The abbreviation you'll keep running into - in every piece of writing about artificial intelligence.
The mechanism is brutally simple, however mystical it sounds from outside. An LLM looks at a sequence of words and guesses which word comes next. Then another one. Then another. A whole sentence, a whole paragraph, a whole article is born from thousands of such small bets, one after another. That's why my register for it is a machine that finishes sentences - not because it's simple, but because that's exactly what it does at every moment, only with an instinct built from reading vast amounts of text a single person couldn't fit into a thousand lifetimes.
Here's the difference from an ordinary program. An Excel formula or a calculator follows rules someone wrote in advance: if this, then that. An LLM has no such rules - it has billions of internal settings, called parameters, that adjust themselves during training, as the model reads text after text and learns to recognize what sounds natural. Nobody sits down and writes 'if the question is about the weather, answer with a forecast'. The model gets there on its own - as statistics, not as logic.
The LLM itself is the engine. The product is the interface put on top of it - the chatbot you talk to. ChatGPT is OpenAI's chat app (an American artificial intelligence company), built on their GPT models, and it reached 100 million people in just two months - the fastest-spreading consumer app up to that point. Claude is the assistant from Anthropic. Behind Google stands Gemini. The different names you run into everywhere - GPT, Claude, Gemini, Llama - are different LLM-based systems, trained by different companies, with different taste and different mistakes.
And here's the catch I like to say straight out. An LLM doesn't check facts - it recognizes what sounds probable. That's why it sometimes sounds completely confident and is completely wrong - a phenomenon that already has its own name, hallucination. The machine doesn't lie on purpose. It finishes the sentence the most probable way, and the probable isn't always the true.
I'll say it as I think it
An LLM is like a very well-read colleague who never admits not knowing something. It'll answer you smoothly, with confidence, even when it's wrong. It's not devious - it just has no mechanism to say 'I'm not sure', unless someone has explicitly taught it to.
I work with these models daily, and the line I hold the whole time is this: I use them to think faster. I don't let them think instead of me. The machine finishes sentences brilliantly. Judgment remains my job.