AI Dictionary

AI dictionary: from basics to advanced terms

An AI glossary for business teams - from fundamentals to agents, RAG, safety, and governance.

This dictionary explains the terms you'll encounter when working with AI tools, models, and agents. It's organized into 14 sections so you can quickly find what you need.

1. Basic concepts

AI / Artificial intelligence: The field of computer science that builds systems capable of performing tasks usually requiring human intelligence - language understanding, image recognition, decision-making, and learning from data.
Machine Learning / ML: A branch of AI in which a system learns from data instead of every rule being hand-coded.
Deep Learning: A type of machine learning that uses neural networks with many layers to recognize complex patterns.
Model: An AI system trained on data that can produce predictions, answers, or decisions.
Algorithm: A set of rules or steps a computer follows to solve a problem.
Data: Information used to train, test, or operate an AI system.
Dataset: An organized collection of data used for training or evaluating a model.
Model training: The process in which a model learns from data by adjusting its internal parameters.
Inference: Using an already-trained model to generate an answer, prediction, or decision.
Prediction: The output a model computes from input data.
Input: The data a user or system sends to an AI model.
Output: The result the AI model returns.
Automation: Using technology to perform tasks without continuous human work.
Intelligent automation: Automation that uses AI for decision-making, language processing, or adapting to context.
Generative AI: AI that creates new content - text, images, audio, video, or code.
Classical AI: AI systems based on rules, logic, classification, or prediction - without necessarily creating new content.
Assistant: An AI system that helps a user with tasks such as writing, analysis, search, or planning.
Chatbot: A program that talks with a user via text or voice.
Virtual assistant: A more advanced chatbot that understands context, executes tasks, and uses tools.

2. Generative AI and language models

LLM / Large language model: An AI model trained on massive amounts of text, capable of understanding and generating language. Examples: ChatGPT, Claude, Gemini, Llama.
Language model: A model that predicts the next word, token, or piece of text based on prior context.
Token: The basic unit of text a model processes. Can be a word, part of a word, character, or whitespace.
Tokenization: The process of converting text into tokens the model can process.
Context: The information the model currently has available in a conversation or task.
Context window: The maximum amount of text, tokens, or information a model can consider in a single request.
Prompt: An instruction, question, or text a user sends to an AI model.
Prompt engineering: The skill of writing high-quality instructions for a model to get better results.
System prompt: The main instruction that shapes the AI assistant's behavior.
User prompt: The message or instruction the user sends to the model.
Few-shot prompting: Giving the model a few examples so it better understands the expected style or format of the answer.
Zero-shot prompting: Asking the model to perform a task without giving any examples.
Chain-of-thought: A method in which the model breaks a problem into steps internally or explicitly.
Reasoning model: A model optimized for more complex reasoning, planning, analysis, and problem-solving.
Hallucination: When AI generates incorrect, fabricated, or unverified information that sounds convincing.
Temperature: A setting that controls model creativity. Lower temperature gives more predictable answers, higher gives more creative ones.
Top-p: A setting that limits the choice of possible next tokens by probability.
Max tokens: The maximum number of tokens the model can generate in a response.
Completion: The text the model generates as a continuation of, or answer to, a prompt.
Embeddings: Numerical representations of the meaning of text, images, or other content. Used for search, similarity, and semantic analysis.
Semantic search: Searching by meaning rather than just exact keywords.
Multimodal model: An AI model that can work with multiple types of data - text, images, audio, video, and code.
Vision model: A model that can analyze images, documents, charts, or video.
Speech-to-text: Converting speech into text.
Text-to-speech: Converting text into speech.
Image generation: Generating images using AI.
Video generation: Generating video using AI.
Code generation: Generating program code using AI models.

