hahaha, never thought about it but post should have come with a warning for anyone with service indicator-related ptsd.

also, being at nationals back-to-back years is impressive!

Yup, same result set across all tests! A lot of it was deduplicating requests, removing feature bloat, smart tuning based on internet speeds, and being much more efficient with memory.

Note: this is as much an indictment of default settings on tools as it is of feature bloat. Even painstaking optimization of the original tool didn't approach these numbers.

truly believe pompeii/herculaneum graffiti should be required reading in school to really emphasize this point

yeah, even the best models in general are pretty fragile with wording when it comes to tool use.

scale.com/leaderboard/...