You’re hearing “language model” everywhere. GPT-4, Claude, Llama, Gemini \u2014 they’re all language models. But what exactly is that?<\/p>\n
In 3 minutes, you’ll understand. Promise.<\/p>\n
A language model is a computer program that does one thing<\/strong>: predict the next word.<\/p>\n You give it “The Montreal Canadiens won the game” and it predicts the next word could be “last,” “against,” “thanks.” It picks the most likely one and continues. Word by word, it builds sentences, paragraphs, entire pages.<\/p>\n It’s the autocomplete on your phone, but on steroids. Your phone predicts one word. A language model predicts entire texts.<\/p>\n It was trained on a staggering amount of text. Books, articles, forums, websites, documents. Imagine your neighborhood library, multiplied by a million.<\/p>\n By reading all of that, the model learned the unwritten rules<\/strong> of language. Not just grammar \u2014 style, tone, associations of ideas, facts, opinions. It’s seen so many texts that it can generate new ones that look like something a human would write.<\/p>\n The word “model” is key. It’s a model<\/strong> of human language. A statistical representation of how humans use words. Not a copy of the brain. Not understanding. A model.<\/p>\n What’s wild is that just by predicting the next word, the model develops abilities nobody taught it. It can summarize, translate, explain, code, reason (a bit), crack jokes, write poetry.<\/p>\n Nobody told it how to write a summary. But it’s read so many summaries that it knows what one looks like. And it can generate a new one that fits your data.<\/p>\n It’s like a musician who’s listened to so much music they can improvise in any style. They didn’t consciously learn the “rules” of jazz \u2014 they absorbed them through exposure.<\/p>\n A language model doesn’t know what’s true<\/strong>. It knows what’s probable<\/strong>. That’s a massive difference. If the most likely text after “The capital of Australia is” is “Sydney,” it’ll say Sydney. Even though the right answer is Canberra.<\/p>\n It doesn’t have real-time Internet access (unless specifically connected). Its knowledge stops at its training date. It doesn’t know what happened yesterday.<\/p>\n And it makes things up. With confidence. Because its job is to produce probable text, not true text. That’s why you should always verify important facts.<\/p>\n A language model predicts words. It learned by reading billions of texts. It’s impressive but not 100% reliable. And understanding that is the foundation for using all these tools well \u2014 ChatGPT, Claude, Gemini, or Sherpa<\/a>, our free AI guide.<\/p>\nHow it knows what to predict<\/h2>\n
Why it’s impressive<\/h2>\n
The limits you need to know<\/h2>\n
There, you know the essentials<\/h2>\n