What is AI?

This article builds on:

Most people treat "artificial intelligence" like it's some new, exotic thing. It isn't. I'd argue you've been using AI your entire life.

Here's my definition: artificial intelligence is human-like intelligence that exists outside a human brain.

That's it. And by that definition, AI is ancient.

Think about a piece of paper and a pencil. You can solve math problems on paper that you can't solve in your head. The paper isn't alive. It isn't thinking. But it holds intermediate results and keeps track of carries, letting you work through logic you'd otherwise lose track of. It's doing cognitive work outside your brain.

An abacus does even more. Slide the beads and it performs arithmetic for you. A shopkeeper in 1200 AD didn't need to multiply in her head. Was that artificial intelligence? By my definition: yes. She offloaded a cognitive task to a non-brain thing, and the non-brain thing handled it.

A book is externalized knowledge. A library is externalized collective memory. A calculator is externalized math. A spreadsheet is an externalized financial analyst, at least for the simple stuff.

The whole history of tools is the history of offloading cognitive work to things that aren't brains.

In the summer of 1956, a group of researchers got together at Dartmouth College and coined the term "Artificial Intelligence." Their definition was much narrower than mine: build machines that could do things we'd call intelligent if a person did them. Play chess, prove theorems, understand language, learn from experience.

The early work was all rule-based. Programmers hand-coded the logic: if the opponent's queen is threatening your king, move the king. They called these "expert systems": you'd interview a human expert, write down all their rules, and teach the computer to follow them.

It worked, kind of. These systems could play decent chess and diagnose certain diseases. But they were brittle: every edge case needed a new rule, and the programmer had to anticipate them all. It turns out most of what humans know is really hard to express as explicit rules. Try writing down instructions for recognizing a face, or understanding sarcasm, or knowing when someone is about to cry. You can't, because you don't actually know how you do those things. You just... do them.

The field went through cycles. A burst of excitement and funding, then the realization that the hard problems were way harder than expected, then funding dried up and researchers scattered. These downturns got a name: AI winters. There were two big ones, in the mid-70s and late-80s. Each time, the surviving researchers quietly kept working and came back with better ideas.

The big idea that pulled AI out of its second winter was simple: stop writing rules. Let the machine figure them out.

Instead of a programmer coding "if pixel pattern X then it's a cat," you show the machine ten thousand photos labeled "cat" and ten thousand labeled "not cat," and it finds the patterns itself.

This is machine learning. The programmer doesn't need to know the rules, or even be able to explain how to recognize a cat. They just need examples.

But how does the machine actually "discover" the rules? The trick is pretty intuitive once you see it.

Imagine you're standing in hilly terrain, blindfolded, trying to reach the lowest point. All you can do is take a step, feel whether you went downhill or uphill, and decide where to step next.

You'd probably do something like this: take a step to the right. Did you go downhill? Great, keep going right. Uphill? Turn around. As the slope gets flatter, take smaller steps so you don't overshoot the bottom. Eventually every direction feels flat, and you stop. You've found a valley.

That's gradient descent, the core technique behind nearly all modern machine learning. Let's play it. You're fishing a murky lake, too cloudy to see the bottom, trying to find its deepest point. Drop your line where you click and it reads the depth right there. You get ten casts. Find the deepest spot:

Now let's try that in two dimensions. Same game, but from the top of the whole lake. You can move anywhere on the surface, not just left and right, so there's a lot more water to search with the same ten casts:

In a real machine learning model, the "terrain" is the model's error: how wrong its predictions are. The "position" is the setting of all the model's internal dials (called parameters or weights). Training means twisting those dials, checking how wrong the model is, and twisting again, over and over, to get the error as low as possible.

Each check of the error is a cast of your line. It costs time and compute, so you want to find the bottom without wasting casts. That's why you follow the slope downhill instead of guessing randomly.

You just did this in one dimension, then two. A real AI model does it in millions. A cat-vs-not-cat classifier has millions of those internal dials, and gradient descent tunes them all at once, feeling downhill through a space with millions of directions instead of two. You can't picture a million dimensions, but the math is identical to the lakes you just fished.

This approach has been quietly running the world for twenty years. Spam filters. Netflix recommendations. Fraud detection at your bank. Every Google search. Machine learning was making decisions about your life long before anyone started talking about ChatGPT.

But each of those models does one thing. A spam filter can't recommend movies. A fraud detector can't filter spam. They're narrow specialists. What changed recently is that people figured out how to train a general-purpose model that can handle almost any language task you throw at it.

When someone in a boardroom says "AI" in 2026, they're almost never talking about abacuses or spam filters. They mean one of three things:

1. Large language models (LLMs) — the technology behind ChatGPT, Claude, Gemini, and the rest. You type a question, it types back an answer. They're trained with the same gradient descent described above, at a staggering scale: trillions of words of training data, billions of internal dials, months of training on thousands of specialized chips.
2. Generative models — systems that create images (Midjourney, DALL·E), video (Sora, Veo), music, and code. Same underlying ideas, different outputs.
3. Agents — LLMs hooked up to tools. An LLM by itself can only produce text. Give it the ability to search the web, run code, read files, send emails, and call APIs, and it can actually do things in the real world. The LLM thinks about what to do, picks a tool, uses it, looks at what happened, then decides what to do next. Loop until the task is done.

These feel qualitatively different from earlier machine learning. The old models were specialists you'd never want to talk to. LLMs are general-purpose and conversational: you talk to them in plain English. That's why they've captured the public imagination in a way spam filters never did.

Understanding these layers, from ancient tools to rule-based systems to machine learning to LLMs and agents, gives you a much better frame for evaluating what AI can and can't do for your company.

When a vendor tells you their product is "AI-powered," you can now ask: what kind of AI? A hand-coded rules engine from the 1980s with a chatbot UI bolted on? A narrow ML model trained on a specific task? Or something built on top of a modern LLM?

You'll know what questions to ask when an employee says "we should use AI for this," and you'll have a better read on whether a competitor's "AI initiative" is substance or marketing.

If you want to go deeper on how LLMs actually work (what's happening under the hood when you type a question into ChatGPT), I wrote a companion piece that builds the intuition from scratch: