AI Is Here — Just Do It — The Essence of AI

# Session 1: How AI Actually Works

## Opening

AI is here — not to take your job, but to unlock your **potential**.

---

## I. The Current Landscape

### Jensen Huang's AI Five-Layer Cake (Davos WEF 2026)

| Layer | Name | Description |
|-------|------|-------------|
| 5 | Applications | Drug discovery, autonomous driving, legal analysis — where economic value is created |
| 4 | Models | AI models that understand language, biology, physics, finance |
| 3 | Infrastructure | Data centers, cooling, networking — AI factories |
| 2 | Chips | Processors that convert energy into compute |
| 1 | Energy | Real-time intelligence needs real-time power |

> This course focuses on Layer 5: Applications — building products with AI to create value.

### AI Market Data

- **$255B** — 2025 global AI market (→ projected $1.5T+ by 2030)
- **150M+** — GitHub developers (Octoverse 2024)
- **1 in 5** new unicorns are AI Agents — hottest track in 2025

### AI Job Explosion

- 6M new AI jobs per year
- 143% YoY growth for AI Engineers
- $136B AI education market by 2035

---

## II. The Boundary Between Humans and AI

### 99% Noise vs 1% Signal

**Noise:** "This tool replaces X", "10x productivity", surface-level changes

**Signal:**
- How AI understands language (principles)
- What you show AI determines its performance (methods)
- Making AI complete tasks autonomously (applications)
- The Last Mile needs you — Human in the Loop

### General World vs Your World

- **General World** = LLM training data (encyclopedias, papers, code…) — getting cheaper
- **Your World** = your unique expertise (client understanding, industry experience) — your moat

### What LLMs Can vs Cannot Do

**Can:** Write/debug/deploy code, analyze documents, multilingual translation, use tools, understand images/audio/video

**Cannot:** Don't know your world (your processes, clients, boss), hallucinate confidently, take no responsibility for outcomes

---

## III. The Essence of AI

### What is an LLM?

A mathematical model trained on all text from the internet, possessing the collective knowledge of humanity.

- **Token** = the unit of AI consumption (~4 English characters per token)
- 1,000 tokens ≈ 750 words (one page), 1M tokens ≈ 5 novels

### How LLMs Are Trained: 4 Stages

1. **Pretraining** — all internet data, months, hundreds of millions of dollars → knows everything, can't communicate well
2. **Supervised Fine-tuning** — human-written good answers → communicates well, may hallucinate
3. **RLHF** — human ratings 👍👎 → understands nuance = birth of ChatGPT
4. **Distillation** — learning from other models' outputs → small models gain big-model capabilities

### Chain-of-Thought (Emergent Ability)

Giving the model a "think before answering" example dramatically improves accuracy. Nobody designed this — it **emerged** on its own.

### Three Stages of AI Agents

1. **Pure LLM Call** — Q&A, you make all decisions
2. **Human-Wired Logic** — code orchestrates the flow, LLM is one component
3. **Agentic Flow** — you set the goal, LLM decides how to achieve it

### Three Layers of an AI Product

| Layer | Trend | Description |
|-------|-------|-------------|
| LLM Model | Trending free ↓ | Trained by big companies, API cost dropping |
| AI Agent Architecture | Trending free ↓ | Servers, databases, frameworks — nearly all free |
| Your ability to direct the LLM | Your edge ↑ | Your prompts determine what AI does — not replicable |

---

## IV. Hands-On Verification

### LLM vs Agent

- **Raw LLM:** No identity, no rules, random quality
- **Agent:** LLM + System Prompt + Memory + Tools = stable, predictable quality

> ChatGPT itself is an Agent — but a generic one, not built for you.

### One Resume, Two Results

Same resume (Sarah: UC Berkeley CS, 2 years at Google):
- "Help me improve my resume" → mediocre result
- Professional prompt → top-tier result

### ★ Core Insight ★

> **An LLM is a super probability model built on human language. It knows everything and has superpowers, but what you say determines what it gives you.**

---

## V. Two Answers

### Answer 1: Change Your Mindset

- Not you vs AI → it's **you + AI** vs people who don't use AI
- AI is not your opponent → AI is your apprentice, you are the master
- Super Individual = one person + AI = an entire team

### Answer 2: Collect Six Dragon Balls

| Ball | Session | Skill |
|------|---------|-------|
| ★ | S1 | Courage |
| ★★ | S2 | Leverage |
| ★★★ | S3 | Freedom |
| ★★★★ | S4 | Signature |
| ★★★★★ | S5 | Refinement |
| ★★★★★★ | S6 | Mastery |

---

## Key Takeaways

1. **LLM and Agent are two different things** — LLM = raw power, Agent = the software wrapper
2. **Your instructions are the key** — same model, different prompts, vastly different results
3. **Your world + your iterations are your moat** — AI capability is cheap, your experience is valuable
4. **AI is in its earliest days — get on board now** — those who wait will be left behind