Understand AI, one concept at a time.
Direct definitions connected to examples, common mistakes, technical decisions, and verified editorial analysis.
Fundamentals
Basic units, capabilities, and limits for understanding how an AI system works.
Multimodality: definition, how it works, and its limits
Multimodality lets AI systems process and relate information from more than one kind of input, such as text, images, audio, and video. This glossary entry explains the technical pipeline, develops examples from healthcare, manufacturing, and autonomous vehicles, and examines challenges including data alignment, computing costs, and privacy.
Parameter
Numerical value learned during training that contributes to transforming inputs into outputs.
Token: Definition, Mechanics, and Limits in Language Models
A token is the smallest unit of text that a language model processes to understand and generate content. It is not equivalent to a word or a character, but rather a fragment defined by the specific tokenizer of each model. This entry explains how text is split, how identifiers are assigned, how tokens influence context limits and API costs, and provides practical examples to optimize their use.
Context window
Maximum number of tokens a model can consider across input, history, tools, and output in a request.
Models and learning
Architectures, training, and techniques that shape model behaviour.
Fine-Tuning: What Changes When You Continue Training a Model—and What It Does Not Guarantee
Fine-tuning is additional training of a previously trained model to adapt it to a specific objective. This entry explains the process, walks through three examples, and distinguishes the concept from other ways of adapting or using models.
Self-Supervised Learning: What It Is and How It Works
Self-supervised learning creates training objectives from the data itself. This guide explains how the approach works, gives examples from language, vision, and audio, and clarifies its limits and how it differs from other methods.
RLHF: What It Is, How It Works, and Where Its Limits Lie
Reinforcement learning from human feedback uses human judgments or preferences to guide model tuning. A common workflow generates responses, collects comparisons, trains a learned reward model, and optimizes a policy; each stage involves choices and limitations that matter.
Synthetic Data: What It Is, How It Is Generated, and What It Does Not Guarantee
Synthetic data is generated artificially to preserve properties of real data that matter for a specific purpose. Its artificial origin does not, by itself, show that it is useful, high-quality, or private.
Knowledge Distillation: How a Student Model Learns from a Teacher—and What It Does Not Guarantee
Distillation transfers training signals from a teacher model to a student. It can take different forms; it does not mean that the student preserves every capability or is more efficient in every environment.
Loss Function: What It Measures During Training—and What It Doesn’t Tell You About a Model
A loss function quantifies the discrepancy between a prediction and a target according to a defined criterion. Understanding what it penalizes—and what falls outside that measure—helps you interpret training without mistaking it for a complete evaluation of the model.
LoRA: How It Adapts a Model with Low-Rank Matrices—and What It Doesn’t Guarantee
LoRA adapts a pretrained model by keeping its original weights fixed and learning additional low-rank matrices. This guide explains the mechanism, compares it with other forms of adaptation, and clarifies what cannot be concluded from the number of trainable parameters alone.
Attention Mechanism: How a Model Relates Different Parts of an Input
Attention is a mathematical operation that combines representations according to their compatibility. This guide explains Q, K, and V, the mechanism’s variants, and its limits—without confusing computational attention with human attention or treating it as an automatic explanation of a model.
Distilled Model: What the Label Means—and What It Does Not Prove
A distilled model is the student produced by a process that transfers some of a teacher model’s behavior. The label describes a training relationship, not a guarantee of equivalence, quality, or efficiency.
Foundation model
Model trained at large scale on broad data and adaptable to multiple downstream tasks.
Pretraining: What a Model Learns Before It Is Adapted
Pretraining is a training phase that comes before certain uses or later adaptations. Its objective, data, and relationship to other stages depend on the model: there is no universal recipe, and pretraining alone does not guarantee that a system is reliable.
Overfitting: Definition, Detection Mechanisms, and Limits of Generalization
Overfitting occurs when a machine learning model memorizes noise and specific patterns in training data, achieving high accuracy on that data but failing to generalize to new inputs. This technical guide explains the phenomenon through the bias-variance tradeoff, details detection methods using loss curves and cross-validation, and contrasts overfitting with underfitting and data leakage. It includes applied examples in fraud detection, medical diagnosis, and logistics, along with a mitigation protocol based on regularization, early stopping, and data augmentation.
Transformer
Neural architecture based primarily on attention that processes relationships between elements of a sequence or multiple modalities.
Generation and interaction
Concepts involved in requesting, controlling, and validating a generated response.
Few-Shot Learning: Definition, Mechanism, and Limits of In-Context Learning
Few-shot prompting is a context engineering technique that includes input-output pairs in the request to guide the model's response without updating its weights. This entry explains the in-context learning mechanism, provides three applied examples in different domains, distinguishes the technique from zero-shot, fine-tuning, and chain of thought, and outlines its operational constraints.
