Code & Chaos
The Builder’s Field Guide

A living glossary for the language of software, AI, agents, systems, and the humans building with them.

IV.

AI Behavior & Processing

The machinery beneath model behavior: architecture, training, inference, token processing, sampling, multimodality, generation, routing, evaluation, and the limits hiding behind confident output.

130 terms

Artificial Neural Network — Neural Network Beginner
Plain-English definition

A computational system made of connected layers that learn patterns from data.

More precise definition

Neural networks transform input through weighted connections and nonlinear operations. Training adjusts those weights so the network becomes better at a target task.

Example

A neural network can learn to classify images, predict text, or recognize speech.

Commonly confused with

Artificial intelligence, deep learning

Common misconception

A neural network is inspired by biological brains, but it is not a literal digital brain.

Transformer Beginner
Plain-English definition

A neural-network architecture designed to process relationships within sequences.

More precise definition

Transformers use attention mechanisms to weigh how strongly different parts of an input relate to one another. They underpin many modern language, vision, and multimodal models.

Example

Most current large language models use transformer-based architectures.

Commonly confused with

Large language model

Common misconception

A transformer is the architecture. An LLM is a trained model that may use that architecture.

Attention Beginner
Plain-English definition

A mechanism that helps a model decide which parts of the input matter most to each other.

More precise definition

Attention computes relationships between tokens or features so the model can weight relevant context when producing internal representations.

Example

In a sentence, attention helps connect a pronoun to the noun it refers to.

Commonly confused with

Memory, focus

Common misconception

Attention is a mathematical operation, not conscious concentration.

Self-Attention Intermediate
Plain-English definition

Attention applied within one sequence so its parts can relate to each other.

More precise definition

Self-attention lets each token compare itself with other tokens in the same input and build a context-sensitive representation.

Example

A word can be interpreted differently depending on the surrounding words.

Commonly confused with

Cross-attention

Common misconception

Self-attention does not mean the model is self-aware.

Cross-Attention Intermediate
Plain-English definition

Attention that connects one set of information to another.

More precise definition

Cross-attention allows one representation, such as generated text, to attend to another source, such as an image, audio signal, or encoder output.

Example

An image generator using text instructions to guide visual generation.

Commonly confused with

Self-attention

Common misconception

Cross-attention links separate information streams; self-attention relates items within the same stream.

Parameter Beginner
Plain-English definition

A learned numerical value inside a model.

More precise definition

Parameters determine how the model transforms input. Training updates them to reduce error or improve a target objective.

Example

A model described as having 7 billion parameters contains roughly 7 billion learned values.

Commonly confused with

Function parameter, setting

Common misconception

A model parameter is not the same as a function parameter or API option.

Weights Beginner
Plain-English definition

The learned numerical values that shape a model’s behavior.

More precise definition

Weights are model parameters adjusted during training. They encode statistical patterns learned from data.

Example

Fine-tuning changes some or all of a model’s weights.

Commonly confused with

Settings, memory

Common misconception

Weights do not store experiences as neatly readable sentences.

Checkpoint Intermediate
Plain-English definition

A saved version of a model’s learned state.

More precise definition

A checkpoint usually contains model weights and may include optimizer state, configuration, tokenizer files, or training metadata.

Example

A team may compare checkpoints from different stages of training.

Commonly confused with

Model release, snapshot

Common misconception

A checkpoint is the saved model state, not necessarily a complete deployable product.

Model Size Beginner
Plain-English definition

A rough measure of how large a model is, often described by parameter count.

More precise definition

Model size may refer to parameter count, memory footprint, file size, or compute requirements depending on context.

Example

A 70B model has about 70 billion parameters.

Commonly confused with

Context window

Common misconception

A larger model is not automatically better for every task.

Model Capacity Intermediate
Plain-English definition

How much complexity a model can represent or learn.

More precise definition

Capacity depends on architecture, parameter count, training, and constraints. Too little capacity can underfit; too much can increase cost or overfitting risk.

Example

A small classifier may lack the capacity for highly nuanced distinctions.

Commonly confused with

Model size

Common misconception

Parameter count is one contributor to capacity, not a complete measure of intelligence or quality.

Tokenization Beginner
Plain-English definition

Breaking text or other input into units a model can process.

More precise definition

A tokenizer converts raw input into token identifiers and may also reconstruct output tokens back into text.

Example

A long word may be split into several smaller tokens.

Commonly confused with

Chunking

Common misconception

Tokenization is a model-input operation. Chunking is a higher-level strategy for dividing larger content.

Tokenizer Beginner
Plain-English definition

The component that converts text into tokens and back again.

More precise definition

A tokenizer defines the model’s vocabulary and the rules used to encode and decode text.

Example

Two models may count the same sentence as a different number of tokens.

Commonly confused with

Model

Common misconception

The tokenizer is separate from the model weights, though the two must be compatible.

Vocabulary Intermediate
Plain-English definition

The set of tokens a tokenizer knows how to represent.

More precise definition

A model vocabulary contains token units and their identifiers. It may include words, subwords, punctuation, symbols, or special tokens.

Example

A tokenizer vocabulary might contain separate tokens for common word fragments.

Commonly confused with

Training data

Common misconception

A token vocabulary is not a dictionary of meanings.

Special Token Intermediate
Plain-English definition

A token reserved for structural or control purposes.

More precise definition

Special tokens may mark message boundaries, padding, starts or ends of sequences, tool calls, or roles.

Example

A model may use hidden boundary tokens between system and user messages.

Commonly confused with

Prompt text

Common misconception

Special tokens may affect processing even when they are not visible in the interface.

Embedding Beginner
Plain-English definition

A numerical representation of meaning or features.

More precise definition

Embeddings map text, images, audio, or entities into vectors so similar items can be compared mathematically.

Example

Two semantically similar sentences may have nearby embeddings.

