AI Dictionary
Plain-language definitions for the acronyms and terms shaping AI today. Each entry is tagged by the AI literacy level at which it most naturally lives.
Agent
Level 2An AI system that can plan, take actions, and use tools autonomously to accomplish a goal, often across multiple steps.
Agentic AI
Level 2AI systems designed to operate with autonomy — making decisions and taking actions in pursuit of objectives rather than just answering questions.
A hypothetical AI with broad, human-level cognitive abilities across most domains, as opposed to narrow AI focused on a single task.
AI (Artificial Intelligence)
Level 1The broad field of building computer systems that perform tasks normally requiring human intelligence.
Alignment
Level 2The research area focused on ensuring AI systems behave in accordance with human values and intentions.
ANN (Artificial Neural Network)
Level 3A computing system loosely modeled on the human brain, made of interconnected layers of nodes (neurons) that learn patterns from data.
A defined way for software to talk to other software. AI APIs let developers send prompts to a model and receive responses programmatically.
Attention Mechanism
Level 3A neural network technique that lets a model weigh the importance of different parts of an input when producing each part of the output — the foundation of modern transformer models.
Backpropagation
Level 4The core training algorithm for neural networks, which adjusts model weights by propagating errors backward through the layers.
Batch Size
Level 4The number of training examples processed before the model's parameters are updated in a single training step.
Benchmark
Level 3A standardized test used to compare AI model performance, such as MMLU, HumanEval, or GSM8K.
Bias (in AI)
Level 1Systematic errors or unfair outcomes in model behavior, often stemming from training data or design choices.
Chain-of-Thought (CoT)
Level 2A prompting and reasoning technique where the model explicitly works through intermediate steps before producing a final answer.
Chatbot
Level 1A conversational interface, often powered by an LLM, that interacts with users through natural language.
Classification
Level 2A machine learning task where the model assigns inputs to one of a set of predefined categories.
A neural network architecture especially well suited to image and spatial data, using convolutional layers to detect local patterns.
Completion
Level 2The text or output that a model produces in response to a prompt.
Compute
Level 2The processing power (typically GPU/TPU hours) consumed by training or running AI models.
Context Window
Level 2The maximum amount of text (measured in tokens) that a language model can consider at once when producing a response.
Corpus
Level 3A large, structured collection of text used as training or evaluation data.
Dataset
Level 2A collection of examples used to train, validate, or test a machine learning model.
Deep Learning
Level 2A subfield of machine learning that uses neural networks with many layers to learn complex patterns from large datasets.
Diffusion Model
Level 3A generative model that learns to create images (or other data) by gradually denoising random noise — the technology behind tools like Stable Diffusion and DALL·E.
Distillation
Level 4A technique where a smaller "student" model is trained to mimic the outputs of a larger "teacher" model, producing a faster, cheaper version with similar quality.
Embedding
Level 3A numerical (vector) representation of text, images, or other data that captures semantic meaning, enabling similarity search and retrieval.
Epoch
Level 4One complete pass of the training algorithm over the entire training dataset.
Evaluation (Eval)
Level 3The process of measuring a model's performance on specific tasks, often using benchmarks or custom test sets.
Few-Shot Learning
Level 2A prompting technique where a model is given a few examples of the desired task in the prompt to guide its response.
Fine-Tuning
Level 3The process of further training a pre-trained model on a smaller, specialized dataset to adapt it to a specific task or domain.
Foundation Model
Level 2A large, general-purpose model trained on broad data that can be adapted to many downstream tasks.
A model architecture where two neural networks (a generator and a discriminator) compete, producing increasingly realistic synthetic data.
Generative AI (GenAI)
Level 1AI systems that create new content — text, images, audio, video, code — rather than just analyzing or classifying existing data.
A family of large language models developed by OpenAI, based on the transformer architecture.
GPU (Graphics Processing Unit)
Level 2Specialized hardware originally for graphics, now the workhorse for training and running large AI models due to massive parallelism.
Gradient Descent
Level 4The optimization algorithm used to train neural networks by iteratively adjusting weights to minimize a loss function.
Grounding
Level 2Connecting a model's output to verifiable, external sources of truth (such as documents or databases) to reduce hallucination.
Guardrails
Level 2Rules, filters, or constraints applied to AI systems to keep their behavior within safe and appropriate bounds.
Hallucination
Level 1When a generative AI model produces output that is fluent and confident but factually incorrect or fabricated.
HITL (Human-in-the-Loop)
Level 2A workflow design where a human reviews, corrects, or approves AI outputs at key points.
Hyperparameter
Level 4A configuration value (such as learning rate or batch size) set before training, as opposed to weights learned during training.
Inference
Level 3The process of using a trained model to generate predictions or outputs from new inputs (as opposed to training).
Instruction Tuning
Level 3A fine-tuning technique that trains a model to follow natural-language instructions across many task types.
Knowledge Cutoff
Level 1The date after which a model's training data ends, beyond which it has no built-in knowledge of world events.
Latency
Level 2The time it takes for a model to produce a response after receiving a request.
Layer (Neural Network Layer)
Level 3A group of neurons in a neural network that processes inputs from the previous layer and passes outputs to the next.
Learning Rate
Level 4A hyperparameter that controls how much model weights change in response to error during training.
