AI Explained in Two Exercises

Most people think AI is either magic or hype. It's neither. It's a different way to build software.

Let me show you what I mean with an exercise a mentor taught me twenty years ago. It's still the clearest way I know to explain this.

The first exercise

Imagine I hand you a blank piece of paper and ask: write down the steps you take to calculate your taxes.

You could do this. It might take a few minutes, but you'd write something like:

1. Gather all your income documents (W-2s, 1099s, etc.)
2. Add up your total income
3. Subtract the standard deduction (or itemize if that's larger)
4. Look up your tax bracket based on the result
5. Calculate your tax by applying the rate for each bracket
6. Compare that number to what was already withheld from your paychecks
7. If you paid more than you owe, you get a refund. If less, you owe the difference.

Now take that piece of paper and hand it to a software engineer. Or code it up yourself in Excel, which many people do. Either way, you end up with software that calculates taxes.

This is how software has been built since the beginning. You describe what you want done. Someone (or you) converts that into instructions a computer can follow. The computer runs those instructions. Done.

The second exercise

Now take another blank piece of paper. Write down the steps you take to recognize a face.

Your mother's face. Your best friend's face. A colleague you've worked with for years.

Go ahead. Write down the steps.

You can't.

You do it instantly, without thinking. A face appears and you just know who it is. But you have no idea how you do it. There are no steps to write down.

This is the problem machine learning solves

There are things we can do that we cannot explain. Recognizing faces. Understanding speech. Reading handwriting. Identifying a song from a few notes. We do these effortlessly, but we can't write down the rules because we don't know what they are.

Traditional programming can't help here. If you can't describe the steps, an engineer can't code them.

Machine learning takes a different approach. Instead of writing instructions that solve the problem, you build a model and show it examples. Thousands of examples. This is Maria. This is John. This is Sarah. The model finds patterns in the data and learns to make the distinction on its own.

Nobody tells the model what to look for. Nobody writes rules like "if the distance between the eyes is X and the nose width is Y, then it's Stefanos." The model figures it out from the examples.

Why this matters

This is the fundamental difference between traditional software and AI. Traditional software follows instructions you write. AI learns patterns from data you provide.

This is also why AI is powerful for some problems and useless for others. If you can write down the steps, traditional software is simpler, cheaper, and more predictable. If you can't write down the steps but you have lots of examples, that's where AI shines.

The ten First Principles of AI I wrote about previously all build on this foundation. Data quality matters because the model learns from examples. Generalization matters because the model has to work on examples it hasn't seen before. The right model for the right problem matters because different approaches learn different kinds of patterns.

It all starts here: AI is not magic. It's a different way to build software, designed for problems where we know the answer but can't explain how we get there.