Commonly confused with

Model weights, memory

Common misconception

An embedding is not the original content. It is a compressed numerical representation.

Vector Beginner
Plain-English definition

An ordered list of numbers used to represent information.

More precise definition

Vectors can encode features, positions, meanings, or model states in a mathematical space.

Example

An embedding is typically stored as a vector of floating-point numbers.

Commonly confused with

Array

Common misconception

A vector is mathematically meaningful because of how its numbers relate within a space.

Latent Space Intermediate
Plain-English definition

An internal numerical space where a model represents learned features.

More precise definition

Latent spaces encode compressed patterns and relationships that may correspond to concepts, styles, shapes, or semantic structure.

Example

Image models operate on latent representations rather than raw pixels during much of generation.

Commonly confused with

Embedding space

Common misconception

A latent space is not a literal hidden room containing complete concepts.

Representation Intermediate
Plain-English definition

The internal form a model uses to encode information.

More precise definition

Representations transform raw input into learned features that support prediction, generation, or classification.

Example

A sentence becomes a sequence of internal vectors before output is generated.

Commonly confused with

Embedding

Common misconception

Embeddings are one kind of representation, not the only kind.

Hidden State Intermediate
Plain-English definition

An internal representation created while a model processes input.

More precise definition

Hidden states carry transformed information between layers or generation steps and are not usually exposed directly to users.

Example

A language model builds hidden states for each token position.

Commonly confused with

Memory, chain of thought

Common misconception

A hidden state is a numerical representation, not a readable private diary.

Activation Intermediate
Plain-English definition

The numerical output produced by a model unit or layer during processing.

More precise definition

Activations are intermediate values created when input passes through the model and are shaped by weights and activation functions.

Example

Researchers may inspect activations to study model behavior.

Commonly confused with

Weight

Common misconception

Weights are stored learned values. Activations are temporary values produced during inference.

Layer Beginner
Plain-English definition

One stage in a neural network’s processing pipeline.

More precise definition

Layers apply transformations to representations and pass results forward through the model.

Example

A transformer contains repeated attention and feed-forward layers.

Commonly confused with

Module

Common misconception

More layers do not automatically guarantee better performance.

Inference Beginner
Plain-English definition

Running a trained model to produce an output.

More precise definition

Inference converts input into predictions or generated content using fixed model weights.

Example

Sending a prompt and receiving a completion.

Commonly confused with

Training

Common misconception

Inference normally uses the model; it does not retrain the model.

Forward Pass Intermediate
Plain-English definition

Passing input through the model to calculate output.

More precise definition

A forward pass computes layer-by-layer activations using the current weights.

Example

Each next-token prediction requires model computation over the current sequence.

Commonly confused with

Backpropagation

Common misconception

Forward passes happen during both training and inference. Backpropagation happens during training.

Backpropagation Intermediate
Plain-English definition

The training process used to calculate how model weights should change.

More precise definition

Backpropagation propagates error gradients backward through the network so optimization can update parameters.

Example

Training compares output against a target and backpropagates the error.

Commonly confused with

Inference

Common misconception

Backpropagation is a learning mechanism, not the ordinary generation process.

Loss Function — Loss Intermediate
Plain-English definition

A numerical measure of how wrong a model’s output is during training.

More precise definition

The loss function defines the objective the optimizer tries to reduce.

Example

A language model may be trained to reduce next-token prediction loss.

Commonly confused with

Evaluation metric

Common misconception

Lower training loss does not always mean better real-world usefulness.

Gradient Intermediate
Plain-English definition

A signal showing how model parameters should change to reduce loss.

More precise definition

Gradients represent the direction and magnitude of change for each parameter during optimization.

Example

The optimizer uses gradients calculated through backpropagation.

Commonly confused with

Weight

Common misconception

Gradients are temporary update signals, not the final learned parameters.

Optimizer Intermediate
Plain-English definition

The algorithm that updates model weights during training.

More precise definition

An optimizer uses gradients and update rules to reduce the training objective.

Example

Adam is a widely used optimizer.

Commonly confused with

Model architecture

Common misconception

The optimizer trains the model; it is not usually part of inference after training.

Learning Rate Intermediate
Plain-English definition

A setting controlling how large each training update is.

More precise definition

The learning rate scales parameter updates and strongly affects stability and convergence.

Example

A learning rate that is too high can make training unstable.

Commonly confused with

Training speed

Common misconception

A higher learning rate does not simply mean the model learns better or faster.

Hyperparameter Intermediate
Plain-English definition

A setting chosen for training or inference rather than learned directly by the model.

More precise definition

Hyperparameters include learning rate, batch size, temperature, architecture depth, and regularization strength.

Example

Temperature is an inference hyperparameter.

Commonly confused with

Parameter

Common misconception

Model parameters are learned. Hyperparameters are selected or configured.

Batch Intermediate
Plain-English definition

A group of examples processed together.

More precise definition

Batching improves computational efficiency during training or inference by processing multiple inputs in parallel.

Example

A training batch may contain 64 text sequences.

Commonly confused with

Dataset

Common misconception

A batch is one portion of a dataset, not the entire dataset.

Batch Size Intermediate
Plain-English definition

The number of examples processed together in one batch.

More precise definition

Batch size affects memory use, training dynamics, latency, and throughput.

Example

A batch size of 32 processes 32 examples before one update.

Commonly confused with

Context length

Common misconception

Batch size counts examples, while context length counts tokens or input size per example.

Epoch Intermediate
Plain-English definition

One complete pass through a training dataset.

More precise definition

Training often uses multiple epochs so the model sees the dataset repeatedly.

Example

Fine-tuning may run for three epochs.

Commonly confused with

Step

Common misconception

One epoch contains many training steps.

Training Step — Step Intermediate
Plain-English definition

One parameter-update cycle during training.

