The glossary
The 88 published entries, sorted by floor: each one gives a one-sentence answer, then says what the term is not. Every floor opens on the terms you enter it by; the rest unfold.
88 entries published
- artificial intelligenceNo. 045Artificial intelligence is the name of an entire field, not of a technique: everything that makes a machine perform tasks once associated with human judgement. Like the word “medicine”, which covers a vaccine as much as a hip replacement.
- context windowNo. 007The context window is the amount of text a model can have in front of it at once: your messages and the documents supplied. Like a worktop: past a certain size, something has to come off before anything else goes on.
- fine-tuningNo. 014Fine-tuning means continuing the training of an existing model on your own examples, in order to change the way it answers: like an experienced professional you spend a few weeks training in your in-house practices, until they become reflexes.
- generative AINo. 017Generative AI refers to the systems that produce new content, text, image, sound or code, rather than classifying or measuring what already exists. Like a draughtsman composing a scene never seen before, where a stamp always reproduces the same imprint.
- hallucinationNo. 011A hallucination is a false answer stated with the same assurance as a true one: an invented date, a quotation that does not exist. Like someone who, rather than admit a gap in their memory, fills it with whatever sounds right.
- large language modelNo. 005A large language model is a machine for predicting what comes next in a text: from what precedes, it proposes the next fragment, then starts again. Like someone who had read everything and would finish your sentences, without checking what they claim.
- modelNo. 004A model is the frozen result of long learning on masses of examples: it keeps from them a way of continuing what it is given. Like an engine out of the factory: it does not start on its own and never changes once built.
- tokenNo. 006A token is the small piece of text that the model handles: not a letter, not quite a word, rather a chunk of a common word. What you write reaches it cut up in this way, like a text read syllable by syllable.
See the other 21 terms
- alignmentNo. 060Alignment is the work that makes a model behave as one wishes: following an instruction, refusing certain requests, adopting a tone. Like the induction that follows a hire, which does not change what a person knows but what they do with it.
- benchmarkNo. 061A benchmark is a standardised test that scores models on a common set of questions, in order to compare them with one another. Like a national examination: it ranks the candidates, it does not say which one will do the job in your team.
- biasNo. 062A bias is a systematic and reproducible deviation in a model’s outputs, inherited from its data or its training. Like a badly calibrated scale: it does not get things wrong at random, it always gets them wrong in the same direction.
- deep learningNo. 047Deep learning is the branch of machine learning that stacks up many layers of computation, each one describing the data a notch more abstractly than the one before. Like a chain of reviewers where the first sees strokes, the next shapes, the last a face.
- determinismNo. 089A system is deterministic when the same input always gives exactly the same output. A language model is not: like an orchestra playing the same score again, the request is identical and the performance never quite is.
- distillationNo. 063Distillation consists in training a small model to imitate the answers of a large one, in order to obtain a far lighter object that keeps the essentials of the behaviour. Like a workshop manual drawn from the experience of a master craftsman: shorter, and sufficient for most cases.
- embeddingNo. 053An embedding is the translation of a text into a long list of numbers, arranged so that two texts of similar meaning give neighbouring lists. Like coordinates on a map, where nearby towns talk about nearby things.
- frontier modelNo. 087A frontier model is one of the most capable models in existence at a given moment: the category is defined in relation to what the field can do, never by a size or a figure. Like a world record: it is not a performance fixed once and for all, it is the one nobody has beaten yet.
- inferenceNo. 051Inference is the moment when an already trained model produces an answer: it reads what it is given and writes the continuation, without learning anything along the way. Like a musician playing a piece, as opposed to the years spent working on the instrument.
- machine learningNo. 046Machine learning consists in no longer writing the rules of a program but giving it examples so that it derives them itself. Like learning to recognise an edible mushroom by seeing a thousand of them, rather than by reading their description.
- multimodalNo. 058A multimodal model handles several kinds of input in the same computation: text, images, sound, without first converting them into a single format. Like a person looking at a chart while listening to the commentary on it, rather than reading its written description.
- neural networkNo. 048A neural network is a long sequence of multiplications and additions, organised in layers, whose coefficients are adjusted until the outputs match the examples. Like a mixing desk with billions of faders, set automatically by ear.
- open sourceNo. 056Open source names a licensing regime that grants four freedoms: to use, to study, to modify and to redistribute, with no discrimination as to use or as to person. Like a recipe published with its list of ingredients, which anyone can cook, change and publish again.
- open weightNo. 055An open weight model is a model whose parameters can be downloaded: you can run it on your own machines, adapt it, integrate it. Like an engine delivered to you that you fit wherever you want, without the plans of the factory that built it.
