In our previous Parts, we discovered that AI begins as a Base Model—a giant "Internet Simulator" that has read trillions of words but doesn't quite know how to be a helpful assistant. If you ask a Base Model "What is the capital of France?", it might simply respond with "And what is the capital of Germany?" because it is mimicking the format of a school quiz it saw online.
To transform this "Know-it-all Professor" into the Virtual Assistant we use today, it must go through a specialized "socialization" phase called Supervised Fine-Tuning (SFT).
1. Training the Assistant (Supervised Fine-Tuning)
To teach the AI how to behave, engineers set aside the chaotic documents of the internet and replace them with a high-quality dataset of conversations.
Instead of letting the AI guess the next word from a random blog post, it is given "ideal examples." It studies hundreds of thousands of pairs of User Prompts and Assistant Responses. Through this process, the AI learns that when a human asks a question, its job is to provide a direct, helpful answer rather than just completing a text pattern.
SFT (Supervised Fine-Tuning): Also known as "instruction tuning," this is the process of teaching a Base Model to act as an assistant by training it on human-written conversation samples.
2. The Golden Rules of Human Teachers
The experts who write these ideal responses are known as human labelers. They follow strict "Labeling Instructions" to ensure the AI adopts a consistent persona.
The three core principles they teach the AI are:
- Helpful: Solve the user's problem directly and clearly.
- Truthful: Provide accurate information and avoid making things up.
- Harmless: Refuse requests to engage in dangerous or illegal activities.
3. The Secret "Script" Behind the Chat
To help the AI distinguish between your words and its own, engineers use Special Tokens that are invisible to the user.
Think of these as stage directions in a play. Symbols like user and assistant act as markers that tell the model: "The human has finished speaking; now it is your turn to act as the helpful assistant".
4. Who Are You Actually Talking To?
When you chat with an AI, you aren't talking to a conscious mind. You are interacting with a statistical simulation of a highly-skilled human labeler.
The AI is essentially asking itself: "Based on my training, what would a professional assistant following these rules write next?". It is an instantaneous remix of the "ideal behaviors" it learned during its SFT phase.
Pro-Tip: Because the AI is simulating a professional assistant, the way you speak to it matters. If you provide a clear, detailed, and professional prompt, the AI is more likely to "enter the role" of a professional and give you a higher-quality result.
Summary of Part 3: The SFT stage takes the raw power of the Base Model and channels it into a useful, polite assistant persona.
In Part 4, we will explore the Context Window—the "working memory" of your AI. You’ll learn why "pasting" information directly into the chat can make your assistant significantly smarter.