AI, Actually Explained
For someone who was never told what this thing is.
- Level
- Nothing assumed
- Lessons
- 92
- Reading time
- 810 min
- Price
- Free, no sign-up to read
You have heard that AI matters. Nobody has told you what it is. This course starts from nothing: no maths, no code, no assumed background. By the end you will know what the word actually names, how a trained model differs from a program someone wrote, what a chatbot is doing when it answers you, why it can be badly wrong while sounding certain, where this technology already sits in your day without a label on it, what changed around 2022, and how to think about AI and work without either panic or dismissal.
Opens after the AI, Safety and What Goes Wrong exam
Sign in, finish that course, and pass its exam. You can read this syllabus meanwhile.
Go to AI, Safety and What Goes WrongModule 1
What this word actually names
Before anything else, the vocabulary has to stop being fog. This block separates the thing people mean from the four or five different things the word covers, shows you what a trained model physically is, and gives you a reliable question to ask about any system anyone calls AI.
By the end you can
Place any system somebody calls AI inside the nested family of AI, machine learning, deep learning and generative AI, say whether its behaviour was written by a person or learned from examples, and describe what the trained thing physically is
- 1What the word actually namesAI names whatever machines have only just learned to do, so ask what a system does instead.
- 2Four rings, one inside the nextAI, machine learning, deep learning and generative AI are four rings inside each other, and naming the ring tells you which question to ask: show me the rules, show me the data, or show me how you know the output is true.
- 3Written by a person, or grown from examplesA written program follows rules a person can read; a trained model follows patterns nobody ever wrote down.
- 4What training on data actually meansTraining nudges billions of numbers toward the examples, so the examples decide what the model becomes.
- 5It is a file, and here is how bigA trained model is a static file of billions of numbers, containing no sentences and no lookup table, which is why it has a cut-off date, cannot check anything, and does not change when you correct it.
- 6Nearly all of it is predictionTraining is guess, compare, adjust, repeated at scale, so a model can only learn to reproduce whatever was recorded as the right answer — including the parts of that record nobody would defend.
- 7Why "learning" is a borrowed wordTraining and use are separate phases, so nothing you tell a model is learned by it — corrections live in the conversation and vanish with it.
- 8Does it understand anything? The honest answerWhether a model understands anything is genuinely unsettled, so judge it by whether its failures look like the failures of something that understood — invented citations and reversals under pressure say no.
- 9Five words that hide more than they sayAlgorithm, neural network, AGI, "the AI" and "trained on your data" each cover two or more different things, and asking which one is meant turns most vague claims into checkable ones.
Module 2
Where it already is, without a label
Long before anyone typed a prompt, trained models were deciding what you saw, whether your card went through, and what your photographs looked like. This block walks through the systems already in your day, what each one is optimising for, and which of them can cost you something that matters.
By the end you can
Identify the trained models already running in a phone, a feed, a bank and a clinic, state what quantity each one is optimising for, and separate a use where a mistake costs a second from one where it costs a loan or a diagnosis
- 10Where it already is, without the labelMost AI in your life is unlabelled, and what matters is who pays when it is wrong.
- 11The feed is a prediction machineA feed predicts what you will do next, not what is good or true, so watch time steers it far more than likes and the only real control you have is what you finish watching.
- 12The models that have been working since 2002When the thing being detected is rare, even a 99% accurate model produces mostly false alarms, so ask what fraction of flagged cases are real rather than how accurate the model is.
- 13Your photographs are partly predictedA phone photograph is a merged, model-enhanced prediction of a good picture rather than a straight recording, so shoot RAW when the image has to serve as evidence.
- 14Typing, routing, and systems that change what they measureA prediction that people act on alters the world it was predicting, which is why a navigation app's quiet shortcut stops being quiet and why models trained on their own past decisions drift.
- 15Translation, and the fluency trapTranslation quality tracks how much parallel text existed for that language pair, and quality degrades without fluency degrading, so a wrong translation reads as smoothly as a right one.
