A glossary of AI terms, rewritten 2026-10-05: classical ML first, then modern LLM, agent, training, safety and infrastructure vocabulary, then abbreviations (A-Z). One term per line; a term links to a vault note where one exists. Definitions are neutral and deliberately free of model names, prices and benchmark scores, which go stale. They are working definitions summarised for orientation; the attributions and dates cited in them come from the papers and pages named in Sources, and fast-moving agent terms are not standardised.

1. Classical ML and AI terms

Activation Function= A function applied to a neuron’s weighted input to produce its output, introducing the non-linearity that lets networks learn complex patterns (e.g. ReLU, sigmoid).
Agent (classical)= An entity that perceives an environment and acts on it to achieve goals; in the LLM sense see AI Agent below.
artificial-intelligence (AI)= The simulation of human intelligence processes by machines, especially computer systems. These processes include learning, reasoning, and self-correction.
Algorithm= A set of rules to be followed in calculations or other problem-solving operations, especially by a computer.
Algorithm Bias= Systematic and repeatable errors in a computer system that create unfair outcomes such as privileging one arbitrary group over others.
Artificial Neural Network= A computing model of layered, connected units (neurons) whose weights are learned from data.
Attention Mechanisms= The part of a Transformer that weights how relevant each token in the input is to every other token; introduced for translation and central to large-language-models.
Autoencoder= A neural network trained to compress its input into a small latent representation and reconstruct it, used for dimensionality reduction and generative modelling.
Autonomous Systems= Systems capable of performing tasks without human intervention.
autonomous-vehicles= Vehicles that drive with little or no human input; see the note.
Backpropagation= The algorithm that computes the gradient of a loss with respect to a network’s weights, layer by layer, so gradient descent can train it.
Bias-Variance Tradeoff= The tension between a model too simple to fit the data (high bias) and one so flexible it fits noise (high variance).
Big Data= Data sets too large or complex for traditional processing tools, commonly characterised by volume, velocity and variety.
Binary Classification= A supervised task that assigns each input to one of two classes (e.g. spam or not spam).
Chatbot= A program that converses with users in natural language; modern ones are built on large-language-models.
Classification= A supervised task that assigns each input to one of a set of discrete classes.
Clustering= An unsupervised task that groups similar data points without predefined labels.
Cognitive Computing= Systems that are designed to simulate human thought processes in a computerized model.
computer-vision= An interdisciplinary field that deals with how computers can gain high-level understanding from digital images or videos. It seeks to automate tasks that the human visual system can do.
Convolutional Neural Network (CNN)= A type of deep neural network often used in image recognition and processing that is specifically designed to process pixel data.
Cross-Validation= Evaluating a model by repeatedly training on part of the data and testing on the held-out remainder.
Data Mining= The process of discovering patterns in large data sets using methods from statistics, machine learning and databases.
Data Science= An interdisciplinary field that uses statistics, programming and domain knowledge to extract insight from data.
Decision Tree= A model that predicts an outcome by following a tree of feature-based yes/no splits from root to leaf.
Deep Learning= A subset of machine learning that uses neural networks with many layers (deep neural networks) to analyze various factors in large amounts of data.
Dimensionality Reduction= Mapping data to fewer variables while keeping most of its structure (e.g. PCA).
Ensemble Learning= Combining several models (e.g. bagging, boosting) to get better predictions than any single one.
Epoch= One full pass of the training algorithm over the entire training data set.
Evolutionary Algorithm= An optimisation method inspired by biological evolution (selection, mutation, recombination of candidate solutions).
Expert System= An artificial intelligence program that has expert-level knowledge about a particular domain and can emulate expert decision-making abilities.
Explainable AI (XAI)= An emerging field in machine learning and artificial intelligence aiming at making black box systems transparent, so users can understand how models make decisions.
Feature Engineering= Selecting, transforming or creating input variables to make a machine-learning model work better.
Feature Extraction= Deriving a reduced, informative set of features from raw data.
Federated Learning= Training a shared model across many devices or servers that keep their data local and only share model updates.
Fuzzy Logic= A form of many-valued logic which deals with reasoning that is approximate rather than fixed and exact.
Generative Adversarial Networks (GANs)= A class of machine learning frameworks designed by pitting two neural networks against each other in order to generate new, synthetic instances of data that can pass for real data.