- parameter (weight)No. 049A parameter is one of the numerical coefficients adjusted during training: together they make up everything the model has retained. Like an instrument being tuned: the settings remain, the tunes played to tune it are nowhere inside.
- pre-trainingNo. 052Pre-training is the first phase, the longest and the most expensive: the model reads immense quantities of text, learning only to predict what comes next. Like years of reading that build a general culture, before anyone teaches you to answer a question politely.
- reasoning modelNo. 059A reasoning model first produces, for itself, a long series of intermediate steps, then writes its answer from them. Like someone covering a sheet with rough calculations before writing the line with the result.
- temperatureNo. 054Temperature sets the degree of chance at the moment the model picks the next fragment: low, it almost always takes the most probable one; high, it allows itself less expected choices. Like a cook who follows the recipe to the letter, or departs from it a little at every move.
- tokenisationNo. 088Tokenisation is the operation that rewrites any incoming text as a sequence of fragments taken from a fixed vocabulary, the only one the model can read. Like a printer’s type case: whatever word is to be set, it has to be set with the sorts that are in the case, and a rare word takes more pieces than a common one.
- training dataNo. 050Training data is the set of examples from which a model was formed: what it can say and what it is ignorant of both depend on their content. Like the library where someone learned everything, and whose gaps become their own.
- transformerNo. 057The transformer is the network architecture that equips almost every current model: at each step, each fragment of text looks at all the others and keeps those that shed light on it. Like a reader who, at every word, goes back to the ones that fix its meaning.
Floor 2 · The Harnessthe block and its bolted-on plates: what gets added to it
Floor 2 · The Harness
22 entries- assistantNo. 002An assistant is a product that waits for your message, answers you, then stops. Like a service desk: you ask a question, you are given an answer, and nothing happens as long as you ask for nothing. You lead the exchange, turn by turn.
- chatbotNo. 015A chatbot is a program you exchange messages with, as at a counter: you ask a question, it answers, and everything happens in the thread. It is a way of presenting a service, not a particular technology.
- contextNo. 008The context is all the text the model has in front of it at the moment it answers: your question, the history, the documents you added. Like the file placed in front of someone before an interview: whatever is not in it does not exist for them.
- harnessNo. 003The harness is everything built around a model to make it useful: instructions, documents, tools, guardrails. Like a car around an engine: without a body, seats or steering wheel the mechanics still turn, but nobody goes anywhere.
- memoryNo. 009Memory is what a product decides to keep between your exchanges, then to give the model to read when it is useful. Like a notebook kept by a colleague: they remember nothing, but they read their notes again before each meeting.
- promptNo. 010The prompt is the text you address to the model: your request, with what is needed to answer it. Like an instruction left for someone who arrives knowing neither you nor the file: the more precise it is, the less they improvise.
- RAGNo. 013RAG means searching for the documents useful to your question, then giving them to the model to read before it answers: like a colleague who, before answering you, goes and pulls out the right binder and rereads it in front of you.
- toolNo. 012A tool is an action the model can request from a piece of software: search the web, read a file, send a message. Like a diner in a restaurant: they order, they do not go into the kitchen, and the dish is brought to them.
See the other 14 terms
- APINo. 040An API is a doorway meant for software: one program asks, another answers, in an agreed format. Like a vending machine: you explain nothing to a shop assistant, you put in what it expects and you receive what it returns.
- chunkingNo. 065Chunking consists in cutting your documents into fragments before indexing them, since it is a fragment, and not a whole document, that the search will return. Like cutting a manual up into index cards: the cut decides what will be found again, and what will remain incomprehensible.
- few-shot promptingNo. 081Few-shot prompting means putting two or three examples of what you expect, input and output, into the request itself rather than describing it. Like handing someone three completed forms instead of explaining how to fill the form in.
- guardrailsNo. 041Guardrails are the limits placed around an AI system so that its errors stay harmless: what it cannot trigger on its own, what a person has to approve. Like the railing on a balcony: it does not stop you leaning out, it stops the fall.
- MCPNo. 039MCP is a standard socket between the software that drives a model and the tools or documents it has to reach. Like a standardised electrical socket: every appliance plugs in without a bespoke cable, and changing one does not mean redoing the wiring.
- moderationNo. 044Moderation is the sorting of the content that enters and leaves an AI system: what you refuse to process, what you refuse to let through. Like the check at the door of a hall: you look at what crosses the threshold, in both directions.
- prompt cachingNo. 066Prompt caching keeps the work already done on the beginning of an input, so that a request starting with the same text does not have to do it again. Like a machine already brought up to temperature: the warm-up is not started again for every batch.
- prompt engineeringNo. 067Prompt engineering is the work of formulating what you ask the model, with its examples and its output format, then checking that the wording holds on real cases. Like briefing a capable temporary worker who knows neither the organisation nor the file.
