Inference
The process of a trained model predicting on new input — every query you make is inference, which drives API pricing and speed.
AI glossary
What does that AI word actually mean? 67 high-frequency terms, each explained in one direct sentence with the context you'll meet it in.
Embedding
Converting text or images into vectors so semantically similar items land close together in vector space — the foundation of search and RAG.
Vector Database
A database built for storing vectors and fast similarity search, powering retrieval over private knowledge bases.
Temperature
A sampling parameter trading determinism for creativity: low for precise tasks, high for brainstorming.
System Prompt
The upfront instructions that set a model's role, tone and constraints — the real engine behind most wrappers.
Prompt Injection
An attack hiding malicious instructions in text or web content to override a model's original instructions — a security boundary for any app ingesting external content.
Large Language Model
A very large neural network trained on massive text corpora that learns language, knowledge and reasoning by predicting the next token. ChatGPT, Claude and Gemini are all LLMs.
Token
The smallest unit of text a model processes — roughly 0.5–0.75 Chinese characters or 0.75 English words. APIs bill by tokens and context windows are measured in them.
Retrieval-Augmented Generation
Retrieve relevant documents from an external knowledge base before generation. It addresses stale knowledge and hallucination, and is the standard architecture for enterprise Q&A.
AI Agent
An AI system that plans multi-step actions, calls tools and adjusts from feedback. A chat model answers; an agent gets things done — booking trips, fixing code, running analyses.
Fine-tuning
Continuing to train a pretrained model on domain data so its style and capabilities specialize to your task. It changes the model's "instincts" more than RAG does, at higher cost.
Reinforcement Learning from Human Feedback
Humans rank model outputs, those preferences train a reward model, and RL optimizes against it — the key alignment technique that makes LLMs helpful and harmless.
Prompt
The instruction text you give a model. Output quality tracks prompt quality; the four basics are role, context, examples and output format.
Multimodal
A model that understands or generates multiple information types — text, images, audio, video. Photo-based tutoring and voice chat are multimodal abilities.
Hallucination
The model confidently inventing facts, citations or data. No complete cure yet; mitigations include requiring sources, RAG and human review of critical claims.
Context Window
The maximum amount of conversation and documents a model can hold at once, measured in tokens. Bigger windows fit longer docs; mid-window recall remains a common weakness.
Model Context Protocol
An open protocol from Anthropic that lets AI apps reach external tools and data sources through one standard interface — a USB-C for AI: integrate once, reuse across clients.
Function Calling
The model outputs which function to call and with which arguments; your code runs it and feeds the result back — the step that turns a chatbot into something that can act.
Reasoning Model
A model that deliberates at length before answering (OpenAI o-series, DeepSeek-R and peers). Clearly stronger at math, code and multi-step planning; slower and pricier per answer.
Chain-of-Thought
Prompting a model to write out intermediate steps before the final answer — a big gain on multi-step math and logic. Reasoning models internalize the same idea in training.
World Model
A model of how an environment evolves — predicting next states, not just next tokens. The backbone of planning in autonomous driving, robotics and video generation.
Diffusion Model
A model family that generates images, audio or video by iteratively denoising random noise (Stable Diffusion, Sora). It learns to add noise, then reverse it — which is what makes output steerable.
Transformer
The neural architecture introduced by "Attention Is All You Need" (2017). Its attention mechanism processes whole sequences in parallel and underpins virtually every modern LLM.
Attention Mechanism
The Transformer's core computation: weighting which other tokens each token should attend to, capturing reference and long-range dependencies — and the main cost driver as context grows.
Quantization
Compressing weights from high precision (16-bit) to lower (8/4-bit) to cut memory and speed up inference, letting large models run on laptops or phones at some accuracy cost.
Knowledge Distillation
Training a small "student" model on a large "teacher" model's outputs to compress capability into fewer parameters — the trick behind many small models punching above their weight.
Synthetic Data
Training or eval data generated by models where real data is scarce, private or costly to label. Standard practice is synthetic plus light human review — while guarding against bias amplification.
Guardrails
Checks around a model — content filters, tool-permission gates, output validation — that productized AI apps use to contain a model's residual unpredictability.
AI Alignment
The field making model goals and behavior match human intent and values — RLHF, constitutional methods, interpretability. In products it shows up as appropriately declining and hedging.
Red Teaming
Deliberately attacking your own model and product to find weaknesses — jailbreaks, prompt injection, data leakage, dangerous capabilities. Reports are becoming a compliance expectation.
Benchmark
Standardized test suites for comparing models (MMLU, GPQA, SWE-bench). Two caveats: scores get gamed, and leaderboard rank rarely predicts your own workload.
Agentic RAG
Retrieval-augmented generation where the model decides what to search, how often and when it has enough — multi-round retrieval with self-evaluation. More accurate, slower, costlier.
Streaming
Pushing tokens to the client as they are generated so text appears word by word — far better perceived latency than waiting for a whole answer, and now the default in chat products.
Computer Use
Agents that operate a GUI directly — reading the screen, moving the mouse, typing — instead of calling APIs. High ceiling, slower and error-prone; useful for systems with no API.
Vision-Language Model
Models that understand images and text together (GPT vision, Claude, Qwen-VL). They read screenshots, tables and charts — the basis of photo Q&A and computer-use agents.
