Training. At the beach it’s the word that reassures: “AI keeps learning, the more you use it the better it knows you, it updates itself.” It sounds so reasonable that nobody checks.
Last time we completed the toolkit: the message, the tools, the loops. Now it’s time to open the bonnet and look at the engine.
Training (and its expiry date, the cut-off).
The myth tells of an AI that grows with you. The reality is cooler: training is an event in the past. The model read an enormous slice of the world, once, at monstrous cost. Then it sealed itself shut.
Imagine a student who read a vast library, right up to graduation day. Then closed the books for good. That graduation has a date: the cut-off. Everything that happened after, as far as they’re concerned, doesn’t exist.
The AI model you use today is that photograph. Blind in two directions. Forward in time: yesterday’s news wasn’t there. And toward you: your company, your clients, your style. Unless you’re a public figure, you’re not in the library. It hasn’t forgotten. It never knew.
Here’s the part nobody mentions at the beach. Where it doesn’t know, the model doesn’t stay silent — it fills in. Ask it about your company and it will answer with something plausible, fluent, invented. We’ll come back to this, because it deserves an article of its own.
“But the new version just came out — it’s learned!” No: someone has taken a new photograph, with a new date. Between one shot and the next, the model you’re using stays still.
The takeaway: the model knows everything about the world up to a date, and nothing about you. Your job, to use AI effectively, is not to leave it the gap.
One legitimate question remains: so why does ChatGPT seem to remember me? Good question. It’s the most elegant trick in the whole architecture.