Generative AI Fundamentals
& How Agents Work
Understanding the technology before using it
What We'll Cover Today
1
What is Generative AI?
GPT, tokens, training
2
Fun Exercises
Creative generation
3
From Autocomplete to Agents
ReAct framework
4
Lab
Hands-on experimentation
Part 1
What is Generative AI?
Understanding AI Training
Two short videos explaining how AI learns:
A Brief History of Generative AI
| Year | Milestone |
|---|---|
| 1957 | Perceptron - First trainable neural network |
| 1961 | ELIZA - First chatbot (early generative AI) |
| 1979 | Neocognitron - First deep learning neural network |
| 1989 | Backpropagation - Deep learning becomes practical |
| 1997 | LSTM - Long short-term memory for speech recognition |
Source: dataversity.net/articles/a-brief-history-of-generative-ai
The Modern Era of Generative AI
| Year | Breakthrough |
|---|---|
| 2014 | GANs - Generate realistic images, video, audio |
| 2017 | Transformers - "Attention Is All You Need" paper |
| 2022 | ChatGPT - LLMs go mainstream |
| 2023+ | Agentic AI - Systems that plan and take actions |
Most of what we call "AI" today happened in the last 10 years
GPT = ?
G
Generative
Creates new content
P
Pre-trained
Learned from massive data before you use it
T
Transformer
The architecture that made this possible (2017)
The Core Insight
"World's Best Autocomplete"
The Core Insight
"World's Best Autocomplete"
The Core Insight
How Generative AI Works (Interactive)
How Training Works
1
Feed billions of text examples
Books, websites, code, conversations
2
Learn to predict next token
Given previous tokens
3
Scale up
More data + more parameters = emergent capabilities
4
Fine-tune for conversation
RLHF (Reinforcement Learning from Human Feedback)
Key Concept: Tokens
Tokens ≠ Words
Tokens are subword pieces (~4 characters average)
| Text | Tokens |
|---|---|
"hello" |
1 token |
"uncomfortable" |
["un", "comfort", "able"] = 3 tokens |
| Code | Often more tokens per line than English |
Why it matters: You pay per token, limits are in tokens
Key Concept: Context Window
The model's "working memory"
System prompt
Conversation history
Your current message
Documents/code
Claude
~200K tokens
GPT-4
~128K tokens
Gemini 1.5
~1M tokens
Key Concept: Temperature
0
Deterministic
0.5
Balanced
1.0
Creative
Lower for code/facts
Higher for creative writing
Key Concept: Hallucinations
Why AI makes things up
Key Concept: Hallucinations
- Model generates plausible next tokens
- Plausible ≠ True
- Confident prediction ≠ Factual information
Always verify important outputs
The Jagged Frontier of AI
The Jagged Frontier - Key Insights
Superhuman at unexpected tasks
Medical diagnosis, complex math, sophisticated code
Struggles with "simple" tasks
Visual puzzles, counting, physical reasoning
Jaggedness doesn't match intuition
Passes bar exam, fails at basic visual tasks
Creates collaboration opportunities
Humans fill AI gaps, AI amplifies human strengths
Source: Ethan Mollick, "The Shape of AI"
The Equation of Agentic Work
The Equation - Key Factors
1
Human Baseline Time
How long would this take YOU to do?
2
Probability of Success
How likely is AI to succeed? (Remember the jagged frontier)
3
AI Process Time
Agents run in background while you work on other things
Management skills become your superpower with AI agents
Source: Ethan Mollick, "Management as AI Superpower"
Management as AI Superpower
Part 2
Fun Generative Exercises
These aren't just games — they reveal how the model works
Exercise: The Dinosaur Rewrite 🦖
Exercise: The Dinosaur Rewrite 🦖
Take this news article: [paste any recent news]
Rewrite it so that a dinosaur is somehow
centrally involved in the incident.
Keep the same journalistic tone and structure.
Exercise: The Tone Dial 🎚️
Original email:
"The project deadline was missed again. This is unacceptable. We need to discuss this."
1. Furious
→
2. Frustrated
→
3. Neutral
→
4. Understanding
→
5. Gracious
Exercise: Format Juggling 🔄
Input:
"John Smith is a 34-year-old software engineer from Seattle. He earns $150,000 at TechCorp..."
