Why not whole words?

So why put up with pieces that land in awkward places? Because the alternative is worse.

It is the difference between a box of standard bricks and a kit with one custom piece per finished object. A vocabulary of whole words never closes. Every name, every typo, every word coined next year needs its own entry, and the list still misses whatever comes next. Fragments do not have that problem. un and bel are not much on their own, and that is exactly why they work. A fixed set of them snaps together into almost any word, including ones nobody has written yet.

How big does that set need to be? For many language models, tens of thousands of pieces. Millions of words, spelled out of a box that size.

The exact number drifts between models and does not matter. The trade is a small box of reusable parts instead of a catalogue that has to contain everything in advance.