Open AI glossary for companies

GlossarIA — Open AI glossary

A curated, bilingual glossary of AI concepts for teams and companies.

Controlled comparison of two versions of a model or system to see which performs better.

experimentos a_b

Obligation to assign and track who is responsible for AI systems' decisions and errors.

responsabilidad etica gobernanza

Activation function

Mathematical function applied to the output of an artificial neuron that determines if it should be activated, introducing non-linearity to the network.

redes matematicas neurona

Adversarial attacks

Techniques that manipulate inputs in ways imperceptible to humans but that fool the model.

seguridad ataques robustez

Agent orchestrator

System that coordinates multiple specialized models or tools to solve complex tasks autonomously.

agentes flujo automatizacion

AI systems capable of planning, reasoning and executing complex tasks autonomously with minimal human supervision.

agentes autonomia razonamiento

Agentic Workflow

Design pattern where the AI iterates, searches for tools, and reflects to complete a task.

agentes procesos automatizacion

Interface that allows integrating AI models into external applications and systems.

api integracion backend

Set of libraries and tools for developing, training and deploying AI models.

framework desarrollo

Framework of policies, processes and controls to manage responsible AI development and use in the organization.

gobernanza estrategia control

AI impact assessment

Prior analysis of risks and effects before deploying an AI system.

impacto riesgo evaluacion

Set of steps that define a model life cycle, from data collection to deployment.

Low-quality, repetitive, generic or valueless content mass-produced by AI that floods websites, social media and documents.

calidad generativa contenido etica

Algorithmic bias

Distortion in model results caused by biased data or design choices.

Algorithmic transparency

Degree to which the functioning of an algorithmic decision system can be explained and understood.

transparencia modelo explicacion

Process of ensuring AI system goals and behaviors match human values and intentions.

etica alineacion valores

Centralized entry point that manages, routes and controls access to multiple AI APIs and backend services.

api arquitectura integracion

Artificial General Intelligence (AGI)

Hypothetical AI system capable of understanding, learning and applying knowledge across any intellectual task at or above human level.

agi inteligencia_general futuro

Artificial intelligence (AI)

Field that aims to make machines perform tasks that normally require human intelligence, such as learning, reasoning or communicating.

fundamentos empresa

Attention mechanism

Technique that allows the model to focus on the most relevant parts of the input when processing information, fundamental in transformer architectures.

fundamentos transformer arquitectura

Audio generator

Model that synthesizes voice or music from text or parameters.

audio voz sintesis

Autonomous agent

Model or system that chains actions to perform complex tasks with little human supervision.

agentes automatizacion

Backpropagation

Fundamental algorithm that calculates the gradient of the loss function with respect to neural network weights, enabling learning.

fundamentos entrenamiento optimizacion

Batch inference

Process of running predictions on multiple examples simultaneously instead of one-by-one, optimizing resources and time.

inferencia produccion optimizacion

Number of training examples processed simultaneously before updating model weights.

fundamentos entrenamiento parametro

Evaluating models by comparing their results with standard metrics or reference datasets.

metricas comparacion

Metrics to evaluate generated text quality by comparing with references: BLEU for translation, ROUGE for summaries.

metricas evaluacion texto

Catastrophic forgetting

Phenomenon where a neural network completely forgets what it learned in a previous task upon being trained on a new task.

entrenamiento memoria problema

Chain-of-Thought (CoT)

Prompting technique that forces the model to reason step-by-step before giving the final answer, dramatically improving accuracy on complex tasks.

prompt razonamiento generativa

Virtual assistant that holds automated conversations with users.

chatbot soporte conversacion

Unsupervised learning technique that groups similar data into clusters without prior labels, revealing hidden patterns.

no_supervisado agrupacion segmentacion

Computer Vision

AI area that enables machines to interpret and understand images and videos.

vision imagen analisis

Confusion matrix

Table showing true positives, false positives, true negatives and false negatives of a classifier.

metricas evaluacion clasificacion

Constitutional AI

Anthropic's approach to align models using explicit ethical principles instead of only human feedback.

etica alineacion seguridad

Additional information provided to the model (usually at the beginning of the conversation or in the system prompt) so it understands the role, domain, rules or specific reality in which it must operate.

prompt contexto dominio personalizacion

Maximum number of tokens (words or subwords) that a language model can process and keep in memory in a single interaction.

llm token memoria contexto

Convolutional Neural Network (CNN)

Neural network architecture designed to process grid-structured data like images using convolutions and pooling.

arquitectura vision imagen

Price charged by LLM providers for each processed token (input + output).

coste facturacion token

Cross-Validation

Method to evaluate model performance by splitting data into subsets and rotating between training and testing.

evaluacion modelo validacion

Data augmentation

Generating additional or synthetic data to improve model training.

datos aumento sintetico

Process of removing errors, duplicates and inconsistent values from data.