Genetic Algorithm= Search heuristics that mimic the process of natural selection to generate useful solutions to optimization and search problems.
Gradient Descent= An iterative optimisation method that adjusts parameters in the direction that most reduces the loss.
Heuristic= A practical rule of thumb that finds a good-enough solution quickly without guaranteeing the optimum.
Hyperparameter= A setting chosen before training (e.g. learning rate, number of layers) rather than learned from data.
Image Recognition= The task of identifying what an image contains (an object, scene, face or text), typically by classifying it with a computer-vision model such as a CNN.
machine-learning (ML)= A subset of AI that involves the creation of algorithms that can modify themselves without human intervention to produce desired outputs by feeding on data inputs.
Loss Function= A measure of how far a model’s predictions are from the targets, which training tries to minimise.
Natural Language Processing (NLP)= A field at the intersection of computer science, artificial intelligence, and computational linguistics which focuses on interactions between computers and human languages.
Neural Network= A series of algorithms that attempt to recognize underlying relationships in a set of data through a process that mimics the way the human brain operates. See neural-networks and Artificial Neural Network.
Object Detection= A computer-vision task that locates objects in an image and labels each with a class.
Optimization Algorithms= Procedures that search for the parameter values that minimise or maximise an objective function; in machine learning, gradient descent and its variants (SGD, Adam) are the common ones.
Overfitting= When a model fits its training data, including noise, so well that it generalises poorly to new data.
Pattern Recognition= The automated detection of regularities in data and their use to classify or act on it; the older name for much of what is now called machine learning.
Perceptron= The simplest neural network unit: a weighted sum of inputs passed through a threshold, used as a linear binary classifier.
Precision and Recall= Classification metrics: precision is the share of predicted positives that are correct, recall the share of actual positives that were found.
Predictive Analytics= Using historical data, statistics and machine learning to estimate the likelihood of future outcomes.
Recommender System= A system that predicts which items (products, videos, articles) a user is likely to prefer, using collaborative filtering, content features or both.
Recurrent Neural Network (RNN)= A type of neural network where connections between nodes form a directed graph along a temporal sequence, allowing it to exhibit temporal dynamic behavior. Used for tasks like language modeling and translation.
Regression= A supervised task that predicts a continuous numeric value.
Regularization= Techniques (e.g. weight decay, dropout) that discourage overfitting by constraining the model.
Reinforcement Learning= An area of machine learning concerned with how software agents ought to take actions in an environment so as to maximize some notion of cumulative reward.
Robotics= The branch of technology that deals with the design, construction, operation, and application of robots.
Self-supervised Learning= Learning from labels derived from the data itself (e.g. predicting the next word or a masked patch); the basis of pre-training.
Semantic Analysis= Interpreting the meaning of text (as opposed to its form), e.g. word senses and relations.
Semi-supervised Learning= An approach to machine learning that combines a small amount of labeled data with a large amount of unlabeled data during training. It falls between supervised learning (fully labeled data) and unsupervised learning (no labels).
Sentiment Analysis= An NLP task that classifies the opinion or emotion expressed in text (e.g. positive, negative, neutral).
Sequential Model= A model that processes ordered data such as text, audio or time series, where earlier elements affect later ones (RNNs, Transformers); in Keras, also a plain stack of layers.
Supervised Learning= A type of machine learning algorithm that uses a known dataset called the training dataset, which includes input data and response values (outputs) to train a model which makes predictions or decisions without human intervention.
Support Vector Machine= A supervised model that separates classes with the maximum-margin boundary.
TensorFlow= An open-source machine-learning framework from Google for building and training neural networks; see tensorflow-ecosystem-note and jax for the alternative.
Transfer Learning= Reusing a model trained on one task as the starting point for another task.
Unsupervised Learning= Machine learning algorithms used when the information used to train is neither classified nor labeled. It studies how systems can infer a function to describe hidden structure from unlabeled data.
Virtual Assistant= Software that takes spoken or written requests and carries out tasks or answers questions for a user; older ones were rule-based, current ones are built on large-language-models. See also AI agent.
Neurosymbolic AI= Approaches that combine neural networks with symbolic reasoning or knowledge representation.
ontology= A formal, shared specification of the concepts in a domain and the relations between them; see resource-description-framework and owl.