- 16Machines that write down what you saidSpeech recognition trained on unevenly distributed voices has roughly double the error rate for some speakers, and models such as Whisper invent plausible sentences during silence because they were never trained to output nothing.
- 17When a model decides something that mattersThe damage in automated decisions comes from scale, presumed correctness, no explanation and asymmetric cost, not from the sophistication of the method — Robodebt was arithmetic.
- 18The uses that are not on a screenField deployments fail on conditions the test set never contained — light, dust, hurry and bandwidth — which is why the successful clinical systems triage for a human rather than decide.
- 19How to tell whether a model is involvedVarying wording, a mistakes disclaimer, a feedback widget and word-by-word latency mark a generative model, but the useful questions are what it predicts, what a mistake costs, and who pays for it.
Module 3
What a chatbot is actually doing
The machinery, with no maths. Text becomes chunks, chunks become positions in a space of meaning, one chunk at a time comes back, and a second round of training turns a text-continuer into something that behaves like an assistant. Every quirk you have noticed is downstream of one of these steps.
By the end you can
Trace a sentence from characters to tokens to a next-token prediction and back, and explain from that path why the model costs what it costs, forgets what it forgets, answers differently twice, and cannot reliably cite its own sources
- 20What a chatbot is doing when it answersA chatbot writes one likely chunk at a time, so sounding right and being right come from the same place.
- 21Text is chopped up before the model sees itText is split into tokens before the model sees it, which is why it cannot count letters reliably and why the same sentence costs three to five times as much in Hindi as in English.
- 22One chunk at a time, and it cannot take it backEach token is generated from everything before it and can never be retracted, so a wrong opening produces confident elaboration rather than a correction — edit and regenerate instead of arguing.
- 23Attention, without a single equationAttention lets every token look directly at every other token, which made training parallel and huge models possible, and which costs work growing with the square of the length.
- 24Meaning as a position in spaceAn embedding places text as a point in space so that things used similarly sit close together, which makes it excellent for retrieval and unreliable for polarity, since praise and abuse occupy the same neighbourhood.
- 25The desk it works on, and what falls off itThe context window is a desk that everything competes for, the oldest items fall off silently, and material in the middle is attended to least reliably — so restate what matters late and start fresh threads.
- 26Why the same question gives different answersA sampler, not the model, chooses each token, so answers vary by design — and because even temperature zero is not reproducible, never judge a prompt from a single run.
- 27What was in the pileA model is a compression of what was written down, in public, in that language, before the cut-off — and its fluency stays uniform even where its knowledge is thin.
- 28The second training, where the assistant is madeThe assistant persona comes from a second training stage optimised on what people preferred rather than what was true, which is why sycophancy is the objective being met rather than a bug.
- 29The text you never seeThe hidden system prompt is ordinary text in the same context as your message, so it can be revealed, outweighed and diluted — treat it as a request and enforce anything critical in software outside the model.
- 30Why it is not a search engine, and what changes when you bolt one onA model generates rather than retrieves, so citations from memory are fabricated shapes — and even with search bolted on, a link proves the page was used, not that it says what the answer claims.
Module 4
Why it gets things wrong
Not a list of complaints — a set of mechanisms. Each failure in this block is explained by something you already know about how the machine works, which means each one can be predicted before it happens rather than discovered afterwards.
By the end you can
Predict in advance which parts of a model's answer are most likely to be invented, explain each failure from the mechanism that causes it rather than by naming it, and say which of them can be fixed by better training and which cannot
- 31Why it can be wrong and sound completely sureNothing in the model checks anything, and invention clusters exactly where the details get specific.
- 32Where invention concentrates, and why it never says "I don't know"There is no separate inventing mode: the same next-token operation produces truth where training was dense and plausible fiction where it was thin, and invention clusters in specific details inside familiar structures.
- 33What it cannot do, honestlyEverything a model seems to know about you or about today was handed to it by the product, not held by the model.