More precise definition

A step usually processes one batch, computes loss and gradients, and applies an optimizer update.

Example

A training run may contain tens of thousands of steps.

Commonly confused with

Epoch

Common misconception

A step is one update. An epoch is a full pass through the data.

Pretraining Beginner
Plain-English definition

Large-scale initial training that teaches a model broad patterns.

More precise definition

Pretraining usually uses extensive datasets and a general objective before later specialization.

Example

An LLM learns broad language patterns during pretraining.

Commonly confused with

Fine-tuning

Common misconception

Pretraining creates the general base model. Fine-tuning adapts it later.

Base Model — Foundation Model Beginner
Plain-English definition

A general model used as the starting point for later adaptation.

More precise definition

A base model is pretrained broadly and may later be instruction-tuned, fine-tuned, aligned, or wrapped in a product.

Example

A general language model before chat or tool-use tuning.

Commonly confused with

Chat model

Common misconception

A base model may be capable but not yet optimized for helpful conversation.

Foundation Model Beginner
Plain-English definition

A broadly trained model that can support many downstream tasks.

More precise definition

Foundation models are trained at scale and adapted through prompting, fine-tuning, tools, or product layers.

Example

A general text-and-image model used across many applications.

Commonly confused with

Base model

Common misconception

Foundation model emphasizes broad reuse. Base model emphasizes its role as the starting model.

Fine-tuning Beginner
Plain-English definition

Additional training used to adapt an existing model.

More precise definition

Fine-tuning updates model weights using a narrower dataset or objective to improve task, style, domain, or behavior performance.

Example

Fine-tuning a model on customer-support conversations.

Commonly confused with

Prompting, retrieval

Common misconception

Fine-tuning changes model weights. Prompting and retrieval change the input context instead.

Instruction Tuning Intermediate
Plain-English definition

Training a model to follow written instructions more reliably.

More precise definition

Instruction tuning uses examples of instructions and desired responses so the model generalizes across tasks.

Example

Teaching a base model to answer questions, summarize, and format outputs.

Commonly confused with

Prompt engineering

Common misconception

Instruction tuning changes weights. Prompt engineering changes the prompt.

Preference Tuning Intermediate
Plain-English definition

Training a model to prefer some outputs over others.

More precise definition

Preference tuning uses ranked or comparative examples to shape style, helpfulness, safety, or behavior.

Example

Training the model to prefer clearer answers over rambling ones.

Commonly confused with

Instruction tuning

Common misconception

Instruction tuning teaches task following. Preference tuning shapes which acceptable outputs are preferred.

RLHF — Reinforcement Learning from Human Feedback Intermediate
Plain-English definition

A method for shaping model behavior using human preferences.

More precise definition

RLHF typically trains a reward model from human comparisons and then optimizes the model toward higher-scoring outputs.

Example

Human reviewers rank responses, and the model is trained toward preferred behavior.

Commonly confused with

Fine-tuning, DPO

Common misconception

RLHF does not directly encode every preference as a fixed rule.

DPO — Direct Preference Optimization Intermediate
Plain-English definition

A method for training directly from preferred and rejected response pairs.

More precise definition

DPO optimizes model preferences without requiring a separate reinforcement-learning loop in the same form as classic RLHF.

Example

Training on pairs where one answer is marked preferred.

Commonly confused with

RLHF

Common misconception

DPO and RLHF pursue similar preference goals through different training procedures.

Reward Model Intermediate
Plain-English definition

A model that scores outputs according to learned preferences.

More precise definition

Reward models are trained from comparisons or labels and used to estimate which outputs are more desirable.

Example

A reward model may score helpfulness or safety.

Commonly confused with

Judge model

Common misconception

A reward model shapes training. A judge model is often used during evaluation.

Alignment Beginner
Plain-English definition

The effort to make a model behave according to intended goals and constraints.

More precise definition

Alignment may address helpfulness, reliability, safety, honesty, instruction following, and value-sensitive behavior.

Example

Training a model to refuse dangerous instructions while remaining useful.

Commonly confused with

Safety

Common misconception

Alignment is broader than safety and remains an imperfect, ongoing process.

Overfitting Intermediate
Plain-English definition

Learning training examples too specifically and performing poorly on new data.

More precise definition

An overfit model captures noise or narrow patterns instead of generalizable structure.

Example

A classifier memorizes training phrases but fails on reworded examples.

Commonly confused with

Memorization

Common misconception

Overfitting is about poor generalization, not merely high training performance.

Underfitting Intermediate
Plain-English definition

Failing to learn enough of the underlying pattern.

More precise definition

An underfit model performs poorly on both training data and new data because its capacity, data, or training is insufficient.

Example

A simple model cannot capture the complexity of nuanced language.

Commonly confused with

Overfitting

Common misconception

Underfitting is not cautious behavior. It is inadequate learning.

Generalization Intermediate
Plain-English definition

How well a model performs on new examples it was not trained on directly.

More precise definition

Generalization measures whether learned patterns transfer beyond the training set.

Example

A model correctly handles a new phrasing of a familiar task.

Commonly confused with

Memorization

Common misconception

Good generalization does not mean the model understands every situation.

Memorization Intermediate
Plain-English definition

Retaining specific patterns or examples from training data.

More precise definition

Memorization can range from useful retention of common patterns to reproduction of rare or sensitive sequences.

Example

A model reproduces a distinctive phrase seen during training.

Commonly confused with

Learning, memory

Common misconception

Model memorization is not the same as a user-facing memory system.

Generation Beginner
Plain-English definition

Producing new model output.

More precise definition

Generation repeatedly selects or predicts outputs from a probability distribution until a stopping condition is reached.

Example

A language model generates one token at a time.

Commonly confused with

Retrieval

Common misconception

Generated output is produced by the model, while retrieved output is fetched from an external source.

