Interactive guide

How AI actually works

Most explanations of artificial intelligence are either marketing or mathematics. This one is neither. Fire a real neuron with your own hand. Watch a language model choose its next word and see the odds it is working from. Train a network and watch it stop being wrong. Then look at what the evidence actually says about work, and what it does not say.

Nothing here is a video of a simulation. It is the simulation. Roughly a 20 minute read
Part one

Start with the thing that already works

Every conversation about artificial intelligence borrows a word from biology. Neural network. Neuron. Learning. Those words came from somewhere real, and the real thing is sitting inside your skull right now, reading this sentence. It is worth understanding first, because once you know what a neuron actually does, the machine version stops being mysterious and starts being simple. Almost disappointingly simple.

86 billion Neurons in an adult human brain. The familiar figure of 100 billion was never measured. This one was, by dissolving whole brains and counting the cell nuclei. Azevedo et al, 2009
~100 trillion Connections between them. Each neuron reaches thousands of others, and the pattern of those connections is where nearly everything you know is stored. Herculano-Houzel, PNAS
516 kcal What your brain costs to run for a day. That is roughly 20 watts, about the same as a dim light bulb, for the most capable general intelligence we know of. Herculano-Houzel, PNAS
2% and 20% The brain is 2 percent of your body mass and takes 20 percent of your energy budget. Thinking is expensive, biologically speaking. It is just far cheaper than the alternative. Herculano-Houzel, PNAS

Here is what one of those 86 billion cells does. It sits quietly at about minus 70 millivolts, slightly negative compared to its surroundings. Other neurons feed it small pushes through its dendrites. Those pushes add up. They also leak away, so a slow trickle of input never gets anywhere. But if enough arrives fast enough to lift the cell to about minus 55 millivolts, something abrupt happens: the neuron fires a spike, at full strength, down its axon, and passes the signal on.

It is all or nothing. There is no half-fire. A stronger stimulus does not produce a bigger spike, it produces more spikes. That single fact, a threshold and an all-or-nothing response, is the seed the entire field grew from.

Interactive 01

Fire a neuron

Push the button to send a signal into the dendrites. Watch the voltage climb in the trace at the bottom right. Turn the stimulus down low and hold it: the charge leaks away faster than it builds and nothing happens. Turn it up and the cell crosses the threshold and fires.

Leaky integrate-and-fire model

How hard each incoming signal pushes, in millivolts. It takes 15 mV to climb from rest to the threshold, so at this setting three quick presses will do it. Drop it to 1 and you can press as fast as you like and never fire: the charge leaks away faster than you can add it.

Used by the steady signal train below.

Voltage-70 mV
Spikes0
StateAt rest

What this shows A neuron is a decision, not a wire. It gathers evidence, forgets it if nothing more arrives, and commits completely once the evidence crosses a line. Keep that shape in mind. You are about to see a machine do a cartoon version of it, billions of times per second.

Part two

Now the cartoon version

In 1943 a neurophysiologist and a logician wrote down a mathematical caricature of the cell you just fired. Strip away the chemistry, the timing, the living tissue. Keep only this: take some inputs, multiply each by a number that says how much it matters, add them together, and if the total is big enough, pass something on. That caricature is the artificial neuron, and it is still what a trillion-dollar industry is built out of.

Interactive 02

The artificial neuron

Three inputs come in on the left. Each has a weight, which is just a number saying how much that input counts. Negative weights argue against. Drag anything and watch the wires thicken, the total change, and the output flip. This is the whole mechanism. There is nothing else hiding behind it.

One artificial neuron. A frontier model contains billions.

The neuron's built-in reluctance, or eagerness, before any input arrives.

Total0.00
Output0.000
FiresNo

The arithmetic, live

What this shows A neuron in a model is one multiplication, one addition, and one squash. Nobody programs the weights. They start as random noise and get nudged, billions of times, by the process in Interactive 05. Intelligence, to whatever extent this counts as intelligence, is an emergent property of a very large number of these arranged in layers.

Language models do not read words.

Before any of that arithmetic can happen, text has to become numbers. A model chops your sentence into pieces called tokens, which are sometimes whole words, often fragments, and occasionally single characters. Each piece is looked up in a fixed vocabulary of about a hundred thousand entries and replaced by its number. That is what the model actually sees: a list of integers.

This is not a technicality. It explains a family of behaviours that otherwise look like stupidity. Ask a model to count the letters in a word and it may get it wrong, because it never saw the letters. It saw two or three chunks with ID numbers attached.

Interactive 03

What the machine sees instead of your sentence

Type anything. Watch it break apart. Notice that common short words survive whole while longer or unusual words shatter into fragments, and that a leading space is part of the token itself.

Tokens0
Characters0
Chars per token0.0

Honest caveat This is a simplified splitter built to show the idea. Real systems use byte pair encoding, learned from the training data rather than hand-written, and the ID numbers come from a fixed vocabulary rather than being computed on the fly. The shape of what happens is the same: your words become a list of numbers, and the model never sees anything else.

Everything else is one trick, repeated.

