3 comments

[ 0.19 ms ] story [ 14.0 ms ] thread
Three numbers from the research that made me record this:

57% / 11.8% : the best coding agent produces functionally correct code 57% of the time… and secure code 11.8% of the time

91% : share of deployed vibe-coded apps containing at least one vulnerability

+441% : the rise in code-review time while generation got faster

> 91% : share of deployed vibe-coded apps containing at least one vulnerability

This is interesting, I wonder what percentage of deployed human-coded pre November-2022 had "at least one vulnerability"

spec driven dev with good search is the only way I have been able to get consistently good output from a cheaper model