AI impacts the quality of my work severely.
Posted by HotJellyfish8247@reddit | ExperiencedDevs | View on Reddit | 58 comments
I was moved to a different team recently. So I have to learn new domain knowledge, new language (moved from C++ to python). My company openly says we should burn tokens, and vibe code, it's only speed that matters. I picked up a new task. I really tried to embrace agentic AI, I had skills written that plan the feature, analyse existing codebase, subagents that guard quality of code, compliance with a spec etc. First problem, this huge system is slow as hell. Secondly, I then have to sift through 3000 lines of code, only to discover, it's all slop, and it shouldn't even be 3000 lines. I feel like I'm also not learning anything. But if I stop burning dollars on cursor, it will negatively impact my performance review. Also, anyone around me does not care, they just push that slop to review, so my 500 lines of good code written in 3 days will look worse on paper than 2000 lines of slop written in one day. Oh, also, I have no structured onboarding because we have ai now, so it is not needed, apparently.
I am just so tired. I like AI as a tool, I am interested in introducing AI systems to my workflow, I'm just tired of this AI psychosis.
ok-computer-x86@reddit
Well we kinda have similar situation and I am already at the point where I no longer care about what I do for my company. My manager also keep pushing me to develop products with agentic AI just for the sake of "fast delivery". I just do the "actual" programming for my hobby projects on my weekends. Most of the companies only care about shipping their end product anyway, and my manager has similar perspective so it feels like it is pointless for me to bring up the “craftsmanship” mindset discussion to the table (which I believe it is still far more important) when the only metric my manager seems to care about is whether the product ships on time.
Watermelonnable@reddit
sad times to be a swe indeed. Endure, cash out and buy a farm
NewFuturist@reddit
I ain't getting farm money
guns_of_summer@reddit
all the farmland is getting bought up for data centers anyway
BlazingThunder30@reddit
All the money is being spent on AI
ymonad@reddit
Yes, and also nobody is reviewing the code. just instantly LGTM.
ultraDross@reddit
Difficult to review a 10k+ PR, or as others in my team say "ask copilot to review it for you". Which begs the question, WTF are we doing?
Illustrious_Pea_3470@reddit
You can’t use a skill to plan the feature, you have to plan it yourself. Otherwise you are just slowing yourself down.
Lucifernistic@reddit
What model are you using? It's really not hard to get high quality code these days. I almost never get slop except from inferior models.
But I also don't write C++, so I'm not sure how it fairs there.
ZunoJ@reddit
Do you have one example of foss software that was written this way and is genuinely well written? Only one, nothing more
Lucifernistic@reddit
Written with AI? Almost nothing is pure AI, but I've seen plenty of Claude/copilot/cursor commits on things.
Off the top of my head, Onyx. They build AI tools and use AI heavily to dev, and their codebase is quite solid.
I have a number of projects my department runs which were heavily assisted by AI, very large in scope. But these are internal / NDA'd, not FOSS.
ZunoJ@reddit
I've checked the last 50 commit of Onyx and there were very little AI assisted commits. Those that did use AI were mostly, documentation, comments and configuration. I also wonder what makes you say their code base is quite solid. I took a quick look at some code and it was pretty shockingly bad. Lots of hard coded strings that are repeated in different parts of the application, string manipulation instead of good abstractions, cyclomatic complexity to hell. All in all a project that looks like it was cobbled together by a bunch of new grads. Either way, this is a very bad example
Lucifernistic@reddit
Not all AI commits show it, most don't, I think, since literally none of mine ever do unless I have an automous cloud agent do the commits.
But regardless, I think their backend is fairly extensible, and easily understandable. I've been able to very easily extend it's functionality and maintain my own fork without issues. Code is fairly reliable, been running my instance for about a year without downtime.
I'm not sure what you prioritize in code, but reliability, readability, and extensibility are big ones for me.
In any case it was just one I personally use that where I know the devs had talked about using cursor a lot.
But I'm sure you'd just rip apart any example I gave no matter what it was. Go read the Django repo and tell me if that looks good to you, cause it's worse than a lot of the code we have in house.
tiajuanat@reddit
If C lets you shoot your foot, then C++ is Hemingway.
C++ is famously 4 languages in a trenchcoat, and nearly all the code online is what not to do, so the training code might compile if you're lucky.
Lucifernistic@reddit
This lines up with what I'd expect, tbh. I would imagine it does better at high level than low level in general, but also fairs better when the language doesn't let you fuck yourself too hard.
Empanatacion@reddit
This sub's orthodoxy on AI just doesn't match reality. I'm making exactly the code I'd have written by hand, but more quickly and with better testing.
The model is able to answer questions about our 50 different repos in 30 seconds. Questions that would have taken me an hour to research myself.
Head_Education9387@reddit
I started using Cursor + Claude 4.7 Opus lately, surrendering to the pressure from the management, after they said that AI usage will be tied to the performance review.
First, right now I'm mostly doing refactoring tasks, and it's mostly just moving the code around. The tasks are well-defined, and its output is easily verifiable. So the code generally works, with minor corrections here and there.
But man, do I hate my work right now. I feel as alienated from my work as I can possibly be. I hate writing prompts, and I hate reviews. I love writing code and structuring my work myself, and sometimes I complete tasks faster than an AI does (if they are purely mechanical). But now, with AI becoming my workflow, I feel like I don't do shit, and despite reviewing more, I don't feel any ownership over the code. I don't feel like I'm contributing shit, and at this point, I'm not even sure if I want to. I feel like jumping the ship and retraining for another profession. But don't know where yet.
brik94@reddit
This is where my head is too sadly.
thethirdmancane@reddit
Omg I've had it with this sub
Long_Egg_8644@reddit
This honestly sounds less like “AI helping developers” and more like management optimizing for output metrics nobody actually believes in.
Good engineers usually care about understanding systems, constraints, and tradeoffs. If AI removes onboarding, learning, and code ownership entirely, people stop growing and the codebase quietly decays under a mountain of fast-generated sludge.
RoosterBurns@reddit
Every single line of code is a risk to the business. Every single line has the possibility of losing money or customer trust.
Investors wanting the company to "embrace AI" then trickling down like a stream of piss to the most basic dumb metric imaginable "spend a lot of money pumping huge amounts of Slop Coding into the thing that makes us money" is genuinely damaging business after business in a way that won't be easily fixed.
Beyond the insanity and the idiocy of doing this the contempt - the actual blazing contempt - for the golden goose stands out to me. How are you supposed to look the people that thought this was a good idea in the eye ever again?
robert4221@reddit
You're confused as to the goal of a company as seen by the owners of companies. It's not to exist or be sustainable or survive for a thousand years. It's to increase stock price or valuation in some relatively short time horizon. Leadership is doing exactly what the people who own the company want and expect.
RoosterBurns@reddit
No it's to maintain a sustainable business right? Otherwise they'd be ripping copper pipes out the walls day one
I think you're thinking of sociopaths
robert4221@reddit
No. Sustainable businesses are not how investors make money. Stock growth is how they make money. If sustainable fuels stock growth then sustainable is valued. If it's not needed then it has zero value and a non zero cost.
Have you ever been in a startup? Or wondered why PE firms do what they do and keep getting money to do what they do?
No, I'm thinking of capitalism which has no morals or ethics or anything except for the growth of wealth. Frankly, cutting some corners on software is like the single least evil outcome of modern capitalism drives. Mass ecological damage, a hundred million dead from cancer, thousands dead from acute poisoning, towns wiped off the map, government overthrow, etc, etc.
edgmnt_net@reddit
The problems is those startups flop all the time and investors lose their money.
Also I'd say this has little to do with capitalism per se, in general. You need more than just capitalism to get this and modern states screw a lot with the money supply (creating cheap money) chasing endless economic growth. Of course a lot of it ends up being unsustainable growth. And I bet this would be much less common under a system with a roughly fixed money supply like gold.
robert4221@reddit
Why is that a problem? Investors keep putting money into VC in massive amounts despite it being literly known that 99% of startups won't do well. In aggregate investors either make money or gain value from the uncorrelated investments. You seem to be projecting your own person views onto another group that doesn't share any of those views.
03263@reddit
I'll be around to make some money fixing it if anyone ever decides to. I'm 50/50 if they'll ever come to their senses or just keep piling on shit and somehow make money because finance is completely detached from reality.
ShazaBongo@reddit
The question is if you will WANT to digg this shit code even for 100X money you get now. Mental heath matters.
03263@reddit
I've fixed stuff written by the lowest paid offshore devs who had zero insight into the product or its users. Doesn't really bother me to fix bad code, actually I enjoy making sense of it, what it's trying to do, and improving it.
RoosterBurns@reddit
Human bad code is obviously bad in a way LLM bad code isn't though, I'm always surprised at how superficially convincing LLM code looks until you drill down and find weird shit like casing differences between functions it wrote and the call site or it assigned the wrong values to reasonably named variables
03263@reddit
That happens with human code too. Debugging is fun.
RoosterBurns@reddit
It does but it does in such a way that you can understand normally what they were attempting to do and also again human "bad" code is obviously bad, if they're pressured or lost it shows in code written by a human hand, it's not a smoothly-commented sociopathic pack of lies
deZbrownT@reddit
Why is the speed only thing that matters? Is the company facing an existential crisis?
3_living_and_3_dead@reddit
My bleak take on all this: The non -technical management belief is that the belief of people with money to invest, whether they are retail investors, institutional investors, or VCs, equate token usage with value, whether that is to give the impression of speed to customers or other observers, or simply to compete on that metric with the other firms doing the same thing.
Anecdotally, a contact at a financial start up was explicitly told this: he needed to get his token usage up to increase the numbers to show investors.
