Devs are officially shown to be 20% faster with AI
https://libguides.csun.edu/data/doubtI'm not saying the data is right or wrong, but I must upload this at least, so that you are aware of it.
>>108900262>all research across all time has been wrong because you can be misleadingprofound
>>108900250>>108900264What is it with these fucking bots that reply with inane greentext that isn't what the post said at all. The post says that sometimes stats can be misleading. You are such a lazy retard, OP, that you didn't even link to the shit in your screenshot. Go get those fucking URLs and paste them here, I'm not doing it for you.
>>108900250faster at what? producing bugs for mythos to catch? L0L
>>108900268It’s a common disinfo tactic. When you recognize them you see it everywhere and most of it is done by bots now. We are all being programmed.
>>108900268reminder if OP can't substantiate the study URL this is a ragebait thread
>>108900268>What is it with these fucking bots that reply with inane greentext that isn't what the post said at all. It's not bots, it's midwits. You can thank wojak culture for representing all of reality as a black and white face screaming to mindbreak people into thinking greentext is a valid form of argument.
>>108900277Only the retarded are being programmed. I read the Cointelpro and consensus cracking guides that were routinely posted here in the early 2000s, i know all these forum sliding and shill techniques
>>108900284The techniques in question
>>108900299fucking hellpaste from atril, since apparently copy image from atril is fucked up on my machinehttps://ghostbin.lain.la/paste/5mvw2
>>108900262I'm not saying this post is right or wrong, but I must upload this at least, so that you are aware of it.
>>108900250I dont code. I vibe code. I have pushed out ~10 different projects, without vibe coding this #w would be 0 because the complexity of the project I created was never a 1 man project in the past unless you had extensive knowledge, capability, and time to do it
https://metr.org/blog/2026-02-24-uplift-update/#wider-adoption-of-ai-has-made-it-more-difficult-to-measure-task-level-productivitySince OP couldn't be bothered to provide an actual link.The experiment was done on only 57 people, and they admit that the sampling was very biased because they couldn't find people willing to do tasks by hand.
>>108900368>57 peopleuselessalso i think the fact the did the study twice and got the exact opposite result should itself show the study just isn't very good
>20% fasterWhat does that mean though?20% more lines of code per time?20% less time to produce the same quality and functionality of code?20% less time to produce a solution which may or may not vary in quality?
>>108900380The original was apparently done with 16 people.
>>108900368>they couldn't find people willing to do tasks by hand.why did this make me laugh
They let the experiment's subjects select their own tasks, too. Holy shit, this is garbage.
>>108900368>>108900407lol whatThis is the rigor of the group so many people use to track AI progress?
>>108900339you're brown
>previous years developers slow down to 80% of the previous speed>the following year they're at 96% of their speed two years prior>this is a huge win
>>108900456you're brown
>>108900368>57 peoplebtw this "study" will be quoted everywhere and will influence billions of dollars of investment
>>108900250I personally think when the dust finally settles it's going to be around 10-20% improvement in throughput but also cost. Anecdotally in the actual corporate scene it's hard to tell what the productivity gains really are because everyone is also working more hours on top of using AI. Either way a long way from the 10x that sloppers are promising.
>>108900546>Either way a long way from the 10x that sloppers are promising.why would you even be proud of that? "i can do more work for the same amount of money!" really? that's supposed to be good for you? is it not bad enough already?
>>108900250and i am also 20% faster with clang-formatbut yes, i am probably faster with ai because it sometimes finds blunders
>>108900596this
>>108900436Maybe they had an AI come up with the study's methodology
>>108900361Interesting, show them. Provided they exist.
>>108900727Claude is 10x better than in 2025
You know, if AI companies really believed their tools would make software development so much more productive, then they would stop selling access and instead they would simply produce all the software themselves and take the whole market.
>>108900727That is funny.
>>108900250How would you even test this? Have a thousand devs all making the same piece of software and timing how long they take with AI and without? Because I can guarantee you that isn't what happened.
>>108900743yup. but then, how would they convince retarded C-levels to spend billions on their shit by investing on them or buying underpriced tokens until they end up depending on AI-generated code to continue doing anything at all?
>>108900596I'm not disagreeing with you, or "proud" of productivity gains either. The growing pains of shitty vibecoding are already fucking up the entire industry and it will get worse before it gets better. But it is nevertheless the way the industry seems to want to go: increasing developer throughput by spending shitloads of time, money, and energy on developer tools that are easily misused. As companies inevitably try to add process and safeguards around these tools to prevent the kinds of accidents that are today becoming common place, the productivity gains will also start to decrease.Personally I find it terrifying that adoption of these tools is so high among managers especially. Managers are making real decisions over whether or not their real employees get to keep paying their bills based on what the next-token predictor hallucinates for them. I don't think any of us have the power to stop it, only prepare for the worst.Speaking from my own experience, the tools are useful as a way to quickly sift through shittily maintained wikis on the company intranet, and for generating wrote code that would normally be handled with copy+paste with some adjustments. Any time I've tried to have it do more, I regret that decision almost immediately.
