Skip to content

FREN

GlossaryOften confused

The frontier pages

Often confused

An ordinary glossary defines terms. Lexigraph takes confusions apart: each page answers in two sentences, compares point by point, offers concrete situations and gives the test that settles it. A frontier page is born from a confusion actually observed, never from an appetite for volume.

Published

19 frontier pages

01AgentorAssistant3 questions · v2026-08An assistant waits for your message, answers, then stops: you are the one driving the exchange. An agent receives a goal and chains the steps itself until it reaches it, deciding at each turn what it does next.
02Fine-tuningorRAG3 questions · v2026-08Fine-tuning modifies the model itself, by continuing its training on your examples: it changes the way it answers. RAG does not touch the model: it fetches the right documents and places them in the context at the moment of the question.
03ContextorMemory3 questions · v2026-08The context is what the model has in front of it while it answers, and it disappears afterwards. Memory is what someone decided to keep between two exchanges, then to place back in the context when it is useful.
04LLMorAgent3 questions · v2026-08A language model is a part: you call it, it produces text, and there it stops. An agent is the system that calls that part in a loop, carries out what it proposes and calls it again with the result, until the goal is reached.
05WorkfloworAgent3 questions · v2026-08In a workflow, the sequence of steps is written in advance by a person: a model may fill one box, it does not choose the next box. In an agent, the plan does not exist before the run, because it is the model that observes the state of the work and decides, turn after turn, what it does next.
06AIorGenerative AI3 questions · v2026-08Artificial intelligence is the name of an entire field, one that holds systems which classify, predict or sort without producing anything. Generative AI is a recent branch of it, the one that makes new content: text, image or sound.
07ModelorProduct3 questions · v2026-08The model is the frozen part that predicts, and nobody uses it directly. The product is what you actually use: the same model, plus instructions, a memory, documents, tools and an interface, all of them chosen by someone.
08ChatbotorAgent3 questions · v2026-08A chatbot is an interface: a conversation window, which can be plugged into anything at all, rules written by hand included. An agent is an architecture: a system that pursues a goal in a loop, and that needs no conversation window in order to exist.
09CopilotorAgent3 questions · v2026-08A copilot assists a person while they work: they keep their hands on the controls, and every suggestion goes through them. An agent pursues a goal in their place: they set the aim and come back to a result, without approving each step.
10Open sourceorOpen weight3 questions · v2026-08Open source is a licensing regime: it grants the right to use, study, modify and redistribute, with no discrimination as to use. Open weight states only a fact: the parameters can be downloaded, under conditions that still have to be read.
11RAGorWeb search3 questions · v2026-08RAG queries a corpus you chose, split and indexed, and that you answer for. Web search queries a public index where you control neither what exists, nor what comes back, nor what changed since yesterday.
12PromptorSystem prompt3 questions · v2026-08The prompt is what you write at each exchange, and it changes every time. The system prompt is the standing instruction placed ahead of everything by whoever built the service, and you never see it.
13TokenorWord3 questions · v2026-08The token is the unit the model handles and the one you pay for: a frequent fragment of text, which does not follow grammar. The word is a unit of meaning, and it has no existence for the model.
14ParametersorTraining data3 questions · v2026-08Parameters are what is left of a model once training is over: numerical coefficients. Training data is what served to set them, and then disappeared from the model.
15HallucinationorError3 questions · v2026-08An error is a breakdown: something did not work as intended, and it can be fixed. A hallucination is the model working normally: it produces a plausible continuation, and nothing in its mechanics tells the true apart from the merely credible.
16AutomationorAugmentation3 questions · v2026-08Automation takes the task away from the person: the system does it from beginning to end. Augmentation equips the person: they go on deciding, but faster and on material that arrives already prepared.
17CloudorOn-premise3 questions · v2026-08Cloud runs the system on someone else’s machines, billed by usage. On-premise runs it on machines you own and operate, and the choice turns on how the data circulates and which models you can reach, rarely on price.
18Machine learningorDeep learning3 questions · v2026-08Machine learning is the set of methods that derive their behaviour from examples rather than from written rules. Deep learning is one family within it, the one that stacks many layers of computation, and generative AI is a narrower subset still.
19Context windoworContext3 questions · v2026-08The context window is a capacity of the model: the largest amount of text it can have in front of it at once, measured in tokens. The context is what the harness actually placed inside it for one request, and that is a choice.

In preparation

The corpus programme