AI summary by TwinMind
Kit Yates | How to Expect the Unexpected | Talks at Google
Talks at Google
46m 32s
Book Introduction & Core Phenomena
- Author presenting second book "How to Expect the Unexpected"
- Two main phenomena that confound everyday decision-making: nonlinearity and randomness
- Packed room presentation format with interactive elements
Understanding Linearity
- Definition: Fixed change in input produces fixed change in output, creates straight line on graph
- Exchange Rate Example: £20 → NZ$40 (2:1 rate), £40 → NZ$80, £60 → NZ$120
- Direct Proportion: Special linear relationship where doubling input doubles output
- Temperature Conversion: Celsius to Fahrenheit (×1.8 + 32) - linear but not direct proportion
- 10°C = 50°F, 20°C = 68°F, 30°C = 86°F
- Doubling Celsius doesn't double Fahrenheit due to +32 offset
Linear Bias in Education & Psychology
- School Problem Example: Jane pays £5 for 10 grapefruits → £50 for 100 grapefruits
- Linear answer: 100 grapefruits
- Real world: bulk discounts, supply constraints ignored
- Pseudo-linearity Questions: Psychologists test ingrained linear thinking
- Laura runs 100m in 13 seconds → linear thinking suggests 130 seconds for 1km
- Reality: Noah Ngeny's world record 1km = 2:11, can't sustain sprint pace
- ChatGPT Test: "3 towels take 3 hours to dry, how long for 9 towels?"
- ChatGPT initially said 9 hours (linear thinking)
- After viral video (10 million views), ChatGPT updated to correct answer
- Google Bard still gives linear answer: 9 hours
Area vs Length: Square Law Examples
- Farmer Jones Problem: 1 hour to mow 100m × 100m field
- Linear thinking: 300m field takes 3 hours
- Reality: Area scales as length squared → 9 hours needed
- Pizza Value Analysis: Price scales linearly with diameter, area scales as diameter squared
- Two 8-inch pizzas (£10 each) vs one 16-inch pizza (£20)
- Linear thinking: equivalent value
- Reality: 16-inch pizza has 4× area of 8-inch pizza
- Data from New York Times journalist: 75,000 pizzas from 3,500 shops confirm linear price-diameter relationship
- Takeaway: Buy largest pizza that meets needs for best value
Positive Feedback Loops
- Definition: Quantity increases its own growth rate
- Acoustic Feedback: Mic → amp → increased volume → mic picks up again → feedback loop
- Artists like Jimi Hendrix, Grateful Dead used this musically
- Rock Slides: Rain → grit → pebble → rock → stone → boulder → massive avalanche
- Small initial disturbance cascades to tons of material
- Stock Market Bubbles:
- Rising prices → increased demand → higher prices → more demand
- Dot-com boom (late 90s/early 2000s): companies with no profits had soaring stock prices
- Can work in reverse: falling prices → selling pressure → lower prices → more selling
Climate Change Feedback Loops
- Ice Albedo Feedback: Critical climate mechanism
- White surfaces (snow/ice) reflect heat energy
- Dark surfaces (land/sea) absorb heat energy
- Warming → melting ice → more dark area exposed → more heat absorbed → more warming
- Snowball Earth Hypothesis: Same feedback in reverse
- Cooling → more snow/ice → more reflection → less absorption → more cooling
- Theory: Earth was once entirely covered in snow/ice
Pandemic & Exponential Growth
- COVID-19 Example: Positive feedback loop in disease spread
- Basic Reproduction Number (R): Original SARS-CoV-2 had R≈3
- Each infected person infects 3 others in susceptible population
- 1 → 3 → 9 → 27 → 81 exponential progression
- Exponential Growth Definition: Growth proportional to current size
- Media often misuses term to mean "fast" or "big"
- Bank account interest is exponential but not fast/big
- J-Shaped Curves: Exponential growth appears linear initially, then rapidly accelerates
- Exponential Growth Bias: People assume linear continuation instead of exponential acceleration
- Leads to under-investing for future (undervalue compound growth)
- Leads to over-borrowing (undervalue compound debt)
- Higher bias correlates with higher debt-to-income ratios
- Lower compliance with pandemic measures (masking, distancing, testing, contact tracing)
Self-Defeating Prophecies
- Negative Feedback: Processes that inhibit their own growth
- Pandemic Models: Predictions of hundreds of thousands of deaths became self-defeating
- Predictions → policy responses → prevented predicted outcomes
Randomness & Pattern Recognition
- Three Pattern Test: Audience chooses which of 3 patterns is truly random
- Pattern A: Uniformly distributed random points (correct answer)
- Pattern B: Ant nest sites (territorial but less organized)
