How models work

How AI learns (and what it can’t)

Training, cutoff dates, and why it doesn’t know your business unless you tell it.

6 min readPlain English

Read it in

In one sentence

A model does its learning before you ever use it, in training, so it knows nothing after its cutoff date, and little or nothing about your business, unless you tell it.

Think of it like…

Imagine a new hire who read an enormous library, then left it on a certain date to come and work for you. They know a lot about the world up to that day, nothing since, and nothing about your customers or prices until you brief them. Briefing them doesn’t change what they read; it just gives them notes to work from.

How it works

Training happens in stages, long before you type anything. By the time you use a model, the learning is done.

1ReadThe model is shown an enormous amount of text (web pages, books, code and more) and practices predicting the next word, over and over.
2AdjustEach wrong guess nudges a vast number of internal settings slightly. Over many rounds, those settings come to capture grammar, facts, styles and ways of reasoning.
3CoachPeople write example answers and rate the model’s replies, teaching it to follow instructions and be helpful. It also practices on problems with checkable answers, like code and math.
4FreezeIts training data ends at a cutoff date. The finished model is released, and its knowledge stays fixed until its maker trains a new version.

Chatting doesn’t retrain it. Depending on your plan and settings, though, your conversations may be kept and used to train future versions.

Good at

General knowledge up to its cutoff
Common business writing and formats
Using what you give it in the chat

Bad at

News since its training ended
Your prices, policies and customers
Saying where it learned something

Where you’ll see it at work

The out-of-date answer. An online shop asks about product labeling rules and gets them as they stood at the model’s cutoff. Check the official source.
Briefing it once. A plumbing firm saves its services, service area and tone in custom instructions, so every draft starts from the right facts.
An answer bank. A bakery keeps 20 real customer questions and approved answers in a file the assistant reads before drafting replies.

Myth vs reality

Myth: It learns from every chat, so it will get to know your business.

Reality: The model doesn’t change while you chat. Some apps save notes about you, but your business knowledge lives in what you give it: instructions, files and examples.