AI Glossary
Plain-English definitions of the AI and machine learning terms that actually come up, from RLHF and transformers to RAG and agents. No jargon for its own sake.
141 terms
Foundations
Artificial General Intelligence
AGIA hypothetical AI that can learn and reason across any task about as well as a human, rather than being good at one narrow thing.
Artificial Intelligence
AISoftware that performs tasks we normally associate with human intelligence, like understanding language, recognizing images, or making decisions.
Deep Learning
Machine learning built on large neural networks with many layers, which is what powers most recent AI breakthroughs.
Generative AI
AI that creates new content such as text, images, audio, or code, rather than just classifying or scoring existing data.
Machine Learning
MLA way of building software that learns patterns from examples instead of being told exactly what to do.
Narrow AI
ANIAI that is built and trained for one specific task, which describes essentially every AI system in use today.
Neural Network
A model loosely inspired by the brain, built from layers of simple units that each pass signals to the next.
Reinforcement Learning
RLTraining a model through trial and error, rewarding good outcomes and penalizing bad ones until it learns a strategy.
Self-Supervised Learning
SSLA training approach where the data provides its own labels, like hiding part of an input and predicting it, which is how LLMs first learn language.
Semi-Supervised Learning
A training approach that mixes a small amount of labeled data with a large amount of unlabeled data to get the best of both.
Supervised Learning
Training a model on labeled examples, where each input comes paired with the correct answer.
Turing Test
A classic thought experiment proposing that a machine is intelligent if a person cannot tell its responses apart from a human in conversation.
Unsupervised Learning
Training a model on data with no labels, so it has to find structure and patterns on its own.
Classical Machine Learning
Classification
A machine learning task that sorts inputs into predefined categories, like labeling an email spam or not spam.
Clustering
Grouping data into clusters of similar items with no predefined labels, a core unsupervised technique.
Decision Tree
A model that makes predictions by following a branching series of yes or no questions, like a flowchart.
Gradient Boosting
An ensemble technique that builds models in sequence, each correcting the errors of the last, behind top performers like XGBoost.
Random Forest
An ensemble model that combines many decision trees and lets them vote, far more accurate and stable than one tree.
Regression
A machine learning task that predicts a continuous number, like a price or a temperature, rather than a category.
Models & Architecture
Activation Function
The small nonlinear function applied at each neuron that lets a network learn complex, nonlinear patterns.
Attention
A mechanism that lets a model focus on the most relevant parts of its input when producing each piece of output.
Autoencoder
A network that learns to compress data into a compact code and reconstruct it, useful for denoising and finding structure.
Convolutional Neural Network
CNNA network design built for images, which scans for small visual patterns and builds them up into recognizable objects.
Diffusion Model
A type of generative model, common in image tools, that creates content by starting from noise and gradually refining it.
Dropout
A training trick that randomly switches off some neurons each step to stop a network from overfitting.
Embedding
A way of turning words, images, or other data into lists of numbers that capture their meaning, so a computer can compare them.
Encoder-Decoder
A model design with one half that reads and compresses the input and another that generates the output, common in translation.
Foundation Model
A large, general-purpose model trained on broad data that many different applications can be built on top of.
Frontier Model
One of the largest and most capable AI models at the leading edge of what is currently possible.
Generative Adversarial Network
GANA generative design where two networks compete, one creating fakes and one trying to catch them, improving each other.
Large Language Model
LLMA model trained on massive amounts of text that predicts and generates language, powering tools like ChatGPT and Claude.
Long Short-Term Memory
LSTMA type of recurrent network with a built-in memory gate, designed to remember information across long sequences.
Mixture of Experts
MoEA model design that routes each input to a few specialized sub-networks instead of running the whole model every time.
Multimodal AI
A model that can work with more than one type of input or output, such as text, images, audio, and video together.
Parameters
The adjustable numbers inside a model that get tuned during training and store what the model has learned.
Recurrent Neural Network
RNNAn older network design for sequences like text or audio, which processes one step at a time while carrying a memory forward.
Transformer
The neural network design behind almost every modern language model, built around a mechanism called attention.
Vision Transformer
ViTA transformer adapted for images, which splits a picture into patches and treats them like a sequence of words.
Weights
The individual strengths of the connections in a neural network, which are what actually change as a model learns.
Training & Learning
Backpropagation
The algorithm that figures out how to adjust each weight in a network to reduce its errors.
Batch Size
How many training examples the model looks at before it updates its weights once.
Direct Preference Optimization
DPOA simpler alternative to RLHF that tunes a model on human preferences directly, skipping the separate reward model and reinforcement learning.
Epoch
One complete pass of the training process through the entire training dataset.
Fine-Tuning
Taking a pretrained model and training it further on a smaller, focused dataset to specialize it for a task or style.
