Steve Wozniak’s re-reading of the initials. His distinction is between actual intelligence and its symptoms, and the symptoms are what a chat produces perfectly: it talks, it sounds sharp, it holds an argument, while underneath it predicts the next word from statistics. Keeping the acronym and swapping the claim turns out to be the cheapest way of saying what the technology is not.
Also called an agentic system: a group of two or more AI agents that share one goal and divide the work between them. The group takes on complex objectives that span many different tasks. People set the limits: what the agents must achieve, which tools each one may use and which actions need a person's approval. Inside those limits, the agents choose their own actions and the order of them. They break the objective into tasks, and each task goes to the agent suited to it. Results pass from agent to agent until the objective is met. The work continues over long periods when the objective requires it.
Artificial General Intelligence, or strong AI. A machine that could match a person across the board rather than in one domain: pick up an unfamiliar field, carry what it learned into a different one, handle a problem nobody prepared it for. It has not been built. The human brain is complex enough that no one can say with confidence how far off it is, or whether the current approach gets there at all.
An AI agent is a program that carries out a task by itself, within a frame people set in advance. Setting up that frame comes first. The person who sets up the agent connects it to the software it may use. That person can be a developer writing code, or a user adding connections in a settings screen. They also write the instructions and limits the agent must follow: which actions it may take, and which cases must stop it and go to someone else, such as a payment above an agreed amount. The user then gives it a goal. From there the agent takes over. It uses a large language model (LLM), the same engine behind AI chat, to decide which steps to take and which tools to use. It carries those steps out itself: it can send an email, update a customer record, edit a web page or search a database. It checks the results, corrects its course and continues until the task is done. When it meets one of those set cases, or fails more times than the limit people set, it stops and hands the work to a person or to another agent set up for that case. Because the agent acts in live systems itself, the access and limits people set decide how useful it is.
A fixed set of steps for getting from data to an answer. Feed the same numbers in twice and the same result comes out both times, because the procedure does not vary. What varies is what you feed it.
An AI chat is a program in which a user writes or speaks a question or an instruction in ordinary words and a large language model (LLM) produces the reply. Common examples are ChatGPT, Claude and Gemini. Chat is one of several ways to use AI. The model behind it can also work inside other software through a programming interface (API), or run as an AI agent that carries out tasks by itself. A conversation in an AI chat runs in turns: the user writes or speaks, the chat answers, and each new reply takes the earlier messages into account. The model builds its replies from patterns it learned in training. Many chats can also search the web or read files the user uploads, and answer from those. The reply can be text, code or speech, and in some chats also an image, music or video made by separate models built for those formats.
What comes out of training. Take an algorithm, run it over a body of data, and what you are left with is a set of numbers, millions of them, that encode the patterns found in that data. That file is the model. It has no logic you could read, only weights. Swap the data and the same algorithm produces a different model.
The AI-written answer that sits above the results on a Google search page. It reads the sources itself, extracts what it needs, and hands back one synthesised answer with a few citations underneath, so most people stop there. Getting cited in that block is now what visibility means, and it is a different job from ranking. There is a fuse built in: on high-risk questions, if no source clears the trust bar, no overview is generated at all.
The working thing a client actually uses. A model on its own does nothing: it needs somewhere to run, something feeding it inputs, rules about what happens to its outputs, a store of knowledge to check against, and a person with authority to approve or reject what it produces. The system is all of that assembled. When automation goes wrong it is rarely the model that failed. It is the wiring around it.
Artificial Narrow Intelligence, also called weak AI. A system built for one task and trained on data from that task alone. Inside those limits it often beats people, because it never tires, never loses focus and handles volumes no team could. Outside them it has nothing. It does not know the task is a task, and it cannot carry what it learned into a neighbouring problem the way a person would after a week on the job.
Application Programming Interface: a connection that a program, an online service or a device offers so that other software can use its data and functions directly. Through this connection, the connecting software sends a request, and the program, service or device carries it out and returns the result. This lets separate systems work together automatically, while each keeps its inner workings closed. Access usually requires an API key, a code tied to an account and checked with every request.
Artificial Super Intelligence. A self-improving system that exceeds human capability in most or all fields, which puts it beyond anything we can usefully picture. Probably centuries out. The ethical argument about how far to push without losing control has already started anyway.
The one letter in E-E-A-T you cannot award yourself. Expertise is what you know. Authoritativeness is whether your field treats you as the place to check, and engines read it from the company you keep.
Datasets too large, too fast-moving or too messy for ordinary tools to handle. The shorthand is volume, velocity, variety: more records than a database can sensibly hold, arriving faster than anyone can process them, in formats that do not line up. The term dates from the era when storage got cheap enough that companies started keeping everything without deciding what it was for, and it named the problem that created. Its answer was to stop moving data to the analysis and move the analysis to the data instead, spread across many machines at once. That shift is what made training on internet-scale text possible, so today’s models are downstream of it.
