How AI Learns

An AI can only learn what's in the examples it's fed, so the amount, quality, and variety of its data shape everything it can and can't do, including its blind spots and biases.

What counts as "data"?

Data is any recorded information a computer can learn from. For AI, that includes:

• Text: books, websites, news articles, encyclopedias, online forums, computer code, transcripts.
• Images: photos, drawings, medical scans, satellite pictures.
• Sound: speech recordings, music, phone calls (with permission, one hopes).
• Video: films, online videos, security and dashboard cameras.
• Numbers and records: sales, weather readings, bank transactions, sensor measurements.
• Behavior: what people click, watch, buy, skip, and search for.

How much food are we talking about?

The amounts are almost impossible to picture. A large modern chatbot is typically trained on trillions of words. If you read one book a day, it would take you many thousands of lifetimes to read that much. Big image systems learn from hundreds of millions or even billions of pictures.

For comparison: a human child hears roughly tens of millions of words in their first several years, and learns to talk just fine. AI needs vastly more examples to learn language than a child does. That's one of the biggest differences between machine learning and human learning, and a mystery researchers are still working on.

Where does it come from?

• The public internet, gathered by automated programs ("web crawlers") that copy pages at huge scale.
• Licensed collections: companies pay publishers, news outlets, photo libraries, and websites for access.
• Product use: what people type, click, and upload, depending on each company's policies and your settings.
• Human labelers: people paid to tag photos, rate answers, and write examples. This hidden workforce numbers in the many thousands worldwide.
• Synthetic data: examples generated by other AI systems, used more and more as fresh human data gets harder to find.

Cleaning the plate: data quality

Raw data is messy. Builders spend a large share of their effort filtering it: removing duplicates, spam, gibberish, and some harmful material, and trying to balance it. Computer scientists have an old saying for what happens when they don't: "Garbage in, garbage out."

Three quality problems come up again and again:

1. Gaps. If a medical AI saw mostly scans from one group of people, it may perform worse for everyone else.
2. Skew. The internet over-represents some languages, countries, ages, and viewpoints, and under-represents others. AI inherits that tilt.
3. Errors and falsehoods. The internet contains plenty of mistakes and myths. An AI can absorb them just as easily as facts.

The big open questions about data

Data is where many of the hottest debates about AI begin:

• Consent and copyright: Did authors, artists, and photographers agree to have their work used for training? Should they be paid? Courts in several countries are hearing these questions, and the answers are still being decided.
• Privacy: Personal information posted online can end up in training data.
• Running out: Some researchers estimate that the supply of high-quality human-written text is being used up, which is one reason companies turn to synthetic data and licensing deals.

An AI's life has two very different stages: a long, expensive training phase where it learns from data, and a fast, everyday inference phase where it uses what it learned to answer you, usually without learning anything new from your conversation.

Phase 1: Training (school)

• What happens: the system is shown enormous amounts of data and adjusts itself, over and over, to get better at its task. (Section 4 shows exactly how.)
• How long: for the largest systems, weeks to months of nonstop computing.
• Where: giant data centers packed with thousands of specialized chips.
• Cost: for the biggest modern models, reported costs run to tens or hundreds of millions of dollars for a single training run, and rising.
• How often: rarely. A major model might be trained once, then refined, then eventually replaced by a new version.

Training has stages too

For chatbots, "training" usually happens in steps, like a general education followed by job training:

1. Pre-training (general education): the model reads a vast library of text and learns to predict the next word. This is where it picks up grammar, facts, writing styles, and reasoning patterns. At this stage it's knowledgeable but not yet a helpful assistant; it might answer a question with another question, or just keep rambling.
2. Fine-tuning (job training): it's trained on carefully written examples of good conversations, so it learns to act like a helpful assistant.
3. Feedback training (coaching): people (and sometimes other AI systems guided by written principles) rate its answers, and it's adjusted toward the better ones. This is where much of its politeness, carefulness, and personality come from.

