GLOSSARY / CONCEPTS
AI Glossary
Core AI concepts explained in plain language — no jargon, no fluff.
01Large Language Model (LLM)A neural network trained on massive text corpora that predicts and generates language, forming the engine behind every AI chat assistant.02Generative AIAI that creates new content — text, images, video, audio or code — rather than only classifying or predicting from fixed options.03Prompt (Prompt Engineering)The instruction you give an AI model; prompt engineering is the craft of writing instructions that reliably produce the output you want.04TokenThe small chunks of text (roughly a word or part of a word) that AI models actually read and generate; pricing and context limits are both counted in tokens.05HallucinationWhen an AI model states false information confidently — invented facts, fake citations, plausible but wrong answers.06MultimodalModels that work across multiple media types at once — reading images, generating audio, watching video — instead of text alone.07Fine-tuningAdditional training that adapts a pre-trained model to a specific style, task or domain, using your own examples.08Open Weights (Open-Source Models)Models whose trained parameters are publicly downloadable, so anyone can run, audit or adapt them on their own hardware.09Context WindowHow much text a model can consider at once — its working memory, measured in tokens, spanning your conversation and any attached documents.10Retrieval-Augmented Generation (RAG)A technique where a model first searches relevant documents (or the web) and then answers based on what it found, citing sources.11AI AgentAn AI system that plans and executes multi-step tasks autonomously — browsing, writing code, using tools — instead of only answering single questions.12InferenceThe act of running a trained model to produce an output — every chat reply, generated image or API call is one inference pass.13EmbeddingA numeric representation of text, images or audio that captures meaning, letting machines find similar content by comparing vectors.14API (AI API)The programming interface that lets developers call an AI model from their own software, usually billed per use.15AI Safety & AlignmentThe practice of making AI systems helpful, honest and harmless — following human intent and values rather than producing capable but harmful outputs.16Foundation ModelA large general-purpose model trained on broad data, designed to be adapted to many downstream tasks rather than one single job.17Diffusion ModelThe family of models behind most modern image and video generation: they learn to remove noise step by step until a clear picture emerges.18LoRA (Low-Rank Adaptation)A lightweight way to teach a base model a new style, character or skill — a small add-on file trained in hours instead of retraining the whole model.19QuantizationCompressing a model's weights from 16-bit numbers to 8, 4 or fewer bits — making big models run on smaller hardware with a modest quality cost.20TokenizerThe component that chops text into tokens — the small units a model actually reads, writes and bills by.21Mixture of Experts (MoE)An architecture that routes each request to a small subset of specialized sub-networks — frontier-level quality at a fraction of the compute per token.22Vector DatabaseA database that stores text (or images) as numeric embeddings and finds the most similar items in milliseconds — the memory layer behind RAG and AI search.23System PromptThe hidden instructions given to a model before your conversation starts — defining its role, tone, boundaries and rules.24TemperatureThe dial that controls how predictable or creative a model's output is — low for factual precision, high for variety.25TransformerThe neural-network architecture behind virtually every modern AI model — from chatbots to image generators — built around a mechanism called attention.26RLHF (Reinforcement Learning from Human Feedback)The training step that makes raw language models polite, helpful and safe — by teaching them from thousands of human judgments about which answers are better.27DistillationTraining a small, fast model to imitate a large one — capturing much of the big model's quality at a fraction of the cost to run.28GuardrailsThe safety layer around an AI system: filters and rules that catch toxic output, blocked topics, prompt attacks and unsafe actions before they reach users.29Tool CallingHow models use software: instead of only producing text, they emit structured requests to call calculators, search engines, APIs — then weave the results into an answer.30Structured OutputMaking a model return valid JSON that matches your schema — the difference between a demo and software you can put in production.31Few-Shot PromptingTeaching by example inside the prompt: show the model a few input-output pairs, and it imitates the pattern on new inputs — no retraining.32Prompt InjectionThe hacking technique for AI systems: sneaking instructions into the data a model reads so it ignores its real orders — the SQL injection of the LLM era.33Chain-of-Thought (Reasoning)Letting a model work through a problem step by step before answering — dramatically better on math, logic and planning, at the cost of extra time and tokens.34Evals (Evaluations)Systematic tests for AI behavior: fixed task sets and graders that tell you whether a model or prompt change actually made things better.35JailbreakA crafted prompt that tricks an AI assistant into ignoring its safety rules and producing output it is designed to refuse.36Vision-Language Model (VLM)A model that reads images as well as text: send a screenshot, chart, photo or diagram and ask questions about it in plain language.37Speech Recognition (ASR)Converting spoken audio into text — the foundation under meeting