AI Terms Glossary

model_training Machine Learning Fundamentals

Machine Learning (ML)

A subfield of artificial intelligence where systems learn from data to improve performance on tasks without being explicitly programmed. ML algorithms identify patterns, make decisions, and generate predictions by training on examples.

Model

A mathematical function trained on data that maps inputs to outputs. In ML, a model is the artifact produced after training — it encapsulates learned patterns and is used to make predictions or decisions on new, unseen data.

Training

The process of exposing a machine learning model to labeled or unlabeled data so it can learn patterns. During training, the model adjusts its internal parameters (weights) to minimize a loss function.

Inference

The process of using a trained model to generate predictions or outputs on new data. Unlike training (which is computationally intensive), inference is typically faster and is the production-time operation.

Supervised Learning

A machine learning paradigm where the model is trained on labeled data — input-output pairs — so it learns to predict outputs from inputs. Examples include image classification, spam detection, and price prediction.

Unsupervised Learning

A machine learning paradigm where the model is trained on unlabeled data and must discover hidden patterns or structure on its own. Common tasks include clustering, dimensionality reduction, and anomaly detection.

Reinforcement Learning (RL)

A machine learning paradigm where an agent learns by interacting with an environment, receiving rewards or penalties based on its actions. Used in robotics, game playing (e.g., AlphaGo), and RLHF for fine-tuning LLMs.

Neural Network

A computational model loosely inspired by the human brain, consisting of layers of interconnected nodes (neurons). Each node applies a weighted transformation and an activation function. Deep neural networks with many layers are the foundation of modern AI.

Deep Learning

A subset of machine learning using neural networks with many hidden layers (deep networks). Deep learning excels at learning representations from raw data such as images, text, and audio, and powers most state-of-the-art AI systems.

Overfitting

When a model learns the training data too well — including its noise and outliers — and performs poorly on unseen data. Regularization, dropout, and more training data are common mitigations.

Underfitting

When a model is too simple to capture the underlying patterns in the data, resulting in poor performance on both training and test sets. Typically addressed by increasing model complexity or training longer.

Loss Function

A mathematical function that measures how far a model’s predictions are from the true values. The training process minimizes the loss function via optimization algorithms such as gradient descent.

Gradient Descent

An iterative optimization algorithm used to minimize the loss function during training. At each step, model parameters are updated in the direction of the negative gradient. Variants include SGD, Adam, and AdaGrad.

Hyperparameter

A configuration setting set before training begins (not learned from data). Examples include learning rate, batch size, number of layers, and number of epochs. Hyperparameter tuning is a key step in model optimization.

Feature

An individual measurable property or characteristic of the data used as input to a model. Feature engineering — selecting, transforming, and creating features — is a critical step in the ML pipeline.

Label

The ground-truth output or target value associated with a training example in supervised learning. Labels are what the model is trained to predict (e.g., spam/not-spam, price, sentiment).

chat Large Language Models (LLMs)

Large Language Model (LLM)

A neural network trained on vast amounts of text data to understand and generate human language. LLMs use the Transformer architecture and can perform a wide range of tasks — summarization, translation, coding, question answering — without task-specific training. Examples: GPT-4, Claude, Gemini, Llama.

Token

The basic unit of text processed by an LLM. A token is roughly 4 characters or 0.75 words in English. LLMs process and generate text as sequences of tokens. Tokenization varies by model and language.

Context Window

The maximum number of tokens an LLM can process in a single request — both input (prompt) and output (completion) combined. Larger context windows allow the model to “remember” more of a conversation or document. Modern LLMs range from 8K to 1M+ tokens.

Prompt

The input text provided to an LLM to guide its response. Prompt design — choosing words, structure, and examples — significantly affects output quality. Also referred to as the user message or query.

Completion

The text output generated by an LLM in response to a prompt. Also called a response or generation. LLMs generate completions by predicting the most likely next token given the context.

Temperature

A sampling parameter that controls the randomness of an LLM’s output. A temperature of 0 produces deterministic, conservative responses; higher values (e.g., 1.0) produce more creative and varied output. For factual tasks, lower temperature is preferred.

