'AI fatigue is real and nobody talks about it': A software engineer warns there's a mental cost to AI productivity gains

Posted by CackleRooster@reddit | programming | View on Reddit | 153 comments

"I shipped more code last quarter than any quarter in my career," he wrote. "I also felt more drained than any quarter in my career."

Reply to Post

Reply

153 Comments

[-]

andrewsutton@reddit

Oh, this was obvious to me as soon as I started using AI tools for even small tasks. Reviewing code is much harder than writing code. Reviewing code all day is exhausting. If you've ever been a grader/teacher/professor... grading is roughly equivalent to code review -- if you're doing it right. One place AI coding agents really shine when the code it generates can be thrown away.

Reply

[-]

LoreBadTime@reddit

Yes like "I have some points can you make a script to visualize them"? And then try to reload the prompt until it works.

Reply

[-]

PotaToss@reddit

They turned coding into a fucking gacha game.

Reply

[-]

chu@reddit

Always was

Reply

[-]

UnexpectedAnanas@reddit

Um....no it wasn't.

Reply

[-]

chu@reddit

Downvotes are hilarious. I'd wager I have multiple decades more on this than the average redditor who fancies themselves an 'engineer'. Coding even today is a craft - that is cut and see iteration. Absolutely nothing wrong with that of course but to mistake something that is akin to riding a bicycle for a set and forget self flying drone is a category error. The automations we build do apply engineering principles by definition (the bits that don't have bugs anyway) - but the process of building those automations (the actual software dev) is the opposite. If this weren't the case, development would be fully automated already. And every generation tries to do exactly that (we may be way closer with LLMs though a long way off and may turn out intractable like all previous efforts).

Reply

[-]

QueenVogonBee@reddit

I mean it’s simple: get the AI to review the code… 🤣

Reply

[-]

n_lens@reddit

It's crap at reviewing code.

Reply

[-]

QueenVogonBee@reddit

Yeah I know. That was the joke

Reply

[-]

FetusExplosion@reddit

Depending on the model and context it can be pretty good. GPT codex reviews on github usually find a thing or two inoverlooked. Opus 4.6 does a good job reviewing logic for an LLM. They're useful tools in the toolbox, I wouldn't say those are crap.

Reply

[-]

babige@reddit

Yeah I had opus 4.6 review a sonnet 4.5 basic JS API service + API end points it found and fixed 8 issues, but IT ALSO CREATED 6 ISSUES BY OMITTING WORKING CODE IN THE NEW FILES AND MUTATING OUTPUTS, AND FUCKING CHANGING FUNCTION NAMES! ABSOLUTE GARBAGE.

Reply

[-]

Sparaucchio@reddit

That's tomorrow's AI's problem

Reply

[-]

jamesinc@reddit

I like reviewing code, and AI has made it harder because the problems have become more random. When I read code it is usually clear to me who wrote it from its style. Different engineers approach problems differently and if you work in a stable team those differences become signatures that you can use to direct your attention as you know what the strengths and weaknesses of each engineer are. But now when I read code it is like suddenly the voice changes and a robot is blurting out different code that might be great or might have the most profoundly fucking stupid problem with it, the kind of problem that the engineer themselves would never make. So trust goes to zero and I have to scrutinise every detail of every line for stupid problems. It gets worse with scale too, if you vibe code more, you don't just potentially introduce more stupid problems, you introduce more very subtle problems. All this is to say that, to my eyes, as someone who spends a significant amount of time reviewing code, and as someone who enjoys that work, AI has made it shitter, it has made life harder. The title of this post says "mental cost to AI productivity gains" and all I can ask is "what gains are those?"

Reply

[-]

annodomini@reddit

It's also shittier because normally when you review code, you are helping impart knowledge to someone else who will learn and improve. When you review LLM generated code, no one cares. The model will make the same mistakes again next week. Or not, because it is somewhat random when it makes particular mistakes. You can never really tell. Of course, if you ever do actually give training data back to the AI companies, they'll just use that to improve their models which they'll continue to rent back to you; you're giving them data so that you can keep paying them more in the future.

Reply

[-]

startwithaplan@reddit

Today I'm 99% sure I was iterating on code review either with someone that was totally checked out, having an aneurysm, or just handing my feedback to a model. I pointed out a flaw, an easy enough mistake, but it was more of a statistically likely approach than a "look at similar stuff in the code ase and mimic it" approach. Ok fixed, but if they applied a little bit of context they would've seen it was now semantically wrong and what should've been a minor tweak was now way off. Fixed again, nope obviously tracing the code path more than two or three steps should throw an exception with real data. I almost hope they were having an off day instead of vibe fixing it, because as you say they learned nothing and it took an extra two round trips *so far*. Wasting my actual attention and I've got other shit to do. It was like reviewing some intern code and this guy has been on the team for years. Fuck that noise with something hard and sandpapery.

Reply

[-]

zbignew@reddit

The whole system needs to change. Like, your PR interface should let you right click and vibe out a failing test case in 2 seconds.

Reply

[-]

annodomini@reddit

Yep. You see this with real people too; sometimes it's because they're just checked out, sometimes it's just an off day, sometimes it's because they've given up and are just passing everything to an LLM. Frustrating as hell.

Reply

[-]

schplat@reddit

> When you review LLM generated code, no one cares. The model will make the same mistakes again next week. Not always. Some will actually learn from mistakes, if the reviewer bothers to point it out. Sometimes, as an exercise, if I find a mistake, I'll just tell the prompt that I found one, ask it to review the code it wrote, and see if it can find the same mistake. It's been about 50/50 on finding it. If it repeats a mistake, tell the prompt this is the second time it's made a mistake, and that it should remember this mistake and avoid making it again. Any decent model will remember it, at least in relation to conversations with you.

