AI Glossary Terms and Definitions
Artificial intelligence dictionary. Clear definitions of technical terms in the AI ecosystem.
A
Agent (AI Agent)
An AI system capable of autonomously performing tasks by making decisions, using tools, and interacting with external systems. Agents often combine LLMs with memory, planning, and tool execution.
AGI (Artificial General Intelligence)
A hypothetical AI system with cognitive abilities comparable to or exceeding those of humans across any intellectual task. Unlike today's narrow AI, AGI would be able to learn, reason, and adapt across domains.
AI (Artificial Intelligence)
A field of computer science dedicated to creating systems that perform tasks typically requiring human intelligence: learning, reasoning, perception, and language.
AI Agent
An artificial intelligence system capable of perceiving its environment, making decisions, and autonomously executing actions to achieve a goal. Modern AI agents combine large language models (LLMs) with external tools, memory, and planning capabilities.
AI Benchmark
A standardized set of tasks and evaluation metrics used to compare AI models objectively. Popular benchmarks include MMLU, HumanEval, GPQA, and Chatbot Arena.
AI Governance
The framework of policies, practices, and standards that guide the responsible development, deployment, and use of AI systems within organizations.
AI Hallucination
When a language model generates false, fabricated, or inaccurate information that appears convincing. Hallucinations are a common challenge in LLMs and require human verification in critical use cases.
Algorithm
A defined set of instructions or rules that a computer follows to solve a problem or perform a task. In AI, algorithms process data to identify patterns and make predictions.
Alignment
The process of ensuring that AI systems behave according to human values, intentions, and safety guidelines. Includes techniques like RLHF and policy constraints.
Anthropic
An AI safety and research company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei. Creator of the Claude family of models and a pioneer in AI alignment and mechanistic interpretability.
API (Application Programming Interface)
A software interface that enables different applications to communicate with one another. AI APIs, such as those provided by OpenAI and Anthropic, allow developers to integrate language models into software products.
API endpoints
Think of API endpoints as 'buttons' on the back of a piece of software that other programs can press to make it do things. Developers use these interfaces to build integrations. As AI agents grow more capable, they are increasingly able to find and use these endpoints on their own, opening up powerful possibilities for automation.
Attention Mechanism
A neural network technique that enables a model to focus on the most relevant parts of the input. It is the foundation of the Transformer architecture used by GPT, BERT, Claude, Gemini, and other modern LLMs.
Autoencoder
A neural network that learns to compress data into a compact representation and then reconstruct it. Common applications include dimensionality reduction, anomaly detection, and generative modeling.
B
Backpropagation
The core algorithm used to train neural networks. It computes how much each parameter contributes to prediction error and updates the network by propagating gradients backward.
Batch
A subset of training data processed in a single iteration during model training.
Benchmark
A standardized test or dataset used to evaluate and compare the performance of AI models across specific tasks.
BERT
Bidirectional Encoder Representations from Transformers. A Google language model that transformed natural language processing by introducing bidirectional contextual understanding.
Bias
Systematic distortions in AI model outputs, often originating from biased training data. AI bias can lead to unfair or discriminatory outcomes.
Bias (in AI)
Systematic errors or unfair tendencies in model outputs, often caused by imbalanced or unrepresentative training data.
Big Data
Extremely large and complex datasets that require specialized tools for storage and analysis. Commonly described by the five Vs: Volume, Velocity, Variety, Veracity, and Value.
C
Chain of Thought (CoT)
A prompting technique that encourages a language model to reason through intermediate steps before producing a final answer. It often improves performance on logical and mathematical tasks.
Chain-of-Thought (CoT)
A prompting technique that encourages models to generate intermediate reasoning steps before producing a final answer. Improves performance on complex reasoning tasks.
Chatbot
A software application designed to simulate human conversation through text or voice. Modern chatbots use AI to understand natural language and generate contextual responses.
ChatGPT
OpenAI's conversational AI assistant built on the GPT family of language models. It can generate text, answer questions, write code, analyze documents, and perform a wide range of language tasks.
Claude
Claude is Anthropic's family of AI assistants. It is recognized for strong reasoning, high-quality writing, long-context document analysis, and coding capabilities. The Claude lineup includes Haiku (fast), Sonnet (balanced), and Opus (highest capability). Anthropic is widely known for its work in AI safety and Constitutional AI.
