Vibe coders don't know shit!
(midwest.social)
Programmer HumorWhat about us caffeinated autistic vibe coders? Slop production through the roof bb!
I put together a new modpack for minecraft this year and from that i can verify that this comment definitely checks out.
Vibe coders not understanding their code because they didn't write it.
Me not understanding my code because I wrote it a month ago.
My workflow is :
Step one: write down a bunch of nonsense, massive block comments of old code in case the new code is worse, and cludges in one massive file.
Step two: rewrite the whole thing: breaking it down into classes so it's readable and maintainable ...
Step 3: come back later and wonder where the hell anything is. Go to step 1
I'm going to lose my mind if I get any more support tickets where someone has cut and pasted the output from Claude instead of telling me what the problem actually was.
My job has basically turned into guessing what bad info Claude has given
My god, had someone send in a merge request with utter garbage, with Claude describing some nonsensical "root cause" as to why a function didn't work. Claude made two changes:
- If the current scheme failed, use an entirely different scheme it claimed the software supports that does not work and was never documented as a way to work.
- If that scheme fails, swallow the error and don't report the problem, leaving the problem silently in place but out of the user's view at least for a short while. This is actually the change that 'fixed' the problem.
So I needed to know the debug data for what really went wrong, but the submitter got all pissy and said Claude already sorted it out and why can't I just accept their slop merge request as-is, refusing to believe that all it did was catch and silently drop any and all errors associated with the code instead of fixing it. Wouldn't run the procedure I stated would show them they had messed up data as a result of their 'fix'.
Absolutely infuriating.
The fabled meat proxies.
Only one step away from becoming a reverse centaur
Just complete the AI slop trough and copy and paste the ticket back into Claude to fix and open an MR and then have Claude review it and merge and then when there is a bug repeat the process.
I honestly don't know what people are thinking, both employees and employers when people start copy pasting LLM shit. Why do I need you if all you are is a glorified macro? I strongly suspect the human is there to be the fall guy when the LLM invariably does something really stupid.
https://dontbeameatproxy.com/
Tom still used AI — and that's fine. Encouraged, even ... This is kinda only half serious. AI is great — use it all day. Just don't be a pipe. 🥩
That site is still promoting AI. So close, so damn close
I strongly suspect the human is there to be the fall guy when the LLM invariably does something really stupid.
This has a term, actually! “Accountability sink”.
Yes, like a heatsink in a computer, your job is to soak up all the damaging and unwanted aspects of what the “real” and “valuable” main components decide to do, so that consequences never reach any decision-makers.
Fun!
Decision makers are often set up as the "fall guy", but the consequences are getting a huge severance package and an even more lucrative c-suite position at another company.
Honestly, until recently the PotUS was a fall guy for most of congress's shit
Work slop! Just love people using LLMs to hand the heavy lifting over to me.
I don’t even get tickets. I get tagged into a slack thread full of slop trying to guess why I am here and what I’m supposed to do.
I have narrowed the containment of the isolation.
Now I’m verifying the resolution path of the indicators.
Permission to accumulate the resolution of the transformation.
I know what all of those words mean individually…
Just caffeine? I thought a sizeable portion of yinz are on adderrrall.
Some of us have to rawdog it, due to the meth myth.
What's the meth myth
That ADHD meds are just fancy meth, so we have to rigorously select who gets them.
It is just speed though
ADHD is wildly common in the field. Everyone in my infosec department has it, which is amazing because we all get the struggle.
Ah, actually, caffeine can be really contra-productive for autistic people.
For me it doesn't make me feel awake, only incapable of sleeping. Which is hell, like being thirsty and not being able to drink
I was only ever observed to have autism, but I highly suspect that I'm audhd. What kind of hell is this where caffeine can help my ADHD focus but harms my autistic focus
Caffeine be like:
I'm doing thousands of calculations per second and they're all wrong
If it's any consolation, I'm AuDHD myself (ADHD diagnosed first) and I would have never made it this far without copius amounts of caffeine.
Edit: To expand on this, my autistic tastebuds have rejected the taste of coffee until I was like 21. Imagine the astonishment following my first cup when I was bustling around actually getting things done for once.
