The language model you use every day guesses the next word, over and over. Once you see that, every strange thing about it, the confidence, the made-up facts, the forgetting, stops being a mystery.
This is the first of three quick looks, one for each of the big three from the map, and we start with the language model because it is the one you touch every day. The plan here is not to make you an expert. It is to get close enough that the model stops feeling like a mind you cannot read, then step back out. There is no mind. There is one simple move repeated at speed, and everything the model does grows out of it.
It guesses the next word
At the bottom it does exactly what the first piece said, it guesses which word is likely to come next. You give it your text, it looks at all of it, and it picks the most probable next chunk, then adds that and picks again, word after word, faster than you can read. There is no plan for the whole answer and no meaning behind it. Each step is only this, given everything so far, what usually comes next. The reason it reads so well is that it learned the shape of good writing from a mountain of it, so the words it strings together land in the shape of a real answer.
It sounds like it knows, because it learned how knowing sounds
Here is the part that trips everyone. Sounding right and being right are two different things, and the model is a master of the first one. It learned what a confident, correct answer looks like, so it produces text in that exact style whether or not the content underneath holds up. That is why it never seems to hesitate. It is not weighing the truth and landing sure. It is doing the one move, next likely word, and a smooth confident sentence is simply the most likely shape.
Sure of itself, and flat wrong
So when it is wrong, it is wrong in the same steady voice it uses when it is right. It has no inner sense of true and false, only of what fits. When the fitting thing happens to match reality, you get a correct answer. When it does not, you get an equally polished sentence that is made up, a fake source, a quote no one said, a fact that is not, delivered with the same calm. People call this hallucination, which makes it sound like a rare glitch. It is not. It is the machine working exactly as it always does, landing this time on a plausible falsehood it cannot tell apart from the truth.
It has happened to me more than once. I asked for sources on a topic I was writing about and got back a tidy list, each one with an author, a year and a title that sounded right. One of them did not exist. Not misremembered and not mislabeled, it had never existed anywhere, and it sat in that list wearing the same confidence as the real ones. That is the behavior to internalize. The tell is not in the tone, because there is no tell in the tone.
It forgets, until the app remembers for it
On its own the model has no memory. Each new chat it wakes up blank, no idea who you are or what you said yesterday, its whole world the text in the window right now. That is the raw model. The apps have moved past it. The big chatbots now keep a running set of notes on you and feed them back into every conversation, so it does remember you these days, your name, your preferences, what you worked on last time. The thing to remember is where that memory lives. It is in the app, not the model, a notebook the app keeps and hands over each time, not the model holding onto you. Which is why it can remember something wrong, or something you would rather it dropped, and why you can open that notebook and correct it.
Its knowledge is dated, not sealed
The model learned from a huge snapshot of text taken up to a certain point, then that learning stopped, so the model itself has a cutoff. Those cutoffs are more recent than people assume, and even so, on its own it cannot know what happened since. Here too the tools have moved. Many now search the web live, mid-answer, and pull in what is happening today, so it is not sealed off from the present the way it used to be. The catch is the one from a moment ago. Reaching the live web does not give it judgment. It still cannot tell a good source from a bad one, so it can serve you something fresh and false in the same sure voice. The time gap closes. The truth gap does not.
The core of it
That is a lot of moving parts, and here is the good news. You can forget most of it. The memory, the web access, the cutoff, those are details that will shift with every update, and this whole closer look is a snapshot, not a manual. The full version of this, how these models work all the way down, is its own project for another day. One thing survives every update and carries the rest. It predicts, it does not know. Remember that single line and you will read this tool right for years, sure of itself whether it is right or wrong, more capable than the horror story and still unable to vouch for a word it says.
That is the tool you use every day, seen clearly. Next we stay in close and look at the tool carrying all that anger from a couple of pieces back, the one that turns your words into a picture.
Next up: A Closer Look at Image Generators →