[-]

Glum-Psychology-6701@reddit

Ok we'll revisit this in a year, such posts are dime a dozen, the same anti AI rants with the same vocabulary .. What makes a human programmer better than Ai ?

[-]

NoRegreds@reddit

Tell me that you never wrote code yourself without telling me you never wrote code yourself.

[-]

Glum-Psychology-6701@reddit

Tell me you don't know what you're talking about

[-]

I had a claude do a complete and utter meltdown during some of my experiences. I've been calling the problem of AI Assistants being easier but not faster the "babysitter problem." I wrote about it here: https://chrisbeckman.dev/posts/the-babysitter-problem

[-]

Complex_Disaster1552@reddit

Shouldn't this technically be "improved" when/if AI coding becomes even better?

Ok_Individual_5050@reddit

Where did you get the idea that you can take a tool that is notoriously good at creating security issues, ask it to review code for security issues and somehow get a better result?

[-]

Complex_Disaster1552@reddit

Ha, makes sense. I would say - especially in Lovable's scenario, it did do a good sweep of Supabase's settings

[-]

Elctsuptb@reddit

Yes, everyone here is coping hard

[-]

lokaaarrr@reddit

It’s pretty clear that current AI tools don’t produce better code than the person using them could have written. In many, but not all, situations they reduce the time needed.

AreWeNotDoinPhrasing@reddit

Shit, I’ve had almost this exact conversaition before with Claude, in CC.

[-]

oppalissa@reddit

Well, that's one way to fix it

[-]

UltimateTrattles@reddit

It regularly fixes type errors by just casting to any.

[-]

oppalissa@reddit

What AI are you using? Copilot is pretty good

[-]

UltimateTrattles@reddit

Copilot uses gpt.

I mostly use Claude via cursor but also use other models. I’ve found gpt5 and Claude to be the most reliable.

Need a good rules file to keep them on the rails and a spec to tell them to build from.

[-]

RadiantHC@reddit

lmao

[-]

Hawkes75@reddit

That too - and I'm like sorry, linter rules.

[-]

JuiceChance@reddit

You will find money, a lot of money that you should charge those CEO, CTO idiots and greedy stakeholders.

[-]

No_Oil_6152@reddit

Agreed.

Bill them double to fix the issues they'd never have had if they hired human professionals to create their app.

Feel no shame for doing so; these same companies would happily see you thrown on the scrap heap because "savings"

[-]

oppalissa@reddit

I believe it's only a matter of time for new gen AI to fix the mess of the old AIs, and again and again.

Time to raise chickens

[-]

Skibidirot@reddit

so true

[-]

AmbitionHopeful7227@reddit

Or "that happens because you have an old version of a library" (not true), or "your environment is corrupted, just reinstall all lol"

[-]

DoctorWaluigiTime@reddit

I'm guilty and have done that on occasion. "Oh this works if setTimeout(0) happens, because something something race conditions or something something new thread happens.

Sea-Employment3017@reddit

The setTimeout (fix) is a classic AI hallmark. I've also seen it suggest any type casting for TypeScript errors and .catch(( => {}) to "handle" promise rejections. The scary part is these "solutions" often work during development, so non-technical stakeholders think it's fine. Then, production breaks under load or under normal usage.

[-]

JakubErler@reddit

Yeah exactly, every time I asked it for a JavaScript patch of a system we are using, it uses setTimeouts and the code is unreliable (it fires too early - the eom is not yet there, it fires too late - it blinks). I had to explain to AI how to replace setTimeout and why. It replaces it and several iterations later it quietly reintroduces it back. Interesting. Maybe it is trained on a low quality code

[-]

DrobnaHalota@reddit

There is a lot of cope in this thread. Next gen AI models will do the clean up for the current gen.

[-]

AntDracula@reddit

lol THIS is cope

[-]

TechnicianUnlikely99@reddit

What should we do? Should we leave the field asap and become nurses, plumbers and electricians?