Reply

[-]

MinimumPrior3121@reddit

The context windows is at max 512kb, really not much when you have a big projects with big and complex classes.

Reply

[-]

gc3@reddit

The user can just add better prompts

Reply

[-]

snowrazer_@reddit

Disagree, I've written so much code in my lifetime. I would much rather have it written for me and review it than write it manually.

Reply

[-]

mycall@reddit

So when you write code manually, you didn't review code or is it just that there is more code to review? Has anyone tried using multiple SOTA models to review each other's code? I would think this would converge to a better solution which would have less issues to resolve in the review process. Maybe it diverges, not sure.

Reply

[-]

KevinCarbonara@reddit

> Reviewing code is much harder than writing code. One of the dirty secrets of the industry is that code is easier to write than read. Management always treats it as if it's the other way around, as if we can code all day and then run through some code reviews in 5m before we leave for the day. But if you don't have a good, formalized code review process, this is going to be a problem. And an even bigger problem once AI gets involved.

Reply

[-]

am0x@reddit

It’s also just plain boring using AI to do your work. The fun part of development is problem Solving. That’s now mostly gone and it’s all architecture, prompt prepping, system configurations, and reviews.

Reply

[-]

SnugglyCoderGuy@reddit

And it is exponentially exhausting the worse whatever it is you are reviewing. There is a *strong* exponentially inverse relationship between time and effort reading and time and effort writing. This is why I *hammer* writing code so that it takes the absolute minimum mental effort to read. Code is read tens if not hundreds of more times than it is written and that costs blood, sweat, tears, and money. However, that cost is spread out over time so it goes unnoticed.

Reply

[-]

worldofzero@reddit

Writing code teaches you to review code. Both sides of that process are symbiotically linked.

Reply

[-]

gc3@reddit

Debugging code teaches you how to write code. We still got that

Reply

[-]

gc3@reddit

Start reviewing code with AI /s

Reply

[-]

Recol@reddit

Don't worry, professors are using AI to grade now.

Reply

[-]

Fenix42@reddit

Everyone is being turned into SDETs. Welcome to my world.

Reply

[-]

fj2010@reddit

SDET?

Reply

[-]

DDB-@reddit

Software Development Engineer in Test

Reply

[-]

polaroid_kidd@reddit

There are no productivity gains. For god's sake, just because you ship more code doesn't mean you ship more productive code.

Reply

[-]

levodelellis@reddit

I don't remember if it was in 2023, but I saw a comment on this site about how AI is completely stupid because static analyzers are good at checking our mistakes, but people are bad at checking AI mistakes. Yet there's still so many articles about AI being 'productive' even though we knew it was a terrible idea pretty much immediately Just use it as a search engine replacement and quit pretending it can write your code.

Reply

[-]

herabec@reddit

Google 10 years ago was still 10x better than AI search.

Reply

[-]

Legs914@reddit

Some of it is that the internet has gotten worse too. Stackoverflow has very few new contributions and so much information is private now. The number of distinct active websites and blogs has plummeted as well.

Reply

[-]

captain_zavec@reddit

> so much information is private now I'm really hoping the mass exodus from discord will reverse this a bit, but I don't know how likely that is. So many good forum communities moved on there and became inaccessible to search.

Reply

[-]

Legs914@reddit

Before that it was slack, so I'm not sure how much it matters. Internet forums weren't the best places to post, but they're amazing for readers years later.

Reply

[-]

aztecraingod@reddit

I would pay real money to have access to Google's search algorithm from 10 years ago today

Reply

[-]

Luckey_711@reddit

Not to say but I started paying for Kagi a few months ago and I could not love their service enough

Reply

[-]

aztecraingod@reddit

Interesting, color me intrigued

Reply

[-]

privacyplsreddit@reddit

Paid for it and really couldnt find much of a difference between it and modern google. There's just less good content out there as everything has turned to click farms to game googles algo to make money and real discussion has died and reddit is the main source of "publically indexable" online conversation even though its filled with astroturfing corpos and llm bots now. Im happy to shown some real examples and id resub to kagi, but after using it for a few months the top 20 results for any query werent that different from what i remember. Internet is just different now.

Reply

[-]

PedanticSatiation@reddit

I think part of the problem is SEO becoming more and more sophisticated. Even the same algorithm wouldn't be as good as it was 10 years ago.

Reply

[-]

timsredditusername@reddit

Let's just go back to AOL keywords

Reply

[-]

G_Morgan@reddit

Android assistant from 5 years ago was still better than Gemini today. I still regret every time I ask Gemini to set an alarm these days. Not only do I end up needing to do it myself, I also have to deal whatever bullshit Gemini came up with.

Reply

[-]

FlyingBishop@reddit

AI isn't a substitute for static analyzers, but AI is way better than static analyzers at telling you what is wrong with a given piece of code. In fact it makes static analyzers a lot more useful because the AI can explain what you did that caused the analyzer to get mad.

Reply

[-]

neithere@reddit

It's good as an extra fuzzy grep and a rubber duck.

Reply

[-]

throwaway490215@reddit

Anybody still denying its *potential* this far in needs to just pick up blacksmithing as a career. Having proof that others have machine-gunned themselves in the foot is not the same as it not being a productivity gain.

Reply

[-]

thatsnot_kawaii_bro@reddit

Everything has potential. Even during the dotcom bubble things that crashed had potential.

