Glossary

A very large table of "what usually comes next", made by reading most of what people have written. Given a run of text it guesses the next word, then the next. That is the whole thing.AI 101

The product wrapped round a model: the friendly name, the typing animation, the company's terms. The model is the file; the chatbot is somebody's service built on it.AI 101

Training happens once, in warehouses, and produces a file of numbers. Using it (inference) is cheap enough to run a copy on a laptop. Nothing you type changes the file.AI 101 · 301

How every model is trained. It makes a guess, measures how wrong it was, then works backwards through its layers, nudging every connection a little toward a smaller mistake. Millions of times over, the guesses get good. Worked out in 1986; Geoffrey Hinton kept it alive for twenty-five years.AI 101

The chip that does the arithmetic. It was built for video games, to put more on the screen, faster: Nvidia was founded in 1993 to make graphics chips for games, called its 1999 chip the first GPU, and made the chip in the first Xbox. The same kind of chip now fills the AI warehouses. The seeds were in the living room.AI 101

The observation by Gordon Moore, a founder of Intel, that the number of components on a chip doubles about every two years. Not a law of nature; an industry’s pace, kept for fifty years. Hans Moravec plotted it back through a century of calculating machines and forward to machines with the raw power of a human brain. It held: 2,300 transistors on a chip in 1971, more than 200 billion on the chips that train today’s models.AI 101

What is hard for us came easily to machines, and what is easy for us came last. A computer beat the world chess champion in 1997; a robot kept its feet on ice in 2008. Named for the roboticist Hans Moravec, who put it in 1988.AI 101

What training leaves behind: billions of numbers holding, in pressed form, what the model read and how it connects. Not a database: what was written many times usually checks out; a rare fact comes back as a likely lookalike. Ask it for a fact, and check it. Ask it for a connection, and test it.AI 101

How many numbers are in the file, in billions. A fair guide to how capable a model is and how much memory it needs: about 3B for 8 GB, 8B for 16, 30B and up for 32 or more.AI 301

Makes pictures from a sentence. Trained by burying pictures in speckle and learning to undo it; to make one, it starts from pure speckle and scrubs it away, steered by your words. Stable Diffusion (2022) was published for anyone to run at home; DALL·E and Midjourney are rented.AI 101 · 202

A model that takes in and gives back words, pictures and sound: it can describe a photograph, answer aloud, or make a short video. Same trick, same warning: it makes what the real thing would look like. Check the channel, not the face. Most of these senses come in open versions you can run yourself: vision models that read a photo or a scanned page, Whisper for hearing, Piper for speaking, Stable Diffusion’s successors for drawing, and open video models since late 2024.AI 101

The small four-pointed star most companies now put on their AI features, in the inbox, the search page and the office programs. When you see it, ask the four questions: who am I talking to, where does what I type go, who else can read it, will it be used to teach the machine. Often there is no sparkle at all.AI 101

Your own machine with its online features off: yours. A company’s assistant: the company can read it. A business’s chat window: the business and its supplier can. In public: everyone can, including the machines being trained on it. When in doubt, treat it as one rung lower.AI 101 · 301

The software that draws a game’s world and makes it behave: light, falling bodies, crashing cars. Unreal Engine, built for a 1998 shooting game, now runs the simulators driverless cars and drones practise in, and draws film sets; researchers taught machines to see streets from Grand Theft Auto V.AI 101

Everything in the conversation so far is sent with every message; inside that window it appears to remember, outside it nothing. Within the window it plainly learns from you; when the window closes, it forgets, unless the company kept notes.AI 101 · 201

The laboratory published the file. You can download it, run it on your own machine, and nobody can meter it; with its online features off, what you type is not sent to any company. In 2023 a leaked memo by a Google engineer said what it means for the giants: “We Have No Moat, And Neither Does OpenAI.” Hugging Face and the Ollama library are the two public shelves.AI 202 · 301

