My unsubstantiated working theory is that LLMs are too verbose because they weighted SEO’d content too much in the training data, probably because that is what was easily crawled and obtained.
Microblog
-
I was bummed when the veggie share/CSA we’d been with for the past four years ended, but the new one we switched to this year has exceeded our expectations. The fruit has been a delightful bonus. Multiple varieties of plums, peaches, pears, and apples, all juicy and flavorful. All from a farm just across the Hudson!
-
I got a new space gray MacBook M4 Pro for work with 48GB of RAM.
I’m not running any local LLMs yet, so the main place I’ve noticed the RAM making a difference is in Numbers.app. On my previous machine with 16GB of RAM, applying filters to large sheets was noticeably slow. Now it is snappy. I still prefer Google Sheets, though.
This space gray machine gets hot noticeably faster in the sun than my previous silver one did.