System Prompt

An instruction provided to an LLM before the user’s message, typically used to set the model’s behavior, persona, tone, and constraints. System prompts are invisible to end users and are a key tool for customizing LLM applications.

Few-shot Prompting

A prompting technique where a small number of input-output examples are included in the prompt to guide the model’s behavior. Helps the model understand the desired format or task without fine-tuning.

Zero-shot Prompting

A prompting technique where no examples are provided — the model is asked to perform a task based on the instruction alone. Works well with capable LLMs on common tasks.

Chain-of-Thought (CoT)

A prompting technique that encourages LLMs to reason step by step before arriving at a final answer. Significantly improves performance on math, logic, and multi-step reasoning tasks. Often triggered by phrases like “Let’s think step by step.”

Embedding

A dense numerical vector representation of text (or other data) that captures semantic meaning. Semantically similar texts have embeddings that are geometrically close. Embeddings power semantic search, RAG, and clustering.

Fine-tuning

The process of further training a pre-trained LLM on a smaller, task-specific dataset to adapt it to a particular domain or behavior. Fine-tuning is more efficient than training from scratch and produces specialized models.

RLHF (Reinforcement Learning from Human Feedback)

A technique used to align LLMs with human preferences. Human evaluators rank model outputs, and a reward model is trained on these rankings. The LLM is then fine-tuned using RL to maximize the reward signal. Used by ChatGPT, Claude, and most modern chat models.

Hallucination

When an LLM generates plausible-sounding but factually incorrect or fabricated information. Hallucinations are a known limitation of LLMs and are mitigated by retrieval-augmented generation (RAG), grounding, and verification layers.

Tokenizer

The component that converts raw text into tokens before it is fed to an LLM. Each model has its own tokenizer (e.g., BPE, SentencePiece). Tokenization affects context length, cost, and multilingual support.

auto_awesome Generative AI

Generative AI (GenAI)

A class of AI systems capable of generating new content — text, images, audio, video, code — rather than just classifying or predicting. GenAI models learn the distribution of training data and sample new examples from it. Includes LLMs, diffusion models, and GANs.

Foundation Model

A large AI model trained on broad, diverse data at scale, which can be adapted to a wide range of downstream tasks via fine-tuning or prompting. Examples include GPT-4, Claude, Llama, DALL-E, and Stable Diffusion.

Multimodal AI

An AI system that can process and generate multiple types of data — text, images, audio, and video — within a single model. Examples include GPT-4o (text + vision), Gemini, and Claude 3 (text + images).

Diffusion Model

A generative model that learns to reverse a process of adding noise to data. During generation, it starts from random noise and iteratively removes noise to produce high-quality images or audio. Powers Stable Diffusion, DALL-E, and Midjourney.

Text-to-Image

A generative AI capability that converts a text description (prompt) into an image. Models such as Stable Diffusion, DALL-E 3, and Midjourney are prominent examples. Widely used in creative content, design, and marketing.

GAN (Generative Adversarial Network)

A generative model architecture consisting of two competing neural networks: a generator (creates fake samples) and a discriminator (distinguishes real from fake). The adversarial training produces highly realistic outputs. GANs were dominant for image generation before diffusion models.

Prompt Engineering

The discipline of designing and optimizing prompts to elicit desired outputs from generative AI models. Effective prompt engineering involves choosing the right instruction format, context, examples, and constraints to maximize output quality and reliability.

Latent Space

A compressed, abstract representation space learned by a generative model where data points with similar characteristics are grouped together. Interpolating in latent space produces smooth transitions between generated outputs (e.g., morphing between two faces).

smart_toy AI Agents & Automation

AI Agent

An AI system that autonomously perceives its environment, reasons about goals, and takes actions to achieve them — often through multiple steps and tool calls. Unlike a simple chatbot, an agent can plan, use external tools, and iterate based on intermediate results.