I seem to have developed some kind of ADHD in adulthood, while also avoiding caffeine in recent years. Now, if I drink two cups of coffee, I come to four fours later with fifty new tabs in the browser, the system deeply in swap, two new projects started and abandoned, and nothing actually done. If I'm lucky, some system reconfiguration that I've been putting off for months, will finally be made.
Depends on use case - I find it helps a lot with things like competitive video games.
a fraction is being too kind to us
from what I've heard from my software engineer friends, a decent coder with standard knowledge of CS and the language they're using will always write better (more optimized, more debuggable, etc.) code than currently available LLMs
even if the "fraction" is closer to 1 than 0, the heavy lifting here is done by "look what they need." vibe coders need gigantic climate destroying machines to write code that isn't as good as what an educated programmer writes running on a burrito and coffee
Depends on your domain. Devops is mostly a solved problem terraform is well documented and most infra follows a far more simple pattern. It’s not replacing engineers but if you aren’t a moron it can be a 5x type tool. It is extremely good diagnosing issues, parsing logs, and frankly writing well formed TF. It’s garbage at architecture and every once in a while it’ll absolutely send you a wtf were you thinking type pr. But it’s 80% good with a 20x in raw velocity.
Our team has fixed up a lot of shitty technical debt, increased automation and our deterministic scripts and workflows, and increased usable test coverage on $30/mo copilot plans.
The big problem is that sometimes it's "good" at diagnosing issues and sometimes it's very bad and in both cases it looks the same if you don't have a way to know yourself.
I've seen some of the fodder people consider 'well formed playbooks' and it's also a bit of a crapshoot there. It can speed up some tedium, but so many users get in over their heads.
Today had a junior proudly provide an untested playbook that just did everything wrong and even if it did what it purported to do, it would have been hard coded to certain things that are not what we want to do.
Unfortunately we live in day and age when corporations prioritize quantity over quality.
And that age is going to be very short as it will end with flaming hacks and downtime.
Been suffering the hacks and downtimes for over two decades waiting for the day when corpos come around on this.
Hard to continue holding out hope.
They always did this. Now they are just doing it faster.
I’m tempted to try to calculate how much energy has been used to produce the currently and previously available library of anime popular amongst coders, because that’s a necessary resource, along with various energy drinks and snackies
Lisdexamfetamine+caffeine, let's gooooooo
Hands shaking but mind is going brrrr
It is by caffeine alone I set my mind in motion. It is by the beans of Java the thoughts acquire speed, The hands acquire shaking, The shaking becomes a warning. It is by caffeine alone I set my mind in motion.
Everything brrr if you coffee enough
can't eat, coding.....
i don't even need caffeine to write code (of varying quality) at high speeds.
I could either take several tedious hours to write out several thousand regex patterns by hand to automatically recognize differing formats of notation for this thing I'm doing IRL......
or I could have an LLM do it in a few seconds, and it'll get 99.99% of the cases, missing maybe one case every hundred thousand...
take your guess as to which option I chose.
LLMs are a tool. I use it as such.
Frankly, your use case sounds like something has gone horribly wrong from the outset. If something needs several thousand regex patterns, then something is very very wrong, and there's zero chance you have a comprehensive solution as it stands. Either trying to use regex for a use case that some natively AI approach might actually be warranted for, or regex is being used poorly and each one is too limited, or some other approach entirely is warranted.
well, it works, and i don't need to touch it.
remaking what is already working would cost more time.
Isn't this a case where regex may be the worse choice?
write out several thousand regex patterns
Oh my! This thing is more likely to gain sentience than any llm.
it is a truly absurd number, simply due to the inane BS variations on terminology and notation.
It might have been more reasonable to train a local LLM on this material and then just query -that-..... but hey 30ish seconds of claude's free tier usage; and i've got a script that's ~99.99% as good as a human in terms of pattern recognition. Pretty neat when you think about it.
And 'truly absurd number' might still be a bit of an understatement here. Saying that Claude can generate several thousand regexes for you is not really a selling point for llms in my book.
On the other hand. Have you tried running them on the bible or something like that? I bet it finds it talks about ww2, 9-11, the nintendo wii, and shit...
If, by bible, you refer to the holy book of the local Building Code; then no, I have not.
and again; the regex in the script generated works, it's just a teeny bit cursed and maybe a little bit sentient. Praise be to the Omnissaiah, and its blessed machine.
