The tiny cast that learned a huge show
Sawdust floated in the work lights while a touring crew shoved flat walls and painted doors into one truck. The director rubbed a thumb over a thick stack of scripts. Big cast, easy memory. Or a small cast, and a lot more rehearsal, so the show can actually travel.
Most people bet on the big cast. More people, more lines. People make the same bet with chatbots: more inner knobs should mean better answers. But a huge cast costs more every night and moves slow. Takeaway: when performance is pricey, extra practice can beat extra bulk.
They chose the compact cast and went hard on rehearsal. A team behind a system called LLaMA did something similar: they built smaller and larger versions, then made the smaller ones read far more than usual. They used only public writing, like cleaned web pages, encyclopedias, books, code, and forums.
Long rehearsal only works if it stays organized. The crew taped clear marks on the floor, kept steady routines, and trimmed awkward bits of dialogue so nobody lost their place. LLaMA got the same kind of behind-the-scenes care so it could stay steady, track where it was in long passages, and train faster without wasting work.
Opening night, the small cast hit scenes that usually need a crowd. LLaMA showed that too: a smaller version beat a much older, much bigger system on many common scorecards, and the biggest LLaMA kept up with top systems that had used more private or larger-scale training. It also did better at math and code when it could try a few answers and pick the cleanest one.
After the applause, the crew listened for what felt off. With a bigger cast, some lines came out harsher, and old stereotypes slipped in. LLaMA can also produce toxic or biased outputs, and it can sound confident while being wrong. The gamble still landed: smaller systems can get strong from longer practice on public writing, but safety and honesty still need real attention.