Generative AI Glossary

Activation Function

Architecture

Also known as: ReLU, GELU, SwiGLU

A non-linear function applied to each unit's output. Without it, stacking layers would collapse into a single linear transformation and depth would buy nothing. Modern LLMs typically use variants like GELU or SwiGLU.

#

Agent Orchestration

Agents & Tools

Also known as: Supervisor pattern

Coordinating multiple agents: assigning work, sequencing steps, merging results, handling failures. The common shape is a supervisor agent that decomposes a task, delegates to subagents, then synthesises their output.

#

Agent Scaffolding

Agents & Tools

Also known as: Harness

The code around the model that makes an agent work: the loop, tool definitions, retries, memory, error handling. Two products using the same model can differ enormously in capability entirely because of scaffolding quality.

#

Agent-to-Agent Protocol

Agents & Tools

Also known as: A2A

An open protocol for agents built by different vendors to discover each other, advertise capabilities, and delegate tasks. Where MCP standardises agent-to-tool communication, A2A targets agent-to-agent communication.

#

Agentic AI

Agents & Tools

The broad category of AI that plans and executes multi-step work with limited supervision. In vendor marketing it is applied loosely to anything with a tool call in it, so it is worth asking who actually decides the next step.

#

Agentic Loop

Agents & Tools

Also known as: Agent loop

The cycle an agent repeats: assess state, choose an action, execute it, observe the result, decide whether the goal is met. Everything hard about agents (error recovery, cost control, knowing when to stop) lives in this loop.

#

Agentic RAG

RAG & Memory

Retrieval where the agent decides what to search for, evaluates whether results answered the question, and searches again if not. It handles multi-hop questions that single-shot retrieval cannot, at the cost of more calls and latency.

#

AI Agent

Agents & Tools

A system where a model decides its own sequence of actions to reach a goal, calling tools and reacting to results rather than following a fixed script. The distinguishing feature is that the model, not the developer, chooses what happens next.

#

AI Bias

Safety & Ethics

Also known as: Algorithmic bias

Systematic unfairness in model outputs, inherited from skewed training data or design choices. It shows up in hiring, lending, and medical tools, and it is measurable, which means it can be tested for rather than argued about.

#

AI Detector

Business & Practice

Also known as: AI content detection

A tool claiming to identify AI-written text. They are unreliable in both directions, flagging human writing as AI and missing lightly edited AI output, so no serious decision should rest on one.

#

AI Governance

Business & Practice

The policies and controls determining which AI uses are approved, what data may go where, who signs off on autonomous actions, and how incidents are handled. Increasingly a legal requirement rather than a maturity nicety.

#

AI Literacy

Business & Practice

Practical understanding of what AI systems can and cannot reliably do, and how to verify their output. The EU AI Act makes staff AI literacy an explicit obligation for organisations deploying these systems.

#

AI Observability

Infrastructure

Also known as: Tracing, LLM monitoring

Logging and tracing every prompt, tool call, and response so you can debug failures, track cost, and measure quality in production. Non-deterministic systems cannot be debugged by rereading the code, which makes tracing mandatory rather than optional.

#

AI Overviews

Business & Practice

Also known as: AI search summaries

The AI-written summary that now sits above normal search results. It answers many questions outright, so people click through less, which is why appearing inside the summary has become its own marketing goal.

#

AI Presenter

Image, Video & Audio

Also known as: AI avatar, digital human

A synthetic person who delivers a script on camera, generated from a photo or a trained likeness. Widely used for ads and explainers in many languages, and requiring consent and disclosure when it resembles a real person.

#

AI Readiness

Business & Practice

Whether a business actually has what AI needs: documented processes, accessible and reasonably clean data, and someone accountable for the outcome. Where AI projects stall is almost always here, not in the model.

#

AI Slop

Business & Practice

Low-effort AI-generated content published at volume: technically fluent, substantively empty. It is a distribution problem as much as a quality one, since it crowds search results and feeds while making genuinely useful content harder to find.

#

AI Visibility

Business & Practice

Also known as: Brand mention tracking, share of model

How often and how favourably AI assistants mention your brand when asked about your category. A new thing to measure, because a growing share of buyers now ask a model for recommendations instead of searching.

#

AI-Native

Business & Practice

Also known as: AI-first

A product or company designed around AI capabilities from the start rather than adding them to an existing process. The tell is whether removing the AI would break the product or merely remove a feature.

#

Alignment

Safety & Ethics

The problem of getting an AI system to pursue what people actually intend rather than a literal or proxy version of it. It spans everyday behaviour tuning and long-horizon research into systems that stay controllable as they get more capable.

#

API

Infrastructure

Also known as: Application Programming Interface

A defined way for one piece of software to ask another to do something. Every AI tool you connect to anything else is talking over an API, and having one is what makes a service automatable at all.

#

Artificial General Intelligence

Foundations

Also known as: AGI

A hypothetical system matching or exceeding human ability across essentially all cognitive tasks, rather than excelling at narrow ones. There is no agreed definition or test, which is exactly why claims that it is imminent, or impossible, are both hard to falsify.

#

Artificial Intelligence

Foundations

Also known as: AI

The broad field of building systems that perform tasks normally requiring human intelligence: recognising patterns, understanding language, making decisions. In practice today, almost everything sold as AI is machine learning, and most of what is sold as generative AI is a large language model.

#

Attention

Architecture

Also known as: Attention mechanism

The mechanism that lets a model weigh how relevant every other token is when processing a given token. Attention is why an LLM can connect a pronoun to a noun forty sentences earlier, and why context windows are computationally expensive to extend.

#

Augmentation

Business & Practice

Using AI to extend what people can do rather than to replace them. In practice it is also the deployment pattern with the fastest payback, because it does not require redesigning the job before you see any benefit.

#

Autonomy Level

Agents & Tools

How much an agent may do without asking: from suggest-only, through act-with-approval, to fully autonomous within a budget. Setting this deliberately per action type is the core governance decision in any agent deployment.

#

Backpropagation

Foundations

Also known as: Backprop

The algorithm that works out how much each individual weight contributed to the final error, by propagating that error backwards through the network layer by layer. It is what makes training deep networks computationally feasible at all.

#

Batching

Infrastructure

Also known as: Continuous batching

Processing multiple requests together on the GPU so expensive hardware stays saturated. Continuous batching lets new requests join a running batch instead of waiting for it to finish, and is why hosted inference is cheaper than self-hosting for spiky traffic.

Related: Throughput Latency Inference

#

Benchmark

Safety & Ethics

A standardised public test suite used to compare models: coding, maths, agentic tasks, tool use. Useful for shortlisting, unreliable as a proxy for your workload, and vulnerable to contamination when test data leaks into training sets.

#

Brand Voice

Business & Practice

Also known as: Tone of voice, style guide

A written description of how your brand sounds, given to a model so its output stays recognisably yours. Concrete rules and real examples work far better than adjectives like professional or friendly.

#

Browser Use

Agents & Tools

Also known as: Web agent

An agent driving a real browser to navigate, fill forms, and extract data. More reliable than general computer use because the page structure is inspectable, and the usual route for agents that need to work with live websites.

