Glossary

Concise definitions of the central terms in the book, in alphabetical order by English term. Each entry links to the module where the concept is developed.

A2A (Agent2Agent) (PT: A2A (Agent2Agent), ES: A2A (Agent2Agent))

Open standard for communication between agents from different origins, allowing them to discover and cooperate with one another. Complements MCP (which connects agents to tools). [M10]

API (PT: API, ES: API)

Programmatic channel through which software (not a person) sends requests to the model and receives responses. [M10]

Activation function (PT: função de ativação, ES: función de activación)

Non-linear function (such as ReLU) that gives the network the power to represent complex relationships. [M3]

Agent (PT: agente, ES: agente)

System in which the model operates in a loop (plan, act, observe, repeat) to complete a multi-step task that it drives autonomously. [M10]

Alignment (PT: alinhamento, ES: alineamiento)

Set of techniques that steer the model’s behavior toward being helpful, honest, and safe (see RLHF, RLAIF, DPO, Constitutional AI). [M5, M11]

Artificial neuron (PT: neurônio artificial, ES: neurona artificial)

Unit that computes a weighted sum of inputs, adds a bias, and applies an activation function. [M3]

Attention (PT: atenção (attention), ES: atención (attention))

Mechanism by which each token weights the relevance of all others to reinterpret itself in context. [M4]

Attention budget (PT: attention budget (orçamento de atenção), ES: attention budget (presupuesto de atención))

Idea that the model’s capacity to “pay attention” is finite and competed for by all tokens in the context. [M10]

Autoregressive generation (PT: geração autorregressiva, ES: generación autorregresiva)

Production of the response one token at a time, feeding each generated token back into the input. [M6]

Backpropagation (PT: backpropagation (retropropagação), ES: backpropagation (retropropagación))

Algorithm that computes the gradients of all weights by propagating the error from the output back to the first layers. [M3]

Benchmark (PT: benchmark, ES: benchmark)

Standardized set of tasks with known answers, used to compare models. [M8]

Chain-of-thought (PT: raciocínio estruturado (chain-of-thought), ES: razonamiento estructurado (chain-of-thought))

Generation of intermediate reasoning steps before the final answer, improving performance on complex problems. [M6]

Chatbot (PT: chatbot, ES: chatbot)

Direct conversational interface with the model, without access to external data or execution of actions. [M10]

Constitutional AI (PT: Constitutional AI, ES: Constitutional AI)

Alignment variant in which feedback is generated by an AI guided by written principles (RLAIF), reducing dependence on human labels. [M5]

Context engineering (PT: context engineering (engenharia de contexto), ES: context engineering (ingeniería de contexto))

Discipline of deciding what information enters the context and how to organize it, shifting focus from the prompt to the information environment. [M10]

Context rot (PT: context rot, ES: context rot)

Degradation of quality as the context grows and accumulates irrelevant or stale information. [M10]

Context window (PT: janela de contexto, ES: ventana de contexto)

Maximum number of tokens the model considers at once. [M4]

DPO (Direct Preference Optimization) (PT: DPO (otimização direta de preferências), ES: DPO (optimización directa de preferencias))

Alignment method that tunes the model directly from preferences, without training a separate reward model. [M11]

Embedding (PT: embedding, ES: embedding)

Vector of real numbers associated with a token, defining its position in a high-dimensional space. [M2]

Evaluation contamination (PT: contaminação de avaliação, ES: contaminación de evaluación)

Leakage of test examples into the training set, inflating scores and measuring memorization rather than competence. [M8]

Fine-tuning (PT: ajuste fino (fine-tuning), ES: ajuste fino (fine-tuning))

Adaptation of an already-trained model to a specific task, style, or domain using additional data. [M5, M11]

Function calling (PT: function calling, ES: function calling)

Model’s ability to request the execution of external functions and incorporate their results. [M10]

Generalization (PT: generalização, ES: generalización)

Ability to perform correctly on new data, not just on examples seen during training. [M3]

Gradient descent (PT: descida do gradiente, ES: descenso del gradiente)

Procedure that adjusts weights in the direction that most reduces the loss. [M3]

Guardrails (PT: guardrails, ES: guardrails)

Filters and rules that validate input and output and restrict the system’s actions. [M12]

Hallucination (PT: alucinação (hallucination), ES: alucinación (hallucination))

Production of content that is plausible but factually false or unsupported by the source. [M7]

Harness (PT: harness, ES: harness)

Runtime that executes the agent loop, managing tools, memory, limits, and errors. [M10]

Inference (PT: inferência, ES: inferencia)

Use of the trained model to generate responses; no weights are altered in this process. [M6]

Instruction tuning (PT: ajuste por instruções (instruction tuning), ES: ajuste por instrucciones (instruction tuning))

Supervised fine-tuning that teaches the model to follow requests in an instruction-response format. [M5]

Intent engineering (PT: intent engineering, ES: intent engineering)

Effort to capture and express the real objective behind a request, which is often implicit. [M10]

Jailbreaking (PT: jailbreaking, ES: jailbreaking)

Circumventing the model’s alignment to obtain responses it was trained to refuse. [M12]

Knowledge distillation (PT: destilação (knowledge distillation), ES: destilación (knowledge distillation))

Training a small model (student) to imitate the outputs of a large model (teacher). [M8]

KV-cache (PT: KV-cache, ES: KV-cache)

Storage of keys and values already computed during attention, speeding up generation at the cost of memory. [M8]

LoRA / PEFT (PT: LoRA / PEFT, ES: LoRA / PEFT)

