Machine learning is the practice of building systems that derive their behaviour from data rather than from explicitly written rules. "Artificial intelligence" is the broader term for systems performing tasks that would otherwise require human judgement; in current practice most of what carries that label is machine learning underneath.
01How a model is produced
The distinction that clarifies most discussions is between training and inference. Training is the process of fitting a model to data — computationally expensive, done ahead of time, and producing a set of learned parameters. Inference is running that finished model on new input, which is comparatively cheap and is what happens when a product feature is used. The two have entirely different cost, latency and infrastructure profiles.
Learning approaches are usually grouped in three ways. Supervised learning fits a model to examples that are already labelled with the desired output, and covers most classification and prediction work. Unsupervised learning finds structure in data with no labels, such as clustering or dimensionality reduction. Reinforcement learning learns from a reward signal produced by acting in an environment rather than from a fixed dataset.
Neural networks are the model family behind most recent capability gains: layered numerical functions whose parameters are adjusted by gradient descent. Their capacity to represent complex relationships is what enables systems that work on language, images and audio without hand-written feature engineering.
02What determines whether a system works
Data quality dominates. A model reproduces the patterns in its training data, including the errors, gaps and biases in it, and no amount of architecture work compensates for a dataset that does not represent the situation the system will face.
Evaluation is the second determinant, and the most commonly mishandled. A model must be measured on data it did not train on, or the measurement reports memorisation rather than capability. Overfitting — fitting the training set so closely that performance on new data degrades — is the standard failure mode, and holding out a genuinely separate test set is the standard defence.
The metric chosen also encodes a decision. Accuracy is misleading when the classes are unbalanced, and precision and recall trade against each other, so the choice between them is a judgement about which kind of error is more costly in context.
03Deployment realities
Generative models produce fluent output regardless of whether that output is correct, and they do not signal their own uncertainty reliably. Any workflow that depends on factual accuracy needs verification outside the model, not a stronger instruction to the model.
Models also drift. The world moves away from the data a model was fitted to, so performance degrades over time without any change to the code, which makes ongoing monitoring a requirement rather than an optional maturity step.
Finally, sending data to a model is a data-handling decision. Where the inference runs, what is retained, and whether inputs are used for further training are questions to answer explicitly before any workflow involving customer or confidential material is built on top of one.