I believe you that genAI is being used for code suggestions, but I wouldn’t call it a genius.
This is anecdotal but over the last couple of years I’ve noticed Visual Studio’s autocomplete went from suggesting exactly what I wanted more often than not to just giving me hot garbage today. Like even when I’m 3-4 lines in to a very obvious repeating pattern it’ll suggest some nonsense variation on it that’s completely useless. Or just straight making up variable and function names that don’t exist, or suggesting inputs to function calls that don’t match the signature. Really basic stuff that any kind of rules-based system should have no problem with.
I wouldn’t call it a genius either, it’s just all over the place. Sometimes it’s scary good and predicts your next move, most of the time it’s just okay, sometimes it’s annoyingly bad
My last job was on a fairly large typescript codebase (few hundred Klocs) which we started some time before LLMs were a thing. While we weren’t into academic engineering patterns and buzzwords, we were very particular in maintaining consistent patterns across the codebase. The output of Copilot, even with early models which were far from today’s standards, was often scarily accurate. It was far from genius but i’m still chasing that high to this day, to me it really indicated that we had made this codebase readable and actionable even by a new hire.
I believe you that genAI is being used for code suggestions, but I wouldn’t call it a genius.
This is anecdotal but over the last couple of years I’ve noticed Visual Studio’s autocomplete went from suggesting exactly what I wanted more often than not to just giving me hot garbage today. Like even when I’m 3-4 lines in to a very obvious repeating pattern it’ll suggest some nonsense variation on it that’s completely useless. Or just straight making up variable and function names that don’t exist, or suggesting inputs to function calls that don’t match the signature. Really basic stuff that any kind of rules-based system should have no problem with.
I wouldn’t call it a genius either, it’s just all over the place. Sometimes it’s scary good and predicts your next move, most of the time it’s just okay, sometimes it’s annoyingly bad
My last job was on a fairly large typescript codebase (few hundred Klocs) which we started some time before LLMs were a thing. While we weren’t into academic engineering patterns and buzzwords, we were very particular in maintaining consistent patterns across the codebase. The output of Copilot, even with early models which were far from today’s standards, was often scarily accurate. It was far from genius but i’m still chasing that high to this day, to me it really indicated that we had made this codebase readable and actionable even by a new hire.