The Curiosity University

Picture a coffee counter that can serve 60 people an hour, with 50 people arriving each hour. On paper the counter has spare capacity. In practice there is often a line, and sometimes a long one. Why?

Fact Queueing theory, the mathematics of waiting lines, gives a blunt answer: arrivals and service times vary. People don’t arrive evenly every 72 seconds. Three arrive together, then nobody comes for five minutes. Some orders take twenty seconds and some take three minutes. Time when the server sits idle is lost for good and can’t be saved for the next rush, so variation turns into waiting even when average capacity is enough.

The key number is utilisation: the arrival rate divided by the service capacity, written ρ (rho). In the simplest textbook model, with random arrivals, random service times and one server, the average number of people in the system is ρ / (1 − ρ). At 50% utilisation that is 1 person. At 80% it is 4. At 90% it is 9. At 95% it is 19. The curve doesn’t rise in a straight line. It bends upwards sharply as you get close to full capacity.

Fact A second, remarkably general result is Little’s law, proved by John Little in 1961: the average number in a system equals the arrival rate multiplied by the average time each person spends there (L = λW). If 50 people an hour arrive and each spends 6 minutes (0.1 hours) in the café, there are on average 5 people inside. The law holds for almost any stable queue, whatever the arrival pattern.

Interpretation This explains why organisations that run “at full efficiency” often feel slow. A hospital ward, a help desk or a road running close to capacity has no slack to absorb surprises, so small fluctuations become long waits. Slack looks wasteful on a spreadsheet, but it is what keeps the line short.

Two other levers matter besides adding capacity. Reducing variability by scheduling appointments, standardising tasks or batching similar work shortens queues even at the same utilisation. Pooling also helps: one shared line feeding several servers usually beats separate lines for each server, because an idle server can take whoever is next.

Two children, one cake. The old rule is that one cuts and the other chooses. It needs no referee and no agreement about what the cake is worth, yet both children end up satisfied in a precise sense. How?

Fact In divide and choose, the cutter splits the cake into two pieces they value equally, by their own judgement. Whichever piece the chooser takes, the cutter gets exactly half of the value as they see it. The chooser picks the piece they prefer, so they get at least half as they see it. Each person is guaranteed a fair share by their own measure, and neither envies the other. Economists call this proportional and envy-free for two people.

Notice what the rule does not need. Nobody has to agree on what the cake is worth. One person can love the chocolate end and the other can be indifferent to it. The procedure makes fairness something each person can check for themselves, rather than a verdict handed down by an authority.

Fact Extending this to more people is surprisingly hard. An envy-free procedure for three people (the Selfridge–Conway method) was found around 1960. A procedure for any number of people with a bounded number of steps was published only in 2016, by Haris Aziz and Simon Mackenzie, and the number of steps it can need is astronomically large.

Interpretation Fairness comes in several kinds that can conflict. Procedural fairness asks whether the process was one everyone could accept beforehand. Outcome fairness asks whether the result is equal, deserved, or matched to need. Divide and choose is a procedural answer. It guarantees each person half by their own lights, but it doesn’t ask who is hungrier, who baked the cake, or whether one child has had no dessert all week.

There is also an information problem, which the experiment shows. A cutter who knows what the chooser likes can cut strategically and keep more than half of what they value, while the chooser still gets a piece they think is at least half. The rule protects each person’s minimum. It does not stop the better-informed person from gaining an advantage.

Most of what we believe comes with a feeling of certainty, not a stated level of confidence. “I’m sure the meeting is on Tuesday.” When it turns out to be Wednesday, we say we were wrong. But how wrong? And were we wrong to be so sure?

A useful move is to treat confidence as a claim in its own right. If you say “70% sure” about many different things, roughly 70% of them should turn out true. If only half do, you were overconfident. If 90% do, you were underconfident. Matching your stated confidence to how often you are right is called calibration.

Fact In 1950 the meteorologist Glenn Brier proposed a simple way to score probability forecasts. For each forecast, take the probability you gave, subtract what happened (1 if it happened, 0 if not), and square the result. Then average over all forecasts. Lower is better: 0 is perfect, and always saying 50% scores 0.25. The Brier score is still used in weather forecasting and in forecasting research today.

The score has a valuable property: it is proper. You get the best expected score by reporting what you actually believe. Exaggerating your confidence to sound decisive, or hedging to sound safe, both make your expected score worse.

Interpretation This changes what “being wrong” means. One wrong answer at 60% confidence is not a failure. You expected to be wrong 40% of the time. What counts against you is a pattern, such as being wrong half the time when you claimed 90%. You can only see that pattern if you write your confidence down before you learn the answer. Memory edits our past certainty to fit what happened, an effect psychologists call hindsight bias.

So the practical habit is a forecasting log. Write the claim, the probability and the date. Check it later. After a few dozen entries you start to learn how far your “sure” can actually be trusted.

Try describing the smell of rain on dry ground, or of a grandparent’s cupboard, without naming the thing itself. Most English speakers fall back on sources (“it smells like cedar”) or on feelings (“it smells nice, it smells like home”). We have few words for the smell itself.

Fact Smell starts with chemistry. Volatile molecules reach receptors high in the nose. Humans have roughly 400 kinds of functional olfactory receptor, and a single odour usually activates a combination of them. The brain reads that pattern, much as a chord is heard from several notes played together.

Fact For a long time, poor odour vocabulary was treated as a human universal. In 2014 Asifa Majid and Niclas Burenhult tested speakers of Jahai, a language of hunter-gatherers on the Malay Peninsula. Jahai speakers named odours as consistently and concisely as they named colours, using dedicated abstract smell words. English speakers in the same task gave long, varied, source-based descriptions. The difficulty seems to depend partly on language and culture, not only on human biology.

