Does anyone else not understand the hype with AI generated code? I almost never use AI at $DAYJOB; only for "search engine" type tasks where I'm looking for an answer or perspective on something that's hard to search for with google - not code generation.
Feels like more of a fight with the AI and less time thinking about the bigger picture of the changes being made (e.g. systems involved, business considerations, documentation, etc..)
I use it like a search engine (or rubber duck) too. I like to test ideas and form hypotheses, but I don't really like to code with an LLM. Occasionally it's helpful to template something but it's rare that I want to iterate on the foundation an LLM created.
I suppose the main reasons are that 1) the code convention likely doesn't mesh with what I'm working on, 2) the lack of context awareness, 3) common performance concerns and 4) potential IP issues which admittedly seems like a rare issue yet I'd rather not contend with it.
There is an exception though. When I'm trying to prototype firmware quickly, I like to use it to generate code for components I haven't used before or recently, and I'm content to rely on it until I've proven out an idea. Past that point though I will almost certainly rip out that code and start fresh using the conventions and patterns I prefer.
Last week I oversaw and architected a number of STDIO Python scripts for ingesting PDFs page-by-page into postgres tsvector, searching, and displaying the text and metadata that was entirely written by an LLM.
I didn’t “fight” with the machine and I focused solely on the bigger picture. I read the output like I was performing a code review.
> Does anyone else not understand the hype with AI generated code?
This can feel maddening, but is basically the approach taken for a few decades in many companies.
Developpers who think long and hard about their code and bring stability to the code base cost money, sometimes a lot. In comparison many businesses can get away with the "cheap labor with meh output but handled by QA and a management layer" approach.
That was Taylorism in the early days, outsourcing to random location also came from the same philosophy, the dream of "code modules" from the EJB area
that non devs would just mix and match, etc.
That dream will always live, if AI isn't it, something else will come up as a magical wand that somewhat generate output from a wide cheap "meh" labor base that only needs to be checked and managed by a few knowledgeable people.
Feels like more of a fight with the AI and less time thinking about the bigger picture of the changes being made (e.g. systems involved, business considerations, documentation, etc..)