AI and ML Core Terminology and Paradigms

Sort the family before you learn the services

Every later decision — which service, which data, which deployment option — rests on a handful of words that are often used as if they meant the same thing. They do not, and questions are written precisely to separate them.

Artificial intelligence (AI) is the broad field of creating or imitating machines that perform tasks traditionally requiring human intelligence, including understanding language, visual perception, decision-making, and problem-solving.

Machine learning (ML) is a branch of AI in which techniques enable computers to learn from data and make decisions instead of relying on explicit programming for every rule. The system generalizes from examples rather than following rules written by hand.

Deep learning is a specialized subset of ML that uses neural networks with multiple layers to analyze complex factors in data.

The three terms nest in one direction only, and that direction is the point: AI is the umbrella, ML is a branch within AI, and deep learning is a subset within ML.

Neural networks are the technique used in deep learning: a deep-learning neural network has multiple layers that analyze complex factors in data. Deep learning is especially good at unstructured content such as images and text.

Reversing the nesting is the classic trap. Deep learning never contains AI, and ML never contains AI.

Seeing and reading: two applied fields

Two ML fields come up constantly because they are named after the kind of input they handle, not after the algorithm inside them.

Computer vision uses ML to see and understand images, such as finding defects or anomalies in visual representations. A manufacturer can point it at production-line photographs to spot a missing component or a tiny surface flaw at a scale that manual inspection cannot match.

Natural language processing (NLP) refers to ML algorithms that help computers understand human language, for tasks such as translation, intelligent search, and text analysis. Support tickets, medical notes, and call transcripts are its natural material.

When a scenario mixes both — images and written reports — it needs both fields, each on its own modality.

Algorithm, training, model, inference

These four words describe one pipeline, and each names a different thing in it. Mixing them up is the single most common terminology error.

An algorithm is a recipe for addressing a machine learning problem.

Training provides a learning algorithm with example data so that it can discover patterns. The training data contains the correct answer, known as the target, and the algorithm finds the patterns that map input attributes to that target.

A model is the artifact produced by training. It captures the patterns that the learning algorithm found — it is the output of the process, never the input.

Inference applies the trained model to new data to generate a prediction when the target answer is not already known.

The trained model is the boundary between the two stages: training creates the model from examples, whereas inference uses that model after training.

A spam filter makes the sequence concrete. You supply emails already marked spam or not spam, training produces a model from them, and inference then judges a new message whose label nobody has written yet.

Bias and fairness are different questions

Both words concern groups of people, which is why they get confused, but they sit at different points of the work.

Bias is an imbalance in the training data or in the prediction behavior of the model across different groups, such as age or income bracket. It can result from the data or from the algorithm used to train the model. A model trained mostly on data from middle-aged people may simply be less accurate for younger and older people.

Fairness considers the effects of an AI system on different stakeholder groups. It is one of the responsible-AI dimensions, and it asks about impact rather than about a single statistic.

Bias can be measured before training, measured again after training, and monitored against a baseline once a model is serving predictions.

Model fit: underfitting and overfitting

Model fit describes how well a predictive model captures the relationship between inputs and targets. You diagnose it by comparing the prediction error on the training data with the error on the evaluation data — one number alone tells you nothing.

An underfitting model performs poorly on its training data, because it cannot capture the relationship between inputs and targets at all. It is often too simple for the problem.

An overfitting model performs well on training data but poorly on evaluation data, because it memorizes the examples it has seen and cannot generalize to unseen ones.

The corrective step follows the diagnosis. Underfitting calls for more model flexibility, such as richer features or less regularization. Overfitting calls for less flexibility, such as fewer feature combinations or more regularization. When both numbers are poor, the algorithm may simply not have had enough data to learn from.

Note what the good-then-poor split is not: it is not fairness, and it is not inference, because at evaluation time the answers are already known.

Language models and generated content

A large language model (LLM) is a deep learning AI model pretrained on a vast amount of data. One LLM can perform many language tasks: answering questions, summarizing documents, translating text into other languages, and completing sentences.

Generative AI (GenAI) refers to AI systems capable of creating new content — text, images, or code — from input data or prompts. GenAI names a behavior, not a particular file: the behavior of producing something that was not in the input.

The two words describe different layers of the same product. The LLM is the pretrained artifact an application calls; creating new prompted text is what the application does with it.

Agents that pursue goals

Agentic AI consists of systems whose autonomous software agents can perceive context, reason over goals, make decisions, and take purposeful actions on behalf of users or systems. Such agents use AI models, integrate with tools and resources, and pursue delegated objectives with memory and intent.

The step beyond generation is action. A generative system asked for a refund email writes the email. An agentic system given the goal "resolve this ticket" decides which approved tool to call and calls it.

These paradigms overlap, but they are not interchangeable.

An LLM can supply the language capability inside a generative or agentic product, but the LLM is a model rather than the whole workflow built around it.

Key takeaways

• The nesting runs one way: AI is the umbrella, ML is a branch of AI, and deep learning is the subset of ML built on neural networks with multiple layers.
• Name the pipeline correctly: the algorithm is the recipe, training discovers patterns from example data, the model is the artifact that stores them, and inference applies that model to new data whose answer is unknown.
• Match the field to the input: computer vision sees and understands images, while NLP helps computers understand human language.
• Keep the quality words apart: bias is imbalance across groups arising from data or algorithm, fairness weighs effects on stakeholder groups, and fit is judged by comparing training and evaluation performance.
• Generation is not agency: GenAI creates new content, while agentic AI reasons over goals and takes purposeful action, often using an LLM as its language engine.