Tool Use / Function Calling

The ability of an LLM to invoke external functions or APIs as part of generating a response. The model decides when and how to call a tool (e.g., search, calculator, database query) and incorporates the result into its reasoning.

Agentic Workflow

A multi-step process where an AI agent autonomously plans, executes actions, observes results, and adjusts its approach to complete a complex goal. Agentic workflows may involve branching logic, parallel tasks, and human-in-the-loop checkpoints.

Orchestration

The coordination of multiple AI agents, tools, or services to accomplish a complex task. An orchestrator directs subagents, manages state, handles errors, and aggregates results. Frameworks like LangGraph and CrewAI provide orchestration primitives.

MCP (Model Context Protocol)

An open protocol developed by Anthropic that standardizes how AI models connect to external tools, data sources, and services. MCP allows LLMs to call tools (file systems, APIs, databases) in a consistent, secure, and composable way across different applications.

RAG (Retrieval-Augmented Generation)

A technique that enhances LLM responses by retrieving relevant documents from an external knowledge base and including them in the prompt context. RAG reduces hallucinations and enables LLMs to answer questions grounded in up-to-date, domain-specific information.

Vector Database

A database optimized for storing and querying high-dimensional embedding vectors. Used in RAG and semantic search to find documents or data points most similar to a query vector. Examples: Pinecone, Qdrant, Weaviate, pgvector.

Semantic Search

A search technique that finds results based on meaning and intent rather than exact keyword matches. Semantic search converts queries and documents into embeddings and retrieves the most semantically similar results.

Human-in-the-Loop (HITL)

An AI system design pattern where humans are integrated into the decision-making process at critical steps — reviewing, approving, or correcting AI outputs before they are acted upon. Essential for high-stakes automation workflows.

Agent Memory

The mechanism by which an AI agent retains information across turns or tasks. Types include in-context memory (within the context window), external memory (stored in a database and retrieved as needed), and episodic memory (records of past interactions).

translate Natural Language Processing (NLP)

Natural Language Processing (NLP)

A field of AI focused on enabling computers to understand, interpret, and generate human language. NLP tasks include text classification, named entity recognition, machine translation, summarization, and question answering.

Transformer

A neural network architecture introduced in the 2017 paper “Attention Is All You Need.” Transformers use self-attention mechanisms to process sequences in parallel, enabling training on large datasets. All modern LLMs (GPT, Claude, Llama, BERT) are built on the Transformer architecture.

Attention Mechanism

A component of the Transformer architecture that allows the model to weigh the importance of different tokens in a sequence when making predictions. Self-attention enables the model to capture long-range dependencies in text.

Named Entity Recognition (NER)

An NLP task that identifies and classifies named entities in text — people, organizations, locations, dates, and quantities. Used in information extraction, search, and document processing pipelines.

Sentiment Analysis

An NLP task that determines the emotional tone or opinion expressed in text (positive, negative, neutral). Widely used in customer feedback analysis, social media monitoring, and product review mining.

Text Classification

An NLP task that assigns predefined categories to text documents. Examples include spam detection (spam/not-spam), topic classification, and intent detection in chatbots.

Summarization

An NLP task that condenses a longer document into a shorter version retaining the most important information. Extractive summarization selects key sentences; abstractive summarization generates new text.

Machine Translation

The automated process of translating text from one human language to another using AI models. Modern neural machine translation (NMT) systems, built on Transformers, achieve near-human quality for major language pairs.

Chunking

The process of splitting a large document into smaller, overlapping or non-overlapping pieces (chunks) before embedding and indexing. Chunk size and overlap are important parameters in RAG pipelines that affect retrieval quality.

dns AI Infrastructure & MLOps

MLOps

A set of practices that combines machine learning, DevOps, and data engineering to deploy, monitor, and maintain ML models in production reliably and efficiently. MLOps covers the full lifecycle from data ingestion to model retirement.

Model Registry

A centralized repository for storing, versioning, and managing trained ML models and their metadata (metrics, parameters, artifacts). Enables reproducibility, auditability, and controlled model promotion to production. Examples: MLflow Registry, Hugging Face Hub.