That feels like the new “learn to code” that will be completely saturated in the next 5-10 years

[-]

HomelanderOfSeven@reddit

Well they better be hurry while “AI” bubble still floating

[-]

Several_Note_6119@reddit

Why do we think that AI slop won’t be cleaned up by AI in a few years? AI is moving fast, so maybe in a few years, it clean itself up to improve performance, maintainability, etc

[-]

AntDracula@reddit

Just 2 more weeks to AGI bro!

[-]

silentjet@reddit

the problem starts even before the code. Do you know what's the problem? Name! What is being called "AI" is not actually AI, not even an element of AI. We, as programmers, know if you'll name an index variable as a sum variable, that would not make it actually a sum. the same here, a great technology, ml, was brainwashed to be an AI, which it is not. But now an expectation and huge selling point is expected to be like AI from futuristic movies, and it cannot give that, because it is not an ai.

The other side of the problem is that ppl start to be scared of the rise of machines, for their jobs, for their future, because it was promoted like that last 50years of movies and that is really bad.

Maintenance for massive, older services are the hardest jobs in the industry. There is negative value in cheaping on those positions.

[-]

plinkoplonka@reddit

They say AI.

it just means Affordable Indians.

[-]

Life-Principle-3771@reddit

This is correct, have worked at multiple FAANG and left G recently.

The truth is that AI is not that really used in FAANG companies for development. Maybe a little bit for stuff like drafting reports/emails and there is certainly a lot of interest in the MCP space...but the truth is that the performance/accuracy of AI code generation is negatively correlated with the size/complexity of a codebase, which is a big time blocker for the massive and highly complex codebases that you deal with in those companies.

It can be a useful process for engineers, too.

plan to throw one away; you will, anyhow

With AI you can quickly create throwaway prototypes.

You should be very proud.

Going by my friends experiences with artists, same thing.

Some guy generates a mockup of "the vibe" of what they want, concepts mostly.

ikee85@reddit

What of waste of time and resorces.

[-]

Mission_Cook_3401@reddit

As is life

[-]

Efficient-Design-174@reddit

rainroar@reddit

Yes you do when you spend a decade telling investors that hiring everyone at all costs was the only path forward. It makes you look pretty dumb when you realize cutting staff is the only path to increased margins.

I’m telling you, both places I’ve worked that claim to be “ai focused” are absolutely lying sure we burn 3 million tokens on a day because we were told to or else, but that isn’t what’s doing the work. Devs are doing the work.

[-]

teslas_love_pigeon@reddit

It's really something how others can't easily see through the farce of what it is. I really wonder if they've even done any critical thinking in school, no not the insult but like a class on critical thinking.

All throughout my public schooling we were taught how to think more thoroughly regarding political news and events. About conflicts of interests.

Maybe they should cram this into the bootcamps?

[-]

shadowsyfer@reddit

Social media and tech influencers are to blame. Not all of them of course, but a large number.

[-]

teslas_love_pigeon@reddit

It plays a part yes. It does feel accurate to say that 95% of tech content tailored to programmers is just engaging in boosterism with no real push back or discussions on the political economy within tech.

Sadly I think that the companies that "vibed up" will disappear once they get hacked, or investors realise they've sunk money into a ChatGPT wrapper.

So they won't be around long enough for people to be needed to fix the code.

[-]

pm_me_ur_happy_traiI@reddit

If you’re right, it’s the end of high salary days for devs. That small shop is going to pay 1/3 of what a bigger company would have, and the rest will be valueless stock options.

[-]

jhaand@reddit

I think a lot of contractors can do the clean-up work from all the vibe coders. You could even make a business model from it.

Create a small company with 5 dudes that specialise in testing, QA, refactoring and architecture archaeology and Devops. You can go from one pile of AI slop to the next one I think.

[-]

SignoreBanana@reddit

Maybe eventually. It's going to be a long few years though

[-]

No-Row-Boat@reddit

Well the numbers aren't backing that thesis unfortunately.

