AI Coding Glossary

The AI Coding Glossary is a comprehensive collection of common AI coding concepts and terms. It’s a quick reference for both beginners and experienced developers looking for definitions and refreshers related to AI coding.

It covers the fundamental concepts, terminology, and patterns that are essential for understanding AI-assisted programming. From core machine-learning concepts like transformers and tokenization to practical coding patterns like prompt engineering and chain of thought (CoT), this glossary helps you navigate the vocabulary of AI programming.

Whether you’re working with large language models (LLMs), implementing RAG systems, or optimizing prompts for better code generation, these terms form the foundation of modern AI-enhanced development practices.

Not sure which term you need? Describe what you’re trying to do, and Mentor AI will point you to the right entries.

  • activation-aware weight quantization (AWQ) A post-training quantization method that compresses a model’s weights to 4 bits, using activation statistics to pick which ones to protect from rounding.
  • activation function A nonlinear mapping applied to a neuron’s weighted sum that enables neural networks to learn complex relationships.
  • Adam optimizer An optimization algorithm that adapts each parameter’s step size using running averages of recent gradients and their squares.
  • agent A system that perceives, decides, and acts toward goals, often looping over steps with tools, memory, and feedback.
  • agentic coding An approach to software development in which AI agents plan, write, run, and iteratively improve code.
  • agent loop A repeating cycle in which an AI agent gathers context, lets the model choose a step, runs it, and feeds the result back until the task is done.
  • anomaly detection A technique that identifies data points, events, or sequences deviating from a dataset’s established pattern of normal behavior.
  • artificial intelligence (AI) The field of building machines and software that perform tasks requiring human-like intelligence.
  • attention mechanism A neural network operation that computes a weighted sum of value vectors based on the similarity between a query and a set of keys.
  • augmented coding A software development approach where a developer uses AI coding agents to write most of the code while keeping responsibility for its quality, tests, and design.
  • autoencoder A neural network that compresses its input into a compact code and reconstructs it, learning useful representations without labeled data.
  • autoregressive generation A method in which a model produces a sequence one token at a time, with each token conditioned on all previously generated tokens.
  • bias A systematic deviation from truth or fairness.
  • BM25 A ranking function that scores how well a document matches a keyword query, weighting rare terms heavily and saturating repeated ones.
  • chain of thought (CoT) The intermediate reasoning steps a model generates before its final answer, whether elicited by prompting or produced natively by reasoning models.
  • computer vision (CV) A field of computer science and artificial intelligence that enables computers to derive meaning from images and video.
  • confusion matrix A table that compares a classifier’s predictions against the true labels, showing how often the model is wrong and which classes it confuses.
  • context engineering The systematic design and optimization of the information given to a model at inference time so it can answer effectively.
  • context window The maximum span of tokens that a language model can consider at once.
  • continuous batching A serving technique that adds and drops requests from a running inference batch after every generation step instead of waiting for the batch to finish.
  • convolutional neural network (CNN) A neural network that uses local receptive fields and shared weights to process structured signals such as images.
  • cosine similarity A measure of similarity between two vectors that depends only on the angle between them, ignoring their magnitudes.
  • cross entropy loss A classification loss function that measures the gap between a model’s predicted probabilities and the correct labels.
  • curse of dimensionality A phenomenon in high-dimensional data where the space grows so fast that points spread thinly and distances stop separating near neighbors from far ones.
  • Data exfiltration The unauthorized transfer of data out of a system, which in AI applications usually means injected instructions steering an agent into leaking private context.
  • decision tree A model that predicts a label or a value by following a branching series of yes-or-no tests on a data point’s features down to a leaf.
  • deep learning A subfield of machine learning in which many-layered neural networks learn their own representations of raw data instead of hand-designed features.
  • embedding A learned vector representation that maps discrete items, such as words, sentences, documents, images, audio, video, or users, into a continuous space.
  • evaluation The process of measuring how well an AI system or model meets its objectives.
  • feature engineering A process that transforms raw data into the input features that a machine learning model learns from.
  • few-shot learning A setting where a model adapts to a new task using only a small number of labeled examples.
  • fine-tuning The process of adapting a pre-trained model to a new task or domain.
  • FlashAttention (v2) An attention algorithm that computes exact attention in on-chip GPU memory tiles, with a 2023 redesign that roughly doubles the original’s speed.