Model Serving

The infrastructure and process of deploying a trained ML model so it can receive requests and return predictions in real time or batch. Common approaches include REST APIs, gRPC endpoints, and serverless functions.

Batch Inference

Running a model on a large dataset offline (not in real time), generating predictions for all records at once. Used for scheduled jobs like nightly scoring, report generation, and data enrichment pipelines.

Online Inference

Running a model in real time to generate predictions for individual requests as they arrive. Requires low-latency infrastructure and is used in production applications such as chatbots, recommendation engines, and fraud detection.

Model Drift

The degradation of a model’s performance over time as the real-world data distribution changes from the distribution it was trained on. Monitoring for model drift and triggering retraining are key MLOps practices.

Data Pipeline

An automated sequence of steps that ingests, transforms, validates, and delivers data for training or inference. Data pipelines ensure consistent, repeatable data preparation and are foundational to reliable ML systems.

Feature Store

A centralized platform for storing, sharing, and serving features for ML models. Feature stores decouple feature engineering from model training, ensure consistency between training and inference, and enable feature reuse across teams.

Experiment Tracking

The practice of logging parameters, metrics, artifacts, and code versions for each ML training run to enable comparison, reproducibility, and auditing. Tools like MLflow, Weights & Biases, and Neptune are commonly used.

GPU / TPU

Specialized hardware accelerators used for training and running AI models. GPUs (Graphics Processing Units, e.g., NVIDIA A100/H100) offer massive parallelism for matrix operations. TPUs (Tensor Processing Units) are Google’s custom AI chips. Both dramatically reduce training time compared to CPUs.

Quantization

A technique that reduces the precision of model weights (e.g., from 32-bit float to 8-bit or 4-bit integer) to decrease memory footprint and accelerate inference. Quantization enables running large models on consumer hardware with minimal quality loss.

balance AI Ethics & Governance

Responsible AI

A framework for developing and deploying AI systems in a way that is safe, fair, transparent, accountable, and aligned with human values. Responsible AI encompasses practices for bias mitigation, explainability, privacy protection, and oversight.

AI Bias

Systematic errors in AI model outputs that result from biased training data, flawed algorithms, or misaligned objectives. AI bias can lead to unfair or discriminatory outcomes, particularly affecting marginalized groups. Detecting and mitigating bias is a core concern in Responsible AI.

Fairness

The property of an AI system where its outputs do not systematically disadvantage individuals or groups based on protected characteristics (race, gender, age, etc.). Fairness is context-dependent and may involve trade-offs between different notions of equity.

Explainability (XAI)

The ability to describe the internal mechanisms or rationale behind an AI model’s predictions in human-understandable terms. Explainability is critical for trust, debugging, and regulatory compliance. Techniques include SHAP, LIME, and attention visualization.

AI Safety

The research field focused on ensuring AI systems behave as intended and do not cause unintended harm. AI safety covers robustness, reliability, alignment, and the long-term risks of advanced AI systems.

AI Alignment

The challenge of ensuring that AI systems pursue goals that are beneficial to humans and consistent with human values and intentions. Misaligned AI may optimize for proxy goals that diverge from true human preferences.

Model Card

A short document published alongside a trained ML model that documents its intended use cases, performance metrics across demographic groups, limitations, and ethical considerations. Model cards promote transparency and informed deployment.

Data Privacy

The protection of personally identifiable information (PII) used in AI training and inference. Data privacy practices include data anonymization, differential privacy, federated learning, and compliance with regulations such as GDPR and PDPA.

AI Governance

The set of policies, processes, and oversight mechanisms that organizations establish to ensure AI systems are developed and used responsibly, legally, and in alignment with business values. Includes model risk management, audit trails, and compliance frameworks.

Guardrails

Constraints and filters applied to LLM inputs and outputs to prevent harmful, inappropriate, or off-topic content. Guardrails may be implemented at the prompt level, via classifiers, or through Constitutional AI techniques.