>>108900250Uhhh I thought they can do 6 months of work in a just a few hours?
>>108900380>>108900407>>108900452Hello retards, the reason the study was done twice with two different results is because it was repeated a year later after AI coding tools actually got goodI think this is one of the failure modes of sceptics - their small and smooth brains can just barely comprehend the present, let alone understand that things can change and that there are trends that can continue
>>108900250its probably true i used to try using chatgpt etc when feeling lazy at work then id have to spent a bunch of time handholding it because it never got anything right and itd take longer than it would have for me to write it. now i can just ask them for some stuff and theyll give me working code
C-suite isn't going to like to hear thisbut the best use case for AI is going line by line "pair programming" with a human, through every single line of code and auditing itNOT generating lines of codeThe reason for this is "AI as a search engine" btw
>>108900944It doesn't need that much supervision. You only need to manually inspect your auth/security relevant code. Business logic you can go blind on all day long
It is also faster to just write down random gibberish instead of writing actual sentences. Doesn't mean that gibberish increases productivity or allows you to replace proper writers.Ai increases the amount of code you can produce. BUT, that code needs to be checked and corrected and thus, any gains from ai are in fact nullified.
>>108900956t. retarded webshitter
>>108900961Do you have a test suite?If you did, you wouldn't have to be a zealot looking at every LOC. That's what the test suite would be for
>>108900964The purpose of a test suite is profilingTDD proves you are a retarded webshitter
>>108900968No, purpose of a test suite is multifacetedWe write feature tests, unit tests. Feature test is for making sure shit works and you didnt break a feature when you introduce a changeYou're grasping at straws, i'm not this "webshitter" boogeyman of yours, and quite frankly I see you thread to thread accusing others of the same. Get over it
>>108900974>multifacetedgay corpo buzzword, cope webshitterThe purpose of tests is for benchmarkingIf you are attempting to check logic via TDD, that merely proves you've been writing webshitter slop before AI was even a thing
>>108900987The purpose of tests is not benchmarking, where the fuck do you work? Unit test - You do deep, edge-casing tests against functions here. Feature test - You guessed it, you test the fucking feature, as I described before. You'd probably put a harness on it, very simple.It has nothing at all to do benchmarking. Benchmarking is not a test, moron. Benchmarking falls under profiling, which would be used to ascertain if there are any poorly written functions, see where we would need to improve DB queries, and so on and so forth. Profiling is entirely separate from your tests.Now, I wanted to fucking help you, but clearly you refuse to listen to common sense from someone who is likely thrice your age. Good luck in the industry, you moron.
Trvke warning.You shouldn’t be able to speak on AI matters unless you are employed in tech.
>>108901017Credentialism is invalid. You don't get to dictate if a fact is true based on who said the thing. It only matters if the claim is true or not.
>>108901017>>108901005Being a webshitter isn't tech employment, you colossal retard
>>108901058non-argument, post github or shut up
>>108901078>webshitter resorts to socialsLOL you can't make this shit upThanks for proving you don't work in tech, you're a glorified HR worker
>>108900800You can also use it for review. You don't have to fully trust it, but it will at least find obvious mistakes.
>>108901108You don't work in software.
People keep bugging me to use AI to save time and eventually I give in an try it and it wastes my fucking time every time.But more importantly, not only does it waste my time, it actively frustrates me by consistently disregarding directives.The fact that it's not even a "you're using it wrong" is not the problem is extra annoying because you can't just get better at it, the models have to get better. It's not like you're going to learn one weird trick that makes the machine actually follow orders correctly, it's just a roll of the dice.
>>108901307Get claude code
>>108901355Claude code is just a frontend, Anthropics models are available outside of that and suffer the same issues.
>metr study comes out saying AI makes devs less productive>every anti AI nigger points to it as evidence>metr study does a followup fixing methodology flaws>shows an increase in productivity>suddenly it's lying with statistics, small sample size, etc>>108900262>>108900544>>108900380Hypocritical disingenuous nocoder faggots you lot are
>>108900250I can do way more than before, but that comes with a lot more complexity to juggle
>>108900250>devsWhat kind of devs>fasterAt doing what?You know the average """AI""" user is an imbecile just from the mindnumbing level of the propaganda targeting them.
