Three doors: one hides a car, two hide goats. You pick door 1. The host—who knows where the car is—opens a goat door, say door 3, then asks: switch to door 2, or stick with door 1?
Nearly everyone blurts out "two doors left, 50/50, it makes no difference." Wrong. Switching wins with probability $2/3$. One sentence cracks it open: your first pick lands on the car only $1/3$ of the time and misses it $2/3$ of the time. And whenever you missed (the $2/3$ case), the host is forced to dodge the car and open the one remaining goat door—so the door left standing must be the car. "Switch" therefore bets on "I picked wrong to begin with," an event whose probability is exactly $2/3$. The host is not opening a door at random; his move leaks information.
Let the car sit behind door $i$, you pick door 1, the host opens door 3. By Bayes: $P(\text{car at }2\mid\text{opens }3)=\dfrac{P(\text{opens }3\mid\text{car at }2)P(\text{car at }2)}{P(\text{opens }3)}=\dfrac{1\cdot\frac13}{\frac12}=\dfrac23$. In the numerator $P(\text{opens }3\mid\text{car at }2)=1$—if the car is at 2 he can only open 3; whereas if the car is at 1 he opens 2 or 3 with equal chance. That asymmetry is what manufactures the $2/3$.
Scale it to 100 doors and it becomes obvious: you pick door 1 (a $1/100$ shot), the host flings open 98 goat doors and leaves only yours and one other. The door he deliberately avoided concentrates the entire probability of the other 99. Intuition fails not at the arithmetic but at ignoring the condition "the host knows"—information changes probability. This is the seed of the Bayesian worldview: probability is not a property of objects but a measure of how much you know.
This is Bayesian updating in miniature. In medical diagnosis, how one test result reshuffles disease probabilities is the very same computation; so are spam filtering, the posterior of an A/B test, and a robot's SLAM tightening its position distribution as sensor readings arrive. It also works as a diagnostic mirror: many "statistical intuition failures" stem from treating an informative observation as a random one.
Put just 23 people in a room and it is already more likely than not that two share a birthday. Twenty-three feels absurdly few—there are 365 days in a year.
What fools you is the wording. You instinctively think "someone shares my birthday," which is indeed hard. But the question asks about "any two people." Twenty-three people form $\binom{23}{2}=253$ pairs, and each pair is one chance for a match. With 253 chances thrown at 365 slots, a hit is no surprise. The point: the quantity to compare is not the number of people but the number of pairs—and that grows with the square of the people.
It is easier to count "all distinct" and subtract: $P(\text{no repeat})=\dfrac{365}{365}\cdot\dfrac{364}{365}\cdots\dfrac{365-n+1}{365}$, so $P(\text{at least one pair})=1-\text{that}$. For modest $n$ there is a clean approximation $P\approx 1-e^{-n(n-1)/730}$—the $n(n-1)/2$ in the exponent is precisely the number of pairs, and $730=2\times365$. Plugging in $n=23$ gives $\approx 0.507$.
It puts "combinatorial explosion" on full display: collision chances scale as $\sim n^2/2$, not $n$. So to make some coincidence more likely than not, the sample size you need is about the square root of the number of slots—$\sqrt{365}\approx 19$, the same order as 23. This "$\sqrt{N}$ threshold" is not special to birthdays; it is the universal rhythm of every "throw balls into $N$ buckets at random" problem. It feels wrong because the human mind extrapolates linearly and is unprepared for quadratic growth.
It is the heart of the cryptographic birthday attack: to find a hash collision you need not try all $2^{n}$ inputs, only about $2^{n/2}$—which is why a 128-bit hash offers only 64 bits of collision resistance, and the mathematical root of MD5's downfall. The same rhythm shows up in hash-table load design, in estimating UUID collision probabilities in distributed systems, and in the sampling analysis behind data deduplication.