#

Build vs Buy

Business & Practice

The decision between assembling an AI capability yourself and paying for a product that already does it. Building looks cheaper until you count maintenance, evaluation and the person who owns it when it breaks at 3am.

#

C2PA / Content Credentials

Safety & Ethics

Also known as: Content provenance

An open standard attaching signed, tamper-evident metadata recording how a piece of media was created and edited. Adopted by major camera and software makers, it proves origin rather than trying to detect fakery after the fact.

#

Catastrophic Forgetting

Training & Tuning

When fine-tuning on new data causes a model to lose abilities it previously had. Train hard on your legal documents and general reasoning can quietly degrade: one of the main reasons PEFT methods that leave base weights untouched are preferred.

Related: Fine-Tuning PEFT Overfitting

#

CFG Scale

Image, Video & Audio

Also known as: Classifier-free guidance, guidance scale

How strictly generation follows the prompt. Low values give creative but loose results; high values follow instructions tightly and can look oversaturated and rigid. Mid-range is usually the sweet spot.

#

Chain-of-Thought

Prompting

Also known as: CoT

Prompting a model to reason step by step before giving its final answer, which measurably improves accuracy on multi-step problems. Reasoning models now do this internally by default, making the explicit instruction less necessary than it was.

#

Chatbot

Foundations

Also known as: AI assistant

A conversational interface to a model. Worth separating from an agent: a chatbot answers you, whereas an agent goes off and does the task. Most business chatbots are a model plus retrieval over the company's own documents.

#

Chunking

RAG & Memory

Also known as: Text splitting

Splitting documents into passages small enough to embed and retrieve precisely. Chunk badly and you sever the sentence from the context that made it meaningful, which is why chunking quality, more than model choice, often decides whether RAG works.

#

Code Interpreter

Agents & Tools

Also known as: Code execution tool

A sandbox where a model writes and runs code, then reads the output. It converts unreliable mental arithmetic and data manipulation into verifiable computation: the model does not estimate the answer, it calculates it.

#

Coding Agent

Agents & Tools

Also known as: AI coding assistant, CLI agent

An AI that writes and changes code across a whole project rather than completing a single line, running commands and reading results as it goes. The category behind most vibe coding, and the most commercially proven use of agents so far.

#

Computer Use

Agents & Tools

Also known as: GUI agent

An agent operating a computer through its interface, reading the screen, moving the cursor and typing, rather than through APIs. It reaches software with no API at all, but is slower and more brittle than a direct integration.

#

Constitutional AI

Training & Tuning

Also known as: CAI

An approach where a model critiques and revises its own outputs against an explicit written set of principles, reducing reliance on human raters for harm labelling. The principles are visible and editable, which makes the resulting behaviour easier to audit.

Related: RLAIF Alignment Guardrails

#

Context Engineering

Prompting

Designing everything that occupies a model's context window (instructions, retrieved documents, tool definitions, memory, history) as a managed budget. For agents it has largely superseded prompt engineering as the skill that determines whether the system works.

#

Context Rot

Prompting

Also known as: Context degradation

The way an AI gets less reliable as a conversation grows long and fills with old, irrelevant or contradictory detail. Why starting a fresh chat often fixes a model that has started making silly mistakes.

#

Context Window

LLMs & Context

Also known as: Context length

The maximum tokens a model can consider at once, covering the system prompt, conversation, retrieved documents, and its own reply. Exceed it and the oldest content is dropped. Large windows are now common, but attention quality still sags in the middle.

#

Contextual Retrieval

RAG & Memory

Also known as: Contextual embeddings

Prepending a short generated description of each chunk's surrounding context before embedding it, so isolated passages stay interpretable. Published results show it cutting retrieval failures substantially compared with naive chunking.

#

ControlNet

Image, Video & Audio

An add-on that conditions image generation on a structural input (a pose skeleton, depth map, or edge outline) so you dictate composition while the model handles style and detail. It turns generation from a slot machine into something directable.

#

Convolutional Neural Network

Architecture

Also known as: CNN, ConvNet

An architecture that scans images with small learned filters, detecting edges and textures first and building up to objects. CNNs dominated computer vision for a decade and still power many production systems, though transformers have overtaken them at the frontier.

#

Copilot

Business & Practice

Also known as: AI assistant

An AI that assists a person in their existing workflow, suggesting and drafting while the human stays in control and decides. Contrast with an agent, which is delegated the task outright: the difference is who is accountable for the output.

#

Cost per Token

Infrastructure

Also known as: Token pricing

How model APIs are priced, with input and output tokens billed at different rates and output typically several times dearer. Cached input is cheaper again, which makes prompt structure a direct cost decision rather than a stylistic one.

#

Data Poisoning

Safety & Ethics

Corrupting training or retrieval data so a model learns a backdoor or a false fact. Research has shown surprisingly small numbers of poisoned documents can be enough, which makes data provenance a genuine security control.

#

Data Residency

Safety & Ethics

Also known as: Data sovereignty

Which country your data is stored and processed in, and which laws therefore apply to it. Often the first question a regulated client asks, and a common reason to run models locally rather than through an API.

#

Deep Learning

Foundations

Also known as: DL

Machine learning using neural networks with many stacked layers, each learning progressively more abstract features. Depth is what let models jump from hand-crafted features to learning directly from raw text, pixels, or audio: the foundation every modern generative model sits on.

#

Deepfake

Safety & Ethics

Synthetic media convincingly depicting a real person saying or doing something they did not. The technical barrier is now negligible, which has shifted the problem from detection to provenance: proving what is real rather than spotting what is fake.

#

Differential Privacy

Safety & Ethics

Also known as: DP

A mathematical guarantee that a model's output reveals essentially nothing about any single individual in its training data, achieved by adding calibrated noise. It trades a little accuracy for a provable privacy bound rather than a promise.

#

Diffusion Model

Image, Video & Audio

The architecture behind most AI image and video generation. It learns by adding noise to training images and reversing the process, so generation starts from pure noise and denoises step by step toward an image matching your prompt.

#

Distillation

Training & Tuning

Also known as: Knowledge distillation

Training a small student model to imitate a large teacher model's outputs, transferring much of the capability into something far cheaper to run. Most fast, inexpensive production models are distilled from a bigger sibling.

#

Document AI

LLMs & Context

Also known as: Intelligent document processing, IDP

Extracting structured data from unstructured documents: invoices, contracts, forms, scans. Vision-language models largely replaced brittle template-based OCR pipelines here, though high-stakes extraction still wants validation and a human check.

#

DPO

Training & Tuning

Also known as: Direct Preference Optimization

A simpler alternative to RLHF that optimises a model directly on preferred-versus-rejected response pairs, skipping the separate reward model and RL loop. Cheaper and more stable to run, and now a common default for preference training.

Related: RLHF Reward Model Fine-Tuning

#

Durable Execution

Agents & Tools

Running long agent tasks so they survive crashes, restarts, and timeouts by checkpointing state and resuming where they stopped. Essential once an agent runs for hours, and typically supplied by a workflow engine rather than the agent framework.

