The Night the Orchestra Got Smaller and Sounded Better
The conductor stared at the wall clock in the quiet hall. One rehearsal slot left, and the stage was already crowded. One choice sat there like a weight: add more musicians, or use the same time to rehearse more pages until the sound tightens.
For years, a lot of people building text predictors followed the “bigger orchestra” instinct. They kept adding more inner parts, but didn’t give the system much more reading practice. The stage looked impressive, but the timing stayed messy.
Then some builders treated rehearsal time as a hard limit and split it on purpose. In their picture, rehearsal-hours are the fixed budget of tiny calculations, musicians are system size, and pages are how much text it reads while practicing.
They tried many mixes: small orchestra, lots of pages; huge orchestra, barely any pages; and everything between. The same pattern kept showing up: with the same total rehearsal time, the best sound came from a balance, not an extreme.
One detail tripped people up. The rehearsal plan has a pace, like pushing hard early and polishing later. If the plan assumes a longer rehearsal than you actually have, you run out of time while still pushing, and the last run-through stays rough.
In a head-to-head run, they spent the same total rehearsal budget as a very large earlier system, but chose a smaller one and gave it far more reading practice. It did better on many kinds of language and problem checks, and it was cheaper to run later. But the cleaner performance didn’t automatically make it safe or unbiased.