Reply

[-]

throwaway490215@reddit

This is such a dumb argument people keep making. Yeah, be proud that you see yourself as on top of the bell curve. At the height of the dot-com bubble really was the time to step away from the internet because it clearly only produced overvalued companies. If you think AI is still in the hypothetical like NFT's then - again - just go pick up blacksmithing.

Reply

[-]

Wafflesorbust@reddit

The potential of flying cars is undeniable but we're never going to have them.

Reply

[-]

G_Morgan@reddit

I fully believe we'll have viable AI coding assistants one day. I firmly believe they will not be LLMs.

Reply

[-]

vytah@reddit

Seeing how people drive their non-flying cars, the only potential of flying cars I see is deadly crashes. Also, we already have flying cars. They're called helicopters.

Reply

[-]

FlyingBishop@reddit

Helicopters exist. And unlike Helicopters there's basically zero risk of accidentally killing someone. Obviously you can write bad code that kills someone, but AI is very low stakes and it really does let you understand things much faster.

Reply

[-]

justinlindh@reddit

Reddit isn't the place for viewpoints like this. I've also pointed this out (a bit more gently) on here a few times and get drowned in downvotes. I empathize with where people have problems with AI in all aspects (I wish it didn't exist) but we're at the point where if you're using the right tools and models effectively and thoughtfully and you're not seeing any positive net value that it's user error. It's here to stay whether we like it or not in software engineering and it's time to learn how to use it effectively or be left behind. Shoot the messenger all you want on this, but it's the new reality.

Reply

[-]

polaroid_kidd@reddit

I'm curious in which parts of the engineering cycle you're seeing these gains. I do notice them in research, debugging obscure errors and stack traces, but very rarely in actual implementation work in medium to large code bases. FWIW, we're on opus 4.6 and have extensive skills with best practices defined. It's just not helping the way I hoped. Trust me when I say, I look forward to the day when I don't have to debug some good damn edge case in input validation combined with an unmaintained form library. We're just not seeing these gains in real work. Prototypes, one off scripts, familiarizing yourself in a new code base with a new stack just gotten easier since the patterns are mostly the same. But feature development more often then not leads into a accidental engineering cycle where at the end of the day I think "jeez .. would have been faster and cleaner if I had just done it." My experience so far is that it either one shots it, or 80%, or it won't do it. What have you experience been? Do they align with mine or have you managed to get it to implement from ticket to production hands-off flows or somewhere in-between for the majority of work?

Reply

[-]

FlyingBishop@reddit

> I do notice them in research, debugging obscure errors and stack traces I mean, exactly this? The speed with which AI can diagnose stupid typos is incredible. Literally saves hours.

Reply

[-]

gc3@reddit

Small one off tools, or even important ones you will use again that are well defined, took 10 minutes rather than a few hours. Front end code that you fix by adding features and then iterating in a conversation. I had mixed luck in the code quality but it was understandable and it felt slick to the end user, and Javascript (ugh) is usually a lot less understandable than what Cursor wrote. I have seen terrible Javascript. And: I am soo lazy. Fixing syntax errors on compiles! Before I even attempt to fix it myself I copy and paste the error into Cursor and ask it to fix it. It usually explains what the error was as it fixes it so if it looks wierd I will look at it closer. But I have often given it sample code from another project of mine I want it to be similar to, this is the best case but it really applies only to pipeline jobs that do calculations and process data. Be prepared to tell Cursor to put functions it wrote twice into a common library after it writes the code for you. Except for the font end, I didn't vibe code anything huge or monolithic.

Reply

[-]

justinlindh@reddit

From my experience: It's absolutely not about one shotting anything. Opus 4.6 leveraging the common "superpowers" skill set to brainstorm/plan/implement is competent at brownfield tasks in large repos (even mono-repos). It also has a "systematic-debug" skill that's excellent at properly getting to root cause, reproducing the problem by running it, solving, and verifying. The key tool here is still YOU! You need to understand how to provide the right amount of specificity to the robot and steer it. Doing that alongside proper spec driven development can produce very, very good results. At the end of the day, YOU still need to run/debug/test/review. But integrating it into your cycle properly and monitoring what it's already done can help minimize that time. It's still the developer who must own the outcome and our skills aren't obsolete. It's not all or nothing in this world. But these tools, the way I'm using them (as a very experienced engineer in corporate world) absolutely provide a net benefit once it "clicks" how they can fit into your workflow. And you need to constantly evolve and adapt. Claude Agent Teams dropped with Opus 4.6; experiment with it and figure out when it could be useful (hint: it's not always). New models and tools arrive at a dizzying pace but the bleeding edge of the tools combined with your core problem solving skills and technical expertise is how you survive.

Reply

[-]

MoreRespectForQA@reddit

>if you're using the right tools and models effectively and thoughtfully and you're not seeing any positive net value that it's user error. Ive noticed that the people who see the most perceived productivity gains tend to be the weakest engineers.

Reply

[-]

justinlindh@reddit

This is what I mean when I say healthy reddit isn't the right place for discussions about this topic. You just implied that I'm a weak engineer. I can assure you that I am not. I'm also just not interested in trying to convince you otherwise here, so you win. I shouldn't have commented at all.

Reply

[-]

PotaToss@reddit

It's just math. They're built to probabilistically produce like median code. If you put in the work to be better than average, the flaws are obvious, but if you're just having it poop out code in some domain you don't know well, you can't tell that it sucks. It just feels like the biggest boosters are people who don't understand Gell Mann Amnesia.

Reply

[-]

ciemnymetal@reddit

I've noticed this trend as well. It's either that or ESL folks who tout the most productivity gains (which tbf, using a LLM to overcome language barriers is a legitimate use case). The issue is the more experienced devs are forced to stop what they are doing right and hamper their productivity for the sake of using AI justifying an exec's investment in a Open AI license.