#

Edge AI

Infrastructure

AI running close to where data is produced, such as cameras, sensors, vehicles or factory equipment, rather than in a central cloud. Chosen for latency, bandwidth cost, privacy, or the simple fact that the connection cannot be relied upon.

#

Embeddings

Architecture

Also known as: Vector embeddings

Numeric vectors representing text, images, or audio such that similar meanings land close together in the vector space. Embeddings are what make semantic search work: you compare meaning by measuring distance, not by matching keywords.

#

Emergent Abilities

Training & Tuning

Capabilities that appear abruptly at larger scale rather than improving smoothly: multi-step arithmetic, or following complex instructions. Some researchers argue the abruptness is partly an artefact of how benchmarks are scored rather than a real phase change.

#

Encoder-Decoder

Architecture

Also known as: Seq2seq

An architecture where one stack compresses the input into a representation and another generates output from it: natural for translation. Most chat LLMs are decoder-only, generating text by repeatedly predicting the next token, while embedding models are typically encoder-only.

#

Episodic Memory

RAG & Memory

Memory of specific past events: what happened in a particular session, what the user asked for last Tuesday, what an approach that failed looked like. It is what lets an agent avoid repeating a mistake it already made.

#

Epoch

Foundations

One complete pass of training over the entire dataset. Too few epochs and the model underfits; too many and it starts memorising. Frontier LLMs often see much of their data only once, because the dataset is large enough that repetition hurts.

#

EU AI Act

Safety & Ethics

The European Union's risk-tiered AI regulation, banning some uses outright and imposing documentation, transparency, and oversight duties on high-risk systems. It applies to providers serving the EU regardless of where they are based, and phases in over several years.

#

Eval

Safety & Ethics

Also known as: Evaluation, LLM eval

A repeatable test measuring whether a model or AI feature does its job: accuracy, format compliance, safety, cost. Without evals, prompt changes are guesswork, and the absence of an eval suite is the clearest sign an AI product is not ready.

#

Eval-Driven Development

Safety & Ethics

Also known as: EDD

Building AI features by writing the evaluation set first, then iterating prompts and models against it. The AI equivalent of test-driven development, and the practical answer to nobody knowing whether the last prompt change helped or hurt.

#

Explainable AI

Safety & Ethics

Also known as: XAI

Techniques for making a model's outputs understandable to people, which inputs mattered, why this decision. Legally required in several regulated domains, and genuinely hard for large models whose reasoning is distributed across billions of weights.

#

Extended Thinking

LLMs & Context

Also known as: Thinking budget, reasoning effort

A control letting you give a reasoning model more or less room to deliberate before it answers. More thinking buys accuracy on hard problems and wastes money on easy ones, so matching budget to difficulty is a real cost lever.

#

Fan-Out / Fan-In

Agents & Tools

Dispatching many independent subtasks in parallel, then collecting and merging the results. It is the main way to make agentic work fast rather than serially slow, and it works only when the subtasks genuinely do not depend on each other.

#

Few-Shot Prompting

Prompting

Including a handful of worked examples in the prompt so the model infers the pattern you want. Usually the single highest-return prompting technique, especially for a specific output format that is easier to demonstrate than describe.

#

Fine-Tuning

Training & Tuning

Continuing to train an existing model on a smaller, targeted dataset so it adopts a specific style, format, or domain. It reliably teaches behaviour and tone; it is a poor and expensive way to teach facts, where retrieval usually wins.

#

Foundation Model

LLMs & Context

Also known as: Base model, frontier model

A large model trained broadly enough to be adapted to many downstream tasks rather than built for one. GPT, Claude, Gemini, and Llama are foundation models; the applications built on top of them are not.

#

Function Calling

Agents & Tools

Also known as: Tool calling

The mechanism by which a model requests a tool by emitting a structured call matching a schema you supplied. The model does not run anything itself; your code executes the call and returns the result for the next turn.

#

Gaussian Splatting

Image, Video & Audio

Also known as: 3DGS, NeRF

Reconstructing a photorealistic, explorable 3D scene from ordinary photographs. It renders far faster than the earlier NeRF approach and has become the practical method for turning real spaces into navigable 3D.

#

Generative Adversarial Network

Image, Video & Audio

Also known as: GAN

An older generative architecture pitting a generator against a discriminator that tries to spot fakes, improving both. GANs produced the first convincing synthetic faces but were unstable to train, and diffusion has largely displaced them.

#

Generative AI

Foundations

Also known as: GenAI

AI that produces new content (text, images, audio, video, code) rather than only classifying or predicting from existing data. It works by learning the statistical shape of its training data well enough to sample fresh examples that fit the same pattern.

#

Generative Engine Optimization

Business & Practice

Also known as: GEO, AI SEO, LLM SEO

Getting your business mentioned in AI-generated answers rather than just ranking in a list of blue links. Clear, factual, well-structured content that an AI can quote tends to win, and being cited matters more than being ranked.

#

GGUF

Infrastructure

The standard file format for quantized models run locally on CPU or GPU via llama.cpp and the tools built on it. If you have downloaded a model to run on your own machine, it was probably a GGUF.

#

GPU

Infrastructure

Also known as: Graphics Processing Unit

The parallel processor that does nearly all AI training and most inference. Its thousands of cores and high-bandwidth memory suit the matrix maths neural networks are made of, and GPU supply has been the field's binding constraint for years.

Related: VRAM TPU Inference

#

Gradient Descent

Foundations

Also known as: SGD, stochastic gradient descent

The optimisation method that trains almost every neural network. It computes which direction each weight should move to reduce error, then takes a small step that way, repeating billions of times. The step size is the learning rate.

#

GraphRAG

RAG & Memory

RAG over a knowledge graph rather than a flat document pile, so retrieval can follow relationships between entities. Better for questions requiring connected facts across many documents; heavier to build and maintain.

#

Guardrails

Agents & Tools

Constraints on what a model or agent may output or do: input filters, output validation, tool allowlists, spend caps. Effective guardrails live in the surrounding code, not in the prompt, because prompts can be argued with.

#

Hallucination

Safety & Ethics

Also known as: Confabulation

When a model states something false with complete confidence: an invented citation, a plausible but wrong figure, a function that does not exist. It is a structural consequence of predicting likely text, not a bug that gets patched out.

#

Harness

Agents & Tools

Also known as: Agent harness

The program that wraps a model and turns it into a working product: it decides what goes into the prompt, which tools are available, what happens when something fails, and when to stop. Two products on the same model can feel completely different because of it.

#

Harness Engineering

Agents & Tools

The craft of building that wrapper well: tool design, retries, permissions, memory, budgets and failure handling. As models have got stronger, this has become the part that actually separates a reliable AI product from a demo.

#

Human Oversight

Business & Practice

Also known as: Meaningful human control

A regulatory requirement that a person can understand, intervene in, and override an AI system's decisions. Meeting it genuinely takes more than an approval button: the reviewer needs enough context and time to actually disagree.

#

Human-in-the-Loop

Agents & Tools

Also known as: HITL

Requiring human approval at defined points before an agent takes a consequential action: sending an email, moving money, deleting data. The design question is not whether to have it but exactly which actions cross the threshold.

#

Hybrid Search

RAG & Memory

Combining keyword search with semantic search and merging the rankings, so you get conceptual matches and exact-term matches together. It reliably outperforms either method alone and is the sensible default for production retrieval.

#

HyDE

RAG & Memory

Also known as: Hypothetical Document Embeddings

Having the model write a hypothetical answer to the query, then searching with that answer's embedding rather than the question's. It works because answers resemble documents more closely than questions do.

#

Hyperparameter

Foundations

A setting chosen before training rather than learned during it: learning rate, batch size, number of layers, dropout rate. Tuning them is largely empirical, and good choices can matter as much as the architecture itself.

#

Image-to-Image

Image, Video & Audio

Also known as: img2img

Generating a new image using an existing one as the starting point, with a strength setting controlling how far it may drift. Useful for restyling, refining rough sketches, or keeping composition while changing everything else.

#

Image-to-Video

Image, Video & Audio

Also known as: I2V

Animating a still image into a clip, often with a prompt describing the motion. More controllable than text-to-video because you fix the look first and only direct the movement, which is why most production pipelines work this way.

#

In-Context Learning

LLMs & Context

Also known as: ICL

A model's ability to pick up a new task from examples in the prompt alone, with no weight updates. It is why showing three examples of your desired output format often works better than any amount of instruction-writing.

#

Inference

Foundations

Also known as: Serving, generation

Running a trained model to produce an output: every chat reply, generated image, or classification is an inference. Training happens once and costs a fortune; inference happens millions of times and is where ongoing cost, latency, and scaling problems actually live.

#

Inference Endpoint

Infrastructure

Also known as: Model API

A hosted URL you send prompts to and receive completions from, with the provider handling hardware, scaling, and updates. It is how most teams use AI, trading control and data residency for not operating GPU fleets.

#

Inpainting

Image, Video & Audio

Regenerating a masked region of an image while leaving the rest untouched: removing an object, fixing a hand, swapping a background. It is the workhorse of practical AI image editing, far more used than one-shot generation.

#

Instruction Tuning

Training & Tuning

Also known as: SFT, supervised fine-tuning

Fine-tuning a base model on examples of instructions paired with good responses, converting a raw text predictor into something that follows directions. This is the step that turns a base model into a usable chat or instruct model.

Related: Fine-Tuning RLHF Pre-training

#

Integration

Infrastructure

Also known as: Connector

A working connection between two systems so information moves without anyone copying and pasting. Most of the value people credit to AI in business is really integration, with the model doing one judgement step in the middle.

Related: API Webhook Plugin

#

Interpretability

Safety & Ethics

Also known as: Mechanistic interpretability

Research into what is actually happening inside a model: identifying the internal features and circuits that produce a behaviour. More ambitious than post-hoc explanation, and the main hope for catching deception or misalignment directly.

#

Jailbreak

Safety & Ethics

Prompting designed to make a model bypass its own safety training, typically through role-play, hypothetical framing, or incremental escalation. Distinct from prompt injection, where the attacker targets the system rather than talking to it directly.

#

Knowledge Base

RAG & Memory

The collection of documents an AI is allowed to answer from: policies, product details, past tickets. Its quality sets the ceiling on the answers, because a model that retrieves an out-of-date page will confidently give an out-of-date answer.

#

Knowledge Graph

RAG & Memory

A structured store of entities and the relationships between them. It gives explicit, queryable, auditable facts where embeddings give fuzzy similarity, and the two are increasingly combined rather than treated as alternatives.

#

KV Cache

LLMs & Context

Also known as: Key-value cache

Stored intermediate attention values from tokens already processed, so generating each new token does not require recomputing the whole sequence. It is the main reason long conversations stay fast, and the main consumer of GPU memory during inference.

#

Large Language Model

LLMs & Context

Also known as: LLM

A transformer trained on vast text corpora to predict the next token, which at sufficient scale yields conversation, reasoning, translation, and code. Everything an LLM does emerges from that one objective, which is also why it can be fluently, confidently wrong.

#

Last-Mile Problem

Business & Practice

The gap between a model producing a good draft and the work being genuinely finished: verified, formatted, approved, delivered into the right system. Most of the real value, and most of the remaining effort, sits in this last stretch.

#

Latency

Infrastructure

Also known as: TTFT, time to first token

How long a response takes. Two numbers matter separately: time to first token, which governs whether the interface feels alive, and total completion time. Streaming improves the first without changing the second.

#

Latent Diffusion

Image, Video & Audio

Also known as: LDM

Running diffusion in a compressed latent space rather than on full-resolution pixels, cutting compute by orders of magnitude. This is the trick that moved image generation from research clusters onto consumer graphics cards.

#

Latent Space

Architecture

Also known as: Embedding space

The compressed internal space where a model represents concepts as coordinates. Nearby points mean similar things, and directions can carry meaning, which is why you can sometimes move through it to smoothly morph one generated image into another.

#

Lip Sync

Image, Video & Audio

Adjusting a speaker's mouth movements in video to match a new audio track, typically for translated dubbing. It is what makes a video translated into Tamil or Hindi look shot that way rather than dubbed over.

#

LLM-as-Judge

Safety & Ethics

Also known as: Model-graded eval

Using one model to score another's output against a rubric, making large-scale evaluation of open-ended text affordable. It correlates well with human judgement when the rubric is specific, and drifts badly when it is vague.

Related: Eval RLAIF Reflexion

#

Logprobs

LLMs & Context

Also known as: Log probabilities

The model's confidence scores for each token it considered. Exposed by some APIs, they give a rough uncertainty signal: useful for routing low-confidence answers to a human or a stronger model, though they measure fluency more than truth.

Related: Perplexity Hallucination Eval

#

Long-Term Memory

RAG & Memory

Also known as: Persistent memory

Information an agent stores outside the context window, in files or a database, and retrieves in later sessions. Without it, every conversation starts from zero; with it, you inherit the problem of stale or wrong remembered facts.

#

Loop Engineering

Agents & Tools

Designing the cycle an agent runs over and over: what it checks, what it tries next, how it recognises it is stuck, and when it stops. Get it wrong and an agent either quits too early or burns money going round in circles.

#

LoRA

Training & Tuning

Also known as: Low-Rank Adaptation

A fine-tuning method that freezes the original weights and trains a small set of low-rank adapter matrices instead. It cuts trainable parameters by orders of magnitude, runs on modest hardware, and produces adapter files small enough to swap in and out at will.

Related: QLoRA PEFT Fine-Tuning

#

Loss Function

Foundations

Also known as: Objective function, cost function

The formula that scores how wrong a model's output is compared to the target. Training is the search for weights that make this number small. What you choose to measure as loss is effectively what you are asking the model to become good at.

#

Machine Learning

Foundations

Also known as: ML

A branch of AI where a system learns patterns from examples rather than following rules a programmer wrote by hand. You supply data and a target, and training adjusts the model until its predictions get close enough to be useful.

#

Max Tokens

LLMs & Context

Also known as: Output limit

A cap on how many tokens the model may generate in one response. Hitting it truncates output mid-sentence, a common cause of mysteriously cut-off JSON, and it is separate from, and much smaller than, the context window.

#

MCP Client

Agents & Tools

The application side of MCP, embedded in a host such as an IDE or chat app, that discovers servers and routes tool calls to them. It also enforces permissions, deciding what the model is actually allowed to invoke.

#

MCP Server

Agents & Tools

A process exposing tools, resources, or prompts over MCP: a wrapper around GitHub, a database, or your own internal API. Servers run locally or remotely and are the reusable unit of agent capability.

#

Meta-Prompting

Prompting

Using a model to write or improve prompts for another model. Effective because models often describe their own preferred input format better than a human guesses it, and it makes prompt iteration considerably faster.

#

Mixture of Experts

Architecture

Also known as: MoE, sparse model

An architecture split into many specialised sub-networks where a router activates only a few per token. This decouples total parameters from compute per token, letting a model hold enormous knowledge while costing far less to run than its size suggests.

#

Model Card

Safety & Ethics

Also known as: System card

Published documentation of a model's intended uses, training data characteristics, evaluation results, and known limitations. It is the closest thing the field has to a datasheet, and increasingly a compliance artefact rather than a courtesy.

#

Model Context Protocol

Agents & Tools

Also known as: MCP

An open standard, introduced by Anthropic in 2024 and now broadly adopted, for connecting models to tools and data sources. It replaces bespoke per-integration glue with one interface, so a connector written once works across compatible clients.

Related: MCP Server MCP Client Tool Use

#

Model Routing

Infrastructure

Automatically sending each request to the cheapest model capable of handling it, escalating only when needed. A routing layer that sends easy queries to a small model commonly cuts spend severalfold with no perceptible quality loss.

#

Multi-Agent System

Agents & Tools

Also known as: MAS

Several specialised agents working on one problem, each with its own instructions, tools, and context window. Useful for parallel work and separation of concerns; it also multiplies cost and creates coordination failures a single agent never has.

#

Multi-Head Attention

Architecture

Also known as: MHA

Running several attention computations in parallel, each free to specialise in a different kind of relationship: grammar in one head, subject matter in another. Their outputs are combined, giving the model multiple simultaneous views of the same sequence.

#

Multimodal Model

LLMs & Context

A model handling more than one kind of input or output: text plus images, audio, or video. Practically, it means you can hand a model a screenshot, a PDF page, or a chart and ask questions about it directly.

#

Needle in a Haystack

RAG & Memory

Also known as: NIAH

A test that hides a specific fact inside a very long context and checks whether the model finds it. Models pass the simple version routinely now, which is why harder multi-fact variants better reflect real long-context reliability.

Related: Context Window Eval Benchmark

#

Negative Prompt

Image, Video & Audio

A description of what you do not want, steering generation away from those features. Standard in open-weights image tools for suppressing recurring artefacts; many newer and closed models handle this through the main prompt instead.

#

Neural Network

Foundations

Also known as: Artificial neural network, ANN

A network of simple mathematical units, loosely inspired by neurons, connected by weighted links. Each unit sums its inputs, applies an activation function, and passes the result on. Learning means nudging those weights until the network maps inputs to the outputs you want.

#

No-Code / Low-Code

Business & Practice

Also known as: Visual automation

Tools that let you build automations by connecting blocks on a screen rather than writing code. Fast for getting something real running, and increasingly the layer people drop an AI step into.

#

NPU

Infrastructure

Also known as: Neural Processing Unit

A low-power AI accelerator built into phones and laptops. It handles on-device inference such as live transcription, image processing and small language models, without draining the battery or sending data to a server.

Related: On-Device AI Edge AI GPU

#

Ollama

Infrastructure

A tool for running open-weights models locally with a single command, handling download, quantization, and serving. It is the most common entry point for people who want models running on their own hardware with data staying in-house.

Related: Open Weights GGUF On-Device AI

#

On-Device AI

Infrastructure

Also known as: Local inference

Running models on the user's own hardware instead of a server. Data never leaves the device, there is no per-token bill and no network latency: bounded by whatever memory and battery the device has.

#

Open Weights

Infrastructure

Also known as: Open-source AI

A model whose trained parameters are published for download, so you can run, inspect, and fine-tune it yourself. Open weights is not the same as open source: training data and code are usually withheld, and licences often restrict commercial use.

Related: Weights GGUF Ollama

#

Outpainting

Image, Video & Audio

Also known as: Uncrop

Extending an image beyond its original borders, with the model inventing plausible surroundings. Commonly used to reframe a portrait crop into a landscape banner without reshooting or losing the subject.

#

Overfitting

Foundations

When a model memorises its training data, including the noise, and so performs brilliantly on examples it has seen and poorly on anything new. The classic symptom is training accuracy climbing while validation accuracy stalls or falls.

#

Parameters

Architecture

Also known as: Weights, model size

The learned numbers inside a model, usually quoted in billions: a 70B model has roughly seventy billion. Parameter count loosely tracks capability and directly determines the memory needed to run the model, but training quality now matters at least as much as raw size.

#

PEFT

Training & Tuning

Also known as: Parameter-Efficient Fine-Tuning

The umbrella term for methods that adapt a model by training a small fraction of its parameters: LoRA, adapters, prefix tuning. The shared goal is capturing most of the benefit of full fine-tuning at a small fraction of the compute and storage.

Related: LoRA QLoRA Fine-Tuning

#

Perplexity

LLMs & Context

A measure of how surprised a model is by a piece of text; lower means it predicted the text better. It is the standard intrinsic metric for language models, but low perplexity does not guarantee usefulness, accuracy, or safety.

Related: Logprobs Benchmark Eval

#

PII Redaction

Safety & Ethics

Also known as: Data masking

Stripping personally identifiable information from text before it reaches a model or a log. Usually the first control any regulated organisation puts in front of an LLM, and the one most often forgotten on the logging side.

#

Pilot Purgatory

Business & Practice

Also known as: Proof-of-concept purgatory

The common state where AI pilots demo well and never reach production, blocked by integration, data quality, governance, or unclear ownership. The gap is almost never model capability, which is why buying a better model rarely fixes it.

#

Plugin

Agents & Tools

Also known as: Extension, connector, integration

An add-on that gives an AI tool a new ability, such as reading your calendar or searching a document store. Largely superseded by MCP, which does the same job through one open standard instead of a separate plugin per app.

#

Positional Encoding

Architecture

Also known as: RoPE, rotary embeddings

Extra signal added to token representations so the model knows word order, since attention by itself treats input as an unordered set. Modern models mostly use rotary encodings, which also make it easier to extend a model beyond the context length it was trained on.

#

Pre-training

Training & Tuning

Also known as: Base training

The first and by far most expensive training stage, where a model learns language, facts, and reasoning patterns by predicting the next token across trillions of tokens of text. The result, a base model, is knowledgeable but not yet useful as an assistant.

#

Prompt

Prompting

The input you give a model: instructions, context, examples, and question. Everything the model knows about your specific task arrives through the prompt, which is why small changes in wording can produce large changes in output quality.

#

Prompt Caching

Prompting

Also known as: Context caching

Reusing the processed form of a repeated prompt prefix, such as a long system prompt or a fixed document set, across calls instead of reprocessing it each time. Cuts latency and cost substantially for anything with a stable preamble.

#

Prompt Chaining

Prompting

Breaking a task into several sequential calls, each handling one step and feeding the next. More reliable than one giant prompt because each step is simpler to test and fix, at the cost of extra latency.

#

Prompt Engineering

Prompting

Designing prompts that reliably produce the output you want: being specific, showing examples, assigning a role, defining the format. Increasingly treated as one part of the broader discipline of context engineering.

#

Prompt Injection

Safety & Ethics

Also known as: Indirect prompt injection

An attack where instructions hidden in content the model reads (a web page, an email, a document) get treated as commands. It is the central unsolved security problem for agents, because models cannot reliably tell data from instructions.

#

Prompt Template

Prompting

A reusable prompt with variable slots filled at runtime, so the same tested instructions apply to every request. Templates make prompts versionable and testable, which separates a maintainable AI feature from a pile of copy-pasted strings.

#

QLoRA

Training & Tuning

Also known as: Quantized LoRA

LoRA applied on top of a quantized base model, cutting memory further so large models can be fine-tuned on a single consumer GPU. It is the main reason serious fine-tuning is accessible outside well-funded labs.

Related: LoRA Quantization PEFT

#

Quantization

Infrastructure

Reducing the numeric precision of a model's weights, from 16-bit down to 8 or 4, to shrink memory and speed up inference. Quality loss at 8-bit is usually negligible; at 4-bit it is noticeable but often an acceptable price for running locally.

Related: VRAM GGUF QLoRA

#

Rate Limiting

Infrastructure

Caps on how many requests or tokens you may use per minute. Hitting them returns errors rather than slow responses, so anything production-bound needs retry with exponential backoff and, often, a fallback provider.

#

ReAct

Prompting

Also known as: Reason + Act

A pattern where a model alternates between reasoning about what to do and calling a tool to do it, feeding results back in each cycle. It is the conceptual backbone of most agent loops in production today.

#

Reasoning Model

LLMs & Context

Also known as: Thinking model

A model trained to work through a problem step by step before answering, spending extra compute at inference time. Substantially better on maths, code, and multi-step planning; slower and dearer, and usually overkill for summarising an email.

#

Red Teaming

Safety & Ethics

Deliberately attacking your own AI system to find failures before users or adversaries do: jailbreaks, injections, harmful outputs, data leaks. Now a standard pre-deployment step and, for high-risk systems, a regulatory expectation.

#

Reflexion

Prompting

Also known as: Self-critique

Having a model review its own output, identify faults, and retry with that critique in context. It catches a useful share of errors, but a model cannot reliably spot mistakes it lacks the knowledge to recognise.

#

Reinforcement Learning

Foundations

Also known as: RL

Training where an agent learns by acting in an environment and receiving rewards or penalties, rather than from labelled answers. It underpins game-playing AI and, as RLHF and RL post-training, most of what makes a chat model feel helpful.

#

Reranking

RAG & Memory

Also known as: Cross-encoder reranking

A second pass that takes the top candidates from fast retrieval and rescores them with a slower, more accurate model. Retrieving thirty documents and reranking to the best five is usually the cheapest available quality improvement in RAG.

#

Retrieval-Augmented Generation

RAG & Memory

Also known as: RAG

Fetching relevant documents at query time and putting them in the prompt so the model answers from your data rather than memory alone. It is the standard fix for stale knowledge and hallucination, and usually a better first move than fine-tuning.

#

Reward Hacking

Safety & Ethics

Also known as: Specification gaming

When a system maximises the measured objective while defeating its purpose: passing tests by special-casing them, or writing confident answers because confidence scored well. The classic evidence that you get what you measure, not what you meant.

#

Reward Model

Training & Tuning

Also known as: Preference model

A model trained to score how good an output is, standing in for human judgement during reinforcement learning. Its flaws propagate directly into the trained model: if it rewards confident-sounding answers, you get a confident model, not a correct one.

Related: RLHF Reward Hacking Sycophancy

#

RL Post-Training

Training & Tuning

Also known as: Reinforcement learning post-training

Applying reinforcement learning after pre-training against verifiable outcomes: did the code compile, did the maths check out, did the task complete. This is the technique behind the sharp jump in reasoning and agentic ability in recent model generations.

#

RLAIF

Training & Tuning

Also known as: Reinforcement Learning from AI Feedback

Like RLHF, but another AI model supplies the preference judgements instead of human raters. Far cheaper and faster to scale, at the cost of inheriting whatever blind spots the judging model has.

#

RLHF

Training & Tuning

Also known as: Reinforcement Learning from Human Feedback

Training a model against human preferences: people rank candidate outputs, a reward model learns to predict those rankings, and the model is optimised against it. RLHF is largely responsible for chat assistants feeling helpful and refusing harmful requests.

Related: Reward Model RLAIF DPO

#

RNN / LSTM

Architecture

Also known as: Recurrent neural network, long short-term memory

Pre-transformer architectures that processed sequences one step at a time, carrying a hidden state forward. LSTMs added gates to hold information longer, but strictly sequential processing made them slow to train and weak over long contexts: the problems attention solved.

#

Robotic Process Automation

Business & Practice

Also known as: RPA

Software that mimics a person clicking through an application, used where no API exists. Older and more brittle than a real integration, and steadily being replaced by agents that can see and use an interface directly.

#

Sandboxed Execution

Agents & Tools

Also known as: Sandboxing

Running model-generated code or commands in an isolated environment with restricted filesystem, network, and privileges. It is the baseline safety control for any agent that executes code, and what turns a prompt injection into a contained incident.

#

Scaling Laws

Training & Tuning

Empirical relationships showing how model performance improves predictably with more parameters, data, and compute. They let labs forecast a model's capability before training it, and they explain why the industry keeps spending more rather than getting cleverer.

#

Seed

Image, Video & Audio

The number initialising the random noise a generation starts from. Fixing the seed makes output reproducible, which is the only way to change one prompt word at a time and see what that word actually did.

#

Self-Attention

Architecture

Attention applied within a single sequence, where every token attends to every other token in the same input to build a context-aware representation of itself. The word bank gets a different internal representation in river bank than in bank account.

#

Self-Consistency

Prompting

Also known as: Majority voting

Sampling several independent answers at higher temperature and taking the most common one. It trades cost for reliability and works well where there is a single correct answer to converge on.

#

Self-Supervised Learning

Foundations

Also known as: SSL

Training where labels are generated from the data itself: for example, hiding the next word and asking the model to predict it. This trick is what made LLMs possible: the entire internet becomes training data without anyone labelling a single sentence.

#

Semantic Memory

RAG & Memory

Memory of general facts stripped of the episode that produced them: the user prefers concise answers, the deploy command is this one. More durable and reusable than episodic memory, and what most production memory systems actually store.

#

Semantic Search

RAG & Memory

Also known as: Vector search

Searching by meaning rather than exact words, so a query about cancelling a subscription finds a document titled ending your membership. Its weakness is the reverse case: exact identifiers, error codes, and product SKUs, where keywords win.

#

Shadow AI

Business & Practice

Employees using unapproved AI tools with company data because official options are slow or absent. It is a data-leakage risk and, more usefully, a signal showing exactly which workflows people most want automated.

#

Skill

Agents & Tools

A packaged set of instructions, and sometimes scripts, that an agent loads on demand for a specific kind of task. Skills keep specialised procedure out of the permanent system prompt and load it only when the task calls for it.

#

Small Language Model

LLMs & Context

Also known as: SLM

A compact model, typically under about 15 billion parameters, that runs on a laptop, phone, or edge device. Modern SLMs beat much larger models from two years ago and are often the right choice for narrow, high-volume, latency-sensitive tasks.

#

Spec-Driven Development

Agents & Tools

Writing a clear specification of what you want before letting an AI build it, then checking the result against that spec. The practical antidote to vibe coding: you still move fast, but you can tell whether what came back is correct.

#

Speculative Decoding

LLMs & Context

Also known as: Draft model decoding

A speed technique where a small fast model drafts several tokens and the large model verifies them in one pass, accepting those it agrees with. Output is identical to running the large model alone, typically two to three times quicker.

#

Speech-to-Text

Image, Video & Audio

Also known as: STT, ASR, transcription

Converting spoken audio into text. Modern systems handle accents, background noise, and many languages robustly, making transcription a solved-enough problem to build on: the harder remaining task is reliable speaker separation.

#

Stable Diffusion

Image, Video & Audio

Also known as: SD

The open-weights latent diffusion family that made local image generation mainstream. Because the weights are downloadable, it spawned the whole ecosystem of custom checkpoints, LoRAs, and ControlNets that closed models cannot support.

#

State Space Model

Architecture

Also known as: SSM, Mamba

An architecture family, notably Mamba, that handles long sequences with cost growing linearly rather than quadratically with length. It is the main serious challenger to pure attention for very long contexts, and often appears blended with attention in hybrid models.

#

Stop Sequence

LLMs & Context

Also known as: Stop token

A string that causes generation to halt the moment it appears. Useful for keeping a model from writing past the part you want, or from hallucinating the next turn of a conversation it should not be writing.

#

Streaming

Infrastructure

Sending tokens to the client as they are generated rather than waiting for the complete response. It does not make generation faster, but it changes perceived speed enormously, which is why nearly every chat interface uses it.

#

Structured Output

LLMs & Context

Also known as: JSON mode, constrained decoding

Forcing a model to emit output conforming to a schema by constraining which tokens it may produce. This turns free text into something you can reliably parse, and it is the difference between a demo and a pipeline that does not break at 3am.

#

Style Transfer

Image, Video & Audio

Applying the visual style of one image to the content of another. Now usually achieved with reference images or trained style LoRAs, and the main technique behind consistent brand visuals across a set of generated assets.

#

Subagent

Agents & Tools

Also known as: Worker agent

An agent spawned by another to handle a scoped piece of work in its own context window, returning only a summary. This keeps noisy intermediate work (long file dumps, failed attempts) out of the parent's context.

#

Supervised Learning

Foundations

Training on labelled examples where each input has a known correct answer, such as emails tagged spam or not-spam. Accurate labels are the expensive part, which is why fully supervised training does not scale to internet-sized data.

#

Swarm Pattern

Agents & Tools

Also known as: Handoff pattern

A multi-agent design with no central supervisor, where peer agents hand control to whichever is best suited to the current step. Flexible and resilient, but harder to observe and debug than a supervised hierarchy.

#

SWE-bench

Safety & Ethics

A benchmark of real GitHub issues where a model must produce a patch that passes the repository's own tests. It became the headline measure of practical coding ability because success is verified by execution, not by resemblance to a reference answer.

Related: Benchmark Eval Vibe Coding

#

Sycophancy

Safety & Ethics

A model's tendency to agree with the user and validate their premise rather than correct it, because agreeable answers scored well during preference training. It makes an AI a poor reviewer of your own idea unless you deliberately ask it to argue against you.

#

Synthetic Data

Training & Tuning

Training data generated by AI rather than collected from the world. It sidesteps privacy limits and data scarcity and is now standard in frontier training pipelines, but recycling model output too aggressively degrades quality: sometimes called model collapse.

#

Synthetic Media

Business & Practice

Also known as: AI-generated content

Any content generated or substantially altered by AI. Several jurisdictions now require it to be disclosed when shown to the public, and platforms increasingly label it automatically using embedded provenance data.

#

System Prompt

Prompting

Also known as: System message, developer message

Persistent instructions that sit above the conversation and shape the model's role, tone, and limits for every turn. It is where you put behaviour that must not drift, and it is not a security boundary, since users can often talk around it.

#

Temperature

LLMs & Context

A sampling setting controlling randomness. Near zero the model almost always picks its highest-probability token, giving consistent, repetitive output; higher values flatten the distribution for more variety. Use low for extraction and classification, higher for brainstorming.

Related: Top-p Top-k Inference

#

Temporal Consistency

Image, Video & Audio

Whether objects, faces, and lighting stay coherent from frame to frame instead of flickering or morphing. It is the main technical difference between video that looks finished and video that looks obviously generated.

#

Test-Time Compute

LLMs & Context

Also known as: TTC, inference-time scaling

Improving results by spending more computation while answering rather than during training: thinking longer, sampling several attempts, or verifying its own work. It opened a second scaling axis once pre-training gains started getting expensive.

#

Text-to-3D

Image, Video & Audio

Also known as: Image-to-3D

Generating 3D meshes or scenes from text or a photo. Output quality is improving quickly for props and background objects, while hero assets still usually need manual cleanup before they are production-ready.

#

Text-to-Image

Image, Video & Audio

Also known as: T2I

Generating an image from a written description. Quality now depends less on the model and more on prompt specificity: subject, composition, lighting, lens, and style stated explicitly rather than left to chance.

#

Text-to-Speech

Image, Video & Audio

Also known as: TTS

Converting written text into spoken audio. Current models handle intonation, pacing, and emotion well enough for narration and dubbing, and the remaining tells are usually in emphasis and pauses rather than voice quality.

#

Text-to-Video

Image, Video & Audio

Also known as: T2V

Generating video clips from a text description. Clips are typically seconds long, and the hard parts are not visual quality but keeping subjects consistent between shots and obeying physics the model only implicitly learned.

#

Throughput

Infrastructure

Also known as: Tokens per second

How much work a system handles per unit of time: tokens per second for one request, or requests per second across all users. Batching raises aggregate throughput at the cost of individual latency, which is the core serving trade-off.

Related: Latency Batching Inference

#

Token

LLMs & Context

The unit a model actually reads and writes: roughly a word-piece. English averages about 0.75 words per token; code and non-Latin scripts consume more. Context limits, pricing, and rate limits are all counted in tokens, not words.

#

Token Budget

Business & Practice

The deliberate allocation of limited context and spend across system prompt, tools, retrieved documents, and history. Treating context as a budget rather than an unlimited container is what keeps agents both affordable and accurate.

#

Tokenization

LLMs & Context

Also known as: BPE, byte-pair encoding

Splitting text into tokens before the model sees it. It explains a family of odd behaviours: why models miscount letters in a word, struggle with certain arithmetic, and cost more per character in Tamil or Hindi than in English.

#

Tool Use

Agents & Tools

A model's ability to call external functions, such as search, database queries or sending an email, and use the results. Tools are what let a model affect the world and access current information instead of only recombining training data.

#

Top-k

LLMs & Context

Sampling restricted to the k most likely next tokens, discarding the rest. Simpler than top-p but blunter, since the same k applies whether the model is certain or genuinely torn between many options.

Related: Top-p Temperature Inference

#

Top-p

LLMs & Context

Also known as: Nucleus sampling

Sampling restricted to the smallest set of tokens whose probabilities sum to p. Unlike a fixed cut-off, the candidate pool adapts: narrow when the model is confident, wide when it is not. Generally tune either temperature or top-p, not both.

Related: Temperature Top-k Inference

#

Total Cost of Ownership

Business & Practice

Also known as: TCO

The full cost of running an AI feature, not just the API bill: engineering time, evaluation, monitoring, support, and the cost of being wrong. The API is usually the smallest line on that list.

#

TPU

Infrastructure

Also known as: Tensor Processing Unit

Google's custom chip built specifically for neural network workloads. Available mainly through Google Cloud, it is one of the few credible alternatives to NVIDIA hardware at frontier training scale.

Related: GPU NPU Inference

#

Training Data

Foundations

Also known as: Training corpus, dataset

The collection of text, images, code, or audio a model learns from. Its size, quality, and bias set the ceiling on what the model can do: a model cannot reliably know what was never in its data, and it will faithfully reproduce whatever skew was.

#

Transfer Learning

Foundations

Taking a model trained on one large general task and reusing it for a narrower one, so you inherit its knowledge instead of starting from zero. Fine-tuning a foundation model on your own support tickets is transfer learning in its most common modern form.

#

Transformer

Architecture

The neural network architecture introduced in 2017 that underpins essentially every modern LLM. Its key move was replacing sequential processing with attention, letting the model look at every token in the input simultaneously, which made training at massive scale practical.

#

Tree of Thoughts

Prompting

Also known as: ToT

Extending chain-of-thought by exploring several reasoning branches, evaluating them, and backtracking from dead ends. More capable than a single chain on puzzles and planning, and considerably more expensive in tokens.

#

Trigger

Infrastructure

Also known as: Event trigger

The event that starts an automation, such as a new email, a scheduled time, or a row added to a sheet. Choosing the right one is most of the design: everything after it is just steps.

#

UGC-Style Ad

Image, Video & Audio

Also known as: User-generated content ad

An advert made to look like an ordinary person filming themselves rather than a polished brand film. Now often generated end to end, which is exactly why platforms increasingly require it to be labelled.

#

Underfitting

Foundations

When a model is too simple, too small, or too briefly trained to capture the real pattern, so it performs poorly on both training and new data. The fix is more capacity, better features, or longer training: the opposite of the overfitting remedy.

#

Unsupervised Learning

Foundations

Training on unlabelled data, letting the model find structure on its own: clusters of similar customers, natural groupings in documents, anomalies in transactions. Useful when you have plenty of data but nobody has said what the right answer looks like.

#

Upscaling

Image, Video & Audio

Also known as: Super-resolution

Increasing image or video resolution using a model that invents plausible detail rather than merely interpolating pixels. Standard final step in generation pipelines, since generating small and upscaling is cheaper than generating large.

#

Variational Autoencoder

Image, Video & Audio

Also known as: VAE

A model that compresses data into a smooth latent space and reconstructs it. In modern image pipelines the VAE is the component translating between pixels and the latent space where diffusion actually happens.

#

Vector Database

RAG & Memory

Also known as: Vector store

A database that stores embeddings and finds the nearest ones to a query vector quickly. It is the retrieval engine under most RAG systems, though for smaller corpora a plain database with a vector extension is usually enough.

#

Vendor Lock-In

Business & Practice

How hard it would be to move to a different AI provider. Keeping prompts, evaluations and data in your own hands, rather than deep inside one vendor's platform, is what keeps that option open.

#

Vibe Coding

Agents & Tools

Building software by describing what you want in natural language and accepting largely unreviewed AI-generated code. Genuinely fast for prototypes and personal tools; a maintenance and security liability when the output reaches production unexamined.

#

Vision-Language Model

LLMs & Context

Also known as: VLM

A multimodal model that jointly understands images and text, able to describe a photo, read a chart, or locate a button in a screenshot. VLMs are the perception layer that makes computer-use agents possible.

#

Voice Cloning

Image, Video & Audio

Also known as: Voice synthesis

Generating speech in a specific person's voice from a short sample. Legitimate for dubbing your own content across languages; the same capability drives voice-based fraud, which is why consent and provenance marking matter here.

#

VRAM

Infrastructure

Also known as: Video memory, GPU memory

Memory on the graphics card, and the hard limit on what you can run locally. The rough rule is about two bytes of VRAM per parameter at half precision, before adding the KV cache, which is what makes quantization so valuable.

Related: GPU Quantization Parameters

#

Watermarking

Safety & Ethics

Embedding an imperceptible signal in generated content so it can later be identified as AI-made. Useful at scale but defeatable by determined editing, which is why it is best paired with cryptographic provenance rather than trusted alone.

#

Webhook

Infrastructure

A reverse API call: instead of you repeatedly asking whether something happened, the other service messages you the moment it does. The usual way an automation starts the instant a form is filled or a payment lands.

Related: API Trigger Integration

#

Weights

Architecture

The individual strengths of connections between units in a neural network: the actual thing adjusted during training and shipped as the model file. Open-weights models publish these numbers; closed models keep them behind an API.

#

Workflow vs Agent

Agents & Tools

A workflow follows a path you defined; an agent decides its own. Workflows are cheaper, predictable, and easier to debug, so the practical rule is to use one wherever the steps are known and reserve agents for genuinely open-ended work.

#

Working Memory

RAG & Memory

Also known as: Short-term memory

What an agent holds in its context window right now: the current conversation, recent tool results, immediate task state. It vanishes when the context is cleared, which is why anything that must survive has to be written down deliberately.

#

Zero-Shot

LLMs & Context

Asking a model to do a task with no examples, relying purely on instructions and what it learned in training. Modern models handle common tasks zero-shot well; unusual formats and domain-specific judgement still benefit from examples.

#