AI for Good — Interactive Mindmaps

Key concepts: Introduction

1. Introduction

The AI Awakening

• Leopold's forgettable YouTube presentation sparks curiosity
• Global debate on AI pause signed by Musk and Wozniak
• AI industry is a hall of mirrors with hype and jargon
• Skeptics profit from speaking fees and consulting gigs

Defining AI is Arbitrary

• AI is a 70-year-old field with no objective threshold
• Debaters are often poor communicators, neurodivergent, or non-native speakers
• Trust drained by same companies behind social media
• Matthew McConaughey's Salesforce ads epitomize confusion

Danny Hillis's Reframing

• Hillis is a computer science legend from early Arpanet days
• Advice: imagine the tech without the tech companies
• His career predates tech megalopolises
• AI experimentation exists beyond chatbot-calendar integration

The AI Counterculture

• Quiet, practical people tinkering with AI to fix real problems
• Many had no prior software expertise
• Stubbornness drives them through AI's weirdness
• AI will get easier and less strange over time

The Choice: Fix vs. Destroy

• Humans resist change due to sentimentality and comfort
• Techno-optimism says 'move fast and break things'
• Counterculture chooses to fix broken systems, not smash them
• They use AI to repair what they love despite limited understanding

The Inevitability of AI and the Choice We Face

• AI misuse and malfunction are real risks
• Defensive avoidance is ineffective; collaboration is key
• Enlightened self-interest means shaping AI for good
• Choice is between fixing or breaking systems

The AI Counterculture and Practical Application

• Non-expert problem-solvers use AI for mission-driven goals
• Focus on fixing education, health care, government, and connection
• Avoidance or destruction is not the genuine response
• Thoughtful collaboration preserves what matters

Key concepts: Chapter 1: Ed’s Dead

2. Chapter 1: Ed’s Dead

EdTech's Failure Pattern

• Ed, LAUSD's AI chatbot, died spectacularly in 2024
• AllHere Education imploded under fraud and bankruptcy
• Ed joins a graveyard of failed EdTech experiments
• Khan Academy stands as the gold standard of integrity

GPT-4's Transformative Revelation

• Khan initially passed on GPT-3 in early 2021
• GPT-4 demo in summer 2022 gave Khan goose bumps
• Model answered AP biology and debunked learning styles myth
• GPT-4 was a turbocharger for content but complex for students

The Internal Debate and Risk Assessment

• Khan convened a Socratic debate at the next hackathon
• Team polarized between scaling potential and dangerous flaws
• Khan understood the wager on Khan Academy's reputation
• Majority aligned: AI was coming, they could navigate it best

Turning Risks into Product Requirements

• Every risk was written down and turned into a feature
• Transparency, guardrails, and moderation were prioritized
• Fixing math and anchoring model on Khan Academy content
• The decision was an informed bet, not an impulsive leap

Key concepts: Chapter 2: How to Train Your Tutor (While Slowly Losing Your Mind)

3. Chapter 2: How to Train Your Tutor (While Slowly Losing Your Mind)

ChatGPT as a Prediction Engine

• Trained on hundreds of billions of words
• Uses transformer architecture for word relationships
• Optimizes for fluency, not truth
• Scale: 10²⁵ floating point operations for GPT-4

Khan Academy's Informal Partnership with OpenAI

• No roadmap or formal rituals, like a Vegas elopement
• Only two engineers with AI experience
• Sent proprietary content, but model was past training
• Real work was prompting, not engineering

The Art and Frustration of Prompting

• Prompting is a new coding language called English
• First prompt: guide student without giving answer
• Probabilistic responses: good result only 6/10 runs
• Endless cycles of tweak, test, and reword

Instability of GPT-4 During Development

• OpenAI updates made earlier progress obsolete
• Hallucinations surfaced unpredictably
• Stability was a promise, not a guarantee
• Building on shifting ground with no reliable prompts

Teaching a Tutor Bot Patience and Encouragement

• Mimic Sal's blend of probing questions and encouragement
• Knowing when to hold back was messy
• Required endless trial, error, and rewording
• Avoid answer-dispensing behavior

Key concepts: Chapter 3: You’ll Be Disappointed for a Long Time Until You’re Not

4. Chapter 3: You’ll Be Disappointed for a Long Time Until You’re Not

The Emotional Arc of Building Khanmigo

• Messy, uncertain, and often frustrating process
• Grinding through failures and misaligned expectations
• Months of feeling broken until suddenly it works
• Core tension: scaling Sal's teaching vs. model limits

Jessica Shieh: The Human Bridge

• Solution strategist connecting OpenAI to Khan Academy
• Self-described 'very nice asshole' with evangelical energy
• Personal mission: grew up watching Sal Khan's videos
• Pushed Khan Academy to aim higher than GPT-3.5

Context Stuffing: The Art of Reminding the Model

• Packing prompts with student info and Socratic instructions
• No persistent memory—every interaction starts fresh
• Expensive: nine cents per interaction, millions of sessions
• Balancing enough context to work without bankrupting mission

The GPT-3.5 Shock and Cheating Panic

• OpenAI released ChatGPT without warning Khan Academy
• Students immediately cheated; schools banned the tool
• Sal Khan stuck under NDA, unable to defend technology
• Public backlash surfaced failure modes for guardrails

Red Teaming: Finding Vulnerabilities on Purpose

• Simulating attacks to discover model weaknesses
• Uncomfortable but trust-building exercise for leaders
• Found appalling casualness on sensitive topics
• Built stricter standards than OpenAI's own

The Prototype: Craftsmanship Over Magic

• Five-day sprint with Shieh at center of engineers
• Three days of nothing clicking, then day four breakthrough
• Gradual refinement: show, tell, build, then ride like bicycle
• No single breakthrough—just persistent iteration

The Math Problem and Last-Minute Launch

• GPT-4 terrible at arithmetic despite conversational prowess
• Used Python for calculations, not AI prediction
• Frantic weekend texts before March 14 launch
• No one sure it would work until delivery