Trained, Not Programmed

Lesson 3 · taught by Pip

Here's the plain idea: older software follows rules a person typed in by hand. Modern AI doesn't. Instead, it learns patterns from mountains of text and examples — the way a person picks things up by experience rather than from a rulebook.

That single difference explains almost everything that feels surprising about AI: why it's so flexible, and why it's never perfectly predictable.

The rulebook vs. the apprentice

Think of two new hires at a front desk.

The first gets a binder: "If a guest asks about parking, read paragraph 4." Every situation needs its own page. Ask something the binder didn't anticipate, and the hire freezes. That's traditional programming — a human spelled out every rule in advance.

The second hire never gets a binder. Instead they shadow your best receptionist for months, listening to thousands of real conversations. Nobody hands them a script. They absorb the patterns — the tone, the judgment, the way a good answer sounds. Soon they can handle a question no one ever wrote down.

That second hire is how modern AI is built. It was "trained" by reading an enormous amount of writing and learning the patterns inside it. No one programmed a rule for every question, because no one could — there are too many questions in the world.

Why this cuts both ways

The upside is flexibility. Because it learned patterns instead of memorizing scripts, AI can answer phrasings it has never seen, in a tone that fits, on topics nobody explicitly prepared. That's why a single assistant can field a thousand differently-worded versions of "are you open Sunday?" without anyone writing each one out.

The trade-off is predictability. An apprentice who learned by absorbing patterns will usually be excellent — and occasionally, confidently, wrong. It isn't looking up a guaranteed-correct rule; it's producing the response that best fits the patterns it learned. Most of the time that's right. Sometimes it sounds just as sure when it's off.

This is why "trained, not programmed" matters for your business. You can't fully eliminate the surprises by demanding tighter rules — that's not how the thing works. What you can do is shape what it learned from. A good business AI is trained and grounded on your facts — your hours, your services, your policies — so its patterns lean on your truth instead of guesses. (That grounding is exactly the work behind a tool like Portmint's Training & Testing setup before a bot goes live.)

Your turn

Picture the strangest, most oddly-worded question a real customer has ever asked you. A rulebook system would need someone to have written that exact case in advance — a trained assistant can take a fair swing at it. Which would serve your customers better at 9pm on a Saturday?

Next, we'll look at why this trained-from-examples approach makes AI sound so remarkably human. 🔦