Your bank blocked a £14 payment at 3am before you’d noticed anything odd. A streaming service queued up a documentary you ended up finishing in one sitting. Neither system was written with rules about what fraud looks like or which films you enjoy. Both were trained on examples of the past and asked to make a judgement about something new.
That is machine learning in one sentence: rather than a programmer writing the rules by hand, a program finds patterns in data and turns them into a model that makes predictions. The vocabulary around it, all those neural networks and gradient descent and feature stores, is detail layered on top of that single idea.
How a model actually learns
Start with a spreadsheet. Each row is a case, each column is something you know about it. In a customer churn project those rows might be 40,000 subscribers, with columns for contract length, monthly spend, support tickets raised and, crucially, whether they eventually cancelled.
Training hunts for a set of weights that turns those columns into a decent guess about the last one. On the first pass the guesses are awful, close to a coin flip. The algorithm measures how wrong it is, nudges the weights slightly, and tries again. Repeat that a few thousand times and the error falls to something useful. The core loop has barely changed in spirit since the 1980s.
What changed is scale. A model with a few hundred million parameters trained on billions of examples can pick up patterns that no hand-written rulebook would ever capture, which is why the same underlying maths now powers spam filters, translation and image recognition.
The three learning styles you’ll keep hearing about
Supervised learning
You have labelled answers to learn from. Someone has already marked 8,000 loan applications as repaid or defaulted, or tagged 50,000 X-rays as clear or suspicious. The model learns the mapping. Most commercial machine learning is this, and it works best when the historical labels are trustworthy and the world stays roughly the same.
Unsupervised learning
No labels, no answers. The algorithm looks for structure on its own: grouping customers into behavioural clusters, compressing a hundred sensor readings into three, or flagging the one server log line that looks nothing like the rest. Useful for exploration, harder to prove value on a slide.
Reinforcement learning
An agent takes actions, receives a reward, and adjusts. DeepMind’s AlphaGo learned Go this way, playing millions of games against itself. The same machinery sits behind recommendation feeds that optimise for watch time, and it is notoriously fiddly because the reward you specify is rarely the behaviour you actually want.
Most of the work happens before the model exists
Practitioners consistently report that 70 to 80 percent of a project is data plumbing. Merging tables, reconciling timezones, deciding what to do with the 12 percent of rows where postcode is blank. The modelling step, the part everyone talks about, is often an afternoon.
Four questions worth answering before you train anything:
- What decision does this change? If nobody alters their behaviour based on the output, the model is a science project.
- What is the naive baseline? Last month’s average, or a single business rule, is often within a few points of a trained model. Beat it before celebrating.
- Is anything leaking? A column like “account closed date” will happily predict churn with 99 percent accuracy and teach you nothing.
- How will you know it drifted? Inputs shift, and a model can quietly rot for six months before anyone notices the accuracy drop.
Where machine learning quietly breaks
The famous pneumonia study is the cautionary tale. Researchers at Carnegie Mellon built a model that seemed to identify low-risk asthma patients, which looked like a breakthrough until they checked why. Asthmatic patients with pneumonia were rushed straight to intensive care, so their outcomes were better, and the model had learned the hospital’s triage routine rather than the biology. Deployed without that check, it would have sent fragile patients home.
Markets are the sharpest version of the same problem. A strategy fitted to ten years of price history can look superb until the regime changes and the relationships it memorised stop holding. Anyone using models to trade should understand the difference between a pattern that persists and one that was merely convenient, which is the theme of what machine learning really means for your portfolio.
Feedback loops deserve their own warning. A recommender trained on what people clicked will show more of what people clicked, which changes what they click next. The training data is no longer a neutral record of the world; your model is writing it.
What a real project looks like
Say a grocery chain wants to predict late deliveries. It has 300,000 past orders with distance, driver tenure, weather, day of week and warehouse backlog. A simple baseline predicts the average delay for every order and lands at about 22 minutes of mean absolute error. A gradient-boosted model gets that to 14 minutes by learning that rain on a Friday afternoon with a new driver is the bad combination.
That 8-minute gain is worth real money if it triggers earlier dispatch. It is also fragile in a specific way: once drivers learn they are being scored, some start accepting only easy routes, and the model’s inputs shift under its feet. Good teams monitor that, retrain quarterly, and keep a human in the loop for the edge cases the model has never seen.
Tools that cut the boilerplate
You no longer need a research team to get a first model out. AutoML platforms handle feature engineering, model selection and hyperparameter tuning behind a point-and-click interface, then hand back an explainable ranking of what mattered. DataRobot is a good example, aimed at analysts who know the business problem but not the maths.
Lighter options exist too. If you want to test whether a predictive feature is worth building, something like Abacus AI will get you from a CSV to a working prototype in an afternoon, which is often all the evidence a budget holder needs.
Skills worth building
The bottleneck in most organisations is not model architecture. It is people who can frame a problem sharply, judge whether the data is trustworthy, and explain the output to someone who will act on it. Those skills come from shipping small projects, not from collecting certificates. There is a decent roadmap in this piece on building AI skills without drowning in hype.
Where this is heading
Today’s systems are, for all their fluency, pattern matchers. They interpolate brilliantly within the territory they were trained on and stumble outside it. The next wave of research is aimed squarely at that gap: systems that build internal representations, plan a few steps ahead, and revise their own reasoning when it looks shaky. That is the territory covered by cognitive AI and machines built to reason rather than recall.
For now, the practical rule holds. Get the data right, define the decision, beat a boring baseline, and watch for drift. The teams that do those four things unglamorously well will outperform the ones chasing the biggest model on the leaderboard.