AI Glossary,from concepts to practice

A deep learning model pretrained on massive text, good at understanding and generating natural language.

Artificial intelligence that matches or exceeds human ability across nearly all cognitive tasks.

The question, instruction or example text a user sends to an AI model.

The smallest unit of text a model processes; it can be a word, subword or character.

The maximum number of tokens a model can process in one request.

Model output that sounds plausible but is actually wrong or fabricated.

A parameter controlling output randomness: higher is more creative, lower is more deterministic.

The learnable numbers inside a model, usually measured in billions (B).

AI that creates content such as text, images, video and code.

A machine learning approach where multi-layer neural networks learn features and patterns from data automatically.

A computational model of many interconnected "neurons" that can approximate complex functions.

The process of steering model behavior toward human intent, values and safety requirements.

Standardized tasks and scoring used to compare model capabilities quantitatively.

Models with publicly available weights that can be freely downloaded and deployed.

Models served only through APIs, with weights kept private.

Giving a few examples in the prompt so the model can generalize to the task.

Asking the model to perform an unseen task from the description alone, without examples.

The cutoff date of a model's training data; events after it are unknown to the model.

A model's ability to reason logically, think in steps and solve complex problems.

A neural network architecture based on self-attention; the foundation of modern LLMs.

A mechanism that lets a model dynamically focus on important parts of the input.

Attention computed between all pairs of elements within one sequence.

Running several attention heads in parallel to capture diverse relations in different subspaces.

A method for injecting position information into a model.

An architecture that splits a model into "expert" subnetworks, activating only a few per token.

A class of generative models that create data such as images and video by iteratively denoising.

A two-part architecture where an encoder reads the input and a decoder generates output.

Caching already-computed key-value vectors during inference to avoid recomputation.

The component that splits text into token sequences—the first and last stop for model I/O.

Generating text one token at a time, each predicted from the previous ones.

Attention approximated by computing only a subset of token pairs.

The low-dimensional semantic space into which data is compressed and encoded.

Learning a new task from examples in the prompt without updating weights.

The training algorithm that computes gradients backward from output error and updates parameters layer by layer.

Training a model on massive unlabeled text to learn language patterns.

Further training a pretrained model on specific data to adapt it to a task.

Training a model on human-labeled instruction-response data to follow instructions.

An alignment method that trains a reward model from human preferences, then optimizes the policy against it.

Efficient fine-tuning that trains only a few low-rank matrices.

A memory-efficient fine-tuning scheme combining 4-bit quantization with LoRA.

A simplified alignment method that optimizes the policy directly from preference data, without a reward model.

A reinforcement learning algorithm commonly used in RLHF; clipping update steps keeps training stable.

Training a model on diverse instruction data so it understands and executes task descriptions.

Continued pretraining on new corpora to inject domain knowledge or new languages.

Training a smaller "student" model on the outputs of a larger "teacher" model.

A model memorizing training data so well that performance on unseen data degrades.

The massive corpora of text, images and more used to train models.

Expanding training data via transformations, rewrites and similar techniques.

Using a trained model to produce output for new input.

The time from sending a request to receiving the first (or full) response.

The number of requests or tokens a model handles per unit of time.

Returning tokens as they are generated, so users need not wait for the full answer.

Making the model output strictly in predefined formats such as JSON.

Representing weights in lower precision (8-bit, 4-bit) to shrink size and speed up inference.

A high-performance LLM serving framework known for high throughput and ease of use.

Spending more compute at inference time in exchange for higher-quality output.

Extra compute allocated at inference time (search, verification, iteration).

The standard interface for applications to access model services, billed by usage.

Provider limits on requests or tokens per unit of time.

Merging multiple requests into one inference batch to raise throughput.

Reusing computed results for identical prefix content to cut repeated input cost.

Inference billed by actual usage, with no GPU cluster to manage.

A technique that lets LLMs retrieve external documents before generating an answer.

An AI system that plans, calls tools and completes multi-step tasks autonomously.

Mapping text to high-dimensional numeric vectors where similar semantics sit closer together.

A database built for storing and searching high-dimensional vectors.

Search based on meaning rather than keyword matching.

The practice of designing prompts to optimize model output.

A prompting technique that guides the model to reason step by step before answering.

An agent paradigm where the model alternates think-act-observe cycles.

A model's ability to emit call arguments matching predefined function signatures.

The model calling external tools (search, code execution, databases) to extend capability.

Pluggable components that extend a model's or app's capabilities.

The ability of an agent to retain and use information across interactions.

Automating multi-step tasks by orchestrating them into fixed pipelines.

A pattern where multiple specialized agents collaborate on complex tasks.

AI assistants embedded in dev tools that help write code.

An AI application that provides services through conversation.

Models that process multiple modalities—text, image, audio—as input or output.

Models combining a vision encoder with a language model for image understanding.

AI that generates images from text descriptions.

AI that generates video clips from text descriptions.

A model's ability to recognize and explain image content.

Technology that converts text into natural-sounding speech.

Technology that converts speech into text.

Technology that recognizes and extracts text from images.

A model's ability to understand video content, timing and events.

Aligning and transferring information across modalities such as text, image and audio.

Replicating a specific voice's timbre and style from a small sample.

Research and practice ensuring AI systems are reliable, controllable and non-harmful.

Systematically probing model weaknesses from an attacker's perspective.

The ability to understand and explain why a model made a decision.

Copyright ownership and compliance for training data and generated content.

Requirements to protect personal data during training and model use.

The computing resources needed to train and run AI models.

Chips with massive parallel compute—the workhorse hardware for AI training and inference.

Licenses governing rights to use, modify and commercialize model weights and code.

Documentation accompanying a model release covering capabilities, limitations and usage.

Rankings of models by evaluation scores.

Laws and governance frameworks for AI development and deployment across jurisdictions.

A security attack that hijacks model behavior via crafted inputs.

Systematic skew or unfair tendencies in model output.