AI Glossary

An AI system that perceives its environment, makes decisions, and takes actions — often using tools like web search, code execution, or external APIs — to autonomously accomplish multi-step goals. Agents go beyond single-turn question answering to execute workflows.

The research field focused on ensuring that AI systems pursue goals and exhibit behaviors that are safe and beneficial to humans, even as they become more capable. Alignment asks: how do we build AI that reliably does what we actually want?

AI systems designed to operate with significant autonomy, planning and executing multi-step tasks over extended time horizons with minimal human intervention. Agentic AI represents a shift from AI as a tool to AI as an autonomous collaborator or worker.

A hypothetical AI system capable of performing any intellectual task that a human can perform, across all domains, without task-specific training. No AGI system exists today; current AI models excel within defined domains but lack general-purpose cognitive flexibility.

A theoretical AI system that surpasses the cognitive ability of all humans combined across every domain, including scientific creativity, social reasoning, and strategic planning. ASI does not exist and may be decades or more away, if achievable at all.

A neural network component that allows a model to dynamically focus on different parts of its input when producing each output token, assigning learned relevance scores across the full context. Self-attention, used in transformers, computes these scores between all positions in the input simultaneously.

Automated machine learning — tools and techniques that automate the selection, configuration, and training of ML models, reducing the need for manual expert tuning. AutoML democratizes model development by enabling non-specialists to build effective models.

A standardized test or dataset used to measure and compare AI model performance on defined tasks. Benchmarks provide a common language for tracking progress and identifying strengths and weaknesses across models.

In neural networks, a learnable offset added to a neuron's weighted input sum, allowing the model to fit data that doesn't pass through the origin. Separately, "bias" also refers to systematic errors or unfair skews in model outputs caused by imbalances in training data.

A prompting technique where the model is instructed or shown examples of reasoning step-by-step before giving a final answer, improving performance on complex reasoning, math, and multi-step tasks. Chain-of-thought prompting essentially asks the model to "show its work."

The AI field concerned with teaching machines to interpret and understand visual information from images and video. Core tasks include image classification, object detection, semantic segmentation, and visual question answering.

Synthetic media — typically video or audio — in which a person's likeness or voice is convincingly replaced or fabricated using AI, often without their consent. Deepfakes pose serious risks for misinformation, fraud, and non-consensual intimate imagery.

A generative AI model that learns to create data — most commonly images — by gradually reversing a noise-addition process. During training, the model learns to denoise progressively noisier versions of real images; at inference it starts from pure noise and iteratively refines a coherent output.

Running AI model inference directly on local devices — smartphones, cameras, sensors, vehicles — rather than sending data to cloud servers. Edge AI reduces latency, preserves privacy, and enables AI in offline or bandwidth-constrained environments.

A dense numerical vector representation of a word, sentence, image, or other data object that captures its semantic meaning in a high-dimensional space. Items with similar meanings have vectors that are close together by distance metrics like cosine similarity.

A distributed machine learning approach where model training happens locally on user devices, and only model weight updates — not raw data — are shared with a central server. Federated learning enables AI improvement while preserving user data privacy.

The ability of a model to perform a new task correctly after seeing only a small number of examples in the prompt, without updating its weights. Few-shot learning is a key emergent capability of large pre-trained models.

The process of continuing training a pre-trained model on a smaller, task-specific dataset to adapt it to a particular domain or behavior. Fine-tuning requires far less compute than training from scratch because the model already has broad world knowledge.

A specific implementation of tool use where the AI model generates structured JSON to invoke a developer-defined function, enabling reliable integration between LLMs and application code. Function calling makes LLM outputs machine-parseable and action-ready.

Generative Pre-trained Transformer — a family of large language models developed by OpenAI that generate coherent text by predicting subsequent tokens. GPT models popularized the pre-train-then-fine-tune paradigm that now dominates NLP research.

A deep learning framework consisting of two networks — a generator and a discriminator — that compete against each other, pushing the generator to produce increasingly realistic synthetic data. Introduced by Ian Goodfellow in 2014, GANs were the dominant generative image model before diffusion models.

Technical mechanisms — filters, classifiers, constitutional rules, or system prompts — that constrain AI model outputs to be safe, appropriate, and within defined policy limits. Guardrails are a practical layer of safety on top of model-level training.

When an AI model generates text that is factually incorrect, fabricated, or unsupported by its input, while presenting it with apparent confidence. Hallucination is a fundamental reliability challenge for large language models.

The process of running a trained AI model on new input data to produce a prediction or output, as opposed to training. Inference is what happens every time you send a message to a chatbot or submit an image for analysis.