3. Data and learning

Training set: Data the model learns from.
Validation set: Data used during development to check the model's performance.
Test set: Data used for the final evaluation of the model after training.
Label: The correct answer or category attached to a data point during training.
Labeled data: Data that has correct answers attached to it.
Unlabeled data: Data without predefined answers.
Supervised learning: The model learns from data that includes correct answers.
Unsupervised learning: The model finds patterns in data without any predefined answers.
Semi-supervised learning: A combination of a small amount of labeled data and a large amount of unlabeled data.
Reinforcement learning: The model learns through rewards and penalties tied to its actions.
Reward: A signal that tells the model whether a decision was good.
Agent: A system that observes its environment, decides, and executes actions toward a goal.
Policy: The strategy by which an agent decides which action to take.
Feature: A measurable characteristic of a data point that the model uses to learn.
Feature engineering: Manually creating or selecting features that help the model learn better.
Normalization: Adjusting data values onto a common scale.
Data cleaning: Removing errors, duplicates, empty values, and inconsistencies from data.
Data augmentation: Artificially expanding a dataset by creating variations of existing examples.
Synthetic data: Artificially generated data that mimics real data.
Bias in data: Imbalance or distortion in data that can lead to unfair model results.
Noise: Random, irrelevant, or erroneous variation in data that obscures the true signal the model is trying to learn.
Ground truth: The reference truth - the correct answer used to evaluate the model.

4. Neural networks and deep learning

Neural network: A model inspired by how the brain works, composed of layers of artificial neurons.
Neuron: The basic unit of a neural network that takes inputs, processes them, and produces an output.
Layer: A group of neurons in a neural network.
Input layer: The first layer that receives data.
Hidden layer: A layer between input and output that learns patterns.
Output layer: The final layer that produces the result.
Weights: Numbers inside the model that determine how important a given input is.
Bias: An extra parameter that helps the model adjust its calculations.
Parameters: The internal values of a model that are learned during training.
Hyperparameters: Settings a person or system chooses before training - e.g., learning rate or batch size.
Activation function: A function applied to a neuron's output that introduces non-linearity by transforming the value (e.g. squashing it into a range or zeroing out negatives), letting the network learn complex patterns.
ReLU: A popular activation function that turns negative values into zero.
Sigmoid: A function that maps a value into the range between 0 and 1.
Softmax: A function that turns model outputs into class probabilities.
Loss function: A measure of how far the model's prediction is from the correct answer.
Gradient descent: An optimization method by which the model gradually reduces error.
Backpropagation: The algorithm by which error is propagated backward through the network to adjust weights.
Epoch: One pass of the model through the entire training set.
Batch: A smaller set of data processed in a single training step.
Batch size: The number of examples in one batch.
Learning rate: A value that determines how fast the model adjusts during training.
Overfitting: When the model learns the training data too well and generalizes poorly to new data.
Underfitting: When the model has not learned enough patterns from the data.
Regularization: Techniques for reducing overfitting.
Dropout: A technique where some neurons are temporarily switched off during training so the model generalizes better.
Fine-tuning: Adapting an existing model to specific data or a specific task.
Transfer learning: Using knowledge from one model or task for another task.

5. Transformer architecture and modern LLMs

Transformer: A neural network architecture that powers modern large language models.
Attention: A mechanism by which the model decides which parts of the input text matter for the current answer.
Self-attention: A mechanism by which the model compares different parts of the same text to one another.
Multi-head attention: Multiple attention mechanisms that look at different relationships in the text in parallel.
Encoder: The part of the model that processes and understands input.
Decoder: The part of the model that generates output.
Encoder-decoder model: An architecture that uses an encoder to understand input and a decoder to generate output.
Decoder-only model: A model that primarily generates text token by token. Many modern LLMs use this approach.
Positional encoding: How the model gets information about the order of tokens.
Embedding layer: The layer that turns tokens into numerical vectors.
Pretraining: The first large training of a model on huge amounts of general data.
Instruction tuning: Training a model to follow human instructions better.
RLHF / Reinforcement Learning from Human Feedback: Reinforcement learning from human feedback. Used to align models with human preferences.
Alignment: The process of making a model more helpful, safer, and more aligned with human intent.
Distillation: Creating a smaller model that learns the behavior of a larger one.
Quantization: Reducing the precision of numbers in a model to use less memory and run faster.
MoE / Mixture of Experts: An architecture that activates only certain experts (sub-models) depending on the task.
Sparse model: A model in which only part of the parameters is used for a given input.
Dense model: A model in which most or all parameters are used for every input.

