What Is AI, Really?

Strip away the hype: AI is software that learns patterns from data instead of being told every rule.

Building with AI · Lesson 1 · 7 min read

Your friend says “AI is going to take everyone's jobs,” your uncle says “it's just autocomplete,” and an ad promises an “AI” that will make you rich. They can't all be right. Before you can build with AI — or judge what's real — you need a clear, jargon-free answer to one question: what actually is this thing? Hold that question; by the end of this lesson you should be able to cut through any of those claims.

Traditional software follows rules. AI learns patterns.

For 70 years we built software by writing explicit rules: if the user clicks this, do that. A programmer foresees the situations and codes the response. That works great for a calculator or a spreadsheet — the rules are knowable.

But how do you write the rules for “is this photo a cat?” You can't list every pixel pattern. So instead of writing rules, we show the machine thousands of examples and let it find the patterns itself. That is machine learning — the engine under almost everything called “AI” today.

Learning from examples is the whole trick

A machine-learning model is trained by adjusting millions of internal numbers until its guesses match the examples it was shown. Nobody hand-codes “a cat has pointy ears.” The pattern is discovered and stored as numbers. That's why modern AI can do things we could never write rules for — and also why it can be confidently wrong in ways traditional software never is: it learned a pattern, not a fact.

Spam filter, two ways:
• Rules: “if the email contains the word FREE, mark spam.” Easy to fool, misses new tricks.
• Learning: show it 100,000 emails already labeled spam / not-spam. It discovers subtle patterns (odd sender, link soup, urgency) on its own — and keeps working on spam it has never literally seen.

Optimism, with eyes open

AI is already a genuine superpower: it tutors, translates, writes code, spots tumors, and drafts the boring first version of almost anything. Used well, it gives an ordinary person leverage that used to require a whole team.

The honest part: because it learns from human data, it can inherit human bias; because it predicts rather than knows, it can hallucinate confident nonsense; and powerful tools can be misused. None of that is doom — it's a to-do list. The whole point of learning to build with AI well is to get the benefits while handling the risks on purpose, not by accident.

Ask “how does Netflix recommend shows?”
1. Nobody wrote “if they watched a heist movie, suggest another heist movie.”
2. Netflix shows a model millions of viewing histories.
3. The model learns patterns like “people who finished X tend to enjoy Y.”
4. It predicts what you might like from what people similar to you watched.
That's learning-from-examples doing a job no human could write rules for — and if your taste is unusual, its confident guesses can be way off. Same engine, same strengths, same failure mode.