Summary: Updating beliefs with evidence: Bayes' theorem

Bayes’ theorem is how to change your mind correctly: start with a belief, weigh the evidence, end with an updated belief. The previous lesson showed that the chance of A given B is not the chance of B given A; Bayes is the machine that converts one into the other, keeping the base rate in the calculation so a scary-sounding result does not run away with your belief. It is the formal engine under lesson 1’s base-rate example and under spam filters. This summary is the scan-in-five-minutes version of the full lesson.

Core ideas

• The idea in one sentence. A new belief equals your old belief, adjusted by how well the evidence fits. You go from a prior to a posterior as evidence arrives.
• Build it from natural frequencies. Count a concrete population. For lesson 1’s disease (1 in 100, 99% accurate test) among 10,000 people: 99 true positives and 99 false positives, so P(disease | positive) = 99/198 = 0.50. The rare base rate matches every true positive with a false one.
• The formula. P(H | E) = P(E | H) x P(H) / P(E). The parts: prior P(H) (base rate), likelihood P(E given H) (how well evidence fits), evidence P(E) (total chance of E, including false positives), posterior P(H given E) (the answer). Formula and counting give the same 0.50.
• The denominator must include false positives. P(E) = P(E given H)P(H) + P(E given not H)P(not H). Forgetting the false-positive piece overstates the posterior.
• Today’s posterior is tomorrow’s prior. Evidence accumulates: a second independent positive test takes the disease example from 1% to 50% to 99%. The same machine handles each update.
• In AI. Bayes underlies spam filtering (naive Bayes, which assumes word independence), the discipline of combining a detector’s hit rate with the base rate to get the real probability, and the mindset of updating beliefs as data arrives. Reading a detector without its prior is base-rate neglect.

What changes for you

You get a disciplined way to react to evidence instead of overreacting to it. When a result lands, a flagged transaction, a positive test, a model’s alert, you no longer let the result alone set your belief. You ask two things: what did I believe before (the prior, the base rate), and how strong is this evidence really (the likelihood)? Bayes combines them. The payoff is immunity to the most expensive mistake in reading AI systems: treating a high hit rate as a high probability. A detector that catches 95% of a rare problem still flags mostly false alarms, and now you can compute, rather than guess, how seriously to take any single flag.