800.000 layoffs, 70k jobs created in the last period?

All the AI optimists should start watching what Geoffrey Hinton is talking about. AI replacing us in the workforce is a more optimistic outcome.

[-]

enzamatica@reddit

That would be incredible. Now if they didnt lowball.

[-]

quentech@reddit

That's the tough part, but if you find one that's willing to open the purse strings, you can have a ton of leverage as the magic fixer-upper in a small company.

[-]

quentech@reddit

two has opened up at a startup that got funding because they vibed up at POC

I'll take it.

Can confirm. Even at big Fortune 100 size companies there's a lot of vibe coded slop that works well for MVP but will absolutely fall apart under stress and within production environments once they scale to more users.

[-]

GWstudent1@reddit

I’m desperately trying to fight off vibe coding at the start-up I work at to avoid this shit. Managers’ brains were permanently broken by agile/MVP//first-to-market product strategies. The only thing on their mind is “Just get it done. Make it work. We’ll make it scalable later.” And that pushes a culture of manufacturing shit code that doesn’t stand up to the pressures of the real world and going make to make stuff scalable is 10x than building it correctly the first time.

[-]

Regular_Tailor@reddit

The times I've heard "gold plating" from people when they mean "might not fail" is amazing.

[-]

krisolch@reddit

> Make it work. We’ll make it scalable later

Maybe they are correct?

Most startups should not care about scaling until they hit the problem of scaling. Until you've been a founder yourself you won't understand this. And yes I'm a 12 year experienced dev

[-]

jl2352@reddit

They are correct, however on the surface that type of talk (get it done and care about scalability later) is easy to say and hard to do. It can lead to very bad results, like the shitty code the person describes.

I find it's better solved at the ticket stage. Where you can design features in ways that the feature does less, or has acceptable restrictions built in, and then you write good and proper code that fits those lower requirements.

[-]

GWstudent1@reddit

That’s workable when you get your code to run for 100 customers and then run into a scaling problem at 1000/2000 customers. I have like ~15 demo customers to work with and I know we’re aiming for hundreds of customers at some point. I’ve been given AI-written code that’s like O(N^2) that could be less complex. I know that’s going to hit the wall when we start selling. And I could keep provisioning more DTUs to keep refreshes from hanging, but why would I increase our bills that high when I could just write better code?

[-]

pavlik_enemy@reddit

If you have funding you should totally just throw money at the problem

The most difficult thing is to be able to tell people "told ya" without pissing them off and then somehow they should accept the fact that optimizing the code will take a long time after plowing through features with breakneck speed

[-]

GWstudent1@reddit

It’s also a startup I own in part, I’m not just an employee drawing a paycheck. It’s my fellow founders (who are not programmers) that are pushing AI slop onto the project that I know is not scalable. And going from $200 a month for database DTUs to $4000 a month is unsustainable for the amount of money we have.

[-]

pavlik_enemy@reddit

I guess you should pull rank. If they don’t trust your technical expertise why are you even working together?

[-]

break_card@reddit

This effect is probably costing economies billions each year. It’s the fundamental problem of enterprise coding. Rough hacky MVPs are needed to survive at early stages, but these will make your dev cycles lethargic and rotten as it scales to a medium sized business. It’s impossible to know the exact right time to revamp your MVP into a long term software system. It’s extremely hard for devs to convince business folks to make this investment.

[-]

jollydev@reddit

Most MVPs never result in products that get to scale to a medium sized business. It's a luxury problem to have a tech indebted product that actually has PMF.

Comparing two sloppy MVPs vs one high quality MVP and I'll take two sloppy ones all day to double my chances of finding PMF in a given timeframe.

[-]

Subject_Health_3182@reddit

This reminded me of a story in early 2000s about a company with smart bot that was better than Google. You typed your question, and while info was “loading”, real human behind that screen googled, and then typed back the answer in chat.

