The words of AI, explained without jargon

The term covers very different things. Today, when an article says "AI", it almost always means a language model trained on vast amounts of text, able to produce plausible answers.

A model that, at every step, computes the most likely next word. Repeated thousands of times, that small calculation produces coherent text, translations, code. All its power and all its limits come from this.

The instruction given to the model. How it is phrased changes the result a great deal: giving a role, an example, an expected format often improves the answer more than switching tools.

The interface, not the brain. The chatbot is the conversation window; behind it a language model does the work. The same model can power a chatbot, a mobile app or an invisible background service.

A model given tools and the right to use them in a loop until a goal is reached. This is the big push of 2025-2026, and the source of most spectacular announcements — and failures.

The family of programs every current AI model belongs to. A network stacks layers of very simple calculations; learning means tuning the billions of coefficients linking them until the outputs are right. Nobody writes the rules: they emerge from the tuning.

The horizon the big labs set themselves: a system able to carry out most human intellectual tasks, not only the ones it was prepared for. The word serves at once as a scientific goal, a fundraising argument and a source of worry.

The basic unit the model reads and writes. A common word is one token, a rare word several. It is also the billing unit: when a service quotes a price, it prices tokens.

The phase where the model adjusts its parameters by going through its data. It happens once, costs a great deal, and freezes the model's knowledge at a date — its "cutoff". After that it does not learn from talking to you.

The numbers that make up the model. They are counted in billions, and that figure is shorthand for a model's size: "a 70-billion" is said the way one might quote an engine's displacement.

How much text the model can process at once, measured in tokens. Recent models advertise enormous windows — a million tokens, several books' worth. Beyond that you must chunk, summarise, or have the model search documents.

Using the already-trained model. Every answer draws a little electricity and costs a fraction of a cent. Multiplied by hundreds of millions of daily users, this is where most of the energy bill is settled.

A model that accepts several kinds of input — text, image, audio, video — and sometimes produces several. It is what lets you photograph a maths problem, have a scene described, or speak aloud to an assistant.

The architecture invented at Google in 2017 and used by nearly every current model: GPT, Claude, Gemini, Llama. Its "attention" mechanism weighs how much each word matters relative to the others, which lets it keep track of a long context. It is the "T" in GPT.

The translation of a word, a sentence or an image into a long list of numbers, a vector, chosen so that two things close in meaning get close numbers. This is what makes searching "by meaning" rather than by keyword possible, and it is the building block of RAG.

A setting between roughly 0 and 2 that decides whether the model always picks the most likely word or allows itself less expected ones. Low for extracting data or coding, higher for brainstorming. Most consumer chatbots set it for you, often somewhere around 0.7 to 1.

Text, images or dialogues generated by one model and used to train another, because high-quality human data is running short or costs too much to collect. Widely used for small models and rare cases (dialects, dangerous situations); most recent models contain some.

The technique behind almost every image and video generator: Midjourney, Stable Diffusion, DALL·E, Sora. The model starts from random noise and "denoises" it over a few dozen steps, guided by your description. That is why an image takes a few seconds, and a video much longer.

Since late 2024, "reasoning models" (o1, o3, DeepSeek-R1, the "thinking" modes of Claude or Gemini) generate a long chain of intermediate steps before the final answer. They are markedly better at maths, code and logic, but slower and dearer: thinking is paid for in tokens.

The hardware behind the whole story. Training and serving models takes tens of thousands of these chips, sold mostly by Nvidia, whose shortages and prices drive the news as much as the models themselves. When an article talks about "compute", giant data centres or export restrictions to China, it is talking about GPUs.

A false statement delivered with confidence: a source that does not exist, an invented date, a quotation attributed at random. It is the best-known flaw, and the reason any information that matters must be checked.

A systematic tendency to favour certain answers: tying a job to a gender, handling English better than other languages, reflecting the majority viewpoint of its data. Not a one-off error but a steady distortion.

Synthetic content imitating a real person. In a few years the technique moved from lab demos to consumer apps, shifting the problem from "can it be done" to "how do we prove it".

Circumventing a model's safety rules. Methods keep renewing themselves: role-play, rare languages, encoding, instructions hidden in a document the assistant will read.

An attack in which text the assistant reads (a web page, an email, a document, a comment) carries instructions that hijack its task: leaking data, sending a message, ignoring its rules. The more assistants become agents that read and act, the heavier the risk. There is no patch for it.

A mark inserted at generation time so that AI-made content can be identified later: a pattern in an image's pixels, a statistically skewed choice of words in a text, signed metadata (the C2PA standard). The EU AI Act requires it for some content from 2026. It is one answer to deepfakes, not the solution.

The set of techniques that make a model behave according to its designers' intentions: helpful, honest, refusing dangerous requests. The main tool is learning from human feedback (RLHF), where annotators rate answers. It is what separates the raw model from the product you are given, and the subject of most AI-safety debates.

The set of limits placed on a model: refusing to help build a weapon, not giving a firm medical diagnosis, not insulting people. Part of it is learned during alignment, part is added around the model by filters that read the question and the answer. That is why the same model refuses more or less depending on the service using it.

A model whose parameters are published. You can install it, adapt it, run it offline. It is the strategic choice of Meta, Mistral and most Chinese players, against the closed models of OpenAI, Anthropic and Google.

The technical means of plugging an AI into software: a website, an app, a spreadsheet. It explains why so many "AI-powered" tools offer the same thing — many call the same models underneath.

Lightly retraining an existing model on specific data: a trade's vocabulary, a brand's tone, a response format. Far cheaper than full training — hours of compute rather than months.

The technique behind most enterprise assistants: find the relevant passages in a document base, slip them into the prompt, then let the model write. It is what allows citing sources and staying current without retraining.

A set of standardised tests — reasoning, code, general knowledge — whose scores feed every announcement. It is the number you will see quoted whenever a new model ships.

A database built to store embeddings and quickly retrieve the ones nearest to a query. In an enterprise assistant, it is what fetches the right passages before the model answers. Pinecone, Weaviate, Qdrant, or a PostgreSQL extension: the market filled up in two years.

An open standard, published by Anthropic in late 2024 and since adopted by OpenAI, Google and Microsoft, describing how an assistant connects to outside tools and data: your calendar, a database, a business application. An "MCP server" written once works with every compatible assistant.

A way of building models so they are very large without being very slow: the parameters are split into "experts", and a router calls on only a handful of them for each token. It is the architecture behind most recent large models, closed and open alike, and one reason prices keep falling.

Slipping a few examples of what you expect, input and output, into the prompt and letting the model continue the series. No retraining: everything happens in the prompt. It is often the most effective way to get a precise format or a steady tone, long before fine-tuning. "Zero-shot" is the opposite: the instruction alone, with no example.

The delay between sending a request and getting the answer. For an assistant it often matters more than a few points of quality: a voice that replies after three seconds feels broken. It is the selling point of "small and fast" models against the big reasoning models, which can think for a minute.

The step that turns a raw model, which can only continue a text, into an assistant that answers a question. Annotators compare answers; a reward model learns their preferences; the main model is then pushed to produce what that judge likes. It is what made the difference with ChatGPT in 2022, far more than model size.

Training a small "student" model on the answers of a large "teacher" model, to keep most of the quality at a fraction of the cost. That is how most "mini", "flash" or "nano" versions are made, and the models that run on a phone or a laptop.