6. AI agents and automation

AI agent: An AI system that can plan, use tools, make decisions, and execute steps toward a goal.
Agentic AI: An AI approach in which the model doesn't just respond, but autonomously organizes actions and uses tools.
Tool calling: The model's ability to use external functions, APIs, databases, or services.
Function calling: A structured way for the model to call predefined functions.
Planner: The component of an agent that breaks a goal into steps.
Executor: The component that runs the planned steps.
Memory: The mechanism by which an agent remembers information from previous interactions or tasks.
Short-term memory: Information available during the current task or conversation.
Long-term memory: Information stored for future use.
Reflection: A process in which the agent analyzes its own results and tries to improve them.
Self-correction: The model or agent's ability to recognize and fix its own error.
Task decomposition: Breaking a complex task into smaller steps.
Workflow: A defined sequence of steps in a business or technical process.
Orchestration: Coordinating multiple models, tools, APIs, or agents in a single system.
Multi-agent system: A system where multiple AI agents collaborate or specialize in different tasks.
Human-in-the-loop: A process where a human supervises, confirms, or corrects AI decisions.
Autonomous agent: An agent that can carry out tasks with minimal human intervention.
Guardrails: Rules and limits that prevent undesirable behavior in an AI system.
Approval step: A point in the process where a human must confirm continuation.
Action space: The set of all actions an agent can take.
Environment: The system, application, or context in which the agent operates.

7. RAG, knowledge bases, and search

RAG / Retrieval-Augmented Generation: A technique in which AI first retrieves relevant information from a knowledge base, then generates an answer based on it.
Retrieval: The process of finding relevant documents, passages, or data.
Retriever: The component that searches for relevant information in a database or document set.
Generator: The model that produces an answer based on the retrieved information.
Vector database: A database that stores embeddings and enables search by meaning.
Vector search: Searching data based on the proximity of their embeddings.
Similarity search: Finding content semantically similar to the query.
Chunking: Breaking large documents into smaller segments for easier search.
Chunk: A smaller part of a document that is stored and searched.
Chunk size: The size of a single text segment.
Chunk overlap: Overlap between segments so that context isn't lost.
Re-ranking: Additional sorting of retrieved results by relevance.
Hybrid search: A combination of classic keyword search and vector search.
BM25: A classic algorithm for keyword-based text search.
Knowledge base: An organized collection of documents, data, and information that AI can use.
Citation: Naming the source from which AI obtained the information.
Source grounding: The practice of tying AI answers to verifiable documents or data.
Context injection: Adding relevant information to the prompt before generating an answer.
Semantic similarity: A measure of how similar two pieces of text are in meaning.

8. Evaluation and metrics

Evaluation: The process of measuring the quality of a model or AI system.
Benchmark: A standardized test for comparing models.
Accuracy: The share of correct predictions out of all predictions.
Precision: How many positive predictions are actually correct.
Recall: How many of the actual positive cases the model found.
F1 score: A combined measure of precision and recall.
Confusion matrix: A table that shows the model's correct and incorrect classifications.
True positive: The model correctly predicted a positive case.
False positive: The model incorrectly predicted a positive case.
True negative: The model correctly predicted a negative case.
False negative: The model incorrectly predicted a negative case.
Latency: The time the AI system takes to return a response.
Throughput: The number of requests the system can process in a given time.
Cost per request: The cost of a single AI call or query.
Token usage: The number of tokens spent on the model's input and output.
Hallucination rate: The share of answers that contain incorrect or fabricated information.
Groundedness: A measure of how tied an answer is to real sources or data.
Faithfulness: A measure of how faithfully an answer conveys the information from its source.
Robustness: The model's ability to perform well under different conditions and on unexpected inputs.
Generalization: The model's ability to perform well on new, unseen data.
A/B testing: Comparing two versions of a system to see which performs better.

