AI Glossary

AI Glossary

Plain-English definitions for the AI terms a small business owner actually runs into — without the buzzwords. Each term has its own page with use cases, pitfalls, and Central Florida examples. Skim the list, click anything that matters to your business, and email me if you want a real opinion on it.

A

• Active LearningA training approach where the model picks which examples it wants labeled next so humans don't waste time labeling easy ones.Read more →
• Agentic AIAI that doesn't just answer — it plans, takes steps, and uses tools to finish a task.Read more →
• AI AdoptionHow a team actually starts using AI in real work, not just buying licenses.Read more →
• AI AgentA piece of software that uses an LLM to make decisions and call tools on its own.Read more →
• AI AlignmentTraining so a model behaves the way humans actually want it to.Read more →
• AI AssistantAn AI that helps a person do work — drafts, summarizes, answers — but the human stays in charge.Read more →
• AI AuditingA formal review of how an AI system makes decisions, what data it uses, and whether it's fair.Read more →
• AI BiasWhen a model treats some people or topics worse because of skewed training data.Read more →
• AI ComplianceMaking sure AI use lines up with the laws and standards your industry already follows.Read more →
• AI CopilotAn AI built into a tool you already use (Word, Excel, your CRM) that drafts alongside you.Read more →
• AI EthicsThe principles and judgment calls behind using AI in ways that don't quietly harm people.Read more →
• AI EvalsTests that measure whether an AI's outputs are actually good enough for your use case.Read more →
• AI GovernanceThe rules a business sets for who can use AI, with what data, for what jobs.Read more →
• AI HallucinationWhen an AI confidently makes up facts that aren't true.Read more →
• AI LiteracyHow well your team understands what AI can and can't do — the foundation under any rollout.Read more →
• AI MemoryHow an AI agent remembers context across turns or sessions instead of starting fresh every time.Read more →
• AI PilotA small, time-boxed test of an AI tool to see if it actually delivers value before rolling it out.Read more →
• AI PipelineThe chain of steps — data in, processing, model, output — that moves work through an AI system.Read more →
• AI ReadinessWhether your data, processes, and team are in shape to actually benefit from AI.Read more →
• AI ROIHonest math on what AI costs vs. what it actually saves or earns you.Read more →
• AI SandboxA walled-off space where staff can experiment with AI safely on non-sensitive data.Read more →
• AI SlopMass-produced low-effort AI content that clutters search results and feeds.Read more →
• AI StrategyA written plan for where AI fits in your business, what it's for, and what it won't be for.Read more →
• AI Use CaseA specific job AI is being put to — "answer after-hours phone calls," "draft proposals," not "do AI."Read more →
• AI Vendor Due DiligenceThe checklist you run on an AI tool's maker before signing — security, data use, support, viability.Read more →
• AI Voice AgentAn AI that answers phone calls, books appointments, or handles inbound voice inquiries.Read more →
• AI WorkflowA multi-step process where AI handles parts of the work and humans handle the rest.Read more →
• AnthropicThe AI safety company behind Claude — known for the constitutional-AI approach to alignment.Read more →
• API KeyThe secret token that authenticates your app to an AI provider — guard it like a password.Read more →
• API Rate LimitingCaps that AI providers put on how many calls or tokens per minute you're allowed.Read more →
• ASR (Speech-to-Text)Tech that turns spoken audio into written text — the listening half of voice AI.Read more →
• Attention MechanismThe trick inside transformers that lets a model decide which words in a sentence matter most for each prediction.Read more →
• AutoGenMicrosoft's open-source framework for building multi-agent AI systems that talk to each other.Read more →
• AutoGPTAn early autonomous-agent project that kicked off the agentic-AI craze in 2023.Read more →
• AutoMLTools that automate the picking, training, and tuning of machine-learning models so non-experts can use them.Read more →
• AI Code GenerationAI writing or completing source code — the engine inside Copilot, Cursor, and similar tools.Read more →
• AI Content ModerationFilters that keep harmful, off-brand, or unsafe outputs from reaching users.Read more →
• AI ExplainabilityHow clearly a model can show why it gave the answer it gave.Read more →
• AI GuardrailsHard rules that block an AI from saying or doing things it shouldn't.Read more →
• AI Image GenerationAI that creates pictures from a text prompt — Midjourney, Stable Diffusion, DALL-E sit here.Read more →
• AI JailbreakA prompt trick that gets an AI to bypass its safety rules.Read more →
• AI Knowledge BaseYour curated docs — wiki, SOPs, policies — fed into a RAG system so AI can answer from them.Read more →
• AI Red TeamingHiring people to attack your AI system on purpose to find weaknesses before bad actors do.Read more →
• AI RefusalWhen a model declines to answer — sometimes for good reason, sometimes overcautious.Read more →
• AI Tool UseWhen an AI calls outside tools (search, calendar, your CRM) to finish a task.Read more →
• AI Training DataThe text, images, or audio fed in during training — quality and licensing matter as much as volume.Read more →
• AI Video GenerationAI that creates moving footage from a prompt or a still — Sora, Veo, Runway.Read more →