Can't be worse than when I was an intern, the product was literally based around parsing arbitrary HTML with regex and not even decent regex, POSIX regex from like the '80s for reasons known only unto the programming gods.
sweet merciful torvalds, that must have sucked.
Glad you made it out of that hellhole.
Cheers, me too! Was educational as an intern to say the least. I think the firm went belly-up a few years later.
When I wrote my numerical parser, people told me to use regexes to detect what kind of numerical value is being parsed. I just did it in one go. Only regex-like checking is for ISO datetime values.
regex is computationally pretty cheap, from what I've read; due to something about it being a simple AND gate or something..... . IDR, and to be honest it doesn't really matter in the era of 16-core/32thread 5GHz processors.... but when you just want it to work..... it does "just work". /shrug.
Not really. There will only ever be about a hundred people who ever use this tool, and it's really just a massive timesaver that keeps people from needing to open a smorgasbord of files to get a few tidbits from each file; then collate it into one doc....
If I get ~99.99% of all the information, and one hundred people use this, odds are, nobody will ever see the .01% edge case of missed info. and even if it does happen, it's not like it's obsfucated information in each file, it's just a pain to open a hundred files and wait for them all to load at once through the disk-heavy GUI application we have to use.
simpler to just read via the regex patterns, and then put it out into the terminal for the info needed.
or I could have an LLM do it in a few seconds, and it'll get 99.99% of the cases, missing maybe one case every hundred thousand...
Even if the LLM does things perfectly 100% of the time, how do you expect to maintain your skills if you offload all of your work onto a chatbot? What happens when you start offloading more and more until you're not actually doing anything except typing in prompts?
My skillset is geared towards meatsacks and things that go boom. LLMs have been a massive boon to the 1% of my job that needs a code monkey.
As long as people still talk to each other, and things need exploding, I'll have a job. Heck, if we somehow get to a point where money isnt needed, I'd still do what I do simply 'cuz I like it.
Yeah I'm not a fan of slop merchants and PRs the size of battleships but anyone solving merge conflicts by hand because they hate LLMs is being a tad masochistic in my opinion. I'm in two minds about how much coding itself ought to be automated, but I'm a strong advocate of automating the tedious donkey work we'd absolutely have got rid of years ago some other way if possible.
It's also good for avoiding tarpits, sometimes I literally don't want to learn something because it's too expensive relative to the value of retaining it. I would rather crawl over a mile of broken glass and hot coals than learn the detailed ins and outs of WSDL just to support a legacy feature which is firmly in the crosshairs of my deprecation gun, to use a totally neutral and not driven by personal bitterness example.
Lots of folks here create absurdly high expectations that most coders don’t meet. They are out there but they ain’t average. Getting it 80% correct with good human reviews to catch the trash and using it to figure out what’s breaking quickly vs tracking down logs has been great for our team. I laugh at some of the garbage it spits out and the dumb way you can see shit like, ‘I have to be honest I did zero testing’.
Even though i hate ai and what it has done to my profession i can confidently say that you are in denial and this is pure cope. Agents can already write much more organized code than any developer in my company. "Knowledge" and "architecture" won't save anyone since llms already have plenty of that. We became a meat proxy and it spreads in other professions too. What a shame.
Agents can already write much more organized code than any developer in my company.
Then you must have nothing but absolute shit developers.
Honestly, have you used an agent ? Comments like your make me think your llm usage consists of copy paste to "chat" instead of prompting an agent. That or all haters are Jake Wharton level and actually produce better code manually fast enough.
I have both used agents and have been the "victim" of heavy agent users.
If you can provide an utterly well bounded problem with perfectly verifiable criteria for it to retry against until the tests pass and the tests are full and valid for the use case, it can work. Similarly, if your code's reality is forgiving and flexible, and a 'close enough' result is good enough to get the human software user in the right ball park, you might be able to extract decent behavior.
However, if there is a means by which the answer can 'look correct' yet be wrong in the real world and the real world scenarios require accuracy and precision, there's huge gaps.
Especially if the agents control the coding and the test case generation, seen plenty of times where it talked itself out of a test case that was failing when the test case was in fact correctly showing a flaw.
