Nobody wrote the rules. Nobody could. A machine learns by adjusting millions of numbers, over and over, until it stops being wrong — and that one idea explains almost everything else. Five steps, in order, each with something you can touch.
Behind this text: a working systolic array. Each tile is lit by the number it holds right now, and the corner shows the multiply it is doing. Move your cursor and you become its workload — leave it alone and it rewrites its own dataflow.
It is a wall of dials. Nothing more mysterious than that. Each dial is a number called a weight, and turning it changes what comes out the other end. Real models have billions of these. Here are nine. Drag one up or down and watch the picture in the middle change under your hand — and watch the error bar, which measures how far it is from the picture on the right.
That is the whole game: find the dial settings that make the error small. Everything on the rest of this page is a consequence of this one sentence.
Nobody turns the dials by hand. Instead, picture a landscape where low ground means fewer mistakes. Every possible setting of every dial is one spot on that landscape. The machine cannot see the map — it can only feel which way is downhill from where it stands, take a small step that way, and repeat. Millions of times. That is training, and there is no more to it. Move your cursor to push the hills around and watch it change its mind.
Everything in the next section was described on paper in 1943, by a neurophysiologist and a logician who had no computer to try it on. The idea then sat almost still for seventy years, not because it was wrong, but because nobody could afford the arithmetic. This is that wait, drawn to scale. Each gridline is a thousand times the one below it — a log scale — because no honest chart of this fits on a screen otherwise. Hover any point to see what it would cost a human.
Building a model and using one are separate economies. The chart above is the building. What actually put this in your hands is the second curve: the price of getting an answer out of a finished model fell faster than almost any price in industrial history — which is why a thing that cost a national-scale budget to train costs a fraction of a cent to ask.
Every “a thousand times cheaper” figure holds an old, low standard fixed and measures the price per word. Both choices flatter the trend. Newer models think before answering, which means they emit far more words per answer — so the price of a word can fall while the price of an answer rises.
Measured per finished task on hard, current work, the fall is a much soberer 5× to 10× a year, and the sticker price of the very best models has been going up 3× to 18× a year. Cheap and frontier are moving in opposite directions. Both facts are true at once, and anyone selling you only one of them is selling you something.
Both sides below run on the same five inputs at the same instants. On the left, a biological neuron: charge arrives along the dendrites, pools in the cell body, and the moment it crosses a threshold the cell fires an all-or-nothing spike down its axon. On the right, the artificial unit that borrowed the idea: each input times its weight, summed, pushed through an activation function, out as a number. Watch the two traces at the bottom — that is the real difference.
It only multiplies numbers. So before any of the layers below can run, whatever you gave it — a photo, a sentence, a sound — has to be turned into a list of numbers. Watch both conversions happen. Nothing clever occurs here: it is bookkeeping, and it is the same bookkeeping every time.
This is a small model that reads handwriting — not a chatbot. It is here because it is the only kind you can actually watch: a chatbot’s layers do the same job on meaning instead of pixels, where there is nothing to see. Stack the units from step 3 into rows, and each row learns something bigger than the row below. Write a number and the machine chops it into 784 squares of light and dark. The first row hunts for edges. The next finds shapes made of edges. The next, whole digits. Nobody told it to work in stages — that is simply what the dial settings with the lowest error turned out to look like. Every tile you see is a real number from a real network, not an illustration. Pick a digit and step through it slowly.
Your question gets chopped into chunks called tokens and laid out on a desk, next to the rules it was given, your earlier messages, and anything it looked up. Then it reads the whole desk and guesses the next chunk — then does it again, and again, until the answer stops. It is not retrieving an answer. It is guessing the next piece, very well, very fast. The lower lane shows where the dial settings came from in the first place: the text it was trained on, and the shaping that came after. Those dials are frozen while it answers you — it is not learning from your question.
Until recently the honest answer was “a wall of numbers, and nobody knows.” That has changed. Researchers can now trace which parts of a model light up while it answers, the way a hospital scanner traces a brain — and what they found is stranger and more specific than either the hype or the dismissals suggested. Three findings, each of which changes how you should use one.
These are not opinions about AI. They are things somebody opened up a model and measured. Tap each one to see what it looks like.
