Probability Distributions

A probability distribution is a map from every possible outcome to its likelihood — and once you can build and read that map, you can reason about randomness, rank things fairly, and make systems that behave predictably instead of accidentally.

The Recommender That Ranked Everything Wrong

Picture this: EngineerPrep ships a new feature that surfaces the "hardest" interview questions. The logic is straightforward — count how many users answered each question incorrectly, then sort descending. In production, the list fills up with questions that thousands of beginners attempted and failed. Questions that only a handful of senior engineers ever tried — and also failed — sit near the bottom. The count for the beginner question is 800. The count for the senior question is 12. But 800 out of 10,000 attempts is only an 8% failure rate. 12 out of 14 attempts is an 86% failure rate. The senior question is far harder — yet it ranked lower. The raw count lied. What we needed was a way to describe the full spread of outcomes: not just how many times something happened, but how likely each outcome really is across all possibilities.…

What a Probability Distribution Actually Is

Imagine a six-sided die. You roll it once. There are six things that could happen: land on 1, 2, 3, 4, 5, or 6. A probability distribution is simply a mapping that pairs every possible outcome with the probability that outcome occurs. For a fair die that looks like: | Outcome | Probability | |---------|-------------| | 1 | 1/6 | | 2 | 1/6 | | 3 | 1/6 | | 4 | 1/6 | | 5 | 1/6 | | 6 | 1/6 | Every row is an outcome. Every probability is a number between 0 (impossible) and 1 (certain). And — crucially — all the probabilities add up to exactly 1, because something must happen. That's the whole idea. A distribution doesn't just capture one number. It captures the entire landscape of what can happen and how often. Back to EngineerPrep: instead of storing a raw wrong-answer count, the recommender now stores a failure rate — wrong answers divided by total attempts.…

Watching a Distribution Take Shape

Step 1 — One attempt. Imagine the EngineerPrep database has a single attempt on a question: the user got it wrong. We draw a number line from 0% to 100% failure rate. We place one dot at 100%. 0% 50% 100% • One dot tells us almost nothing. We have no distribution yet — just a single data point. --- Step 2 — Ten attempts, three wrong. Now ten users have tried the question. Three failed. We update our estimate: 30% failure rate. We move the dot. 0% 50% 100% • Still just one dot. The estimate shifted, but we still can't see a shape . --- Step 3 — One thousand questions. Now imagine all one thousand questions in EngineerPrep's bank each have at least 100 attempts. We plot a dot for every question's failure rate on the same line. 0% 50% 100% ··· ·· ··•••••••••••• ··· · Dots cluster between 20% and 60%. A few questions sit at the extremes. The cluster has a shape — a hump.…