Machine learning (ML) is the discipline of building algorithms that improve at a task by being exposed to examples, rather than by being explicitly programmed. If a traditional program is rules in, answers out, an ML system is examples in, rules out — the rules being a learned set of statistical weights inside a model.
ML is usually divided into three families. Supervised learning trains on labelled examples — emails marked as spam or not, images tagged as cat or dog — and learns to predict the label for new inputs. Unsupervised learning finds structure in unlabelled data through clustering, dimensionality reduction or topic modelling. Reinforcement learning trains an agent to take actions in an environment, getting a reward signal for good behaviour; this is what powers game-playing systems and robotic control, and a variant (RLHF) is used to align large language models.
In a typical US business pipeline, an ML project looks like this: collect data, clean and label it, split into train/validation/test sets, train a model, measure performance against a baseline, and ship the best version behind an API. Fraud detection at a fintech, churn prediction at a SaaS, demand forecasting at a retailer — these are classic supervised ML workloads still in production every day.
Modern deep learning is a sub-field of ML built on neural networks with many layers, and it is what powers the LLM era. But not every problem is a deep-learning problem. For tabular data with a few thousand rows, gradient-boosted trees (XGBoost, LightGBM) often beat neural networks while being cheaper and easier to debug. A serious ML practitioner picks the smallest model that solves the problem.
The most common reasons ML projects fail in production are not algorithmic. They are: the training data does not match production data (distribution shift), the metric chosen does not reflect business value, or no one owns retraining when the world changes. Building for the second model — the one you ship six months in once reality has drifted — is the hallmark of a mature ML team.
For builders, learning ML in 2026 means three things: enough statistics to read a confusion matrix, enough Python to handle pandas and scikit-learn, and enough engineering to deploy and monitor. The model itself is the easy part.