9. Safety, ethics, and risks

AI safety: The field concerned with preventing harmful, unpredictable, or undesirable behavior in AI systems.
AI ethics: The field that studies the moral, social, and legal consequences of using AI.
Bias: Systematic non-objectivity in a model due to data, design, or how it's used.
Fairness: The goal of ensuring an AI system doesn't discriminate against certain user groups.
Transparency: Clarity about how an AI system works and how it makes decisions.
Explainability: The ability to explain why the model produced a particular result.
Interpretability: A measure of how much a person can understand the model's internal workings.
Privacy: Protecting personal and sensitive data in AI systems.
Data leakage: In machine learning, when information from outside the training set (e.g. test data or the target variable) leaks into training, inflating measured performance while the model actually generalizes poorly; more broadly, the unintended disclosure of private or confidential information.
Prompt injection: An attack in which a user or document tries to alter the AI model's behavior with malicious instructions.
Jailbreak: An attempt to bypass the AI model's safety constraints.
Adversarial attack: A deliberately crafted input that tries to fool the model.
Model misuse: Using AI for harmful purposes such as fraud, spam, or manipulation.
Deepfake: AI-generated or manipulated audio, video, or image that falsely depicts a person or event.
Content moderation: Automatic or manual filtering of harmful, illegal, or inappropriate content.
PII / Personally identifiable information: Data that can identify a person - e.g., name, address, ID number, email, or phone number.
Data retention: Rules about how long data is kept.
Audit log: A record of activity in a system for monitoring, security, and compliance.
Compliance: Adherence to laws, standards, and internal rules.
GDPR: The European regulation on the protection of personal data.

10. Business application of AI

AI transformation: Introducing AI into the processes, products, and business model of an organization.
AI adoption: The process by which an organization starts using AI tools and systems.
AI readiness: An assessment of how technically, organizationally, and data-wise an organization is prepared for AI.
Use case: A concrete business problem or task in which AI can create value.
ROI / Return on investment: A measure of the benefit a business gets relative to the money invested.
Productivity gain: An increase in work efficiency from using AI.
Cost reduction: Using AI to reduce operational costs.
Decision support: An AI system that helps people make better decisions by analyzing data.
Customer support automation: Automating customer support using chatbots, agents, or knowledge bases.
Document automation: Automatic processing, analysis, generation, or classification of documents.
Lead scoring: Estimating the quality of a potential customer using data and AI.
Personalization: Adapting content, offers, or experiences to the individual user.
Recommendation system: An AI system that suggests products, content, or actions.
Predictive analytics: Using data to predict future events or behavior.
Churn prediction: Predicting which customers might stop using a product or service.
Fraud detection: Using AI to identify suspicious activity.
AI governance: Rules, processes, and responsibilities for safe and effective use of AI in an organization.
Model governance: Control of versions, evaluation, security, and use of AI models.
AI policy: An internal policy that defines how employees may use AI.
Shadow AI: Informal or unauthorized use of AI tools inside an organization.

11. Development, integrations, and infrastructure

API: An interface through which applications communicate with other systems or AI models.
AI API: An API that lets you send queries to an AI model and receive responses.
SDK: A set of tools and libraries for easier integration of AI into applications.
Endpoint: A URL or access point to an API.
Request: A message an application sends to an API.
Response: A message an API returns to the application.
Streaming: Sending the response in chunks as it's generated, rather than waiting for the whole response.
Rate limit: A limit on the number of requests in a given time window.
Caching: Storing results for faster and cheaper reuse.
Queue: A system for organized execution of tasks, one at a time or in parallel.
Worker: A process that handles tasks from a queue.
Webhook: A mechanism by which one system automatically notifies another about an event.
Serverless: An architecture where code runs without direct server management.
Edge computing: Running code closer to the user to reduce latency.
GPU: A graphics processor often used for training and running AI models.
TPU: A specialized processor for machine learning.
VRAM: Graphics card memory - important for running large models.
Model hosting: Running an AI model on a server or cloud platform.
Self-hosted model: A model an organization runs on its own infrastructure.
Cloud model: A model used via a cloud service provider.
On-premise AI: An AI system that runs inside an organization's own infrastructure.
MLOps: Practices for developing, deploying, monitoring, and maintaining ML models in production.
LLMOps: Practices for developing, evaluating, monitoring, and maintaining applications based on large language models.
Monitoring: Tracking performance, costs, errors, and quality of AI systems.
Observability: Deep insight into a system's operation through logs, metrics, and execution traces.
Logging: Storing system events and activity.