>>108900299>>108900302What's the supposed source for this? It's written just a little too close to how a false flag would read for my liking, or a libertarian forum ad.>This is the 'ultimate victory' as the fonum is no longer participated with by the general public and no longer useful In maintaining their freedomsThe government is openly talking about eroding freedoms! Keep using my forum to stick it to them!>Remember these techniques are only effective if the forum participants 00 NOT KNOW ABOUT THEM. Once they are aware of these techniques the operation can completely fail, and the forum can become uncontrolled.Top secret! Do not share around and increase the popularity of my forum as a source for this!
>>108900727>Initial participants: 16Can't believe such a study ever had any influence.
>>108900436lol OP officially shown to be 20% more of a faggot.
>>108901017What if I spend 16 hours a day generating gay furry porn?
>>108900386I would assume it means 20% higher velocity (stories/bugs fixed/reviewed/done.)
>>108900302>demonstrate that only a minimal number of operatives can be used to eventually andeffectively gain a control of a 'uncontrolled forum'.WTF is with the grammar here? Is this written by a child?
>>108900250and costs 0.2 dev salary, we all know 20% faster dev = 20% more profit gogogogo
>>108900250So 100k a year for dev and 200k a year for tokens and it's only 20% faster.
>>108901383>>metr study does a followup fixing methodology flawsYou didn't read the paper or the thread.
>>108900250>officially
>>108900727>likely
>>108903637Yes. CIA actually stands for Child Internet Access.
>>10890025020 percent... and at what cost to quality and long-term maintainability?
>>108900743I believed that too, but actually it is probably smarter to let everyone race themselves to the bottom paying billions to anthropic on the way down just to fight with everyone else to slop out and copy the same software/features. Anthropic wins either way massively.
i'm not buying whatever you're sellingbuy an ad
>>108901383>fixing methodology flawsUnless you used fixing as an euphemism for cheating no they have not. See >>108900436
>>108900250one useful thing with slop is that you can go into a codebase youve never seen, in a language youve never used, and make a small qol change that only you would ever give a shit about
>>108900905The sample size is far too small for us to be able to conclude anything
>>108904698This is true not just for QoL but also actual fixesI used claude to fix Jellyfin broken Dolby Vision on WebOS after version 10.11.xI could submit the fix for everybody but fuck them got mine
study that agrees with me is right. study that disagrees with me flawed.
>>108904770only a shill would disagree that 16 and 57 are tiny sample sizes
I'm just happy to talk to my virtual reddit buddy :)
>>108900250>100>20% slower>80>20% faster>96So they haven't caught up to 2024 yet?
>>108900250> field that has never been able to measure productivity suddenly claims a big measurable increase in productivity due to <new hot buzzword>
>>108900669No
the constant shock from this board at corporate metrics being retarded because they're designed to demonstrate success not actually do anything practical is another one of the tells that none of you have ever actually been employed in tech ever.
>>108900302I wonder if anyone has ever devised a protocol for a forum resistant to that.Even if you hand out invites in-person, a bunch of bad actors can get and fuck it up without you noticing.
>>108905892the protocol is just public awareness and commit to naming, shaming, and doxing [within the law\ (not on 4chan, keep the rules of 4chan in mind) people who engage in forum shilling. this is already done by lots of watchdog organizations
>>108905892>>108905907the protocol is just eugenics to ensure only people with a minimum level of intelligence and criticism are allowed to exist
This just in: Software Developers are 20% faster while using a prompt based search engine that gathers and compiles all the correct code for you from stackoverflow automatically. dats crazybtw this new captcha is fucking cancer aids wtf happened to this place god damnit it sucks so much now
>>108901211I'm really wary of that. As you say, it can find obvious mistakes, and with the right context can do a decent enough job, but from my experience, people are already getting lazy on actual reviews because the AI review is good enough 90% of the time. It can be useful, but nobody seems to care about mitigating the risk.
>>108907890Ship now, fix the bugs later. What could go wrong?
>>108907574A 20% speed increase is fuckingg massive and it doesn't matter what methods it uses to get there. I'm relatively neutral on the whole AI thing and don't want it to take jobs, but you cannot try to talk down a 20% increase in productivity.
>>108907574Do the email verification and you only get 3 and they're easier. But it's pure aids to get to that point
>>108908011>A 20% speed increase is fuckingg massiveMake an algorithm that creates gibberish code. 1000%INCREASE IN SPEED!!!Speed is only used by vibecoders and ceo's as a metric for programming productivity.
>>108900250>"we changed our methodology">"the new estimate is not reliable"lollmao even
>>108900436Also they paid people $50/hour this time around, it was $150/hour last time. This selects for much worse devs.
