Are there any more recent studies that are half-decent and analyze whether use of AI helps in SWE/SRE?Feels like current LLMs either one-shot the task (coding or research), in which case you save time (at the price of tokens) or they fuck it up in many subtle ways that would never pass code review and any quality standards. For research tasks they simply leave you misinformed unless you go and re-verify.The time it takes to review, fix the code and/or verify the output ends up in a net negative for productivity overall (+wasted money on tokens/local grid power draw).Ergo, LLMs are only good if you (and your colleauges + superiors) don't give a shit and are fine with not understanding/learning anything new.>t. Trying to make AI actually work in job and privately since GPT-3.5
>>109677980I use it for as an advanced search engine and for changes that are mostly trivial Using it for actual codegen is completely retarded Most people have stopped caring about quality but they also waste plenty of time writing COE's for the outages they cause
also genimi 3.8, astra are nice but overhyped like literally every other model ever
>>109677980I technically spend as much time giving context/logs/support docs in place as I would've just reading and tracing a request but it means I won't need to read java code written by Indians which helps my overall sanity and long term productivity
>>109678063You may not realize it, but you are also shitting fecal code all over the place now. If you can't beat them join them ,I guess
If you know how to prompt and verify the code AI generates, AI can greatly speed up development.
>>109677980It's strange because personally, I feel I am more productive. However, if I step back and look at the entire company I work at, it doesn't seem like we're getting any more stuff done than normal.
AI produces a ton of code quickly goving the illusion of saved time and then a month in you realize you have infinite bugs, terrible design and the code is an unreadable monsteosity so the only way to make changes is using ai to fix ai.
>>109682948/thread
>>109683676Skill issue
>>109677980>current LLMs either one-shot the task (coding or research), in which case you save time (at the price of tokens) or they fuck it up in many subtle waysYes. Guess what, it's a guessing machine. It never understood what you want form it, it just got very good at guessing.Current LLMs are literally "thousand monkeys with a typewriter". Only difference is that once the monkey fucks up janitors shoot it and bring a new one. So far the monkey supply seems endless.
>>109678012>Using it for actual codegen is completely retardedWish Jeets at Microslop understood that as well.
