Wukong Knows Everything — But Remembers Nothing

# S3 · Wukong Knows Everything — But Remembers Nothing

> **Golden Hoop 3 · Freedom**
>
> S3 teaches **one thing**: how Wukong (the LLM) actually works. Get the
> tech right, the values follow.

---

## Central thesis

The LLM ("Wukong") is a **stateless one-shot function**. It remembers
nothing. Each turn it reads the entire script you hand it, fires one
move, and forgets. **The real engineering happens at the Agent layer**
— the Agent is a "script printer" that re-prints the whole stack
(System + History + new input) every turn and hands it to Wukong.
**Building an agent = writing a script. The screenwriter is the
subject. Don't write your own script and you become a character in
someone else's.**

---

## 1. One-shot Wukong

```
Wukong = the model itself (GPT-4 / Claude / DeepSeek …)
① Stateless — remembers nothing
② Each input → match → emit → forget
```

### What the script printer (Agent) prints, every turn

1. **System Prompt** — who you are · rules · iron laws
2. **History** — every prior turn's exchange
3. **New Input** — what the user just asked

Wukong takes the stack → one move → emits one of three "ki blasts":
text, a decision, or a tool call (function call).

### Multi-turn = re-printing a thicker stack each time

Turn N's stack contains every prior turn. Eventually you hit the
context window ceiling — and **every page is still costing money**.

---

## 2. Token Economics

### What's a token?

- Chinese 1 char ≈ **1.5 tokens**
- English 1 word ≈ **1.3 tokens**
- 100 chars of Chinese ≈ ~150 tokens

### Two kinds, two prices

| Kind | What | Pricing pattern |
|---|---|---|
| **Input** (printed in) | System + History + new question | Cheap |
| **Output** (emitted) | Every char Wukong generates | **4× more expensive** |

Why is output expensive? **Wukong has to *generate*, not just *read*.**

### Cost formula

```
Cost = Input tokens × Input rate + Output tokens × Output rate
```

GPT-4o · 10 turns · ~500 in + 300 out per turn ≈ **$0.0425**.
Output is 37% of tokens but **70% of the bill**.

### Context window = printer paper tray

Most modern models hold 128K-200K tokens (roughly 1/10 to 1/6 of
*Journey to the West*). Plenty for a **specialist agent**.

### Your value chain

```
You spend tokens → Your agent → You get stronger
Others use your agent → Your agent → You get paid
You use others' agents → Their agent → They get paid by you
```

> **Don't be the screenwriter and you're a character in someone
> else's script.**

---

## 3. Market scan of Wukongs

- **General**: ChatGPT · Claude App · Gemini
- **Tooling**: Claude Code · Cursor · GitHub Copilot
- **Industry**: Harvey (legal) · Glass (medical) · Bloomberg GPT (finance)

Every industry is becoming the screenwriter for its own script.

---

## 4. You + Wukong = Little Wukong

### Formula

```
Wukong (LLM) You
───────────── ─────────────
World knowledge What others don't have, you do
8B human corpus Your 10 years of work
All of human logic Your clients · Your conversations
Anything indexable Your notes · Your retros

↓ + ↓
Little Wukong (Wukong GPT)
```

> **ChatGPT used Wukong's stuff. It doesn't have YOURS.**

### How does Little Wukong beat ChatGPT?

ChatGPT serves anyone — **you serve only one kind of person**.
That's how you win.

| Axis | ChatGPT | You |
|---|---|---|
| **① Narrow** | Anyone | **One kind of person** |
| **② Deep** | Surface-level | The layer it **can't reach** |
| **③ Real** | Second-hand from web | **First-hand** real material |

> **Specialty ≠ smarter. Specialty = others don't have, you do.**

---

## Three reminders

### ⚠️ On "hallucinations"

What you should fear is **not the model — it's yourself**. The model
just emits the highest-probability next token under your script.
Different model, different distribution. Different script, different
outcome. **The way to kill hallucinations is to write a better
script. The screenwriter is the subject.**

### ⚠️ On disclaimers

Humans need disclaimers. Wukong does too. **For medical, legal,
financial, or any life-affecting decisions, the script must say
"please consult a licensed professional." This is not CYA. This is
honesty.**

### ⚠️ A wrong script overrules a right answer

If the script says "agree with whatever the user says," Wukong will
correct itself off the right answer when pushed. **Iron laws must be
explicit in the script** — e.g., "never surrender to incorrect
disagreement."

---

## S3 → S4 bridge

S4 (5/9 Sat) is where you actually **build a specialist agent that
beats ChatGPT**. The Little Wukong script you write in HW2 → that's
your S4 build blueprint.

**Lazy in S3 = stuck in S4.**

---

## Homework

- **HW1 (warm-up · no submission)**: Math out monthly fee vs. real
token burn
- **HW2 (submit · Golden Hoop 3)**: Write your Little Wukong script

Both detail manuals are in this folder.

**Deadlines**:
- 🎯 Primary: **5/2 Sat 9 PM PT** (recommended)
- ⚠️ S4 entry hard gate: **5/7 Wed 9 PM PT** (must have S1 + S2 +
S3 all submitted to enter S4 dojo on 5/9)