Dhananjay Nene

People asked me why I cancelled. LLMs ruined my code. Once I prompt and the LLM starts spewing code, it sets the direction of my thoughts. And it rarely arrives at the elegant solution. Because verbosity is not a problem for the LLM, it provides verbose solutions. And in coding verbosity is inversely correlated with elegance and also correctness. In a well factored system, verbosity should not be a thing. In general I found the LLM never pointed to a refactoring that would make the code less verbose. And this was painful as it happened. And it is going to be 10x more painful in teams and large codebases and so on. When I apply my attention, I come up with solutions that reduce the surface area of the code. I tend to refactor so the change is easy, comprehensible, elegant. Working with the LLM felt like building channels for effluent. Somehow contain the slop. I used to think that LLMs would be useful in the hands of masters. But masters have no need for verbosity. Once you line up the system correctly, you rarely need to specify much. If my system is poorly specified, I find it much more edifying to read documentation until I understand how to specify it correctly rather than deal with a badly factored system. I went back to chatting with Claude as a documentation copilot. That's enough for me. Some basic copilot autocompletion is enough. Programming through prompting has negative RoI for the moment. Of course this can change in future. I'll be waiting.