Given the list of numbers so far, produce a probability for every possible next number. Pick one. Add it to the list. Do it again. That is what a large language model does, and it is the only thing it does. Every essay, every apology, every line of code you have ever seen it produce came out one token at a time, each one chosen from a ranked list of options.

Below is a real language model. Not a picture of one. It is a small n-gram model that your browser builds when this page loads, by reading a passage of ordinary English and counting which words follow which. It is genuinely tiny, and the difference between it and a frontier system is scale, architecture and training, not concept. Watch it choose.

Interactive 04

Watch it pick the next word

The bars are real probabilities from a real model. Click any word to accept it yourself, or let the model write on its own. Then drag the temperature. At zero it always takes the top bar and gets repetitive. High, and it starts reaching for unlikely words and falls apart. Every AI product you have used has this dial set somewhere.

Low is cautious and repetitive. High is inventive and unreliable.

What could come next

Context matched3 words
Words written0

Matched on:

Training words0
Distinct words0
Learned patterns0
GPT-3 parameters175,000,000,000

What this shows The model has no plan for the sentence. It has no idea how the sentence ends. At every step it looks back at what has been written, produces a ranked list of possible next words, and picks one. The sense of intention you feel when you read the output is manufactured entirely in your head, one word at a time, after the fact.

Which raises the obvious question. If nobody sets those probabilities by hand, where do they come from? The answer is the least glamorous and most important idea in the field: the model starts out random and useless, and is then corrected, over and over, by being shown how wrong it was.

Interactive 05

Training is just being wrong, less

The blue dots are real answers. The purple line is what a small network currently believes. Right now it believes nonsense, because its weights are random. Press train. Real backpropagation runs in your browser: forward pass, measure the error, push the correction back through every weight. Nothing here is pre-recorded.

8 hidden units, tanh activation, batch gradient descent.

How big a correction to make each round. Too small and it crawls. Push it past about 2 and watch it overshoot and blow up. Finding this number is a real part of the job.

Rounds0
Error0.0000
StateUntrained

What this shows Learning, in this field, is a search for numbers that make the error smaller. That is the entire definition. There is no understanding step, no moment of insight. Scale this up to trillions of words and billions of weights and the behaviour that emerges is startling, but the mechanism on screen is the mechanism, all the way up.

Interactive 06

Why it makes things up without blinking

Same model as Interactive 04. The first button starts from a phrase it has seen. The second starts from one it has never seen in its life. Read both outputs and try to tell, from the writing alone, which is which.

Stated confidenceAlways 100%
Actual groundingStrong

What this shows A hallucination is not a malfunction. It is the machine working exactly as designed. The model is built to always return its most likely next token, and it has no separate circuit for reporting that the evidence behind that token was thin. Fluency is produced by the same process whether the underlying knowledge is there or not, which is precisely why confident nonsense reads so well. When the answer matters, the burden of verification is on the human. Every time.

Rows of illuminated servers in a data centre
The other side of the comparison Your brain does what it does on about 20 watts. The machines do not. Global data centre electricity use was around 415 terawatt hours in 2024, roughly 1.5 percent of all electricity on Earth, and the International Energy Agency's central case has it more than doubling to about 945 terawatt hours by 2030. AI's share of that load, 5 to 15 percent in recent years, is projected to reach 35 to 50 percent. Whatever these systems are, they are not efficient in the way you are. IEA, Energy and AI
The honest part

It would be neat to end part two by saying we fully understand these systems because we built them. We do not. We know the architecture and we chose the training procedure, but the meaning of any particular weight among billions is not something anyone wrote down. It is discovered, afterwards, by research.

In 2024 Anthropic managed to identify millions of interpretable concepts inside a production model, including a feature that responds to the Golden Gate Bridge whether it appears in English, Japanese, Greek or an image. Turning that feature up caused the model to insist it was the bridge. That work is genuinely impressive, and the fact that it counted as a landmark tells you how much of the inside of these systems remains unmapped. Anthropic, 2024

Part three

So what does this do to work

This is the part everyone actually came for, and it is the part where the public conversation is worst. The headline number you have seen, whichever one it was, almost certainly measured something narrower than the sentence it was wrapped in. So here is the careful version, with the measurements separated from the predictions.

80% of US workers are in an occupation where at least 10 percent of tasks could be done substantially faster with a language model. Note the words task and faster. Eloundou et al, Science 2024
19% are in occupations where that applies to at least half their tasks. A much smaller and much more serious number. GPTs are GPTs
34 occupations in the same study had no tasks flagged as exposed by any measure at all. Not few. None. GPTs are GPTs, Table 11
25% of all US work hours could in principle be automated by AI, on Goldman Sachs Research's estimate, with roughly 300 million jobs globally exposed to some degree. Goldman Sachs Research

One word is doing an enormous amount of work in all of those figures, and it is exposure. In the Science paper, a task counts as exposed if a language model could cut the time it takes by at least half. The authors state plainly that this does not mean the task can be fully automated, and it certainly does not mean the job disappears. A job is a bundle of tasks, plus accountability, plus judgement, plus the fact that somebody has to be answerable when it goes wrong.

Look up your own work below. Then read the number as what it is: an estimate of how much of the daily task list gets faster, not a countdown.

Interactive 07

Task exposure by occupation

Every figure here is from the same peer-reviewed study. Occupations are grouped by which measure flagged them, because the measures ask slightly different questions and mixing them silently would be the exact sort of thing this page exists to argue against.

The study covered 750 occupations. This is a readable slice of the extremes, taken directly from its published tables.

Mathematicians

Highly exposed
Share of tasks exposed 0%

Read it carefully Notice the pattern in the bottom group. The work that came back at zero is overwhelmingly physical, skilled, and performed in an unpredictable environment: line workers, diesel mechanics, masons, short-order cooks. For twenty years the story was that automation comes for manual labour first and leaves thinking work alone. That story has been running in reverse.

Customer support agents wearing headsets at a shared desk
Where it is landing first Support, coordination and routine written communication. Highly exposed on every measure, and also where the first measurable employment effects have shown up.
A welder working with sparks flying
Where it is not Skilled hands in an unpredictable physical space. Language models are extraordinary with language, and a welding seam is not language.
A nurse taking a patient's blood pressure
Where demand is growing The World Economic Forum's employer survey puts nursing, social work, counselling and personal care among the fastest growing roles to 2030, alongside the technology jobs everyone expects.

The first real measurements are in.

Everything above is about potential. Potential is cheap. The harder question is whether anything has actually happened in the labour market yet, and for the first time there is payroll evidence rather than speculation.

Stanford's Digital Economy Lab went through ADP payroll records covering millions of American workers. In the occupations most exposed to AI, employment for 22 to 25 year olds fell about 6 percent between late 2022 and September 2025, while employment for older workers in the same occupations rose 6 to 9 percent. In less exposed occupations, young workers were fine. The researchers called them canaries in the coal mine, and the detail that matters most is this: the declines cluster in places where AI automates the work rather than assists it. Where it assists, entry-level employment held up. Stanford Digital Economy Lab

That distinction between automating and assisting is not academic, and it is now measurable at scale. Anthropic publishes an index of how its own model is actually used. As of November 2025, 52 percent of consumer conversations were augmentation, where a person is iterating, learning or getting feedback, against 45 percent automation, where the task is simply handed over. In programmatic API traffic, automation dominates. The same technology lands very differently depending on how an organisation chooses to deploy it. Anthropic Economic Index

Part four

What comes next, and how much to trust it

Everything up to this point was measured. Everything below is forecast, which is a different kind of statement wearing similar clothing. These are the most serious projections available, each with its author and date attached, because a forecast without a source is just a mood.

Interactive 08

Drag through the next decade

Move the slider forward and the projections light up as you pass their target year.

20262032203820442050
2028

Superforecasters expect AI to handle a full day of expert software work

Asked when a model will hit 80 percent success on software tasks that take a human expert eight hours or more, superforecasters give a median of 2028. Domain experts in the same exercise say 2030. Both groups have moved earlier with each round of the survey.

2030

170 million jobs created, 92 million displaced

The World Economic Forum's employer survey projects a net gain of 78 million jobs by 2030, with 22 percent of all jobs disrupted and 39 percent of the skills a typical job requires expected to change. Net positive, which is not the same as painless for the 92 million.

2030

Data centres draw about 945 terawatt hours

The IEA's central scenario has global data centre electricity use more than doubling from 2024, with AI's share of that consumption rising from 5 to 15 percent toward 35 to 50 percent. The constraint on AI may end up being electrical, not algorithmic.

2047

Median researcher estimate for a 50 percent chance of high-level machine intelligence

From a 2023 survey of thousands of AI researchers. Worth holding beside one detail: the same survey a year earlier gave 2060. Thirteen years of expectation evaporated in twelve months, which tells you as much about the reliability of these estimates as about the technology.

How to hold these Forecasts about AI have a poor track record in both directions. The honest position is that the near-term measurements are solid, the five-year projections are informed guesses from people with incentives, and anything past that is a genre of writing rather than a form of knowledge. Plan for the measured part. Stay flexible about the rest.

What actually seems worth doing

No prediction, just the pattern that falls out of the evidence above. The work holding up best is work where somebody has to be accountable, where the environment is physical and unpredictable, or where the value is in the relationship rather than the artefact. The work under most pressure is routine written output produced at a desk, especially at entry level.

The most useful thing an individual can do is unglamorous: use these tools enough to develop real judgement about where they fail. The people who will be fine are not the ones who avoided this and not the ones who trusted it blindly. They are the ones who know exactly when to check its work, because they have watched it be confidently wrong and learned the tells.

Part five

Every source on this page

No figure above is unattributed and none of it came from a press summary of a press summary. If a claim here is wrong, it should be traceable to the paper that made it.

Photography and footage from Pexels. The neuron simulation, the artificial neuron, the language model, the training loop and the exposure explorer were all written for this page and run entirely in your browser. No data leaves your device.

Understanding this well is now part of the job

I am Kevin Champlin, founder and CEO of Champlin Enterprises. We build production software and AI systems, which mostly means knowing precisely where these tools are extraordinary and precisely where they will confidently hand you something wrong. This page is what I wish more people had been shown before they were asked to have an opinion.