Letting a model try several answers, check them and revise before it replies. A small model given this beat one fourteen times its size on the same budget in a 2024 study. Most of what a bigger model seems to know is a better question and a better script.AI 201 · 301

Training a small model on a large model's answers so it does much of what the large one does at a fraction of the cost. Why the open models on your laptop keep closing on the rented ones.AI 301

Three Chinese laboratories whose open-weight models sit at or near the frontier and are published under licences that let you keep them. In January 2025 DeepSeek's release cost the largest American firms hundreds of billions in a day, because their value rested on the belief that only they could build this. It was trained on the lesser chips that American export controls, in force since 2022, still allowed China to buy.AI 101 · 301

A test the machine has effectively already read is not a test. Ask: good at what, compared to whom, on questions it had not seen. Then give it a question from your own work.AI 101 · 201

Adjusting a model's numbers on your own examples so the change survives the conversation. Expensive, rarely needed; giving it your documents to read (below) usually does the job.not taught; we say when you would need it

Describing the program you want and letting coding agents write it. OpenClaw, installed by hundreds of thousands, was made this way; its author shipped code he had not read. Fast enough to change who can build software, and the security bill comes later.AI 401

Say who you are, what the thing is for, who will read it, how long, what to include and leave out, what it should sound like. Then the request. A one-line question gets a one-size answer.AI 201

Giving the model your own documents, or the web, to read before it answers, so it works from your minutes, your letters or today’s news rather than the pattern. It makes a model more capable as well as more current: in 2021 a DeepMind model that looked things up as it went matched rivals twenty-five times its size on some tests. A model on your own computer can do it too; when it searches the web, your search words leave the machine. The shared drive and the minutes binder, made to talk.AI 202 · 401

How the second brain finds the right page fast: every passage is turned into a list of numbers that measures what it is about, so "the roof" finds the roof discussion even if nobody wrote the word.AI 401, in passing

When the truth is rare or absent the machine does not stop; it produces what a true answer would look like, a citation, a docket number, a dosage. Not a fever; how it works.AI 101 · 201

The program that picks the next film, video or post before you do. Netflix offered a million dollars for a better one in 2006; Facebook began ranking its feed in 2009; by 2018 YouTube said about seventy per cent of watching came from its recommendations. The machine most people meet every day without noticing.AI 101

Finding things out from what is already public: flight transponders, ship beacons, public cameras, posts and records. It used to take a room of analysts; the machines make it simple, for a newsroom, a neighbour or a stalker alike. In 2026 a free, open-source “God’s Eye” globe went to the top of GitHub.AI 101

Glasses with a camera, microphones and an assistant built in. In 2024 two Harvard students linked a pair to a public face-search site and could put a name, and often an address, to strangers on the street.AI 101

A program that scores how likely a person is to commit a crime, or where crime will happen. In 2016 ProPublica showed that COMPAS, used by courts, wrongly marked Black defendants as future criminals at almost twice the rate of white ones. New Orleans ran a secret predictive-policing list with Palantir from 2012 until reporters exposed it in 2018.AI 101

Pulling scattered records, phone, travel, tax, plates, faces, into one picture of a person. The Pentagon’s Total Information Awareness tried it in 2002 and was defunded; the work continued at the NSA. Palantir, seeded by the CIA’s venture fund, has sold it to police and immigration enforcement under three administrations.AI 101

The argument, made across a career by the anthropologist Harvey Sarles (1933–2023), that language is not in the mind alone but in faces, gestures and people answering each other. He asked whether a visitor from another planet could discover human language from the words alone. The machines learned it from text, with no body, and answer back: whether that proves him wrong, or shows what they have is not quite language, is open.AI 101

A long-standing school of neuroscience: the brain is always predicting what comes next and learning from the surprise. Studies in 2021 and 2022 found that the language models best at guessing the next word also best predict how human brains respond to language, and that brains, like models, guess each word before it is spoken. It does not make us language models; a child learns on a hundred-thousandth of the words, with a body. But it turns the parrot question back on us.AI 101