Phase 2: Inference (the job)

• What happens: you type a question; the finished model runs your words through its learned patterns and produces an answer.
• How long: seconds.
• Cost: a tiny amount per question, but multiplied by hundreds of millions of people, inference adds up to a very large share of AI's total energy use.

The "knowledge cutoff"

Because training ends on a certain date, a model's built-in knowledge stops there. This is called its knowledge cutoff. Ask about something that happened afterward and it may not know, or worse, may guess. Many chatbots now get around this by searching the web during inference, the way our frozen-memory doctor might look something up in a current journal before answering.

"But it remembered what I said!"

Within a single conversation, a chatbot can refer back to what you said earlier. That's not learning; it's more like the doctor reading your chart during your appointment. The information sits in the conversation's short-term "working space" (called the context window). Some products add memory features that save notes about you between conversations, but that's a filing system added on top, not the model itself changing. Whether your conversations are later used to train future versions depends on the company's policy and your settings.

AI finds patterns in past examples and uses them to make a best guess about something new, whether that's the next word, the next song, or whether a charge is fraud, and nearly everything impressive it does is built from that one move.

Finishing the sentence

Try this. Finish each line in your head:

• "Once upon a…"
• "Peanut butter and…"
• "The capital of France is…"

You probably said time, jelly, and Paris without effort. You just made three predictions from patterns you've absorbed over a lifetime. A chatbot does the same thing, one small piece at a time.

How a chatbot writes a whole answer

A chatbot doesn't plan its entire reply and then type it out the way you'd write a letter. At its core, it:

1. Looks at everything so far (your question plus anything it has already written).
2. Calculates how likely each possible next piece of text is. (These pieces are called tokens: usually a word or part of a word.)
3. Picks one, often the most likely, sometimes with a little deliberate randomness so answers don't sound robotic.
4. Adds it to the text, and repeats. Thousands of times per answer.

That's why you can ask the same question twice and get slightly different answers: a small roll of the dice happens at every step.

How can "just predicting" do so much?

Here's the surprising part. To predict the next word really well across trillions of words, a system has to pick up a lot along the way. To predict the end of "The ball fell because of…," it helps to have learned something about gravity. To predict the next line of a legal argument, it helps to have absorbed how legal reasoning works. To predict the answer to a math problem, it helps to have learned the steps.

Nobody programmed those skills in directly. They emerged as by-products of getting very good at prediction. Researchers still debate how deep that "understanding" goes, but there's no doubt it's more than simple memorization.

Prediction beyond words

Where prediction goes wrong

• Plausible isn't true. The most likely-sounding answer can be false. Ask for a book on an obscure topic and the AI may invent a realistic title and author, because that's what a book reference usually looks like. This is a "hallucination."
• The future isn't always like the past. Patterns learned before a big change (a pandemic, a new law, a market crash) can mislead afterward.
• Correlation isn't cause. AI notices what tends to go together, not necessarily why. Ice cream sales and drownings rise together, but ice cream doesn't cause drowning; summer causes both.

A neural network is a huge stack of layers made of simple math units connected by adjustable "dials"; training means nudging billions of those dials, a tiny bit at a time, until the network's guesses stop being wrong.

The pieces

• Neurons (units): tiny calculators. Each one takes in some numbers, combines them, and passes a number along. One unit alone is almost laughably simple.
• Connections with weights (the dials): every connection between units has a number, a "weight," that controls how strongly one unit influences the next. These weights are the dials, and they're what training changes. Together with a few other adjustable numbers, they're called parameters.
• Layers: units are arranged in rows. The first layer takes in the raw input; the last layer gives the answer; the layers in between (the "hidden layers") do the work. "Deep" learning simply means many layers.

What the layers learn

In an image-recognition network, researchers have peeked inside and found something like a production line:

1. Early layers learn to notice simple things: edges, lines, patches of color.
2. Middle layers combine those into shapes: circles, corners, textures like fur.
3. Later layers combine shapes into parts: ears, eyes, noses, wheels.
4. The final layer weighs it all up: "dog, 94% sure."

No one told it to look for ears. It discovered that noticing ears helps it be right.

How training nudges the dials, step by step

1. Guess. Show the network an example (say, a photo) and let it make a guess with its current dial settings.
2. Check. Compare the guess to the right answer. Measure how wrong it was. (This "wrongness score" is called the loss.)
3. Trace the blame. Work backward through the layers to figure out how much each dial contributed to the mistake. (The technique is called backpropagation.)
4. Nudge. Turn every dial a tiny amount in the direction that would have made the answer less wrong.
5. Repeat. Billions of times, over trillions of examples.

How it's like a brain, and how it isn't

The "black box" problem

Here's something that surprises most people: even the builders can't fully explain why a network gives a particular answer. They know the recipe for training it, and they can see all the dial settings, but billions of numbers don't come with labels. It's like having every page of a book written in a language no one speaks. A growing research field called interpretability is working on translating it, and has made real progress in finding what some groups of dials represent. But we're far from being able to read the whole thing.

AI learns in three main ways: from examples with the answers attached (supervised), by finding patterns in unlabeled data on its own (unsupervised), or by trial and error with rewards (reinforcement), and modern chatbots use a blend of all three.

1. Supervised learning: learning with an answer key

The AI gets examples that people have already labeled: "this email is spam," "this X-ray shows pneumonia," "this house sold for $310,000." It learns to predict the label for new examples.

• Everyday uses: spam filters, recognizing faces, reading handwriting, predicting home prices, flagging medical images for a doctor's review.
• Strength: clear goal, easy to measure how well it's doing.
• Weakness: labels are expensive; often people must make them by hand. And if the labels are biased or wrong, the AI learns the bias and the mistakes.

2. Unsupervised learning: finding patterns without an answer key

The AI gets data with no labels and discovers structure on its own: groups, similarities, oddities.

• Everyday uses: grouping customers with similar shopping habits, spotting unusual bank activity that doesn't match any known pattern, organizing news stories by topic, finding similar songs.
• Strength: works on huge amounts of unlabeled data, which is most of the data in the world.
• Weakness: the groups it finds may not mean anything useful to people. A human has to interpret them.

A clever in-between: self-supervised learning

This is the secret behind chatbots, and it's worth knowing by name. The AI takes ordinary text, hides a word, and tries to guess it. The text itself provides the answer, so no human labeling is needed. "The cat sat on the ___" → check: "mat." Because it doesn't need people to write labels, it can learn from trillions of words. This is how pre-training works (Section 2).

3. Reinforcement learning: trial, error, and reward

The AI takes actions, gets a reward or penalty, and gradually learns which actions lead to more reward.

• Everyday uses: game-playing AI like AlphaGo (which played millions of games against itself), robots learning to walk or grasp, managing cooling in data centers, and training chatbots to give better answers.
• Strength: can discover strategies no human thought of. AlphaGo's famous "Move 37" against Lee Sedol in 2016 stunned experts.
• Weakness: it chases whatever the reward actually measures, which isn't always what we meant. (See "Watch out for" below.)

Reinforcement learning from human feedback

Chatbots are polished with a version called reinforcement learning from human feedback (RLHF). People compare pairs of answers and pick the better one. The AI is adjusted to produce more answers like the winners. Some companies also use AI-assisted feedback guided by a written set of principles. This is the "coaching" stage from Section 2, and it's a big reason chatbots are helpful and polite, and also a reason they can be overly agreeable.

How a modern chatbot uses all of them

1. Self-supervised pre-training on trillions of words → broad knowledge.
2. Supervised fine-tuning on example conversations → acts like an assistant.
3. Reinforcement from feedback → more helpful, careful, and pleasant.

Researchers discovered that making neural networks bigger, training them on more data, and using more computing power made them better in surprisingly steady and predictable ways, so companies scaled all three up dramatically, and new abilities appeared along the way.

