Concept Glossary: L¶

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

label consistency¶

• Meaning: The degree to which labels with the same meaning repeat for the same event or for cases under similar conditions. It checks not just whether a label column exists, but whether the judgment meaning remains stable across reviewers, time, and criteria.
• Why it matters: Even when a label-candidate column exists, it is hard to raise it directly to a stable learning problem if reviewers attach it differently or its meaning changes after a rule shift. Label consistency separates there is a column from the meaning repeats enough to use as a target label.
• Related concepts: supervised learning label, target, human oversight, model evaluation design
• Core Section: P3-9.6
• Appears in: P3-9.6

label prediction¶

• Meaning: A stage where the problem is framed as predicting an already defined target label for an input case. It is a stronger problem setup than an alert or review candidate, and it requires stable labels plus a learning and evaluation structure.
• Why it matters: If label prediction begins just because there are a few result notes, weak label quality can be hidden behind a complex classification frame. Label prediction should be considered only after earlier comparison signals and review structure have made the target label sufficiently stable.
• Related concepts: supervised learning label, target, classification, review candidate
• Core Section: P3-9.1
• Appears in: P3-9.1, P3-9.2

language model¶

• Meaning: A language model estimates how likely words or tokens are in sequence, often by predicting what can come next from previous context.
• Why it matters: This concept places LLMs in the older lineage of probabilistic language prediction rather than treating them only as chat products.
• Related concepts: language modeling, statistical language model, token
• Core Section: P1-11.1
• Appears in: P1-10.2, P1-11.2, P1-11.3, P6-20.1, P6-20.2

language modeling¶

• Meaning: Language modeling is the task of estimating the probability structure of token or word sequences.
• Why it matters: It explains why next-token prediction and sequence probability are central to modern LLMs and their earlier statistical and neural predecessors.
• Related concepts: direct lineage, Transformer, embedding
• Core Section: P1-9.3
• Appears in: P1-10.1, P1-11.1, P1-11.3, P6-20.2

learned representation¶

• Meaning: A learned representation is an internal form a model builds from data so useful differences for a task become easier to use.
• Why it matters: The concept distinguishes hand-designed features from internal representations learned by a model for the current task.
• Related concepts: representation, representation learning, feature
• Core Section: P1-9.1
• Appears in: P5-10.1, P5-10.2

learning-based approach¶

• Meaning: A learning-based approach builds a model's judgment criterion by fitting the relationship between inputs and outputs from case data instead of hand-writing every rule.
• Why it matters: It keeps machine learning from being misunderstood as simply discovering human-readable rules. Because the criterion is fitted from data, data quality, labels, and evaluation design become central.
• Related concepts: machine learning, feature, supervised learning label, model, model evaluation design
• First explained in: P4-1.2
• Appears in: P4-1.2

least privilege¶

• Meaning: Least privilege is the security principle of giving a person, tool, or AI agent only the access required for the current task.
• Why it matters: AI systems often connect to tools that change real state. Restricting access reduces the damage from mistakes, misuse, or prompt injection.
• Related concepts: permission, security, AI agent, tool use
• Core Section: P1-15.3
• Appears in: P1-14.5, P1-14.6, P7-6.2

limit¶

• Meaning: A limit describes what value a function approaches as the input gets close to a point or condition.
• Why it matters: Limits support the language of change, continuity, and derivatives, especially when direct substitution is not enough.
• Related concepts: convergence, function, rate of change
• Core Section: P2-2.3
• Appears in: P2-4.1, P2-4.2

linear regression¶

• Meaning: Linear regression models the relationship between input features and a continuous output with the simplest linear form first.
• Why it matters: It gives a simple baseline for regression before more complex models are justified, and it supports coefficient-based interpretation.
• Related concepts: regression, rate of change, error
• Core Section: P4-10.1
• Appears in: P4-10.2, P4-index, P4-10.3, P4-11.1

LLM¶

• Meaning: An LLM is a large language model trained on large-scale text to predict and generate token sequences.
• Why it matters: LLM is central to generative AI experience, but it should be separated from the full chatbot product, tool system, or all of AI.
• Related concepts: generative AI, machine learning, deep learning, language model, conversational LLM
• Core Section: P1-1.3
• Appears in: P1-9.3, P4-1.1, P1-11.3, P6-1.2, P6-20.1, P6-summary

