Concept Glossary: G¶

This page lists English glossary entries for this letter. Entry bodies are assembled from term source files.

generalization¶

Meaning: Generalization is the state in which a model captures a relationship that still works reasonably well on data it did not see during training. It means the model has learned a broader pattern instead of merely memorizing the training examples.
Why it matters: Generalization is the key difference between learning and memorization. It explains why validation and test data are needed even when training scores look good, and why overfitting and underfitting are both forms of generalization failure. Model performance is not just about fitting known examples; it is about holding up in similar new situations.
Related concepts: overfitting, underfitting
Core Section: P1-3.2
Appears in: P5-8.1, P4-index, P4-3.2, P4-4.1, P4-5.1, P4-5.2, P4-6.1, P4-7.2, P4-9.1, P5-8.2

Meaning: Generation is the task of producing new text, images, speech, code, or other output from instructions, conditions, examples, and context. It is not choosing one item from a fixed list; it is building an output through a sequence of smaller choices.
Why it matters: Generation separates generative AI from classification or ranking tasks. The quality of a generated result depends not only on a final choice, but also on context, sampling settings, and the sequence of intermediate decisions that form the output.
Related concepts: recommendation task, ranking, large language model, sampling, next-token prediction
Core Section: P1-10.1
Appears in: P1-1.2, P1-13.3, P5-15.1, P5-15.2, P5-15.3

Meaning: Generative AI is the broad category of models and services that create new content such as text, images, audio, video, and code. It is defined by the character of the output, not by one specific model architecture.
Why it matters: This concept keeps deep learning methods, LLM model families, and user-facing AI services from being collapsed into one vague idea. It also helps readers see why evaluating generated content requires different standards from evaluating a simple classification label.
Related concepts: generation, deep learning, large language model
Core Section: P1-10.1
Appears in: P1-1.3, P4-1.1

Meaning: A generative model learns patterns or candidate distributions from data and can produce new samples or next outputs. It does more than map an input to a fixed label; it can draw an output from what it has learned about possible data.
Why it matters: Generative models mark the shift from deciding what category is correct to producing something new. They require readers to consider diversity, consistency, conditioning, and distribution fit, not only accuracy.
Related concepts: generation, sampling, next-token prediction
Core Section: P5-15.2
Appears in: P5-15.1, P4-1.1, P5-15.3

Meaning: GPT is a family of language models based on Transformer decoders and pretrained to generate the next token from preceding context. It is best read as a generative model family that builds output one token at a time.
Why it matters: GPT-style models form a direct lineage for much of the modern LLM user experience, including chat, code generation, summarization, and tool-using workflows. The concept also helps distinguish decoder-based generation from encoder-centered representation models.
Related concepts: decoder, pretraining, in-context learning
Core Section: P6-5.1
Appears in: P1-11.3

Meaning: A group split separates data by dependent groups, such as the same person, device, store, patient, or account, so that records from the same group do not appear on both the training side and the validation or test side.
Why it matters: If similar records from the same entity appear on both sides, the model may look strong because it recognizes a familiar entity rather than generalizing to a new one. A group split closes the question of what should count as an independent new case.
Related concepts: evaluation design, data leakage, sample unit
Core Section: P3-9.13
Appears in: P3-9.13

Meaning: A gradient is a vector of partial derivatives that shows how a value, usually loss, changes with respect to multiple parameters. It acts like a direction signal for how parameter changes would affect the loss.
Why it matters: Gradients are the link between loss and learning. They explain why parameters are not changed randomly, but updated in directions that are expected to reduce loss. Without gradients, gradient descent and backpropagation cannot be read as concrete computation.
Related concepts: partial derivative, vector, model training
Core Section: P2-4.3
Appears in: P2-4.4, P2-6.3, P5-5.1, P5-5.2, P5-6.1, P5-7.1

Meaning: Gradient descent is an iterative optimization method that changes parameters little by little in a direction expected to reduce loss. It does not solve for the best parameters at once; it repeatedly follows a local direction signal.
Why it matters: Gradient descent shows how gradients become actual learning updates. It helps readers see training as movement across a loss landscape, where the model adjusts values step by step rather than discovering a final answer in one calculation.
Related concepts: gradient, optimization, hyperparameter
Core Section: P2-6.3
Appears in: P2-15.2

Meaning: A policy, rule, or technical constraint that limits unsafe or unwanted AI system behavior.
Why it matters: It helps readers see safety as an explicit control layer rather than as a vague hope that the model will behave well.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-14.5