What ML actually is
Cut through the hype. Here's what machine learning is, what it isn't, and where it sits in the modern data stack.
What you'll learn
- The minimal definition of ML that holds up
- When ML is the right tool — and when it isn't
- The three flavors (supervised, unsupervised, reinforcement) and what they're each for
Before you start
How does your inbox know a brand-new message is spam, when nobody wrote a rule for that exact email? It learned the pattern from millions of past messages people marked as spam or not. That, in one sentence, is the whole idea: machine learning is software that learns rules from examples instead of having them coded by hand. Everything else is implementation detail.
A traditional program looks like:
input + rules → outputML flips this around:
input + output (lots of examples) → rules (the model)Then at prediction time, the learned rules turn new inputs into outputs.
The three flavors
| Flavor | What you give it | What it learns | Example |
|---|---|---|---|
| Supervised | Pairs of (input, correct answer) | A function mapping input → answer | Spam classifier, house-price predictor |
| Unsupervised | Just inputs, no labels | Hidden structure (clusters, components) | Customer segmentation, anomaly detection |
| Reinforcement | An environment and a reward signal | A policy that maximizes reward over time | Game-playing agents, robotics, ad bidding |
In industry, supervised learning is 95% of what gets deployed. The other two show up in narrower domains. We’ll cover all three but spend most of our time on supervised.
When NOT to use ML
(The transaction example needs no ML at all — amount > 10000 is an exact, stable
rule. Reaching for a model there adds cost and unpredictability for nothing.)
ML is the right tool when:
- You can’t write the rules by hand (image recognition).
- The rules are stable enough that examples won’t go stale next month.
- You have lots of labeled data — or can get it cheaply.
- A wrong prediction has bounded cost (not life-critical without safeguards).
ML is the wrong tool when:
- A simple deterministic rule works. (Don’t use ML to detect even numbers.)
- You have 50 training examples. (Use heuristics, or get more data first.)
- The cost of a wrong answer is catastrophic and unreviewable.
- The system needs to explain its reasoning legally (some regulators forbid black-box models).
ML vs “AI”
The terminology is a mess. A rough mapping:
- Machine learning — the field. Linear regression to deep neural nets.
- Deep learning — ML with deep neural networks (multiple hidden layers).
- Generative AI — deep learning models that generate text/images/etc.
- LLMs — generative AI for language.
- “AI” in business contexts usually means generative AI + maybe some agents on top.
If a colleague says “AI,” they probably mean LLMs. If they say “ML,” they probably mean classical models (regression, trees, boosting). The distinction matters because the tools and skills differ.
In one breath
- Machine learning is software that learns rules from examples instead of having
them hand-coded — the inversion: traditional code is
input + rules → output, ML isinput + outputs (examples) → rules. - Three flavors: supervised (labeled pairs → a mapping), unsupervised (inputs only → hidden structure), reinforcement (environment + reward → a policy). Supervised is ~95% of what ships.
- ML is the right tool when rules are too messy to write by hand, you have lots of (cheap) labels, and a wrong prediction has bounded cost.
- ML is the wrong tool when a simple deterministic rule works, data is tiny, errors are catastrophic-and-unreviewable, or the law demands explainability.
- “AI” in business usually means LLMs / generative; “ML” usually means classical models (regression, trees, boosting) — different tools, different skills.
Quick check
Quick check
What’s coming in this section
The next lessons cover the scikit-learn API (the convention every ML library copies), the train/test split (the most important habit), and then we’ll work through regression, classification, trees, and ensembles — the techniques that win on most tabular problems today.
Practice this in an interview
All questionsClassification predicts a discrete class label; regression predicts a continuous numeric value. The choice is determined by the nature of the target variable, not by the algorithm family — many algorithms (e.g., decision trees, neural nets) handle both.
Supervised learning trains on labeled input-output pairs to predict a target. Unsupervised learning finds structure in unlabeled data. Reinforcement learning trains an agent to maximize cumulative reward through trial-and-error interaction with an environment.
Generalization is a model's ability to perform well on unseen data drawn from the same distribution as the training set. It is controlled by the interplay of model capacity, dataset size, regularization, and distributional shift between training and deployment.
AutoML automates parts of the ML pipeline such as data preprocessing, feature engineering, model selection, hyperparameter tuning, and sometimes neural architecture search, lowering the barrier to building models. It falls short on problem framing, data quality, domain feature engineering, careful evaluation against leakage, fairness, and deployment concerns, which still need human expertise. It's best as an accelerator and strong baseline generator, not a replacement for an ML engineer.