Gradual change in the distribution of real-world data compared to training data, causing performance degradation.

monitorizacion calidad produccion

Data governance

Strategic management of the data lifecycle in an organization, including quality, access and security.

gobernanza datos estrategia

Process of annotating data with correct labels to train supervised models.

datos anotacion supervisado

Centralized repository where raw data from multiple sources is stored.

datos almacen nube

Structured repository optimized for analytical queries that stores processed, clean and organized data from multiple sources.

datos almacen analitica

Structured collection of data used to train AI models.

datos fundamentos

Technique based on neural networks with many layers that can process complex data such as images, speech or text.

fundamentos redes vision

Process of putting a trained AI model into production so it can be used by real applications, users or systems.

mlops produccion despliegue

Differential privacy

Technique that protects individual identity by adding controlled noise to aggregated data.

privacidad datos anonimizacion

Diffusion models

Generative architecture that creates images by progressively adding noise and then learning to remove it.

arquitectura generativa imagen

Set of principles and rules that protect people in the digital environment, including AI systems.

derechos ciudadania datos

Regularization technique that randomly deactivates neurons during training to prevent overfitting, forcing the model to be more robust.

regularizacion entrenamiento overfitting

Running AI models directly on local devices without always relying on the cloud.

edge dispositivos iot

Numeric representation of text, image or audio that allows measuring semantic similarity.

vector busqueda semantica

Emergent abilities

Abilities that appear in large models but not in smaller versions, without being explicitly trained.

llm escalado capacidades

Ensemble Methods

Techniques that combine multiple models to improve accuracy and robustness, such as Random Forest or Boosting.

ensamble modelo combinacion

One complete pass of the model through the entire training dataset.

fundamentos entrenamiento proceso

ETL (Extract-Transform-Load)

Process of extracting data from multiple sources, transforming it to a common format and loading it into a destination for analysis or training.

datos pipeline integracion

European Union legal framework that classifies AI systems by risk levels and defines obligations.

regulacion ue ley

Ability of a model to justify or explain its decisions in an understandable way.

explicabilidad modelo

Ethical principle ensuring AI models do not discriminate against groups based on protected attributes like gender or ethnicity.

equidad etica sesgo

Variable or attribute in the data used to train a model.

Feature engineering

Selection and transformation of variables that improve model learning.

features diseno

Centralized repository that stores, documents and serves features for model training and inference.

mlops datos features

Federated Learning

Distributed training technique where the model learns from decentralized data without it leaving the original devices, preserving privacy.

distribuido privacidad entrenamiento

Few-shot learning

Learning from only a small number of labeled examples.

pocos_datos ejemplos

Adapting a base model with specific data for a concrete use case.

ajuste modelo dominio

Future of work with AI

Changes in professional roles and skills driven by the adoption of artificial intelligence.

futuro_trabajo competencias empresa

General Data Protection Regulation (GDPR)

EU legal framework for personal data protection, applicable to AI systems processing sensitive information.

regulacion privacidad ue

Generative Adversarial Network (GAN)

Model consisting of a generator and discriminator that compete to create realistic data, such as synthetic images.

arquitectura generativa adversarial

Gradient Descent

Optimization algorithm that iteratively adjusts model parameters to minimize the loss function.

fundamentos optimizacion entrenamiento

Approach that prioritizes energy efficiency and carbon footprint reduction in AI projects.

sostenibilidad energia huella

Technique to anchor model responses to real, up-to-date data (usually via RAG) to prevent hallucinations.

rag precision actualizacion

Set of rules, validations and controls that limit and guide AI model behavior to prevent dangerous or inappropriate outputs.

seguridad control validacion

Phenomenon in generative models where plausible but false or invented responses are produced.

generativa error llm

Human-in-the-loop

Approach in which a person reviews, corrects or validates the model's decisions.

human_in_the_loop revision

Hyperparameters

Parameters set manually before training that control the model's learning process.

fundamentos modelo ajuste

Image generator

Model that creates images from natural language descriptions.

imagen generativa creatividad

Imbalanced data

Situation where some classes have many more examples than others, causing model bias.

datos calidad sesgo

Use of a trained model to generate results or predictions.

proceso produccion

Integration of AI into connected devices for real-time analysis and automation.