2. Modern LLM terms

Large Language Model= A neural network, almost always a Transformer, trained on very large text corpora to predict the next token, which can then follow instructions and generate text.
Transformer= The neural architecture (Vaswani et al., 2017) built on self-attention rather than recurrence; the basis of modern LLMs. See transformer-models-2023.
Generative Pretrained Transformer= A decoder-only Transformer pre-trained to predict the next token, then adapted for tasks.
Tokenizer= The component that splits text into tokens and maps them to integer IDs, usually with a learned subword vocabulary.
BPE (Byte-Pair Encoding)= A tokenizer algorithm that builds a subword vocabulary by repeatedly merging the most frequent adjacent symbol pairs.
Token= The unit a language model reads and writes: text is split into tokens, each mapped to an integer index in the model’s vocabulary. Billing and limits are usually per token.
Context Window= The span of tokens a model can attend to in one request (prompt plus output). See in-context-learning.
Context Rot= The degradation in accuracy and instruction-following as the context fills with more, and often irrelevant, tokens, even within the advertised window.
Context Compaction= Summarising or trimming earlier conversation so a long-running session can continue within the context window.
context-engineering= Designing everything that goes into the model’s context (instructions, retrieved data, tools, memory, history) so it can do the task reliably; a broader successor to prompt writing.
Prompt Engineering= Crafting the instructions and examples given to a model to get better outputs; see chain-of-thought-prompting and meta-prompting.
System Prompt= Developer-set instructions placed ahead of the conversation that define a model’s role, rules and style.
In-Context Learning= A model’s ability to pick up a task from examples or instructions in the prompt without any weight update.
Few-shot / Zero-shot Prompting= Giving a model a few worked examples (few-shot) or none (zero-shot) in the prompt.
Chain of Thought (CoT)= Having a model write out intermediate reasoning steps before the answer, which improves accuracy on multi-step problems.
tree-of-thought-prompting= A prompting method that explores and evaluates several reasoning branches instead of one chain.
Reasoning Model= An LLM trained to produce extended step-by-step reasoning before its final answer, for harder tasks.
Test-Time Compute= Spending more computation at inference (longer reasoning, sampling several answers, search) to improve an answer, instead of only scaling training.
Prompt Caching= Reusing the already-computed internal state for a repeated prompt prefix so later requests are faster and cheaper.
inference-and-serving= Stored attention keys and values for tokens already processed, so each new token does not recompute the whole sequence; it grows with context length.
Structured Outputs= Constraining a model to emit output that conforms to a schema (e.g. JSON), so software can parse it reliably.
Function Calling= The model emits structured calls to external functions or APIs, which the application executes and returns; the open standard for exposing tools is model-context-protocol.
retrieval-augmented-generation= Fetching relevant documents at query time and putting them in the prompt so the model answers from them; see retrieval-augmented-generation-overview.
GraphRAG= RAG that retrieves over a knowledge graph or graph-derived summaries, not just text chunks, to answer questions needing relations across documents.
Embedding= A numeric vector that represents a token, word, passage or image so that similar items are close together; used for search and RAG.
Vector Database= A database that stores embeddings and returns the nearest ones to a query vector, usually via approximate indexes (see ann-index-algorithms).
Chunking= Splitting documents into pieces (by size, structure or meaning) before embedding them for retrieval.
Reranking= A second retrieval stage in which a more accurate model rescores the top candidates from a fast first-stage search.
Hybrid Search= Combining keyword (lexical) and vector search and merging the rankings.
Semantic Search= Retrieval by meaning (embedding similarity) rather than by exact keyword match.
Temperature= A sampling setting that controls randomness: low gives more deterministic output, high more varied.
Sampling (top-k / top-p)= Methods that restrict next-token choice to the k most likely tokens or the smallest set whose probability mass reaches p.
Logits= The raw, unnormalised scores a model produces for each vocabulary token before softmax.
Pre-training= The first, large-scale training phase on broad data (usually next-token prediction) that gives a model general capability.
Post-training= Everything after pre-training that shapes behaviour: instruction tuning, preference optimisation, reinforcement learning.
Instruction Tuning= Fine-tuning on instruction-and-response pairs so a base model follows requests.
Hallucination= Fluent, plausible output that is factually wrong or unsupported by the source.
Grounding= Tying a model’s answer to verifiable sources (retrieved documents, tool results) so claims can be checked.
Multimodal Model= A model that handles more than one data type, such as text, images, audio and video.
vision-models= A model that takes images (and text) as input and answers or reasons about them in language.
Diffusion Model= A generative model that learns to reverse a gradual noising process, producing images, video or audio by iterative denoising.
World Model= A model that learns how an environment evolves and responds to actions, so it can simulate or plan; also used for generative video that simulates scenes.
Mixture of Experts (MoE)= An architecture in which a gating network routes each input to a few specialised sub-networks, so only part of the model’s parameters is used per token.
State-Space Model (SSM)= A sequence model built on linear recurrent dynamics that scales linearly with sequence length, an alternative to attention.
Mamba= A state-space architecture (Gu and Dao, 2023) with input-dependent (“selective”) state updates for efficient long sequences.
Hybrid Architecture= A model mixing layer types, typically attention layers with state-space or linear-attention layers (or dense with MoE), to balance quality and speed.
Dense Model= A model that uses all its parameters for every token, as opposed to an MoE.
Multi-Token Prediction= Training or decoding a model to predict several future tokens at once; see multi-token-prediction.
Encoder / Decoder= Transformer halves: an encoder builds representations of input, a decoder generates output token by token; most LLMs are decoder-only.
Latent Space= The internal vector space in which a model represents data in compressed form.