LLM response generation¶

• Meaning: LLM response generation is the execution in which an LLM receives a prompt and produces a natural-language response. It can be read as the process of using learned parameters, the input prompt, and sampling settings to continue output tokens.
• Why it matters: An LLM inference result often appears as natural language rather than as a class label or number, so it is easy to confuse model execution with an actual thinking process. This concept helps keep factuality, evidence, and logical connection separate from the fluency of the generated response.
• Related concepts: inference, prompt, sampling, generation
• Core Section: P1-5.2

log loss¶

• Meaning: Log loss is a loss that becomes larger when the model assigns low probability to the correct answer. It uses a logarithm so confident wrong predictions are penalized strongly.
• Why it matters: Log loss appears often in classification models, logistic regression, and neural-network output interpretation. It helps explain why training pushes the model to raise the probability of the correct class, and why being confidently wrong is costly.
• Related concepts: logarithm, loss function, logistic regression, maximum likelihood estimation
• Core Section: P2-2.4
• Appears in: P4-11.3

logistic regression¶

• Meaning: Logistic regression is a linear classification model that turns a linear score into a probability-like value for class decisions.
• Why it matters: It is a simple, interpretable classification baseline and helps separate model score, probability-like output, threshold, and final class.
• Related concepts: classification, threshold, decision boundary
• Core Section: P4-11.1
• Appears in: P4-index, P4-3.2, P4-4.2, P4-11.2, P4-11.3, P4-11.4

long-context¶

• Meaning: Long-context refers to the design problem of preserving and using information from a long input within one task.
• Why it matters: Long inputs are not just a larger token limit. They require cost, latency, retrieval, summarization, and memory-management decisions.
• Related concepts: context window, retrieval-augmented generation, RAG, sparse attention
• Core Section: P6-4.5
• Appears in: P5-14.2, P6-4.2, P6-4.3, P6-4.4

long-term dependency¶

• Meaning: Long-term dependency is the problem where information from much earlier in a sequence remains important but is hard for a sequential model to preserve.
• Why it matters: It explains why basic recurrent models struggled with distant information and why gated memory designs and attention became important.
• Related concepts: RNN, recurrent neural network, hidden state, Attention
• Core Section: P5-12.2
• Appears in: P1-9.3, P5-12.1, P5-13.1, P5-13.2, P5-14.1, P5-14.2

loss curve¶

• Meaning: A loss curve is a line plot showing how loss changes across training steps or epochs.
• Why it matters: It reveals training flow, instability, plateauing, and overfitting more clearly than a single final loss value.
• Related concepts: visualization, loss function, accuracy
• Core Section: P2-13.2
• Appears in: P2-13.3, P2-15.1

loss function¶

• Meaning: A loss function turns prediction error into a number that training can try to reduce.
• Why it matters: Training needs a numeric objective to reduce. The loss function defines what kind of error the model treats as costly.
• Related concepts: error, objective function, gradient descent
• Core Section: P2-6.2
• Appears in: P2-6.3, P2-13.2, P2-15.1, P5-4.1, P5-4.2

mathematical logarithm¶

• Meaning: The inverse operation of exponentiation, used to turn multiplicative scale into additive scale.
• Why it matters: It appears in loss functions, probability, information measures, and numerical transformations throughout AI.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-2.4

labeled example¶

• Meaning: A training example that pairs an input with the target label or answer used for supervised learning.
• Why it matters: It shows that supervised learning needs not only data, but data whose answer standard has been defined.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-8.1

linear algebra¶

• Meaning: The branch of mathematics that studies vectors, matrices, linear transformations, and spaces.
• Why it matters: It is a core language for reading data representation, model parameters, neural network computation, and embeddings.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-3.1

linear transformation¶

• Meaning: A transformation that preserves addition and scalar multiplication, often represented by matrix multiplication.
• Why it matters: It helps readers see matrix multiplication as a structured way of moving or reshaping vectors.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P2-3.3

loss¶

• Meaning: A numerical signal showing how far the model output is from the target or learning objective.
• Why it matters: It is the quantity that training usually tries to reduce, so it connects model output, labels, gradients, and optimization.
• Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
• Core Section: P1-5.1
• Appears in: P2-1.1, P2-1.2, P2-2.2, P2-2.3