12. Advanced concepts

Foundation model: A large model trained on a broad data set that can be adapted to many tasks.
General-purpose model: A general-purpose model that can handle a wide range of tasks.
Domain-specific model: A model specialized for a particular industry or domain - e.g., law, medicine, or finance.
Small language model / SLM: A smaller language model optimized for speed, cost, or local deployment.
Open-source model: A model whose code, weights, or documentation are publicly available under a specific license.
Closed-source model: A model that users access via an API but whose internal weights or code aren't visible.
Model weights: The learned numbers that represent the model's knowledge.
Checkpoint: A saved state of the model at a specific point during training.
LoRA / Low-Rank Adaptation: An efficient fine-tuning technique that uses a smaller number of additional parameters.
Adapter: A small additional module that adapts a large model to a specific task.
PEFT / Parameter-Efficient Fine-Tuning: A set of methods for fine-tuning without changing all of the model's parameters.
Context engineering: The design and management of the information given to a model so the result is accurate and useful.
Prompt chaining: Linking multiple prompts in a sequence of steps.
Recursive prompting: A technique where the result of one prompt is used as input for the next.
Tree of Thoughts: An approach in which the model explores multiple possible reasoning paths.
Graph of Thoughts: A more advanced approach where ideas, steps, or conclusions are organized as a graph.
Toolformer-style model: A model that has been trained or tuned to know when and how to use tools.
Neural symbolic AI: A combination of neural networks and symbolic reasoning.
Knowledge graph: A structure that represents entities and their relationships.
Ontology: A formal description of concepts and relations in a domain.
Causal AI: AI that tries to understand cause-and-effect relationships, not just correlations.
Continual learning: A model's ability to keep learning over time without full retraining.
Catastrophic forgetting: A problem where the model forgets old knowledge after learning something new.
Federated learning: Training a model across multiple devices or locations without centrally collecting data.
Differential privacy: A privacy-protection technique that adds controlled noise to data or results.
Model compression: Reducing model size while losing as little quality as possible.
Pruning: Removing less important parts of the model to speed it up or reduce its size.
Speculative decoding: A technique to speed up text generation by using a smaller model to propose tokens.
KV cache: Memory that stores prior transformer computations for faster generation.
Inference optimization: Optimizing speed, cost, and memory usage when running a model.
Model drift: A change in model quality over time as real-world data or user behavior shifts.
Data drift: A change in the distribution of input data compared to the data the model was trained on.
Concept drift: A change in the relationship between input data and the correct answers over time.

13. The key difference: classical AI, generative AI, and agents

Type	What it does	Example
Classical AI	Classifies, predicts, recognizes patterns, or makes decisions based on data.	Fraud detection, product recommendation, credit scoring.
Generative AI	Creates new content.	Writing text, generating an image, producing code.
AI agent	Plans, uses tools, and executes tasks across multiple steps.	An agent that reads an email, analyzes a document, updates the CRM, and sends a draft response.

14. Quick glossary for presentations

AI: A computer that can perform tasks that look intelligent.
Machine Learning: AI that learns from examples.
Deep Learning: Advanced ML that uses layers of neural networks.
LLM: An AI model that understands and writes text.
Prompt: An instruction we give to AI.
Token: A small piece of text that AI processes.
Context: Everything the AI is currently looking at.
Hallucination: When AI invents an incorrect answer.
RAG: AI that searches documents before answering.
Embedding: A numerical representation of text meaning.
Agent: AI that doesn't just respond, but also executes tasks.
Tool calling: When AI uses an external tool or API.
Fine-tuning: Additional training of a model for a specific task.
Inference: Using the model after training.
AI governance: Rules for safe use of AI in a company.

AI dictionary: from basics to advanced terms

Need AI education for your team?