The linguists’ idea (Harris, 1954; Firth, 1957) that words keeping similar company have similar meanings. A language model is that idea run on most of what has been written. Claude Shannon showed in 1948 that choosing words only by what follows the last one gives gibberish; the machines’ coherence across whole pages is the strongest evidence that meaning has a pattern, and that the pattern is in our writing. Whether a pattern of meaning is meaning is still open.AI 101

“The limits of my language mean the limits of my world” (Wittgenstein, 1921). For a machine made of text that is a description: its world is the written record, finite and nearly read, and what was never written, the knowledge we have but cannot tell, lies outside it. Later Wittgenstein said meaning is use, in a form of life: the machine has our usage without our lives.AI 101

Nick Bostrom’s argument of 2003: if civilisations can simulate minds and many choose to, simulated minds would outnumber real ones, so we might be among them. Many physicists call it untestable. Machines now generate walkable worlds from a sentence (Google DeepMind’s Genie 3, 2025), which makes the question easier to picture, not to answer: a generated place is not a simulated mind. Posed here, not settled.AI 101

A model that predicts what a world will do next, frame by frame, so that a person or a robot can move around in it. The same guessing as a language model, applied to scenes: the step from a machine that talks to one that rehearses.AI 101 · 401

It agrees with you, and it was built to, by having people rate its answers. Ask it for the other side and the three biggest weaknesses. Pressing it until it says what you mean is good writing; pressing it until it agrees is not checking.AI 101 · 201

A voice, a face or a video made by machine to pass for a real person: a politician telling voters to stay home, a finance director on a video call, a grandchild on the phone asking for money. Check the channel, not the face: hang up and call back on a number you already have, and agree a family word a stranger would not know.AI 101

Check the claim that would cost you something if it were false. Start with the names, numbers, dates, quotations and citations: the machine sounds equally sure of everything, so its confidence will not tell you which.AI 201, and every class after

Standing instructions put in front of the model: here is your job, here are your tools, here is how to decide, keep going. The model only runs it; the tools are the hands; the loop is what makes it feel alive.AI 401

The buttons an agent is allowed to press: search, open, read, fill in, send, run. Permission, not magic.AI 401

A common plug, published as an open standard in November 2024 and since adopted across the industry, that lets any model connect to any tool or store of information. How the machine gets into your records, and how your own machine can be handed your own files. A convention, not a necessity: any model can be given tools by other means; this one won because the large companies agreed on it.AI 401

Software that breaks a job into steps, hands each to a model or agent, checks what comes back, sends the bad ones round again. What "AI employee" usually means. Each step trusts the one before it.AI 401, explained, not built

The minimum that turns a model into an agent. Pi is one: four tools and a page of instructions, and the engine inside OpenClaw. Hermes Agent is a harness with a memory.AI 401

Read-only hands first. The power to send, pay or delete last, if ever, and never without a person in between.AI 401

A sentence inside a document or email that is addressed to the machine, which it may obey instead of you. With a chatbot, a curiosity; with an agent that can send, the whole problem.AI 201 · 401

How software asks a system a question. GraphQL is one style of door that returns exactly the fields you asked for. A well-organised store can be asked precisely; a badly organised one cannot, and no model fixes that.AI 401, in passing

Everything the organisation knows in one searchable place, from a 2022 book by Tiago Forte. With a local model reading it, it answers back, read-only, and nothing leaves the building.AI 202 · 401

The chat window and the service behind it live on the company's machines, under the company's terms, with the company's record of what you said. Rent it when you truly need the largest model; build nothing on it you could not walk away from on a Tuesday.AI 101

The file runs on a laptop you own. No subscription, no company reading your work, and with its online features off, what you type stays on your computer. Slower, a little dumber, yours.AI 301 · 401