That's actually my biggest issue with it, it often tries to get out of scope and touch something it shouldn't. But thats a low bar for an entire profession to exist. I wish llms never happened, ruined our field and I'm upset about that. I honestly cannot imagine a CS field it wont disrupt.
You know I keep seeing this, and I keep trying AI, and my experience really just does not track with the grandiose claims y'all AI fans (with 2 comment accounts that have been dormant for six months...) keep making. Maybe you're some canned code shop with crap devs, but even then I'm real curious what the use case is where AI agents can take over from real devs like you're claiming. Sounds like marketing hype more than anything.
I keep trying AI
When you try AI, how do to do it? And do you try for hours? Days? Weeks?
There's a learning curve to use these tools productively, and they often feel unproductive early on while building your proficiency.
I think the learning curve is often misunderstood or underestimated, possibly because we interface with these tools using natural language. We overestimate how "easy" it "should" be. (This was my own experience anyway, when I first started learning to use AI)
I know there's a temptation to call "skill issue" when you see anecdotal evidence that runs counter to your own experience, but I can say with some degree of professional certainty and a smidge of personal snark that no, I in fact do know what I'm doing. These tools have their uses, but their general applicability is vastly overstated (if I were the kind of cynical person that enjoyed being right, I would say intentionally overstated in an effort to attain profitability on bootstraps made of hype). And I wanna clarify that I do indeed use some AI tools (ex: security checkers) with regularity. They do absolutely have worth - they just don't have trillions of dollars worth... of worth. Ahem.
The hype-doomer thing that some people have about this, oh woe it's going to destroy an entire conceptual industry, that just isn't what's happening, but it's very much what the advertising would like you to believe it's doing.
You may try to attack my character by implying a bot or throwaway account but my point stands and all metrics agree with what i said. I literally used the newest model of a certain company (not naming them) and the agent finished a weeks work in 2 sessions. It even found lib incompatibilities and a weird bug in codebase without even prompting for that. The models became better but the harness got really good now. Your experience may be a skill issue, your prompts may suck or even your codebase might be too big, or cope, either way we are screwed, please don't give a false hope by moving the goalpost each time a llm evolves further, it's better for programmers to swim with the wave and evolve with the profession than be stuck in denial and becone obsolete.
Y'all have been saying "the problem is you're not using the newest model/harness" to any criticism for 3 years now lol
"But have you considered that it might not be a fundamental misrepresentation of the capabilities of our system that's been borne out repeatedly in real-world testing, but instead a skill issue?"
We didn't even have proper harnesses 3 years ago. Claude Code which more or less kicked everything off is under 2 years old, before that you would've had to use Cursor or something. And we now have local models better than the frontier ones of 3 years ago if you have a good graphics card or an SoC with lots of shared memory. The cheapo Chinese models now beat the super expensive American ones from just a few months ago.
Whole thing is changing so fast if you haven't tried it in a few months, you don't really know what it entails now.
GPT 3.0 couldn't create gta 6 using a single prompt therefore all future models are bad. /s Maybe get out of your cave and test it yourself instead of coping in denial.
The problem is that every time, in the moment, people advocate and then when it comes up short, come back with "Oh, you did XYZ 3.2? Yeah that's busted, 3.3 can really do it" and rinse and repeat and it's hard to take that argument credibly when it's a treadmill of dissing yesterday's tech but swearing today's is different.
The current best of breed I have very limited exposure to (too rich for my blood), but it didn't seem overwhelmingly a slam dunk in my interaction. Incremental value going from more modest models to those don't seem to justify the price tag.
And every developer that is merely curating fully agentic workflow I have dealt with has pretty shit functionality and no idea what the hell they are doing. Whatever the rhetoric is, the results are just utter shit. Somewhat better if the problem domain can be infinitely retried without consequence with results that can be perfectly verified to let it drive eternal retries (while burning through token budget), but generally I just see pretty shit software.
On the other hand, a lot of these groups that I say has pretty shit AI software formerly had pretty shit normal software. Problem being clueless management now thinking they must be smart because they say things more aligned to the AI hype.
Every technology we have today started as being barely usable to "pretty good" across it's lifespan. Why would you think llms are different ?
I have tinkered with it, and I've also had to deal with coworkers using it. I have actual work to do instead of playing with toys that make my job harder.
