Lexicon
The words of AI
A term is not understood through its definition, but through its place in the system. Each entry gives the essential in one sentence, an analogy, a concrete example, and what the term does not mean.
The map of notions
Four layers, from the engine to the organization. See where every term sits, and which ones masquerade as each other.
I · The model
What gets trained: the engine.
- LLMAlso: large language model
- The engine: a statistical model trained on vast amounts of text, able to produce language. It is the thing every generative-AI product puts to work.
- Frontier modelAlso: frontier AI
- A model from the most advanced generation of the moment, the one pushing the frontier of what AI can do, at a training cost only a handful of players can afford.
- Open weightAlso: open-weight model
- A model whose weights are published: anyone can download it and run it on their own machines, without ever holding the full recipe of how it was made.
- WeightsAlso: parameters
- The billions of numerical parameters adjusted during training: the very substance of the model. Whoever holds the weights file holds the model.
- TrainingAlso: pre-training
- The manufacturing phase of a model: months of computation over vast amounts of data to adjust its weights. It happens once, and it is what costs fortunes.
II · Access
How you talk to it: the counter.
- APIAlso: application programming interface
- The counter through which one piece of software talks to another. For AI, it is the channel your products and internal tools use to send requests to the model, on demand, billed by the token.
- TokenAlso: text token
- The model's unit of account: a fragment of text worth about three quarters of a word on average. Everything is measured in tokens: document length, limits, and the bill.
- Context windowAlso: context
- Everything the model has in front of it when it answers: your request, the documents provided, the history of the exchange. Whatever is not in it does not exist for the model.
- PromptAlso: instruction
- The brief you give the model: the request, the context, the constraints and the expected format. The quality of the answer is largely decided by the quality of that brief.
- Inference
- The moment the model works: it receives a context and produces its answer, token by token. Every request is one inference, and each one has a cost.
III · Orchestration
How you put it to work: the workshop.
- HarnessAlso: scaffolding
- The working environment built around an AI model: its tools, its access rights, its rules and its guardrails.
- AgentAlso: AI agent
- An AI model allowed to chain actions toward a goal (act, observe the result, correct) instead of answering once and stopping.
- Agentic loopAlso: loop
- The cycle that turns AI from talk into action: act, observe the result, correct, repeat, until the goal is reached or the blockage is acknowledged.
- ToolAlso: tool use, function calling
- A capacity for action the harness puts in the model's hands: search the web, read a file, write to the CRM, trigger a payment. Without tools a model talks; with them, it acts.
- MCPAlso: Model Context Protocol
- An open standard acting as a universal socket between models and your systems: instead of one custom integration per tool and per product, one common socket everyone knows how to plug into.
IV · The organization
How people work with it: the team.
- AssistantAlso: copilot
- AI in an advisory posture: it answers when asked, one request at a time, and the human keeps their hands on the wheel between steps. Most enterprise usage starts here.
- Human oversightAlso: human in the loop, HITL
- The organization of control when AI does the work: who checks what, at which moments, with what power to stop. It is what turns a delegation into measured trust.
- Agentic workflow
- A business process redesigned so that agents own entire steps of it, under defined oversight, rather than a chat bolted onto the side of the existing process.