AI-101 — Systems by Layer

[OUT OF SCOPE] Math, neural networks, or AI training, what product is trendy, AI hype app

[IN SCOPE] How systems is formed: who controls the servers, how you get charged, how it run.

[CORE OUTCOME] Look at any AI tool and immediately know what piece of the puzzle it actually is.

⚠️ SYSTEMIC OVERVIEW — READ FIRST

This is a systemic mapping of the AI ecosystem, structured like an OSI/RFC layer model.

Every serious engineer will eventually go neck-deep into some AI tool or vendor ecosystem. This exists to help with getting to know what kind of mess you are going to walk into — layer by layer:

what you own
what you rent from vendor
what you borrow from Open Source toy-grade project
what disappears the moment vendors want to screw you over for their profit

Without understanding the block, you are not making technical decisions. You are an engineer trusting marketing and Facebook KOLs.

WE WILL NOT BE TALKING ABOUT THIS, ASK YOUR CHATBOT (OR PERPLEXITY) :

How training models actually work, weights, decoders, or transformer math.
Product hype, Google lastest toys vs ChatGPT new toys, making slideshows, or which models are trending bullcrap.
How to install newest fancy agent, use OpenClaw, use Claude Code, or mess around with Antigravity.

User

Part 5 — Surface

Chat

IDE

CLI

Web

Mobile

API

"Where you talk. Not where it thinks."

5 modules Tier 1

Part 2 — Runtime Layer

The Control Plane

System Rules

Task Runner

Session Memory

Tool Caller

Approvals

Multi-Agent

"The model is the data plane. The runtime is the control plane."

The model only sees what the runtime feeds it
Executes memory, loops, and external tools
Same model + different runtime = different behavior

6 modules Tier 1

6. Control

"Is this allowed?"

Auth · TOS · Official path vs opportunistic · Sub ≠ API

5 mod

7. Economics

"Can you pay?"

API vs Sub vs Local · Cost routing · Flat-rate ≠ portable

6 mod

4. Context / RAG

"Not memory. Fuel."

Retrieved, ranked, budgeted, injected — never stored
RAG ≠ memory (stale source = stale output)
Wrong chunk = confidently wrong answer

8 mod · Tier 1

1. Inference

Stateless Processor

"Receives. Processes. Returns. Forgets."

Stateless processor — no memory between calls
1M context ≠ 1M reliable recall (degrades at high fill)
Runtime manages state, not the model

6 mod · Tier 1

3. Ext Services

Search Service Image Processing Web Scraper Code Sandbox File Processor

"Consumed, not owned."

Consumed by runtime — not owned
Swappable via MCP, HTTP, or SDK
OCR (capability) ≠ Router ≠ Protocol