Most software follows instructions a programmer wrote down. A tax program applies the rules it was given, and a calculator does exactly what its circuits say. The systems now called AI mostly work another way, and that difference explains a great deal about both their power and their strange mistakes. This level builds a plain picture of what goes on inside, without any math.

Learning from examples

Imagine trying to write rules for spotting spam: flag messages that mention prizes, come from unknown senders or shout in capital letters. Spammers adapt, and the rules never keep up. Machine learning takes a different route: instead of following hand-written rules, the system learns patterns from many examples. Shown millions of emails marked spam or not, it tunes itself to tell them apart, using clues no programmer listed.

Most modern systems are neural networks, loosely inspired by the brain. They are made of layers of simple calculations connected by billions of adjustable numbers. A model's parameters, also called weights, are the adjustable numbers set during training. Training means showing the network an example, measuring how wrong its output is, nudging every parameter slightly to be less wrong, and repeating this an enormous number of times. Nobody sets the parameters by hand, which is why even the builders cannot fully explain why a model gives a particular answer.

Predicting the next word

The chatbots most people meet are built on large language models. At their core, such models are trained to predict the next piece of text in a sequence, usually a word or part of a word called a token. Fed a vast amount of writing from books, websites and code, a model learns that "The capital of France is" is usually followed by "Paris." To predict well across billions of sentences it has to absorb grammar, facts, styles of argument and some patterns of reasoning, because all of those help guess what comes next.

A raw prediction engine is not yet a helpful assistant, so companies shape it with further training. One common method is reinforcement learning from human feedback, or RLHF, in which people rate the model's answers and it is tuned toward the kinds they prefer. This makes models more polite, more useful and more willing to refuse harmful requests. As later levels show, it can also teach them to please.

Training and use

It helps to separate two phases. Training builds the model, and inference is using the finished model to answer. Inference happens every time you type a question. The model's parameters do not change while you chat with it, although the app around it may store your conversation or add tools.

Because the training data was gathered up to some date, every model has a knowledge cutoff: the date after which its training data ends. Ask about later events and a model with no search tool may guess, or describe the world as it used to be. Many products now connect models to web search or to an organization's own documents, a technique called retrieval, so the model can draw on fresh sources rather than memory alone.

Predictive and generative systems

Chatbots and image makers are generative: they produce new text, pictures, audio or code. Older and quieter systems are predictive: they produce a score or a category, such as the chance a loan will be repaid, whether a payment looks like fraud or which video you might watch next. Predictive systems shape decisions about credit, insurance, policing and welfare, often without the affected person knowing. Both kinds learn from data, and both inherit its flaws.

Do they understand?

Here thoughtful people disagree. In 2021 the linguist Emily Bender, the computer scientist Timnit Gebru and two coauthors published a widely read paper on "stochastic parrots." They argued that language models stitch together forms of language without any grasp of meaning, so fluent output should not be mistaken for understanding. On this view the model is a remarkable mimic and nothing more.

Others point to evidence of internal structure. In a 2023 study often called Othello-GPT, researchers trained a model only on lists of moves from the board game Othello, never showing it a board. Probing its insides, they found that the model had built an internal map of the board and used it to choose legal moves. Defenders of this view say prediction at scale can push a model to build working pictures of the world. The fair summary is that these systems are neither simple parrots nor minds like ours, and what should count as understanding remains an open question.

Worked example: the library question

Dana asks a chatbot when her town library opens on Saturdays. The model has no search tool. It has seen thousands of library web pages, so it predicts a likely answer: Saturdays, 10 a.m. to 5 p.m. The answer is fluent, specific and invented. The real library opens at 9 and closes at 1.

Nothing malfunctioned: the model produced a plausible continuation, which is what it was trained to do. Dana tries a different app that searches the web first. It retrieves the library's own page and quotes it, with a link. The second answer is more trustworthy not because the model is smarter but because it rests on a source Dana can check. Knowing which kind of tool you are using is the first habit of living with AI.