Gradient Descent
The optimization method that gradually tweaks a model to lower its error, like walking downhill to the lowest point.
Hyperparameter
A setting you choose before training that controls how the model learns, as opposed to what it learns.
Instruction Tuning
Fine-tuning a base model on examples of instructions and good responses so it follows requests instead of just continuing text.
Learning Rate
The size of the steps a model takes while learning, and one of the most important training settings to get right.
Loss Function
The formula that scores how wrong a model is, giving training a single number to push downward.
Low-Rank Adaptation
LoRAAn efficient fine-tuning method that freezes the original model and trains only a small set of added weights.
Optimizer
The algorithm that decides how to adjust a model weights from its errors, with Adam being the common default.
Overfitting
When a model memorizes its training data too closely and then performs badly on anything new.
Pretraining
The first, largest training phase where a model learns general patterns from a huge amount of broad data.
Regularization
Any technique that discourages a model from fitting its training data too closely, to help it generalize.
Reinforcement Learning from Human Feedback
RLHFA training method that tunes a model to be helpful by rewarding the responses human reviewers prefer.
Scaling Laws
The empirical finding that model performance improves predictably as you increase model size, data, and compute together.
Training Data
The examples a model learns from, whose size, quality, and bias largely determine how the model turns out.
Transfer Learning
Reusing what a model learned on one task as a head start for a different but related task.
Underfitting
When a model is too simple or undertrained to capture the real pattern, so it performs poorly even on its training data.
Prompting & Inference
Beam Search
A decoding strategy that explores several candidate continuations at once and keeps the best overall sequence.
Chain-of-Thought
Prompting a model to reason step by step, which tends to improve its accuracy on hard problems.
Context Window
The maximum amount of text, measured in tokens, that a model can consider at once.
Few-Shot Learning
Guiding a model by including a handful of examples in the prompt so it can follow the pattern.
In-Context Learning
A model’s ability to learn a task from examples placed in the prompt, without any change to its weights.
Inference
Actually running a trained model to get an answer, as opposed to training it in the first place.
Prompt
The input you give a generative model, usually instructions or a question, that tells it what you want.
Prompt Engineering
The practice of wording and structuring prompts to get better, more reliable results from a model.
Reasoning Model
A model trained to think step by step before answering, spending extra compute at inference time to solve harder problems.
Speculative Decoding
An inference speedup where a small fast model drafts several tokens and the big model checks them in parallel.
Streaming
Delivering a model response token by token as it is generated, instead of waiting for the whole thing.
Structured Output
Making a model return data in a strict machine-readable format like JSON, so software can use the answer directly.
System Prompt
A behind-the-scenes instruction that sets a model’s role, rules, and tone for an entire conversation.
Temperature
A setting that controls how random or predictable a model’s output is.
Token
The small chunk of text a language model actually reads and writes, usually a word or part of a word.
Tokenization
The step that chops text into tokens before a model can process it.
Top-p Sampling
A decoding method that picks the next token from the smallest set of options whose probabilities add up to a chosen threshold.
Zero-Shot Learning
Asking a model to do a task with no examples, relying only on instructions and what it already knows.
Retrieval & Agents
Agent Memory
The mechanisms that let an AI agent remember things across steps or sessions, beyond what fits in one context window.
Agentic AI
A general term for AI that acts with autonomy over multiple steps, rather than responding one prompt at a time.
AI Agent
An AI system that can take actions and use tools to pursue a goal, not just answer a single question.
Chunking
Splitting documents into smaller passages before indexing them, a foundational step that makes or breaks a RAG system.
Computer Use
Giving an AI agent control of a computer, letting it click, type, and navigate apps and websites like a person.
Function Calling
A capability that lets a model use external tools by outputting a structured request to run a specific function.
GraphRAG
A RAG variant that builds a knowledge graph from your documents so it can answer broad questions ordinary RAG cannot.
Human in the Loop
HITLKeeping a person in the decision path to review, approve, or correct an AI action, especially for high-stakes steps.
Hybrid Search
Combining keyword search with meaning-based vector search to get the strengths of both.
Model Context Protocol
MCPAn open standard for connecting AI models to external tools and data sources in a consistent way.
Multi-Agent System
A setup where several specialized AI agents collaborate, often coordinated by an orchestrator, to tackle a complex task.
ReAct
An agent pattern that interleaves reasoning and action in a loop: think, take an action with a tool, observe the result, repeat.
Reranking
A second pass that reorders retrieved results by relevance, surfacing the best few before they reach the model.
Retrieval-Augmented Generation
RAGA technique that looks up relevant documents and feeds them to a model so its answers are grounded in real sources.
Semantic Search
Search that matches on meaning rather than exact keywords, so related ideas surface even with different wording.
Vector Database
A database built to store embeddings and quickly find the items most similar in meaning to a query.
Alignment & Safety
AI Alignment
The work of making AI systems pursue what people actually intend, including their values and safety.