The three ingredients, turned way up

Lesson 1 named them. Here's what happened to each:

1. Data. The internet, digitized books, and online video created an ocean of examples, far more than existed even twenty years ago.
2. Computing power. Graphics chips (GPUs), originally built to draw video games, turned out to be excellent at the kind of math neural networks need: millions of simple multiplications at once. Companies now link tens of thousands of these chips together. The computing used for the largest AI training runs has grown by a factor of several times every year for over a decade.
3. Model size. The number of dials grew enormously.

The "scaling laws" discovery

Around 2020, researchers published findings that surprised the field: as you increase model size, data, and computing together, a model's error drops along a smooth, predictable curve. That meant companies could forecast roughly how much better a bigger model would be before spending the money to build it. It turned AI progress from a gamble into something closer to an engineering plan, and set off a race to build bigger.

The design breakthrough that made scaling work

Scale alone wasn't enough. The 2017 Transformer design (from Lesson 1's timeline) was key because it lets a model pay "attention" to how every word in a passage relates to every other word, and because it can be trained efficiently across thousands of chips at once. Older designs read text one word at a time, like reading through a keyhole; Transformers see the whole page at once.

Surprise abilities

As models grew, some skills seemed to appear rather suddenly: doing arithmetic, translating between languages they saw little of, writing code, following complicated instructions. Researchers debate whether these abilities truly "switch on" at certain sizes or grow gradually and only look sudden because of how they were measured. Either way, bigger models kept surprising their own builders.

Beyond just bigger

Recently, progress has also come from other directions:

• Thinking time: letting a model work through a problem step by step before answering (spending more computing during inference, not just training) improves results on hard problems.
• Better data, not just more: carefully chosen and filtered examples can beat sheer volume.
• Smaller, efficient models: clever techniques let compact models run on a phone or laptop while keeping much of the ability of larger ones.

The costs of scale

• Money: only a handful of companies and governments can afford the biggest training runs, which concentrates power.
• Energy and water: data centers use large amounts of electricity and, in many cases, water for cooling. This is a growing concern for local communities and climate goals.
• Chips: advanced AI chips are made by very few companies, making them a matter of national security and trade policy.

When training ends, what's left is "the model": a very large file containing billions of numbers (the final dial settings), which, run through the right software, turns your question into an answer. It's not a database of sentences, and it's not a little mind in a box.

What's inside the file

If you opened a model file, you'd see something like this, continuing for billions of entries:

0.0231, −0.1874, 0.0042, 0.3310, −0.0967, 0.1128, −0.2203, 0.0519…

That's it. No words, no pictures, no facts written in plain language. Just numbers, each one a dial setting from training. The knowledge is spread across all of them at once, the way a photograph's image is spread across millions of pixels.

How big is it?

• A small model that can run on a phone or laptop may be a file of a few gigabytes, around the size of a movie download.
• The largest models are hundreds of gigabytes or more, too big to run on a home computer, so they run in data centers and you reach them over the internet.

Compression, not a copy

A model is much smaller than the data it trained on. A chatbot trained on trillions of words doesn't store those words. It stores patterns distilled from them, more like the lessons a lifelong reader carries in their head than a library on a shelf.

That's also why models sometimes get facts wrong: the knowledge is a blurry summary, not a precise record. It's a bit like trying to redraw a famous painting from memory: the overall picture is right, but details drift.

There's a nuance: models can sometimes memorize and reproduce specific passages, especially text that appeared many times in training (famous quotes, common code, popular articles). This is part of the ongoing copyright debate.

The model vs. the product

When you use a chatbot app, you're using much more than the model:

So when two apps behave differently, it might be a different model, or the same model with different instructions and tools.

Open and closed models

• Closed models: the company keeps the file private; you can only use it through their app or service.
• Open-weight models: the company publishes the file of numbers so anyone can download, run, and modify it. This spreads access and allows independent study, but also means safety controls can be removed by whoever downloads it. Thoughtful people disagree on where the balance should be.

Versions and names

Companies release models in families with version numbers, much like car models: a bigger "luxury" version and smaller, faster, cheaper versions. When a new version comes out, it's a new file, trained again, often from scratch or from an earlier version. The old one doesn't "grow up"; it gets replaced.