CNN (Convolutional Neural Network)
A neural network architecture specialized for processing grid-like data such as images. CNNs use convolutional filters to detect features including edges, textures, and objects.
Coding agents
A specialized version of an AI agent applied to software development. Rather than simply suggesting code for a human to review, a coding agent can write, test, and debug code autonomously. Think of it like hiring a very fast intern who never sleeps — though a human still needs to review the work.
Compute
Compute generally refers to the vital computational power that allows AI models to operate. This type of processing fuels the AI industry. The term is often a shorthand for the kinds of hardware that provides the computational power — things like GPUs, CPUs, TPUs, and data center infrastructure.
Computer Vision
The field of AI focused on enabling computers to interpret and understand visual information. Applications include facial recognition, autonomous vehicles, industrial inspection, and medical imaging.
Context Window
The maximum number of tokens a language model can process in a single interaction. Larger context windows allow models to analyze longer documents and maintain more conversational history.
D
DALL·E
OpenAI's image generation model that creates images from natural language descriptions. Recent versions provide high-quality visual generation with strong prompt understanding.
Data Augmentation
A technique that artificially expands a training dataset through transformations such as rotation, cropping, scaling, or noise injection. It improves model robustness and generalization.
Data Privacy
The protection of sensitive information used in AI systems, ensuring compliance with regulations and safeguarding user data.
Data Science
An interdisciplinary field that combines statistics, programming, and domain expertise to extract insights from data. Modern data scientists increasingly work with machine learning, LLMs, retrieval systems, and generative AI alongside traditional analytics.
Dataset
A structured collection of data used for training, validating, and testing machine learning models.
Decoder
The part of a model (especially in sequence-to-sequence architectures) responsible for generating outputs based on encoded input representations.
Deep Learning
A subfield of machine learning based on multi-layer neural networks capable of learning hierarchical representations. Deep learning powers advances in computer vision, natural language processing, speech recognition, and generative AI.
DeepSeek
A Chinese AI research company and family of open-weight language models. DeepSeek gained worldwide attention for delivering highly competitive reasoning models with significantly lower training costs than many proprietary alternatives.
Diffusion Models
Generative models that create data (especially images) by learning to reverse a noise-adding process. Foundation of Stable Diffusion, Midjourney, and DALL-E 3.
E
Edge AI
The deployment of AI models directly on local devices (such as smartphones or IoT devices), reducing latency and dependence on cloud infrastructure.
ELO Score
A scoring system derived from chess, adapted to compare AI models based on head-to-head matchups evaluated by humans. Used by Chatbot Arena (LMSYS) to create human preference rankings.
Embedding
A numerical representation (vector) of data such as words, images, or concepts in a high-dimensional space. Captures semantic meaning and enables mathematical operations.
Encoder
The component of a model that processes input data and transforms it into an internal representation that the model can use.
Epoch
A complete pass through the training dataset during model training. Multiple epochs are required for the model to properly learn patterns.
Evaluation Metrics
Quantitative measures used to assess model performance, such as accuracy, precision, recall, F1 score, BLEU, or perplexity.
Explainability (XAI)
Techniques and methods that make AI model decisions more transparent and understandable to humans, especially in high-stakes applications.
F
Federated Learning
A training approach where models are trained across multiple decentralized devices or servers without sharing raw data, improving privacy.
Few-Shot Learning
A technique where models learn new tasks with only a few examples. LLMs demonstrate few-shot learning when they receive examples in the prompt to guide their responses.
Fine-Grained Control
Techniques that allow precise control over model outputs, such as system prompts, temperature tuning, constraints, and tool usage.
Fine-Tuning
Fine-tuning is the process of continuing the training of a pre-trained model using a specific dataset to adapt its behavior to a particular task or domain. Instead of training a model from scratch, which requires enormous computational resources, fine-tuning starts from a model that has already learned rich language representations and specializes it with relevant examples. For example: a generic LLM can be fine-tuned with Brazilian legal contracts to improve its accuracy in legal terminology; a customer service model can be fine-tuned with a company’s conversation history to adopt its tone and policies. The result is a customized model that combines the general knowledge of the base model with domain-specific expertise. Modern techniques such as LoRA and QLoRA have made fine-tuning more accessible: it is possible to adapt models with tens of billions of parameters on consumer GPUs in a few hours, using training datasets with hundreds to a few thousand examples. For Brazilian companies, fine-tuning is the difference between a generic assistant and a specialist in the product, market, and language of the business.