B

• Benchmark (AI)A standard test (like MMLU, HumanEval) used to compare AI models head-to-head.Read more →
• BPE (Byte-Pair Encoding)The most common way LLMs chop text into tokens — a balance between letters and whole words.Read more →

C

• Chain of ThoughtA prompting trick that asks the model to show its reasoning step by step.Read more →
• ChatbotA text-based conversational interface — older rule-based ones and modern LLM-powered ones both count.Read more →
• ChatGPTOpenAI's consumer chat product — the app that put generative AI on every desk in 2022.Read more →
• ClaudeAnthropic's family of AI models — known for long context, careful tone, and coding.Read more →
• Closed-weights ModelA model whose internal weights aren't published — you use it via someone else's API.Read more →
• Cloud AIAI run on someone else's servers — fastest path to capability, with the trade-offs of sending your data out.Read more →
• Constitutional AIAnthropic's training method that teaches a model to grade and improve its own answers against a written set of principles.Read more →
• Context WindowHow much text a model can hold in mind at once for a single request.Read more →
• Continual LearningLetting a model keep learning from new data without forgetting what it already knew.Read more →
• Conversational AIAny AI built around back-and-forth dialog — chat, voice, or both.Read more →
• Cosine SimilarityThe math vector databases use to score how close two pieces of meaning are.Read more →
• CrewAIAn open-source framework for orchestrating teams of AI agents with assigned roles and tasks.Read more →

D

• Data GovernanceThe rules for what data you have, who can touch it, and how AI is allowed to use it.Read more →
• Data LabelingThe grunt work of tagging examples so a supervised model can learn from them.Read more →
• Decision IntelligenceA discipline that combines data, AI, and decision theory to actually improve business choices, not just dashboards.Read more →
• Deep LearningThe branch of machine learning that uses many-layer neural networks — what powers modern AI.Read more →
• DeepSeekA Chinese open-weights AI lab whose models compete with OpenAI and Anthropic at much lower cost.Read more →
• Differential PrivacyMath that lets you analyze data without exposing any single person's record.Read more →
• Diffusion ModelThe architecture behind most image generators — it learns to undo noise step by step into a picture.Read more →
• Discriminative AIAI that classifies or predicts — "is this email spam?" — the older sibling of generative AI.Read more →
• Document AIAI that pulls structured data out of contracts, invoices, forms, and PDFs.Read more →