iot dispositivos integracion

Techniques to bypass an LLM's safety restrictions and make it generate prohibited content.

seguridad prompt vulnerabilidad

Knowledge Graph

Data structure that represents entities and their relationships as a graph, enabling complex reasoning and semantic searches.

grafo conocimiento relaciones

Language model (LLM)

Algorithm trained on large volumes of text to generate natural language.

texto generativa empresa

Time the model takes to generate a response after receiving the request.

rendimiento produccion tiempo

Abstract and compressed mathematical representation of data, where nearby points represent semantically similar concepts.

representacion matematicas generativa

LoRA (Low-Rank Adaptation)

Efficient fine-tuning technique that trains only low-rank matrices instead of all model parameters.

fine_tuning eficiencia optimizacion

Mathematical function that measures the difference between model predictions and actual values, guiding training.

fundamentos entrenamiento optimizacion

Low-code / No-code AI

Platforms that allow building AI solutions with little or no code.

low_code no_code automatizacion

Machine learning

Subfield of AI that trains algorithms to learn from data without being explicitly programmed.

fundamentos modelo empresa

MCP (Model Context Protocol)

Open protocol that standardizes how AI models connect with data sources and tools.

estandar datos herramientas

Mixture of Experts (MoE)

Architecture that divides the model into multiple specialized experts, activating only the relevant ones for each input.

arquitectura escalabilidad eficiencia

Practice that combines development and operations to manage AI models in production.

mlops devops produccion

Phenomenon where models trained on data generated by other AI models progressively degrade in quality, losing diversity and accuracy.

calidad degradacion sintetico

Model distillation

Technique to transfer knowledge from a large expensive model to a smaller faster one while keeping nearly the same accuracy.

optimizacion eficiencia modelo

Model evaluation

Measuring model accuracy and quality using metrics such as accuracy, recall or F1-score.

modelo metricas

Model monitoring

Continuous tracking of performance, quality and behavior of models in production.

mlops produccion calidad

Model poisoning

Attack that corrupts training data so the model learns malicious behaviors.

seguridad ataques datos

Centralized repository that stores, versions and manages metadata for all ML models in the organization.

mlops gestion versionado

Measures to protect AI models from data leaks, adversarial attacks and misuse.

seguridad ataques proteccion

Model traceability

Ability to know how, with which data and with which versions a model was trained and changed.

trazabilidad auditoria

Model versioning

Practice of maintaining records of different model versions with their changes and metadata.

mlops gestion control

Multiagent systems

Architecture where multiple AI agents collaborate, negotiate or compete to solve complex problems.

agentes colaboracion coordinacion

Ability of a model to understand and combine different data types such as text, image or audio.

multimodal imagen texto

Named Entity Recognition (NER)

NLP task that identifies and classifies entities in text such as people, organizations, locations or dates.

nlp texto extraccion

Natural Language Processing (NLP)

AI subfield focused on interactions between computers and human language, including text analysis and translation.

texto nlp lenguaje

Natural Language Understanding (NLU)

NLP subfield focused on making the machine understand meaning, intent, and entities in human text.

texto nlu entendimiento

Mathematical model inspired by the human brain, made of interconnected nodes that learn relationships between data.

fundamentos redes

Adjusting data values so they follow a common comparable scale.

OCR (Optical Character Recognition)

Technology that converts images of text into editable digital text.

vision texto documento

Open-source model

AI model whose weights and architecture are public and can be freely downloaded and modified (Llama-3, Mistral, Gemma, etc.).

modelo open_source licencia

When a model learns the training data too well and fails to generalize to new data.

Internal model variables (weights and biases) that are automatically adjusted during training from the data.

pesos entrenamiento modelo

PEFT (Parameter-Efficient Fine-Tuning)

Family of techniques that allow fine-tuning large models by modifying only a small fraction of their parameters.

fine_tuning eficiencia optimizacion

Metric that measures how surprised a language model is by new data; lower values indicate better performance.

metricas evaluacion llm

Personally Identifiable Information (PII)

Information that can be used to directly identify a person (name, ID, email, phone, etc.).

privacidad datos cumplimiento

Initial phase where a language model is trained on massive amounts of public internet text to learn general language patterns and world knowledge.

entrenamiento llm fundacion

Precision and Recall

Classification metrics: precision measures accuracy of positive predictions, recall measures ability to find all positives.

metricas evaluacion clasificacion

Instruction or text that guides the model's response.

prompt instrucciones

Prompt engineering

Design and optimization of prompts to improve generative model outputs.

