AI jargon for non-engineers

Key facts · Part 1 Essentials

• Model: the engine. ChatGPT is the car.
• LLM: autocomplete at enormous scale, plausible text, not looked-up facts
• Prompt: output quality tracks input quality
• Tokens: the meter that vendors bill by
• Context window: the desk, working memory for one conversation
• Hallucination: confidently wrong, so verify
• RAG: open-book exam using your own documents
• Agent: a chatbot tells you how to book the flight, an agent books it
• Fine-tuning: a training course that changes behavior, not a way to keep facts current
• Who makes what: OpenAI makes ChatGPT, Anthropic makes Claude, Google makes Gemini, Microsoft sells Copilot, Meta makes Llama

Key facts · Part 2 Deeper

• Inference: the meter running every time someone hits enter; training happens once
• Reasoning models: think step by step first, slower and costlier, better at hard problems
• Open-weight vs closed: Llama-style models you can run yourself vs services you reach through a vendor
• A.P.I. vs chat app: the A.P.I. plugs a model into company software, billed by tokens
• Shadow A.I.: employees using unapproved tools, so don't paste confidential data
• M.C.P.: USB-C for A.I. tools, the plug that connects agents to tools and data
• Storylines: bills climb with inference and reasoning, hallucinations are reduced but not removed, agents raise permission and security questions

Key facts · Part 3 Drill

• The model is the engine, and ChatGPT is the car
• Tokens are the meter, the context window is the desk, and inference is the meter running
• Hallucination means confidently wrong, so verify
• RAG is an open-book exam, and an agent acts instead of just answering, ideally with a human in the loop
• OpenAI makes ChatGPT, Anthropic makes Claude, Google makes Gemini, Microsoft sells Copilot, Meta makes Llama, and don't paste confidential data into unapproved tools