A language model is autocomplete with a very good education.

Five steps, repeated for every word

It reads in chunks, not words

Before anything else, your text is cut into small pieces called tokens. Common words are one piece. Rarer words get split. Each piece gets an ID number, because computers only work with numbers.

Every piece becomes a spot on a map of meaning

Each token is swapped for a long list of numbers: coordinates. Words that are used in similar ways end up close together. Nobody told the model that cats and dogs are related. It worked that out from how people write.

Words look at each other to work out context

The same word can mean different things. Before predicting, the model lets every word check which other words matter to it. This is called attention. Change one word in the sentence and watch what "it" points to.

It scores every word, then picks one

The output is not one answer. It is a list of likely next words with odds. A setting called temperature controls how adventurous the pick is. Low means it always takes the favorite. High means it will sometimes take long shots.

Where the skill comes from: practice at huge scale

Inside the model are billions of tiny dials called parameters. At first they are set randomly and the guesses are nonsense. Then the model is shown text with the last word hidden, guesses, gets told the real word, and every dial is nudged slightly. Repeat this trillions of times.

Where the training text comes from

A model is only as good as what it read. The text comes from several sources, and each one raises its own questions. Click a source.

Why models sometimes get things confidently wrong

A model is built to produce a plausible next word. Plausible and true usually match, because true statements fill its training text. When they don't, it still sounds sure. Below, the question is about a book that does not exist.

Six common myths

Click a myth to see what is actually going on.

What makes one model better than another

Every model runs the same basic loop. They differ in how big they are, what they learned from, how they were coached afterward, and how much effort they spend on each answer. Move the levers and watch the trade-offs.

How people teach a model its manners (RLHF)

After reading, a model can only continue text. It has no idea what a helpful reply looks like. So people compare its answers and pick the better one. This is called reinforcement learning from human feedback, or RLHF. You be the rater for three rounds.

Reasoning models: thinking before answering

A newer kind of model writes out its working before it gives the final answer. That working is extra text in the context, so each later word can build on the steps before it. Try a famous trick question.

Multimodal: reading pictures, sound and video

Many models now take in more than text. The trick is the same as with words: cut the input into small pieces and turn each piece into numbers. A picture becomes a grid of patches. Click a patch.

Right-sizing: the biggest model is not always the right one

Many companies now use smaller models for narrow jobs and save the largest for hard ones. Pick a task and see which size fits.

Open or hosted: who holds the model?

With a hosted model, you use it through a company's service and cannot see inside it. With an open-weight model, the finished model files are published, so you can download, run and adapt them yourself. Strictly open-source models also share the training code and data, which is rare. Pick a priority.

Context: give it the facts before you ask

A model only knows what it absorbed in training, plus whatever is in front of it right now. That second part is the context. Put the right facts in the context and the odds shift toward the right answer. No retraining needed.

Prompts: how you ask changes what you get

The prompt is the part of the context you write yourself. Vague in, generic out. Each detail you add narrows the set of likely next words toward the answer you wanted. Tick the ingredients and compare.

RAG: fetch the right pages automatically

A company may have thousands of documents. They won't all fit in the context, and pasting them by hand is slow. RAG stands for retrieval-augmented generation. First a search step finds the few relevant pieces. Then those pieces are placed into the context, and the model writes the answer from them.

Fine-tuning or RAG: change the dials, or change the desk?

RAG and prompts change what the model sees. Fine-tuning goes the other way and changes the model itself. It continues training on your own examples, so the dials shift slightly. Pick a need and see which tool fits.

Tools: let it use a calculator, a database, the web

A model is great with language and poor at exact math or live data. So it can be given tools. It writes a short, structured request. Ordinary software runs the request and returns the result. The model then writes the answer using that result. The model never runs anything itself.

MCP: one plug for every tool

Every tool used to need its own custom connection to every AI app. MCP, the Model Context Protocol, is an open standard that Anthropic introduced in late 2024 to fix that. Any AI app that speaks MCP can use any tool that speaks MCP. It works like a USB-C port. Change the numbers and watch the wiring.

Agents: tools in a loop, working toward a goal

An agent is a model that keeps going. It looks at the goal, picks a tool, reads what came back, and decides the next move, repeating until it is done or it needs a person. Step through an example.

Automations: AI you can rely on every day

An agent decides its own next step. An automation follows a set path: something triggers it, it runs defined steps, and AI handles only the parts that need language or judgment. People approve the risky steps, and every run is logged. Most everyday business use of AI looks like this, because it is repeatable and easy to check. Walk a new lead through one.

Teams of agents: specialists passing work along

Instead of one agent doing everything, a task can be split among several, each with its own job and instructions. One plans, one researches, one writes, and one checks the work.

Memory: how an assistant remembers you

A model does not learn from chatting with you. Its dials stay fixed. When an assistant seems to remember, it is saving short notes and pasting them into the context next time. Untick a note and see the answer change.

Testing quality: how do you know it's good?

Models are often compared on public tests called benchmarks. Those scores are a starting point. What matters is how a model does on your own work. Compare two made-up models on both.

AI in everyday work: jobs are bundles of tasks

AI rarely replaces a whole job. It takes on some tasks, helps with others, and leaves the rest to people. Pick a role to see one way the tasks might split.

People in the loop: before, during and after

People are involved at every stage. Raters coach the model before release. You read, edit and approve its work while you use it. Reports, reviews and audits catch problems afterward. What changes from job to job is how much checking is needed.

Prompt injection: when the content gives the orders

A model reads instructions and content in the same stream of text, and it cannot always tell them apart. If an assistant reads an email or web page that contains hidden instructions, it may follow them. This is called prompt injection, and it matters most when the assistant has tools.

Keeping AI under control

The practical dangers today are mistakes, misuse, and handing AI too much power with nobody watching. Testers also look for subtler problems, such as misleading answers or actions outside the task, and researchers study longer-term risks from far more capable systems. Experts disagree about how large those risks are. The common answer is the same: stack several safeguards and keep people in charge of high-stakes decisions.

Regulation: rules scale with the risk

Governments are writing rules for AI. The best known is the EU AI Act, which sorts AI uses into four levels of risk. The higher the risk, the stricter the rules. Click a level.

Cost and value: does it pay for itself?

Analysts warn that many AI projects get cancelled because the costs are higher than expected or the value is unclear. The hidden costs are usually review time and upkeep. Adjust the numbers to see how quickly a project turns negative.

Energy: what a question costs the planet

AI runs in large data centers. Training a model is a big one-time cost. Using it, called inference, is small per question but adds up across millions of people. Heavier tasks use more.

Plain-English glossary