressed.