A camera sensor does not record a picture. It counts light.

Behind the lens sits a chip of silicon covered in millions of tiny light detectors, called photosites. Each one is a bucket. When the shutter opens, light falls into the buckets for a fraction of a second. When it closes, the camera reads how full each bucket is. That grid of numbers is the photo.

Everything else, including colour, sharpness and the look of the final image, is built on top of that count.

Light becomes charge

Light arrives as photons. When a photon hits the silicon in a photosite, it can knock an electron loose. Those electrons are collected and held in a small well. More photons in means more electrons stored.

A photosite can only hold so many. A small pixel in a phone holds on the order of ten thousand electrons or fewer. A large pixel in a big camera holds many times that. Once a well is full, extra light has nowhere to go, and that pixel is pure white. That is what "blown highlights" means.

From charge to a number

At the end of the exposure, the sensor turns each well's charge into a voltage, amplifies it, and passes it to a converter that writes it down as a number. A raw file from a modern camera stores that number with 12 or 14 bits per photosite, which means thousands of possible brightness levels.

Photosites are colour-blind

Here is the surprise. A photosite only knows how much light arrived, not what colour it was.

To get colour, a grid of tiny filters sits on top of the sensor. Each photosite sits under a red, a green or a blue filter, in a repeating pattern called a Bayer pattern. In every block of four, two are green, one is red and one is blue. Green gets half because our eyes are most sensitive to it.

So each pixel records only one colour. The other two are missing. The camera fills them in by looking at the neighbours and making an educated guess. That step is called demosaicing, and it means a lot of the colour in your photo is estimated, not measured.

Why dark photos are noisy

Light comes in discrete packets, and photons do not arrive on a schedule. If a pixel should catch about 100 photons, it might catch 92 one time and 108 the next. This randomness is called shot noise, and it is physics, not a flaw in the camera.

The useful part is how it scales. The random variation is roughly the square root of the number of photons. A pixel that catches 100 photons has noise of about 10, so the signal is ten times the noise. A pixel that catches 10,000 photons has noise of about 100, so the signal is a hundred times the noise.

More light means a cleaner picture. This is why bigger sensors tend to look better in the dark: a bigger sensor catches more total light from the same scene. It is also why a phone adds a lot of software in low light. It has less light to work with.

Why fast motion can look bent

Most modern sensors are CMOS chips. Each pixel has its own tiny amplifier, and the sensor reads the grid one row at a time. Reading all the rows takes a moment, so the top of the frame is captured slightly earlier than the bottom.

If something moves fast across the frame, like a propeller or a passing car, it shifts between rows, and the shape looks skewed. This is rolling shutter. Sensors that capture every row at the same instant exist. They are called global shutter sensors, and they are harder to build.

What I've simplified

This is the short version. Real sensors add a tiny lens on top of every photosite to funnel more light in. Many move the wiring behind the light-sensitive layer so it does not block anything. Cameras also do a lot of processing after the read, such as noise reduction and sharpening, so the final image is never a pure count.

But the core idea holds. A sensor is millions of buckets catching light, and a photo is the count.

A language model does one thing, over and over: it predicts what comes next.

Give it the start of a sentence and it works out how likely every possible next piece of text is. It picks one, adds it to the text, and does the same again. A long answer is just that loop running many times.

That sounds too simple to produce anything useful. It is worth looking at how the loop works, because the details explain both what these models are good at and where they go wrong.

Text becomes tokens

A model does not read letters or whole words. It reads tokens, which are small pieces of text. Common words are often one token. Rarer words are split into several. "Understanding" might be one piece, while an unusual name might be three or four.

Each token is a number, and each number is looked up to become a long list of values the model can calculate with.

One step of the loop

For every step, the model reads all the tokens so far and produces a score for every token in its vocabulary, which can be tens of thousands of options or more. Those scores are turned into probabilities that add up to one.

If the text so far is "The capital of France is", the token for " Paris" gets a very high probability, and most other tokens get close to zero.

Then the model picks one. It does not always take the single most likely token. A setting called temperature controls how adventurous it is. At low temperature it almost always takes the top choice, so its answers are steady and repetitive. At higher temperature it will sometimes take a less likely token, so its answers are more varied and more likely to wander.

