> In terms of software quality, I would say the code created by the AI was worse than code written by a human–though not drastically so–and was difficult to wor

> In terms of software quality, I would say the code created by the AI was worse than code written by a human–though not drastically so–and was difficult to work with since most of it hadn’t been written by the people whose job it was to oversee it. This is a key insight. The other insight is that devs spend most of their time reading and debugging code, not writing it. AI speeds up the writing of code but slows down debugging... AI was trained with buggy code because most code out there is buggy. Also, when the codebase is complex and the AI cannot see all the dependencies, it performs a LOT worse because it just hallucinates the API calls... It has no idea what version of the API it is using. TBH, I don't think there exists enough non-buggy code out there to train an AI to write good code which doesn't need to be debugged so much. When AI is trained on normal language, averaging out all the patterns produces good results. This is because most humans are good at writing with that level of precision. Code is much more precise and the average human is not good at it. So AI was trained on low-quality data there. The good news for skilled developers is that there probably isn't enough high quality code in the public domain to solve that problem... And there is no incentive for skilled developers to open source their code.