In-Context Learning (ICL): How a Model Responds to Prompt Examples
In-context learning lets you guide a model’s response with instructions and examples included in its input. Learn what it means, how to test it, and how it differs from training and related methods, with examples from support, invoices, and customer reviews.
Chain of Thought in AI: What It Is—and What It Doesn’t Prove
A chain of thought is a sequence of intermediate steps expressed in text that a model generates or receives as an example. It can help structure a task and make some steps visible, but coherence does not prove that those steps are correct or faithfully reflect the internal process that produced the answer.
Context Engineering: How to Select and Maintain the Information an AI System Receives
Context engineering means deciding what information a model needs for a task, how to obtain it, and how to keep it relevant while the task is underway. It is not the same as writing a longer prompt: it also involves selecting documents, using tools, and managing state.
Prompt
Set of instructions and context provided to a model to guide a task and its output format.
Reasoning in Models: How It Works, Examples, and Limits
A rigorous definition of reasoning capabilities in language models, a step-by-step explanation of their operation, three applied examples, evaluation methods, and an analysis of documented limitations from primary literature.
Structured Output: Definition, Coercion Mechanisms, and Limits of Automatic Validation
Structured output constrains LLM generation to a formal schema (JSON, XML) to enable programmatic consumption. This entry distinguishes between prompt-based suggestion and technical coercion via constrained decoding, analyzes the limits of syntactic validation versus semantic correctness, and provides applied examples in data extraction, agent orchestration, and configuration generation.
Temperature: Definition, Mechanics, and Limits of the Sampling Parameter in Language Models
Temperature is an inference parameter that modifies the probability distribution of logits to control output randomness. This entry explains its mathematical mechanism, its interaction with top-k and top-p, and its limits for reproducibility and evaluation.
Data and knowledge
How data is represented, retrieved, and connected to verifiable answers.
Vector Database: What It Stores and How Similarity Search Works
A vector database stores numerical representations and retrieves those closest to a query. It does not necessarily generate embeddings or decide on its own which results are relevant.
Semantic search
Information retrieval based on meaning relationships, usually combined with filters or lexical search.
Embedding
Numerical representation that places elements in a space where certain relationships can be measured by proximity.
Document Chunking: How It Works, Strategies, and Limits
Document chunking splits content into retrievable units for AI systems. This guide explains the process, key strategies (fixed, semantic, hierarchical), size and overlap parameters, and how to evaluate results without falling into common pitfalls.
Primary Source: What It Is, How to Identify It, and What It Can Show
A source is primary in relation to a question when it presents directly the object, data, method, or decision being studied. Its status does not guarantee that it is true, impartial, or sufficient: you still need to assess what it claims and what its limits are.
Retrieval-augmented generation (RAG)
Pattern that retrieves relevant external information and incorporates it into the context before generating a response.
Grounding: Definition, Mechanics, and Limits in AI
Grounding connects an AI model's output to verifiable external evidence. This guide explains how it works, how it differs from RAG and fine-tuning, and what its practical limitations are.
Reranking: What It Is, How It Works, and Its Limits in Retrieval Systems
Reranking is a second scoring stage in retrieval systems that evaluates query–document pairs to refine an initial set of candidates. This glossary entry explains how it works, distinguishes two-stage retrieval from vector search alone, and examines operational limits in RAG and semantic search.
Agents and automation
Tools, memory, and patterns for systems capable of carrying out tasks.
AI Agent: What It Is, How It Works, and How It Differs from a Chatbot
An AI agent is a system that pursues a goal by observing its environment and choosing actions. This guide explains its operating cycle, its limits, and how it differs from a chatbot or a fixed automation.
Tool calling
Mechanism by which a model selects an operation and proposes structured arguments for the application to execute.
Agent memory
Mechanism to preserve and retrieve relevant state between steps or sessions of a model-based system.
Model Context Protocol (MCP): What It Connects, How It Works, and What It Doesn’t Guarantee
MCP is an open integration standard for connecting AI applications to external systems, data, and tools. This guide explains what can be established from the available sources, offers conditional examples, and distinguishes the protocol from an API, tool calling, and an agent.
Assessment
Methods for measuring quality, cost, robustness, and usefulness under reproducible conditions.
AI benchmarks: what they measure and what a score demonstrates
A benchmark is an evaluation defined by tasks, data, a protocol, and metrics. Its score describes performance under those specific conditions; by itself, it proves neither general capability nor production performance.
Data contamination
Total or partial presence of evaluation data in the training, tuning, or prior context of a model.
AI evaluation
Reproducible test that measures a behavior through cases, conditions, metrics, and an explicit rubric.
Generalization in machine learning: definition, mechanics, and limits
Generalization is a model's ability to keep useful behavior on examples that do not exactly match the training data. This entry explains how it is measured, why independence assumptions matter, and what empirical studies show about its limits.