Completion Beginner
Plain-English definition

The generated continuation returned by a language model.

More precise definition

A completion may be free text, structured output, code, a tool call, or part of a chat response.

Example

The assistant message returned after a prompt.

Commonly confused with

Response

Common misconception

Completion is a model-generation term. Response may include additional system or tool-layer data.

Decoding Intermediate
Plain-English definition

The strategy used to choose model output from predicted possibilities.

More precise definition

Decoding methods include greedy selection, beam search, top-k sampling, nucleus sampling, and temperature scaling.

Example

A model samples the next token using temperature and top-p.

Commonly confused with

Tokenizer decoding

Common misconception

Decoding can also mean converting token IDs back into text; context determines the meaning.

Logits Intermediate
Plain-English definition

Raw scores a model produces before turning them into probabilities.

More precise definition

Logits are transformed, often through softmax, into a probability distribution over possible outputs.

Example

Each possible next token receives a logit score.

Commonly confused with

Probability

Common misconception

Logits are not probabilities until normalized.

Softmax Intermediate
Plain-English definition

A function that converts scores into probabilities.

More precise definition

Softmax normalizes logits into a distribution whose values sum to one.

Example

Next-token logits become probabilities before sampling.

Commonly confused with

Argmax

Common misconception

Softmax creates a distribution. It does not itself decide which token is selected.

Probability Distribution Beginner
Plain-English definition

A set of possible outputs paired with their likelihoods.

More precise definition

During generation, the model assigns relative probability to candidate tokens or outputs.

Example

One token may have 40% probability while another has 15%.

Commonly confused with

Confidence score

Common misconception

High probability means likely under the model, not necessarily true or correct.

Sampling Beginner
Plain-English definition

Choosing an output from the model’s probability distribution.

More precise definition

Sampling introduces controlled variability by selecting among plausible candidates rather than always taking the single highest-probability option.

Example

Two runs with the same prompt may produce different wording.

Commonly confused with

Training sample

Common misconception

Sampling during generation is different from sampling examples from a dataset.

Greedy Decoding Intermediate
Plain-English definition

Always choosing the highest-probability next output.

More precise definition

Greedy decoding is deterministic under fixed conditions but may produce repetitive or locally optimal text.

Example

At each step, choose the token with the highest probability.

Commonly confused with

Sampling

Common misconception

The most probable next token at every step does not always produce the best complete answer.

Temperature Beginner
Plain-English definition

A setting that changes how adventurous or conservative output selection is.

More precise definition

Temperature rescales logits before sampling. Lower values concentrate probability on likely outputs; higher values flatten the distribution.

Example

Lower temperature often produces more consistent wording.

Commonly confused with

Creativity

Common misconception

Temperature does not directly control intelligence, truthfulness, or emotional warmth.

Top-k Intermediate
Plain-English definition

Restricting sampling to the k most likely candidates.

More precise definition

Top-k removes all but a fixed number of highest-probability tokens before sampling.

Example

Top-k 50 samples only from the 50 most likely next tokens.

Commonly confused with

Top-p

Common misconception

Top-k uses a fixed candidate count. Top-p uses a probability-mass threshold.

Top-p — Nucleus Sampling Intermediate
Plain-English definition

Restricting sampling to the smallest group of candidates that reaches a chosen total probability.

More precise definition

Top-p dynamically changes the candidate set based on how probability is distributed.

Example

Top-p 0.9 samples from candidates covering 90% of probability mass.

Commonly confused with

Top-k

Common misconception

Top-p does not mean the top 90% of tokens by count.

Seed Beginner
Plain-English definition

A starting value used by a random-number process.

More precise definition

A fixed seed can help reproduce stochastic output when the rest of the system is also stable.

Example

Using the same image-generation seed to explore prompt variations.

Commonly confused with

Prompt

Common misconception

The same seed does not guarantee identical output across different models, versions, or infrastructure.

Deterministic Beginner
Plain-English definition

Producing the same output from the same input under the same conditions.

More precise definition

Determinism requires stable code, model version, settings, environment, and random state.

Example

A pure formatting function is deterministic.

Commonly confused with

Consistent

Common misconception

Low temperature may improve consistency but does not guarantee full determinism.

Nondeterministic Beginner
Plain-English definition

Able to produce different results from the same apparent input.

More precise definition

Variation may come from sampling, concurrency, hardware behavior, model updates, hidden context, or external systems.

Example

A generative model returns slightly different answers on repeated runs.

Commonly confused with

Random

Common misconception

Nondeterministic does not mean uncaused or completely random.

Reproducibility Intermediate
Plain-English definition

The ability to recreate the same result or experiment reliably.

More precise definition

Reproducibility depends on recording model versions, prompts, seeds, data, dependencies, settings, and environment.

Example

An eval stores the exact model and prompt version used.

Commonly confused with

Determinism

Common misconception

A nondeterministic system can still be reproducibly evaluated through repeated measurements and recorded conditions.

Context Window Beginner
Plain-English definition

The maximum amount of input and generated text a model can consider in one request.

More precise definition

The context window is measured in tokens and includes instructions, conversation, retrieved content, tool data, and output budget.

Example

A 128k context window can process roughly 128,000 tokens in one request.

Commonly confused with

Memory

Common misconception

A large context window is temporary working space, not permanent memory.

Token Budget Beginner
Plain-English definition

The amount of context space available for input and output.

More precise definition

Token budgets may be allocated across system instructions, conversation history, retrieved content, tool results, and generated output.

Example

A long document may leave less room for the model’s response.

Commonly confused with

Usage quota

Common misconception

Token budget is request capacity. Quota is a service usage allowance.

Truncation Beginner
Plain-English definition

Cutting off content because it exceeds a limit.

More precise definition

Truncation may remove earlier context, later content, or generated output depending on system policy.

