Learning from examples, not rules

Imagine teaching a child to recognize cats. You don't hand them a definition ("four legs, fur, pointy ears") — you point at hundreds of cats and a few dogs, and gradually they notice the patterns themselves. Show them enough, and they'll spot a breed they've never seen before.

Machine learning works the same way. Instead of a programmer writing the rules, the model is shown a mountain of examples and slowly adjusts millions (or billions) of internal numbers called parameters until its guesses match reality. Each parameter is like a tiny dial; learning is the process of nudging every dial a little, over and over, until the whole system behaves.

The two phases: training and inference

There are two completely separate moments in a model's life, and confusing them causes a lot of misunderstandings.

1. Training is the slow, expensive phase where the model learns. It can take weeks, thousands of specialized computers, and enormous amounts of data. This happens once, before you ever touch the product, and is done by the model's creators.
2. Inference is using the finished model to answer your question. This is what happens every time you hit "send." It's fast and cheap by comparison — usually just a few seconds.

A useful analogy: training is the years a doctor spends in medical school; inference is the ten-minute appointment where they apply what they learned. You only ever meet the doctor at the appointment.

What this explains

• Why models have a "knowledge cutoff": training ended on a certain date, so the model doesn't natively know what happened afterward unless it can search the web.
• Why they don't learn from your chat: for most consumer tools, your conversation doesn't retrain the model. It may be reviewed later to improve future versions, but the model in front of you is frozen.
• Why answers come back instantly: the hard, slow work already happened during training.

Training is the years of study; inference is the conversation. You're always talking to a finished, frozen model — not one that's learning from you in real time.

Try this: Ask an assistant "What's today's date?" or about a very recent event. If it hesitates, guesses, or offers to search, you've just bumped into the line between training and inference.