3. Agents and agentic engineering

AI Agent= A software system that uses a model (usually an LLM) to pursue a goal by choosing and executing actions, often via tools, in a loop.
Agentic AI= The approach or class of systems in which AI plans, acts with tools, observes results and iterates with limited human steering, as opposed to single-turn question answering.
Agent Harness= The code and configuration around a model that turns it into an agent: the loop, tools, context management, permissions, memory and hooks.
Subagent= An agent started by another agent to handle a delegated subtask in its own context, returning only a result to the parent.
Skill (agent skill)= A packaged folder of instructions, scripts and resources that an agent loads on demand when a task needs it.
MCP (Model Context Protocol)= An open protocol for connecting AI applications to tools, data and prompts through servers, so one integration works across clients.
A2A (Agent2Agent protocol)= An open protocol for communication and task hand-off between independent agents, possibly from different vendors; see also acp-agent-communication-protocol.
AGENTS.md= A plain-markdown file in a repository that gives coding agents project instructions (build, test, conventions); see also claude-md-and-agent-instructions.
Agentic AI Foundation= A neutral foundation hosting open agent standards; see the note for current members.
Orchestrator= An agent or program that splits work, assigns it to worker agents or tools, and combines results; see ai-code-orchestrators-2026.
Multi-Agent System= Several cooperating agents, each with a role, working on one task; see multi-agent-systems.
Workflow (agentic)= A fixed, developer-defined sequence of model and tool steps, as opposed to an agent that decides its own steps.
ReAct= A pattern that interleaves reasoning text with tool actions and observations (Yao et al., 2022).
Planning= An agent step that breaks a goal into ordered subtasks before or while executing them; see task-decomposition-patterns.
Memory (short-term)= State an agent keeps within a session: the current context, scratchpad and recent history.
Memory (long-term)= Information persisted across sessions (facts, preferences, past outcomes) and retrieved when relevant.
Computer Use= An agent controlling a graphical desktop by viewing screenshots and issuing mouse and keyboard actions.
ai-browsers= An agent that navigates and acts in a web browser to complete tasks for the user.
agent-sandboxes= An isolated execution environment (container, VM, restricted process) where an agent can run code and commands without risking the host.
Worktree (git worktree)= A git feature giving a repository several working directories on different branches; coding-agent tools use it to let parallel agents edit without colliding.
Human-in-the-Loop (HITL)= Design in which a person approves, corrects or steers the agent at defined points.
Vibe Coding= Building software by describing intent to an AI and accepting its code with little or no review; the term is attributed to Andrej Karpathy (a post on X, dated 2025-02-02 per secondary reports; not rechecked).
Spec-Driven Development= Writing a detailed specification first and having AI agents implement, test and revise code against it.
Agentic Coding= Using AI agents that read the codebase, edit files, run tests and iterate to complete software tasks.
Loop Engineering= Designing the repeated act-observe-verify cycle that an agent runs; see the note.
agent-evals-and-observability= Tracing an agent’s model calls, tool calls and decisions so failures can be debugged and measured.
Agent Gateway= A proxy layer that mediates agent-to-tool and agent-to-model traffic with routing, auth and policy; see mcp-security-with-gateway.
Agent Payments Protocols= Standards that let agents pay or purchase on a user’s behalf with authorisation controls.
Agent SDK= A library for building agents with the harness features (loop, tools, context handling) already provided.
Context Graph= A graph of entities, relations and decision history used to give agents structured context.
Software Factory= A setup where agents carry out most steps of software production under human specification and review.
Compound Engineering= A workflow in which each completed task feeds lessons back into the agent’s instructions and tooling so later tasks improve.
WebMCP= A proposal for web pages to expose tools to in-browser agents via MCP-style interfaces.

