Ground Your Wukong — Model + Knowledge

# Session 3 · Ground Your Wukong — Model + Knowledge

**Trait: Freedom.** Today: how a chatbot system really works — once you see the machinery, no vendor can mystify you again.

## Part 1 · What happens when you hit Send

Three players:

- **YOU** — typing in the chat box at wukongdojo.ai.
- **THE AGENT** — the dojo behind the screen; packs everything Wukong needs to see, every single turn.
- **WUKONG (the LLM)** — a brain in a jar: brilliant, but no hands, no memory. Reads once, answers once, forgets everything.

Before Wukong sees anything, the Agent gathers three things: ① the persona card (system prompt — who your agent is) ② the whole chat so far (history) ③ your new message. Three things printed into **ONE stack** — that stack is the **context**.

The flow: you type → agent gathers the 3 things → prints one stack → hands it to Wukong (**this hand-off is "the LLM call"**) → Wukong writes the answer word by word → the answer lands on your screen and is filed into history.

Multi-turn chat = the agent **reprinting the whole stack every turn**. The paper tray has a limit — the **context window** — and every sheet costs money.

**Three truths:**
1. "Memory" is a trick — it lives in the agent's stack, not in Wukong.
2. Every sheet costs money — re-sent, re-read, re-billed every turn.
3. Whoever writes the stack controls the answer — hallucination is a weak script; you are the editor.

## Part 2 · Token economics

- A token ≠ a word. Rough math: 1 Chinese character ≈ 1.5 tokens · 1 English word ≈ 1.3 tokens.
- Input is cheap, output is several times more expensive — you pay for what Wukong writes.
- Your $20/month subscription is a package on top of raw token prices. Warm-up homework: run the numbers and see the gap.

## Part 3 · The model menu

**Closed models — three sizes per house:**

| House | Fast/cheap | Day job | Hard stuff |
|-------|-----------|---------|------------|
| Claude | Haiku $1/$5 | Sonnet $3/$15 | Opus $5/$25 |
| GPT | GPT-5 mini $0.25/$2 | GPT-5 $1.25/$10 | o3 $2/$8 |
| Gemini | 2.5 Flash-Lite $0.10/$0.40 | 2.5 Flash $0.30/$2.50 | 2.5 Pro $1.25/$10 |

*(prices per million tokens, input/output)*

**Open-source wave:** DeepSeek is #1 on OpenRouter by token share; 8 of the top-10 open models are Chinese (Qwen · Kimi · GLM); Llama/Gemma are free to run. Up to 20× cheaper · weights in your hands · your data stays home.

**The punchline: WukongDojo lets you pick any good LLM — your Wukong stays yours.**

## Part 4 · You + Wukong = Your Wukong

How does Your Wukong beat ChatGPT? Three axes:

- **① Narrow** — ChatGPT serves anyone; you serve just one kind of person, and you win.
- **② Deep** — ChatGPT knows the surface; you know the layer it can't reach.
- **③ Real** — ChatGPT is secondhand from the web; you are firsthand, the real thing.

Proprietary ≠ clever. Proprietary = **what others don't have, you do.**

## Lab · Ground Your Wukong

1. Ask your agent a hard domain question — watch it guess. Fluent, but not grounded.
2. Attach a knowledge .md file — your notes, your FAQ, your domain facts.
3. Ask the same question again — now the answer comes from YOUR file.

**Same model. Same question. Your file makes the difference.**

## Homework

- **HW1 (warm-up, not submitted):** run the numbers — monthly fee vs actual token burn.
- **HW2 (submit · `ENS3-FREEDOM` · due Wed 8/5 9 PM PT):** Ground Your Wukong — ① your edge (narrow·deep·real) ② your knowledge .md (pasted) ③ before/after proof on the same question ④ model pick + one S4 blueprint line. See the S3 Homework Guide in materials for full details and examples.

Next: **S4 · Aug 15 — walk into the dojo and build your agent.** Your S2 hero + your S3 knowledge file are the raw material.