Strip away the marketing and AI is a plain idea: a set of techniques for getting computers to accomplish tasks. That's it. The trouble is that the word has been stretched, redefined, and oversold so many times that it now means almost nothing without context. So let's rebuild a definition you can actually reason with.

The AI Effect#

There's a strange pattern in how we talk about AI. Once a computer reliably does something that used to seem impossible, people stop calling it AI. Optical character recognition, spam filtering, route planning, beating humans at chess, all of these were "artificial intelligence" until they worked, at which point they became "just software." This is the AI Effect: the goalposts move the instant the goal is scored. It's worth knowing about, because it means the label "AI" tracks our sense of mystery more than any technical boundary.

Narrow AI is all we have#

It helps to separate two very different things people mean by AI:

• AGI (artificial general intelligence) is hypothetical general intelligence, a system that could reason across any domain the way a person can. It does not exist. Nobody has built it, and despite confident headlines, we are not on the verge of it.
• Narrow AI excels at one task or a narrow band of tasks. This is all that exists today, every bit of it, including the chatbots that feel general because they can talk about anything.

That last point trips people up. A modern chatbot feels general because language touches everything, but it's still narrow AI. It's extraordinarily good at one thing, producing plausible text, and that one capability gives the illusion of broad intelligence.

What an LLM is really doing#

Large language models are narrow AI specialized for generating text from prompts, and they're remarkably flexible at it. But it's worth being precise about what's happening under the hood, because the precision changes how you use them.

Their "reasoning" is best understood as statistical reasoning over text. It's pattern recognition and probabilistic chaining, predicting what tends to come next given everything so far, rather than conceptual understanding of the world. A few consequences follow directly:

• No world model. The model doesn't carry an internal map of how reality works. It has a map of how language tends to flow.
• Error-prone in multi-step logic. Because each step is a probabilistic continuation rather than a verified deduction, long chains of reasoning can drift off course in ways that look confident and read fluently.

None of this makes LLMs useless, far from it. It just means you should treat their output as fluent, pattern-driven generation that needs verification, not as the conclusions of something that understands.

They don't make decisions#

A related misconception is that these systems make decisions. They don't, at least not in the way we mean when a person decides. An LLM has no goals, no desires, no agency, and, by default, no memory. It reacts to input. Give it a prompt and it produces a continuation; that's the whole interaction.

What it can do is genuinely useful: it can suggest decisions and weigh trade-offs when you hand it the criteria. Ask it to compare two architectures against cost, latency, and team familiarity, and it will lay out a reasoned-looking comparison. But the deciding is still yours. The model proposes; you dispose. Keeping that boundary clear protects you from over-trusting a tool that has no stake in being right.

Machine learning, the practical core#

The most practical slice of AI for everyday engineering is machine learning. The core idea is a reversal of normal programming. Instead of hand-writing every rule, you let the system learn patterns from data. You show it many examples, and it infers the rules implicitly.

This matters because some problems have too many rules, or rules nobody can fully articulate. How do you write explicit if statements to tell a cat from a dog in a photo, or to flag a fraudulent transaction among millions of legitimate ones? You can't enumerate those rules cleanly. ML lets you skip the enumeration and learn from examples instead.

A decision lens#

So when should you reach for ML versus ordinary code? A clean heuristic:

• Use conventional, rules-based code when you know and can cleanly define all the rules. If the logic is fully specifiable, write it out. It'll be faster, cheaper, more predictable, and far easier to debug than any model.
• Use machine learning when the variables or rules are huge in number, unknown, or impossible to specify. When you genuinely can't write the rules down, learning them from data is the move.

That lens keeps you from the two common mistakes: reaching for ML to solve a problem a simple if statement would handle, and trying to hand-code rules for a problem that's fundamentally too fuzzy to enumerate. Knowing what AI actually is, narrow, statistical, and reactive, is what lets you place that line correctly.