Inside these models there is a circuit that declines by default. It stays on unless a separate “I know about this” feature fires and switches it off. Hallucination is what happens when that recognition feature fires for something the model does not actually know — the refusal lifts, and the machine answers fluently about nothing.
That is why hallucination is not a bug someone forgot to fix. It is the same mechanism that lets the model answer anything at all, misfiring. And it is why “are you sure?” is a weak check and “show me the source” is a strong one.
A raw guess from memory is a scattergun. Ask a good model a question about something that happened last week and, with nothing but its training to go on, it gets it right about one time in ten. Everything that turns that scattergun into a rifle is scaffolding bolted on around the model — not extra intelligence inside it. Switch the pieces on below and watch the spread close.
This is the single most useful thing a non-engineer can know. When an assistant is accurate, it is usually because someone gave it the right material and a way to check itself. When it is confidently wrong, one of these pieces was missing — or, worse, was present and wrong.
Give a model a document that contains a wrong answer and it will usually repeat the wrong answer — in a controlled test, 66.1% of the time, even when its own training said otherwise. Retrieval does not make a model truthful. It makes the model faithful to whatever it was handed.
So the question to ask about any AI answer is never “is this model good?” It is “what did it read, and can I see it?” An assistant that shows its sources is not being polite. It is the only part you can actually audit.
Four hops, repeated for every single chunk of the answer. Multiply this by a billion people and you get the second half of this page.
Your words are split into tokens the moment they land, and slotted into the batch of requests already in flight.
The weights for each layer stream out of memory into the chip's own local store, one wavefront ahead of the activations that will need them.
Numbers march across a systolic array: values flow left to right, running totals build top to bottom, and every cell does one multiply-and-add per tick, forever.
The scores for every possible next token are turned into a choice on the chip itself, and that chunk is on its way to you before the next layer has finished loading.
Forget scores out of a hundred. Here is a measure anyone can feel: give the machine a job, and ask how long that job would take a person. In March 2023 the best models finished jobs that take a human about four minutes. By April 2026, about seventeen hours — the same test, the same people scoring it. The length of job a model can finish half the time has been doubling roughly every four months.
The line below is not a forecast. Every dot is a model that shipped, plotted on the day it shipped. Notice the scale on the left — each step up is ten times longer, and the line is still close to straight.
Most benchmark charts you see are broken. The most-quoted coding benchmark, SWE-bench Verified, was retired by OpenAI in February 2026 after most of its test cases were found to be flawed. Others are saturated — every good model scores 90-something, so the chart flatlines and tells you nothing.
This measure survives because its units are human: minutes and hours of work. But it measures reliability at half odds on text-shaped work, and nothing else. On a broader test of real economic tasks judged by human experts, the best model still only wins or ties about 48% of the time. Both things are true at once.
There are two races, not one. The first is between the frontier labs, whose best models you can only rent through their own servers. The second is between open-weight models, which anybody can download and run on their own hardware. The gap between them is now the single most important number in this industry, and it is about four months.
Four months is not a technical detail. It decides whether capability stays rentable from three companies or becomes something a hospital, a bank or a government can hold on its own premises. Amber bars below are models you can download today.
A benchmark is a school exam, and these models have read the internet — including, very often, the exam. That is not cheating by anyone in particular; it is what happens when a test becomes famous. So the numbers that survive scrutiny are the ones measured on problems written after the model was trained, scored by people, on tasks nobody has an incentive to leak.
The practical version: a model that is two points higher on a leaderboard is not noticeably better for you. A model that is fifteen points higher probably is. And a model you can run inside your own building may beat a better one you are not allowed to send your data to.
Two numbers that point in opposite directions, and almost everybody has only heard one of them. One short question costs about as much electricity as a 10-watt bulb burning for a minute and a half, and about five drops of water. That is far less than the internet believes. But a billion people are asking, and the buildings that answer now use roughly as much electricity as the whole of Japan.
Your own use is almost a rounding error. The construction is not. Both of those sentences have to sit in your head at the same time, and most arguments about AI and the environment are two people each holding one of them.
0.24 Wh is a median for a short text answer. Ask a model to think before it answers, or hand it a long document, and one answer can use 10 to 100 times more — up to roughly 40 Wh. Quoting the small number on its own is the most common mistake made about this topic, in both directions.
