Post

I don't get how so many software folks cling to LLMs, like, I'm sorry but every time I've given them a try the result is dogshit. & I promise I'm not holding it wrong! Even excluding externalities, it's just not good at what it's sold to do! If you think it is, I worry about your judgement.

2 reposts42 likes

If you write code today in Python - you focus on the python - but behind the scenes it ends up being machine code. You care about the python. You don't worry about the machine code. Some developers will write in LLM prompts. They won't care about how good the machine code is either. It will suck

Eh, it's a false equivalency. It's more like if they were writing Python but the runtime was a group of 100 people who've known a single programming language for 4 weeks trying to execute your code one line at a time without talking to each other.

Absolutely - but they will write their 'spec' as prompts, and have it generate code. They won't read the code, or tweak the code. They will tweak the spec, and completely regenerate completely new code. The waste is going to be immense - but they won't see it - and they won't care.

100%!! my company is pushing it hard and nothing i have seen either in my own testing or in any code written makes me feel comfortable. even worse nobody understands the codebase, cannot troubleshoot it, and cannot explain to me simple security related info (they llm it and it's incorrect)

I have yet to see any LLM produce code that I would put my name on, even with editing and revision. I'm not gonna plate a turd and garnish it with fresh parsley and send it to the table with my god damn name on it. No way! & now I got haters because I don't wanna make literal shit!

I swear some of y'all sound like "soon it'll be normal for everyone to make and eat shit!" Okay! Have fun! Just please don't eat shit and then talk in my face!

& y'know, there's legitimately useful and interesting applications for LLMs, but god damn if it hasn't become the erstwhile hammer to every nail. The science is neat, the business is horrifically misleading, exploitative, and dangerous.

Gotta be honest that I’m having a blast with them. I think I’m often looking for “good enough”. And the domain matters. Doing a CRUD web front end to a simple database? They are really good IME. Doing some novel algorithm work? They kinda suck. ¯\_(ツ)_/¯

The only thing I’ve found it useful for so far is a search engine for things my own spelunking has failed at. I still won’t trust it with writing actual code.

I’ve been having Claude churn out some boilerplate code lately that has turned out well. It is a task of copying snippets that already exist and reassembling them, so nothing too hard. But, yeah, asking it to do something novel means I spend a lot of time fixing vibe code shit.

LLMs produce really bad code, and are complete trash at architecture and design. But, there are niches where they’re genuine magic. Worked recently on some code for file processing where inputs were messy and variation was huge. Using AI to find generalizable heuristics was an enormous time saver.

for document processing work, it seems good. for code...i'm seeing the allure. turn brain off, write prompt, get result. it'll kinda work, and for a lot of usecases that's close enough. ofc ppl aren't thinking too far ahead with what is promised by this becoming a reality, and their role in it.

i absolutely am fucking tired of arguing with claude by proxy, or fifty thousand line claude MRs, and the general slop trend. ai;dr no, i am not going to read your claude conversation. it's wrong, more context doesn't change that.

LLMs do seem capable enough in belching out UI stuff. tools like lovable have let our less-technical folks iterate quickly. it's still slop though. tools like claude will just add entire modules that duplicate other work and not clean up. tech debt as a service. this is the honeymoon phase.