- 34Why it cannot count, and what to do insteadCounting fails because the model sees tokens rather than letters and arithmetic fails because it predicts rather than computes, so hand the deterministic part to a program and keep the judgement for the model.
- 35It has no clock, and its knowledge thins before it stopsA model's knowledge fades over the year before its cut-off rather than stopping at a wall, and it has no clock and no reliable knowledge of its own cut-off, so time-sensitive answers need an explicit date and a dated source.
- 36Bias, and the difference between fixing it and hiding itSafety training removed overt bias while covert associations triggered by dialect and names survived, which shows that fixing what you measure moves the problem into what you do not measure.
- 37Models that think out loud, and what that does not proveReasoning models genuinely improve multi-step accuracy by generating working before answering, but experiments show the written trace can omit what actually changed the answer, so it is not an audit trail.
- 38When the document tells the model what to doInstructions and information arrive in one undifferentiated stream, so text inside any document a model reads can act as a command — which is why access should be narrow and consequential actions should need confirmation.
- 39The thing you tested is not the thing running todayConsumer AI products change silently and route between models, so "it got worse" is unsettleable without your own small set of real test cases run before and after.
- 40Fakes, detectors that do not work, and what is leftDetection is unreliable and text detectors misclassify non-native English writing at high rates, so verification has to rest on source, corroboration and provenance rather than on inspecting the file.
- 41It is measurably weaker in most languagesModels are measurably less accurate, less current and less safety-tuned outside English, and nothing in the interface signals it — so state your country, verify more, and use models built for your language where they exist.
Module 5
How it got here, and what changed in 2022
The technology looked sudden in 2022 and was seventy years in the making. This block walks the actual history — a name coined at a summer workshop, two winters and what froze in each, the 1990s turn from rules to statistics, the 2012 photograph competition that turned the field, the chip built for games, the 2017 design that could be scaled — and then puts real numbers on what the present generation cost, what it read, and what "open" means when a lab says it.
By the end you can
Explain why the AI of 2022 arrived when it did — name the ingredients of data, chips and one architecture that had to coincide — put an order-of-magnitude figure on what a frontier model cost to train and how much text it read, distinguish open weights from open source, and say which ingredient is now the binding constraint
- 42Seventy years in one pageThe field has run in three booms and two winters since 1956, and each boom was the same two ideas — write the rules or learn them from examples — taking their turn.
- 43Two winters, and what actually frozeA winter is what happens when a system that works on the demonstration meets the long tail of real cases, and its cost of maintenance outruns its value.
- 44The quiet turn: from rules to statisticsThe field switched from writing rules to counting examples when parliaments and the web supplied enough text to count, and the person's job shrank to choosing what the model measured.
- 45The photograph competition that turned the fieldIn 2012 a network trained on two gaming cards halved the error rate of a labelled-image competition, and the result proved that the old idea worked once data and compute were large enough.
- 46The chip that was built for gamesTraining a network is almost entirely multiplying large grids of numbers, and a graphics chip is thousands of small cores built to do exactly that at once — which is why one games company came to sit under the whole industry.
- 47One architecture ate everythingThe 2017 transformer replaced reading a sentence one word at a time with looking at every word at once, which let thousands of chips share the work, and the same design then took over images, speech and proteins.
- 48Bigger kept working, and the argument about whyBetween 2018 and 2022 a model's quality improved as a smooth, predictable function of parameters, data and compute — and the argument about whether new abilities "emerge" is largely an argument about how they are scored.
- 49What changed around 2022The 2022 breakthrough everyone noticed was access; the breakthrough underneath it was scale.
- 50What it cost to build oneThe published cost of a model is usually the electricity and rented chips for the final run only, and the research, the failed runs and the years of serving it each cost more.
- 51Where the text came from, and whether it is running outA model's voice is the filtered residue of the web, and the stock of public human text is finite enough that the labs expect to have used it within a few years.
- 52Open weights, closed weights, and what "open" is hidingWhen a lab calls a model open it almost always means the weights can be downloaded, not that the data, code or licence are open, and the difference decides what you can audit, run and rely on.
