What a Large Language Model Actually Is

Over deze cursus

The machine behind ChatGPT, Claude and Gemini, taken apart and handed to you one moving part at a time. No hype. No doom. You turn every dial yourself.

Het cursusmateriaal is Engelstalig.

Wat je krijgt

• 1 PE-uren voor je PE-portfolio
• Online, in je eigen tempo
• Een eindtoets, en daarna je PE-certificaat

Modules

1. What is a large language model

   Behind ChatGPT, Claude and Gemini sits one small move, repeated fast. The model reads the words in front of it and guesses the next word. Then it does the same thing again. Stack up enough of those guesses and you have a whole answer. That’s it. That is all those fancy AI models really are.

   8 min

2. How it's trained

   So where do these guesses come from? A model is built by reading: more text than a person could get through in a thousand lifetimes, from web pages, books, code repositories, Reddit threads and Wikipedia. What it keeps is the patterns, and the patterns it saw most are the ones it does best.

   10 min

3. What made it possible

   People have chased talking computers since the 1950s. So why did it click around 2020, and not decades earlier? Three things had to show up together: a smarter way to read a whole sentence at once, chips fast enough to crunch it, and enough writing on the internet to learn from. Pull any one of them out and there is no ChatGPT.

   10 min

4. A predictor, not a database

   You have seen what it reads and how it learns. Now the thing itself: the base model, the box in the middle of the diagram. Type a question, hit enter, and it does not go and look anything up. It just guesses the next token. Then the next. Then the next, until the answer is finished. That is the whole engine, and the G in GPT: it generates. Every reply you have ever gotten was built exactly this way, one guess at a time.

   10 min

5. Pretrained and frozen

   Two things people take for granted about AI are both wrong. The first: that it learns from your chats. Hear “self-learning” and it is easy to assume the model picks things up on the spot, so people spend real time correcting it mid-conversation, get annoyed when it repeats the same mistake tomorrow, and write it off as a dumb model. The second: that it knows what happened in the world last week. Neither is true, and the reason is the same for both. Training happens once, in advance. Then the weights freeze, and after a few more steps, like testing and safety checks, the model ships. From then on the weights do not move while you use it. That is what the P in GPT stands for: pretrained.

   10 min

6. From base model to assistant

   After its first big training run the model can predict words, but it is still raw. As an assistant it is close to useless: ask it a question and it just carries on writing more text at you. So it goes through a second, much shorter round called post-training, and this is where it learns to act like a helpful assistant instead of an autocomplete. It is also where the eager, agreeable streak is born. One more thing shapes every reply, and it has nothing to do with training: how much the model can actually see at once. That is the context window, its short-term memory for the conversation, and you will fill it up yourself further down.

   10 min

7. Why it feels intelligent

   You have seen the whole machine now, and it still feels like someone who understands you. That feeling is not an accident. The last chapter trained it to sound helpful and sure of itself, and that same confidence is exactly what makes it so easy to trust when it is wrong. This chapter lets you watch it fail on purpose, so the feeling loosens its grip.

   10 min

8. Final assessment

   10 min