Autocomplete Studies#
Week 1 — before the screen, the keypad
Twelve keys. Twenty-six letters.#
Every letter after the first press of a key is a guess.
That guess is predictive text. It always has been.
Multi-tap#
Press 4 once for G, twice for H, three times for I.
Slow. Unambiguous. Nobody predicts anything.
T9#
Press 4663 once, in order.
The phone looks up every word that mapping could spell: good, home, hood, goof…
…and picks one. That pick is the whole course.
“Text on 9 keys” — the name says exactly what it does, and nothing about how strange that is to actually build.
The tension, stated once#
- a small, fixed vocabulary (the dictionary)
- an ambiguous signal (the keypress sequence)
- a ranking someone built, resolving the ambiguity
Every week this semester changes one of these three things.
Week 2 changes the vocabulary’s confidence#
Autocorrect: the same ranking, applied uninvited.
Week 5 changes the signal#
Code completion: the signal is a program, not a word.
Week 7 changes the scale#
A chat model: the same ranking, one token at a time, run long enough to look like an answer.
Today’s lab#
Build the T9 disambiguator. 4663 should return good and
home — and you should be able to defend which one comes first.
Whose dictionary?#
Keep that question. It doesn’t go away for the rest of the semester — it just changes what counts as the dictionary.