A structured representation of entities and the relationships between them, stored as a graph of nodes (entities) and edges (relationships). Knowledge graphs encode factual world knowledge in a machine-queryable form that complements the statistical knowledge inside LLMs.

A type of AI model trained on vast amounts of text data to understand and generate human language. LLMs predict the next token in a sequence, a simple objective that gives rise to surprisingly broad reasoning and language capabilities.

The discipline of applying DevOps principles — continuous integration, deployment, monitoring, and automation — to the machine learning lifecycle. MLOps bridges the gap between experimental data science and reliable production AI systems.

Massive Multitask Language Understanding — a benchmark covering 57 academic subjects from elementary mathematics to professional law and medicine, used to measure the breadth of world knowledge in language models. MMLU has become one of the most widely cited evaluation suites for LLM capability.

An open standard developed by Anthropic that defines a universal interface for connecting AI models to external tools, data sources, and services. MCP replaces ad-hoc integrations with a consistent protocol any model or application can implement.

AI systems that can process and generate multiple types of data — such as text, images, audio, and video — within a single model. Multimodal models can answer questions about images, generate images from text, or transcribe and summarize audio.

The field of AI focused on enabling computers to understand, interpret, and generate human language. NLP encompasses tasks such as translation, summarization, sentiment analysis, question answering, and named-entity recognition.

A computational system loosely inspired by the human brain, composed of layers of interconnected nodes (neurons) that learn to recognize patterns from data. Adjusting the connection weights during training allows the network to improve its predictions over time.

A formal specification of concepts, categories, and the relationships between them within a domain, providing a shared vocabulary that machines can reason over. Ontologies are the structured knowledge layer beneath knowledge graphs.

A learnable numerical value inside a neural network — typically a weight or bias — that is adjusted during training to minimize the model's prediction error. The count of parameters is the most common shorthand for model size.

A statistical measure of how well a language model predicts a sample of text — lower perplexity means the model assigns higher probability to the actual next tokens and is therefore a better fit. Perplexity is a standard intrinsic evaluation metric for language models.

The practice of crafting input text to elicit desired outputs from a language model, without changing the model weights. Effective prompts can dramatically improve accuracy, tone, format, and reasoning quality.

A structured adversarial testing process where people or automated systems attempt to find failure modes, safety vulnerabilities, or policy violations in an AI model before it is deployed. Red teaming is a core practice in responsible AI development.

A machine learning paradigm where an agent learns by taking actions in an environment and receiving reward or penalty signals based on the outcomes. Over many iterations the agent develops a policy that maximizes cumulative reward.

A training method where human raters compare model outputs and their preferences are used to train a reward model, which then guides further RL training. RLHF aligns model behavior with human values and instructions far more effectively than pure supervised learning alone.

The process of fetching relevant documents, passages, or data from an external store in response to a query, used to ground AI model responses in specific, up-to-date information. Retrieval is the first stage of RAG pipelines and semantic search systems.

A technique that combines a language model with a live retrieval system, fetching relevant documents from an external knowledge base before generating a response. RAG grounds LLM outputs in up-to-date, verifiable facts rather than relying solely on trained parameters.

Artificially generated data that mimics the statistical properties of real-world data, used to train or evaluate AI models when real data is scarce, sensitive, or imbalanced. Synthetic data is increasingly used to bootstrap model training and augment edge cases.

The basic unit of text that a language model processes — typically a word, sub-word, or character, depending on the tokenizer. Models read and generate sequences of tokens rather than raw characters or words.

The ability of an AI model to call external functions, APIs, or services during inference, extending its capabilities beyond what is encoded in its weights alone. Tool use allows models to retrieve current information, run code, query databases, and interact with the world.

The process of adjusting a model's parameters by exposing it to data and minimizing a loss function using gradient-based optimization. Training is the computationally intensive phase that produces a model capable of making useful predictions.

A machine learning approach where a model trained on one task or dataset is adapted for a different but related task, leveraging knowledge already encoded in its weights. Transfer learning dramatically reduces the data and compute needed for specialized applications.

A neural network architecture that uses self-attention mechanisms to process sequences of data in parallel, enabling highly efficient training at scale. It is the foundation of virtually all modern large language models.

A database optimized for storing and querying high-dimensional embedding vectors, enabling fast approximate nearest-neighbor search at scale. Vector databases are the storage backbone of most RAG and semantic search systems.

A specific type of model parameter representing the strength of the connection between two neurons in a neural network. Weights determine how strongly one neuron's output influences the next layer's computation.

A model's ability to perform a task it has never seen examples of, relying entirely on instructions and pre-trained knowledge. Zero-shot capability indicates that a model has learned robust enough representations to generalize to novel problem types.