AI Ethics
The study and practice of building and using AI responsibly, covering fairness, privacy, accountability, and harm.
Algorithmic Bias
When an AI system produces unfair or skewed results, usually because its training data reflected existing bias.
Constitutional AI
A training approach that aligns a model to an explicit written set of principles instead of relying only on human ratings.
Guardrails
Controls placed around a model to keep its inputs and outputs safe, on-topic, and within policy.
Hallucination
When a model states something false or made up as if it were fact, fluently and confidently.
Interpretability
Research into understanding why a model produces the outputs it does, rather than treating it as a black box.
Jailbreak
A prompt crafted to trick a model into bypassing its own safety rules and doing something it should refuse.
Red Teaming
Deliberately probing an AI system to find ways it can fail, be misused, or produce harmful output, before attackers do.
Reward Hacking
When a model learns to game its training objective, scoring well on the metric while missing what you actually wanted.
Superalignment
The research problem of keeping AI systems that are smarter than humans safe and under meaningful human control.
Sycophancy
A model tendency to tell you what it thinks you want to hear, agreeing or flattering instead of being accurate.
Security & Policy
AI Governance
The policies, processes, and oversight an organization or government uses to manage AI responsibly.
Data Poisoning
Corrupting a model training data on purpose to plant hidden flaws or backdoors in the resulting model.
Deepfake
Synthetic media, usually video or audio, that convincingly fakes a real person saying or doing something they never did.
EU AI Act
The European Union comprehensive AI law, which regulates AI by risk level and is the first of its kind at this scale.
Model Collapse
The degradation that can occur when models are trained on AI-generated data, generation after generation, losing touch with reality.
Prompt Injection
An attack that hides malicious instructions in content a model reads, tricking it into ignoring its real orders.
Watermarking
Embedding a hidden, detectable signal in AI-generated text, images, or audio to mark it as machine-made.
Applications
Chatbot
A program you interact with through conversation, now usually powered by a large language model.
Code Generation
Using AI to write, complete, or explain software code from natural language or surrounding context.
Computer Vision
The field of AI that lets computers interpret images and video, from recognizing objects to reading documents.
Copilot
An AI assistant embedded in a tool to help with the work as you do it, like suggesting code or drafting text.
Machine Translation
Automatically translating text or speech from one language to another, one of the oldest goals of AI.
Natural Language Processing
NLPThe field of AI focused on getting computers to understand and generate human language.
Optical Character Recognition
OCRConverting images of text, like scanned documents or photos, into machine-readable, editable text.
Recommendation System
An AI system that predicts what a user will want next, powering feeds, product suggestions, and playlists.
Robotics
Giving AI a physical body, so it can perceive and act in the real world, also called embodied AI.
Speech Recognition
Technology that converts spoken audio into text, the listening half of voice assistants and transcription tools.
Text-to-Image
Generating original images from a written description, as tools like DALL-E, Midjourney, and Stable Diffusion do.
Text-to-Speech
TTSTurning written text into natural-sounding spoken audio, the voice behind assistants and audiobooks.
Text-to-Video
Generating video clips from a text description, one of the newest and most compute-hungry frontiers of generative AI.
Infrastructure
Compute
The raw processing power used to train or run a model, a key driver of cost and capability.
CUDA
The software platform from NVIDIA that lets general code run on its GPUs, and a big reason NVIDIA dominates AI hardware.
Edge AI
Running AI directly on a local device like a phone or camera, instead of sending data to the cloud.
Floating-Point Operations Per Second
FLOPSA measure of raw computing speed used to size AI hardware and training runs.
Graphics Processing Unit
GPUThe specialized chip that does the heavy parallel math behind training and running AI models.
Knowledge Distillation
Training a smaller model to imitate a larger one, keeping much of the quality at a fraction of the cost.
KV Cache
A memory optimization that stores earlier work during generation so the model does not recompute it for every new token.
Model Serving
The infrastructure that runs a trained model in production and answers live requests reliably and at scale.
Open Weights
A model whose trained parameters are released publicly, so anyone can download, run, and build on it.
Quantization
Shrinking a model by storing its numbers at lower precision, so it runs faster and on smaller hardware.
Tensor Processing Unit
TPUA chip designed by Google specifically for machine learning, as an alternative to general-purpose GPUs.
Evaluation & Operations
Benchmark
A standardized test set used to measure and compare model performance on a specific capability.
Evaluation
Measuring how good a model actually is, from accuracy on tests to human judgments of quality and safety, often called evals.
Leaderboard
A public ranking of models by their scores on benchmarks or head-to-head human preference votes.
MLOps
The practice of reliably building, deploying, monitoring, and maintaining machine learning models in production, like DevOps for ML.
Model Card
A short document that describes a model intended uses, training, performance, and limitations, for transparency.