>AI coding tools can slow down seasoned developers by 19%>We conduct a randomized controlled trial (RCT) to understand how early-2025 AI tools affect the productivity of experienced open-source developers working on their own repositories. Surprisingly, we find that when developers use AI tools, they take 19% longer than without—AI makes them slower. We view this result as a snapshot of early-2025 AI capabilities in one relevant setting; as these systems continue to rapidly evolve, we plan on continuing to use this methodology to help estimate AI acceleration from AI R&D automation>Methodology>To directly measure the real-world impact of AI tools on software development, we recruited 16 experienced developers from large open-source repositories (averaging 22k+ stars and 1M+ lines of code) that they’ve contributed to for multiple years. We pay developers $150/hr as compensation for their participation in the study.>Core Result>When developers are allowed to use AI tools, they take 19% longer to complete issues—a significant slowdown that goes against developer beliefs and expert forecasts. This gap between perception and reality is striking: developers expected AI to speed them up by 24%, and even after experiencing the slowdown, they still believed AI had sped them up by 20%.https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
>>107786804Slightly faster, as it is way easier to ask chad than to google something and wade through jeet SEO spam. But if you're running with generated code you're gonna be way slower. The quality of the code produced by models is inversely proportional to the popularity of the language/framework used. This is known as "Jeet's Law"
>>107786804I don't know if I'm faster but I'm definitely less stressed. Writing good prompts and reviewing generated code still requires skill but it's easier and less burdensome than writing it all myself.I suspect that if I wanted to I could move faster thanks to AI, but I wouldn't get paid more so that sounds like a bootlicker wageslave approach to me. I'd rather keep my output at 1x, get paid the same, but be happier and less burned out.
>>107786804im way faster when im learning somethingi use shatbots to generate minimal working examples, and i ask it to explain what it does line by linealso with boilerplate- its pretty decent at thatbut with code?its just fucking worthlessits not that i work faster or slowerit just doesnt fucking work, i end up rewriting everything, its fucking dyslexic>inb4 take the premioomit shits itself at the stage of ingesting whats in the manualwhen you start dealing with the finer things of life there just isnt enough data to train the shatbots to acceptable levels
AI makes me faster as I can use it as a reference and for writing small snippets of code, but I still need to double check what it outputs.
>>107787170>>107787159these guys get it. I use chat gpt to learn about new features or a "different" way to solve problems, claude to write snippets, combine and fix manually, have gpt write tests, fix tests.it's all about laziness desu, if I was less lazy I'd use less ai. but as it stands, they're "decent" at writing snippets and unit tests so long as you tell them you can use a testbed via {testbed creation here}, but I'd say the unit tests are normally like 70-80% with some cases being outright wrong. it's really only good at doing small pieces at a time or learning about how you "could" solve something and then pretty much rewriting it yourself.
I've been "vibecoding" for the last 4 days because I'm doing a project that I simply do not give a fuck about but still have to finish and here are my personal results:Searching for structs, basic data, explaining hex contents into actual values in a struct and small code examples?Speedy Gonzales, what would've taken me an hour of scraping githubs and stackoverflow it just threw on my plate without question and put a cherry ontop.Generating actual code?Waste of time. Waste of time. Waste of time. Waste of time. Waste of time. Waste of time.HOLY FUCK is it a waste of time, I've got to baby this, double triple quadruple check every change it makes, force it to not touch anything outside of functions or snippets of changes.God forbid it gets something wrong and then gets into an infinite loop of "fixes" that don't do anything in a chain that it just iterates over and over again through because it SEEMS like the most common answer/solution.
>>107787159I think I remember you from another thread.What model are you using, and with what tools?>ingesting whats in the manualHave you tested that whatever RAG system you're using is working as expected, i.e. by asking it to reproduce something from the manual verbatim?I have a suspicion that you might be trying to program by copying and pasting snippets back and forth with the free version of the chatgpt website. If I'm right then please understand that you're wasting your own time plus the time of anyone you discuss AI with. My apologies in advance if I'm wrong.
>God forbid it gets something wrong and then gets into an infinite loop of "fixes" that don't do anything in a chain that it just iterates over and over again through because it SEEMS like the most common answer/solution.sometimes it get the right thing after like ten iterations and you're like "yeeeaaaaah finally" and then it will fuck up again so you gonna ask to get back but it get back to when it didn't work kekllm are funny for me ,it's just hasardous shit.Too much hasardous wich means comptletly no use case except for things with large margin error.
>>107787453>>107787508>infinite loop of fixesWhen it fucks up, don't bother arguing with it or asking for corrections. Just throw away whatever it's done and start again, but this time change the plan document to be more specific and pre-empt whatever the problem was.
>>107786804Only using local LLMs. It's a mixed bag, but vaguely good. Way faster when learning. Faster when asking for reasonable amounts of code. Slightly faster when telling the AI almost every single step. Slower when asking for too much unspecified code in one go.It's not magic, and there's a learning curve, but it does slightly increase my productivity already. My methodology is gradually improving too, so eventually it probs will be an ok buff.
>>107786804The latest opus 4.5 is pretty good, and I have been just prompting since my company lifted the usage cap.
>>107787159This is my basic experience. I'm not a dev or anything, but I've been asking grok to solve bash problems. If all it has to do is arrange arguments with flags into a simple conditional from reading manpages, it has no problem. Even then, you have to lead it to water. Anything just outside of that starts breaking down very quickly. It has a habit of using deprecated features in it's responses, which indicate some rather serious flaws in it's ability to reliably identify relevant information.
Really good, it's made my gooning 700% more efficient and 1000% more frequent, no idea about coding because I only use it for gooning
>>107786804It's usually faster to apply or formulize theory with, but terrible at actually implementing it.
>>107786974>to ask chadjeet
>>107789384no u fgt
In my daily life it’s easier to google or just keyword directly to what I want. AI doesn’t get me to my music, video games, porn, shopping or work any faster and actually gets me side tracked because it produces shitty results.
>>107786804>Our north star is 1 developer, 1 month, 1 millions lines of code.
>>107790711>We are beginning to distinguish between 'spectacle' and 'substance'. We now have a clearer sense of where the tech is headed
>>107790718>qwen2.5-coder-14b
I can do everything, finally
AI is worse. It wastes my time. It's like fucking clippy. You don't notice it at first, it seems like it knows what it is talking about. But it's just confidently giving you answers that are 10 years out of date and sending you down rabbit hole after rabbit hole.
>>107793626>it's just confidently giving you answers that are 10 years out of date and sending you down rabbit hole after rabbit hole.It depends on the model and version, no? I don't think newer, premium models do that, at least not very often.
>>107786974If you still need google or AI to help you while writing code then you are still learning how to code.
>>107794790>The new expensive models don't have this problem!Almost all AI is functionally the same. It's a crude technology that is only somewhat functional because of the insanely vast amount of data it has been trained on. The tech just isn't there to do what it is being promised.
>>107786804Probably faster, but I only really use it as a search engine replacement because search engines have gotten so fucking bad.
>>107796222>>107796263I can't wait until chat bots start displaying ads between chat messages.
