What is a model? Why are they large? Who is Claude?

At the core of most AI tools you’re using, whether it’s ChatGPT, Claude, or Gumloop, underneath those is a model. This is what’s processing the text, image, or audio you’re sending in and getting you a response.

How these models work is actually pretty simple to understand, but let’s be clear, extremely hard to build and vastly more complicated than what I’m about to explain.

Large language models are fundamentally next-word predictors. You write something like “Who is the first president of the United States?” and the model takes that, maps it against all the vast data it’s analyzed, and tries to predict the right next word, word by word. Every time it picks a word, it looks at your prompt plus what it’s written so far, and then predicts the next word. It does that until it’s answered your question.

This is how chatbots are built. You prompt them and they go word by word responding to you.

How do they go from one answer to a conversation? It’s actually just more of the same. You feed the whole conversation, every message you’ve sent so far, back to the model, and it predicts the next word. When you start a new conversation, the model starts fresh. It has no memory of what came before.

Now in Gumloop, you can pick from most of the models out there. How are they different? Well, models sit on a spectrum: intelligence on one axis and speed on the other, which are inversely related. The more capable the model, the slower you should expect a response, but with fewer mistakes. And the further up the curve, the more you should expect to pay.

Anthropic, the creator of Claude models, has three options. Opus sits on one end, thinks deeply, responds slowly, like your grandpa. Then there’s Sonnet in the middle, your capable coworker. And finally Haiku, your eager teenager ready for a quick answer. OpenAI and Google’s Gemini models have similar offerings, each along this axis.

So what should you pick in Gumloop? Well, that depends on your task, but what I always recommend is start with an advanced model and then move down to simpler models. As long as you’re still happy with the results, keep going until you find that perfect balance: quality you’re satisfied with, but with the least capable model.

So now we understand what models are. They’re next-word predictors. We can go back and forth with them. And chatbots are simply large language models that keep the conversation going. But how do we give AI access to our day-to-day tools so these large language models can actually do things for us, so we don’t just receive text? That’s in the next lesson.