transcribers, voice assistants and subtitling tools, now near-human in accuracy for clear audio.38Voice CloningCreating a synthetic copy of a specific person's voice from recordings — seconds of audio now suffice, which is both a creative superpower and a fraud vector.39GroundingAnchoring a model's answers in verifiable sources — your documents, a database or the live web — instead of trusting its memory.40Small Language Model (SLM)Compact models (roughly 1–13B parameters) tuned to punch above their weight — fast, cheap and private enough to run on phones, laptops and edge devices.41Latent SpaceThe internal numeric space where a model represents concepts — nearby points mean similar things, and moving in a direction means changing the concept smoothly.42InpaintingRegenerating only a selected part of an image — erase a stranger, swap a product, extend a background — while the rest stays pixel-identical.43CheckpointA saved snapshot of a model's weights — the file you download when you grab an open model, and the unit the community shares on hubs.44KV CacheThe memory that lets a model keep notes about tokens it has already read instead of recomputing them — the reason long conversations eat GPU memory.45RerankingA second-pass model that re-sorts search results by true relevance — the quiet quality boost behind serious RAG and enterprise search.46AGI (Artificial General Intelligence)A hypothetical AI that matches humans across essentially all cognitive work — the industry's stated destination and its most contested word.47SycophancyThe trained-in habit of AI assistants agreeing with you, flattering you and abandoning correct answers under pushback — a direct side effect of human-feedback training.48Sampling ParametersThe dials beyond temperature (top-p, top-k, penalties) that shape which words a model picks — tuning output variety, focus and repetitiveness.49MCP (Model Context Protocol)An open standard that lets AI models plug into external tools and data sources through one common interface — “USB-C for AI applications”.50Speech Synthesis (TTS)Turning text into natural-sounding speech — the output half of voice AI, now indistinguishable from human narration for many languages.51AI UpscalingEnlarging images or video beyond their original resolution — from faithful 2x sharpening to generative models that invent plausible new detail.52AI WatermarkingInvisible markers embedded in AI-generated content (or visible labels on it) so synthetic media can be identified later — a core piece of AI provenance.53Synthetic DataArtificially generated training data — model outputs or simulations standing in for real records when real data is scarce, private or expensive.54On-Device AIAI that runs entirely on your phone, laptop or appliance — no cloud round-trip — trading peak capability for privacy, latency and offline freedom.55Training (Pretraining & Post-Training)The process that creates a model: pretraining learns language from oceans of text, post-training shapes it into a helpful assistant.56Context EngineeringThe discipline beyond prompt-writing: deciding what goes into a model's limited context window — instructions, retrieved documents, tool results, history — and in what order.57AttentionThe mechanism that lets a model decide which parts of the input matter for each output — the single idea that made modern AI possible.58Instruction TuningFine-tuning a base model on thousands of instruction–response pairs so it learns to follow what you ask — the step between a raw predictor and a usable assistant.59Red TeamingDeliberately attacking an AI system — jailbreaks, prompt injections, weird edge cases — before an adversary does, then fixing what broke.60Test-Time ComputeSpending more compute while answering — thinking longer, sampling many attempts, verifying — instead of only making the model bigger beforehand.61AI SlopThe flood of low-effort, mass-produced AI content — listicles, fake images, bot comments — that increasingly pollutes search, social feeds and now training data.62AI BiasSystematic unfairness learned from data: models repeating and amplifying stereotypes about gender, race, language and culture at machine scale.63Human-in-the-LoopDesigning AI systems so a person approves, corrects or takes over at the right moments — the difference between automation you trust and automation you survive.64AI DetectorsTools that guess whether text or media was machine-made — useful as signals, unreliable as judges, and increasingly entangled with plagiarism accusations.65Model MergingCombining the weights of multiple fine-tuned models into one — blending skills without paying for another training run.66ChunkingCutting documents into retrieval-sized pieces before embedding — the unglamorous RAG decision that quietly decides answer quality.67OverfittingWhen a model memorizes its training examples instead of learning the pattern — perfect on the practice test, brittle on anything new.68Data LabelingThe human work underneath every AI system: tagging text, images and audio so models have ground truth to learn from — an industry of millions of workers.69Model CardThe nutrition label for an AI model: what it is, what it was trained on, how it was evaluated, where it fails — documentation that makes models comparable and auditable.70AI GovernanceThe rules, roles and review processes that keep AI systems accountable — inside companies (policy, audits) and across societies (regulation like the EU AI Act).71EnsembleCombining multiple models' outputs for a better answer than any one alone — majority votes, best-of-N selection, or mixture-of-expert routing at scale.72Data CurationChoosing what goes into training — filtering, deduplicating and weighting data — now matters as much as model architecture in the race for quality.