# Glosario — Inteligencia Artificial Aplicada y Segura

Términos que el curso usa y que el examen puede pedir.

| Término | Definición en una línea |
|---|---|
| **Entrenamiento** | Adjusting parameters with data until the model minimizes its error. |
| **Inferencia** | Using the already-trained model to predict. |
| **Sobreajuste** | Memorizing the training set; fails with anything new. |
| **Feature** | Input variable the model uses to predict. |
| **LLM** | Large Language Model: predicts the next token from context. |
| **Token (linguistic)** | Text unit that the model processes; not the same as credential token. |
| **Temperatura** | Sampling randomness: low = conservative, high = creative. |
| **Ventana de contexto** | Maximum text the model can consider at once. |
| **Alucinación** | Plausible but false output: statistical property, not a bug. |
| **Prompt engineering** | Designing the instruction: ROLE + TASK + CONTEXT + FORMAT. |
| **Grounding** | Answering only based on cited sources. |
| **Prompt injection** | Untrusted input that the system executes as an instruction. |
| **Ataque adversarial** | Minimal input change that fools the model. |
| **Poisoning** | Contaminating training data to corrupt the model. |
| **Safeguard** | Technical limitation of what the assistant can do. |
| **Human-in-the-loop** | A person validates before the output has an effect. |

> Regla de estudio: si no sabes explicarlo en una frase a alguien de fuera, aún no lo sabes.