Letterforms

Letters and numbers, as a model sees them.

Three small neural nets, each pointed at one quiet thing about written symbols: how we decide which letters to make big, and what comes next when you follow a sequence past its end. Draw in them; they answer in glyphs.

Capitals A net that learned casing as a transformation — and what it does when you feed it things that have no case. open → After Nine A model that learned to count by shape. Ask for the digit after 9 and watch what it reaches for when the sequence runs out. open → After Zed The same idea, aimed at letters: the one after Z, and past the end of Latin, Greek, Cyrillic, and Armenian. open →