Code or product quality, or even in some cases shipping the product, be damned, it was never about that even before AI. The veneer of caring about quality or even safety for customers has been shorn off, and the non-technical management now has more business leverage with engineering to not have to horse trade around precious engineers being given the space to address tech debt in exchange for shipping some feature a consultant promises will make a great press release. Rather, it has become do it or the sword of Damocles will come down and we’ll lay off the precious engineers in favor of the token arsonists. Because value.
deZbrownT@reddit
That just makes no sense to me. Why spend real time and money when you can just fake any token usage value. Companies have been doing fake it till you make it since forever. Companies have no choice but to live in the real market. Focusing on speed for the sake of speed on account of squandering resources, I don’t know. I can see it happening in very large organisations that have resources to throwaway, but for average companies, resource usage is a product of planning.
ShazaBongo@reddit
It's management BS. Almost all of tech companies suddenly count milliseconds. Doesn't matter the product has customers or not.
YK5Djvx2Mh@reddit
To be clear, milliseconds of engineer time, not execution time. They dont give a fuck about code efficiency
Zealousideal-War2807@reddit
I dunno. They probably don’t have to work that hard to justify themselves unlike the people who have to deal with them. Hopefully leadership can be reasoned with, but sometimes they can’t be
deZbrownT@reddit
I also don’t know. The post (to me) reads like:
a) incompetent leadership b) ai slop
Alternative-Papaya57@reddit
This solves all your problems https://github.com/dtnewman/burn-baby-burn
a_slay_nub@reddit
I know this is supposed to be satirical, but it almost feels not.....
big-papito@reddit
Hang in there. These labs need to IPO ASAP before things get really hairy for them. Once the real cost of AI kicks in, the calculus with change and vibe coders will be in the corner.
See: https://www.youtube.com/watch?v=aGBuhX1MYtw&list=PLADd6sStSis77HKfbf4KCY6SvthfxeUgn&index=1
roger_ducky@reddit
Your system needs to have additional guardrails.
Break down a story to about 1/3rd to 1/5th the size of what you’d take on.
Unit tests (interface first, unit tests with failing assertions second, passing tests third) to get proper(ish) conditions.
Test style: “Living documentation” — should document typical uses, expected primary error responses, and secondary assertions for each unit under test. Written as use cases, but with 90+% line coverage.
Mutation testing on the unit tests to catch improper mocking.
Linters to format code and point out overly complex code.
Once that’s done, a reviewer subagent to the main one. It should explain issues with the code with line numbers. Reviewer should flag any module > 500 lines as an issue.
Main agent should summarize the issues and what it thought about them, then you decide on the actual changes to do.
This entire thing takes a bit to do but ends up saving you 2 days out of 10, with similar code quality to hand written by a high performing junior dev.
Once this works, you can parallelize it by using worktrees and having coding agents run stories with no dependencies in parallel. Gives you more to review, but should then boost your velocity a bit more with more planned out.
I’m not sure if that’d fly where you are though, since my thing is about similar/slightly higher quality at a smallish time savings.
mvrckhckr@reddit
Talk to your manager about the conflicting objectives. If they won’t react, find another better place to work.
ShazaBongo@reddit
BURN baby BURN... the time when idiots run companies... Wait till they ask why people feel disengaged.
gordonnowak@reddit
this bullshit alone is going to bounce me out of the industry. fucking insane
gajop@reddit
Imo it's fairly easy to get AI to self improve the codebase (especially in self-refiew), and in that way burn a lot of tokens. Self review might find actual issues and help reduce the time it takes you to review + clean it up - I use it often on my sub plans when I'm too busy to actually focus on something and I know I'll have tokens to spare. The ROI is not great.. improvements are marginal at best, even with strict guidelines.
Really really stupid performance indicator.
Stellariser@reddit
Yep, if that’s what they want just set up a workflow with 10 agents arguing amongst each other about how to refactor and improve the code. Then you can ignore the results if you don’t like it anyway, or maybe it’ll produce something useful. Win/win!
UnderstandingDry1256@reddit
You need to show it who’s the boss all the time haha.
Tell agent this code is shit and make it rewrite in dead simple and bulletproof manner. Keep iterating until you like it. You’ll be surprised how good it can do. Use the latest opus or gpt and nothing else - the difference is huge.
Disastrous_Poem_3781@reddit
Don't use agentic AI??
BoBoBearDev@reddit
I have senior dev 10 years ago shitting on me using code gen from xsd file and he wanted me to hand writes all the models and parsers. So, I don't see how AI is worse than this.
Reasonable_Working47@reddit
What is the AI pricing deal you have, fixed cost or usage based?
HotJellyfish8247@reddit (OP)
Usage based, so sky is the limit.
drnullpointer@reddit
> My company openly says we should burn tokens, and vibe code, it's only speed that matters.
Okay, here is the thing.
Quality *IS NOT* the opposite of speed.
Quality is what enables speed. There is a sweet spot that will allow you best speed and if you overdo it and focus too much on quality it will slow you down again.
The general rule I have is that the longer the project, the larger the codebase, the more people involved, the more effort you need to spend on quality.
vocal-avocado@reddit
Big company?
HotJellyfish8247@reddit (OP)
Not FAANG, but globally recognizable.
vocal-avocado@reddit
Yikes. How long before you can quit?
Shazvox@reddit
Don't know what to say except; Yup.