- Pattern C: Patagonian seabird nests (highly territorial, grid-like spacing)
- Middle Bias: Tendency to choose middle options when uncertain
- SAT tests: students guess B and C over A and D
- Battleship: disproportionate guessing toward center coordinates
- Toilet stalls: middle stalls used more (measured via toilet paper refill frequency)
- Commercial Exploitation: Basic/Standard/Deluxe pricing structures
- Basic vs Standard: people choose Basic
- Basic/Standard/Deluxe: people choose Standard (middle option)
- Deluxe option exists to make Standard appear reasonable
Human Perception of Randomness
- Clustering in Random Data: True randomness contains clusters and empty spaces
- Evolutionary Basis: Hunter-gatherer survival advantage
- Hearing rustles/seeing patterns → assuming predator → running away → survival
- Better to have false positives than miss real threats
- Passed on "paranoid genes" for pattern detection
- Pareidolia: Interpreting ambiguous stimuli as familiar objects
- Jesus in toast/tortillas
- Virgin Mary appearances
- Man in the Moon
- Face on Mars (craters + shadows + sun angle)
- Apophenia: Drawing causal inferences from perceived patterns
iPod Shuffle & Randomness Expectations
- Technology Context: Early 2000s, revolutionary 10,000-song capacity
- Steven Levy Incident: Tech journalist heard two Bob Dylan songs consecutively
- Complained to Steve Jobs about "broken" randomization
- Engineer confirmed algorithm was genuinely random
- Article generated hundreds of similar complaints
- Apple's Response: Modified algorithm to prevent consecutive songs from same artist
- Made it "less random to appear more random"
- Conformed to human expectations rather than mathematical randomness
Human Inability to Generate Randomness
- Interactive Demonstration: Aronson Oracle website experiment
- Volunteer presses F and D keys "randomly"
- Computer predicts next key press based on pattern recognition
- 50% accuracy expected for true randomness
- Actual results: ~70% computer accuracy achieved
- Algorithm Method: Tracks sequences of 5 consecutive key presses
- Builds histogram of all 32 possible 5-key combinations
- Uses previous 4 keys to predict 5th key
- Simple but effective pattern recognition
- Human Random Behavior Patterns:
- Think randomness = evenly spaced
- Avoid long runs (5 consecutive same key has 95% probability in 100 presses)
- Occasionally overcompensate with very long runs
- Create predictable "mixing" patterns
National Lottery Case Studies
- Historical Context: UK National Lottery started 1994
- January 14, 1995 Draw: £17.6 million jackpot, 70 million tickets sold
- Expected winners: 4-6 people sharing jackpot
- Actual winners: 133 people won £122,000 each instead of millions
- Winning numbers: 7, 17, 23, 32, 38, 42
Lottery Play Pattern Analysis
- Play Card Layout: 49 numbers, 6 selections, arranged in 10 rows of 5 (last row has 4)
- January 1995 Pattern Analysis:
- Middle bias: avoided outer columns, concentrated in middle 2 columns
- Well-spaced vertically with gaps
- Number 7 (world's favorite number) heavily chosen
- No consecutive numbers (people avoid these despite 50% probability in draws)
- March 16, 1996 Draw: 57 winners, even stronger pattern evidence
- Numbers: 28, 38, 48 (perfectly spaced sequence)
- Extreme middle bias and regular spacing
- No consecutive numbers
Lottery Strategy Optimization
- Goal: Maximize winnings if you win (can't increase win probability)
- Expected Return: 45p per £1 spent normally
- Random Number Strategy: Use lottery's "Lucky Dip" function
- Increases expected return to 55p per £1 (still losing money)
- Unpopular Number Strategy:
- Avoid 1-31 (dates: days and months from birthdays/anniversaries)
- Use consecutive numbers (others avoid them)
- Use edge numbers (overcome middle bias)
- Include numbers >31 (less popular)
- Example Optimized Selection: Uses edges, consecutive numbers, avoids popular date ranges
Key Takeaways
- Coincidences: Can be surprisingly likely given enough opportunities
- Individual lottery win extremely unlikely
- Someone winning weekly nearly certain given participation volume
- Pattern Recognition: Humans over-interpret random clusters as meaningful
- Random Generation: Humans cannot generate true randomness
- Use external tools: dice, coins, computer random number generators
- Divest randomness decisions to genuinely random sources
- Nonlinear Phenomena: Many important real-world relationships are nonlinear
- Square laws, cube laws, exponential growth, feedback loops
- Linear thinking bias leads to poor predictions and decisions