Interpretation That makes scent an unusually clear case of meaning built in layers. The first layer is chemical: which molecules are present. The second is perceptual: which pattern of receptors fires. The third is learned and cultural: what this pattern has been paired with, such as a festival, a hospital, a person or a warning of spoiled food. Perfumers, sommeliers and chefs train the second and third layers, and build shared vocabularies so they can talk precisely.

This is also why scent can feel so personal. Two people may detect the same molecules but attach completely different associations. A smell that means “clean” to one person means “waiting room” to another.

Peel an orange and try to press the peel flat on the table. It tears or wrinkles. A map has the same problem with the Earth.

Fact Mathematicians proved this in the 19th century: no flat map can show every distance on a sphere correctly. Each projection keeps some properties and gives up others. Some keep angles (conformal maps), some keep areas (equal-area maps), and some compromise between them. None keeps everything.

Fact The Mercator projection, published by Gerardus Mercator in 1569, keeps angles. A line of constant compass bearing (a rhumb line) is straight on the map, which made it extremely useful for sea navigation. To manage this, it stretches the map more the further you go from the equator. At latitude φ, lengths are scaled by 1 / cos φ, so areas are scaled by its square. At 60° north, lengths are doubled and areas are four times larger than at the equator.

That is why Greenland looks about as big as Africa on many wall maps. In fact Africa is about 30 million km² and Greenland about 2.2 million km², so Africa is roughly fourteen times larger.

Interpretation A projection is a choice about what matters. A sailor needs true angles. Someone comparing farmland or population density needs true areas. A web map that zooms seamlessly into streets needs local shapes to look right, which is one reason many online maps use a Mercator variant. The “lie” isn’t a mistake. It is a trade-off that is easy to forget.

The wider lesson is that any representation, whether a chart, a summary, a ranking or a map, keeps some things and loses others. A good question to ask of any picture of the world is: what was this made to preserve?

How many fish are in a lake? You can’t drain it, and fish don’t line up to be counted. Ecologists face this question all the time with butterflies, newts, bears and birds. The answer is a small, clever idea.

Fact In mark and recapture, you catch a sample, mark every animal (a tag, a clipped fin, a dab of paint) and release them. Later you catch a second sample and count how many are already marked. If 50 fish were marked and 10% of the second catch carries a mark, then the 50 marked fish are probably about 10% of the whole lake, which means around 500 fish. This is the Lincoln–Petersen estimator: N ≈ (marked × second catch) / recaptured.

The idea rests on proportional reasoning. The second catch acts as a small window onto the lake, and the fraction of marked fish in it estimates the fraction in the whole population.

Fact The method depends on assumptions. The population must be closed, with no births, deaths or migration between samples. Marks must not fall off. Most importantly, every animal must be equally likely to be caught, and being caught once must not change that. Statisticians also use a corrected version (Chapman’s estimator) because the simple formula is biased for small samples.

Interpretation The same logic appears well beyond ecology. Epidemiologists estimate how many cases surveillance systems miss by comparing overlapping lists, and human-rights researchers have used “multiple systems estimation” to estimate unrecorded deaths from several partial lists. Whenever you have two imperfect views of a hidden group, the overlap between them carries information about everything neither view saw.

Ask a room of people to write down a “random” sequence of 30 coin flips, and ask a few others to flip a real coin 30 times. A practised teacher can usually tell the two kinds of sheet apart at a glance. The real ones have longer streaks.

Fact In 30 fair flips, the longest run of heads or tails is usually about four to six, and runs of seven or more are not rare. Real sequences switch between heads and tails about half the time, around 14–15 switches in 30 flips. Studies of people producing “random” sequences have found that they tend to switch too often and avoid long runs. They over-correct towards a pattern that looks balanced.

Interpretation Our sense of randomness seems to be tuned to “no visible pattern”, but genuine chance produces patterns all the time. This matters outside the classroom. A cluster of illnesses in one street, a run of good days in the market, or three lucky interviews in a row may need no special explanation. Before looking for a cause, it helps to ask how often chance alone would produce something this clumpy.

The opposite mistake matters too. Real clusters, such as a polluted well or a skilled player, do exist. The task is not to dismiss every pattern but to compare it with a sensible baseline of what chance would do.

Two hospitals publish survival rates. Hospital A: 78%. Hospital B: 86%. Which would you choose?

It depends on who they treat. Suppose Hospital A takes most of the region’s severe cases. Within mild cases A does better than B, and within severe cases A also does better, yet because it treats so many more severe patients its overall rate is lower. This reversal is Simpson’s paradox. It is not a trick of arithmetic. It happens whenever groups are mixed in different proportions.

Fact A famous real case is graduate admissions at the University of California, Berkeley, in 1973. Overall, men were admitted at a noticeably higher rate than women. When Peter Bickel and colleagues looked department by department, most departments showed little difference, and some slightly favoured women. Women had applied more often to departments that admitted a small share of all applicants. The overall figure combined two different stories.

Interpretation An average is a weighted summary, and the weights quietly carry information. When you see an average, it is worth asking: averaged over what mix, and would a different mix change the story? Batting averages, school results, drug trials and wage gaps can all hide reversals or exaggerations when groups are combined.

The paradox doesn’t tell you which number is “true”. That depends on the question being asked and on how the variables cause each other. Sometimes splitting into groups gives the right answer. Sometimes the combined figure does, for example when the grouping variable is itself caused by the treatment you are studying.