You've heard that systems "train a model." Here is what the noun actually refers to.
A model is a giant mathematical function — a machine that takes numbers in and pushes numbers out — with millions or billions of adjustable dials, called parameters (or weights). Before training, the dials are random and the function's outputs are noise. Training nudges the dials, a tiny amount per example, in whichever direction makes the function's guesses slightly less wrong. Repeat a few billion times. The final setting of all those dials is the model. On disk, it's literally just the numbers.
Compression, not storage
A frontier language model trains on far more text than fits in its parameters. It physically cannot keep a copy. What survives is whatever regularities were worth keeping: grammar, facts repeated often enough to leave a dent, styles, structures. Think of it as the summary of the data, written in a notation no human reads.
This explains a family of behaviors at once:
- It can write a sonnet it never saw — the pattern of sonnets got in, not the sonnets themselves.
- It misremembers rare facts — one mention leaves a shallow groove; a thousand mentions leave a deep one.
- It can't tell you its sources — there are no sources inside, only settings.
"All models are wrong, some are useful"
Statisticians have said this for fifty years and it survives the trip to AI intact. A model is a theory of its data — necessarily simplified, faithful only where the data was. The interesting question is never "is the model right?" It's "where does its theory hold, and where does the world stop resembling the data it summarized?"
What to read next
A model inherits its theory from the data — including the parts nobody meant to put there. Bias in data is where that bill comes due.