  • function calling A model feature that lets the model choose a tool and emit structured arguments, then your app or the provider runs the call and returns the results.
  • generative model A model that learns a data distribution so it can generate new samples or assign probabilities to observations.
  • generative pre-trained transformer (GPT) Autoregressive language models that use the transformer architecture and are pre-trained on large text corpora.
  • GGUF A binary file format that packages a model’s weights, tokenizer, and metadata into a single self-contained file for fast local inference.
  • GPTQ A post-training quantization method that compresses a trained model’s weights to 3 or 4 bits in a single pass, without retraining.
  • gradient descent A first-order iterative optimization method
  • guardrails Application-level policies and controls that constrain how a model or agent behaves.
  • hallucination When a generative model produces confident but false or unverifiable content and presents it as fact.
  • harness A software layer around a language model that manages its context, runs the tools it requests, and enforces permissions so it can act as an agent.
  • Hybrid search A retrieval technique that combines keyword and vector search over the same corpus and fuses their ranked results into a single list.
  • in-context learning (ICL) A behavior where a model performs a new task by conditioning on instructions and possibly a few input-output examples.
  • inference The phase where a trained model processes new inputs to produce outputs.
  • jailbreak A method of prompting that bypasses model safety constraints to elicit disallowed or unintended behavior.
  • key-value cache (KV cache) A memory buffer that stores the key and value vectors a transformer has already computed so that generating each new token doesn’t repeat the work.
  • k-nearest neighbors (k-NN) algorithm A non-parametric method that classifies or predicts a value for a new data point using the k stored examples closest to it.
  • large language model (LLM) A neural network trained to predict the next token, used for general-purpose language, multimodal, and tool-using tasks.
  • large reasoning model (LRM) A language model optimized for multi-step problem-solving.
  • latency The elapsed time between a request and the observable response.
  • linear regression A machine learning model that predicts a continuous target value as a weighted sum of input features, fit by minimizing the squared prediction error.
  • LLM observability The practice of collecting and correlating telemetry about large language model applications.
  • logistic regression A classification model that estimates the probability of a categorical outcome by passing a weighted sum of input features through a sigmoid curve.
  • loss function A scalar objective that measures prediction error and shapes gradients to guide model training.
  • low-rank adaptation (LoRA) A parameter-efficient fine-tuning method that freezes a model’s pretrained weights and trains small low-rank adapter matrices instead.
  • machine learning A subfield of AI that builds models that improve their performance on a task by learning patterns from data.
  • machine learning operations (MLOps) A set of practices that applies DevOps automation to the machine learning lifecycle, covering data, training, deployment, and production monitoring.
  • Model Context Protocol (MCP) An open, client-server communication standard that lets AI applications connect to external tools and data sources.
  • naive Bayes A family of probabilistic classifiers that apply Bayes’ theorem while assuming that features are independent of one another given the class label.
  • natural language processing (NLP) A field of computer science and artificial intelligence that enables computers to analyze, interpret, generate, and interact with human language in text and speech.
  • nearest neighbor The data point in a reference set that has the smallest distance to a query point.
  • neural network A computational model composed of layered, interconnected units that learn input-to-output mappings.
  • one-hot encoding A representation that encodes a categorical value as a vector of zeros with a single one marking which category is present.
  • optical character recognition (OCR) A process that converts images of text, such as scanned pages or photographs, into machine-readable character data.
  • OWASP LLM Top 10 A community-maintained list of the ten most critical security risks in applications built on large language models.
  • PagedAttention An attention algorithm that stores the key-value cache in fixed-size, non-contiguous blocks, borrowing paging from operating-system virtual memory.
  • parameter A learned internal value of a model, such as a weight or bias.
  • prompt The text, structured message, or multimodal input that tells a generative model what to do.
  • prompt engineering The practice of designing and refining prompts for generative models.
  • prompt injection Input that alters a model’s or model-integrated app’s behavior so that it ignores its original instructions and performs unintended actions, usually crafted deliberately by an attacker but sometimes triggered inadvertently.
  • quantization A compression technique that reduces the numerical precision of a model’s weights to shrink memory use and speed up inference.
  • quantized low-rank adaptation (QLoRA) A memory-efficient fine-tuning method that quantizes a model’s frozen weights to 4 bits and trains small low-rank adapters on top.