E

• Edge AIRunning AI on the device where data is collected — phone, camera, sensor — instead of sending it to the cloud.Read more →
• ElevenLabsThe leading AI voice synthesis platform — voice cloning and natural-sounding TTS.Read more →
• EmbeddingTurning text into a list of numbers so a computer can compare meaning, not just words.Read more →
• Embedding ModelThe model that produces embeddings — separate from the LLM that does the talking.Read more →
• Encoder-Decoder ArchitectureA neural-network shape with one half that reads input and another that writes output — used in translation, summarization, speech.Read more →
• EU AI ActEurope's law that classifies AI use cases by risk and sets rules for each tier.Read more →

F

• Feature StoreA central library of pre-built data signals ("features") that ML models reuse instead of recomputing each time.Read more →
• Federated LearningTraining a shared model across many devices without the raw data ever leaving each device.Read more →
• Few-shot PromptingGiving a model 2-5 examples in the prompt to teach it the format you want.Read more →
• Fine-tuningTaking a pre-trained model and training it further on your own data so it speaks your domain.Read more →
• Foundation ModelA big general-purpose model trained on massive data — the base layer everything else sits on.Read more →
• Frontier ModelThe most capable model class at the moment — GPT-5, Claude 4, Gemini 2 Pro tier.Read more →
• Function CallingLetting an LLM trigger your code — e.g., look up a customer record or send an email.Read more →

G

• GAN (Generative Adversarial Network)A pair of networks where one generates fakes and the other tries to spot them — early image-generation tech.Read more →
• GeminiGoogle's family of AI models — strong on multimodal (text + image + video) tasks.Read more →
• Generative AIAI that creates new content — text, images, audio, code — instead of just classifying or predicting.Read more →
• Golden DatasetA small set of hand-checked correct examples you use to grade an AI's outputs.Read more →
• GPTOpenAI's family of language models — the household name behind ChatGPT.Read more →
• Grounding (AI)Forcing an AI to answer using your documents only, instead of guessing from training data.Read more →

H

• HIPAA and AIWhat healthcare-adjacent businesses have to do before piping protected health info into any AI tool.Read more →
• Human in the LoopA workflow where AI drafts but a person reviews and approves before anything ships.Read more →
• Hybrid SearchCombining keyword search and semantic (vector) search so you get exact matches and meaning matches in one ranking.Read more →
• HyperparameterA knob you set before training (like learning rate or batch size) — different from what the model learns.Read more →

I

• Inference (AI)The act of an AI generating an answer — the part you actually pay for at runtime.Read more →
• Inference CostWhat you pay each time an AI runs — usually per million tokens, in and out.Read more →
• Inference ServerSoftware that hosts a model and serves predictions — how a model gets put behind an API.Read more →

K

• Knowledge GraphA network of facts (entities and relationships) AI can traverse — adds structure where LLMs only see text.Read more →

L

• Labeled DataExamples that have the right answer attached — what supervised learning needs.Read more →
• LangChainAn early framework for chaining LLM calls, tools, and memory into apps — popular but opinionated.Read more →
• Large Language Model (LLM)A model trained on enormous amounts of text that can read and write in plain language.Read more →
• Latency (AI)How long the model takes to start answering and to finish — matters for voice and chat UX.Read more →
• LlamaMeta's open-weights LLM family — the most-used base model for self-hosted AI.Read more →
• LlamaIndexA framework focused on connecting your documents and data to LLMs — RAG-first counterpart to LangChain.Read more →
• LLMOpsThe operational discipline of running LLMs in production — versioning prompts, evals, logging, cost control.Read more →
• LoRAA cheap way to fine-tune a giant model by training only a small adapter on top — minutes instead of days.Read more →
• Loss FunctionThe score that tells a model during training how wrong it just was — what gradient descent shrinks.Read more →

