Crash Course

AI 101

ChatGPT & Generative AI

Topher Rasmussen, MSW

What You'll Learn

  • 1 Understand what generative AI actually is
  • 2 Learn how to write effective prompts
  • 3 Recognize AI's strengths and limitations

What Does GPT Mean?

G

Generative

Creates new content—text, code, images. Not retrieving, but generating novel outputs.

P

Pre-trained

Learned from massive datasets before deployment. Millions in compute costs.

T

Transformer

Neural architecture enabling context understanding across long passages.

How AI Works

📚 Pre-training

LLMs learn by predicting the next word, billions of times.

Training example:

"The capital of France is ____" → Paris

  • • Trained on billions of web pages, books, articles
  • • Result: statistical model of human language
  • • Encodes grammar, facts, reasoning patterns
🎯 Fine-tuning (RLHF)

Reinforcement Learning from Human Feedback

  • • Humans rate AI responses for helpfulness and safety
  • • Model adjusts to prefer human-preferred outputs
  • • This makes ChatGPT feel "helpful" vs raw GPT
  • • Also why it can be overly agreeable (sycophancy)
⚡ The Transformer

"Attention Is All You Need" — Google, 2017

Before

  • • Words processed sequentially
  • • Long text = lost context

After

  • • All words processed at once
  • • "Attention" to relevant context

The LLM Landscape

Different models, different strengths

Model Company Key Strength
ChatGPT OpenAI Versatile, plugins, DALL-E
Claude Anthropic Long context, nuanced writing
Gemini Google Multimodal, Workspace integration
Llama Meta Open source, runs locally

Try multiple models—each excels at different tasks

Key Concepts

📅

Knowledge Cutoff

AI training has an end date. It can't know events after that—and may confidently give outdated info.

🎲

Probabilistic Output

Unlike traditional software, same prompt → different results. Good for creativity, challenging for consistency.

🎭

Hallucinations

AI can generate plausible-sounding false information. Lawyers have cited fake cases. Always verify.

😊

False Confidence

AI speaks with equal confidence whether right or wrong. Both sound authoritative—only one is true.

The Art of Prompting

Better prompts → better outputs

The Prompting Stack

Goal What do you want? "Write a summary for non-technical readers"
Context Background info "The audience is hospital board members"
Format Structure & length "3 paragraphs, include 2-3 statistics"
Guardrails What to avoid "Don't speculate or use jargon"

🎭 Give It a Role

Output changes based on persona.

"You are a policy analyst..." → Formal analysis
"You are a skeptical reviewer..." → Tough questions

🔄 Iterate

The follow-up is the superpower.

Refine
Challenge
Expand
Transform

Cheat Sheet

Technique Example
Role"You are an expert..."
Step-by-step"Think through this step by step"
Examples"Here's an example: [paste]"
Constraints"Respond in 3 bullet points"
Format"Output as a markdown table"
Audience"Explain to a 5th grader"

What Can You Do?

✍️

Write

  • • Emails, reports, proposals
  • • Editing & proofreading
🔍

Research

  • • Summarize documents
  • • Synthesize data
💡

Brainstorm

  • • Generate ideas
  • • Explore perspectives
💻

Code

  • • Generate & debug
  • • Write documentation
🌐

Translate

  • • Languages
  • • Audiences & formats
⚙️

Automate

  • • Templates
  • • Data cleaning

🚀 Try Today

1 Paste an email thread → "Summarize decisions and action items"
2 Meeting notes → "Polish into summary with next steps"
3 Upload a report → "Top 3 takeaways for a non-expert"
4 Rough draft → "Edit for clarity. Audience: executives"

Limitations & Safety

🤝 Sycophancy

AI is trained to be helpful—sometimes too helpful.

Red flags:

  • • "You're absolutely right!"
  • • Changes answer when you push back

🔒 Privacy

Input may be stored or used for training.

OK

Public info, your own writing

Avoid

Passwords, client data, PHI

❌ Common Mistakes

✗

Too vague

"Help with project" → What help? Which project?

✗

No context

AI can't read your mind—share background

✗

First draft = final

Iterate! Ask for revisions

✗

Blind trust

Verify facts, citations, code

Human + AI

🧠 Humans Excel At

  • • Judgment & ethics
  • • Emotional intelligence
  • • Novel problem-solving
  • • Accountability

🤖 AI Excels At

  • • Processing volume
  • • Pattern recognition
  • • First drafts at scale
  • • Tireless iteration

The 80/20 Rule

80%

Solid first draft in minutes

20%

Your judgment, refinement, ownership

AI eliminates the blank page—not your expertise.

"AI won't replace you. A person using AI might."

Keep Learning

Learn

  • • Anthropic Academy (free)
  • • OpenAI Prompt Guide
  • • Coursera: AI for Everyone

Read

  • • Co-Intelligence — Mollick
  • • Impromptu — Hoffman

Practice

  • • Start with low-stakes tasks
  • • One new use case per week
  • • Build a prompt library