AI glossary: the words you keep running into
Only the common ones, two sentences each. Look up as you go — no memorizing.
AI
Software assistants that chat, write, see images, research — some operate computers.
Why it matters: That's what this site means by AI.
大模型 / LLM
The "brain" behind these AIs.
Why it matters: "New model released" = a stronger brain.
Token
How AI counts text — more text, more tokens.
Why it matters: APIs usually bill by tokens.
上下文 / Context
How much the AI holds in mind at once.
Why it matters: "1M context" = it can swallow a lot in one go.
提示词 / Prompt
What you say to the AI.
Why it matters: No templates needed — clarity beats magic words.
Agent
Give it a goal; it plans, uses tools, finishes.
Why it matters: The biggest direction of 2026.
幻觉 / Hallucination
AI confidently making things up.
Why it matters: This is why you verify important facts.
多模态
Understands images, audio, video — not just text.
Why it matters: That's why screenshots just work.
推理 / Reasoning
Thinking longer before answering — slower, more accurate.
Why it matters: The "think deeply" toggle is this.
API
How programs call AI, pay-per-use.
Why it matters: Skip it if you don't code.
API Key
The password string for API access.
Why it matters: Leak it and others spend your money.
CLI / 命令行
The black text window where tools like Claude Code live.
Why it matters: Scary-looking; copy-paste is enough.
GitHub
Software hub and project warehouse.
Why it matters: Can't read it? Send the link to your AI.
开源 / Open Weights
Model files anyone can download and run.
Why it matters: Why Qwen/DeepSeek get so many community tests.
Benchmark
Standardized exams for comparing models.
Why it matters: High scores ≠ good in practice — check real tests.
RAG
Look things up first, then answer from what was found.
Why it matters: How most company-knowledge bots work.
Skill
A how-to manual you hand the AI.
Why it matters: Gets work done your way, every time.
MCP
The universal adapter connecting AI to outside tools.
Why it matters: One sentence is enough until you need it.
微调 / Finetune
Re-training someone's model into a variant.
Why it matters: Community "custom builds" are this — quality varies.
量化 / Quantization
Compressing a model so ordinary computers can run it.
Why it matters: The Q4/Q8 tags you see everywhere in local AI.