That may be your experience but it's clearly the exception. Every developer i know doesn't even use ide anymore and have not written a line of code for months. Most of us hate it and prefer we went back to the old ways of manually typing code, but sadly the llms won't go back in the bottle. You are in the denial stage and it's ok, we were all there.
Perhaps your experience is the exception?
(Dude you've been trying the "smug condescension" thing for like two hours and nobody's playing along, give it up already it's tedious)
Dude, come on - what are you actually trying to accomplish here? All you're doing is throwing a tantrum because people don't agree with you, and that's pretty unpleasant behavior.
Arguing with facts is now a tantrum ? Please refrain from ad hominems if you cannot argue honestly.
... Huh. So you don't really have a goal here beyond just offering up bait?
I implied you were neither of those things - I just quite openly implied you're a shill. Perhaps you should ask an AI for tips on improving your reading comprehension?
Please give me 1 thing that a mid developer can create but the latest model agents can't. There is coping and there's delusion, i think you're heading towards the second.
It's genuinely depressing you're so creatively stifled that you haven't encountered a task that meets those conditions yourself and instead have to retreat to these middleschool-tier insults, but hey, if you're gonna throw down a mitten in challenge, who am I to refuse?
Novel reporting across diverse data systems! That's not even a mid developer task, that's intern shit - and good luck getting an agent to deal with that one lmao.
The more i use it the clearer it becomes it can do much more than a human. One of my hobbies are electronics and embedded development. The llm designed the circuit and even the firmware in 5-6 prompts, that took me a week on a similar project. With clean requirements and scope it may do almost any task, it may struggle on huge projects and monoliths but looking back to each improved model i can guess that won't take a long time to accomplish either. In less than 2 years we went from buggy gpt 3 to multi agents workflow that work almost autonomously.
That you've gone from "I bet you can't name even one thing AI can't do", to being told something it can't do, and now immediately you're ignoring that interaction in favor of telling anecdotes about how great AI is? That says a great deal about your conviction to your own arguments.
You basically did tho lol
Hmm. No, but thank you for your contribution!
It's not up for debate. I'm literally telling you how people are perceiving your response. It's not something you can disagree with. Just because you can't understand that other people are reading you differently doesn't allow you to give it a "no thank you" hahaha
No, (but) thank you (for taking the time to clarify)!
I think the truth is somewhere in the middle, but I could be wrong. I work in a non-programming technical field and I am also impressed and concerned by the progress the models have made over the last few years, especially in terms of being able to "do my job".
Like the other guy was saying, there are still issues with its implementations, and the second you begin to think it's perfect or immutable, you notice issues that would have been show stoppers if they made it through. And at the end of the day, at least for what I do, if nobody is here to press the buttons, the AI will never be able to do its thing to begin with.
I also personally predict that the cost of tokens and utilizing AI services will eventually become overly expensive and encumbered, as does everything under capitalism. Name one thing that hasn't gotten shittier and more expensive overall. The AI companies haven't even begun to enter their enshittification stage yet.
I maintain that AI in its current form will NOT be able to produce truly novel thought or original work, as it can only put together the Legos that it already has ingested. Perhaps grunt work programming will require less manpower, but cutting-edge developments are still beyond the LLM's grasp at the moment.
Anyways, I welcome other opinions
Software development is by far the most affected profession. Every time a new model gets released another part of my occupation can be done by it. People are moving the goalpost to focus on every mistake the llm makes to cope and fool themselves they are still not expendable and better than the llms. With enough confidence i can say that the vast majority of programmers are not.
Yeah. And the thing that people are completely ignoring is that humans are absolutely not infallible either. This is why we're seeing so many vulnerabilities pop up now, because the work performed by humans over the last many decades is being picked apart by a relentless onslaught of automated calculation and critique by these LLMs.
I can see the passion a lot of programmers have for what they do and it reminds me of how artists feel when people try to produce art with an LLM. I hope the future brings a reasonable balance where programmers can actually use LLMs to advance their efforts instead of being replaced by it. There are some things that are just mindless brute-force effort and at the very least, hopefully that bit goes away and leaves programmers (and the rest of the world's technical workforce) with the ability to exercise creativity instead of doing the annoying parts all day.
Beautifully said, i agree completely !