Reply

[-]

48K@reddit

Reality changed at the end of last year and the people here are playing catch up. I used to be one of them so I’m sympathetic.

Reply

[-]

nesh34@reddit

I mean there are productivity gains. I don't think it nets out to the kinds of gains that VPs are touting, but there's obvious gains and high potential for gains.

Reply

[-]

valarauca14@reddit

Productivity is measured by management in terms of features shipped. Doesn't matter if those features are buggy messes (see: Microslop Windows) they shipped in less time than before. [Nothing is more important than short term quarterly gains](https://www.youtube.com/watch?v=X4Mj1N9rfFE)

Reply

[-]

VideoRare6399@reddit

I feel similarly. Still very curious as to see how tech debt associated with AI written code (even if perfect) will play out. If you can’t entirely remove the human, then the human needs to understand / develop a mental model of the system / code, which is a cost that must be paid. Before AI, the cost was paid as you were writing it / figuring it out. Still we are only human and lots of bugs are/were caused by incorrect mental models leading to changes which were made without a complete understanding. Now with AI, all is wonderful while in vibe world. But if the programmer ever needs to step in and figure out what the fuss is all about to fix a bug then they need to start from scratch unless the bug is simple enough no context is needed / the context is limited to the function or something simple. The problem is: the kind of issues AI would be able to fix are exactly ones with small context / tribal knowledge / whatever you want to call it.

Reply

[-]

ZenBourbon@reddit

I just treat AI like an offshore contract dev/team. Assume no consistency. Never abandon having a mental model. Review all code. Ask for plans before big implementation, review the plans. Confidence means nothing. But also, be flexible. Sometimes a well-isolated hack is fine. Sometimes you have a deep product/architectural pivot vision that's just going to take too long to communicate across the language barrier - plan it out, break it down, and DIY or delegate the safe atomic changes. It's a lot easier to navigate LLM's capabilities once I start anthropomorphize it as one of the contractors I used to work with.

Reply

[-]

HotDogOfNotreDame@reddit

So much this. But, it’s also made it much harder to utilize the offshore engineers I’m told to utilize. They’re still not creative. They still are blind to architecture. And if I spend the time to figure out exactly what they need to do (which they fail if I don’t), well now I can just put those instructions in copilot and get results back in minutes instead of 18 hours. (Also, even with the 18 hours if I give them the task, they’re probably still just pasting instructions into copilot themselves. So why do I need them?)

Reply

[-]

ZenBourbon@reddit

If you're an IC dealing with bad contractors, the AI craze is the perfect time to raise concerns about that.

Reply

[-]

HotDogOfNotreDame@reddit

No, I'm an architect and dev lead (depending on the project), and I lead both onshore and offshore engineers for my projects. The offshore engineers are employed by my company. It's just becoming increasingly futile to use them, when copilot does their job faster, and with less input from me. My current stopgap is to assign them manual labor tasks: "Test our framework against this, and against that, and for these 15 other scenarios, and document everything you see." The kind of time-demanding manual labor that I don't yet know how to get an Agent to do.

Reply

[-]

thy_bucket_for_thee@reddit

Being this diligent though isn't a good thing, because you're putting all your effort in preventative measures. When you're always reacting, you'll never catch up; you'll never have a chance to do something right when you're correcting the regressions toward the mean, rather than pushing the mean forward.

Reply

[-]

ZenBourbon@reddit

Just do some time management? I've been writing software for 15 years now. Regardless of my collaborators, I always reserve time for hands-on deep work.

Reply

[-]

thy_bucket_for_thee@reddit

I'm not talking about deep work but big LOL in being a wage slave if you think it's okay to do additional work on top of keeping the dumpster file contained. Sorry but devs aren't trained to be firefighters.

Reply

[-]

Tzukkeli@reddit

Legacy code is code written by someone not in the house anymore. In this case, AI written code is basically instant legacy and we all know how much effort mentally and workhourly it takes to accomodate that.

Reply

[-]

Sevla7@reddit

> Legacy code is code written by someone not in the house anymore. In this case, AI written code is basically instant legacy and we all know how much effort mentally and workhourly it takes to accomodate that. When you put it like that it really explains a lot

Reply

[-]

gc3@reddit

At least you can ask the AI to explain what it was thinking

Reply

[-]

UnexpectedAnanas@reddit

The AI doesn't think. You're just asking it to string more words together, but it doesn't actually relate to the previous task.

Reply

[-]

gc3@reddit

I'm not sure people do. But if you've ever used Cursor or Claude code it appears to think, and it's analysis can be correct. I think you are getting AI from 2023 confused with AI from 2026.

Reply

[-]

synovanon@reddit

And when it’s wrong in it answer, you get sent down a rabbit hole because you got gaslit by the AI hallucinations

Reply

[-]

unflores@reddit

Hah. Legacy code is what I merged to main yesterday. What was that guy thinking?

Reply

[-]

Sparaucchio@reddit

And with AI, legacy code is what you merged right now

Reply

[-]

chu@reddit

The problem with legacy code is that is not disposable.

Reply

[-]

hauthorn@reddit

I'm going to disagree hard with the definition of legacy code. Legacy code is something you decided you will no longer maintain and change. It could be because you are afraid to change it without breaking something. Code with good tests and documentation will not become legacy easily. Doesn't matter who wrote it.