Parameter-efficient fine-tuning: trains small additional matrices instead of the entire model (LoRA); QLoRA adds quantization. [M11]

MCP (Model Context Protocol) (PT: MCP (Model Context Protocol), ES: MCP (Model Context Protocol))

Open standard for connecting AI applications to data and tools, like a “USB port for AI.” [M10]

Mixture of Experts (MoE) (PT: Mixture of Experts (MoE), ES: Mixture of Experts (MoE))

Sparse architecture in which a router activates only a few “experts” per token, decoupling total capacity from per-response cost. [M8]

Mode collapse (PT: mode collapse, ES: mode collapse)

Tendency of responses to converge to a repetitive pattern, losing diversity. [M10]

Model card (PT: model card, ES: model card)

Standardized document describing a model’s intended use, performance, training data, and limitations. [M12]

Multi-head attention (PT: multi-head attention, ES: multi-head attention)

Several attention operations in parallel, each capturing a different type of relationship. [M4]

Multimodality (PT: multimodalidade, ES: multimodalidad)

Model’s ability to process more than one type of signal (text, image, audio) in an integrated manner. [M13]

Neural network (PT: rede neural, ES: red neuronal)

Set of artificial neurons organized in layers, capable of learning representations from data. [M3]

Non-determinism (PT: não-determinismo, ES: no determinismo)

Variation of responses for the same input, resulting from sampling during generation. [M1, M6]

Open-weight vs. proprietary (PT: open-weight vs. proprietário, ES: open-weight vs. propietario)

Models whose weights are published and runnable locally, versus models accessible only via API. [M11]

Overfitting (PT: sobreajuste (overfitting), ES: sobreajuste (overfitting))

Memorizing training examples (including noise) to the point of generalizing poorly on new data. [M3]

Positional encoding (PT: positional encoding, ES: positional encoding)

Vector added to the embedding to register each token’s position in the sequence. [M4]

Pre-training (PT: pré-treinamento, ES: preentrenamiento)

Phase in which the model learns to predict the next token over vast corpora, acquiring language and knowledge. [M5]

Prompt engineering (PT: prompt engineering, ES: prompt engineering)

Art of formulating the instruction sent to the model (words, examples, format). [M10]

Prompt injection (PT: prompt injection, ES: prompt injection)

Malicious instruction embedded in external data that hijacks the model’s behavior. [M12]

Quantization (PT: quantização, ES: cuantización)

Reduction of the number of bits per weight, lowering memory usage and cost with a small quality loss. [M8]

RAG (Retrieval-Augmented Generation) (PT: RAG (recuperação aumentada de geração), ES: RAG (generación aumentada por recuperación))

Retrieving relevant external information and generating a response grounded in it, compensating for the model’s static knowledge. [M9]

Red teaming (PT: red teaming, ES: red teaming)

Deliberately attacking one’s own system to discover and fix flaws before real adversaries do. [M12]

RLHF / RLAIF (PT: RLHF / RLAIF, ES: RLHF / RLAIF)

Alignment via reinforcement from human preferences (RLHF) or from AI feedback guided by principles (RLAIF). [M5, M11]

Self-attention (PT: self-attention (autoatenção), ES: self-attention (autoatención))

Attention in which each token queries all others (query and key) and weights their values according to relevance. [M4]

Semantic vector space (PT: espaço vetorial semântico, ES: espacio vectorial semántico)

Geometric space in which proximity between vectors reflects closeness of meaning. [M2]

Skill (PT: skill, ES: skill)

Packaged, reusable capability that extends the model for a specific task, activatable on demand. [M10]

SLM (Small Language Model) (PT: SLM (modelo de linguagem pequeno), ES: SLM (modelo de lenguaje pequeño))

Lean, specialized model that can outperform a large generalist on its task, at lower cost. [M11]

Specification engineering / Spec (PT: specification engineering / Spec, ES: specification engineering / Spec)

Discipline of transforming objectives into explicit specifications (rules, acceptance criteria); the Spec is the resulting control artifact. [M10]

Temperature (PT: temperatura, ES: temperatura)

Parameter that rescales the probability distribution, controlling the balance between predictability and diversity in generation. [M6]

Test-time compute (PT: test-time compute, ES: test-time compute)

Spending more computation at response time (more reasoning) to improve the result. [M6]

Token (PT: token, ES: token)

Smallest unit into which text is sliced for processing; can be a word, a subword, or a character. [M2]

Tokenization (PT: tokenização, ES: tokenización)

Process of splitting text into tokens; using subwords ensures a finite vocabulary and coverage of unseen words. [M2]

Tool use (PT: tool use, ES: tool use)

Model’s use of external tools (search, computation, APIs) to obtain data or execute actions. [M10]

Transformer (PT: transformer, ES: transformer)

Architecture based on the attention mechanism that made modern LLMs possible. [M4]

Vector (PT: vetor, ES: vector)

Ordered list of numbers; in the context of LLMs, the way tokens and meanings are represented. [M2]

Vector database (PT: banco vetorial, ES: banco vectorial)

System that stores vectors and rapidly finds nearest neighbors, the foundation of search in RAG. [M9]

Vision Transformer (ViT) (PT: Vision Transformer (ViT), ES: Vision Transformer (ViT))

Architecture that splits an image into patches and treats them as tokens, bringing the Transformer to vision. [M13]

Weight (PT: peso (weight), ES: peso (weight))

Adjustable parameter that measures the importance of an input; learning is adjusting weights. [M3]