This series explains how large language models work in plain language. Before the mechanics, I want to start where I started.
A friend mentioned the US military was using AI, and I had not even heard of it.
In 2022, a friend came back from the US and told me the military was taking up AI. I was taken aback. Until that sentence, I was not even aware a technology called artificial intelligence existed. I read articles afterwards, hit a lot of conjecture about job loss and the death of thinking, read the negative sentiment first, and stopped there. The phase faded. I went back to work.
No fear spiral. No late-night rabbit hole. A sentence, some noise, then back to my desk. If you are reading this and you have heard of AI but cannot get past the jargon long enough to form a real opinion, that is where I was in 2022. This is for you.
If that sounds familiar, you are not behind. You are exactly where the gap lives.
A Pew Research survey from 2023 measured this gap. Just 14% of all U.S. adults say they have used it for entertainment, to learn something new, or for their work. That means roughly 86 out of every 100 people who have heard the name have not opened the tool for any purpose, let alone understood what it does.
I am not citing that to make the gap sound insurmountable. The number tells me the barrier is not motivation. It is the jargon itself. "Large language model." "Transformer architecture." "Hallucination." "Fine-tuning." Words that make anyone outside computer science feel they need a degree before they can have an opinion.
Four posts. No prerequisites. The goal is AI literacy for professionals who need to make real decisions about these tools, not pass an exam about them.
Your phone's autocomplete, trained on every book and article ever written, is the closest real-world picture of what an LLM does.
You have used your phone's keyboard today. Every time you type a word, your phone proposes the next one. It is not looking up facts. It is reading your pattern and offering the most probable continuation.
A token is roughly a word or part of a word, the unit the model reads and predicts at each step. An LLM does exactly what your phone does, except it learned from most of the written internet rather than your messages. Billions of documents, compressed into a system with one job: given everything before this point in the text, what comes next?
That is the whole engine. Same mechanism, different scale.
OpenAI's own technical report names the mechanism, and it is not as mysterious as the jargon suggests.
The mechanism is called next-token prediction. The people who built these models describe it plainly.
The GPT-4 technical report from OpenAI puts it in a single sentence. "GPT-4 is a Transformer-based model pre-trained to predict the next token in a document."
Not retrieve. Predict. The distinction is the whole story.
Predicting text and retrieving facts are as different as guessing what someone is about to say versus looking it up in a library. A search engine finds the answer. An LLM produces text that looks like the answer would look. Most of the time, on well-covered topics, those two things land close together. That is why LLMs feel useful. The gap opens when the model is in territory where plausible-sounding text and accurate text are not the same thing.
An OpenAI article on instruction-following described the original training objective. "GPT‑3 is trained to predict the next word on a large dataset of Internet text, rather than to safely perform the language task that the user wants." That is the gap between what the model was trained for and what a professional needs from it.
The GPT-4 technical report is direct about this. "GPT-4 can also be confidently wrong in its predictions, not taking care to double-check work when it's likely to make a mistake." The model does not check itself. It produces the most probable continuation, delivers it in fluent prose, and stops. Confidence is a product of the generation, not a signal of accuracy.
You do not need to understand the engine to drive, and that objection is more reasonable than it sounds.
The obvious pushback: you drive a car every day without understanding the combustion cycle. You use a search engine without knowing how it indexes pages. You use Excel formulas without knowing how the calculation engine works behind them. And you are productive in all three.
That is a fair objection. For many day-to-day tasks, treating an LLM as a black box produces adequate results. Routine emails, meeting summaries, quick definitions - those all work without knowing how the engine runs.
Here is where the comparison breaks. The moment an output matters, the moment it is heading to a customer or informing a decision, you need to be behind the wheel in the actual sense: able to judge the output, not just receive it. Knowing that the model predicts rather than retrieves is the one piece of context that makes that judgment possible. Without it, you either verify everything, which is too slow, or you verify nothing, which is dangerous. The mechanics are not interesting for their own sake. They are load-bearing for the judgment call.
Gemini recommended a phone that never existed, and it took me a moment to understand why.
I was planning to buy a new phone. I gave Gemini my requirements and intended use and asked it to find the best deal. It came back with a specific phone model, complete with model numbers, specifications, and price comparisons that sounded fully authentic. I checked the original source myself. The phone had never existed. Gemini had fabricated the recommendation, start to finish, with the confidence of a retailer reading from a catalogue.
Before I understood the mechanism, I was confused and annoyed. Once I did, it made complete sense. The model was doing exactly what it is designed to do. It predicted the most plausible phone recommendation for my requirements, token by token, with no internal registry of what is on sale. The same mechanism that produces a perfect paragraph also produces a perfect-sounding model number that was never manufactured. The confidence was a feature of the generation, not evidence of accuracy. The lesson is mechanical, not motivational.
The danger is not learning that it predicts. It is forgetting, once, when the stakes are high.
Once you understand that AI hallucinations are a mechanical outcome, not a glitch, the tempting overcorrection is to verify every output manually. That is not a usable workflow.
The more realistic risk runs the other way. You absorb the rule. You apply it when the output is routine. Then one day the model returns something that sounds authoritative, fully formed, exactly like what you needed to hear, and you wave it through. That is the moment the rule needed you most.
A fluent, well-formed AI response is the most credible-looking thing the model produces. It is also the thing that most needs a second look. The rule is not "distrust AI." The rule is this: the confidence of the output tells you nothing about its accuracy. Keep that rule live especially when the text is convincing.
Welcome to the foundation. Three more posts follow from here, and no equations are required for any of them.
The next post covers training vs. fine-tuning vs. RAG vs. prompting. After that, tool selection. Then a 30-day roadmap for putting this into practice.
Does the jargon still feel like a ten-headed devil? It did to me, right up until I understood what the engine does. Welcome to the foundation. No equations required.
