How models are made · Part 1 — The AI Dictionary

The first, most expensive phase: teaching a model language by having it predict the next token across trillions of tokens.

Pre-training is where a model absorbs grammar, facts, reasoning patterns and coding skills from an enormous dataset. It takes thousands of GPUs for weeks or months. The result — a "base model" — is knowledgeable but not yet a helpful assistant; it just continues text.

Think of it like: Reading an entire library before anyone teaches you manners or how to answer questions.

Extra training on a smaller, focused dataset to specialize a model's style, format or skill.

Fine-tuning takes an already-trained model and continues training it on curated examples — customer-support transcripts, legal drafting, a company's tone of voice. It changes the model's default behavior without starting over.

Think of it like: A general doctor doing a residency in dermatology.

Training a model on human ratings of its answers so it becomes more helpful, honest and safe.

After pre-training, people compare pairs of model answers and pick the better one. Those preferences train a "reward model," which then steers the main model toward answers humans prefer. It's a big part of why assistants are polite, follow instructions and decline harmful requests.

Think of it like: A comedian refining their set based on which jokes got laughs.

Training a small 'student' model to imitate a big 'teacher' model — most of the smarts at a fraction of the cost.

In distillation, a large, expensive model generates answers (sometimes including its reasoning), and a smaller model is trained to reproduce them. The student learns the teacher's behavior far more efficiently than it could learn from raw data alone.

Think of it like: A master chef training an apprentice. The apprentice never reads the whole library of cookbooks — they learn by copying the master's dishes.

Example: "Mini", "Flash" and "Haiku"-style models are typically distilled or trained with help from their bigger siblings.

Training data generated by AI models rather than written by humans.

As high-quality human text runs short, labs increasingly generate training data with models: worked math solutions, code with tests, rewritten documents, simulated conversations. Done carefully — with filtering and verification — it's a powerful way to teach specific skills.