AI Glossary

A

• Agent — an AI system that can reason, choose actions/tools, and act toward a goal in a loop, rather than just answering a single prompt. See AI Agent Tutorials.
• AGI (Artificial General Intelligence) — a hypothetical AI with human-level ability across virtually any intellectual task, as opposed to today's models which excel at specific tasks.
• Alignment — the practice of training a model to behave in ways that match human intent and values (helpful, honest, harmless).
• API (Application Programming Interface) — the interface through which software (including AI models) is accessed programmatically.
• Attention Mechanism — the core building block of the Transformer architecture; lets a model weigh how relevant every other token is when processing a given token.
• Autoencoder — a neural network trained to compress data into a smaller representation and then reconstruct it, often used for anomaly detection.

B

• Backpropagation — the algorithm used to train neural networks, propagating the error backward through the layers to update weights.
• Batch Size — the number of training examples processed together before the model's weights are updated once.
• Benchmark — a standardized test used to measure and compare model performance on a specific capability.
• Bias (model bias) — systematic skew in a model's outputs, usually inherited from patterns in its training data.

C

• Chain-of-Thought (CoT) — prompting or training a model to reason step-by-step before giving a final answer, which tends to improve accuracy on complex tasks.
• Checkpoint — a saved snapshot of a model's weights at a point in training.
• Classification — predicting which category an input belongs to (e.g. "malicious" vs "benign").
• Clustering — grouping similar data points together without predefined labels (unsupervised learning).
• Context Window — the maximum amount of text (measured in tokens) a model can consider at once, including the prompt, conversation history, and any retrieved documents.
• Corpus — a large collection of text used for training or analysis.
• Cosine Similarity — a common way to measure how similar two vector embeddings are, used heavily in semantic search and RAG.

D

• Dataset — a structured collection of data used to train, validate, or test a model.
• Deep Learning — machine learning using multi-layer ("deep") neural networks.
• Diffusion Model — a generative model (commonly used for images) that learns to reverse a gradual noising process to produce new content.
• Distillation (Knowledge Distillation) — training a smaller "student" model to mimic a larger "teacher" model, producing a faster, cheaper model with much of the original's capability.

E

• Embedding — a numeric vector representation of text (or images, audio, etc.) that captures semantic meaning, allowing "similar" content to have mathematically similar vectors.
• Ensemble Learning — combining multiple models to get better predictions than any single model alone.
• Epoch — one full pass through the entire training dataset.
• Explainability (XAI) — techniques for making a model's decisions understandable to humans.

F

• Feature — an individual measurable input variable used by a model (e.g. "packet size" in a network traffic model).
• Few-shot Learning — giving a model a handful of examples in the prompt so it can perform a task without additional training.
• Fine-tuning — further training a pre-trained model on a smaller, specific dataset to specialize its behavior.
• Foundation Model — a large, general-purpose model (like an LLM) trained on broad data, meant to be adapted to many downstream tasks.
• Function Calling / Tool Calling — a model's ability to invoke external functions or APIs as part of generating a response.

G

• GAN (Generative Adversarial Network) — two neural networks (a generator and a discriminator) trained against each other to produce realistic synthetic data.
• Generative AI — AI systems that create new content (text, images, audio, code) rather than just classifying or predicting from fixed categories.
• GPU (Graphics Processing Unit) — specialized hardware that dramatically speeds up the parallel math used in training and running neural networks.
• Gradient Descent — the optimization algorithm used to adjust a model's weights to minimize its loss (error).
• Ground Truth — the verified, correct answer used to measure or train a model.
• Guardrails — rules, filters, or checks placed around a model to constrain its outputs and behavior.

H

• Hallucination — when a model generates fluent but factually incorrect or fabricated content.
• Hidden Layer — any neural network layer between the input and output layers.
• Hyperparameter — a training configuration value set before training begins (e.g. learning rate), as opposed to a weight the model learns.

I

• Inference — running a trained model on new input to produce an output; this is what happens every time a model responds to a prompt.
• Instruction Tuning — fine-tuning a model specifically to follow natural-language instructions well.
• Intent Recognition — identifying what a user is trying to accomplish from their input text.

J

• Jailbreak — a prompt or technique designed to bypass a model's safety training and get it to produce restricted content. See AI Security.