Water “withdrawn” and water “consumed” differ by roughly ten times. Withdrawn water is borrowed and largely returned. The alarming headline figures are almost always withdrawal quoted as if it were consumption.
The famous “a bottle of water per 20 questions” claim came from a draft that its own authors have since revised. The current version of that paper reports far smaller per-response figures.
League tables of company water efficiency are close to meaningless. Cooling with more water uses less electricity and vice versa — the numbers measure different trade-offs, not different virtue.
Because the scarce thing is no longer ideas, and it stopped being chips a while ago. It is electricity, memory and heavy steel equipment. Around $700 billion of company money is scheduled to be spent on this build-out in 2026 — roughly the size of Belgium’s entire economy — and it is arriving faster than the physical world can absorb it.
The money moves at the speed of a press release. A high-voltage transformer takes two and a half to five years to build. That gap is the whole story of the next decade.
Of the single most-publicised AI construction programme — nine or more gigawatts announced, half a trillion dollars attached — roughly 3% is actually built and running: about 0.3 gigawatts. Announced capacity and delivered capacity are different industries with different physics.
So treat every enormous number in this section, including the $700 billion, as an intention. Transformers, turbines, substations and skilled electricians are the real constraint, and none of them can be summoned by a funding round. When you next read that a company will spend $100 billion on AI, the useful question is not whether they have the money. It is whether anyone can deliver the equipment.
You can forget every diagram above and still explain this at dinner.
There is no list of instructions inside saying what a cat looks like or how a sentence should end. There is a wall of numbers that got adjusted until the mistakes got small. That is why nobody can fully explain any single answer it gives — including the people who built it.
It picks the next chunk that best fits everything in front of it. When the training text supported an answer, the guess is excellent. When it did not, the guess is still delivered in the same confident voice. That is exactly why it makes things up — not a bug bolted on, a direct consequence of how it works.
The useful question is never “is the AI smart?” It is “was there good material about this, and can I verify the answer?” Recipes, summaries and first drafts: fine. Your specific medication, a legal deadline, a number in a contract: check it. Every time.
Not in the way you do. It has built an extraordinarily detailed map of how ideas sit next to each other in human writing, and it navigates that map. Whether that counts as understanding is a real argument among serious people. What is not in dispute: there is no inner voice checking whether the sentence is true.
Because making things up and being right are the same operation. It always produces the most plausible next chunk. When the training material supported that chunk, you call it an answer. When it did not, you call it a hallucination. The machine cannot tell the two apart from the inside, which is why it sounds equally confident either way.
Not while it is answering you. The dials are frozen. Your conversation can be fed back later into training a future version, depending on the product and the settings you chose — but the model you are talking to right now does not change as you talk to it.
The honest answer is that the measurement above cuts both ways. Task length is multiplying fast. But in a controlled trial, experienced developers given AI help were 19% slower while believing they were faster. Tasks are being automated at speed; whole jobs are mostly bundles of tasks plus accountability, and accountability does not transfer.
For a normal amount of personal use, no — a year of daily questions is a couple of litres of water and about as much electricity as boiling a few dozen kettles. The serious environmental question is not your usage, it is the pace of data-centre construction and where the grid gets the power. Those are policy questions, not personal ones.
Everything it can see is on that desk. It has no memory of you, no access to your situation, and no way to ask itself what you meant. More context on the desk means a narrower, better-aimed guess. This is also why it does well on a task with an example attached and badly on the same task without one.
Anything where lots of good examples exist and you can check the output cheaply: first drafts, summaries, translation, code you will run anyway, explaining a document you already have. It is weakest where being wrong is expensive and checking is hard — medical, legal and financial specifics about you.
Every figure links to a primary source in the grey box beneath its section: METR for capability timing, the IEA and Lawrence Berkeley National Laboratory for energy, Google’s own published measurement for per-question energy and water, and company filings for spending. Where sources disagree, the range is printed rather than the flattering end of it.
Almost every “learn AI” list is a pile of links with no order and no honesty about effort. This one is ordered, and each entry says what it costs you. Start at the top and stop whenever you have had enough — the first two alone will put you ahead of most people arguing about this online.
Everything on this page runs on hardware somebody had to find, power and cool. MachineLearningLogistics is a control tower for exactly that: which GPUs are actually deployable, in which region, on which date, at what contract, from which provider.