I don't think this article is correct exactly, but I do feel that I'm less proud of my work. Less likely to go the extra mile. At first, I tried to do all the due diligence - reading and understanding all of the black box's output. But its clear what my workplace wants - more velocity, more code. If you take time reviewing, you're a blocker. If you lgtm that 3k LoC PR, that's great responsiveness. If you spend two days on a "simple fix" that involves broad cross cutting changes to the system and multiple library updates, you should be doing something else. We are all working across more areas of the system, less specialization, less understanding.
And it is great. It does produce fixes, produce a facimilie of understanding. It answers my questions, and is often right. And tinkering with the process of it is satisfying. Integrating more and more data, writing better specs, you can get better results. Its tempting to think that it could be sustainable, this way of working, but also so scary to lose the understanding, to not have the confidence in how things work. Finding duplicated stacks using different libraries, or even the same library, is becoming more and more common. Even our debugging tools, our tracing grow fragmented and unstandardized.
I liked the old way of working. It was fun for me, if often frustrating. It was solving hard sudoku on the train. This new way is lower friction, but more stress. It's steering a rocket ship using chopsticks to hold the wheel. You desperately want to slow things down and work methodically, to be sure, and safe. But you won't get anywhere near as far if you do that.
> Finding duplicated stacks using different libraries, or even the same library, is becoming more and more common. Even our debugging tools, our tracing grow fragmented.
Same - literally found a re-build of a library feature for use with the library the other day (e. g. MyCustomFooProviderFor(Bar) but Bar already literally has a `.foo` method.) No, it didn't need to be there.
7 years experience as a developer or thereabouts. Its probably been a year since the agentic coding stuff has become really widespread, picking up pace a lot around jan. Even the old hands, 20 years plus at the company and those few holdouts who refused to use AI before are deep in it now.
We're at a product company, not a consultancy. Hard to say exactly about tech, the tech is namely the product, but its b2b, so massive contracts move like glaciers, customer purchase decisions are often as much or more about the claims we made as the reality of the code.
As for outages, its the same as it always was. We have our testing, in layers. unit tests, integrations, e2e, staging envs. Layers and layers before it reaches the customer. If there ever is something that reaches there, as has happened, its so hard to pin the blame on AI, and of course we run a blameless culture here anyhow. Tickets are assigned, emergency patches are made, and the behemoth lumbers on. I don't think AI makes our defences much better though. We catch the things we always caught, and miss those we always missed, in greater and greater volume.
I don't pin blame on stupid management or whatever, I think this is complacency rather than a specific effort to push ai, as some claim. AI has just made it easier to work on more and understand less, and this is the result, no external intervention needed. I don't have a solution other than observing that trying to stop this is fighting the tides. People used to hate working on legacy codebases, where the original developers werent around to explain themselves, now everything is a legacy codebase right from inception - even if you personally don't use ai, the job is fundamentally different.
No problem. If anyone is feeling anxious about this stuff and wants to chat, my email is in my profile. I figure the only way through from here is forward, and talking about things is a great way to get a better understanding and approach solving them. It's long been a catchphrase of mine in meetings and discussions when people are speculating about some other person or team to say "let's go and ask them!". You'd be surprised how often people don't consider it.
"Developers talk not just about how the AI output is often flawed, but that using AI to get the job done is often a more time consuming, harder, and more frustrating experience because they have to go through the output and fix its mistakes."
This has not been my experience. Sure it feels like more work to fix the AI code problems sometimes - it is a different skillset than writing code from scratch. But the speed that I can deliver software has significantly increased by using coding agents.
Honestly, the effectiveness of LLMs in coding depends a lot on what you're working on. If you're dealing with a software package like Odoo that's been around for literal decades an LLMs output can be borderline useless. The problem is that in its training data it has examples from every version that's ever been released and each succeeding major version makes breaking changes to the previous one, so pretty much what happens is that the LLM can't accurately tell what in its training data belongs to which version before concocting a reply.
You're going to keep seeing this because people don't like AI adoption.
But the fact is this is not how it is. Every competent developer I know is delivering significantly more after being AI enabled.
Anyone seriously using the tools without a chip on their shoulder is going to say the same.
Are the tools delivering perfect code 100% of the time, no, of course not. But that's the new skill. Guiding them so they deliver good enough code at 5-50x the velocity.
As the models improve and the ecosystem tries out new workflows, the skill changes and the output gets better and better.
What we're capable of delivering now is incredible and would have been unimaginable just a few years ago.
Of course your claim of 5-50x velocity is not born out in any metrics which track industry software velocity and you need to bend yourself backwards to come up with reasons to explain why they aren't.
> tinygrad I picked on purpose. geohot narrates rejections in public, and a narrated rejection is data; a silent close is noise. Thirteen PRs, one merged, twelve closed. His comments tell the escalation story:
>> be careful with AI usage, we never trade complexity for speed
>> You need to stop with AI PRs, you will be banned.
>> Last warning about low quality PRs before I ban you from our GitHub.
>> I don’t even understand what this does. I’m not reading anything written by AI
> Each line a little more done with my shit than the last.
> Some of those PRs had real bugs with real fixes. The MATVEC pattern rejected equal-range elementwise reduces, a genuine correctness issue. But by that point the maintainer had stopped reading code and started reading provenance. “We never trade complexity for speed” is a valid engineering principle. “I’m not reading anything written by AI” is not.
