AI glossary
Public SeekBox information. Choose a topic for its full explanation and update date.
Agentic AI
AI systems that can autonomously plan, use tools, and execute multi-step tasks with minimal human supervision, going beyond simple question-answering.
AI Alignment
The research field focused on ensuring AI systems act in accordance with human intentions, values, and ethical principles, especially as systems become more capable.
API (Application Programming Interface)
A set of protocols and tools that allow software applications to communicate with AI services programmatically, enabling integration of AI capabilities into products.
Artificial Intelligence (AI)
Systems that perform tasks requiring human-like cognition, using techniques like machine learning, natural language processing, and computer vision.
Attention Mechanism
A mechanism in neural networks that computes weighted relevance scores between elements of a sequence, allowing the model to focus on the most pertinent information for each output.
Benchmark
Standardized evaluation datasets and metrics used to compare AI model performance across tasks like reasoning, coding, math, and language understanding.
Bias (in AI)
Systematic errors in AI outputs that reflect prejudices present in training data, model design, or evaluation criteria, potentially leading to unfair outcomes.
Chain-of-Thought (CoT)
A prompting strategy where the model is encouraged to produce intermediate reasoning steps before arriving at a final answer, improving performance on complex tasks.
Chatbot
A conversational AI interface that processes user messages and generates contextually appropriate responses, often powered by LLMs with dialogue-specific fine-tuning.
Computer Vision
A subfield of AI focused on enabling machines to interpret visual data from the world, including object detection, image classification, and scene understanding.
Context Window
The maximum number of tokens an LLM can process in a single inference call. Modern models support windows from 8K to over 1M tokens.
Data Labeling
The process of annotating data with ground-truth labels used for supervised learning, often requiring human annotators to classify, tag, or rate examples.
Deep Learning
A subset of machine learning using multi-layered neural networks to learn hierarchical representations of data, enabling breakthroughs in vision, NLP, and generative tasks.
Diffusion Model
A generative model that creates images (or other data) by learning to reverse a gradual noising process, producing high-quality outputs from random noise.
Edge AI
Deploying AI models on local devices (phones, IoT, embedded systems) rather than cloud servers, enabling low-latency inference and data privacy.
Embedding
A dense vector representation of text in a high-dimensional space where semantic similarity corresponds to geometric proximity. Used for search, clustering, and recommendation.
Emergent Behavior
Capabilities that appear in large models but are absent in smaller ones, arising from scale rather than explicit training — such as in-context learning or chain-of-thought reasoning.
Few-Shot Learning
A technique where a model learns to perform a task from a small number of examples provided in the prompt, without updating its parameters.
Fine-Tuning
The process of further training a pre-trained model on a smaller, task-specific dataset to adapt it for particular use cases.
Foundation Model
A large-scale model trained on broad data that can be adapted to a wide range of downstream tasks through fine-tuning, prompting, or other techniques.
GAN (Generative Adversarial Network)
A generative model architecture consisting of a generator and discriminator that are trained adversarially — the generator creates samples while the discriminator distinguishes real from generated.
GPT (Generative Pre-trained Transformer)
OpenAI's family of autoregressive language models, trained on large text corpora then fine-tuned with RLHF. GPT-4o is their flagship multimodal model.
GPU (Graphics Processing Unit)
A parallel processing chip optimized for matrix operations, making it ideal for training and running neural networks. NVIDIA's H100 is the dominant AI GPU.
Grounding
The practice of anchoring AI outputs in verifiable external data sources (documents, databases, search results) to reduce hallucination and improve factual accuracy.
Guardrails
Safety mechanisms implemented in AI systems to prevent harmful, biased, or policy-violating outputs, including content filters, system prompts, and output classifiers.
Hallucination
A phenomenon where LLMs generate plausible-sounding but factually incorrect or fabricated information, often with high confidence.
Inference
The phase where a trained model processes new inputs to generate predictions or outputs, as opposed to the training phase where it learns from data.
Knowledge Distillation
A training technique where a smaller "student" model learns to mimic the outputs of a larger "teacher" model, achieving competitive performance with fewer parameters.
Large Language Model (LLM)
A neural network with billions of parameters trained on massive text corpora to understand and generate human language. Examples include GPT-4, Claude, and Gemini.
Latent Space
A compressed, abstract representation space learned by a model where input data is encoded as vectors, capturing meaningful structure and relationships.
Machine Learning (ML)
A subset of AI where algorithms improve their performance on tasks by learning from data rather than being explicitly programmed.
MCP (Model Context Protocol)
An open protocol that standardizes how AI applications connect to external data sources and tools, enabling models to access real-time information and take actions beyond text generation.
Multimodal AI
AI models capable of processing and generating multiple types of data — text, images, audio, and video — within a single system.
Natural Language Processing (NLP)
A field of AI focused on the interaction between computers and human language, encompassing tasks like translation, summarization, sentiment analysis, and question answering.
Neural Network
A computational model composed of layers of interconnected nodes (neurons) that process data through weighted connections, learning to approximate complex functions.
Open Source vs. Closed Source
Open-source models release their weights and often their training code publicly (e.g., Llama, Mistral), while closed-source models (e.g., GPT-4, Claude) are only accessible via API.
Overfitting
A condition where a model learns patterns specific to its training data that don't generalize to unseen data, resulting in high training accuracy but poor real-world performance.
Parameter
The learnable weights and biases in a neural network that are adjusted during training. Model size is often expressed in parameter count (e.g., 7B, 70B, 405B).
Prompt
The input text given to a language model that frames the task, provides context, and guides the model's output. Prompt engineering is the practice of crafting effective prompts.
Quantization
A model compression technique that reduces the numerical precision of weights and activations (e.g., from 32-bit to 4-bit), decreasing memory usage and speeding up inference.
RAG (Retrieval-Augmented Generation)
An architecture that combines a retrieval system with an LLM — the model first retrieves relevant documents, then generates answers grounded in that retrieved context.
Reinforcement Learning from Human Feedback (RLHF)
A training technique where a reward model trained on human preference data is used to fine-tune an LLM via reinforcement learning, aligning it with human values.
Scaling Laws
Empirical relationships showing that model performance improves predictably as a power law of compute, parameters, and data, guiding resource allocation for AI training.
Sentiment Analysis
An NLP task where models classify text by emotional tone or opinion polarity, commonly used for analyzing customer reviews, social media, and brand perception.
Synthetic Data
Artificially generated data used for training models when real-world data is scarce, expensive, or raises privacy concerns. Can be created by other AI models.
Temperature
A parameter (typically 0–2) that controls the randomness of an LLM's output by scaling the logits before sampling. Lower values make output more deterministic.
Token
The basic unit of text that LLMs process. Text is split into tokens (subwords) by a tokenizer, and the model predicts the next token in a sequence.
Training Data
The dataset used to train a machine learning model. For LLMs, this typically includes web pages, books, code, and other text corpora totaling billions of tokens.
Transfer Learning
A technique where knowledge learned from one task or domain is applied to a different but related task, reducing the data and compute needed for the new task.
Transformer
A neural network architecture that uses self-attention mechanisms to process sequential data in parallel, forming the foundation of most modern LLMs.
Zero-Shot Learning
A model's ability to perform tasks it wasn't explicitly trained on, using only a natural language description of the task without any demonstration examples.
Sources
- NIST Trustworthy and Responsible AI Resource Center: Glossary — National Institute of Standards and Technology. Accessed 2026-07-20.