... and then there's me. I know it's capable, but I have not yet found a good heuristic to decide when to write or generate code. A good heuristic would comfort me. It would assure me that I'm not missing chances to improve the system.
I've been getting curious about LLM code generation lately. I've been doing experiments over these past 4 years to try to adopt them at my own discretion. My codebase is around 25,000 lines of a popular C-like language.
My conclusion is; two is better than one. I wouldn't be super upset if I had a junior on my project and they submitted this code. I'd be sure to tell them they could've done it with a lot less, so they get an eye for simplicity. I can't do that for Opus, and I find that a bit sad. But ultimately, it got the job done. How much of this kind of complexity can I let rack up? Is that what I ought to do? This is an interesting comparison, but that strategy reminds me of garbage collection but for code instead of data. Do I just let the mess pile up and go clean it up every week or so? Maybe that's what I'll try next, I don't know. I do think I can accomplish more using this as a tool, but I am just finding it to be a bit of a problem figuring out how I'll do that in a manner that is authentic to myself.
... and then there's me. I know it's capable, but I have not yet found a good heuristic to decide when to write or generate code. A good heuristic would comfort me. It would assure me that I'm not missing chances to improve the system.
I've been getting curious about LLM code generation lately. I've been doing experiments over these past 4 years to try to adopt them at my own discretion. My codebase is around 25,000 lines of a popular C-like language.
I decided to race Opus 5 High the other day. I asked it to add a grid to help me perceive depth in the 3D space of my application. There were some clear flaws, it took several prompts and ~32 minutes. I did not inspect the code for quality, I did not read closely. I then stashed the change and implemented it myself. It took me 35 minutes to get happy with my results (I chose to do analytical filtering to reduce jaggies and moiré.) Upon inspecting both diffs, I had added about 60 lines of code (including comments) and Opus added 500. It looked like just under half of what Opus generated was comments, fine, I expect that... I ended up liking the way I went about things. Opus tried to reuse just about every little utility in the codebase, and actually duplicated the definition of some in-line (an alpha premultiply comes to memory.)
My conclusion is; two is better than one. I wouldn't be super upset if I had a junior on my project and they submitted this code. I'd be sure to tell them they could've done it with a lot less, so they get an eye for simplicity. I can't do that for Opus, and I find that a bit sad. But ultimately, it got the job done. How much of this kind of complexity can I let rack up? Is that what I ought to do? This is an interesting comparison, but that strategy reminds me of garbage collection but for code instead of data. Do I just let the mess pile up and go clean it up every week or so? Maybe that's what I'll try next, I don't know. I do think I can accomplish more using this as a tool, but I am just finding it to be a bit of a problem figuring out how I'll do that in a manner that is authentic to myself.