Your name appearing near sources the engine already trusts, .gov and .edu especially. Enough of that and you get classed as authoritative yourself. Note what changed: an unlinked mention now beats a backlink.
Kevin Kelly’s word, from The Inevitable in 2016, built on the electrification template: take anything that already works and add intelligence to it. Cognified music, cognified laundry, cognified real estate.
Groups of individuals acting collectively in ways that seem intelligent. Malone’s definition, and the standing one in the field.
A program that visits your pages and reads them so a search engine or AI system knows they exist. Googlebot is the familiar one.
Raw values with no meaning attached. Numbers, text, images, signals, rows in a table. A reading of 47 is data. It tells you nothing until you know 47 of what, measured when, against what baseline. Every AI system starts here, both in training and in what it retrieves at runtime, and the quality of everything downstream is set at this layer.
Analysis in layers, each one feeding the next. The first layer finds something crude, the second finds patterns in what the first found, and so on up the stack, until relationships appear that no one could have written a rule for. Nothing in there was told what to look for. The depth is the whole trick, and it is all “deep” means.
E-E-A-T stands for Experience, Expertise, Authoritativeness, Trustworthiness. It is the scorecard an AI engine uses to decide whether a source is worth quoting. Four things get measured: how densely you pack verifiable facts, how often your name appears alongside government and academic sources, whether you exist as a confirmed organisation in the registers the engines check, and how hard you are selling. Promotional language counts against you. E-E-A-T outranks the technical work: clean formatting will not lift a weak source into an answer, and a trusted source gets quoted without it.
Search by breeding. The system starts with a population of random, mostly bad solutions, scores each one, discards the worst, combines the best, and adds small mutations to stop the process settling on a dead end. Repeat for enough generations and a strong answer surfaces.
A case the automation stops and hands to a person.
A way of computing with degrees rather than absolutes. Classical logic makes everything true or false, which breaks the moment the information is ambiguous, and most real information is. Fuzzy logic allows fractional values, so a reading can be largely one thing and partly another, and the messy middle stays computable instead of being rounded away.
High Performance Analytics. The discipline that was already wringing value out of runaway data volumes before anyone called it AI: data warehouses in the nineties, data mining, then big data platforms, then the precision models retail and e-commerce run on now. Its defining instinct is worth borrowing. Value tends to appear when you connect sources nobody had thought to treat as related.
Data with context, structure and meaning added, so a person or a system can use it. 47 returns this week, against a weekly average of 30. Now it says something. For an AI engine this is the unit that gets retrieved, weighed and quoted, and its value depends on more than being correct: how current it is, how complete, how densely packed with facts, and whether it connects to anything else.
What your source adds that the pool does not already contain. Repeat what is known and you earn nothing, because the engine already has it a thousand times over. Original statistics, your own expert commentary, a chart with a description a machine can read: that is what registers. In practice, optimising for it means packing your source with facts nobody else has.
An idea in use. Schumpeter drew the line first, separating the invention from its introduction, and the Oslo Manual made it a criterion: it counts once it has been brought into use, not when it was thought of or built. A prototype is not an innovation. A patent is not an innovation. The thing running in an operation is. New to whom is the second question, and the answer is narrower than the word suggests. In the measured definition it means new to the unit adopting it, not new to the world. An operator on legacy infrastructure who changes how the work runs has innovated, whatever the sector did ten years earlier.
Plain RAG fetches documents. Integrated RAG sits inside a working pipeline: the engine first rewrites the question into the terms the knowledge base actually uses, matches on meaning rather than wording, and only then retrieves. That translation step is the whole point, because people ask in language that drifts while the underlying facts sit under fixed terminology. Once the passages come back, position inside the generated answer is decided by fact density. The passage carrying the most verifiable specifics gets quoted first.
Content that only appears after a browser runs the page’s scripts. Crawlers that do not execute them see an empty page, and an empty page is not indexed. Same outcome when the server is too slow to answer and the request times out: the resource is skipped entirely. Not scored lower. Skipped.
Information plus the relationships between it. Not just what the numbers are, but what causes what, which rule applies when, what follows from what. That is what makes it usable for reasoning rather than lookup.
Language as the humanities mean it: shared meaning between people who both know what they are talking about. Worth separating from the “language” in Large Language Model, which is the computational sense, text as a statistical sequence of tokens. Same word, different claim. One promises comprehension, the other counts symbols.
Large Language Model. Take the words apart and the mystique goes with them.
