Summary — “Art of Uncertainty” talk (lecture highlights) 🎲📚

Speaker background

• Long career in AI, Bayesian methods and public-engagement science (TV, books).
• Wrote books including The Art of Uncertainty (covers probability, luck, risk, prediction).

Key definition

• Uncertainty = the conscious awareness of ignorance — a personal, subjective relationship between you and the outside world (epistemic, not metaphysical).

Main themes

• Subjective probabilities and judgement are useful even when imperfect.
• Numerical probabilities are built on assumptions; communication must make those assumptions explicit.
• Use multiple independent analyses/teams to reveal disagreement and avoid overconfidence.

Illustrative historical examples

• Bay of Pigs (1961): vague wording (“a fair chance”) replaced clearer probability — poor communication contributed to disaster.
• Osama bin Laden raid (2011): multiple independent intelligence teams produced differing probabilities (30–90%); decision-maker (Obama) benefited from knowing the disagreement rather than a single composite estimate.
• COVID R estimates (UK): many different models (12+) produced widely varying R estimates; publishing individual model outputs + composite was good scientific practice and revealed model uncertainty.

Practical points about probabilities

• Probabilities used by experts can be subjective (judgmental) and still valuable.
• Overconfidence is common because statistical intervals assume the model is true — but “all models are wrong,” so intervals are often too narrow.
• “Mandated science” (forced to give numbers) should sometimes refuse (or give low confidence) when evidence is insufficient.

Calibration exercise / scoring rule (forecasting training) ✅❌

• Audience quiz: choose A/B, give a confidence (0–10).
• Scoring: asymmetric quadratic loss with max 25 points for perfect certainty/correct; large penalties for high-certainty wrong answers.
• Purpose: train calibration — maximize expected score by honestly expressing subjective probability (mathematically optimal under squared-error scoring).

Luck — types & examples 🍀⚡

Philosophical decomposition of luck:

• Constitutive luck — who you are (genes, family, era). Example: speaker’s grandfather born into WWI generation.
• Circumstantial luck — being in the right/wrong place/time (e.g., plane crash survivor seating, being on a doomed flight).
• Outcome (dumb) luck — how events turn out despite the circumstances (surviving a crash; passing an exam that led to a later fortunate outcome).
• Moral: make the best of the hand you’re dealt; acknowledge the role of luck in life outcomes.

Coin / magic & waiting-time intuition

• Coin covered after flip: epistemic uncertainty vs. objective chance; hidden info changes the nature of uncertainty.
• Derren Brown coin flips: event of 10 heads in a row is rare (p = 1/1024 per run). Waiting time follows a geometric distribution (mean ≈ 1024 flips); long waits are possible — distinguishes being “lucky” vs. “unlucky” in achieving rare outcomes.

Coincidences, birthday paradox & intuition-busters 🎂

• Birthday paradox: with 23 people probability ≥ 50% that two share a birthday.
  • Standard derivation: compute probability of no match = product(365 − k)/365 for k=0..22 → ≈ 0.49 → match prob ≈ 0.51.
  • Alternative intuition: expected number of matching pairs = (n choose 2)/365 → use Poisson approximation: P(no matches) ≈ e^(−M).
• Practical variants:
  • Last-two-digits of phone numbers (1/100 chance per pair) → with 23 people expected matches ≈ 2.5 → ~94% chance at least one match.
  • Demonstrations: audience exercises showing matches in phone-number endings and birthdays.
• Key lesson: randomness is clumpy; rare events and coincidences are more likely than naive intuition suggests.

“Snap” / matching-card paradox (Treize)

• Two shuffled equal piles, flip cards in parallel: probability of some matching rank at same position ≈ 1 − 1/e ≈ 0.63 (surprisingly independent of deck size beyond small limits).
• Intuition: expected number of position-matches ≈ 1, Poisson approximation → P(at least one) ≈ 1 − e^(−1).

Large numbers & uniqueness of shuffles

• 52! (≈ 8.07 × 10^67) possible permutations of a deck — effectively unique across human history; thus a given shuffle is almost certainly a never-before-seen order.

Practical lessons / recommendations

• Insist on clarity in probabilistic language (define words like “likely”).
• Use multiple independent teams/models to expose structural uncertainty and model disagreement.
• Encourage calibrated judgments: quantify confidence and be prepared to show uncertainty and explain assumptions.
• When evidence is poor, refuse to produce precise numbers or explicitly report low confidence.
• Recognize types of luck; avoid simplistic moralizing about outcomes.

Entertaining endings / anecdotes

• Double-yolk eggs anecdote — warns against naive multiplication of small probabilities; selection bias / procurement matters (you can buy double-yolk eggs purposely).
• “Wipe Out” obstacle-course story — applied simple stats to aim for qualifying time; outcome: fun anecdote that stats don’t guarantee practical success.

Key takeaway: Uncertainty is unavoidable and subjective — but by being explicit about assumptions, using multiple independent analyses, calibrating judgments, and communicating probabilities clearly (with confidence levels), we can make better decisions and avoid dangerous overconfidence.