created: 2026-04-24
updated: 2026-04-24
tags: [source, video, youtube, bayesian, probability, education]
type: source
url: https://www.youtube.com/watch?v=HZGCoVF3YvM
author: "AI Engineer" (Grant Sanderson)
published: 2019-12-22
Bayes Theorem: The Geometry of Changing Beliefs¶
Summary¶
3Blue1Brown (Grant Sanderson) explains Bayes' theorem using geometric intuition and the famous Steve the librarian/farmer example from Kahneman and Tversky. Focuses on understanding the formula through representative samples and area diagrams rather than memorization.
Key Takeaways¶
The Steve Example (Kahneman & Tversky)¶
Description: "Steve is very shy and withdrawn, invariably helpful but with very little interest in people or the world of reality. A meek and tidy soul, he has a need for order and structure, and a passion for detail."
Most people say Steve is more likely to be a librarian. But the farmer-to-librarian ratio in the US is about 20:1. Even if a librarian is 4× as likely to fit this description, the prior overwhelms the evidence: - 10 librarians × 40% fit = 4 - 200 farmers × 10% fit = 20 - Probability Steve is a librarian: 4/24 = 16.7%
The Key Mantra¶
"New evidence does not completely determine your beliefs in a vacuum. It should update prior beliefs."
Bayes' Theorem Components¶
| Term | Symbol | Meaning |
|---|---|---|
| Prior | P(H) | Belief before seeing evidence |
| Likelihood | P(E|H) | Probability of evidence given hypothesis is true |
| Likelihood (not H) | P(E|¬H) | Probability of evidence given hypothesis is false |
| Posterior | P(H|E) | Belief after seeing evidence |
The Geometric Diagram¶
Think of all possibilities as a 1×1 square: - Hypothesis occupies the left portion (width = prior) - Evidence restricts the space — but not evenly between hypothesis and not-hypothesis - Posterior = proportion of hypothesis in the restricted wonky shape - Irrelevant evidence doesn't change your beliefs (equal likelihoods → no change)
Making Probability Intuitive¶
Representative samples work better than percentages: - Kahneman & Tversky found that asking "40 out of 100" instead of "40%" dropped errors from 85% to 0% in the Linda problem - People's intuitions kick in better with concrete counts than abstract percentages
Area diagrams are more flexible: - Continuous (not discrete counts) - Easy to sketch on paper - Both sides of Bayes' theorem tell you the same thing: "look at cases where evidence is true, consider proportion where hypothesis is also true"
Broader Applications¶
- Scientific discovery — analyzing how new data validates/invalidates models
- Machine learning and AI — explicitly modeling a machine's beliefs
- Treasure hunting — Bayesian search tactics found a ship with $700M in gold
- "Bayes' theorem has a way of reframing how you even think about thought itself"