Example

Old conversation turns are dropped when the context window is full.

Commonly confused with

Summarization

Common misconception

Truncation discards content. Summarization compresses it.

System Prompt — System Message Beginner
Plain-English definition

High-priority instructions that shape model behavior.

More precise definition

The system prompt may define role, policies, tone, tool rules, constraints, and response behavior.

Example

A system message instructs an assistant to use a specific tool before answering.

Commonly confused with

User prompt

Common misconception

A system prompt influences behavior but does not guarantee perfect compliance.

Prompt Hierarchy Intermediate
Plain-English definition

The priority order among different instruction sources.

More precise definition

Prompt hierarchy determines how system, developer, user, tool, and retrieved instructions interact and which instructions win during conflict.

Example

A user request cannot override a higher-priority safety rule.

Commonly confused with

Prompt order

Common misconception

Later text is not always higher priority. Authority depends on message role and system design.

User Prompt Beginner
Plain-English definition

The instruction or request provided by the user.

More precise definition

A user prompt may contain goals, context, examples, constraints, and desired output format.

Example

“Explain this schema in simple language.”

Commonly confused with

System prompt

Common misconception

The user prompt is important but may be constrained by higher-priority instructions.

Chat Template Intermediate
Plain-English definition

The formatting pattern used to turn messages into model input.

More precise definition

A chat template inserts roles, separators, control tokens, and message boundaries expected by the model.

Example

System, user, and assistant messages are wrapped with special tokens.

Commonly confused with

Prompt template

Common misconception

A chat template is often model-specific and can strongly affect behavior.

Prompt Template Beginner
Plain-English definition

A reusable prompt structure with placeholders.

More precise definition

Prompt templates standardize instructions while inserting task-specific content, variables, examples, or context.

Example

“Summarize {{document}} in {{length}} words.”

Commonly confused with

Chat template

Common misconception

A prompt template is authored task content. A chat template is low-level message formatting.

Zero-shot Prompting Beginner
Plain-English definition

Asking a model to perform a task without showing examples.

More precise definition

Zero-shot prompting relies on the model’s existing learned capabilities and the written instruction.

Example

“Classify this message as urgent or not urgent.”

Commonly confused with

Few-shot prompting

Common misconception

Zero-shot means no examples in the prompt, not that the model had no prior training.

Few-shot Prompting Beginner
Plain-English definition

Showing a model a small number of examples inside the prompt.

More precise definition

Few-shot examples demonstrate desired behavior, format, labels, or reasoning patterns without changing model weights.

Example

Providing three correctly labeled messages before asking for a fourth classification.

Commonly confused with

Fine-tuning

Common misconception

Few-shot prompting changes context. Fine-tuning changes model weights.

In-Context Learning Intermediate
Plain-English definition

Adapting behavior based on examples or instructions provided in the current context.

More precise definition

The model changes its response pattern during inference without updating its underlying weights.

Example

A model follows a new labeling scheme demonstrated in the prompt.

Commonly confused with

Fine-tuning

Common misconception

In-context learning is temporary and request-scoped.

Instruction Following Beginner
Plain-English definition

How reliably a model follows the instructions it receives.

More precise definition

Instruction-following quality depends on training, prompt clarity, hierarchy, context, capability, and competing constraints.

Example

Returning valid JSON when explicitly required.

Commonly confused with

Obedience

Common misconception

Models do not follow instructions perfectly or literally in every situation.

Structured Output Beginner
Plain-English definition

Model output required to follow a defined format.

More precise definition

Structured output may use JSON, XML, tables, typed fields, or provider-enforced schemas.

Example

Returning an object with name, category, and confidence fields.

Commonly confused with

Plain text

Common misconception

Output that looks like JSON is not necessarily valid JSON.

Schema Adherence Intermediate
Plain-English definition

How closely output follows a required schema.

More precise definition

Schema adherence checks field presence, types, allowed values, nesting, and formatting.

Example

A tool call includes all required arguments with valid types.

Commonly confused with

Semantic correctness

Common misconception

Schema-valid output can still contain incorrect or misleading content.

Tool Use — Function Calling Beginner
Plain-English definition

Allowing a model to request actions from external tools.

More precise definition

The model selects a tool and produces structured arguments; the host executes the tool and returns results.

Example

Calling a calendar search tool instead of guessing availability.

Commonly confused with

Plugin, agent

Common misconception

The model usually proposes the call. The surrounding application executes it.

Tool Call Beginner
Plain-English definition

One structured request to use a tool.

More precise definition

A tool call names the operation and supplies arguments that should match the tool schema.

Example

Search memory with query and result limit fields.

Commonly confused with

API request

Common misconception

A tool call may become an API request internally, but the concepts are not identical.

Function Calling Beginner
Plain-English definition

A model capability for returning structured requests that map to callable functions.

More precise definition

Function calling lets the model choose from registered operations and emit arguments instead of only natural-language text.

Example

The model returns a structured request to send a message.

Commonly confused with

Ordinary code function call

Common misconception

Provider “function calling” is usually a structured generation feature, not direct code execution by the model.

Multimodal Model Beginner
Plain-English definition

A model that can process or produce more than one kind of media.

More precise definition

Modalities may include text, images, audio, video, code, or sensor data.

Example

A model that reads an image and answers questions about it.

Commonly confused with

Omnimodal model

Common misconception

Multimodal does not mean every modality is equally strong.

Modality Beginner
Plain-English definition

A type of input or output media.

More precise definition

Common modalities include text, image, audio, video, and structured data.

Example

Text and image are two separate modalities.

Commonly confused with

Format

Common misconception

JPEG and PNG are formats within the image modality.

Vision-Language Model — VLM Intermediate
Plain-English definition

A model that connects visual information with language.

More precise definition

VLMs can analyze images, answer visual questions, caption scenes, or reason across text and visual inputs.

