People are predictable. Language models know it.
Or on misuse cases a researcher had to imagine.
40 users wrote 273 routines in plain English, plus hints about when each one runs.
Each user’s hints place their routines into hourly slots, giving one ordered, month-long event sequence.
An n-gram language model learns the patterns, then writes new scenarios one event at a time.
Lower than both English text and source code.
for up/4-gram routines, beating the 63.75% for users’ own routines.
specified from 27 unsafe scenarios, without imagining a single case.
than a graph-based baseline that can’t rank which event comes next.
SmartThings sometimes skipped events, including locking the door.
“Deleted” routines kept running in the background.
A smart lock stayed unlocked after the door closed.