Every AI model starts by understanding data — and that starts with five numbers. Play through 7 quick levels (including two hands-on variance & spread labs) to master the building blocks of statistics that power machine learning, then take on the Boss Round.
The balancing point of the data — add it all up, divide by how many.
Sort the values low to high — the one in the middle is the median.
Whichever value shows up most often — the tallest stack wins.
How far values stray from μ — squared, so far-off points count much more.
How far a typical point sits from the mean — watch it breathe: tight = small σ, spread out = big σ.
Every AI model starts with a pile of numbers. These five ideas are how you make sense of that pile — before a computer can learn anything from it.
The mean is the "balancing point" of a group of numbers — add every value up, then divide by how many there are. Mean Squared Error, one of the most common ways to score a model, is literally the mean of its squared mistakes.
In real life: your report card GPA is a mean — every grade gets added up and divided by how many classes you took.
Sort the values from smallest to largest, and the median is the one sitting right in the middle. When a few extreme values would throw the mean off — incomes, wait times, anything with a long tail — the median stays steady.
In real life: the "median household income" you hear about on the news is the middle income — so one billionaire down the street can't drag the number up.
One super-high value (faded bar) drags the mean μ to the right — but the median M barely moves. That gap is how you spot skewed, outlier-heavy data.
The mode is whichever value shows up most often. It's the simplest guess a model can make — "just predict the most common answer" — and it's exactly the baseline every smarter model has to beat.
In real life: the most popular pizza topping in your class is the mode — whichever one gets picked more than any other.
Variance measures how far values wander from the mean, then squares those distances so the far-off ones count much more. That squaring is deliberate — it's why a single outlier can dominate the number. A tight cluster of values has low variance; a scattered, spread-out group has high variance.
In real life: if everyone on a team scores close to the same on a test, variance is low. If scores are all over the place — some acing it, some failing — variance is high.
Standard deviation is just the square root of variance, brought back down to the same units as the original data — so instead of an abstract "squared" number, you get a distance you can actually picture. It's the standard way to describe how spread out something is.
In real life: a weather forecast saying "usually between 68–72°F" is basically giving you the mean plus or minus one standard deviation.
You can't fairly compare a 90-out-of-100 test with an 18-out-of-20 quiz just by looking at the raw numbers. Normalization subtracts the average and divides by the standard deviation, turning any score into "how many steps above or below typical" — so scores from completely different scales become directly comparable.
In real life: it's how colleges compare SAT and ACT scores side by side, or how a running app compares your pace to everyone else's — different raw scales, same "steps from average."
Image generators like Stable Diffusion add random noise to a picture on purpose, following a "variance schedule," then train a network to reverse that process step by step. Chatbots use this too: a setting called "temperature" controls how spread out their word choices are — low temperature keeps answers predictable, high temperature lets them wander further, exactly like a bigger standard deviation.
In real life: when an image generator turns a blurry cloud of random noise into a photo, it's reversing this exact process, one small denoising step at a time.
Regression, classification, neural networks, transformers — none of the topics that come later make sense without first being fluent in "typical" and "spread." That's why every one of them is built right on top of these five ideas.
In real life: it's like learning to add before you learn algebra — every AI topic after this one assumes you already know "typical" and "spread."
Think of it like darts on a target: variance is how far off your throws land, squared so bad misses count extra. Standard deviation brings that back down to a distance you can actually picture — a small σ means tightly grouped darts, a big σ means scattered ones.
These candies were weighed in grams. Find the mean (average) weight.
Click the tiles from smallest to largest to sort the data.
Players rolled a die several times. Which number came up most often (the mode)?
Here's a dataset with mean μ = . Each bar shows how far a value is from the mean. Which value below is the correct variance σ²?
Both datasets below have the same mean. Which one has the larger standard deviation (more spread out)?
These 5 points sit around the mean μ. Drag the slider to spread them out or pull them in, and watch the standard deviation react live. Match the target to lock it in.
Mean μ = . Tap Square it on each bar and watch how much taller the bigger distances grow once squared — that's why variance is so sensitive to outliers.
One dataset. Five answers. Solve them in order for bonus points.
Final score: 0 / 0