LLM (Large Language Model)
Level 1A neural network trained on vast amounts of text to predict and generate human-like language.
LLMOps
Level 3The practices and tooling for deploying, monitoring, and maintaining LLM-powered applications in production.
LoRA (Low-Rank Adaptation)
Level 4A parameter-efficient fine-tuning method that adds small, trainable matrices to a frozen base model, dramatically reducing training cost.
Loss Function
Level 4A mathematical function that measures how far a model's predictions are from the correct answers, used to guide training.
Machine Learning (ML)
Level 1A subfield of AI in which systems learn patterns from data rather than following explicitly programmed rules.
MCP (Model Context Protocol)
Level 3An open protocol that standardizes how AI applications connect to external data sources, tools, and services.
Model
Level 2The trained mathematical artifact (weights and architecture) that performs a task such as generating text or classifying images.
Model Card
Level 3A standardized document describing a model's intended use, training data, performance, limitations, and risks.
MoE (Mixture of Experts)
Level 4A model architecture in which many specialized sub-networks ("experts") share the work, with a router selecting which experts process each input.
Multimodal
Level 2Describes models that can process and/or generate more than one type of data — for example, text, images, audio, and video together.
Narrow AI
Level 1AI designed and trained for a single, specific task (e.g., spam filtering), as contrasted with AGI.
The branch of AI focused on enabling computers to understand, interpret, and generate human language.
Neural Network
Level 2A model composed of layers of interconnected nodes that learn to map inputs to outputs through training.
Open Source Model
Level 2An AI model whose weights (and often training code) are publicly released, allowing inspection, modification, and self-hosting.
Overfitting
Level 3When a model learns training data too closely, including its noise, and consequently performs poorly on new, unseen data.
Parameter
Level 2A learned weight or value inside a model. Modern LLMs have billions to trillions of parameters.
A family of fine-tuning techniques (including LoRA) that update only a small fraction of a model's parameters.
Pre-training
Level 3The initial phase of training a foundation model on a large, broad dataset before any task-specific fine-tuning.
Prompt
Level 1The input text or instructions given to a generative model to elicit a desired response.
Prompt Engineering
Level 2The practice of designing and refining prompts to reliably produce high-quality outputs from a model.
Prompt Injection
Level 2A security attack where malicious instructions hidden in input data trick a model into ignoring its original instructions.
A technique that combines a generative model with a retrieval system, letting the model pull in relevant external documents at query time.
Reasoning Model
Level 2A model designed to spend additional compute "thinking" through a problem before answering, often improving accuracy on complex tasks.
Recall
Level 3A metric that measures the fraction of relevant items that a system successfully retrieved or identified.
Regression
Level 2A machine learning task that predicts a continuous numerical value (e.g., a price) rather than a category.
Reinforcement Learning (RL)
Level 3A training paradigm where an agent learns by taking actions in an environment and receiving rewards or penalties.
A training method that uses human preferences to shape a model's behavior, widely used to make LLMs more helpful and safe.
RNN (Recurrent Neural Network)
Level 3A neural network architecture designed for sequential data, processing inputs one step at a time while maintaining a hidden state.
Safety (AI Safety)
Level 2The discipline of designing AI systems that behave reliably, avoid harm, and remain under meaningful human control.
Self-Supervised Learning
Level 3A training approach where the model generates its own labels from unlabeled data — for example, predicting the next word in a sentence.
Semantic Search
Level 2Search that matches meaning rather than exact keywords, typically powered by embeddings.
SLM (Small Language Model)
Level 2A language model with relatively few parameters, optimized for speed, cost, or on-device deployment.
SOTA (State of the Art)
Level 3The current best published performance on a given task or benchmark.
Supervised Learning
Level 2A training approach where the model learns from labeled examples (input/output pairs).
System Prompt
Level 2A foundational instruction given to a model that defines its role, behavior, and constraints across a conversation.
Temperature
Level 2A sampling parameter that controls randomness in a model's output — lower values are more deterministic, higher values more creative.
Token
Level 2A chunk of text (a word, subword, or character) that a language model processes as a single unit.
Tokenizer
Level 3The component that splits raw text into tokens (and converts tokens back to text) for a language model.
Tool Use
Level 2The capability of an AI model to call external functions, APIs, or services to gather information or take actions.
Top-k / Top-p Sampling
Level 3Sampling strategies that limit the model's word choices to the most probable options, controlling output variety.
TPU (Tensor Processing Unit)
Level 3Custom AI accelerator hardware developed by Google, optimized for tensor operations used in deep learning.
Training
Level 2The process of teaching a model by exposing it to data and adjusting its parameters to minimize errors.
Transfer Learning
Level 3Reusing a model trained on one task as the starting point for training on a related task.
Transformer
Level 3The neural network architecture, introduced in 2017, that underpins modern LLMs and most state-of-the-art generative models.
Unsupervised Learning
Level 2A training approach where the model finds structure in unlabeled data (e.g., clustering similar items).
Vector Database
Level 3A database optimized for storing and searching embeddings, enabling fast similarity search at scale.
Weights
Level 3The numerical parameters inside a neural network that are adjusted during training and determine the model's behavior.
XAI (Explainable AI)
Level 2Techniques and practices for making AI model decisions understandable to humans.
Zero-Shot Learning
Level 2A model's ability to perform a task without having seen any examples of it during training or in the prompt.