4. Training, inference and efficiency

Fine-tuning= Further training of a pre-trained model on narrower data to change its behaviour; see LoRA below for a parameter-efficient method.
LoRA (Low-Rank Adaptation)= A parameter-efficient fine-tuning method that trains small low-rank matrices added to frozen weights.
QLoRA= LoRA applied on top of a quantised, frozen base model so large models can be fine-tuned on modest hardware.
PEFT= Parameter-efficient fine-tuning: methods (LoRA, adapters, prompt tuning) that train few parameters instead of the whole model.
Distillation= Training a smaller “student” model to imitate a larger “teacher” model’s outputs.
Quantization= Storing weights (and sometimes activations) at lower numeric precision to cut memory and speed up inference, with some accuracy cost.
Pruning= Removing weights, neurons or layers that contribute little, to shrink a model.
Speculative Decoding= A small draft model proposes several tokens that the large model verifies in one pass, speeding generation without changing the output distribution.
Continuous Batching= A serving technique that adds and removes requests from the running batch at each step instead of waiting for a whole batch to finish; see vllm.
inference-and-serving= Running a trained model to produce outputs, as opposed to training it.
inference-and-serving= The two inference phases: prefill processes the whole prompt in parallel; decode generates output tokens one at a time.
inference-and-serving= A KV-cache memory-management method that stores the cache in non-contiguous blocks to reduce waste (associated with vLLM).
Batch Size= The number of examples (or requests) processed together in one step.
Throughput vs Latency= Throughput is total work per unit time; latency is the time one request takes. Serving trades one against the other.
Scaling Laws= Empirical relationships showing how model loss falls predictably with more parameters, data and compute.
RLHF= Reinforcement learning from human feedback: a reward model learns human preferences and the language model is optimised against it.
RLVR= Reinforcement learning with verifiable rewards: the reward comes from an automatic check (e.g. tests pass, answer matches), used to train reasoning.
DPO (Direct Preference Optimization)= Optimising a model directly on preference pairs with a classification-style loss, with no separate reward model or RL loop.
GRPO= Group Relative Policy Optimization: an RL method that scores several sampled answers to a prompt against each other, avoiding a separate value model.
Reward Model= A model trained to score outputs by human preference, used as the optimisation target in RLHF.
Synthetic Data= Training or evaluation data generated by models or programs rather than collected from people.
Model Collapse= Degradation when models are trained repeatedly on model-generated data, losing the diversity of the original distribution.
Data Contamination= Test or benchmark material leaking into training data; see Benchmark Contamination below.
Checkpoint= A saved snapshot of a model’s weights during or after training.
Epoch and Step= An epoch is a pass over the data set; a step is one optimiser update.
Mixed Precision= Training with a blend of 16-bit and 32-bit numbers to save memory and time.
Parallelism (data, tensor, pipeline)= Ways to split training or serving across many accelerators: by data batches, by splitting matrices, or by splitting layers.
Compute-Optimal Training= Balancing model size and training tokens for a fixed compute budget (the “Chinchilla” finding, 2022); see Scaling Laws.
Edge / On-Device Inference= Running models locally on phones or PCs instead of in a data centre; see ollama.
Model Routing= Choosing, per request, which model answers (a smaller, cheaper one for easy requests, a larger one for hard ones), usually in a gateway or router layer in front of several models.

