Coders spent more time prompting and reviewing AI generations than they saved on coding. On the surface, METR’s results seem to contradict other benchmarks and experiments that demonstrate increases in coding efficiency when AI tools are used. But those often also measure productivity in terms of total lines of code or the number of discrete tasks/code commits/pull requests completed, all of which can be poor proxies for actual coding efficiency. These factors lead the researchers to conclude that current AI coding tools may be particularly ill-suited to “settings with very high quality standards, or with many implicit requirements (e.g., relating to documentation, testing coverage, or linting/formatting) that take humans substantial time to learn.” While those factors may not apply in “many realistic, economically relevant settings” involving simpler code bases, they could limit the impact of AI tools in this study and similar real-world situations.
Having to repeatedly tweak and review AI generations is a code smell. Your gut could be telling you to start using your brain to build your code if you’re at this stage.
The main issue i have with AI coding, hasn’t been the code. Its a bit ham fisted and overly naive, it is as if it’s speed blind.
The main issue is that some of the code is out of date using functions that are deprecated etc, and it seems to be mixing paradigms and styles across languages in a very frustrating? way.
Yep I’ve got a working iOS app, a v.2 branched and on the way, with a ton of MapKit integrations. Unfortunately I’m getting depreciation errors and having to constantly remind the AI that it’s using old code, showing it examples of new code, and then watching it forget as we keep talking.
Still, I have a working iOS app, which only took a few hours. When Jack Dorsey said he’d vibe coded his new app in a long weekend, I’m like, hey me too.
LLMs can’t forget things because they are not capable of memory.
They can hold session memory including 10+ source files, and a looong chat, but when you run into the wall, suddenly it’s eating its own memory to keep going, rather than forcing me to reset the session. Which is interesting, like co-coding with a mild amnesiac. “Hey remember when we just did that thing 2 minutes ago?” I should have started a new session when I branched.
I don’t doubt this is true. I’ve been playing with an A.I and some fairly simple python scripts and it’s so tedious to get the A.I. to actually do something to the script correctly. Learning to prompt is a skill all it’s own.
In my experience it’s much more useful for doing things like in AWS like create a Cloudformation template or look through user permissions for excess privileges or setup a backup schedule, like at scale when you have lots of accounts and users, etc.
Removed by mod
deleted by creator
Their sample size was 16 people…
Who are in the process of learning to do something new, versus the workflow that they’ve been trained in and have a lot of experience in.
Where was the sample of non-coders tasked with doing the same thing, using AI to help or learning without assistance?
Where was the sample of coders prohibited from looking anything up and having to rely solely on their prior knowledge to do the job?
It might help refine what’s actually being tested.
I’m not really sure why it was such a small sample size. It definitely casts doubt on some of their conclusions. I also have issues with some methodology used. I think a better study that came out a week or two ago was the one that showed visible neurological decline from AI use.
Where the most experienced minority only had a few weeks of using AI inside an IDE like Cursor.
I got flamed pretty hard for pointing out that this sample size really needs to be in the title, but it needs to be said. Thank you. Sixteen people is basically a forum thread, and not a very popular one.
It’s still useful information and a good read, but a lot of people don’t click through to the article, they just remember the title and move on.
Great as an assistant for boring tasks. Still needs checking.
Can also help suggest improvements, but still needs checking.
Have to learn when to stop interacting with it and do it yourself.
A “junior” project manager at my company vibe coded an entire full stack web app with one of those LLM IDEs. His background is industrial engineering and claims to have basically no programming experience.
It “works”, as in, it does what it’s meant to, but as you can guess, it relies on calls to LLM APIs where it really doesn’t have to, and has several critical security flaws, inconsistencies in project structure and convention, and uses deprecated library features.
He already pitched it to one of our largest clients, and they’re on board. They want to start testing at the end of the month.
He’s had one junior dev who’s been managing to keep things somewhat stable, but the poor dude really had his work cut out for him. I only recently joined the project because “it sounded cool”, so I’ve been trying to fix some flaws while adding new requested features.
I’ve never worked with the frameworks and libraries before, so it’s a good opportunity to upskill, but god damn I don’t know if I want my name on this project.
A similar thing is happening with my brother at a different company. An executive vibe coded a web application, but this thing absolutely did not work.
My brother basically had one night to get it into a working state. He somehow (ritalin) managed to do it. The next day they presented it to one of their major clients. They really want it.
These AI dev tools absolutely have a direct negative impact on developer productivity, but they also have an indirect impact where non-devs use them and pass their Eldritch abominations to the actual devs to fix, extend and maintain.
Two years ago, I was worried about AI taking dev jobs, but now it feels like, to me, we’ll need more human devs than ever in the long run.
Like, weren’t these things supposed to exponentially get better? Like, cool, gh copilot can fuck up my project files now.
These AI dev tools absolutely have a direct negative impact on developer productivity, but they also have an indirect impact where non-devs use them and pass their Eldritch abominations to the actual devs to fix, extend and maintain.
Sounds like the next evolution of the Excel spreadsheet macro. Or maybe it’s convergent evolution toward the same niche. (I still have nightmares about Excel spreadsheet macros.)
This does not seem surprising to me:
“Overall, the developers in the study accepted less than 44 percent of the code generated by AI without modification. A majority of the developers reported needing to make changes to the code generated by their AI companion, and a total of 9 percent of the total task time in the “AI-assisted” portion of the study was taken up by this kind of review.”