Foundation Model
An AI model trained on massive datasets that serves as a base for multiple applications. Examples: GPT-4, Claude, Llama. Can be adapted to specific tasks.
Function Calling
The ability of an LLM to identify when it should call an external function and generate the necessary parameters in structured format (JSON). Enables integration with APIs, databases, and real-time tools.
G
GAN (Generative Adversarial Network)
An architecture with two competing neural networks: a generator creates fake data while a discriminator tries to distinguish it from real data. Used to generate realistic images.
Gemini
Gemini is a family of multimodal language models developed by Google DeepMind, launched in December 2023 as a direct response to ChatGPT. Gemini models are natively multimodal, trained jointly on text, images, audio, and video from the start, unlike approaches that add modalities separately. The family includes versions such as Flash (fast and cost-efficient), Pro (balanced), and Ultra/Exp (maximum capability). A key technical advantage is the massive context window, reaching up to 2 million tokens in recent versions, allowing it to process entire books, code repositories, and long recordings in a single session. Gemini is integrated into the Google ecosystem: available in Google AI Studio, Workspace (Docs, Sheets, Gmail), Android, and Search. For Brazilian users already using Google Workspace, Gemini Advanced (included in Google One AI Premium) is the most natural integration. Gemini Flash stands out as the best cost-benefit model per token among major providers, ideal for high-volume applications.
Generative AI
AI systems capable of creating new content: text, images, code, music, videos. Includes LLMs (ChatGPT), image models (Midjourney), and audio (Suno).
GPT (Generative Pre-trained Transformer)
A language model architecture by OpenAI trained to predict the next word. GPT-4 and GPT-5 are the most advanced models, capable of complex reasoning.
Gradient
A vector that represents the direction and magnitude of change needed to minimize the model’s error during training.
Gradient Descent
An optimization algorithm used to minimize the error function by adjusting model parameters. It iteratively moves in the direction of greatest error reduction.
Grok
A language model from xAI (Elon Musk’s company), integrated into X (Twitter). Known for its irreverent tone and access to real-time information from the platform.
Grounding
A technique that anchors LLM responses in verifiable and up-to-date information sources, such as search results or specific documents. Reduces hallucinations by connecting the model to real-world data.
H
Hallucination
Hallucination is the phenomenon in which a language model generates incorrect, fabricated, or unfounded information with apparent confidence. The model does not “know” it is wrong—it produces the most statistically likely text regardless of truth. Common examples include citing nonexistent scientific papers, inventing dates and real people, or describing events that never happened. Causes are multiple: gaps in training data, ambiguity in the user’s question, or the probabilistic generation mechanism itself. To minimize hallucinations, techniques such as RAG (Retrieval-Augmented Generation), which grounds the model in real documents, and RLHF, which penalizes incorrect responses during training, are used. In the Brazilian business context, hallucination is a critical risk in legal, medical, and financial applications, where fabricated information can cause real harm. Human verification remains essential in any critical workflow using generative AI.
Hugging Face
A leading platform for sharing AI models, datasets, and tools. It is the hub of the open-source ML community, with the widely used Transformers library.
Human-in-the-Loop (HITL)
A system design approach where human feedback is incorporated into the AI workflow for validation, correction, or decision-making.
Hyperparameters
Configuration settings defined before training (such as learning rate or batch size) that influence how a model learns.
I
In-Context Learning
The ability of an LLM to learn a new pattern or task from examples provided directly in the prompt, without updating the model’s weights. The basis of few-shot prompting.
Inference
The process of using a trained model to make predictions on new data. Unlike training, inference does not update the model’s weights.
Inference Optimization
Techniques used to improve the speed and efficiency of model inference, including quantization, pruning, batching, and hardware acceleration.
K
Knowledge Cutoff
The point in time after which a model no longer has training data. Information beyond this date may be incomplete or unavailable unless external tools are used.
L
LangChain
An open-source framework in Python and JavaScript for building applications with LLMs. Provides abstractions for chains, agents, memory, document retrieval, and integrations with dozens of AI providers.