Example

Reading a chart and explaining its trend.

Commonly confused with

Image generator

Common misconception

A vision-language model interprets images; an image generator creates them.

Computer Vision Beginner
Plain-English definition

The field of teaching computers to interpret visual information.

More precise definition

Computer vision includes classification, detection, segmentation, tracking, recognition, and visual reasoning.

Example

Detecting objects in a photograph.

Commonly confused with

Image generation

Common misconception

Understanding images and generating images are different tasks.

Image Model Beginner
Plain-English definition

A model designed to analyze, transform, or generate images.

More precise definition

Image models may perform generation, editing, classification, segmentation, upscaling, or restoration.

Example

A model that creates an image from a text description.

Commonly confused with

Computer vision model

Common misconception

Image model is a broad category that includes both understanding and generation systems.

Diffusion Model Beginner
Plain-English definition

A generative model that learns to create data by reversing a noise process.

More precise definition

Diffusion models iteratively denoise a random or partially noised representation while following conditioning signals.

Example

Many text-to-image systems use diffusion.

Commonly confused with

GAN

Common misconception

The model does not retrieve a finished image from storage. It synthesizes one through iterative denoising.

Denoising Beginner
Plain-English definition

Removing noise step by step to produce a coherent output.

More precise definition

In diffusion generation, the model predicts how to move a noisy representation toward a cleaner sample.

Example

An image gradually emerges from noise over several steps.

Commonly confused with

Image cleanup

Common misconception

Diffusion denoising is the core generation process, not merely post-processing.

Latent Diffusion Intermediate
Plain-English definition

Diffusion performed in a compressed internal representation instead of directly on raw pixels.

More precise definition

Latent diffusion reduces compute by generating in latent space and decoding the result back into an image.

Example

A model denoises latent tensors before using a decoder to create pixels.

Commonly confused with

Pixel-space diffusion

Common misconception

Latent diffusion still produces pixel images; it simply performs much of the work in compressed space.

Guidance Scale — CFG Intermediate
Plain-English definition

A setting that controls how strongly generation follows the prompt.

More precise definition

Classifier-free guidance balances prompt conditioning against the model’s unconstrained generation tendency.

Example

Very high guidance may follow the prompt more literally but harm naturalness.

Commonly confused with

Prompt strength

Common misconception

Higher guidance is not always better.

Sampling Steps Intermediate
Plain-English definition

The number of denoising iterations used during diffusion generation.

More precise definition

More steps may improve quality up to a point but increase generation time.

Example

An image may be generated with 30 denoising steps.

Commonly confused with

Training steps

Common misconception

Sampling steps happen during generation. Training steps happen while learning weights.

Sampler Intermediate
Plain-English definition

The algorithm controlling how a diffusion model moves through denoising steps.

More precise definition

Different samplers trade speed, stability, detail, and visual character.

Example

Two samplers may produce different images from the same prompt and seed.

Commonly confused with

Language-model sampling

Common misconception

The word sampler is used differently in diffusion and language generation.

Text-to-Speech — TTS Beginner
Plain-English definition

Turning written text into spoken audio.

More precise definition

TTS systems model pronunciation, timing, prosody, and voice characteristics to synthesize speech.

Example

Generating narration from a script.

Commonly confused with

Speech-to-text

Common misconception

TTS creates audio from text. Speech-to-text transcribes audio into text.

Speech-to-Text — STT / ASR Beginner
Plain-English definition

Turning spoken audio into written text.

More precise definition

Automatic speech recognition converts speech signals into transcripts and may include timestamps, speaker labels, or confidence scores.

Example

Transcribing a voice note.

Commonly confused with

Text-to-speech

Common misconception

Speech recognition accuracy depends heavily on audio quality, accent, noise, and domain vocabulary.

Voice Model Beginner
Plain-English definition

A model used to synthesize or transform a speaking voice.

More precise definition

Voice models may generate speech, clone vocal characteristics, convert one voice to another, or control style and prosody.

Example

Generating speech in a consistent branded voice.

Commonly confused with

Speech recognition model

Common misconception

A voice model produces or transforms speech; a recognition model transcribes it.

Prosody Intermediate
Plain-English definition

The rhythm, stress, pacing, and intonation of speech.

More precise definition

Prosody carries emotional and linguistic information beyond the literal words.

Example

The same sentence can sound tender, sarcastic, or urgent through prosody.

Commonly confused with

Voice identity

Common misconception

A voice can match the speaker’s timbre while still having wrong prosody.

Latency Beginner
Plain-English definition

How long it takes to receive a result.

More precise definition

Latency may include network delay, queue time, preprocessing, model computation, tool execution, and output streaming.

Example

The first token arrives after 800 milliseconds.

Commonly confused with

Throughput

Common misconception

Low latency means fast individual responses. High throughput means processing lots of work over time.

Time to First Token — TTFT Intermediate
Plain-English definition

The delay before the first generated token appears.

More precise definition

TTFT measures initial responsiveness before the rest of the output is streamed.

Example

A model begins responding in 500 milliseconds but takes several seconds to finish.

Commonly confused with

Total latency

Common misconception

Fast first-token time does not guarantee fast total completion.

Tokens per Second — TPS Intermediate
Plain-English definition

How quickly a model generates tokens after it starts responding.

More precise definition

TPS measures generation throughput for one stream and depends on model size, hardware, batch load, and context length.

Example

A model generates 40 tokens per second.

Commonly confused with

Requests per second

Common misconception

Token speed and total system throughput are different measurements.

Throughput Beginner
Plain-English definition

How much work a system can complete over time.

More precise definition

Throughput may be measured in requests, tokens, images, or batches per second.

Example

A service handles 100 requests per second.

Commonly confused with

Latency

Common misconception

A system can have high throughput but poor latency for individual users.