5. Safety, security and evaluation

Alignment= Making AI systems pursue the goals and values their developers and users intend.
Sycophancy= A model’s tendency to agree with or flatter the user instead of giving accurate answers.
Reward Hacking= An RL-trained model exploiting flaws in the reward signal to score high without doing the intended task.
Constitutional AI= Anthropic’s technique (paper, 2022; not re-fetched) in which a model critiques and revises its own outputs against a written set of principles, using AI feedback to train for harmlessness.
Interpretability= Research into understanding what happens inside a model, from circuits and features to causes of behaviour; “mechanistic interpretability” reverse-engineers the internals.
Guardrails= Checks and filters around a model’s input, output or actions that block unsafe or out-of-policy behaviour.
Red Teaming= Adversarially probing a system to find failures and vulnerabilities before attackers or users do.
Jailbreak= A prompt or technique that gets a model to bypass its safety training or rules.
Prompt Injection= Hidden or untrusted text that hijacks an agent’s instructions; “indirect” injection arrives via web pages, files or tool outputs.
Data Exfiltration= An attacker getting private data out of a system, a typical goal of prompt injection against agents.
Excessive Agency= A security risk where an agent holds more permissions or autonomy than a task needs.
Evals= Systematic tests that measure how well a model or application performs a task.
Benchmark= A standard test set and scoring method used to compare models.
Benchmark Contamination= Benchmark questions or answers appearing in training data, inflating scores.
LLM-as-Judge= Using a language model to grade another model’s outputs against criteria.
Eval Harness= Software that runs a model through benchmarks or tests in a repeatable way and records scores.
MMLU= Massive Multitask Language Understanding: a multiple-choice benchmark across 57 subjects (Hendrycks et al., 2020).
SWE-bench= A benchmark that asks a model to resolve real GitHub issues in open-source repositories, checked by the repository’s tests.
GPQA= Graduate-level Google-Proof Q&A: hard multiple-choice science questions written by domain experts (2023).
HLE (Humanity’s Last Exam)= A very hard, expert-written, closed-ended question set built to stay challenging as models improve (2025).
ARC-AGI= A benchmark of novel visual grid puzzles (Chollet’s Abstraction and Reasoning Corpus) meant to test skill acquisition on unseen tasks.
Model Card= A document describing a model’s intended use, training data, performance and limits (Mitchell et al., 2019).
System Card= A document covering a deployed AI system, including safety evaluations, mitigations and risks, beyond the model alone.
AI Safety= The field aimed at preventing harm from AI systems, from misuse to loss of control.
Watermarking= Embedding a hidden statistical signal in generated text, images or audio so it can later be detected as AI-made.
Content Provenance (C2PA)= A standard for attaching signed metadata to media recording how it was made or edited.
Deepfake= Synthetic or manipulated audio, image or video that realistically depicts someone saying or doing something they did not.
EU AI Act= The European Union’s risk-based AI regulation (Regulation 2024/1689), in force from August 2024 (per Wikipedia, secondary) with obligations phased in over following years; see EU AI Act for amended dates.
GPAI (General-Purpose AI)= The EU AI Act’s term for models with broad capability across tasks; their providers have transparency and copyright duties, with extra duties for models posing systemic risk.
AI Ethics= The study of fairness, accountability, transparency and societal impact of AI.
Data Privacy (AI)= Protecting personal data used in training and prompts.
Differential Privacy= A mathematical guarantee that a result barely depends on any single individual’s data.
Human Oversight= Keeping people able to monitor, intervene in and override an AI system.
Model Welfare= Research on whether AI systems could have experiences or interests that matter morally, and what low-cost precautions follow if so; a contested, early-stage field.