Latency
The time it takes for a model to generate a response after receiving input. Critical for real-time applications.
Latent Space
Latent Space is the compressed internal representation that an AI model learns to encode complex data. Instead of working directly with image pixels or text words, the model creates a lower-dimensional coordinate map where similar examples are close together. For example, in an image generation model, the latent space might have 512 dimensions, where each point corresponds to a possible image. Faces might cluster in one region; landscapes in another. The power of latent space lies in continuity: moving gradually between two points generates coherent intermediate examples (interpolation), which would not be possible with raw data. Diffusion models like Stable Diffusion operate in latent space (latent diffusion), making them more efficient than models that operate directly on pixels. The concept is central to autoencoders, VAEs, GANs, and diffusion models. In LLMs, the internal representations of intermediate Transformer layers form a latent space of language.
LLM (Large Language Model)
A type of AI model trained on massive amounts of text data to understand and generate human-like language. Examples include GPT, Claude, and Llama. LLMs power chatbots, assistants, and many generative AI applications.
LoRA (Low-Rank Adaptation)
A parameter-efficient fine-tuning technique that adapts large models by training only small, low-rank matrices instead of updating all parameters. Significantly reduces computational cost and hardware requirements.
Loss Function
A mathematical function that measures how far a model’s predictions are from the actual values. The goal of training is to minimize this loss.
M
Machine Learning (ML)
A subset of artificial intelligence focused on building systems that learn patterns from data and improve over time without being explicitly programmed for each task.
Memory (in AI Systems)
Mechanisms that allow models or agents to retain and use information across interactions, enabling more coherent and context-aware behavior.
Memory cache
Memory cache refers to an important process that boosts inference. Caching is an optimization technique designed to cut down on the number of calculations a model might have to run by saving particular calculations for future user queries. KV (key-value) caching works in transformer-based models to increase efficiency and speed.
Midjourney
A generative AI model specialized in creating high-quality images from text prompts. Known for its strong artistic style and widely used in design, marketing, and creative industries.
Mixture of Experts (MoE)
A model architecture that splits a neural network into many smaller specialized sub-networks, or 'experts,' and only activates a handful of them for any given task. An MoE model has a built-in router that picks just the right specialists for the job, making it possible to build enormous models that stay relatively fast and cheap.
Model Context Protocol (MCP)
Model Context Protocol, or MCP, is an open standard that lets AI models connect to outside tools and data — your files, databases, or apps — without a developer building a custom connector for every single pairing. Think of it as a USB-C port for AI.
Model Distillation
A technique where a smaller model (student) is trained to replicate the behavior of a larger, more complex model (teacher). The goal is to retain performance while improving efficiency, speed, and cost.
Model Quantization
A process that reduces the precision of a model’s numerical parameters (e.g., from 32-bit to 8-bit), making it smaller and faster to run, often with minimal impact on accuracy.
Multimodal
Refers to AI models that can process and generate multiple types of data—such as text, images, audio, and video—within a single system.
N
Neural Network
A computational model inspired by the human brain, composed of interconnected nodes (neurons) organized in layers. Neural networks are the foundation of modern AI systems.
O
Open Source Models
AI models whose weights and architecture are publicly available for use, modification, and distribution. Examples include Llama, Mistral, and Stable Diffusion.
Orchestration
The coordination of multiple AI components—such as models, tools, databases, and APIs—into a unified workflow or system.
Overfitting
A modeling error that occurs when a model learns the training data too closely—including noise and irrelevant details—resulting in poor generalization to new data.
P
Parallelization
Parallelization means doing many things at the same time instead of one after another. In AI, parallelization is fundamental to both training and inference: modern GPUs are specifically designed to perform thousands of calculations in parallel, which is why they became the hardware backbone of the industry.
Parameters
The internal numerical values of a model that are learned during training. In LLMs, parameters can range from millions to hundreds of billions and determine the model’s capabilities.
Prompt
The input given to an AI model (usually text) that guides its response. Prompt design is critical to controlling output quality and behavior.
Prompt Engineering
The practice of designing and refining prompts to achieve more accurate, relevant, or structured outputs from AI models.
Prompt Template
A reusable structured format for prompts that ensures consistency and reliability across different inputs and use cases.
Q
QLoRA
An extension of LoRA that uses quantization to further reduce memory usage, enabling fine-tuning of very large models on consumer-grade GPUs without significant loss in performance.
