I review a *lot* of code, daily, and a new metric has popped up in the back of my mind.
-
I review a *lot* of code, daily, and a new metric has popped up in the back of my mind. I call it "Time to First WTF".
For a high quality PR, I'd say the #TTFWTF is around 5 minutes. On any non-trivial PR there will always be something that looks weird on first sight.
The interesting thing is that I have observed that #genai based PRs, or AI-assisted PRs, generally have a TTFWTF around 10s. I am not exaggerating.
The TTFWTF has actually become the primary way I can recognize AI-based work
-
S simonjust@mstdn.dk shared this topic
R reynir@social.data.coop shared this topic
-
I review a *lot* of code, daily, and a new metric has popped up in the back of my mind. I call it "Time to First WTF".
For a high quality PR, I'd say the #TTFWTF is around 5 minutes. On any non-trivial PR there will always be something that looks weird on first sight.
The interesting thing is that I have observed that #genai based PRs, or AI-assisted PRs, generally have a TTFWTF around 10s. I am not exaggerating.
The TTFWTF has actually become the primary way I can recognize AI-based work
@kamstrup How about the secondary metric, ATBWTF: "Average Time Between WTFs"?
-
I review a *lot* of code, daily, and a new metric has popped up in the back of my mind. I call it "Time to First WTF".
For a high quality PR, I'd say the #TTFWTF is around 5 minutes. On any non-trivial PR there will always be something that looks weird on first sight.
The interesting thing is that I have observed that #genai based PRs, or AI-assisted PRs, generally have a TTFWTF around 10s. I am not exaggerating.
The TTFWTF has actually become the primary way I can recognize AI-based work
Even on 20 line PRs this holds true. It is almost always the case that AI selects the most mind boggling way to solve a trivial matter. Looks passable, if you don't know anything, but if you do...
The real tinfoil-hat-moment comes when you realize that it is in the interest of the large AI companies to produce the worst code possible that is still passable


-
B bogwitch@social.data.coop shared this topic