Treatment A beats B among mild cases, and beats B among severe cases too—A wins in every group. Yet merge the two groups and B's overall cure rate comes out higher. This is not an arithmetic slip; it recurs in real data (the famous 1973 Berkeley graduate-admissions "bias against women" was exactly this).
The trick hides in "who ended up in which group." Suppose A was mostly assigned the severe cases (hard to cure to begin with) and B mostly the mild ones—then A's overall score is dragged down by the difficulty it shouldered. The group proportions, not the efficacy, drive the merged numbers.
At its core is an elementary but lethal inequality: even if $\dfrac{a_1}{b_1}>\dfrac{c_1}{d_1}$ and $\dfrac{a_2}{b_2}>\dfrac{c_2}{d_2}$, it does not follow that $\dfrac{a_1+a_2}{b_1+b_2}>\dfrac{c_1+c_2}{d_1+d_2}$. Fractions cannot be added term by term. Geometrically, two fractions are the slopes of two vectors in the plane; "merging" is vector addition, and the slope of the sum is reweighted by the lengths (sample sizes) of the components.
It is the sharpest geometric demonstration that "correlation is not causation." The data does not lie; the question "which layer should I look at?" simply has no purely statistical answer—it must be settled by causal structure. Whether to split depends on whether the grouping variable is a confounder (split on it—like severity) or a mediator (do not split—it would block the real effect). Judea Pearl uses it to show: from the same numbers, different causal graphs yield opposite correct conclusions. Statistics was forced, for the first time, to admit that data alone is not enough.
It is the number-one trap in A/B testing and observational studies: the overall metric may rise while every user segment falls (only because the traffic mix shifted). In ML fairness, a model can be "calibrated" within every subgroup yet systematically biased in aggregate. Clinical trials, economic policy evaluation, and the segment analysis behind recommender systems must all draw the causal graph first and only then decide the aggregation level—otherwise they will confidently reach the opposite of the truth.
Draw a chord at random in a circle. What is the probability it is longer than the side of the inscribed equilateral triangle? It sounds like a question with a single answer. Yet three equally reasonable ways to "draw at random" give three different numbers: $\tfrac12$, $\tfrac13$, $\tfrac14$.
Nobody miscalculated. The trouble is that "draw a chord at random" is never actually pinned down. "Random endpoints," "a random midpoint," "a random direction plus a random distance from the center"—all deserve the word "random," yet each sprinkles a different uniform distribution over the space of chords. Change how you sprinkle the points and the answer changes.
The three parameterizations correspond to three different probability measures. Using "the chord's midpoint" as the coordinate: the random-endpoints method makes midpoint density thinner near the center, giving $1/3$; the random-midpoint method spreads midpoints uniformly by area over the whole disk, giving $1/4$; the random-radius method makes midpoints uniform along a radius, giving $1/2$. The event "the chord is longer" simply has those three measures under those three measures. Without specifying a measure, the word "probability" has no definition.
It slices straight through the soft spot of the "principle of indifference"—"with no reason to favor any outcome, assume they are equally likely": "equally likely" in a continuous space depends on which variable you make uniform. Uniform in the endpoints is not uniform in the midpoint. This is precisely why measure theory (Day 21) must serve as probability's foundation: probability does not grow automatically out of a "sample space"; you must additionally impose a measure, and it is a separate thing from the space itself. With one small geometry puzzle, Bertrand drags that deep necessity into the open.
It is the hidden reef of every Monte Carlo simulation: which variable your "uniform sampling" is uniform over directly decides whether the result is right—especially deadly on spheres and on parameter manifolds. The choice of an "uninformative prior" in Bayesian statistics is the same ghost: a uniform prior stops being uniform under a change of parameters, which is why we have Jeffreys priors and maximum-entropy priors that chase some invariance. E. T. Jaynes argued that adding the physical constraint "the answer should be invariant under translation and scaling" makes Bertrand's problem converge to the unique answer $1/2$—pinning down "random" by appealing to symmetry.