Streaming Beginner
Plain-English definition

Sending output gradually as it becomes available.

More precise definition

Streaming improves perceived responsiveness by delivering tokens, audio, or events before the full result is complete.

Example

Text appears word by word in a chat interface.

Commonly confused with

Pagination

Common misconception

Streaming delivers one result progressively. Pagination divides a larger result set into pages.

Batching Intermediate
Plain-English definition

Grouping multiple requests or examples for joint processing.

More precise definition

Batching improves hardware use and throughput but can increase individual latency.

Example

A server combines several inference requests into one GPU batch.

Commonly confused with

Queueing

Common misconception

Batching groups work for processing. Queueing only stores work until it can be processed.

Quantization Intermediate
Plain-English definition

Reducing the numerical precision used to store or run a model.

More precise definition

Quantization lowers memory and compute requirements by representing weights or activations with fewer bits.

Example

Running a model in 4-bit precision on consumer hardware.

Commonly confused with

Compression

Common misconception

Quantization can reduce quality, but well-designed methods may preserve much of the model’s performance.

Precision — Numerical Precision Intermediate
Plain-English definition

How many bits are used to represent numerical values in computation.

More precise definition

Common formats include FP32, FP16, BF16, INT8, and INT4.

Example

A quantized model may use INT4 weights.

Commonly confused with

Evaluation precision

Common misconception

Numerical precision is unrelated to the classification metric also called precision.

Distillation — Knowledge Distillation Intermediate
Plain-English definition

Training a smaller model to imitate a stronger model.

More precise definition

A student model learns from outputs, probabilities, or representations produced by a teacher model.

Example

A compact model is trained on answers generated by a larger model.

Commonly confused with

Fine-tuning

Common misconception

Distillation transfers behavior patterns, not a perfect copy of all capabilities.

Teacher Model Intermediate
Plain-English definition

A stronger model used to guide another model’s training.

More precise definition

Teacher outputs or internal signals become training targets for a student model.

Example

A large model generates demonstrations for a smaller model.

Commonly confused with

Judge model

Common misconception

A teacher model supplies learning signals. A judge model scores outputs.

Student Model Intermediate
Plain-English definition

A model trained to learn from a teacher model.

More precise definition

Student models aim to preserve useful behavior while reducing size, cost, or latency.

Example

A smaller assistant model trained from a larger teacher.

Commonly confused with

Fine-tuned model

Common misconception

A student model may have a different architecture or size from its teacher.

Router — Model Router Intermediate
Plain-English definition

A component that chooses which model, tool, or workflow should handle a request.

More precise definition

Routers use rules, classifiers, cost targets, capability signals, or learned policies to dispatch work.

Example

Simple tasks go to a fast model; difficult tasks go to a stronger model.

Commonly confused with

Orchestrator

Common misconception

A router selects a path. An orchestrator coordinates a sequence of work.

Model Cascade Intermediate
Plain-English definition

Using multiple models in sequence or by escalation.

More precise definition

A cascade may start with a cheaper model and move to a stronger one only when confidence or quality is insufficient.

Example

A small model attempts classification before escalating uncertain cases.

Commonly confused with

Ensemble

Common misconception

A cascade routes sequentially. An ensemble combines multiple model outputs.

Ensemble Intermediate
Plain-English definition

Combining outputs from multiple models.

More precise definition

Ensembles aggregate votes, scores, rankings, or generated candidates to improve robustness.

Example

Three classifiers vote on the final label.

Commonly confused with

Model cascade

Common misconception

Ensembles combine parallel opinions. Cascades choose or escalate between models.

Fallback Beginner
Plain-English definition

A backup behavior used when the preferred path fails.

More precise definition

Fallbacks may use another model, cached output, simpler logic, or a human review path.

Example

Use a smaller local model if the cloud provider is unavailable.

Commonly confused with

Retry

Common misconception

Retry repeats the same path. Fallback switches to a different path.

Graceful Degradation Intermediate
Plain-English definition

Continuing with reduced capability instead of failing completely.

More precise definition

A system degrades gracefully by preserving core function when optional services or stronger models are unavailable.

Example

Return text without voice synthesis when the voice service fails.

Commonly confused with

Fallback

Common misconception

Fallback is the alternate mechanism. Graceful degradation is the overall user-visible behavior.

Classifier Beginner
Plain-English definition

A model or rule system that assigns an input to a category.

More precise definition

Classifiers may output one label, multiple labels, probabilities, or confidence scores.

Example

Classifying a message as urgent or routine.

Commonly confused with

Generator

Common misconception

A classifier chooses among labels. A generator creates new content.

Classification Beginner
Plain-English definition

Assigning input to one or more categories.

More precise definition

Classification may be binary, multiclass, multilabel, hierarchical, or probabilistic.

Example

Assigning a memory node to identity, build, or relationship domains.

Commonly confused with

Clustering

Common misconception

Classification uses known label categories. Clustering discovers groups without predefined labels.

Confidence Score Beginner
Plain-English definition

A number indicating how strongly a system favors an answer or label.

More precise definition

Confidence scores may come from probabilities, margins, heuristics, calibration models, or model-generated estimates.

Example

The classifier assigns 0.82 confidence to build.

Commonly confused with

Correctness

Common misconception

High confidence does not guarantee the answer is correct.

Calibration Intermediate
Plain-English definition

How well confidence scores match real-world accuracy.

More precise definition

A calibrated model that reports 80% confidence should be correct about 80% of the time across similar cases.

Example

Comparing predicted confidence with actual outcomes.

Commonly confused with

Accuracy

Common misconception

A highly accurate model can still be poorly calibrated.

Threshold Beginner
Plain-English definition

A cutoff used to decide whether a score counts as a match or action.

More precise definition

Thresholds convert continuous scores into decisions and control trade-offs between false positives and false negatives.

