Just

I’ve been watching some videos on how LLMs work. I understand most of the conceptual stuff, enough for the work I do, and I spend some percentage of my time demystifying AI. Lately I’ve been noticing how often demystification works by making the thing sound smaller.

I know this move from writing. People say writing is just hard work, or just getting words on the page, or just learning some craft and keeping your butt in the chair. Most of that advice is useful. I give versions of it myself. People do need to stop waiting for inspiration sometimes, and a bad draft gives you something to work with.

Keep reducing writing, though, and eventually you get a strange account of it. Learn some craft. Produce enough words. Revise them. Repeat. At some point you have explained writing in a way that makes it hard to explain why some writing is much better than other writing, why some sentences stay with us, or why experienced writers can know exactly what they want to do and still fail to do it.

The same thing seems to happen with LLMs. I keep running into versions of these.

1. An LLM is just advanced autocomplete.

2. An LLM is just a prediction machine.

3. An LLM is just probabilistic reasoning.

4. An LLM is just pattern matching.

5. An LLM is just statistics.

6. An LLM is just a technology.

Most of these descriptions point to something real. Someone who thinks ChatGPT searches a giant database for answers needs to understand token prediction, and someone who imagines a little person inside the machine probably needs a lot more mechanism.

The word just does extra work. “The model predicts the next token” describes something important about how the system works. “The model is just autocomplete” usually carries a larger claim about how seriously we should take the system, how complicated it is, and whether anything interesting is happening.

Writing itself is just a technology, if we want to use that level of description. Libraries store and retrieve information. Mathematics manipulates symbols. These descriptions are true enough, and they still leave most of the interesting questions open.

Weather gives me a useful analogy. Imagine a room full of smart people who do not know much about meteorology.

“How many of you have heard something like this?”

“There wasn’t supposed to be rain today.”

“Yeah, the meteorologists are just rolling the dice. Honestly, if they’re wrong this often, what’s the point?”

Most people would spot the problem. Meteorologists measure the atmosphere, run models and issue probabilistic forecasts. Rain on a day with a 30 percent chance of precipitation does not, by itself, show that the forecast was bad.

Now keep reducing the explanation. Meteorologists take measurements. Models process those measurements. Computers run calculations. The calculations estimate what may happen next. Forecasts assign probabilities.

Keep going and you eventually get something like “meteorologists use numbers to guess what happens next.” Every step contains some truth. The final sentence still gives you a much worse account of meteorology.

The analogy only goes so far. Weather models and LLMs solve different problems and use different kinds of models. What interests me is the rhetoric around prediction. “This system makes probabilistic predictions” can slide very quickly into “this system is basically guessing.”

Something similar happens with next-token prediction. Yes, LLMs predict tokens. That fact matters, and any useful explanation of LLMs should include it. The phrase starts doing too much work once it becomes an answer to questions about what the model has learned.

A phone keyboard predicts what I might type next. An LLM can explain a legal concept, translate a paragraph, write code, identify a problem in an argument, imitate a genre, summarize a text and answer questions about the summary. All of those behaviours come from systems trained through prediction.

That seems worth explaining.

What does a model have to learn in order to become very good at that kind of prediction? What kinds of patterns does it represent, how abstract are they, how well do they transfer, and where do they fail? Researchers still disagree about parts of this, and “next-token prediction” does not settle those questions.

“Pattern matching” has the same problem. Human beings also learn patterns constantly. Writers absorb patterns in syntax, genre, rhythm and argument. Experts notice patterns that novices miss. Calling something pattern matching tells us very little until we know what kinds of patterns are involved and what the system can do with them.

“Statistics” does not get us much further. LLMs are statistical systems. Statistics can represent very complicated relationships, and modern machine-learning systems can build internal structures that researchers themselves struggle to interpret.

My concern is fairly narrow. A useful explanation can become a verdict without anyone noticing the change.

AI needs demystification. People should understand probabilistic generation, hallucination, training, token prediction and the limits of fluent output. None of that requires us to make the technology sound stupid.

I want enough room left after the explanation to ask what these systems can do, what they learn, where they fail, and why some of their capabilities appear at all. “Just autocomplete” tends to close those questions before we get to them.