> I went there for maximum surprise and got it. He had a review queue and a quality bar to protect; I had a clanker and a question. The price was his afternoon, three warnings, an account ban, and real bugs left unfixed.
Because this is Facebook-level "let's make people angry on the internet and see what happens" levels of treating people as if they were means to an end rather than an end in themselves. And you should stop.
No, of course not? I don't even disagree with your main premise but obviously "raw number of merged PRs" is not a high signal metric, even more so in the age of agentic/vibe coding.
I’d we are talking about adding code to a large existing production code base then there’s no way I can see to get 5x, let alone 50x. My experience and the data I’ve seen is more like a 10-20% improvement. We see the volume of code increase more than that but a lot of is bug fixes only necessary because the initial commit wasn’t adequately tested and reviewed. So the net effect is less.
Now if you mean generating some one off script or playing around with a prototype in some area you don’t know then I can see more like 5-10x but these are typically not the bottleneck for shipping software.
This is a business mindset, though. AI is great if you care about "delivering" stuff and "velocity" and you don't care what that stuff looks like. I got into computer programming because I like to program computers, not whatever this is. So glad I changed roles away from software development and only do programming at home as a hobby.
Woodworkers and furniture factories live happily side by side.
The unfortunate fact is that your boss or your customers never cared what your code looked like.
The craft will live on, no doubt, but the fact is that we're in the age of industrial programming.
Spending too much time twiddling line spacing, abstraction names, and dialing everything in just so is now for fun and not for profit.
Although to be honest, AI enables you to do that at scale too. It's never been easier to rename or refactor tens of thousands of lines to your hearts content.
They're expensive to perform and rarely are reproducible but I'll wait for the empirical studies before believing any claims.
We can't even decide if type systems have made us more productive. It's barely been studied. Same with test-driven development.
What it sounds like we'll see, from your description of AI-enabled developers, is a commensurate (perhaps linear) increase in the rate of errors reaching production systems. Every line of code is a liability. Now everyone has a fire-hose they can aim at a production environment.
At least time and effort prevented some bad ideas and potentially bad code from reaching production.
I'm sure the platforms providing these tools are going to be happy with the results when every business writing code this way becomes dependent on them and have no exit strategy. The prices increase, the service gets worse, and you're locked in. Sounds real productive.
I'm always wondering who has the time to consume all the new code that is being produced. Like sure, you can produce at 5-10X the speed, but is someone using those features? Not sure if the typical consumer mind can keep up with such speed of changes.
My embedded systems company has a very strict code quality standard that requires every single patch to be reviewed. This can be very time consuming especially when reviewing the deeper tech modules that have few experts. Imagining what it looks like to now be required to review 10x the code per day is a little dire... I can imagine developers simply turning into code reviewers for the bots. Is this the future? Or do we eventually turn the bots loose on code review as well?
We have been using Copilot for a year or two but it's not required. Any developer who asks for a license gets one. So far I haven't seen anyone get to the point of prompting it to write entire features at a time.
Lol, it's not sustained 50x for every task every minute.
But there are definitely many tasks that used to take a very long time that now take almost no time at all, and that can be delivered in parallel with other tasks.
> What we're capable of delivering now is incredible and would have been unimaginable just a few years ago
What I mean is - are there concrete examples, real world "things" that came from AI programming, that are incredible, and someone can talk about and point to how AI led to the thing being possible?
They’ve said competent. Obviously you’re not competent enough to understand it. Try to feed their message into latest Opus and ask “explain like I’m five”. Good luck!
I'm someone with 20 years in software and the last 10 in management. I have good instincts, good design pattern knowledge, and understand system design well. But my actual coding skills are rusty, I can do it but it takes a lot of time to RTFM because specific libraries and syntax aren't on the top of my mind.
With AI I can build. I'm having so much fun turning ideas into code. I can do a week's worth of work before lunch. I can ask AI to add comments so detailed that my code becomes a refresher tutorial.
It's so exciting to be able to bring my ideas to life, make use of my experience, and not be hobbled by my somewhat atrophied hands-on coding skills. I for one welcome this revolution.
I keep seeing this point made, but what happens when there is not enough demand for all this code that's being produced?
I'm always seeing these "I can finally make my projects and slack off at work!" but I just can't help but feel like people aren't thinking about what comes next
Who said anything about demand. Just let the addict scratch his OCD. It’s a standard value distribution from middle class to hyperscalers. System works as intended.
5-50x huh? Years of AI hype and yet still to this day, not a single person or organization can provide any kind of reputable evidence that it has significantly increased their productivity.
> At Meta, Google, Microsoft, and others, leadership says that AI generates a growing share of the overall code
Probably because they mandate its adoption. And while there are plenty of developers who will happily comply and see it as a good thing. There are others who will do it because they have to or risk losing their jobs.
It's a bit of a silly thing to claim. "We made everyone use it, so they did, and now adoption is going up!"
A lot of the code generated isn’t for the products shipped to users. At Meta someone told me that around 80% of generated code is for internal tools and dashboards (most of which have no usage).
I have just stepped down from a CTO job where I built a FinTech's stack from the ground up. I leveraged a Claude Max plan for about 8 months and I can say with absolute certainty I would not have been as productive without it. I have barely written any code by hand during that time, but I did read almost every single line of code produced. My role was much more that of an "editor", as another article posted here mentioned. There is no doubt in my mind that you can be very highly productive with AI, it's just not the magical silver bullet some people market it as. I have a lot of notes describing the process that I'm considering publishing.
Just yesterday I was interviewing for a very interesting job and I completely flunked the coding question in an unacceptable way for my level of experience. The question was easy, I just couldn't get past some syntactic issues. For 8 months, Claude wrote all of my Python classes and Pydantic types. Now I had to write a dataclass, and because I always just resorted to standard classes before the advent of LLMs, I stumbled. And froze. And panicked. And that was it. Of course you could say I should have just scrapped the dataclass and written it as a simple class. The point is I felt very, very stupid. LLMs suddenly felt like a huge disadvantage.
All this to say I disagree with LLMs "rotting" my brain. Quite the opposite, I know that it's possible to use LLMs to be efficient and correct. It's more the actual mechanical act of writing that gets rusty.
It's pretty interesting because if you look at the net value over the last 2 to 3 years, you'd expect to see a flurry of high value/high complexity/high velocity software being delivered. But we haven't. We've seen outages almost normalized now and the only new thing being built are more AI integrations - and I have to ask: for whom? To me, there's a pretty big gap between the delivery and claim of AI, and the question is - do you stay an IC or do you become an agent manager (in which case you will lose your technical edge for sure).
Is this substantially different, cognitively or skills-wise, than moving into management and directing a team to write code, but no longer writing code oneself?
Yesterday I had been talking to an AI all day, and left work with a feeling of non-accomplishment even though I probably did slightly more than I would do normally, though time will tell if the maintenance costs will be higher or not.
I think Andrej Karpathy's quote summarizes well what all software engineers are going through:
,,you can outsource your thinking but not your understanding''
There's just no way to not generate much more amount of code with LLMs than we would do as humans, so well structuring code gets much more important than ever before.
The compounding issue is that understanding atrophies without thinking constantly reinforcing it. Shed no tears for the boilerplate, but the exercise is useful.
I've used LLMs to help with code since before vibe coding was termed.
Experienced mental pains I never felt with any other activity except watching tiktok reels for hours.
Got into points of no returns on numerous side projects, ai slop neither ai or myself could touch.
I've developed a better mental loop. I simply review every lines of code it spits out, and refine the loop to get less code produced. But always demand the full file again.
I commit each change. And inspect the diff for review.
I don't feel drain or pain.
LLMs still aren't standalone developers, but they can be tamed to execute well on well defined scope. If we review what they do, every time.
The only use case for AI is for looking up historical references and current events. The latter is probably the most used part, which is why models are only useful if they scrape news sites.
You can also use it for regurgitating manuals, but generative AI for coding is counterproductive. Only the tool and gaming addicted people like it and pretend to be more productive, for which there is no public evidence. I don't see any software improving at any faster rate.
> Only the tool and gaming addicted people like it and pretend to be more productive, for which there is no public evidence. I don't see any software improving at any faster rate.
I’m curios, which models have you used? I’ve been using Shmopus 69 and it’s out of the world, it’s so good that I don’t understand how people existed before it.
I wonder how much this is correlated to token budgets? I'd be curious to see a split between $20/$100/$200/$500+ usage and see if there is different responses. I'm in the $400 range with Claude + Cursor subscription, use Opus exclusively, and my experience is wildly different from this.
Totally this. My employer provides $2k in tokens and I personally maxed out 3 credit cards just to continue spawning agents. Never felt so smart and enlightened before! Remove the last sentence before posting.
In my case it's less about actual "rotting" and more about the feeling that any mine attempts to write code are futile and meaningless - if my LLM limits are exhausted it's actually more productive to go do something else (or write specs on how it should be done) and return back later and do LLM assisted coding than coding without it because in literally minutes I can then produce equivalent of hours-long "manual" coding session.
AI agents have made me far more productive, but the work now feels like drudgery. The most intellectually stimulating parts of the job were automated away first, and I am getting increasingly sick of typing into a chat bot all day.
I got into software engineering because I was always fascinated by getting computers to do stuff, and I really enjoyed the manual task of programming. It's been a dream to earn a living doing something I would do in my spare time. I was pretty good at it too.
I'm not having fun any more, so I've decided to leave the field and become a teacher. I won't earn nearly as much money but I expect to feel more fulfilled, and I hope I can help make a difference to some young people.
I've had an extraordinarily privileged career, and many people never get the luxury of enjoying their work at all. But I'd rather try to enjoy what I do day to day than persist in something that's lost its spark.
I am still coding interesting and complex stuff manually. I just make the AI do the boring stuff like data processing scripts or setting up tests. So it works kind of well for me
That's what I'd like to do in my spare time. My job has become intolerant of that slow pace though now they've drunk the kool aid. I work at a startup and we're expected to produce game changing new features every day.
I'm curious, how do you think people around you there are taking it? I just can't help but feel like that is unsustainable and everyone is just going to burn out
I think everyone is getting a bit sick of it - one person left 2-3 months ago, another said he was working on his CV. I'm sure there would be more sane companies to work for in 2026, but I'm not sure about 2030.
Fact is, I can see in myself that I can produce more business value using AI, and I don't enjoy it. So even if I could get away with a slower pace somewhere else for a bit, I don't think we're ever truly getting back to the way things were before.
I've had the same experience. I used to enjoy doing software dev, I was good at my job and liked it and did good work. Then the AI push happened and now whenever I'm typing code I think "I wonder if AI could do this for me" except using AI is infantilizing and boring and I don't want to do it. So I feel bad if I use it and I feel bad if I don't use it. So mostly I go post on HN or something instead of working and my productivity has tanked in the last year. Luckily I'm nearer the end of my career than the beginning so it won't be a big financial impact to me when I finally leave the industry, hopefully later this year.
Commercial software development will increasingly become dominated by the 'get shit done' types who had less appreciation for the craft. The slop will flow and no one will care, because the people who cared for the craft will have left or been pushed out. A shame.
My gf, who’s multiple levels below me in exp, is already expected to write full stack + CI for the same salary. I still have the luxury of seniority to be able to say “fuck off”, but not for long.
You’ll do the work of 10 people and be happy, now you’re all 10x developers for 1x pay, rejoice!
404 media tends to put out quality articles in my opinion but this one feels a bit like clickbait.
It seems like they're overgeneralizing quite a bit here and focusing on a narrow subset of the population while ignoring the people who are actually thriving with their new AI-enabled dev workflows.
LLMs are not a panacea by any means and they have lots of cons. But I for one would find it difficult to go back to a world where I can't lean on LLMs in my day-to-day.
One very specific example that could not possibly contribute to the brainrot mentioned in this article: AI saves time and reduces the headache of having to pore through pages of documentation (if there even is any) to find how that one method works or what arguments it can take. This alone is immensely helpful and can keep you in a state of flow instead of sending you off on a potentially fruitless side quest that derails your whole train of thought.
It's also taken me quite a bit of time, effort, and experimentation to find the right tools and the right ways to work AI into my workflows which I would bet that the developers mentioned in this article have not explored too deeply if at all.
Claiming AI is rotting your brain because you can't one-shot an entire app or even a single feature is a straw man fallacy.
One of my engineers used ai yesterday to write a thing without having to "pore over the docs". Engineer and ai decided the third party api was "inconsistent" and "nondeterministic" so they wrote a script to hit it over and over and over hoping to "catch" any "missed" entries. Built a whole cache and mini db to track all the seen IDs as it hammers the api. Luckily I was there to point out how strange it would be for the api to work that way, at which point they read the docs and saw they weren't using the pagination cursor it requires.
Thousands of lines and hours of wasted time and this was the lucky path because a DIFFERENT human happened to be in the loop and asked the right question.
This isn't a general claim about what ai does or doesn't do, but it is a real life anecdote about a very well paid professional.
Someone posted a great quote above that you can outsource your thinking but you can't outsource your understanding.
> Engineer and ai decided the third party api was "inconsistent" and "nondeterministic"
Sounds like it was the AI that decided this and the engineer didn't bother questioning it which I'd classify as using AI incorrectly. AI is a smart intern, not a smart engineer.
So if I were to continue with your logic, our walking and running skills or level of physical fitness have atrophied because we can drive or take public transit. Should we no longer drive or take public transit anywhere?
To your point, my documentation reading skills have certainly atrophied.
I'm not coding as much so my coding skill has likely atrophied to an extent.
It does take intentional effort to counteract this, which is why I will force myself to write code by hand still. Or why I carefully parse PR diffs and will not approve it unless I can explain what it's doing and why it's doing it that way.
I can understand why you would jump to such an assumption though. There is nuance to everything.
I have usually positive experience with AI. What it excels in are tasks which are having clear boundaries and proper context.
I have also worked in customer support for some time and I have found that huge problem for some people (often times developers) is that they are lacking theory of mind. Like they literally can't comprehend that I don't see into their heads and they need to articulate their question with correct context otherwise I can't help them.
AI is like a litmus test for it. People who have theory of mind, are capable of putting together a question which will give them good results out of AI. On the other hand people who are struggling with the fact that AI can't see what you mean unless it is in a context window will have bad time with it. These people also usually suck in managing other people because - once again - they are unable to provide tasks with enough context and properly set boundaries. At best they will give you some vague poorly defined tasks and get mad when you will do it differently than they had in their mind.
I don't think this article is correct exactly, but I do feel that I'm less proud of my work. Less likely to go the extra mile. At first, I tried to do all the due diligence - reading and understanding all of the black box's output. But its clear what my workplace wants - more velocity, more code. If you take time reviewing, you're a blocker. If you lgtm that 3k LoC PR, that's great responsiveness. If you spend two days on a "simple fix" that involves broad cross cutting changes to the system and multiple library updates, you should be doing something else. We are all working across more areas of the system, less specialization, less understanding.
And it is great. It does produce fixes, produce a facimilie of understanding. It answers my questions, and is often right. And tinkering with the process of it is satisfying. Integrating more and more data, writing better specs, you can get better results. Its tempting to think that it could be sustainable, this way of working, but also so scary to lose the understanding, to not have the confidence in how things work. Finding duplicated stacks using different libraries, or even the same library, is becoming more and more common. Even our debugging tools, our tracing grow fragmented and unstandardized.
I liked the old way of working. It was fun for me, if often frustrating. It was solving hard sudoku on the train. This new way is lower friction, but more stress. It's steering a rocket ship using chopsticks to hold the wheel. You desperately want to slow things down and work methodically, to be sure, and safe. But you won't get anywhere near as far if you do that.
Somewhere quiet, the tech debt demon smiles.
> Finding duplicated stacks using different libraries, or even the same library, is becoming more and more common. Even our debugging tools, our tracing grow fragmented.
Same - literally found a re-build of a library feature for use with the library the other day (e. g. MyCustomFooProviderFor(Bar) but Bar already literally has a `.foo` method.) No, it didn't need to be there.
I have so many questions.
How long have you been doing this?
Are you at a product company, a consultancy, a place where technology is an enabler but not core, or somewhere else?
What happens when there are bugs or an outage due to that 3k LoC PR?
7 years experience as a developer or thereabouts. Its probably been a year since the agentic coding stuff has become really widespread, picking up pace a lot around jan. Even the old hands, 20 years plus at the company and those few holdouts who refused to use AI before are deep in it now.
We're at a product company, not a consultancy. Hard to say exactly about tech, the tech is namely the product, but its b2b, so massive contracts move like glaciers, customer purchase decisions are often as much or more about the claims we made as the reality of the code.
As for outages, its the same as it always was. We have our testing, in layers. unit tests, integrations, e2e, staging envs. Layers and layers before it reaches the customer. If there ever is something that reaches there, as has happened, its so hard to pin the blame on AI, and of course we run a blameless culture here anyhow. Tickets are assigned, emergency patches are made, and the behemoth lumbers on. I don't think AI makes our defences much better though. We catch the things we always caught, and miss those we always missed, in greater and greater volume.
I don't pin blame on stupid management or whatever, I think this is complacency rather than a specific effort to push ai, as some claim. AI has just made it easier to work on more and understand less, and this is the result, no external intervention needed. I don't have a solution other than observing that trying to stop this is fighting the tides. People used to hate working on legacy codebases, where the original developers werent around to explain themselves, now everything is a legacy codebase right from inception - even if you personally don't use ai, the job is fundamentally different.
Thanks a lot for sharing your story. It's incredibly valuable to hear specifics like that.
No problem. If anyone is feeling anxious about this stuff and wants to chat, my email is in my profile. I figure the only way through from here is forward, and talking about things is a great way to get a better understanding and approach solving them. It's long been a catchphrase of mine in meetings and discussions when people are speculating about some other person or team to say "let's go and ask them!". You'd be surprised how often people don't consider it.
"Developers talk not just about how the AI output is often flawed, but that using AI to get the job done is often a more time consuming, harder, and more frustrating experience because they have to go through the output and fix its mistakes."
This has not been my experience. Sure it feels like more work to fix the AI code problems sometimes - it is a different skillset than writing code from scratch. But the speed that I can deliver software has significantly increased by using coding agents.
Honestly, the effectiveness of LLMs in coding depends a lot on what you're working on. If you're dealing with a software package like Odoo that's been around for literal decades an LLMs output can be borderline useless. The problem is that in its training data it has examples from every version that's ever been released and each succeeding major version makes breaking changes to the previous one, so pretty much what happens is that the LLM can't accurately tell what in its training data belongs to which version before concocting a reply.
>This has not been my experience.
>But the speed that I...
First-hand experience is perfectly valid.
I agree with the parent; I'm able to produce more. And with proper documentation and unit tests in place, I don't feel I need to review every line.
You're going to keep seeing this because people don't like AI adoption.
But the fact is this is not how it is. Every competent developer I know is delivering significantly more after being AI enabled.
Anyone seriously using the tools without a chip on their shoulder is going to say the same.
Are the tools delivering perfect code 100% of the time, no, of course not. But that's the new skill. Guiding them so they deliver good enough code at 5-50x the velocity. As the models improve and the ecosystem tries out new workflows, the skill changes and the output gets better and better.
What we're capable of delivering now is incredible and would have been unimaginable just a few years ago.
Of course your claim of 5-50x velocity is not born out in any metrics which track industry software velocity and you need to bend yourself backwards to come up with reasons to explain why they aren't.
are merged PRs a measure of velocity? github.com/kimjune01/
This you?
https://june.kim/speedrunning-open-source
> tinygrad I picked on purpose. geohot narrates rejections in public, and a narrated rejection is data; a silent close is noise. Thirteen PRs, one merged, twelve closed. His comments tell the escalation story:
>> be careful with AI usage, we never trade complexity for speed
>> You need to stop with AI PRs, you will be banned.
>> Last warning about low quality PRs before I ban you from our GitHub.
>> I don’t even understand what this does. I’m not reading anything written by AI
> Each line a little more done with my shit than the last.
> Some of those PRs had real bugs with real fixes. The MATVEC pattern rejected equal-range elementwise reduces, a genuine correctness issue. But by that point the maintainer had stopped reading code and started reading provenance. “We never trade complexity for speed” is a valid engineering principle. “I’m not reading anything written by AI” is not.
> I went there for maximum surprise and got it. He had a review queue and a quality bar to protect; I had a clanker and a question. The price was his afternoon, three warnings, an account ban, and real bugs left unfixed.
Because this is Facebook-level "let's make people angry on the internet and see what happens" levels of treating people as if they were means to an end rather than an end in themselves. And you should stop.
Lil bro thinks he’s Mario Zechner, lol.
No, of course not? I don't even disagree with your main premise but obviously "raw number of merged PRs" is not a high signal metric, even more so in the age of agentic/vibe coding.
No, not any more than lines of code written are measures of velocity
I could easily enjoy a 10x improvement when working on something I'm not familiar with once you factor in the learning time.
I’d we are talking about adding code to a large existing production code base then there’s no way I can see to get 5x, let alone 50x. My experience and the data I’ve seen is more like a 10-20% improvement. We see the volume of code increase more than that but a lot of is bug fixes only necessary because the initial commit wasn’t adequately tested and reviewed. So the net effect is less.
Now if you mean generating some one off script or playing around with a prototype in some area you don’t know then I can see more like 5-10x but these are typically not the bottleneck for shipping software.
This is a business mindset, though. AI is great if you care about "delivering" stuff and "velocity" and you don't care what that stuff looks like. I got into computer programming because I like to program computers, not whatever this is. So glad I changed roles away from software development and only do programming at home as a hobby.
Woodworkers and furniture factories live happily side by side.
The unfortunate fact is that your boss or your customers never cared what your code looked like.
The craft will live on, no doubt, but the fact is that we're in the age of industrial programming.
Spending too much time twiddling line spacing, abstraction names, and dialing everything in just so is now for fun and not for profit.
Although to be honest, AI enables you to do that at scale too. It's never been easier to rename or refactor tens of thousands of lines to your hearts content.
They're expensive to perform and rarely are reproducible but I'll wait for the empirical studies before believing any claims.
We can't even decide if type systems have made us more productive. It's barely been studied. Same with test-driven development.
What it sounds like we'll see, from your description of AI-enabled developers, is a commensurate (perhaps linear) increase in the rate of errors reaching production systems. Every line of code is a liability. Now everyone has a fire-hose they can aim at a production environment.
At least time and effort prevented some bad ideas and potentially bad code from reaching production.
I'm sure the platforms providing these tools are going to be happy with the results when every business writing code this way becomes dependent on them and have no exit strategy. The prices increase, the service gets worse, and you're locked in. Sounds real productive.
I'm always wondering who has the time to consume all the new code that is being produced. Like sure, you can produce at 5-10X the speed, but is someone using those features? Not sure if the typical consumer mind can keep up with such speed of changes.
My embedded systems company has a very strict code quality standard that requires every single patch to be reviewed. This can be very time consuming especially when reviewing the deeper tech modules that have few experts. Imagining what it looks like to now be required to review 10x the code per day is a little dire... I can imagine developers simply turning into code reviewers for the bots. Is this the future? Or do we eventually turn the bots loose on code review as well?
We have been using Copilot for a year or two but it's not required. Any developer who asks for a license gets one. So far I haven't seen anyone get to the point of prompting it to write entire features at a time.
> 50x the velocity
:blinks: You are producing in a week what used to take you a year?
Lol, it's not sustained 50x for every task every minute.
But there are definitely many tasks that used to take a very long time that now take almost no time at all, and that can be delivered in parallel with other tasks.
That's a very silly claim to make, because you can make that same claim about writing a bash script.
Yes, but what _percentage_ are they? Or is this the XKCD optimization graph all over again?
And at the same time they talk about "competent developer"s
> Guiding them so they deliver good enough code at 5-50x the velocity.
Huge problem with this is the rate at which anyone can take accountability for the code produced.
Of course you can let AI do reviews, but my experience so far is that it's, broadly speaking, not working.
Can someone translate this?
> What we're capable of delivering now is incredible and would have been unimaginable just a few years ago
What I mean is - are there concrete examples, real world "things" that came from AI programming, that are incredible, and someone can talk about and point to how AI led to the thing being possible?
They’ve said competent. Obviously you’re not competent enough to understand it. Try to feed their message into latest Opus and ask “explain like I’m five”. Good luck!
After going back and forth, I stopped using AI for coding at all.
Maybe I am not "competent" developer, but the point has some merit.
I feel the same way.
Even if I'm reviewing more, I built the feature without even opening my editor.
My workflow is:
1. Plan mode 2. Read thoroughly or skim if its an easy task 3. /draft command that puts a draft PR on github 4. Review closely then send to team
5. Cover while some poor sod is about to explode from another 3k lines slop PR
I'm someone with 20 years in software and the last 10 in management. I have good instincts, good design pattern knowledge, and understand system design well. But my actual coding skills are rusty, I can do it but it takes a lot of time to RTFM because specific libraries and syntax aren't on the top of my mind.
With AI I can build. I'm having so much fun turning ideas into code. I can do a week's worth of work before lunch. I can ask AI to add comments so detailed that my code becomes a refresher tutorial.
It's so exciting to be able to bring my ideas to life, make use of my experience, and not be hobbled by my somewhat atrophied hands-on coding skills. I for one welcome this revolution.
I keep seeing this point made, but what happens when there is not enough demand for all this code that's being produced?
I'm always seeing these "I can finally make my projects and slack off at work!" but I just can't help but feel like people aren't thinking about what comes next
Who said anything about demand. Just let the addict scratch his OCD. It’s a standard value distribution from middle class to hyperscalers. System works as intended.
I love AI, but it's possible that we are in an temporary golden age of software development because of 3 things:
1. The software is simple because lowly humans wrote them and debugged them and maintained them.
2. The humans are competent in software engineering.
3. All of a sudden we now have help from AI.
Point 3. is here to stay, but 1. and 2. could disappear.
5-50x huh? Years of AI hype and yet still to this day, not a single person or organization can provide any kind of reputable evidence that it has significantly increased their productivity.
I’ll release the metrics in 2 years (after vesting), but for now I’ve never been so productive! Like 100x!
> At Meta, Google, Microsoft, and others, leadership says that AI generates a growing share of the overall code
Probably because they mandate its adoption. And while there are plenty of developers who will happily comply and see it as a good thing. There are others who will do it because they have to or risk losing their jobs.
It's a bit of a silly thing to claim. "We made everyone use it, so they did, and now adoption is going up!"
A lot of the code generated isn’t for the products shipped to users. At Meta someone told me that around 80% of generated code is for internal tools and dashboards (most of which have no usage).
I have just stepped down from a CTO job where I built a FinTech's stack from the ground up. I leveraged a Claude Max plan for about 8 months and I can say with absolute certainty I would not have been as productive without it. I have barely written any code by hand during that time, but I did read almost every single line of code produced. My role was much more that of an "editor", as another article posted here mentioned. There is no doubt in my mind that you can be very highly productive with AI, it's just not the magical silver bullet some people market it as. I have a lot of notes describing the process that I'm considering publishing.
Just yesterday I was interviewing for a very interesting job and I completely flunked the coding question in an unacceptable way for my level of experience. The question was easy, I just couldn't get past some syntactic issues. For 8 months, Claude wrote all of my Python classes and Pydantic types. Now I had to write a dataclass, and because I always just resorted to standard classes before the advent of LLMs, I stumbled. And froze. And panicked. And that was it. Of course you could say I should have just scrapped the dataclass and written it as a simple class. The point is I felt very, very stupid. LLMs suddenly felt like a huge disadvantage.
All this to say I disagree with LLMs "rotting" my brain. Quite the opposite, I know that it's possible to use LLMs to be efficient and correct. It's more the actual mechanical act of writing that gets rusty.
Doesn't open as it is for me.
https://archive.is/2vjJm
It's pretty interesting because if you look at the net value over the last 2 to 3 years, you'd expect to see a flurry of high value/high complexity/high velocity software being delivered. But we haven't. We've seen outages almost normalized now and the only new thing being built are more AI integrations - and I have to ask: for whom? To me, there's a pretty big gap between the delivery and claim of AI, and the question is - do you stay an IC or do you become an agent manager (in which case you will lose your technical edge for sure).
Is this substantially different, cognitively or skills-wise, than moving into management and directing a team to write code, but no longer writing code oneself?
Yesterday I had been talking to an AI all day, and left work with a feeling of non-accomplishment even though I probably did slightly more than I would do normally, though time will tell if the maintenance costs will be higher or not.
And I used to love my work :(
AI is not required for this to happen
I think Andrej Karpathy's quote summarizes well what all software engineers are going through:
,,you can outsource your thinking but not your understanding''
There's just no way to not generate much more amount of code with LLMs than we would do as humans, so well structuring code gets much more important than ever before.
The compounding issue is that understanding atrophies without thinking constantly reinforcing it. Shed no tears for the boilerplate, but the exercise is useful.
I've used LLMs to help with code since before vibe coding was termed.
Experienced mental pains I never felt with any other activity except watching tiktok reels for hours.
Got into points of no returns on numerous side projects, ai slop neither ai or myself could touch.
I've developed a better mental loop. I simply review every lines of code it spits out, and refine the loop to get less code produced. But always demand the full file again.
I commit each change. And inspect the diff for review.
I don't feel drain or pain.
LLMs still aren't standalone developers, but they can be tamed to execute well on well defined scope. If we review what they do, every time.
The only use case for AI is for looking up historical references and current events. The latter is probably the most used part, which is why models are only useful if they scrape news sites.
You can also use it for regurgitating manuals, but generative AI for coding is counterproductive. Only the tool and gaming addicted people like it and pretend to be more productive, for which there is no public evidence. I don't see any software improving at any faster rate.
> Only the tool and gaming addicted people like it and pretend to be more productive, for which there is no public evidence. I don't see any software improving at any faster rate.
I’m curios, which models have you used? I’ve been using Shmopus 69 and it’s out of the world, it’s so good that I don’t understand how people existed before it.
This reminds of Facebook will go the way of Myspace any day now rhetoric.
I wonder how much this is correlated to token budgets? I'd be curious to see a split between $20/$100/$200/$500+ usage and see if there is different responses. I'm in the $400 range with Claude + Cursor subscription, use Opus exclusively, and my experience is wildly different from this.
Totally this. My employer provides $2k in tokens and I personally maxed out 3 credit cards just to continue spawning agents. Never felt so smart and enlightened before! Remove the last sentence before posting.
In my case it's less about actual "rotting" and more about the feeling that any mine attempts to write code are futile and meaningless - if my LLM limits are exhausted it's actually more productive to go do something else (or write specs on how it should be done) and return back later and do LLM assisted coding than coding without it because in literally minutes I can then produce equivalent of hours-long "manual" coding session.
AI agents have made me far more productive, but the work now feels like drudgery. The most intellectually stimulating parts of the job were automated away first, and I am getting increasingly sick of typing into a chat bot all day.
I got into software engineering because I was always fascinated by getting computers to do stuff, and I really enjoyed the manual task of programming. It's been a dream to earn a living doing something I would do in my spare time. I was pretty good at it too.
I'm not having fun any more, so I've decided to leave the field and become a teacher. I won't earn nearly as much money but I expect to feel more fulfilled, and I hope I can help make a difference to some young people.
I've had an extraordinarily privileged career, and many people never get the luxury of enjoying their work at all. But I'd rather try to enjoy what I do day to day than persist in something that's lost its spark.
I am still coding interesting and complex stuff manually. I just make the AI do the boring stuff like data processing scripts or setting up tests. So it works kind of well for me
That's what I'd like to do in my spare time. My job has become intolerant of that slow pace though now they've drunk the kool aid. I work at a startup and we're expected to produce game changing new features every day.
I'm curious, how do you think people around you there are taking it? I just can't help but feel like that is unsustainable and everyone is just going to burn out
I think everyone is getting a bit sick of it - one person left 2-3 months ago, another said he was working on his CV. I'm sure there would be more sane companies to work for in 2026, but I'm not sure about 2030.
Fact is, I can see in myself that I can produce more business value using AI, and I don't enjoy it. So even if I could get away with a slower pace somewhere else for a bit, I don't think we're ever truly getting back to the way things were before.
I've had the same experience. I used to enjoy doing software dev, I was good at my job and liked it and did good work. Then the AI push happened and now whenever I'm typing code I think "I wonder if AI could do this for me" except using AI is infantilizing and boring and I don't want to do it. So I feel bad if I use it and I feel bad if I don't use it. So mostly I go post on HN or something instead of working and my productivity has tanked in the last year. Luckily I'm nearer the end of my career than the beginning so it won't be a big financial impact to me when I finally leave the industry, hopefully later this year.
The emperor has semi-transparent clothes.
Commercial software development will increasingly become dominated by the 'get shit done' types who had less appreciation for the craft. The slop will flow and no one will care, because the people who cared for the craft will have left or been pushed out. A shame.
My gf, who’s multiple levels below me in exp, is already expected to write full stack + CI for the same salary. I still have the luxury of seniority to be able to say “fuck off”, but not for long.
You’ll do the work of 10 people and be happy, now you’re all 10x developers for 1x pay, rejoice!
404 media tends to put out quality articles in my opinion but this one feels a bit like clickbait.
It seems like they're overgeneralizing quite a bit here and focusing on a narrow subset of the population while ignoring the people who are actually thriving with their new AI-enabled dev workflows.
LLMs are not a panacea by any means and they have lots of cons. But I for one would find it difficult to go back to a world where I can't lean on LLMs in my day-to-day.
One very specific example that could not possibly contribute to the brainrot mentioned in this article: AI saves time and reduces the headache of having to pore through pages of documentation (if there even is any) to find how that one method works or what arguments it can take. This alone is immensely helpful and can keep you in a state of flow instead of sending you off on a potentially fruitless side quest that derails your whole train of thought.
It's also taken me quite a bit of time, effort, and experimentation to find the right tools and the right ways to work AI into my workflows which I would bet that the developers mentioned in this article have not explored too deeply if at all.
Claiming AI is rotting your brain because you can't one-shot an entire app or even a single feature is a straw man fallacy.
One of my engineers used ai yesterday to write a thing without having to "pore over the docs". Engineer and ai decided the third party api was "inconsistent" and "nondeterministic" so they wrote a script to hit it over and over and over hoping to "catch" any "missed" entries. Built a whole cache and mini db to track all the seen IDs as it hammers the api. Luckily I was there to point out how strange it would be for the api to work that way, at which point they read the docs and saw they weren't using the pagination cursor it requires.
Thousands of lines and hours of wasted time and this was the lucky path because a DIFFERENT human happened to be in the loop and asked the right question.
This isn't a general claim about what ai does or doesn't do, but it is a real life anecdote about a very well paid professional.
Someone posted a great quote above that you can outsource your thinking but you can't outsource your understanding.
> Engineer and ai decided the third party api was "inconsistent" and "nondeterministic"
Sounds like it was the AI that decided this and the engineer didn't bother questioning it which I'd classify as using AI incorrectly. AI is a smart intern, not a smart engineer.
Well it’s certainly not “clickbait” since the title clearly tells you what it’s about.
> But I for one would find it difficult to go back to a world where I can't lean on LLMs in my day-to-day
It is difficult for me to read this and believe your brain isn't rotted
If you've become reliant on it, then your skills have atrophied. Your brain has rotted.
So if I were to continue with your logic, our walking and running skills or level of physical fitness have atrophied because we can drive or take public transit. Should we no longer drive or take public transit anywhere?
To your point, my documentation reading skills have certainly atrophied.
I'm not coding as much so my coding skill has likely atrophied to an extent.
It does take intentional effort to counteract this, which is why I will force myself to write code by hand still. Or why I carefully parse PR diffs and will not approve it unless I can explain what it's doing and why it's doing it that way.
I can understand why you would jump to such an assumption though. There is nuance to everything.
I have usually positive experience with AI. What it excels in are tasks which are having clear boundaries and proper context.
I have also worked in customer support for some time and I have found that huge problem for some people (often times developers) is that they are lacking theory of mind. Like they literally can't comprehend that I don't see into their heads and they need to articulate their question with correct context otherwise I can't help them.
AI is like a litmus test for it. People who have theory of mind, are capable of putting together a question which will give them good results out of AI. On the other hand people who are struggling with the fact that AI can't see what you mean unless it is in a context window will have bad time with it. These people also usually suck in managing other people because - once again - they are unable to provide tasks with enough context and properly set boundaries. At best they will give you some vague poorly defined tasks and get mad when you will do it differently than they had in their mind.
I feel the opposite, but then again, I do not use AI to actually write the code for me. It's like the faster StackOverflow search.
I'm experiencing the opposite.
The days of typing code manually into an IDE are over. Cope harder, and learn the tools.