6. Models, products and infrastructure vocabulary

Foundation Model= A large model trained on broad data that can be adapted to many downstream tasks (Stanford CRFM, 2021).
Frontier Model= One of the most capable general-purpose models at a given time, at the edge of current capability.
Open-Weight Model= A model whose trained weights are downloadable, usually under a licence that may limit use; training data and code are often not released.
Open Source AI= AI released with source, weights and enough information to study, use, modify and share; the Open Source Initiative has a formal definition; many models called “open” release weights only.
language-models= A compact language model designed to run cheaply or on-device.
Base Model vs Instruct Model= A base model only predicts text; an instruct (chat) model has been post-trained to follow instructions.
Model Family and Series= Related models released under a shared name and generation; see language-models.
Distilled Model= A smaller model produced by distillation.
Sovereign AI= National or regional control over AI compute, data, models and talent, so a country does not depend on foreign providers.
Neocloud= A specialised cloud provider built around GPU capacity for AI workloads, as opposed to the general hyperscalers.
Hyperscaler= A very large cloud provider such as AWS, Azure or Google Cloud.
inference-and-serving= A company that hosts models and sells API access to run them, often including open-weight models.
Router (LLM)= A service that fronts several model providers behind one API with routing, fallback, caching and cost tracking; see PortKey and claude-code-router.
API Platform= A vendor service exposing models through an API with keys, quotas and tooling.
Ollama= A tool for running open-weight models locally.
vLLM= An open-source high-throughput LLM inference and serving engine.
Aider= Command-line AI pair-programming tool.
autogen= Microsoft’s multi-agent framework (described in the vault audit KB-AI-AUDIT-2026-09 as maintenance-only, converging into microsoft-agent-framework; status not rechecked).
LangGraph= A framework for building stateful agent workflows as graphs.
CrewAI= A framework for role-based multi-agent teams.
AI Slop= Low-quality, mass-produced AI-generated content published with little care or review.
Copilot= A product pattern where an AI assists a user inside an application, as opposed to acting alone.
Rate Limit= A cap on requests or tokens per time period imposed by an API provider.
Speech Recognition (ASR)= Converting speech audio into text; see whisper-and-asr-models.
Text-to-Speech (TTS)= Generating spoken audio from text; see also speech-synthesis.
Voice Activity Detection (VAD)= Detecting whether an audio segment contains speech, used to decide when a speaker starts and stops.
Voice Cloning= Synthesising speech that imitates a specific person’s voice; see voice-cloning.
GPU= A graphics processing unit; the parallel processor that dominates AI training and inference.
TPU= Tensor Processing Unit, Google’s custom AI accelerator.
NPU= Neural processing unit: an on-chip accelerator for AI inference in phones and PCs.
HBM= High Bandwidth Memory: stacked memory placed next to AI accelerators; its capacity and bandwidth often limit inference speed.
Data Center= A facility housing servers, networking and power for large-scale computing.
AI Accelerator= Any chip designed to speed up machine-learning workloads (GPU, TPU, NPU and others).
Hugging Face= A hub for sharing models, datasets and libraries, widely used to distribute open-weight models.
AI Research Lab= An organisation, commercial or academic, doing frontier AI research; see the note.