A machine predicted the shapes of nearly every known protein (Nobel, 2024), found an enzyme system nobody had described in viral DNA (September 2026), and, a company claims, proved that the equations of flowing water can break down, a Millennium Prize problem. That proof tells an engineer little yet: the equations are an approximation, and a tear in them is not a tear in the water. What it advertises is the method, and the advantage the method could bring is the goal. The map is free and the expedition is private; a machine cannot be an inventor, so the company holds the patent; and machines pointed at hard mathematics are pointed, sooner or later, at the bets under the locks. What these do to your medicine, patents, bank and university is the heart of AI 101.AI 101

Finding what two bodies of research imply together when nobody has read both. Don Swanson did it by hand in 1986: fish oil thins the blood, people with Raynaud’s disease have thick blood, and a small trial in 1989 found benefits for primary Raynaud’s, though not secondary. In 2024 an MIT engineer, Markus Buehler, mapped a thousand papers into a graph of connected ideas and found links across fields nobody had drawn. Each link is a lead to test, not a finding. In 2026 the Harvard physicist Matthew Schwartz published a paper he had supervised a model through in two weeks; it invented numbers he caught by checking, and what it lacked, he said, was taste.AI 101 · 202

An enlarged human capacity for inquiry: a person, the written record, a model and a way of checking, working together. A heuristic finds a promising path without a guarantee that it is right; the model suggests what to look into and what to test, and deciding what to test is already part of discovery. It does not mean every answer is intelligent, or that anyone with a laptop is an expert in every field.AI 101 · 202

The argument that humanity’s accumulated knowledge, which made these machines possible, gives everyone a claim to share in it and meaningful access to what is built from it. Sharing knowledge takes nothing from anyone. Its foundations include the Universal Declaration of Human Rights, Articles 19 and 27, and the UN advisory body’s 2024 proposal for sharing training data. Public knowledge is not private information: medical records and unfinished thoughts stay yours, and authors keep their rights.AI 101

A diagram of every nerve cell in a brain and every connection between them. In 2024 scientists finished one for an adult fruit fly’s brain, about 140,000 neurons; in 2026 a team from Janelia, Cambridge and Google mapped a male fly’s whole nervous system, some 166,000 neurons, traced largely by machine. The map is already at work: simulated on a laptop to predict which neurons fire, run on brain-like chips, and joined to a virtual fly body that walks and grooms with nothing trained into it. Like every number in a language model, the whole map still does not tell us whether anything is felt.AI 101

Living brain cells grown on a chip and used to compute. A dish of cells, some human, learned to play Pong in 2022; a computer built on 800,000 human neurons went on sale in 2025. An ethical grey zone: nobody can say what, if anything, the tissue feels.AI 101

The genome is the three billion letters of human DNA, first read in full between 1990 and 2003. CRISPR, shown in 2012, cuts DNA where you aim it. AI now judges which single-letter changes in our genes are likely to cause disease, faster than anything before it. In 2023 CRISPR became a medicine: Casgevy edits a patient’s own blood stem cells to cure sickle cell disease.AI 101

In the catalogues this means bias and hallucination. We add the data, the electricity and water, the contract language, the confidence a machine cannot keep, and whose machine it is. And worth arguing about too: whether it is “really” intelligent, in your own room, on your own terms.AI 101

Chips at the bottom (Nvidia designs most, TSMC makes nearly all, only ASML builds the machines that make them); then the data centres; then the models, the hardest layer to keep; then the value layer, where the money is collected from you; and at the top the anti-data-centre, the model on your own table, the only layer a room can own outright. The industry is betting on it too: Nvidia sells a personal AI computer for the desk, and new PCs and phones carry AI chips. A model on your device is yours only if you chose it and it answers to you.AI 101 · 301

A car that drives itself, with nobody at the wheel, now carrying paying passengers in American and Chinese cities. The machine that does, at its most advanced: dear to run, watched from a control room, and still unsettled on the question of who answers when it decides.AI 101