- prompt injectionNo. 096Prompt injection means slipping into a text the system reads an instruction that takes precedence over the designer’s own, because the model receives both as a single stretch of text and cannot tell which one carries authority. Like an unsigned note slipped into a file handed to an employee: they read it along with the rest, and nothing on the paper says it did not come from management.
- structured outputNo. 043A structured output is a response constrained to follow a precise shape, so that another piece of software can read it directly. Like a form rather than a free letter: the boxes are imposed, which says nothing about the accuracy of what you write in them.
- system promptNo. 037The system prompt is the permanent instructions a product places ahead of every request: the tone, the role, the prohibitions. Like the rules posted at the entrance of a swimming pool: they hold for everyone, before you have said a word.
- tool callingNo. 038Tool calling is the moment when the model requests an action instead of writing: finding a file, sending a message. Like a customer filling in an order form: they write down what they want, someone else fetches it and brings it back to them.
- vector databaseNo. 042A vector database stores texts by their meaning rather than by their words, so as to retrieve those that resemble your question. Like a shop aisle where neighbouring items serve the same purpose: you find what you need without knowing the exact name.
- web searchNo. 064Web search is the tool that lets a model go and fetch pages online and read them before answering, instead of sticking to what it has learned. Like a librarian who goes out to consult the day’s newspapers rather than quoting from memory.
Floor 3 · The Agentthe block caught in a loop: it starts again until it gets there
Floor 3 · The Agent
19 entries- agentNo. 001An agent is a system that puts a model to work in a loop to reach a goal: it plans, acts, sees the result and starts again, like a person following a recipe, tasting and adjusting at each step.
- agentic loopNo. 020The agentic loop is the cycle an agent repeats: it looks at where it stands, chooses an action, carries it out, takes note of the result, then starts again. Like someone looking for an address: they walk on, look up, correct their direction.
- autonomyNo. 021Autonomy is the share of decisions that a system takes without asking you. It is not a switch but a slider, set gesture by gesture: like lending your car, where the permitted journey is agreed before the keys are handed over.
- evalsNo. 024An eval is a series of test cases replayed at every change, to see whether the system still answers correctly. Like driving the same test route again after every adjustment: it does not prove that you drive well everywhere, only that you have broken nothing.
- human in the loopNo. 027Human in the loop refers to the points where a system stops to ask for a person’s approval before acting. Like a bank transfer waiting for your confirmation: the machine has prepared everything, it does not take the last step on its own.
- traceNo. 025A trace is the record of what a system did, turn by turn: what it read, what it decided, what it executed. Like the black box of an aircraft: nobody consults it as long as all is well.
- workflowNo. 016A workflow is a series of steps written in advance by a person, then executed automatically: like a recipe followed to the letter, where the order of the moves is fixed before anyone has even entered the kitchen.
See the other 12 terms
- compounding reliabilityNo. 069Compounding reliability is the fact that a multi-step system only succeeds if every step succeeds: the rates multiply instead of adding up. Like a chain whose every link holds ninety-five per cent of the time, and which breaks almost every other time at ten links.
- error recoveryNo. 073Error recovery is what an agent does when an action fails: noticing it, understanding why, and changing its approach rather than repeating. Like a locksmith who, the key not turning, tries a different grip before forcing the same one thirty times.
- goalNo. 068A goal is the outcome an agent has to reach, stated once and held for the whole loop, whatever steps it takes to get there. Like an order placed with a craftsman: you say what you want to end up with, not the sequence of movements.
- LLM as a judgeNo. 105An LLM as a judge is a model tasked with scoring another system’s answers against written criteria, where no mechanical comparison can do the job. Like a marker to whom you hand an essay along with the marking scheme: they are quick and they never tire, but the mark is only worth what the scheme is worth.
- multi-agent systemNo. 071A multi-agent system has several agents work together, each with a role, its own tools and a separate context. Like a team where everyone holds their post, with the coordination overhead that any team implies.
- observabilityNo. 026Observability is the ability to understand what a system is doing from what it records, without having to open it up again. Like a car dashboard: the needles repair nothing, they tell you where to look.
- orchestrationNo. 028Orchestration is the organisation of work between several pieces that do not act at the same time: who calls whom, in what order, and what to do when one of them fails. Like a department head who hands out the case files and picks up the ones that get stuck.
- planningNo. 022Planning is the moment when a system breaks a goal into steps before acting. Like writing a shopping list before going into the shop: the list guarantees nothing, but it saves you from wandering the aisles.