prompt diseno plantillas

Prompt Injection

Vulnerability where a malicious user manipulates inputs to alter model behavior.

seguridad prompt ataque

Optimization technique that removes unimportant connections or neurons from the model to reduce size and speed up inference.

optimizacion eficiencia compresion

Reduction of numerical precision of model weights (from 32-bit to 8-bit or 4-bit) to use less memory and run faster.

optimizacion hardware eficiencia

RAG (Retrieval-Augmented Generation)

Technique that combines information retrieval with text generation.

rag busqueda generativa

Reasoning Model

AI model designed to 'think' before responding, breaking down complex problems into logical steps.

razonamiento llm complejidad

Recommendation System

Tool that suggests items or content based on user preferences and historical data.

recomendacion personalizacion

Recurrent Neural Network (RNN)

Architecture for processing sequential data, maintaining memory of previous inputs; includes variants like LSTM for long dependencies.

arquitectura secuencias tiempo

Process of intentionally testing an AI model with malicious or tricky prompts to uncover vulnerabilities and improve robustness.

seguridad pruebas ataques etica

Reinforcement learning

Model learns by trial and error, receiving rewards for correct decisions.

refuerzo politicas

Ethical, transparent and safe use of artificial intelligence in products and processes.

etica riesgo gobernanza

Process of searching and extracting relevant information from a knowledge base before generating a response.

rag busqueda informacion

RLAIF (Reinforcement Learning from AI Feedback)

Variant of RLHF where feedback to train the model comes from another AI system instead of humans, scaling the alignment process.

rlhf feedback_ia alineacion

RLHF (Reinforcement Learning from Human Feedback)

Training models with human feedback to improve their responses.

rlhf feedback_humano

Additional rules and models that block or modify LLM responses to prevent harmful, illegal or inappropriate content.

seguridad moderacion contenido

Semantic search

Search technique that understands the meaning and intent of the query, rather than just matching literal keywords.

busqueda embeddings nlp

Sentiment analysis

NLP technique that identifies emotions or attitudes (positive, negative, neutral) in text.

nlp texto analisis

Unauthorized use of AI tools (ChatGPT, Gemini, etc.) by employees without IT or compliance approval.

gobernanza riesgo control

A specific function or capability that an AI agent can execute to perform a task.

agentes herramientas capacidades

Speech-to-text / Text-to-speech

Converting speech to text (STT) and text to speech (TTS) using AI models.

audio voz transcripcion

Supervised learning

Model learns from labeled data where the correct answer is known.

supervisado etiquetas

Artificially generated data that mimics the statistical properties of real data, used when real data is scarce or to protect privacy.

datos sinteticos privacidad

Fixed initial instruction that defines the model’s role, tone, rules and personality throughout the entire conversation.

prompt contexto personalizacion

Hyperparameter that controls randomness in model responses (0 = deterministic, 1 = very creative).

parametro creatividad control

Model that writes text based on instructions or prompts.

texto generativa contenido

Smallest unit of text processed by a model, which can be a word or a fragment.

texto coste tokens

Process of splitting text into smaller units (tokens) that the model can process, such as words or subwords.

texto procesamiento nlp

Tool calling / Function calling

Model ability to decide when and how to call external tools (APIs, databases, calculators) instead of trying to answer everything itself.

funciones api agentes

Top-p (Nucleus sampling)

Sampling method that only considers tokens whose cumulative probability exceeds p (usually 0.9–0.95).

parametro muestreo nucleus

Process by which a model learns from data and adjusts its internal parameters.

Transfer Learning

Technique that reuses a pre-trained model on a related task to speed up training on a new problem.

aprendizaje transferencia modelo

Architecture that allows efficient parallel processing of sequences such as text.

arquitectura text transformer

Situation where a model is too simple and fails to capture underlying patterns in the data, leading to poor performance on both training and test data.

fundamentos modelo calidad

Unsupervised learning

Model identifies patterns in data without prior labels.

no_supervisado clusters

Variational Autoencoder (VAE)

Model that learns to compress data into a latent space and then reconstruct it, useful for generation and dimensionality reduction.

arquitectura generativa compresion

Database optimized to store embeddings and perform similarity search.

vector busqueda almacen

Video generator

System that produces video clips or animations from text, images or templates.

video generativa

Vision Transformer (ViT)

Transformer-based architecture applied to images by splitting them into patches and processing them as sequences, often outperforming CNNs.

vision transformer arquitectura

Zero-shot learning

Ability of a model to handle tasks without specific prior examples.

generalizacion tareas