How data science ate the world

Before the mathematics, the story — because every formula in this curriculum was invented by someone trying to answer a question the world was asking loudly, and knowing which question makes the formula worth learning.

A missing planet and a stubborn question

On the first night of 1801, the astronomer Giuseppe Piazzi spotted a faint object — Ceres — and tracked it for forty nights before it slipped behind the Sun. The astronomy of the day could not say where it would re-emerge; the observations were too few and too noisy. A 24-year-old Carl Friedrich Gauss combined the noisy measurements by minimizing the sum of squared errors, predicted where Ceres would reappear — and in December 1801 it was found within half a degree of his prediction. (Legendre published the method first, in 1805; Gauss claimed a decade’s prior use; the priority fight was bitter and the method — least squares, lesson one of the Supervised Learning module — outlived them both.)

That episode is the whole field in miniature, two centuries early: messy measurements, a model, an optimization principle, a prediction that sticks its neck out — and a verdict from reality. Everything since is that loop, run faster, on more data, with better models.

• 1801–09Least squares finds Ceres; Legendre and Gauss formalize learning from noisy data.
• 1880s–1900sGalton and Pearson invent regression and correlation — the first machinery for relationships in data.
• 1920sAt Rothamsted’s farm plots, R. A. Fisher builds likelihood, experimental design and modern statistics to decide which fertilizer actually works.
• 1940sWar industrializes the loop: Bletchley Park’s Bayesian codebreaking, operations research, and the first computers built to run it.
• 1958Rosenblatt’s perceptron — the first machine that learns its own decision rule from examples, weights adjusted by error. The New York Times promises walking, talking machines; Minsky and Papert prove a single layer cannot even represent XOR (1969), funding collapses, and the field learns its other recurring lesson: the AI winter that follows overpromising.
• 1962John Tukey’s “The Future of Data Analysis” argues analysis of data is its own science, not a branch of mathematics.
• 1986Rumelhart, Hinton and Williams show backpropagation trains multi-layer networks — the chain rule, industrialized. XOR falls, hidden layers learn their own features, and the exact algorithm in this curriculum’s Deep Learning module still trains every frontier model today.
• 1998Two students rank the web with an eigenvector (PageRank) — a Foundations-module lesson becomes a trillion-dollar company.
• 2008–12”Data scientist” is coined at LinkedIn/Facebook; commodity clusters (MapReduce/Hadoop) make web-scale data workable.
• 2012AlexNet halves the ImageNet error rate — 1986’s backpropagation meeting 2012’s GPUs and data. The deep-learning era opens, twenty-six years after its algorithm was ready.
• 2016–17AlphaGo beats Lee Sedol with reinforcement learning (this platform’s home turf); the Transformer architecture arrives.
• 2020–22AlphaFold effectively solves protein-structure prediction; ChatGPT puts a large language model in everyone’s browser.

Why now? The three explosions

For most of that timeline, the ideas outran the means. Fisher had the likelihood principle in 1922 and computed with paper. What changed — and why “data science” became a profession around 2010 rather than 1962 — is that three curves went vertical simultaneously:

Data. Society digitized itself as a side effect of convenience: every purchase, search, sensor ping, medical image and satellite pass became a record. Data stopped being something scientists collected and became something civilization exhales.

Compute. Moore’s law compounded for fifty years, then GPUs — built to render video games — turned out to be matrix-multiplication engines, which is (as the Linear Algebra module makes precise) all a neural network needs. The same hardware lineage now trains everything in this curriculum’s playground.

Algorithms. Statistics, optimization and computer science merged their toolboxes. Gradient descent (Foundations) scaled to billions of parameters; the bias–variance trade-off (Supervised Learning) got navigated by architectures rather than formulas; reinforcement learning (this platform’s founding track) went from gridworlds to data centers.

The compounding of the three is the answer to “why is this knowledge so important now”: prediction became cheap infrastructure. Like electricity a century earlier, learning-from-data stopped being a capability and became a layer — one that most of society now runs on without noticing.

What the algorithms are actually doing out there

Every example below runs on mathematics taught in this curriculum — that is the criterion for inclusion, and the tag on each card tells you which lesson to blame.

And beyond the cards, quieter revolutions with the same DNA: logistics networks shaving fuel by learned routing, precision agriculture spraying grams instead of litres, astronomy triaging a night sky’s worth of transients per hour, particle physics filtering a billion collisions a second down to the interesting dozen. None of these announce themselves. That is the point — infrastructure doesn’t.

The part an honest curriculum must include

A field with this much leverage fails at scale too, and the failures are as instructive as the wins. Recidivism-risk scores deployed in US courts showed racially skewed error rates (the COMPAS controversy); early face-recognition systems misclassified darker-skinned women at rates an order of magnitude higher than lighter-skinned men (Gender Shades); the Dutch government fell in 2021 after a tax-authority algorithm wrongly branded tens of thousands of families as welfare fraudsters; a widely-deployed hospital sepsis model missed most sepsis cases while flooding clinicians with false alarms. Read those cases against this curriculum and a pattern appears: almost every one is an evaluation failure — unrepresentative training data, metrics blind to subgroups, no error bars, no monitoring after deployment. Which is why the driest lesson in this curriculum (Evaluation & cross-validation) is also the one with the highest stakes, and why this platform teaches measurement with the same rigor as modeling.

Why this is worth your effort

The premise of everything that follows: this layer of civilization is still being built, mostly by people who learned it the way you are about to — theory with derivations, then hands on real running systems. The curriculum ahead teaches the same least-squares principle that found Ceres (Supervised Learning), the same eigenvector idea that ranked the web (Linear Algebra), the same likelihood logic that ran Bletchley’s cryptanalysis (Probability), and the same reward-maximization loop that cools data centers (the RL track — with the playground where you train it yourself). None of it is historical decoration; every one is deployed, salaried knowledge. Start at Foundations, or jump straight into the playground and meet the algorithms before their equations.