“Oh shit git” seems to have been succeeded by “Oh shit GitHub Copilot”: ‘Downward Pressure on Code Quality’
Posted by jpluimers on 2024/01/29
Not sure about you, but when I write code I want it to be better – way beter even – than average code.
The problem with any LLM based Generative AI is that it generates text based on the average of the past corpus they were trained with at the time they were trained.
It is exactly why I have been advocating for a while: be careful when using Generative AI, as you get generated text based on the combination of averaging over the LLM corpus with the relatively small prompt you phased trying to reflect a tiny bit of the model of the reality you are trying to write software for.
So I was not at all surprised by this article: [Wayback/Archive] New GitHub Copilot Research Finds ‘Downward Pressure on Code Quality’ — Visual Studio Magazine.
It isn’t by accident that Wikipedia warns when incorporating LLM based text into Wikipedia entries: Source: Wikipedia:Large language models – Wikipedia.
So, if you ever decide to use code generated by tools like GitHub Copilot, that code needs a thorough review to verify it is fit for the actual real world problem you are trying to automate.
Via [Wayback/Archive] Fefes Blog: Ihr AHNT ja nicht, was die gerade über Codequalität bei KI-assistiertem Programmieren herausgefunden haben!
Ach. Ach was! Wird gar nicht besser, wenn man weniger Arbeit reinsteckt? NA SOWAS!
Scheiße, Bernd! Hätte uns doch nur jemand vorher gewarnt!1!!
So there is a new incarnation of Oh shit git named Oh shit GitHub Copilot.
--jeroen
The Wiert Corner - irregular stream of stuff 
Leave a comment