Uncertainty: Definition, Types, and Quantification Methods in Predictive Models
Uncertainty measures the degree of lack of certainty regarding a prediction or measurement. This entry explains the fundamental distinction between aleatoric (inherent noise) and epistemic (lack of knowledge) uncertainty, details quantification methods such as Monte Carlo and Gaussian processes, and provides applied examples in meteorology, engineering, and social sciences to assess model reliability.
LLM-as-a-judge: what it evaluates and which biases to check
An LLM-as-a-judge approach uses a language model to assess responses or actions against explicit criteria. This guide explains the process, distinguishes different types of judgment, and shows what to verify before relying on the results.
Reproducibility
Ability to repeat an evaluation with documented versions, data, configuration, and procedure and obtain consistent results.
Robustness: Definition, Functionality, and Limits in Machine Learning Systems
Robustness is not an absolute property but a relationship between a target, a modifier, a variation domain, and a tolerance. This entry explains how to operationalize the concept, how to audit models for distribution shifts, and why high performance in laboratory conditions does not guarantee reliable behavior in the real world.
Assessment Rubrics: Definition, Design, and Limits When Scoring AI Outputs
A rubric organizes criteria and performance levels to make evaluations easier to interpret. This guide explains its components, distinguishes analytic from holistic rubrics, and examines how rubrics can be used—and where they fall short—in evaluating AI outputs, academic work, and machine-learning systems.
Safety and governance
Risks, controls, and responsibilities throughout the lifecycle.
AI Hallucinations: What They Mean and How to Recognize an Unsupported Answer
A fluent answer is not necessarily a correct one. This guide explains what AI hallucination means, how to distinguish it from other errors, and what checks can help assess specific claims.
AI Guardrails: What They Control, Where They Act, and What They Don’t Guarantee
A guardrail is a control that checks a system’s inputs, outputs, or actions against a defined policy. Its scope depends on what it reviews and when it intervenes: it can reduce certain risks, but it cannot by itself prove that the entire system is safe.
Human-in-the-Loop (HITL): What a Person Decides—and What Their Involvement Does Not Guarantee
HITL describes a design in which a person intervenes at a defined point in an automated system. The label alone does not say what authority they have or guarantee that they will detect errors.
Instruction injection
Attack or interference in which untrusted content attempts to divert the system from its rules or legitimate objective.
Moderation: Definition, Mechanics, and Limits in Digital Platforms
Moderation comprises the automated and human actions platforms use to detect, identify, and address content that violates their terms or the law. This glossary explains its operation, regulatory frameworks like the DSA, and the technical and ethical challenges posed by automation.
The Principle of Least Privilege: What It Means and How to Apply It to AI Systems
The principle of least privilege limits each identity—person, process, or service—to the permissions needed for its task. In AI systems, it guides access for agents, tools, and services, but does not replace authentication or, by itself, guarantee that a model behaves safely.
Adversarial Testing: Definition, Methodology, and Limits in AI Systems
Adversarial testing in AI is an intentional evaluation discipline that seeks to break a system's guarantees through hostile inputs or designed scenarios. This entry defines the concept, distinguishes manual practice from automation, details the technical lifecycle, and analyzes its inherent limits regarding coverage and reproducibility.
Human Oversight: Definition, Control Mechanisms, and Operational Limits in AI Systems
Human oversight is not merely the presence of an operator, but the effective capacity to understand, monitor, intervene in, and stop an AI system. This entry distinguishes autonomy levels, explains the technical and organizational mechanisms required by the AI Act, and delineates the cognitive and operational limits where oversight becomes ineffective.
Fact-checking
Process of checking verifiable claims against identifiable, current, and sufficiently authorized evidence.
Operations and infrastructure
Techniques for deploying, optimising, and observing AI systems in production.
Context Caching in AI: What It Reuses—and Why It Doesn’t Mean “Remembering”
A cache can preserve model computations so they do not have to be repeated. This guide distinguishes KV caching from prefix caching, explains when each can be reused, and clarifies their limits.
Quantization in AI: What Changes When Numerical Precision Is Reduced
Quantization represents numbers using fewer levels to reduce resource use, but “4-bit” alone is not enough to predict a model’s memory use, speed, or quality.
AI Inference: What Happens When a Model Processes an Input
Inference is the use of an already-trained model to calculate an output from an input. It can produce a class, score, representation, or generated content; it does not guarantee that the result is correct.
Latency in AI: What Time Is Measured and Why the Response Milestone Matters
Latency is the time between a defined start and a defined end. Learn to distinguish the first result, the complete response, and the time experienced by someone using an AI application.
Processing performance
Amount of work a system completes per unit of time, for example requests or tokens per second.
Try another word, an alias, or a different area.