  • query expansion A retrieval technique that enriches a search query with extra terms so a system finds relevant documents even when the original wording doesn’t match.
  • random forest An ensemble model that combines the predictions of many randomized decision trees to improve accuracy and control overfitting.
  • reasoning model A generative model tuned to solve multi-step problems.
  • recurrent neural network (RNN) A neural network that processes sequences by applying the same computation at each step.
  • reinforcement learning from AI feedback (RLAIF) A training technique that aligns large language models by having another AI model, rather than human annotators, rank model outputs.
  • reinforcement learning from human feedback (RLHF) A training technique that aligns large language models with human preferences using ranked comparisons of model outputs.
  • reinforcement learning (RL) A learning approach where an agent improves decisions by interacting with an environment and maximizing cumulative reward.
  • reranker A model that reorders retrieved documents by relevance to a query, forming the second stage of two-stage retrieval in RAG pipelines.
  • retrieval-augmented generation (RAG) A technique that improves a model’s outputs by retrieving relevant external documents at query time and feeding them into the model.
  • sandbox A restricted environment that limits which files, programs, and network destinations an AI agent can reach while it works.
  • self-attention A mechanism that compares each token to all others and mixes their information using similarity-based weights.
  • skill A packaged folder of instructions and optional scripts that an AI agent loads on demand when a task needs specialized know-how.
  • speculative decoding An inference technique that has a small draft model propose several tokens at once so the larger model can verify them in a single pass.
  • stochastic gradient descent (SGD) An optimization algorithm that updates a model’s parameters from one training example or a small batch at a time rather than the full dataset.
  • structured output Model responses that conform to a specified format, such as a JSON Schema.
  • subagent A separate AI agent that a main agent launches to handle a side task in its own context window and then return a summary.
  • supervised learning A machine learning paradigm that trains a model on labeled examples so it can predict the label for new, unseen data.
  • system prompt A block of instructions that outranks user inputs and establishes a model’s role, goals, constraints, and style.
  • tagging The process of assigning one or more discrete labels to data items so that models and tools can learn from them.
  • t-distributed stochastic neighbor embedding (t-SNE) A dimensionality-reduction algorithm that projects high-dimensional data onto a two- or three-dimensional map, keeping points close to their nearest neighbors.
  • telemetry The automated collection and transmission of measurements or event data from remote or distributed systems.
  • temperature A decoding parameter that rescales model logits before sampling.
  • tensor parameter A learned multi-dimensional array that a model updates during training to shape its computations.
  • term frequency-inverse document frequency (TF-IDF) A weighting scheme that scores how important a word is to one document by balancing its frequency there against how common it is across a collection of documents.
  • text corpora Collections of machine-readable text that serve as data resources for linguistics and natural language processing, ranging from small curated collections to web-scale crawls.
  • throughput The rate at which a system completes useful work.
  • token The smallest unit of data processed and generated by NLP systems and language models.
  • tokenization The process of converting raw text into a sequence of discrete tokens.
  • tool use The ability of a language model to call external functions or services during generation.
  • training The process of fitting a model’s parameters to data by minimizing a loss function.
  • transfer learning A machine learning approach that reuses the knowledge a model gained on one task to improve its performance on a different but related task.
  • transformer A neural network model that uses self-attention to handle sequences without recurrence or convolutions.
  • transformer architecture A neural network design that models sequence dependencies using self-attention instead of recurrence or convolutions.
  • unsupervised learning A machine learning paradigm that finds structure in unlabeled data, discovering groupings, compressed representations, and outliers without human annotation.
  • vector An ordered array of numbers that represents a point, magnitude, and direction.
  • vector database A data system optimized for storing and retrieving embedding vectors.
  • vector space A set of objects called vectors that can be added together and multiplied by scalars.
  • vibe coding An AI-assisted programming style where a developer describes goals in natural language and accepts AI-generated code without reviewing or fully understanding it.
  • vision transformer (ViT) A neural network architecture that applies the transformer to images by splitting them into fixed-size patches treated as a sequence of tokens.
  • weight A learned scalar or tensor that scales signals in a model and is updated during training to shape predictions.
  • zero-shot learning (ZSL) A machine learning setting where a model handles classes or tasks that were not encountered during training.
Reference Share Feedback