M

• Model DistillationTraining a small model to copy a big model — cheaper, faster, almost as good.Read more →
• Machine LearningThe umbrella discipline where software learns from data instead of being explicitly programmed.Read more →
• Model Context Protocol (MCP)An open standard for letting AI assistants connect to tools, data, and apps cleanly.Read more →
• Meta AIMeta (Facebook)'s AI division — best known for the open-weights Llama models.Read more →
• MidjourneyA popular AI image generator known for cinematic, painterly outputs — runs inside Discord.Read more →
• MistralA French AI lab making efficient open-weights models popular for self-hosting.Read more →
• Mixture of Experts (MoE)A model that splits work across specialist sub-models — bigger total brain, lower running cost.Read more →
• MLOpsThe DevOps cousin for traditional machine learning — pipelines, model versioning, monitoring.Read more →
• Model CardA short doc the model maker publishes covering what it's good at, where it fails, and risks.Read more →
• Model RegistryThe internal catalog where you track every model version, its training data, and where it's deployed.Read more →
• Model ServingThe plumbing that takes a trained model and exposes it as something other software can call.Read more →
• Multi-Agent SystemA team of AI agents that talk to each other to finish a job no single agent could do alone.Read more →
• Multimodal AIAn AI that can read, look, and listen — handling text, images, and audio in the same prompt.Read more →
• Model TrainingThe expensive process of teaching a model from scratch — done by labs, not by most businesses.Read more →

N

• Named Entity Recognition (NER)Pulling out the names of people, places, companies, and dates from a block of text automatically.Read more →
• Neural NetworkA web of math units loosely modeled on brain cells — the engine inside almost every modern AI.Read more →

O

• OCR (Optical Character Recognition)Reading text out of images and scans — the base layer underneath most document AI.Read more →
• OllamaA free tool that runs open-weights LLMs locally on your laptop — the easiest path into self-hosted AI.Read more →
• On-Device AIAI that runs entirely on the user's hardware — better privacy, lower latency, no cloud bill.Read more →
• Open-weights ModelA model whose weights are published — you can run it on your own hardware, free.Read more →
• OpenAIThe lab behind ChatGPT, GPT-5, and DALL-E — kicked off the modern generative-AI era.Read more →
• OverfittingWhen a model memorizes its training data instead of learning the general pattern — looks great on practice, fails in real life.Read more →

P

• Parameters (AI)The internal numbers a model learns during training — more isn't always better, but it's the rough size measure.Read more →
• PEFT (Parameter-Efficient Fine-Tuning)An umbrella for fine-tuning techniques (LoRA, adapters) that touch only a tiny slice of the model.Read more →
• Persona PromptingTelling the model who it is ("you are a senior copy editor…") to steer its tone and style.Read more →
• Pre-trainingThe huge first training phase where a model reads the internet to learn language.Read more →
• Predictive AIOlder-school AI that forecasts a number or category — churn, fraud, demand — instead of generating content.Read more →
• Process AutomationUsing software (often AI) to handle steps in a business process that used to be manual.Read more →
• Prompt ChainingWiring multiple prompts together so one model's output becomes the next prompt's input.Read more →
• Prompt EngineeringThe craft of writing instructions that get the AI to give you the answer you actually wanted.Read more →
• Prompt HygieneDiscipline around what you put into prompts — no client PII, no secrets, no internal-only docs unless approved.Read more →
• Prompt InjectionAn attack where a hidden instruction in user input or a webpage hijacks the AI's behavior.Read more →
• Prompt LibraryA shared set of approved prompt templates your team uses for common tasks.Read more →
• Prompt TemplateA reusable prompt with fill-in-the-blank slots — keeps output consistent across staff.Read more →

Q

• QuantizationCompressing a model by storing its numbers at lower precision — cheaper to run, slightly lower quality.Read more →

