Speak AI: the jargon, decoded.
AI conversations are full of terms that make smart people feel dumb — which is a tax on your decisions, because vendors exploit exactly that feeling. Fifteen definitions, plain English, each with the reason it matters to you.
01 The core mechanics
| Term | Plain English | Why you care |
|---|---|---|
| Token | The word-chunks AI reads and writes (~¾ of a word) | It's how API usage is billed and limits are measured |
| Context window | How much text the model can 'hold in mind' at once | Why long documents work in some tools and get 'forgotten' in others |
| Inference | The act of the model generating an answer | 'Inference costs' = what running AI actually costs vendors — and eventually you |
| Parameters | The model's internal dials, set during training | Mostly a bragging number; capability ≠ parameter count anymore |
| Training vs. fine-tuning | Building the brain vs. giving it a specialty course | 'We fine-tuned a model' is a real (modest) claim; 'we built our own AI' usually isn't |
02 The capability layer
| RAG | Fetching relevant documents and handing them to the model before it answers | How 'chat with your data' products actually work — retrieval quality is the whole product |
| Agent | AI that takes multi-step actions toward a goal, using tools | The word every vendor now uses; ask "what actions, with what supervision?" |
| Multimodal | Handles images/audio/video, not just text | Why you can photograph a form and ask about it |
| Reasoning / thinking modes | Generating deliberate steps before answering | Slower, better on hard problems — you're buying rigor with seconds |
| MCP / connectors | Standard plugs connecting AI to apps and data | The power AND the security surface — see our connector-security lessons |
03 The evaluation layer
| Hallucination | Confident, fluent, wrong | Not a bug being fixed next quarter — a property to manage forever |
| Benchmark | A standardized model test | Directionally useful, gameable, and rarely about YOUR task |
| Open vs. closed model | Downloadable weights vs. vendor-hosted API | Control-versus-convenience; relevant when data can't leave the building |
| Prompt injection | Malicious instructions hidden in content the AI reads | THE attack to understand before connecting AI to email and files |
| AGI | Hypothetical human-level-at-everything AI | A debate, not a product. Anyone selling it is selling something else |
Fluency's real payoff is the follow-up question: "Agentic how — what actions can it take, and what's the human checkpoint?" "RAG over which sources, updated when?" Watch pitches get honest fast when the jargon stops working as fog.
Take the last AI product pitch you received and reread it with the tables above open. Highlight every term doing fog-work. Draft the two follow-up questions that would clear it.
Open the Playground →This week's challenge
Use three of these terms correctly in real conversations this week — and once, politely ask a vendor or colleague to define one they used. Fluency plus the confidence to ask: that's the whole upgrade.