K

• Knowledge Base — a structured or unstructured collection of information a system can draw on, often the source data for a RAG pipeline.
• Knowledge Graph — a network of entities and their relationships, used to represent structured knowledge.
• k-NN (k-Nearest Neighbors) — a method that classifies or retrieves items based on the "k" most similar items in a dataset.

L

• Latency — the time it takes a model to produce a response.
• LLM (Large Language Model) — a deep learning model trained on massive text datasets, capable of understanding and generating human language.
• LoRA (Low-Rank Adaptation) — an efficient fine-tuning technique that trains a small number of additional parameters instead of the whole model.
• Loss Function — the formula used to measure how wrong a model's predictions are during training; training aims to minimize it.

M

• MCP (Model Context Protocol) — an open protocol standardizing how AI applications connect to external tools and data sources. See MCP Explained.
• Model Card — documentation describing a model's intended use, training data, and known limitations.
• Model Weights — the millions/billions of learned numeric parameters that define a trained model's behavior.
• Multimodal — a model that can process more than one type of input/output, e.g. text and images.

N

• NLP (Natural Language Processing) — the broader field concerned with computers understanding and generating human language.
• Neural Network — a model made of layered, interconnected nodes loosely inspired by biological neurons.

O

• One-shot Learning — learning a task from a single example.
• Overfitting — when a model learns the training data too specifically (including its noise) and performs poorly on new, unseen data.

P

• Parameter — an individual learned value (weight) inside a model; model size is often described by parameter count (e.g. "a 70-billion-parameter model").
• Perplexity — a measure of how well a language model predicts a sample of text; lower is generally better.
• Pre-training — the initial, large-scale training phase where a model learns general patterns from broad data.
• Prompt — the input text given to a model to elicit a response.
• Prompt Engineering — the practice of crafting inputs to reliably get useful, accurate output from a model.
• Prompt Injection — an attack that manipulates a model via crafted input (direct or embedded in retrieved content) to override its intended instructions. See AI Security.

Q

• Quantization — reducing the numeric precision of a model's weights (e.g. 16-bit to 4-bit) to shrink its size and speed it up, usually with a small accuracy trade-off. Central to running models locally — see Ollama & Local AI Setup.

R

• RAG (Retrieval-Augmented Generation) — retrieving relevant external information at query time and feeding it to a model as context, so it can answer using current or private data. See RAG Tutorials.
• Reasoning Model — a model specifically trained to perform extended, step-by-step reasoning before answering, often improving accuracy on complex problems.
• Regularization — techniques used during training to reduce overfitting.
• Reinforcement Learning (RL) — training a model through trial and error using rewards and penalties.
• RLHF (Reinforcement Learning from Human Feedback) — using human preference ratings as the reward signal to align a model's behavior.

S

• Self-Supervised Learning — a model generates its own training labels from raw data (e.g. predicting the next word), avoiding the need for manual labeling.
• Semantic Search — searching by meaning (via embeddings) rather than exact keyword matching.
• Supervised Learning — training a model on labeled input/output pairs.
• System Prompt — a special instruction set given to a model before the user's conversation begins, shaping its behavior and constraints.

T

• Temperature — a setting that controls how random or deterministic a model's output is; higher values produce more varied (and sometimes less reliable) responses.
• Token — the basic unit of text a model processes — roughly a word or part of a word.
• Tokenization — the process of splitting text into tokens before feeding it to a model.
• Training Data — the dataset used to teach a model.
• Transfer Learning — reusing a model trained on one task as the starting point for a related task.
• Transformer — the neural network architecture (introduced in 2017) using attention mechanisms, underlying virtually all modern LLMs.

U

• Underfitting — when a model is too simple to capture the patterns in the data, performing poorly even on training data.
• Unsupervised Learning — training a model on data with no labels, so it must find structure on its own.

V

• Validation Set — a portion of data held back from training, used to check a model's performance on unseen examples during development.
• Vector Database — a database optimized for storing and searching embeddings by similarity; a core component of RAG systems.

W

• Weak Supervision — training with noisy, indirect, or programmatically generated labels instead of fully hand-labeled data.

Z

• Zero-shot Learning — a model performing a task correctly without having seen any examples of that specific task before.