ChatGPT, Gemini and Claude sometimes feel like magic: ask a question in Hebrew and get a clear, well-phrased answer, sometimes with humor. Behind all of them is the same basic technology, a 'large language model', or LLM. Here is what happens inside, without equations.
At the most basic level, a language model does one thing: it guesses the next word, or more precisely the next piece of text. When you write 'The capital of France is', the model calculates that the most likely continuation is 'Paris'. It does this again and again, word after word, until a full answer is built. It sounds simple, but to guess the next word well in every possible context, the model has to 'learn' an enormous amount about the world, language and reasoning.
How does it learn? Through training. The model goes through huge amounts of text, books, websites, code and documents, each time trying to guess the next word and correcting itself when wrong. That training needs thousands of advanced chips running for weeks or months, so it costs hundreds of millions of dollars or more. The model then goes through another stage in which people teach it to answer helpfully, politely and safely.
That is also where the best-known problem comes from: 'hallucinations'. Because the model guesses what sounds plausible rather than checking facts in a database, it can produce an answer that sounds completely convincing yet is wrong, including numbers, quotes and sources that do not exist. Newer models err less, and some can search the web and cite, but the problem has not gone away.
What does it mean for you? A language model is an excellent tool for writing, summarizing, translating, explaining and brainstorming, but not a reliable source of facts without checking. A good rule of thumb: use it like a very fast, smart assistant who is sometimes overconfident. Check every important fact, especially numbers, dates and names, before relying on it.
