AI in one sentence

AI stands for artificial intelligence. It's software that can do things we used to need a human brain for, like writing an email, summarizing a contract, answering a question, or sorting through a pile of files.

That is the whole idea. Everything else is a flavor of that.

So what is an LLM?

LLM stands for large language model. It is the type of AI that powers ChatGPT, Claude, Gemini, and most of the AI tools you see today. When people say "AI" in a business context in 2026, they almost always mean an LLM.

Think of an LLM as an extremely well-read assistant. It read a huge chunk of the internet, plus a lot of books, articles, and code. From all that reading, it learned how language works. It learned grammar, facts, reasoning patterns, and how people explain things to each other.

When you type a question, the LLM doesn't look up an answer in a database. It predicts what words should come next, one piece at a time, based on everything it learned. That is the trick. It's word-by-word prediction at a scale so massive it feels like thinking.

A useful mental model

Imagine an extremely well-read intern who is fast, available 24/7, and never gets tired, but also has no memory of yesterday's work and sometimes invents things if you don't check.

What an LLM can actually do

  • Write and rewrite. Emails, proposals, blog posts, contracts, summaries, marketing copy, scripts.
  • Read and explain. Drop in a 40-page document and ask for the key points, the risks, or a plain-English version.
  • Translate. Between languages, or between "lawyer speak" and "what does this actually mean."
  • Brainstorm. Names, angles, headlines, agenda items, what-if scenarios.
  • Answer questions. About almost any topic it learned during training.
  • Write code. If you ask it to. It can also explain code line by line.
  • Follow a process. If you describe a step-by-step workflow, it can run it.

What an LLM cannot do (or shouldn't be trusted to do alone)

  • Know the latest news. Most models have a knowledge cutoff. They don't know what happened yesterday unless they can search the web.
  • Remember you between sessions. Unless the tool has a memory feature turned on, each new chat starts fresh.
  • Be 100% accurate. LLMs sometimes make things up. The polite term is "hallucinate." Always verify anything that matters.
  • Make real decisions for you. They are great at drafting and analyzing. They should not be the final word on legal, medical, or financial calls.
  • Replace human judgment. Especially in regulated work. AI helps the professional. It doesn't become the professional.

The different "sizes" of LLM, in plain terms

You will see model names like "Haiku," "Sonnet," "Opus," "GPT-5.5," "GPT-5.5 Pro," "Gemini Flash," "Gemini Pro." These are different sizes from the same company, each tuned for a different job.

  • Small and fast. Cheap, quick, good for simple tasks like sorting emails or short summaries. Examples: Claude Haiku, GPT-5.5 Nano, Gemini Flash.
  • Medium and balanced. The everyday workhorse. Good for most writing, analysis, and back-and-forth. Examples: Claude Sonnet, GPT-5.5.
  • Large and deep. Slower and more expensive but better at hard reasoning, complex analysis, and nuanced work. Examples: Claude Opus, GPT-5.5 Pro.

You don't always need the biggest model. For most everyday tasks, the mid-size one is plenty.

The intermediate take: what AI is good at vs great at

AI is good at almost any text-based task. It is great at tasks where the rules are clear, the answer can be verified, and you can give it strong context. The more specific you are about what you want, who it's for, and what good looks like, the better the output.

Vague request in, vague output out. Specific request in, surprisingly good output out.

The advanced take: how it actually works under the hood

An LLM is built on something called the transformer architecture. The model has hundreds of billions of internal numbers called parameters. During training, those numbers get tuned by feeding the model massive amounts of text and asking it to predict the next word, over and over, trillions of times. Once trained, those parameters are frozen and the model uses them to generate responses to whatever you type.

When you send a prompt, the model breaks your text into pieces called tokens (roughly word fragments), runs the math, and produces one token at a time until it finishes its response. That is why responses sometimes "stream" out word by word. You are watching the prediction happen live.

You don't need to know any of that to use AI well. But it explains why AI sometimes gets confused on long inputs, why it costs money per token, and why prompting it clearly matters.