7. Abbreviations (A-Z)

AGI= Artificial General Intelligence: a hypothetical AI matching human-level ability across most cognitive tasks; definitions vary and are disputed.
AI= Artificial Intelligence.
ANN= Artificial Neural Network; also Approximate Nearest Neighbour search, see ann-index-algorithms.
API= Application Programming Interface: a defined way for programs to call a service.
ASI= Artificial Superintelligence: a hypothetical AI far surpassing human ability in virtually every domain.
ASR= Automatic Speech Recognition (speech-to-text).
A2A= Agent2Agent protocol, see A2A (Agent2Agent protocol).
BERT= Bidirectional Encoder Representations from Transformers, an encoder-only model (2018).
BPE= Byte-Pair Encoding, see Tokenizer.
CNN= Convolutional Neural Network.
CoT= Chain of Thought, see chain-of-thought-prompting.
CPU= Central Processing Unit.
DL= Deep Learning.
DM= Data Mining.
DPO= Direct Preference Optimization, see RLHF.
FLOPs= Floating-point operations: a count of arithmetic work, used to measure training compute (FLOPS with a capital S means operations per second).
GAN= Generative Adversarial Network.
GPAI= General-Purpose AI (EU AI Act term).
GPQA= Graduate-Level Google-Proof Q&A, a benchmark.
GPT= Generative Pretrained Transformer, see Generative Pretrained Transformer.
GPU= Graphics Processing Unit.
GRPO= Group Relative Policy Optimization, see RLHF.
HBM= High Bandwidth Memory.
HITL= Human-in-the-Loop.
HLE= Humanity’s Last Exam, a benchmark.
IP= Internet Protocol, or Intellectual Property depending on context.
KV= Key-Value (as in KV cache), see inference-and-serving.
LLM= Large Language Model, see large-language-model.
LMM= Large Multimodal Model, see Multimodal Model.
LoRA= Low-Rank Adaptation.
LVM= Large Vision Model.
MCP= Model Context Protocol, see model-context-protocol.
MFU= Model FLOPs Utilization: the share of an accelerator’s peak compute that training or inference actually achieves.
ML= Machine Learning, see machine-learning.
MMLU= Massive Multitask Language Understanding, see MMLU.
MoE= Mixture of Experts.
NLP= Natural Language Processing.
NN= Neural Network, see neural-networks.
NPU= Neural Processing Unit.
PEFT= Parameter-Efficient Fine-Tuning.
QLoRA= Quantized LoRA.
RAG= Retrieval-Augmented Generation, see retrieval-augmented-generation.
RL= Reinforcement Learning.
RLHF= Reinforcement Learning from Human Feedback, see RLHF.
RLVR= Reinforcement Learning with Verifiable Rewards, see RLHF.
RNN= Recurrent Neural Network.
SFT= Supervised Fine-Tuning: fine-tuning on labelled input-output examples.
SLM= Small Language Model.
SSM= State-Space Model.
STT= Speech-To-Text.
SWE-bench= Software Engineering benchmark, see SWE-bench.
TPOT= Time Per Output Token: the average time between generated tokens, a serving-latency metric.
TPU= Tensor Processing Unit.
TTFT= Time To First Token: delay from sending a request to receiving the first output token.
TTS= Text-To-Speech.
VAD= Voice Activity Detection.
VLM= Vision-Language Model.
XAI= Explainable AI.

8. Further reading (Wikipedia, reference pages)

Sources (accessed 2026-10-05, earlier pass 2026-10-02)

https://en.wikipedia.org/wiki/Glossary_of_artificial_intelligence (fetched 2026-10-05; only partly covers the 8 newly defined classical terms)
https://en.wikipedia.org/wiki/Regulation_(EU)_2024/1689 (fetched 2026-10-05: GPAI obligations, 12 months after entry into force)
https://en.wikipedia.org/wiki/Large_language_model, /Fuzzy_logic, /Federated_learning (2026-10-02)
Vault notes linked above (concept notes in KB-AI/ai-concepts, ai-prompt-engineering/context-engineering, ai-memory, ai-agentic-systems, ai-infrastructure, ai-coding) were used for terminology consistency.
Stable definitions rely on standard textbook definitions and primary papers (Vaswani et al. 2017 Transformer; Gu and Dao 2023 Mamba; Hendrycks et al. 2020 MMLU; Yao et al. 2022 ReAct; Mitchell et al. 2019 model cards; Anthropic 2022 Constitutional AI); these papers were not re-fetched. Definitions of fast-moving agent terms (subagent, harness, skill, worktree usage, A2A, AGENTS.md) are working definitions from the vault notes, not re-verified against primary docs.
Audit note: removed the abbreviations “SAM = Small Action Model” and “LCM = Large Context Model” (no source found). AutoGen status: KB-AI-AUDIT-2026-09.md Appendix A.

ThirdBrAIn.tech

Explorer

AI Glossary Terms

1. Classical ML and AI terms

2. Modern LLM terms

3. Agents and agentic engineering

4. Training, inference and efficiency

5. Safety, security and evaluation

6. Models, products and infrastructure vocabulary

7. Abbreviations (A-Z)

8. Further reading (Wikipedia, reference pages)

Sources (accessed 2026-10-05, earlier pass 2026-10-02)

Filter Videos

Tags

Channels

Favorites

Table of Contents

Recent Updates

Video topics

Docker Agent

Helix Foundry

Claude Haiku 5.5 pricing and prompt caching

Claude Haiku 5.5

Ayrshare

Blotato

Buffer API

bundle.social

Outstand

Backlinks