R
RAG (Retrieval-Augmented Generation)
A technique that combines language models with external data retrieval. The model first retrieves relevant documents, then generates responses grounded in that information, improving accuracy and reducing hallucinations.
RAMageddon
RAMageddon is a term for the ever-increasing shortage of random access memory (RAM) chips. As the AI industry has blossomed, big tech companies are buying so much RAM for their data centers that there’s a supply bottleneck for consumer electronics, gaming, and general enterprise computing.
Recursive self-improvement
In the RSI scenario, AI models start improving themselves without human intervention, leading to a huge acceleration in capabilities and autonomy. RSI describes a basic capability — can an AI model design its own successor? — representing the next frontier for AI research and development.
Reinforcement Learning
A machine learning paradigm in which an agent learns through trial and error by receiving rewards or penalties. Commonly used in robotics, games, and recommendation systems.
Reinforcement Learning (RL)
A training approach where an agent learns by interacting with an environment and receiving rewards or penalties based on its actions.
Retrieval
The process of searching and fetching relevant information from external sources, often used in RAG systems.
RLHF (Reinforcement Learning from Human Feedback)
A method for aligning AI models with human preferences by training them using human evaluations. Widely used to improve helpfulness, safety, and reliability in LLMs.
S
Safety Guardrails
Rules, filters, and constraints applied to AI systems to prevent harmful, unsafe, or undesired outputs.
Scalability
The ability of an AI system to handle increasing workloads, users, or data efficiently without performance degradation.
Stable Diffusion
An open-source diffusion model for generating images from text. Known for its flexibility, customization, and ability to run locally.
Synthetic Data
Artificially generated data used to train models when real data is scarce, sensitive, or expensive to obtain.
System Prompt
A high-level instruction that defines the behavior, tone, and constraints of an AI model during a session.
T
Temperature
A parameter that controls randomness in model outputs. Lower values produce more deterministic responses; higher values increase creativity and variability.
Token
A unit of text processed by a model. Tokens can be words, subwords, or characters. Pricing, context limits, and performance are often measured in tokens.
Token throughput
Throughput refers to how much can be processed in a given period of time, so token throughput is essentially a measure of how much AI work a system can handle at once. High token throughput is a key goal for AI infrastructure teams to serve users simultaneously and quickly.
Tool Use (Tool Calling)
The ability of an AI model to interact with external tools, APIs, or systems to perform actions beyond text generation—such as retrieving data, executing code, or triggering workflows.
Training
The process of teaching an AI model by exposing it to large datasets and adjusting its parameters to learn patterns and relationships.
Training Data
The data used to train a model. The quality, diversity, and size of the training data directly impact model performance and bias.
Transfer Learning
A technique where a model trained on one task is reused and adapted for a different but related task, reducing training time and resource requirements.
Transformer
A neural network architecture introduced in the paper “Attention Is All You Need.” It uses self-attention mechanisms to process sequences efficiently and is the foundation of modern LLMs.
U
Unstructured Data
Data that does not follow a predefined format, such as text, images, audio, and video. Most real-world data is unstructured.
Use Case
A specific application or scenario in which an AI system delivers value, such as customer support automation, fraud detection, or content generation.
V
Validation loss
Validation loss is a number that tells you how well an AI model is learning during training — and lower is better. Researchers track it closely as a real-time report card. It helps flag overfitting, a condition in which a model memorizes its training data rather than truly learning patterns.
Validation Set
A subset of data used during training to evaluate model performance and tune hyperparameters, helping prevent overfitting.
Vector Database
A specialized database designed to store and search embeddings (vectors). Enables semantic search and is a key component in RAG systems.
Vector Search
A search method that finds information based on semantic similarity rather than exact keyword matches, using embeddings stored in vector databases.
W
Weights
The learned numerical values within a neural network that determine how inputs are transformed into outputs. Adjusted during training.
Workflow Automation
The use of AI and software systems to automate sequences of tasks, reducing manual effort and increasing efficiency.
Z
Zero-Shot Learning
The ability of a model to perform a task without having seen explicit examples during training or prompting, relying solely on prior knowledge.
Zero-Shot Prompting
A prompting approach where the model is asked to perform a task without providing examples, relying entirely on its prior training.