Video Generation Model
Models that generate video clips from text or images (Sora, Kling, Runway). Practical today for short clips, ad assets and storyboards; long-shot consistency and physics stay hard.
Voice Agent
Speech AI that listens and talks in real time, can be interrupted and calls tools to get things done. End-to-end latency, not transcription accuracy, is what makes it usable.
Prompt Engineering
Systematically designing inputs for reliable outputs — role, context, examples, output schema, plus evaluation and iteration. Less incantation, more interface design.
Long Context
Very large single-pass inputs (100K–1M tokens) that can swallow whole books or repos. It costs more and suffers mid-context recall loss — put key facts near the ends.
Agent Memory
How an agent retains facts and preferences across sessions — short-term in context, long-term in vector or structured stores, with forgetting policies to limit noise.
OpenAI-Compatible API
An API mirroring OpenAI's request/response shape so the same client code works with other vendors or local models (DeepSeek, Qwen, Ollama) — the cheap path to staying model-agnostic.
Few-Shot Learning
Putting a few input-output examples in the prompt so the model imitates the pattern (zero-shot = none, few-shot = 2–5). Example quality beats quantity; consistent formatting matters most.
DeepSeek-V3
DeepSeek's open-source flagship model that matched frontier closed models at a fraction of the training cost (~$5.6M), famous for aggressive API pricing.
Claude Code
Anthropic's terminal-first AI coding agent that reads whole codebases and executes multi-step engineering tasks — complementing editor tools like Cursor.
Gemini CLI
Google's open-source terminal AI coding agent powered by Gemini, with a generous free tier (1,000 requests/day) — an alternative to Claude Code.
GitHub Copilot
The first mass-market AI coding assistant, deeply integrated with VS Code/JetBrains for completion and chat; newer versions add agent capabilities.
NotebookLM
Google's AI notebook that answers strictly from your uploaded documents (closed retrieval, low hallucination); famous for generating podcast-style audio overviews.
Sora
OpenAI's text-to-video model generating clips up to ~20 seconds — a leap toward narrated short films, though physics consistency remains hard.
Cursor
An AI-native code editor built on VS Code with whole-repo understanding, multi-file edits and an agent mode — among the most commercially successful AI dev tools.
NotebookLM Alternatives
Closed-document-QA tools similar to NotebookLM: AnythingLLM, Open WebUI (local models), Kimi document chat. Choose local for privacy, cloud for collaboration.
Vercel AI SDK
The de-facto frontend toolkit for AI features: one unified API across providers, streaming and tool-calling built in — swap models without rewriting UI code.
MCP Server
A server implementing the MCP protocol that wraps a tool or data source (database, browser, GitHub) into a standard AI-callable interface, reusable across MCP clients.
Open-Source LLM
LLMs with publicly downloadable weights (Llama, Qwen, DeepSeek). Pros: data privacy, custom fine-tuning, no API lock-in; cons: you manage compute and ops.
Local LLM
LLMs run on your own machine via tools like Ollama or llama.cpp. Full data privacy and zero API cost, limited by local hardware.
Model API
Cloud vendor interfaces for calling models, billed per token. Choose by capability, price, speed and reliability; most vendors offer OpenAI-compatible formats.
Tokenizer
The tool that splits text into the tokens a model processes. Token counts vary by tokenizer and directly affect API cost and context usage; Chinese runs ~0.6-1 token per character.
Fine-tuning vs RAG: which to use
Use RAG to give a model knowledge and facts (updatable, traceable); use fine-tuning to teach format and style (internalized). Most business cases start with RAG — cheaper and faster to iterate.
AI Coding Assistant
AI tools for coding in three forms: in-editor completion (Copilot), AI-native editors (Cursor), and terminal agents (Claude Code). Pick by workflow: completion, refactoring, or automation.
Prompt Template
Prompt skeletons separating fixed structure from variables ("You are a {role}, summarize {content} as {format}"). Shared templates make AI output consistent and reproducible across a team.
Embedding Model
Specialized models that convert text to vectors (distinct from generative LLMs) — the foundation of semantic search and RAG. Choose by language support, dimensions and speed.
Agent Framework
Frameworks for building AI agents (LangChain, LlamaIndex) that wrap model calls, tool orchestration and memory. Great for prototypes; plain APIs are lighter for simple cases.
Inference Engine
High-performance serving software for LLMs (vLLM, SGLang, TGI) optimizing memory and concurrent throughput — the core choice when self-hosting model services.
Ollama
The most popular tool for running LLMs locally: one-line model downloads (ollama run qwen3:8b), automatic model management and an OpenAI-compatible API. The de-facto entry point for local AI.
RAG vs Agent
RAG bolt a knowledge base onto a model with one-way retrieval; an agent is a full system that plans and acts autonomously. Agents may use RAG internally, but they solve different problems.
Temperature Parameter
Controls output randomness: 0-0.3 for code/translation/facts (stable); 0.7-1.0 for creative writing (divergent). Most products default to 0.7; set 0 for reproducible results.
Context Window Explained
The maximum text a model can read in one pass, in tokens (128K ≈ a 200-page book). Beyond it the model is blind; even within it, mid-context recall weakens — put key facts first or last.
AI Hallucination Solutions
No cure exists; five mitigations: require cited sources, ground answers with RAG, low-temperature sampling, human review of critical claims, and guardrails. Trustworthy systems layer them.