JSON
YAML
Bullet points
SQL INSERT
Haiku 🎋
Movie trailer 🎬
Same information, endless formats!
Exercise: The Accordion 🪗
Start with: "The server crashed."
Incident report (1 paragraph)
Post-mortem (3 paragraphs)
↘️ Compress
Tweet (280 chars)
↘️ Compress
Single emoji
Part 3
From Autocomplete to Agents
The conceptual leap that changes everything
The Limitation
LLMs can only produce text
The Limitation
✗
Can't browse the web
✗
Can't run code
✗
Can't read files
✗
Can't call APIs
"All talk, no action"
The Solution: Tools
Give the LLM ability to request actions
User
"What's the weather in Seattle?"
↓
LLM thinks
"I need weather data..."
↓
LLM outputs
{"tool": "get_weather", "location": "Seattle"}
↓
System executes
{"temp": 52, "condition": "rainy"}
↓
LLM responds
"It's 52°F and rainy in Seattle."
The ReAct Framework
Reasoning + Acting
The ReAct Framework
Reasoning + Acting
THINK
Reason about task
→
ACT
Tool call
→
OBSERVE
See result
REPEAT
Why ReAct Works
1
Explicit reasoning
Prevents rushing to wrong actions
2
Observation step
Allows course correction
3
Loop continues
Until task is complete
4
More reliable
Than single-shot generation
Activity: Be the LLM
Activity: Be the LLM
1 student = The LLM
Reads prompt, generates response
1 student = Tool Executor
Runs tools, returns results
LLM outputs either:
TOOL: [name], INPUT: [value]
or
ANSWER: [response]
Prompt → LLM (Round 1)
SYSTEM:
You are a helpful assistant. Follow this loop:
1. THINK out loud about what you need to do
2. ACT by calling a tool if needed
3. OBSERVE the result
4. REPEAT until you can answer
Tools: calculator(expression), web_search(query)
USER:
What is the square root of 65536?LLM Response (Round 1)
[LLM reasoning]
TOOL: calculator
INPUT: sqrt(65536)Tool Executor: run the calculation
Tool Result
256
Prompt → LLM (Round 2)
SYSTEM:
You are a helpful assistant. Follow this loop:
1. THINK out loud about what you need to do
2. ACT by calling a tool if needed
3. OBSERVE the result
4. REPEAT until you can answer
Tools: calculator(expression), web_search(query)
USER:
What is the square root of 65536?
ASSISTANT:
[LLM's reasoning from Round 1]
TOOL: calculator
INPUT: sqrt(65536)
TOOL RESULT:
256LLM Response (Round 2)
[LLM reasoning]
ANSWER: The square root of 65536 is 256.What Makes an "Agent"
LLM
"the brain"
+
Tools
"the hands"
+
Reasoning Loop
"the process"
=
Autonomous Agent
Common Agent Tools
| Tool Type | Examples |
|---|---|
| File system | Read, write, search files |
| Web | Fetch pages, search |
| Code execution | Run scripts, tests |
| APIs | External services |
| Browser | Playwright for web interaction |
Why Agents Matter for Development
| Before | After |
|---|---|
| AI suggests code | AI READS your code |
| You run tests | AI RUNS your tests |
| You research libraries | AI RESEARCHES for you |
| You fix issues | AI FIXES and verifies |
Transforms AI from "assistant" to "autonomous developer"
Why Agents Matter for Development
Part 4
Lab Time
Put these concepts into practice
Lab Exercises
1
2
Creative Generation
10 min
Rewrite your project description in 3 styles
Or: Technical concept as children's story / rap / news
3
Agent Thinking
5-10 min
Write out ReAct steps for researching a library
Key Takeaways
1
LLMs = sophisticated autocomplete
Predicting tokens, not "thinking"
2
Tokens ≠ words
Understanding tokens helps efficiency
3
Agents = LLM + tools + loop
What makes AI useful for dev
4
ReAct: Think → Act → Observe → Repeat
The foundational agent pattern
5
Always verify
Hallucinations happen
Before Next Time
☐
Read: "Management as AI Superpower" - oneusefulthing.org
☐
Think: What would you want an AI agent to research for your project?