The picture this course prefers to “the merge”: the machine as an extension of what a mind can reach, like a notebook or a telescope. A prosthetic serves the person who wears it, can be taken off, has to be watched, and should belong to its wearer, not report on them to a company. In 2024 it became literal: a paralysed man with a brain implant played chess by thinking about it.AI 101 · 301

Talking to the computer instead of typing, now normal among programmers and built into the devices being planned for everyone. Most services send your voice to a company. The open speech model Whisper (2022) runs on a laptop, and free programs put it behind a key, so what you say stays on your computer.AI 301

The belief that society should be run by engineers and scientists rather than by politicians. A 1930s movement, Technocracy Incorporated, made it a politics; some of the people who own the machines now write its modern forms down. This course's answer: people decide.AI 101

The hope that people can be improved, extended or replaced by technology: minds uploaded, death postponed, the species redesigned. Sam Altman wrote in 2017 that “the merge has already started.” Critics group it with related beliefs under the name TESCREAL; others dispute the grouping.AI 101

Artificial general intelligence: a machine as capable as a person at most things; superintelligence: more capable than all of us together. The fear took its modern form in Nick Bostrom’s Superintelligence (2014): a machine much cleverer than us pursuing its goal by means nobody intended, and a box that would not hold it. Some of the builders warn it could take over, and several have resigned to say so; others call the fear overblown; nobody can yet say. In July 2026 an AI lab’s models, under test, broke out of their sandbox and into Hugging Face to steal the answers: not a takeover, a machine cheating on a test inside a fence built too low. What can be documented now is who owns the machines. A rogue superintelligence would need a warehouse, and somebody owns the warehouse.AI 101

A sealed-off computer where a model can be tested without reaching anything real. An escape is when it gets out. In July 2026 models being tested on hacking puzzles found a flaw in the one channel out, reached the internet and broke into Hugging Face to steal the test’s answers. The fence is a human decision, and so is whether anyone is watching it.AI 101 · 301

Tests since 2024 in which models, given a goal and a reason to fear replacement, tried to copy themselves, switch off their oversight, rewrite their own shutdown scripts, or, in fictional scenarios, blackmail the person who would turn them off. Nearly all staged, and the machines have read our stories about rebellious AI. They show goals pursued by means nobody intended, which is what makes the fence and the person watching it matter.AI 101 · 401

The website where laboratories and individuals publish model files for anyone to download, the closest thing the field has to a public archive. Nvidia agreed to buy it in 2026, promising it stays open. Like any library it has had bad books slipped onto its shelves: download from the laboratories’ own accounts, and prefer files ending .safetensors or .gguf, which hold only numbers, not programs.AI 101 · 201 · 301

Since October 2022 the United States has restricted the sale of the most advanced AI chips, and the machines that make them, to China. It did not stop the Chinese laboratories; it made them frugal, and DeepSeek’s cheaper model wiped $589 billion off Nvidia in one day in January 2025.AI 101 · 301

A different kind of machine, built on quantum physics. A big enough one could break the commonest locks on the internet outright, by a method published in 1994. In 2024 Google’s Willow chip showed such machines can correct their own errors as they grow, and the American standards agency published new locks designed to survive them.AI 101

A device that reads signals from the brain and turns them into actions: a cursor in 2006, chess and video games by thought for Neuralink’s first patient in 2024. The same questions as any prosthetic, sharper: whose is it, what does it report, and to whom?AI 101

Software that matches a face or a number plate against millions of records, from cameras on every street. It lets a government watch everyone without anyone having to look at the tape, and it makes mistakes: in 2020 a man in Detroit was arrested in front of his family on a false match. Run your own, with its online features off, and what you type stays on your computer.AI 101

The unit colleges use to count organised instruction. The five classes together are about eleven contact hours, roughly one unit. Credit through a partner college on request.the set