AI Dictionary: Key Terms & Concepts
Artificial intelligence has its own vocabulary — and it's easy to get lost in the jargon. This quick-reference guide breaks down the most essential AI and machine learning terms, from foundational concepts like neural networks and training to modern buzzwords like LLMs, RAG, and hallucination. Each entry includes a plain-language definition and a real-world example, so whether you're a beginner or just need a refresher, you can quickly get up to speed.
| Term | Definition | Example |
|---|---|---|
| AI (Artificial Intelligence) | The broad field of building systems that perform tasks normally requiring human intelligence. | A voice assistant like Siri understanding and responding to spoken questions. |
| Machine Learning (ML) | A subset of AI where systems learn patterns from data instead of being explicitly programmed. | An email spam filter that improves as it sees more labeled spam/not-spam examples. |
| Deep Learning | ML using multi-layered neural networks to learn complex patterns from large datasets. | A system that identifies cats in photos by learning layers of visual features (edges → shapes → objects). |
| Neural Network | A model loosely inspired by the brain, made of connected layers of nodes ("neurons") that process data. | A network with an input layer (pixel values), hidden layers, and an output layer (digit 0-9) for handwriting recognition. |
| LLM (Large Language Model) | A neural network trained on massive text datasets to understand and generate human language. | Claude or GPT-4 writing an essay or answering a question in natural language. |
| Transformer | The neural network architecture (introduced in 2017) behind most modern LLMs, built around the attention mechanism. | The architecture underlying models like Claude, GPT, and BERT. |
| Attention Mechanism | A technique letting a model weigh the importance of different parts of the input when producing output. | In "The cat sat on the mat because it was tired," attention helps the model link "it" back to "cat." |
| Token | A chunk of text (word, part of a word, or character) that a model processes as a basic unit. | The word "unbelievable" might split into tokens like "un," "believ," "able." |
| Parameters | The internal numerical values a model learns during training; more parameters generally means more capacity. | A model described as "70 billion parameters" has that many learned weights. |
| Training | The process of adjusting a model's parameters using data so it learns to perform a task. | Feeding millions of images labeled "dog" or "not dog" so a model learns to classify them. |
| Inference | Using a trained model to generate outputs or predictions on new input. | Typing a question into ChatGPT and getting a response — that response is generated via inference. |
| Fine-tuning | Further training a pre-trained model on a smaller, specific dataset to specialize its behavior. | Taking a general LLM and fine-tuning it on legal documents to make a legal-assistant model. |
| Prompt Engineering | Crafting inputs (prompts) to guide a model toward better or more specific outputs. | Adding "Answer step by step" to a math question to get more accurate reasoning. |
| Context Window | The maximum amount of text (in tokens) a model can consider at once. | A model with a 200K-token context window can process roughly a 500-page book in one go. |
| Embedding | A numerical vector representation of text, images, or other data that captures meaning/similarity. | The words "king" and "queen" end up as nearby vectors because they're semantically related. |
| RAG (Retrieval-Augmented Generation) | A technique where a model retrieves relevant external documents to inform its response. | A customer support bot that looks up your company's help docs before answering a question. |
| Hallucination | When a model generates plausible-sounding but false or fabricated information. | A model confidently citing a fake research paper that doesn't exist. |
| Zero-shot Learning | A model performing a task it wasn't explicitly trained/shown examples for, using only instructions. | Asking an LLM to translate slang it's never seen labeled examples of, based on instructions alone. |
| Few-shot Learning | Giving a model a handful of examples in the prompt to guide its response. | Showing 3 examples of "review → sentiment" pairs before asking it to classify a new review. |
| Supervised Learning | Training a model on labeled data (input-output pairs). | Training a model on house features (input) and sale prices (labeled output) to predict prices. |
| Unsupervised Learning | Training a model to find patterns in data without labeled outputs. | Clustering customers into groups based on purchase behavior, with no predefined categories. |
| Reinforcement Learning (RL) | Training a model by rewarding or penalizing its actions to encourage desired behavior. | An AI learning to play chess by getting rewarded for wins and penalized for losses. |
| RLHF (RL from Human Feedback) | Using human preference ratings to fine-tune a model's behavior via reinforcement learning. | Humans ranking two chatbot responses, and the model being trained to prefer the better-ranked style. |
| Overfitting | When a model learns training data too closely and performs poorly on new, unseen data. | A model that memorizes training exam answers but fails on a slightly different exam. |
| Generative AI | AI systems that create new content (text, images, audio, code) rather than just classifying data. | DALL-E generating an image from the prompt "a cat astronaut on the moon." |
| Alignment | The effort to make AI systems act in accordance with human values and intentions. | Training a model to refuse harmful requests even if technically capable of fulfilling them. |
| Bias (in AI) | Systematic skew in a model's outputs, often inherited from patterns in its training data. | A hiring model favoring resumes similar to past hires, disadvantaging underrepresented groups. |
| Multimodal | A model capable of processing/generating multiple data types, like text, images, and audio together. | Uploading a photo of a receipt and asking the model to summarize the expenses. |
| Agent (AI agent) | An AI system that can autonomously take actions, use tools, and pursue multi-step goals. | An AI that books a flight by searching sites, comparing prices, and completing checkout on its own. |
| Weights | The learned numerical values connecting neurons in a network; often used interchangeably with "parameters." | The strength of the connection between two neurons, adjusted during training via backpropagation. |