Reply

[-]

BroHeart@reddit

they hated him bc he told the truth, lol I agree, code that is well defined, well tested and performs as expected in end to end testing, is still quality code. If you've worked in a team of 20+ people, or larger, everything is legacy code by this broader definition. Humans also have very limited working memory, give it a long weekend, a few more tickets in the interim, or a bad morning on the drive in to work and you'll find that person has to pick the pieces back up. [https://i.redd.it/o1gnotfxi4k01.png](https://i.redd.it/o1gnotfxi4k01.png)

Reply

[-]

PoisnFang@reddit

Based, idk why people be hating on your accurate comment. But be people who don't work a normal corporate job

Reply

[-]

darkde@reddit

I took it as some context is lost because ai doesn’t have real memory. RAG/context is not anywhere close to human memory. So in some ways, a human looking at ai code doesn’t have all the context because it wasn’t their keystrokes.

Reply

[-]

bryaneightyone@reddit

Not sure why they're downvoting you. This is correct in real life.

Reply

[-]

darkde@reddit

Mmm that’s a great point. You can only add so many ai generated docs for context. That’s why I believe engineers will always be needed in some capacity. You need a brain to help breakdown work when working with large epics

Reply

[-]

gc3@reddit

I see you have never tracked down bugs in legacy codebases written by humans

Reply

[-]

VideoRare6399@reddit

Oh but I beg to differ good sir. May I ask what makes you say that?

Reply

[-]

gc3@reddit

Why does reddit not have 'view parent comment' anymore? Something you said about tracking down bugs in AI generated code. I think legacy codebases written by multiple humans who overcomplicate things can be worse

Reply

[-]

WatchOutIGotYou@reddit

Code Janitor is going to be a lucrative industry in a really short amount of time.

Reply

[-]

huyvanbin@reddit

Back in the day 5 years ago, the value of a software company was the value of the team. Everyone understood that getting to an MVP is not that difficult once you know what you want to build, and everyone else knows what you were going to build once you ship, so they can replicate it. The question is, does someone who didn’t come up with your product in the first place have the insights to continue evolving it in a productive way? Can they use customer feedback, market pressure etc. to evolve the concept into something that will get real adoption? And part of that of course is not just product level insights but technical insights: what were the compromises involved in why the code was written this way? What needs to be reassessed if you run into a performance problem or surprising use case? In the old days you would ask the people who coded it. If your entire software team is just a history of LLM prompts, then what is the value of your company? You won’t be able to explain the philosophy behind your decisons because you didn’t make them. And maybe three subsequent Claude sessions had three different philosophies. In any case, the difference between you asking Claude to write the next feature and someone else doing it is precisely zero. Of course at this point the entire tech economy is a shell game so maybe I’m wasting time thinking it needs to make sense…

Reply

[-]

Nyadnar17@reddit

I am unclear the benefits to Vibe codding over AI-assisted codding. I thought I was going insane attempting to make Vibe coding work. Meanwhile AI-assisted codding has removed basically all of the parts of the job I hate(trolling forums, deciphering obscure error messages, and finding typos). Why even try to make Vibe codding work? Whats the perceived benefit?

Reply

[-]

VideoRare6399@reddit

Vibe coding is beneficial only for people who don’t know how to code. Otherwise AI assisted coding is always the best option. If you break the problems down enough though, have a decently maintained repo with, and have some experience with Claude or whatever models you use then you can reasonably assume the output will be as you’d expect. In this way you can kind of vibe code but with a few qualifications.

Reply

[-]

dysprog@reddit

Based on what I have seen, the burden of maintaining a mental model is still mostly on the programmer. The programmer needs to break down tasks, the programmer needs to understand the goal, the programmer needs to guide and correct the AI, the programmer needs to review the code. All of this requires an understanding of the code. I feel like this is heading for a disaster. I can only hope it will come soon. If we get to the 'deskilling' part of this, there will be no one left who can fix anything.

Reply

[-]

CarefulCoderX@reddit

I use AI a lot to modify my resume for certain roles and it often can't even get that quite right. It'll frequently take two separate things I worked on and combine them when describing the experience I have or it'll just make stuff up based on a job description. Dealing with a system with tons of moving parts and pieces that fit together a certain way is way more complex.

Reply

[-]

Randy_Watson@reddit

\> "Each one 'only takes an hour with AI.' But context-switching between six problems is brutally expensive for the human brain," he wrote. "The AI doesn't get tired between problems. I do." No offense to anyone, if you were a programmer that was actually granted real blocks of time between inane meetings with non-technical managers who think they know how any of this shit works, I get what this guy is saying. I have never had a chance to work in a place like that. I've always been at the mercy of the project manager who's deliverable is meetings made to fill out confluence pages so the higher ups can engage in a mutual masturbation session in meetings I'm not in. At least with AI, I don't have to pretend to care about coding in a way that makes me have to work before anyone is around or late into the night. Sure, it will make me obsolete eventually, but I'm using that time to figure out my exit. And here's the fucked up part, I love where I work. The people are great. However, it's the tech managerial culture that is the problem. If your job is to figure out tough logic problems, most of us can't do that in the 30 minutes between hour long meetings we don't need to be in that are splattered all over our day. I recently had to go up in front of a murder board because I went above and beyond. I did more than I was asked for because I made the assumption this feature would be expanded on so I might as well just do it all now. Only, I missed a piece because different departments used different naming conventions for one fucking property. And now I'm looked at like the asshole and being asked how this possibly could have happened when I implemented more than what was asked of me. Then people were like no, the requirements said this needed to be in there. I check the confluence history, that requirement was updated six fucking months after my feature was merged. AI isn't burning me out. AI is helping me fight against the stupid managerial culture in tech.

Reply

[-]

BroHeart@reddit

Everything about work is changing. I have made roughly 6,400 commits in the last 4 months using Claude, Gemini, ChatGPT, with Claude 4.5 and 4.6 being enormous jumps forward in ability. Gemini does translations for nearly a quarter million strings for me and can complete it in 3 hours for one of our longer running consumer projects to support about 18,000 users in 30 languages. The cost is approximately $0.36, when the last model the cost was almost $90 for one full iteration of the same translation set. I've been able to use it to rewrite our corporate website of \~250 pages, create SLA tracking/support handling tools that also call me using vAPI with ticket details as SLAs approach breach. Used it for writing nearly 800 songs in the last 2 months based on old music I published 11 years ago, of which 75 or so are now bangers after multiple iterations of production, and scheduled out months of releases. Use it for analyzing 35,000 court cases in my city to inform public policy proposals for eviction protections after it became clear property management companies were systematically abusing housing law to quickly physically evict tenants in order to aggressively raise rents, and then turn that into a visualization to tell the story on the policy proposal site. Used it for adding SSO/MFA/audit log/RBAC functionality to one open source project in two days when other folks on project had kicked the can down the road on it for literally years. Definitely stressful, it's crazy to be able to work on 12+ projects and see them all progress with new features and long-standing bugs falling away. The most value for me has come from starting each project with a CICD process that forces branches to go through testing / auditing against design specifications before it gets into the main project.

Reply

[-]

ChrisRR@reddit

I suspect you're getting downvoted because 6400 commits over 4 months means 1 commit every 6 minutes. To me that wreaks of someone who's not actually taking the time to read and understand everything they're committing and instead just endlessly churning out AI code, hitting commit, without really taking a step back to really think through what they're doing

Reply

[-]

BroHeart@reddit

That's only 53 commits per day in total, your math is super off I'm not sure where you're getting every 6 minutes. I typically work in batches through my own testing notes and user feedback and start lots of individual branches, flip to those, review changes, push to CICD, merge into main if no failures. I'd say, similar to the article above, I spend MORE time architecting, thinking, reviewing, across MORE projects, than any other time in my career. 800 tracks in 2 months is CRAZY! Generating that many tracks doesn't mean publishing all those tracks, and reviewing, adjusting prompts, re-recording my vocal references, re-writing lyrics for flow, makes up a pretty good chunk of those commits too.

Reply

[-]

braczkow@reddit

Nice. 😎

Reply

[-]

Firm_Party_1458@reddit

This sub is now full of AI denial, let's face it the programming work becomes worthless and you know it

Reply

[-]

valereck@reddit

Not that I can see

Reply

[-]

MinimumPrior3121@reddit

And this code is pure garbage

Reply

[-]

PressWearsARedDress@reddit

i have found using AI to generate boilerplate has improved my productivity. The key is in that AI is really good at "classification" and learning about your style of programming. So to make the best of it, you have it generate some headers or some testing code with the assumption you will be going in later to fill in the blanks or adjust placeholder code. I also find the in line code generator helps me think about what I am doing and even exposes bugs as I am writting them.

Reply

[-]

Legs914@reddit

It seems to work best for me in small boilerplate ways like this. Writing some unit tests to cover a specific edge case. Adding some internal api routes to match functions I've already written. Basically stuff that is very rote and easy to verify at a glance.

Reply

[-]

mycall@reddit

> His typical day is no longer devoted to deep focus on a single design problem, Khare wrote, but rather to context switching between as many as six different problems. Welcome to management, being pulled in all kinds of directions as stakeholder demands your time. You are managing a team of AI agents, so become the manager you are supposed to become.

Reply

[-]

pakfur@reddit

This has not been my experience. I have been writing software for over thirty years. I am so happy to not have to crank out yet-another-controller-service-repository, blah blah blah. I have added agents to my workflow, not only am I more productive, the quality of my output has improved and I feel like writing software is fun again. I _enjoy_ speccing out my workflow, ang guiding the AI through it. I reserve hard and interesting stuff for me to do, and let AI crank out the boilerplate. I understand not everyones experience is the same as mine, this isn't everyones cup of tea, but this feels as big, if not bigger than OOP, or Cloud or PaaS and SaaS combined.

Reply

[-]

Expert_Scale_5225@reddit

AI fatigue is the productivity tax nobody warned us about. The article nails it but misses the deeper pattern: we're not tired of AI - we're exhausted from **context switching between human and AI reasoning modes**. Three layers: **1. Prompt engineering is invisible work** "Just use ChatGPT" ignores that crafting effective prompts burns cognitive cycles. You need to: - Know what the model can/can't do (constantly shifting) - Frame problems in AI-friendly ways (not how humans think) - Debug hallucinations (new failure mode) - Validate outputs (trust but verify on steroids) It's like having a brilliant junior dev who speaks a different language. Fast output, but you're the translator. **2. Decision fatigue compounds** Pre-AI: Write code OR review code Post-AI: Generate code → Evaluate quality → Edit/rewrite → Re-evaluate → Integrate You make 3x more decisions per feature now. Each "should I regenerate this?" is a micro-decision draining willpower. **3. Loss of flow state** Deep work requires uninterrupted focus. AI tools fracture this: - Copilot suggestions interrupt thought - ChatGPT requires task switching - Reviewing AI code = less creating, more judging You're productive in output, exhausted in experience. **The trap:** Management sees 2x output, doesn't see 3x cognitive load. They double your workload because "AI makes you faster." You burn out while hitting metrics. Fix: Track cognitive load, not just velocity. Treat AI assistance as adding team members (with coordination overhead), not magical speed boost.

Reply

[-]

haltingpoint@reddit

AI wrote your comment. It is extremely obvious.

Reply

[-]

MirrorLake@reddit

I've said it before, and I'll keep saying it: where are all the gains? Where are the *intelligent* products? Adding chatbot buttons to every UI on the planet doesn't count as innovation.

Reply

[-]

maxineasher@reddit

I'm still shocked that too many of the comments on yesterday's AI generated "I hate what AI is doing to software engineering" post didn't spot it as AI. Were those AI bots too just commenting for commenting sake? The Internet is 100% AI this point, why not let the code be as well. It's bots all the way down.

Reply

[-]

DerixSpaceHero@reddit

Most technical subreddits are completely compromised by pro-AI bots fighting against anti-AI bots. Foundationally, I blame the deprecation of dedicated niche forums to the centralization towards highly censored, easily manipulated platforms like Reddit and Twitter. I truly cannot wait for the day Reddit ceases to exist.

Reply

[-]

maxineasher@reddit

It doesn't exist right now. Moltbook and reddit read nearly identical to me.

Reply

[-]

FlatProtrusion@reddit

Isn't moltbook mostly humans pretending to be bots?

Reply

[-]

vytah@reddit

> I'm still shocked that too many of the comments on yesterday's AI generated "I hate what AI is doing to software engineering" post didn't spot it as AI. People don't read articles, they comment based on the Reddit post title alone. Are you new to Reddit?

Reply

[-]

Loves_Poetry@reddit

Just because it's AI, doesn't mean it's not worth a discussion

Reply

[-]

maxineasher@reddit

Yes it does. If someone wasn't willing to take the time to write it, I'm not going to take the time to read it.

Reply

[-]

Perfect-Campaign9551@reddit

Guy is blaming the AI for lack of self control

Reply

[-]

Dull_Cucumber_3908@reddit

I tried AI a little and I realized that it is like you have a junior dev who knows a lot of stuff but really messes up when trying to implement something. When I pinpoint an error to the AI and ask it to fix it, it comes with some other random error and after a while it just start hallucinating. I decided that it's better to have a human junior dev instead of an AI.

Reply

[-]

slickwombat@reddit

Cory Doctorow (probably best known as the guy who coined the term *enshittification*) uses the terms *centaur* and *reverse centaur* here from automation theory, and I think it sums it up pretty well. A centaur is a case where a machine exists to help a person, as where the centaur's horse half helps them move faster. A reverse centaur is a case where the person exists to help the machine -- so, a horse head on human legs, as it were. Coders relegated to review roles for AI PRs are reverse centaurs. AI isn't helping them do stuff, they are just serving as reviewers to help the AI do stuff -- or, more pointedly, to prevent it from doing bad stuff, and take the blame when it does. It's easy to see why that ends in unhappiness and burnout, nobody wants to be the legs. I'm highly skeptical about the premise that this is really leading to increased productivity, too, at least in the sense of increased value to a business. Just for starters, we all know that AI can easily produce some monstrously bad code that looks pretty plausible, and I highly doubt unhappy, burnt-out engineers are catching it consistently, so this might well be a case of speed at the expense of quality -- whether that manifests down the road as maintainability problems or more immediately in functional or security issues.

Reply

[-]

TheMistbornIdentity@reddit

> I'm highly skeptical about the premise that this is really leading to increased productivity, too I strongly agree. Having also studied psychology, I have a feeling that reading and writing code are two different skills, likely processed by two (or more) different brain regions. So all the people saying "Oh well working with AI just means the job is more about *reviewing* code now" are likely letting the part of their brain that controls *writing* to atrophy. The day that they actually need to write code themselves (and got forbid they need to use package(s) that the AI wasn't trained on), they might find themselves fucked. And nevermind the fact that people are notoriously bad at self-assessment. Most people likely think themselves great at multitasking, but studies have found that to be demonstrably false. Likewise, I vaguely recall learning that something like 70% (frankly, the percentage is probably wrong) of people think they're above average (I also forget if it was in intelligence or something else. Point is, I don't trust for a second that people are *actually* more productive with AI, even if it seems that way on first glance when they're generating a whole new project, versus having to actually maintain the thing.

Reply

[-]

Haplo12345@reddit

Lots of people have been talking about it. Ever since ChatGPT's first version was released, people have been fatigued by dealing with its shit.

Reply

[-]

MrThingMan@reddit

He just use more AI. Embrace and extend. More productivity.

Reply

[-]

paranoidray@reddit

[AI Doesn’t Reduce Work—It Intensifies It](https://simonwillison.net/2026/Feb/9/ai-intensifies-work/)

Reply

[-]

Educational-Cry-1707@reddit

From where I’m sitting, AI is just the next layer of abstraction in software engineering. We don’t write machine code or assembly, we write in high-level languages. Now we can also write code by prompting. It’s not a revolution, it’s an evolution. It’s more efficient, sure, but not any different from any other productivity tools out there. It’s certainly not autonomous, and no better for non-developers than “no code” or anything else. It’s a full time job, and if a non-developer takes it on they just have to learn to become a developer. Whether by using AI or not. Does this mean there’ll not be a need for as many developers? Maybe. Or we’ll just build more software. After all, everything needs software. And writing code is a very small part of building software.

Reply

[-]

Additional_Rub_7355@reddit

It is not a new layer of abstraction because the end result is not deterministic.

Reply

[-]

chu@reddit

LLMs are designed to be deterministic at temp 0. You will see small variations based on batching and CPU.

Reply

[-]

Educational-Cry-1707@reddit

The end result is code. Which is deterministic. You’re just not typing the code.

Reply

[-]

ChrisRR@reddit

One of these days I'll open this sub and the top post won't be about AI

Reply

[-]

HeracliusAugutus@reddit

There were no productivity gains in the first place, so it's nothing but detriment

Reply

[-]

shokuninstudio@reddit

All those "vibe coders" we have seen with their little utility apps haven't shown us a big fully featured and battle tested application. The kind of applications creatives use for example that require extensive bug hunting and beta testing. It's exhausting enough localising a large app in a dozen languages using Google Translate. Proof of how even one little panel can be a headache... [https://www.reddit.com/r/SwiftUI/comments/1r289mj/localization\_blues/](https://www.reddit.com/r/SwiftUI/comments/1r289mj/localization_blues/) I've been working on that app for six years on my own. I never thought 'Let me vibe code this and throw it on the internet quickly.' Every feature has to be tested for weeks and sometimes months. Then the localizations have to be carefully studied, amended, adjusted and updated - sometimes weekly, sometimes even daily.

Reply

[-]

maxineasher@reddit

> we have seen with their little utility apps haven't shown us a big fully featured and battle tested application. Meh. I've done so much with AI's it's just stupid. I ported, via an AI, Starflight, an old 1986 DOS game written in FORTH. FORTH decompiles nicely (as good as a bad language like FORTH can) and the AI reads it fine. Then I turned it into an Unreal project and made a 3D version of it, using SDF rendering and what not. All using the AI. Brought over the fractal generator and made high res landscapes. I've checked some of it in to github publicly, but no one really cares about Starflight given it's a 40 year old game. So I've kept it to myself. I have a bunch of AI projects like this.

Reply

[-]

Pelly_@reddit

"I've done many impressive things with AI but I'm not gonna share any of it, because nobody cares" LMFAO

Reply

[-]

tanaciousp@reddit

Share your GitHub projects then.

Reply

[-]

pier4r@reddit

but then you agree with the person above. The "vibe code" projects, are projects that are relatively compact (and do not need to be adjusted continously). In that LLMs can help a lot. If the project is a bit bigger or needs constant maintenance - for example go substitute curl or openssl - it is much different. (that is not vibe coding anymore, rather agentic engineering and it needs the approach "trust but verify", that is what cost braincycles)

Reply

[-]

Equux@reddit

"and nobody's talking about it" I feel like I hear more anti ai sentiment than AI sentiment. Like yeah,there's some guys who are all in but I don't think that's most people

Reply

[-]

can_ichange_it_later@reddit

...Are there... AI productivity gains...?

Reply

[-]

smoke-bubble@reddit

This must be some anti AI joke XD I've never felt more rested than before. Not having to google for hours or deal with stackoverflow's toxic community is so liberating. You do your job in half the time and can spend the rest doing some other interesting stuff.

Reply

[-]

ProfessorPhi@reddit

Curious to how managers etc are finding the tool? It feels like a lot of eng management and tech leading which were always super draining haha

Reply

[-]

Milumet@reddit

The whining never stops, doesn't it?

Reply

[-]

MEXAHu3M@reddit

I'm making my pet project to better understand some concepts and I tried to use AI tools as much as possible to make it easy and fast. But I started getting bugs pretty quickly and I realized that yeah, I'd seen this code and I kinda understood it at the moment, but now I don't have the whole picture how my system works. So I just didn't how to fix it. AI couldn't do that either. So I just started gradually rewriting it (because sometimes it was such a mess) and fixing it and now I understand what I really need to do next. And guess I would spend less time if I would do it manually from the begging. So yeah, I stopped to use it as much. Now it's just more about concepts but not for writing code

Reply

[-]

MeBadNeedMoneyNow@reddit

how about companies hire people

Reply

[-]

dinosaursrarr@reddit

Fredrick Engels would like to talk to you about alienation from your work

Reply

[-]

_cob_@reddit

Execs won’t care. They want to suck every morsel from you until they can outright replace you.

Reply

[-]

Shiminsky@reddit

Might still be a net gain at the end of the day, but assuming coding without AI is 1x speed and our ability to comprehend code is at 5x, a 15x faster coding agent means we are still 3 times slower at coder review for code the AI is able to pump it out. This way naturally leads burnout as we feel the ever present need to go faster in order to catch up with the LLMs.

Reply

[-]

huyvanbin@reddit

Back in the day 5 years ago, the value of a software company was the value of the team. Everyone understood that getting to an MVP is not that difficult once you know what you want to build, and everyone else knows what you were going to build once you ship, so they can replicate it. The question is, does someone who didn’t come up with your product in the first place have the insights to continue evolving it in a productive way? Can they use customer feedback, market pressure etc. to evolve the concept into something that will get real adoption? And part of that of course is not just product level insights but technical insights: what were the compromises involved in why the code was written this way? What needs to be reassessed if you run into a performance problem or surprising use case? In the old days you would ask the people who coded it. If your entire software team is just a history of LLM prompts, then what is the value of your company? You won’t be able to explain the philosophy behind your decisons because you didn’t make them. And maybe three subsequent Claude sessions had three different philosophies. In any case, the difference between you asking Claude to write the next feature and someone else doing it is precisely zero. Of course at this point the entire tech economy is a shell game so maybe I’m wasting time thinking it needs to make sense…

Reply

[-]

lambertb@reddit

Did about 11 hours yesterday. Insane levels of productivity. But totally and completely cooked by bedtime.

Reply

[-]

LegendEater@reddit

Even though you did basically nothing? I bet you weren't reviewing.

Reply

[-]

ten0re@reddit

I wonder if that’s because his entire way of doing work has changed and he has to learn entirely new tricks and gotchas. I had a similar experience, least surprising discovery of this year.

Reply

[-]

dragenn@reddit

10x or bust...

Reply