When you type a question and hit “go,” you’re triggering something remarkable—years of research, training, and technology come together to generate a response in seconds. Tools like ChatGPT or DeepSeek-R1 don’t “understand” your question like a person would. Instead, they look at your words and predict what should come next, one piece at a time. This ability comes from three key stages: Pretraining, Supervised Fine-Tuning, and sometimes Reinforcement Learning.

It starts with your question getting broken down into small parts—called tokens. These could be words or even pieces of words. The model treats this as the beginning of a sentence and starts filling in the rest, one token at a time. The response you see isn’t pulled from a database or a Google search. It’s generated on the fly, based on how the model has been trained to “sound” helpful.

That training begins with Pretraining. The model reads and learns from a massive amount of text—things like books, websites, and articles—billions of words in total. It learns how language flows, which words usually follow others, and how ideas connect. But at this point, it’s just mimicking internet text. It doesn’t know how to be useful yet.

That’s where Supervised Fine-Tuning comes in. Here, real people write thousands of examples showing how a helpful assistant should respond to different questions. The model studies these examples and starts to copy the tone, structure, and helpfulness of those human-written replies. So when you ask something, it responds like someone who’s been trained to help—instantly.

Still, just copying examples isn’t enough. That’s where Reinforcement Learning makes a difference. In areas where answers can be clearly checked—like math or coding—the model tries different solutions, and the good ones get reinforced. Over time, it learns to reason better: checking its work, breaking down problems, even trying new approaches. These aren’t things we directly teach it—it figures them out through practice.

For things that don’t have a clear right answer—like writing a poem or summarizing an article—there’s no obvious way to score responses. So instead, people rank a few different outputs, and the model learns which ones humans prefer. This process is called Reinforcement Learning from Human Feedback (RLHF). It works, but it’s not perfect. Sometimes the model learns to game the system and gives high-scoring nonsense. It’s clever, but not always in the way we want.

Even with all this training, LLMs have quirks. They might solve a tough algebra problem but then mess up a basic spelling question. That’s because they don’t actually think—they don’t know facts or hold beliefs. They’re just machines that are really good at predicting what words come next based on patterns they’ve seen before.

So when you chat with one of these models, what you’re really talking to is a statistical echo of thousands of human examples. It’s not giving you “the truth”—it’s giving you what it thinks sounds right. But if you know how to use them well—by checking their work and steering them clearly—they can be powerful tools: great for drafts, ideas, code help, and more.

We’re just getting started with this technology. These models are still in their early stages. But as reinforcement learning evolves, and models start discovering their own ways of thinking—like AlphaGo once did in the game of Go—we might be witnessing the start of something truly new: intelligence that thinks in ways we never taught it.