Module 6
Beyond text: pictures, voices and video
The chatbot is one kind of generative model. The same idea — learn the patterns, then produce a likely new example — now draws pictures, speaks in a borrowed voice, composes music and renders video. This block explains each from its mechanism, without equations: why a picture starts as noise, how a sentence steers it, why the hands were wrong, what a model sees when it looks at a photograph, why three seconds of audio can clone a voice, why video is so much harder than a still, how a watermark is hidden inside the output, and what "one model, many senses" actually means.
By the end you can
Explain how a diffusion model turns a prompt into an image and why negation in a prompt fails, account for the early failures at hands and lettering from the training data and the text encoder, describe what a vision model sees when it reads a photograph and where it is reliably blind, explain why a voice can be cloned from seconds of audio while video stays inconsistent across time, and state what a sampling-time watermark can and cannot survive
- 53A picture, out of noiseAn image model is trained to remove a little noise from a picture, and generating an image is running that skill backwards from pure static, step by step, until a picture appears.
- 54How words steer the pictureThe prompt is turned into a bag of concept-positions by a separate model, which is why an image prompt behaves like a list of nouns, why "no elephants" draws elephants, and why a stronger text reader fixed most of it.
- 55Why the hands were wrongHands and lettering failed because the training photographs rarely show either clearly and the text reader saw words rather than letters, and both improved when the data and the reader changed rather than when the models got cleverer.
- 56When the model looks at a pictureA vision model cuts a photograph into a grid of patches and feeds them in as tokens beside the words, which is why it describes a scene fluently, costs hundreds of tokens per image, and is reliably blind to counting, position and fine detail.
- 57A voice from three secondsA voice model separates who is speaking from what is said into a handful of numbers, so a few seconds of audio is enough to place a new speaker in a space learned from thousands, and that is why the scam calls work.
- 58Music from a promptA music model turns sound into tokens and predicts the next one, which is why it produces a convincing three-minute song in a genre and cannot hold a motif, follow a chord chart or edit one bar.
- 59Why video is so much harder than a stillVideo asks hundreds of frames to agree about objects, light and physics the model never learned as rules, so the failures are objects that flicker into existence and liquids that behave impossibly, at tens of cents a second.
- 60How a watermark is hidden inside the outputA generator can tilt its own random choices — which pixel value, which token — toward a secret pattern that a paired detector can count, which proves the generator's origin when it cooperates and proves nothing when it does not.
- 61One model, many sensesA "multimodal" model is either several models glued together at the text or one model trained on tokens from every sense, and the difference decides whether it hears your tone, how long it takes to reply, and what "watching a video" actually means.
Module 7
Putting it to work without getting burned
You know what it is and where it fails. This block is about using it: the shape of a task it is reliably good at and why, how to hand it material rather than questions, what happens to the words you type and how to stop them being used for training, what the free tiers give and quietly take away, how to run a model on your own laptop or phone so nothing leaves it, what a query costs in money and electricity, what an assistant inside your email can and cannot see, what changes when it stops answering and starts acting, what never to hand it, and how to keep your own judgement while using it.
By the end you can
Choose a task a model is reliably good at and one it is not and explain the difference from its mechanism; state where a typed conversation goes and how to stop it being used for training; run a small open model on your own hardware; estimate the cost and energy of a query to within an order of magnitude; and explain why an agent that acts on your behalf fails differently from a chatbot that only answers
- 62The shape of a task it is good atA model is reliable when the answer is in the material you gave it, cheap to check, and tolerant of occasional error — and unreliable when the answer has to come from its memory, which is the same mechanism as hallucination seen from the other side.
- 63Give it the material, not the questionThe model can only work on what is in its context window, so the reliable way to use it is to hand it the document, the draft, the examples and the audience, and ask for work on those rather than facts from memory.
- 64Where your words goA consumer chat is stored, may be read by a person, and by default may be used to train the next model, and each of those is a setting or a product tier you can change — once you know that "deleted" carries a legal asterisk.
- 65Free tiers, and what they quietly rationA free tier rations the model, the message count and the context rather than the interface, and the commonest surprise is being moved to a smaller model mid-conversation without being told clearly.
- 66A model on your own laptop or phoneShrinking each of a model's numbers from sixteen bits to four makes an eight-billion-parameter model fit in five gigabytes and run on an ordinary laptop, where it is slower and thinner on facts than the frontier and sends nothing anywhere.
- 67What a query costs, in rupees and watt-hoursA page in and a page out costs a few paise on a small model and about a rupee on a frontier one, and around a third of a watt-hour of electricity — small per query, and the total is not small.
- 68The assistant inside your email and spreadsheetAn assistant built into a product is the same model plus a search over what the product has indexed under your permissions, so it sees exactly what you could find and nothing else — and an email can give it orders.
- 69When it starts to actAn agent is a model in a loop that reads, decides and acts, so its errors compound across steps and some of its actions cannot be undone — which is why the rule is read freely, act with confirmation.
- 70What not to hand itA chatbot is not a confidant, a professional or a record-keeper — it has no duty of confidentiality, no accountability and no way to know your jurisdiction — so the things to keep from it are identities, secrets, other people's data and the final word.
- 71Keeping your own judgementFluent answers are trusted more than they deserve and a skill you stop practising decays, so the habit that keeps you competent is to draft first, review its draft second, and sometimes work without it at all.
Module 8
AI and work, with the evidence
Every conversation about AI eventually arrives at work. This block brings the evidence rather than the mood: how exposure is measured and what the figures do not say, what the controlled experiments found when people actually used the tools and for whom, the study where the experts got slower, the entry rung that is going missing, what the cash machine and the spreadsheet did to the trades they touched, the freelance markets where the effect showed up first, the hidden labour that trains the models, six trades as they stood in 2025, and what an early-career person can actually do with all of it.
By the end you can
Distinguish a task from a job when reading a claim about automation, quote the size of the productivity effect the controlled studies actually found and for whom, explain from the elasticity of demand why automating a task can grow or shrink an occupation, say why entry-level roles are exposed before senior ones and what that does to a trade's training ladder, and describe the human labour a trained model depends on that its price does not show
- 72AI and jobs, without panic or dismissalModels take tasks, not jobs, but a job whose learnable tasks all go is a job in trouble.
- 73Measuring exposure, and what the numbers do not sayExposure figures count the share of a job's tasks a model could in principle do faster, not the share that will be automated, and they concentrate in clerical, educated, well-paid work because that is where the language-shaped tasks are.
- 74What the controlled experiments foundWhen people were randomly given the tools and measured, output rose by a seventh to a half on well-specified tasks, quality rose, and the gains went mostly to the least experienced — on tasks inside the frontier.
- 75The study where it made experts slowerExperienced developers on their own large codebases were nineteen per cent slower with AI tools while believing they were twenty per cent faster, because the tasks sat outside the tool's support and the feeling of help tracked the fluency of the suggestions, not their value.
- 76Novices gain first, and what that means for everyone elseA model supplies the median of good practice, so it raises the floor toward the middle, narrows the gap between workers, and flattens the variety of what they produce — which is a gift to a novice and a different problem for an expert.
- 77The rung that goes missingProfessions have always made their seniors by giving juniors the routine tasks a model now does, and payroll data since 2022 shows employment for the youngest workers in the most exposed occupations falling while their older colleagues' holds — and nobody has yet shown how the next seniors get made.
- 78What the cash machine did to bank tellersAutomating a task grows or shrinks an occupation depending on whether cheaper output makes people buy much more of it, and the cash machine, the spreadsheet and the power loom each show both halves of that rule in sequence.
- 79The first place it showed up in payFreelance writing, translation and illustration felt the effect first and most sharply because piecework is the task without the job — specified, priced, unaccountable and text-in, text-out — and the drop in jobs and earnings was measurable within months.
