A spam filter and a chatbot look nothing alike and run the same basic process. Once you can see that, you can look at any new AI product and make a decent guess at what it is doing.
Two products that look nothing alike: a spam filter quietly sorting your inbox, and a chatbot writing your emails for you. Underneath, they run the exact same process. Feed the system a huge pile of examples, let it find the patterns, then have it apply those patterns to something new it has not seen before. Different jobs, same underlying trick.
Nobody sat down and wrote a rule like "the word 'viagra' means spam" or "this sentence sounds professional." The spam filter worked out what spam tends to look like from thousands of examples of it; ChatGPT worked out what good writing tends to look like the same way, from a huge slice of the internet. Neither one was handed the rule. Both found it.
Keep this in mind the next time a new AI product claims to do something entirely novel. Odds are it is the same trick, pointed at a new pile of examples — not a different idea underneath.
"Learning," mechanically, means this: the system has an enormous number of adjustable internal settings, and each time it sees an example, those settings get nudged a little so its next guess is a little less wrong. Do that across a genuinely enormous number of examples and you end up with a system whose guesses have gotten good.
Here is the part people usually get backwards: that process finishes before you ever open the app. By the time you are typing into it, training is done. It is not learning from your conversation as you have it — it is running the finished pattern it already built. Ask it something today and something identical tomorrow, and you are talking to the same fixed pattern both times.
An LLM — large language model — is a machine-learning system trained on an enormous amount of text so it can predict what word is likely to come next. ChatGPT, Claude, and Gemini are all LLMs. "Large" is doing two jobs in that name at once: an enormous number of internal settings, and a training diet that is a meaningful chunk of the public internet.
"Language model" is the more important half, though, because it names the actual job: predicting text, not understanding it the way you do. That one job, done at scale, turns out to be enough on its own to hold a conversation, write code, and draft an email — which is a strange thing to sit with.
It also explains the failure mode the next module is entirely about. An LLM does not "know" facts the way a database does — there is no lookup table to check. Everything it says is a prediction shaped by patterns in its training, and that same mechanism produces a right answer most of the time and a confidently wrong one occasionally. Nothing tells it, in the moment, which one it is doing.
A spam filter and a chatbot are the same trick pointed at different piles of examples: learn the pattern, then apply it. That learning — training — is finished long before you type anything; using the tool is just running the pattern it already built. An LLM is one specific version of this aimed at predicting text, which is powerful enough to hold a conversation and also the exact reason it can sound sure while being wrong.