Post AWqNWybz7sTbdjYu6C by louis@emacs.ch
 (DIR) More posts by louis@emacs.ch
 (DIR) Post #AWqNWxrttDc5Ko06Ge by thees@emacs.ch
       2023-06-18T20:20:19Z
       
       0 likes, 0 repeats
       
       I'm still unsure where exactly I stand within the AI as a means to "outsource" the more mundane coding tasks discussion.Today I played around with a copilot AI, I got access to, a bit and have yet to find an exact use case for it. While coding I have the following ways to deal with "standard" cases:- if it's a pattern I use often, it should be a template to autocomplete- if it's a problem I solve often in a similar way, it should be a template- if it's a function I use in more than a handful projects, it should be part of a library- classes should always follow the SOLID mnemonic, thus being reusable should be the default - if my programming language requires a specific skeleton, it should be a templateI should also mention that I am heavily biased in regards of actually having more than a decade of templating, personal/project libraries, and a toolbox to solve common problems with I can rely on.If I would be at the beginning of my carreer now, I'd probably would use a copilot AI as I wouldn't have got anything else to built my work upon. This would be because of a lack of better options.
       
 (DIR) Post #AWqNWybz7sTbdjYu6C by louis@emacs.ch
       2023-06-19T08:38:34Z
       
       0 likes, 0 repeats
       
       @thees I've played around with Copilot and was amazed with it for a while because it could infer what I was going to type anyway. But then Copilot started to interfer with my line of thought and and suppred the regular LSP suggestions. Now I find it annoyning to get constant suggestions from Copilot. It is like someone is looking over your shoulders while you program and constantly tries to correct you. No Copilot anymore.But I make use of ChatGPT to ask for specific solutions where I would need too much time to do my own research. "How do I create a new superuser role in Postgres" and bam, I have the solution. Yes I could have looked that up in the docs but Postgres docs are very hard to navigate.However, for my AI can be a great assistant if used sparingly and correctly. It will never take over my actual coding. Becaues if I leave that to an AI, I could just quit my job and do something enitrely different.
       
 (DIR) Post #AWqNWzdRJxEaoXFa0O by thees@emacs.ch
       2023-06-18T20:27:15Z
       
       0 likes, 0 repeats
       
       I'd argue that using LLMs for coding is one of the less viable and resourceful options one could choose. I'd love to see more projects using the locally available code (read that as in: your projects folder) base to make suggestions; be it by utilizing parse trees to suggest skeletons/templates of functions/classes with a similar name; or just by detecting pattern and make suggestions based on that (which would be based on a small language model)... I don't know if there are any projects in that direction; but that'd be actually lovely to see and super cool!
       
 (DIR) Post #AWqP8dJAsbsqWjHd8i by efim@emacs.ch
       2023-06-19T08:56:37Z
       
       0 likes, 0 repeats
       
       @louis @thees i tried to use "codeium" - similar to copilot project, targeting code generation.It wasn't helpful enough on few tries with my projects, but what I've thought to do - set completion with its company function to a separate hotkey in emacs, so that it wouldn't interfere with LSP and was a bit more intentional.Its possible that then I'm missing out on important part of ai completion, but maybe this setup would suit you use case.
       
 (DIR) Post #AWqcC8FdadTolecz6e by thees@emacs.ch
       2023-06-19T11:22:56Z
       
       0 likes, 0 repeats
       
       @louis I do prefer the regular LSP suggestions/ my snippet expansions over the Copilot ones. But that’s mainly because they work better for my coding style & especially my snippets are usually closer to what I want than the autopilot suggestions are. I also have a code-generation toolbox I’ve written years ago (csv to sql insert, json to go struct and so on, csv to python dataclass structure); so I tend to rather use that then ask a language model. But I do see how it’s a timesaver for specific questions and more efficient than googling for some questions; I think that especially when one’s solo-working a project LLMs be it ChatGPT or specific code generation AIs like Autopilot can be a timesaving factor and overall a net plus.