AI, Machine Learning, and LLMs — The Big Picture for Non-Developers

These days you hear about AI everywhere. But because words like AI, machine learning, LLM, and generative AI get mixed together, it’s easy to lose track of how they differ. When “can’t we just do this with AI?” comes up in a meeting, it’s often unclear exactly what that AI refers to.

In this post, I’ll sort out the relationship between these words as a big picture, and unpack what today’s tools do well and poorly, without any code. Rather than digging deep into the internals, I’ll focus on drawing a map that lets a non-developer join the conversation.

Inside AI is machine learning, and inside that, LLMs #

Let’s start with how the three words relate. They aren’t rival terms set side by side; they’re nested, like smaller circles inside a bigger one.

The biggest circle is AI, or artificial intelligence. It’s a broad term for the whole field of getting machines to do what human intelligence used to do. One branch inside it is machine learning. Instead of a person spelling out every rule, you give it lots of data and let the machine learn patterns on its own. And within machine learning, a model trained at an enormous scale to handle language is an LLM, or large language model. The chatbots we use are exactly these LLMs.

In short, AI is the broadest field, machine learning is a core method that makes it up, and an LLM is the language-focused model built with that method.

An LLM picks the next word by probability #

The principle behind the smart-looking answers an LLM produces boils down, surprisingly, to a single sentence: it looks at the text so far and picks the most plausible next word by probability, stringing them together. Repeat this very fast and a long passage takes shape.

Because it learned patterns of “this kind of word usually follows that kind of word” while training on vast amounts of text, it produces sentences as natural as a person’s. How this prediction works is covered in more depth in How LLMs Predict the Next Word. What to remember here is that, rather than answering by understanding meaning, an LLM is very good at stringing together plausible words.

What it does well and poorly splits along that line #

Knowing this principle, the strengths and weaknesses of today’s tools come into view naturally.

What it does well is the realm of language. It’s good at summarizing, translating, drafting, polishing awkward sentences, and sketching the overall shape of code. The more a kind of task has piled up in the world as a pattern, the better it does.

The weak spots are just as clear. The one you run into most often is hallucination — the phenomenon of confidently making up things that aren’t true as if they were facts. Since stringing together plausible next words is its native behavior, it arises from prioritizing a smooth sentence over checking what’s true. It also doesn’t know information newer than its training cutoff, and it often gets exact calculation or rigorous logic wrong. So results an LLM produces should go through a step where a person checks the facts, rather than being trusted as is.

Why this makes work easier for non-developers #

• You put the tool in its place. Hand it what it does well, like drafting or summarizing, and have a person verify work that needs fact-checking or exact numbers, and you cut wasted effort.
• You don’t get fooled by hallucinations. Knowing that a plausible answer isn’t always a correct one builds the habit of double-checking anything important.
• You hold realistic expectations. When someone asks “can’t we just do this with AI?”, you can weigh what’s actually possible against what’s still hard.

Wrapping up #

Today we sorted out the relationship between AI, machine learning, and LLMs as a big picture. AI is the broadest field, machine learning is the method of learning patterns from data, and an LLM is the language model built with that method. An LLM is good at stringing together the next word by probability, so it handles language tasks well, but it carries clear weaknesses like hallucination. Knowing these strengths and weaknesses as you use it is the first step to wielding the tool well.

If you’d like to know more about how an LLM picks the next word, read How LLMs Predict the Next Word; if you’re curious about who builds and runs these models, read What Do Developers Actually Do.