The chosen token is added to the text, and the whole step runs again.

How it learned

The model starts with billions of numbers, called parameters, set to nearly random values. Training adjusts them.

The process is simple to describe. Take a piece of real text. Hide the next token. Ask the model to predict it. Compare its prediction with the real one. Nudge every parameter slightly so that the right answer would have been a little more likely. Repeat this across a huge amount of text, an enormous number of times.

Nobody writes rules for grammar, facts or code. Whatever the model ends up knowing, it picked up because it made its predictions better.

What attention does

Predicting well often means looking back at an earlier part of the text. In "The trophy would not fit in the suitcase because it was too big", the word "it" refers to the trophy, and the model has to notice that.

Most modern models are built on an idea called attention, which lets each token look at the earlier tokens and decide which ones matter for the current prediction. That is how a model can connect a word to something said many sentences before.

From text-continuer to assistant

A model trained only to continue text is not yet an assistant. Ask it a question and it might continue with more questions, because that is what text often does.

To make it helpful, it gets more training on examples of good answers to instructions, and on human feedback about which answers are better. This stage shapes how it behaves, not just what it knows.

Why it can be wrong and sound sure

The training goal is to produce likely text, not to check what is true. When a model does not know something, a fluent, plausible-sounding answer is still the most likely kind of continuation, so it can state something false in a confident voice.

Models also only know what was in their training data unless they are given tools to look things up, and they can only consider a limited amount of text at once, called the context window.

What I'm not sure about

People still argue about what this adds up to. The model learns to predict, and to predict well it seems to build useful internal structure for things like grammar, facts and logic. Whether that deserves the word "understanding" is an open question, and so is exactly what the internal numbers represent. Researchers are working on reading them, but we cannot yet look inside and read them like code.

What is clear is the mechanism: predict the next token, check, nudge, repeat. Almost everything else comes from doing that very well, at huge scale.

When a company announces a "3 nm" chip, it is natural to picture something measuring three nanometres. There is nothing in the chip that measures three nanometres. The number is a name.

It used to mean something

For a long time, the node name was tied to a real size. In the 1990s and early 2000s, the labels 250 nm, 180 nm, 130 nm and so on tracked a real dimension of the transistor, such as the length of its gate. Each new generation shrank those features by roughly 30 percent, and the name followed.

Then the shrinking changed shape

Around the 2010s, making transistors smaller stopped being a matter of scaling the same design down. Engineers changed the design itself. Transistors went three-dimensional, with a thin fin of silicon standing up so the gate could wrap around it. Later designs wrap the gate around the channel from every side.

Once the improvements came from new shapes and new materials as much as from smaller dimensions, a single size no longer described a generation. But companies kept the naming habit. The number kept going down with each generation, even though it no longer matched a measurement.

What the real sizes are

The smallest features on a chip marketed as "3 nm" are measured in tens of nanometres. Commonly cited figures put the spacing between transistor gates at roughly 45 to 50 nanometres and the spacing between the finest metal wires at roughly 20 to 25. Sources differ, and the companies publish different numbers, so treat these as rough.

The point is that the label is much smaller than anything you could find with a ruler.

The names are not comparable

Because the number is a marketing label, "3 nm" from one manufacturer is not the same as "3 nm" from another. In 2021 Intel renamed its nodes to line up better with what its competitors were doing, so a name like "Intel 7" does not mean 7 nanometres either.

So a bare node name tells you the rough era of a chip, and not much more.

What to compare instead

A few measures say more.

• Transistor density. How many millions of transistors fit in a square millimetre. This is closer to a real, comparable measure of how much a manufacturer can pack in.
• Performance and power. How fast the chip runs at a given power, and how much energy it needs for the same work. This is what you actually feel.
• The design itself. A well-designed chip on an older node can beat a poor design on a newer one.

The short version

"3 nm" is a label for a generation of chip making. It is not a size. Next time you see a node name in a headline, read it as "the newest generation", then look for density and performance numbers.

Making any of these chips is still remarkable. The lithography machines that print the finest layers use light with a wavelength of 13.5 nanometres, called extreme ultraviolet. How that works is a story for another Deep Dive.