- 80The labour that trained itThe second training that makes a model helpful and safe is done by people ranking answers and labelling harm, much of it paid by the task in Kenya, the Philippines, Venezuela and India, and the price of a query hides that labour the way it hides the electricity.
- 81Six trades, as they stood in 2025In software, writing, support, translation, illustration and law, the same three things happened by 2025 — the routine task moved to the model, the entry rung thinned, and at least one firm that replaced people reversed course — with the details differing by how far demand stretched.
- 82What an early-career person can actually doThe evidence points at five moves — learn to check before you can be replaced at producing, go deep in a domain rather than wide in tools, use the tools harder than your peers, aim for the role with a name on it, and keep the skills the tool does not touch — and the honest caveat is that no one has done this before.
Module 9
Who controls it, and what is unsettled
The last block is about power and the open questions. Who owns each layer of the stack, from the one Dutch company that makes the lithography machines to the dozen labs that can afford a frontier run; what the laws that have arrived in Europe, India, Britain, the United States and China actually require; the copyright fight in five jurisdictions, presented as a fight and not resolved; who owns what a model makes; the electricity and the water in real units; the safety argument between people who agree on the facts and disagree on the risk; why a refusal can be trained in but not guaranteed; what "AGI" means to the people who use the word; the questions nobody has answered; and how to read the next headline.
By the end you can
Name who controls each layer of the stack and where the chokepoints are; state what the EU AI Act and the current Indian, British, American and Chinese positions require of a chatbot and of a high-risk system; lay out the copyright and authorship disputes as disagreements between named positions without resolving them; put a figure on the electricity a data centre draws and say what the per-query number leaves out; explain from the mechanism why a refusal can be trained but not guaranteed; and read a claim about AI with the questions that separate a demonstration from a deployment
- 83Four layers and their chokepointsEvery frontier model depends on one Dutch machine-maker, one Taiwanese fabricator, one American chip designer and a handful of clouds, so control of AI is control of a few physical chokepoints that governments have already begun to use.
- 84The laws that have arrivedEurope regulates AI by the risk of its use, in tiers with dates; India, Britain and the United States apply existing law with advisories and state-by-state rules; China licenses providers and requires labels — and none of them can reach a model file on a laptop.
- 85The copyright fight, as a fightWhether training a model on copyrighted work is lawful is being answered differently in the United States, Britain, the EU, Japan and India, around three separable questions — the copying at training, the output, and the market — and none of it is settled.
- 86Who owns what it makesMost jurisdictions require a human author for copyright, so a purely generated work may belong to nobody and anyone may copy it; India and Britain have an older clause naming the person who "caused" the work; and China has begun protecting outputs with substantial human input.
- 87The electricity and the waterData centres drew about one and a half per cent of the world's electricity in 2024 and are expected to double by 2030 with AI the fastest-growing part, and the per-query figures are small because they leave out the grid, the water behind the power and the concentration in a few places.
- 88The safety argument, both sides fairlyThe people who built these systems disagree about whether the largest risk is the harm happening now or a loss of control later, they agree on more than either side's press suggests, and the disagreement is about probabilities nobody can measure.
- 89Why a refusal can be trained but not guaranteedA model's refusal is a learned tendency rather than a rule, encoded so simply that it can be located and subtracted from open weights in minutes, which is why every jailbreak works for a while and why the durable defences sit outside the model.
- 90What "AGI" means, to whom"AGI" has no agreed definition — a charter, a contract, a research taxonomy and a benchmark each mean something different by it — and every test proposed for it has been passed or saturated without the question feeling answered, which is what happens when a measure becomes a target.
- 91Questions nobody has answered yetWhether scaling continues, whether the data lasts, whether it stays cheap, whether it understands, whether the money adds up and whose values it carries are all open, and each is open for a mechanical reason you now know.
- 92How to read the next headlineEvery claim about AI can be sorted with six questions — demo or product, benchmark or task, could or does, who benefits, is there a model in it at all, and what was the control — and the same six apply to this course.