- sandboxNo. 070A sandbox is the closed environment in which an agent carries out its actions: it has access to what has been placed there, and to nothing else. Like a workshop from which everything that must not be touched has been removed, rather than an instruction asking that it not be touched.
- skillsNo. 103Skills are sets of instructions and resources filed away separately, which the system loads into its context only when the task calls for them. Like the ring binders on an office shelf: their spines stay visible at all times, but one is only opened on the day it is needed.
- stopping conditionNo. 072The stopping condition is what puts an end to an agent’s loop: either the goal is reached and verified, or a limit is crossed. Like an oven timer paired with a thermometer: one says that it is cooked, the other that it has gone on long enough.
- subagentNo. 023A subagent is an agent launched by another agent to handle part of the work, then hand back its result. Like sending someone on an errand: you give the instruction, they go, they come back with the bag, not with the journey.
Floor 4 · The Organisationa mosaic: AI is only one tile among many
Floor 4 · The Organisation
18 entries- AI ActNo. 029The AI Act is the European regulation that governs artificial intelligence according to the risk of each use: like the safety rules of a building, stricter for a lift than for a doorbell. What it looks at is the use, not the technology.
- AI governanceNo. 033AI governance is the set of decisions an organisation takes to know who may use what, on which data, and who answers for the result: like delegated signing authority, which is not there to show distrust but to make clear who commits the firm.
- augmentationNo. 018Augmentation means equipping the person who does the work rather than taking the task away from them: like an electric drill, which does not assemble the furniture in your place but lets you assemble more of it, better, and without wearing yourself out.
- automationNo. 034Automation means taking a task out of human hands: the machine carries it out from end to end, like a dishwasher that you fill and empty, but that washes without you. The person decides beforehand and checks afterwards, no longer during.
- copilotNo. 035A copilot is an assistant built into the tool where the work happens, which makes suggestions while the person keeps the controls: like the co-driver on a rally, who calls the bend out loud but never holds the wheel.
- GDPR and AINo. 030The GDPR is the European regulation that protects people’s data, including when it is an AI that processes it: like an address book, where you write down a number for a particular reason, and keep nothing just in case.
- personal dataNo. 031Personal data is any information that makes it possible to trace back to a person, directly or by cross-referencing several elements: it is not only the name on the letterbox, it is also the floor they live on, the timetable and the car, which together designate someone just as surely.
- shadow AINo. 019Shadow AI is the use of AI tools by employees outside any official framework: like a personal spreadsheet that ends up running an entire department, without anyone ever having decided that it should exist.
See the other 10 terms
- acceptable use policyNo. 078An acceptable use policy is the short document that says to whom, on what data and for what tasks AI is permitted within the organisation, and what is forbidden. Like the rules posted at the entrance of a workshop: few lines, known to everyone.
- adoptionNo. 036Adoption is the moment a tool actually enters the work: not when it is installed, but when people use it without thinking about it. Like a gym membership: paying the joining fee is not training.
- change managementNo. 075Change management is the support work that makes a tool genuinely enter into practice: training, rewriting procedures, arbitrating resistance. Like fitting a new kitchen in a restaurant without stopping service.
- cloudNo. 076Cloud refers to running a service on someone else’s machines, billed by usage, with no hardware to buy or to operate. Like taking the train rather than buying a car: you pay for the journey, not for the garage.
- cost per tokenNo. 074The cost per token is the unit price of what the model reads and writes: it turns a use into a bill, and it is paid on every call, indefinitely. Like a water meter rather than a subscription: what costs is not the installation, it is the turning of the tap.
- digital sovereigntyNo. 032Digital sovereignty is the ability to keep control of your data and your tools: like renting premises where the landlord keeps a spare set of keys and can raise the rent, you work there without setting the rules.
- high-risk useNo. 107A high-risk use is a use of an AI system that European law holds liable to weigh heavily on a person’s rights or safety, and that it subjects for that reason to reinforced obligations. Like the same vehicle that does not obey the same rules depending on whether it carries parcels or children: it is the use that commands the regime, not the mechanics.
- intellectual property and AINo. 080Intellectual property applied to AI raises two distinct questions: what you have the right to give a model to read, and what legal standing whatever comes out of it has. Like a copy shop: what you bring in and what you take out fall under different rules.
- on-premiseNo. 077On-premise means running a system on machines the organisation owns and operates, within its own walls or in a data centre it rents. Like having your own boiler: more control, and the maintenance is yours.
- return on investmentNo. 079Return on investment sets what a use brings in against what it costs, everything included: licences, calls, integration, review and support. Like the calculation for a machine tool, where the purchase price is rarely the decisive item.
A missing term is not an oversight: the corpus is written week after week, and an entry is only published once it holds up under both readings. The often confused pages and the paths are the two other ways into the corpus.