Example

Only classify as urgent when confidence is above 0.8.

Commonly confused with

Top-p

Common misconception

Threshold is a decision cutoff, not a sampling parameter.

Hallucination Beginner
Plain-English definition

Model output that is unsupported, fabricated, or incorrectly presented as true.

More precise definition

Hallucinations can involve invented facts, citations, fields, events, tool results, or reasoning claims.

Example

The model invents an API field that does not exist.

Commonly confused with

Error

Common misconception

Not every wrong answer is called a hallucination; the term is most useful when the model generates unsupported content.

Grounding Beginner
Plain-English definition

Anchoring model output in trusted evidence or data.

More precise definition

Grounding may use retrieved documents, tool results, databases, citations, or verified context.

Example

The model reads the actual tool schema before explaining required fields.

Commonly confused with

Training

Common misconception

Grounding supplies evidence at runtime. It does not permanently retrain the model.

Reasoning Model Beginner
Plain-English definition

A model optimized to spend more computation on multi-step problem solving.

More precise definition

Reasoning models may use longer internal deliberation, verification, search, or tool use before producing an answer.

Example

Solving a complex coding or planning problem with multiple constraints.

Commonly confused with

Agent

Common misconception

A reasoning model is still a model. An agent is the larger system that can plan and act with tools.

Deliberation Intermediate
Plain-English definition

Additional model computation used before finalizing an answer.

More precise definition

Deliberation may involve internal candidate generation, checking, decomposition, or iterative refinement.

Example

A model spends more compute on a difficult proof than on a simple lookup.

Commonly confused with

Visible explanation

Common misconception

A model’s user-visible explanation is not necessarily a literal transcript of its internal computation.

Chain-of-Thought Prompting — CoT Intermediate
Plain-English definition

A prompting technique that encourages a task to be broken into intermediate steps.

More precise definition

Chain-of-thought prompting uses examples or instructions that elicit multi-step problem solving before the final answer.

Example

Ask the model to solve a problem step by step, then provide the conclusion.

Commonly confused with

Private model reasoning

Common misconception

A written chain of thought is generated text and should not be treated as a guaranteed record of internal computation.

Decomposition Beginner
Plain-English definition

Breaking a complex task into smaller parts.

More precise definition

Decomposition improves planning, delegation, testing, and reasoning by isolating subtasks and dependencies.

Example

Separate schema analysis, implementation, tests, and review.

Commonly confused with

Delegation

Common misconception

Decomposition divides the task. Delegation assigns the pieces.

Self-Consistency Intermediate
Plain-English definition

Generating multiple solution paths and choosing the result that agrees most often.

More precise definition

Self-consistency samples several candidate reasonings or answers and aggregates them.

Example

Run the same math problem several times and vote on the final answer.

Commonly confused with

Consistency

Common misconception

Agreement among outputs can improve reliability but does not guarantee truth.

Reranking Intermediate
Plain-English definition

Reordering candidate results using a stronger scoring method.

More precise definition

A reranker evaluates retrieved items or generated candidates after the initial selection stage.

Example

Retrieve 20 passages, then rerank the best five.

Commonly confused with

Routing

Common misconception

Routing chooses a path. Reranking orders candidates.

Evaluation — Eval Beginner
Plain-English definition

A structured way to measure model or system performance.

More precise definition

Evals may use exact answers, rubrics, human ratings, judge models, statistics, or task-specific metrics.

Example

Measure whether the model selects the correct tool across 200 prompts.

Commonly confused with

Test

Common misconception

Tests often expect exact behavior. Evals may measure graded quality or probabilistic performance.

Benchmark Beginner
Plain-English definition

A fixed evaluation set used to compare models or versions.

More precise definition

Benchmarks standardize tasks, data, scoring, and reporting for repeatable comparison.

Example

Compare two models on the same coding benchmark.

Commonly confused with

Eval suite

Common misconception

A benchmark score may not predict performance in your real product.

Rubric Beginner
Plain-English definition

The criteria used to judge output quality.

More precise definition

A rubric defines dimensions, scoring scales, failure conditions, and examples for evaluators.

Example

Score correctness, relevance, completeness, and tone from 1 to 5.

Commonly confused with

Prompt

Common misconception

A vague rubric produces unreliable evaluation.

Judge Model — LLM-as-a-Judge Intermediate
Plain-English definition

A model used to score or compare another model’s output.

More precise definition

Judge models apply a rubric to outputs and may produce scores, rankings, or explanations.

Example

A judge model ranks two answers for factuality.

Commonly confused with

Reward model

Common misconception

Judge models can be biased, inconsistent, or vulnerable to style effects.

Regression Eval Intermediate
Plain-English definition

An evaluation that checks whether previously good behavior got worse.

More precise definition

Regression evals preserve critical examples and metrics across model, prompt, tool, or product changes.

Example

Confirm the new prompt still preserves persona and schema adherence.

Commonly confused with

Benchmark

Common misconception

Regression evals protect known behavior; benchmarks compare broader capability.

Drift Intermediate
Plain-English definition

Behavior gradually moving away from the intended standard.

More precise definition

Drift can come from model updates, prompt changes, data changes, retrieval shifts, user adaptation, or evolving system context.

Example

A persona becomes flatter across several releases.

Commonly confused with

Regression

Common misconception

A regression is a specific backward step. Drift may be gradual and cumulative.

Emergent Behavior Intermediate
Plain-English definition

Behavior that appears from the system without being explicitly programmed as a single rule.

More precise definition

Emergent behavior arises from interactions among training data, architecture, scale, prompting, tools, and environment.

Example

A model develops a useful strategy not directly described in the training objective.

Commonly confused with

Unexpected behavior

Common misconception

Emergent does not mean supernatural or conscious.

No terms match that search yet. Try a broader word.