It sounds about right. The AI should be acting as an assistant. The big question to me is if the code that comes out 19% slower is at all of higher quality. Since the coder is doing more correction and review does it act a bit like a second set of eyes or a pho sort of collaboration. If so it could still be helpful. Granted my experience so far is that most of what it does can be done with plugins to an ide but like it is sorta handy to have it all set and going after an installation without having to find and start using the plugins. Im still worried about energy usage with these things but hoping that can be worked out and honestly im not sure if the energy usage for something integrated with an ide or such is as bad.
Ad a fairly senior developer, I’m not at all surprised. AI speeds me up in some circumstances like writing boilerplate; things like kubernetes manifests. It does not speed up my coding, but it does help me explore options, expand my knowledge, and point me down the right track on new methods and packages. It also lets me do things I wouldn’t normally bother with, but which are good practice like finding edge cases for unit tests, packaging for multiple architectures, writing scripts to profile my code, etc.
Essentially, I’m likely slower writing code with AI assistance but I think the code is higher quality because it let’s me quickly assess many options and implement best practices that are normally tedious to implement manually.
I almost never accept code AI has written without modification, but I think I gain a lot from its use.
pho
faux
I like to think typos like that confirm my humanity :)
shhh don’t let the bots in on our secret
also now I’m hungry for phở
With enough training data from me and chatbots will spell like shit. Bad grammar as well.
The future has not been written. There is no fate but what we make for ourselves.
Maybe they’re making soup
You can tell it’s code soup by the smell
Sounds reasonable. The time and energy ive lost on trying very confident chat gpt suggestions that doesnt work must be weeks at this point.
Sometimes its very good though and really helps, which is why its so frustrating. You never know if its going to work before you go through the process.
It has changed how me and coworkers work now also. We just talk to chat gpt instead of even trying to look something up in the docs and trying to understand it. Too slow to do that now, it feels like. There is a pressure to solve anything quickly now that chat gpt exists.
You have to ignore the obsequious optimism bias LLM’s often have. It all comes down to their training set and if they have seen more than you have.
I don’t generally use them on projects I’m already familiar with unless it’s for fairly boring repetitive work that would be fiddly with search and replace, e.g. extract the common code out of these functions and refactor.
When working with unfamiliar code they can have an edge so if I needed a simple mobile app I’d probably give the LLM a go and then tidy up the code once it’s working.
At most I’ll give it 2 or 3 attempts to correct the original approach before I walk away and try something else. If it starts making up functions it APIs that don’t exist that is usually a sign out didn’t know so time to cut your losses and move on.
Their real strengths come in when it comes to digesting large amounts of text and sumerising. Great for saving you reading all the documentation on a project just to try a small thing. But if your going to work on the project going forward your going to want to invest that training data yourself.
True and not true at the same time. Using agents indeed often don’t work, mostly when I’m trying to do the wrong thing. Because then, AI agent does not say “the way you do it is overly complicated, it does not make any sense”, but instead it says: “excellent idea, here are X steps I need to do to make it happen”. It wasted my time many times, but it also guided me quickly though some problems that would take hours to research. Some of my projects wouldn’t have been finished without AI.
Some of my projects wouldn’t have been finished without AI.
This says way more about you than it says about AI tools
I was talking mostly about side projects. I don’t have much time for them right now. Thanks to LLMs, I can spend those few hours a week on doing instead of reading what is the best way to do X in ever changing world of web front-end frameworks. I just sit down, ask: “how is it usually done?”, tweak it a bit and finish.
Example: I have published an app on flathub a while ago. Doing it from scratch is damn complicated. “Screw it” is what I would say in pre LLMs era after a few hours ;)
Just make sure you’re validating everything you produce with it.
On a different note: is it just me or do images with this color scheme (that blue and black) also have a weird 3d look to them to you?
This article confirms my own experiences with AI. I spend a lot more time reviewing, reprompting, and tweaking than I save on coding. Having to double check or fight it to get what I want is not a time saver. Not to say that it doesn’t save time when it is right, but the thing that I never seem to get across to proponents of AI is that anytime I need to reprompt or refine, I have lost. I have officially wasted time at this point compared to simply referencing the documentation. Unless I’m generating a significant portion of code which only needs minor tweaks. I’m generally not saving time.
Slowing you down is the main benefit!
It helps you to keep more brain time on solving the actual problem, and less on boring syntax crap. Of course, then it gets the syntax crap wrong and you need to waste a lot of time fixing it.
They can’t read your mind. A professional painter is going to make the exact image they want in far less time and with more accuracy than repeatedly prompting a black box to make small changes.
But if you’re an amateur and don’t really know what you want, or you’re not very picky or care about quality, then meh good enough. High level software developers know what they want. They are like painters. And at that point, the LLM isn’t really solving problems for you. At best, it’s putting the paint to the canvas. That is, saving you typing time.
But time spent typing is definitely not the limiting factor for productivity in software.
They can’t read your mind. A professional painter is going to make the exact image they want in far less time and with more accuracy than repeatedly prompting a black box to make small changes.
and this is the exact reason why I hate IDEs that relentlessly “do things” for me.
I don’t need my editor maintaining my includes or updating my lock files. I don’t need them to auto complete words or fix syntax for me.
I know exactly what I’m doing. If I don’t then-- AND ONLY THEN, will I lookup what I need and fix it myself.
if there’s a problem with formatting a linter will pick it up. if there’s a problem with syntax the runtime/compilation will pick it up. if there’s a problem with content uat will pick it up.
we don’t need to be MORE productive, we need to be more skilled and using tools like these only soften the mind and dull the spirit.
Studies show that the electric drills drill faster than a manual, hand-cranked drill.