R

• Retrieval-Augmented Generation (RAG)Giving an AI access to your documents so it answers from your facts, not its training data.Read more →
• Re-rankingA second pass that reorders search results using a smarter, slower model — common in RAG.Read more →
• ReAct PromptingA pattern where the model alternates Reasoning steps with Acting steps (tool calls) — basis of many agents.Read more →
• Reasoning ModelA model trained to slow down and work through problems step-by-step — better on math, code, planning.Read more →
• RegularizationTraining tricks (dropout, weight decay) that stop a model from memorizing and force it to generalize.Read more →
• Reinforcement LearningTraining where the model takes actions and learns from rewards or penalties — the basis of game AIs and RLHF.Read more →
• Responsible AIAn umbrella label for fairness, transparency, and safety practices in AI deployment.Read more →
• RLHFA training step where humans rank model answers, teaching it which kinds of replies are preferred.Read more →
• RPA (Robotic Process Automation)Software bots that mimic clicks and keystrokes to automate desktop work — older sibling to AI agents.Read more →

S

• Self-AttentionThe mechanism that lets every word in a sentence look at every other word to figure out meaning.Read more →
• Self-Supervised LearningTraining where the model creates its own labels from unlabeled data — how LLMs learn from raw text.Read more →
• Semantic SearchSearch that matches meaning instead of exact keywords — "car repair" finds "auto mechanic."Read more →
• Sentiment AnalysisAI that scores text as positive, negative, or neutral — used on reviews, support tickets, social posts.Read more →
• SOC 2 and AIWhat you need to know about feeding controlled data into AI tools when your business is SOC 2 audited.Read more →
• Stable DiffusionAn open-weights image generator that runs on consumer GPUs — the foundation of most self-hosted image AI.Read more →
• Supervised LearningTraining a model on examples that already have the right answer attached.Read more →
• Synthetic DataAI-generated training data used when real data is scarce, sensitive, or imbalanced.Read more →
• System PromptA hidden setup instruction that shapes how the AI behaves for an entire conversation.Read more →

T

• Test-Time ComputeSpending more compute at the moment the model answers (instead of during training) to get better results.Read more →
• Text-to-ImageAI that generates a picture from a written prompt.Read more →
• Text-to-VideoAI that generates short video clips from a written prompt — Sora, Veo, Runway sit here.Read more →
• Tokens (AI)The little chunks (usually 3-4 chars) that LLMs read and write in — what you're billed on.Read more →
• Token CostWhat you pay per million input or output tokens — varies wildly by model.Read more →
• TokenizationThe step that splits raw text into tokens before the model can read it.Read more →
• Transfer LearningTaking a model trained on one job and adapting it to a related job — saves enormous amounts of training cost.Read more →
• TransformerThe neural-network architecture that powers GPT, Claude, Gemini — basically all modern LLMs.Read more →
• Tree of ThoughtsA reasoning technique where the model explores multiple solution branches and picks the best, instead of one straight line.Read more →
• Trustworthy AIA label for AI that is fair, transparent, secure, and reliable — overlaps heavily with responsible AI.Read more →
• TTS (Text-to-Speech)Tech that turns written text into natural-sounding audio — the talking half of voice AI.Read more →

U

• Unsupervised LearningTraining a model on raw data with no labels — it has to find structure on its own.Read more →

V

• VAE (Variational Autoencoder)A neural-network shape that learns to compress and regenerate data — used in image generation pipelines.Read more →
• Vector (AI)A list of numbers that represents the meaning of a piece of text or image — the format embeddings live in.Read more →
• Vector DatabaseA database designed to store embeddings and find the closest matches fast — the engine behind RAG.Read more →
• Vibe CodingBuilding software by describing what you want in plain English and letting AI write the code.Read more →
• Vision ModelAn AI that can look at images or video and describe, label, or reason about them.Read more →
• vLLMA high-performance open-source server for running LLMs efficiently in production.Read more →
• VoicebotA chatbot you talk to instead of type to — common shorthand for AI voice agents on phone or app.Read more →

W

• Whisper (AI)OpenAI's open-weights speech-to-text model — the default starting point for transcription.Read more →
• Workflow OrchestrationTools and patterns for coordinating multi-step jobs across people, systems, and AI agents.Read more →

Z

• Zero-shot PromptingAsking the model to do a task with no examples — just instructions.Read more →