The older idea of reason: understanding, judgment, the grasp of why something follows from something else. It is the standard a machine is being measured against when anyone asks whether AI thinks. It does not meet it, and the word is useful mainly for naming what is absent.
The part of AI where the software works out the rules itself. Instead of a person writing the logic, the system is shown examples and finds the pattern that fits them, then applies it to cases it has never seen. Everything downstream is a variation on that: which examples you supply, whether anyone has labelled them, and how the system is scored when it gets one wrong.
Machines replacing muscle as the power behind work.
The structure most modern AI is built on, loosely modelled on how brain cells connect. It is made of artificial neurons: tiny functions that take numbers in, weigh them, and pass a result to the next one along. Arranged in layers, millions of them, what the system knows is held in the strength of the connections between them rather than in any rule someone wrote.
Making a process better without changing what it is.
What the model learned during training, frozen into its parameters the day training stopped. No system can run on this alone. Language keeps mutating, and stable questions develop answers that expire within the hour.
Work that repeats, in a fixed order, across more than one person or system. A task is one step. A process is everything between the steps, which is also where it goes wrong: the handoff, the wait, the second person opening the same file to check what the first one entered.
RAG stands for Retrieval-Augmented Generation. Instead of answering from memory, the engine goes and fetches: it queries a search index or a document store, pulls the passages that match, and writes the answer from those. That is what makes an AI system auditable, because every statement traces back to a document.
Learning by doing rather than from a dataset. There is nothing collected in advance. The system acts, gets scored on the result, and adjusts to earn more over time.
Structured data. Machine-readable code sitting invisibly in your page that states outright what the content means: this is the company, this is the price, this is the author, this is the date. Without it the engine has to infer all of that from your prose, which it will sometimes get wrong. With it, you hand over a finished conceptual map, ready to cite. As AI search takes over more of the traffic, the gap between pages that do this and pages that do not will widen.
The sources an engine starts from rather than searches for. Government sites, universities, industry bodies. They enter with the highest baseline trust an algorithm can assign, which buys them an asymmetric advantage: when a company page and a trade association page disagree, the model takes the association. Every time. The technical optimisation everyone else has to do barely moves the needle for them.
Language will not sit still. People invent new ways to ask the same thing, and the same question quietly changes what counts as a correct answer. Ask what the best laptop is and the words have not moved in a decade, but the answer expires every few months. A model trained once cannot keep up with either half of that, which is why live retrieval exists.
A measure of how disordered information is. Practical version: the more possible messages, the harder it is to tell which one matters. Data volume now doubles on a timescale measured in minutes, and every new phrase people invent lands on top of the noise already there. That is the problem an AI engine solves for a living. Finding the one relevant source in a pool that grows faster than anyone can read it.
Service Level Agreement. The contract that says what “we keep it running” actually means, in numbers you can hold someone to. Uptime, how fast we respond when something breaks, how long we have to fix it, what support you get and when.
Learning from examples someone has already labelled. You supply a few thousand invoices with the total marked, or emails tagged spam and not spam, and the system works out the rule that separates them.
Support Vector Machine. A supervised method that sorts things into categories by drawing the boundary between them, positioned as far as possible from the nearest cases on either side. Those nearest cases are the support vectors, and they alone define the line. Everything further away is irrelevant to it.
The same problem, solved without a central search. Hundreds of simple agents move through the space of possible answers at once, each remembering its own best find and signalling it to the rest, and the group converges on the promising region. Bird flocks and ant colonies are the model, and the behaviour is genuinely borrowed rather than a metaphor.
The model reaching outside itself for a live value: an exchange rate, a timetable, a stock level. It queries a named provider and returns what that provider says. Worth knowing who decides which tools get called. Not the model. The engineers who wired up the application, in server-side code.
The architecture under every AI chat in use today. Before it, models read text one word at a time and lost the thread over long passages. The transformer reads everything at once and works out, for each word, which other words in the text matter to it. That is what “attention” means. Published by Google researchers in 2017, and the reason ChatGPT, Claude and Gemini exist.
Time to First Byte. How long a server takes to start answering. Two or three seconds is tolerated if the content is worth waiting for; only a timeout kills you.
A process that runs reliably and is written down nowhere.
Learning with nothing labelled. You hand over raw data and the system finds the groupings on its own.
A computational knowledge engine, launched in 2009 and built on Mathematica, the system Stephen Wolfram had been developing since the eighties. It was never a search engine. Ask it a question and it did not hand back a list of pages that might contain the answer. It computed the answer, from curated data and encoded mathematical rules, and showed the working.
Your Money or Your Life. Money, health, law, safety. Ask about these and the engine drops into a stricter mode: institutional sources win by default, and thin evidence produces no answer rather than a guess.
Nothing under that letter yet.