[-]

DecodeBytes@reddit

I am building an AI Agent Framework at the moment, so frequenting that world a lot. The tidal wave of crap is flabbergasting. My project is quite new, but already far more clean and structured, yet it see's no where near traffic and hype of all the slop out there.

I often look at these projects and its easy to spot them. README's full of emojis. Lots of extravagant promises, like a project which is less than a week old describing itself as 'Enterprise Grade'.

Reams and reams of documentation, some of it hilarious. For example, I came over one yesterday that was just a ton of prompts wrapping open-AI APIs and tons of functions that were not even called, it was described as a 'Cutting Edge, Enterprise Grade, at-scale, Multi Agent Swarm'. In the docs it makes claims about SLA's and uptime. Yet it's CLI!??!

The interesting this is though, we have a very low star count (I know, vanity metric, but helps make you discoverable), yet those who are turning up are seeing the potential and actually contributing. So in the end, I know we will come out on top of the slop.

[-]

fmgiii@reddit

Oh you're missing the boat if you're not making this absolutely intentional. It is such a fantastic opportunity that you can't possibly pass this up. Four decade+ IT architect here. Just signed my first ever severance agreement in my entire career. I'm essentially done in 7 months. Being replaced by a crew of outsourced devs from India. Very nice people. I have no qualms with these folks. People need to earn a living around the world, and I am just fine with that.

So I'm tasked with 'knowledge transfer' and of course 'getting sprint tasks completed' because nobody else can yet. Indeed the PR's are coming in from overseas, all AI generated. No surprises here. I fully expect that. The folks from overseas are not stupid. They know how to play the game. Especially with US based money flow.

So (insert Mark Manson's 'Subtle Are of Not Giving a F*k') I start thinking to myself, why do I care so much? Is it because I poured my heart and soul into this code base? Giving it every bit of attention and honor that one would give to something that truly mattered? Or because I cared about the users of the app? Their lives? The impact they would have on others because this code base was of the best quality I could possibly produce? Yes, indeed that was the reason...but

...capitalism doesn't care. So WHY...SHOULD...I?

Introducing a $25.00/month Claude Code Pro subscription. Perfect! They want AI. They're going to get AI. And honestly it's been a JOY to BEHOLD such a fitting REVENGE!!! And so f'ing easy.

I just let Claude Code SLOP IT OUT FOR ME. And it gets reviews and checked in!!!! The code works! Is it extendable? No. Does it fit into the context of the larger architecture? Just barely. Does it fit any overall standards? No. Does one function resemble another in terms of coding style and approach? Absolutely not! Is it code that will stand the test of time? Hell no. Will it be maintainable in a year? Not by anyone will less than 3 decades into the industry. And the list goes on, and on, and on of how this SLOP CRAP violates everything we have held to quality up until now.

Are the capitalists happy? Of course they are! Look at how 'productive' I am now! I can bang out a new radio-set by lunchtime!!! Full stack!!!

But the Beast lies waiting, for the future. Salivating. When the wall hits. When no one knows, anywhere how the overwhelmingly expensive codebase, got to the point where no change can be made to it, without the entire castle, falling to the ground like a house of cards.

Live the dream! SLOP YOUR WAY OUT! We deserved better!

[-]

pavlik_enemy@reddit

Given that very few people comment the code, if there are comments at all it's AI generated

[-]

DogOfTheBone@reddit

LLM comments are generally hilarious it'll be like

var playerSpeed = 500 // This is the player's speed

[-]

pavlik_enemy@reddit

And then there will be something like speed = speed & 0x7F << 15 without any explanation

[-]

Sleisl@reddit

john carmack is stuck in the AI

[-]

Zywoo_fan@reddit

Prompt: "You're John Carmack ..."

[-]

TangerineSorry8463@reddit

// evil floating bit point level hacking //what the f?

[-]

perk11@reddit

This is one of the 3 things I ask ChatGPT not to do in my custom instructions, and adding it there seems to work well.

If providing CLI commands, do not break them into multiple lines. When outputting code, make sure comments only describe things not evident from the code itself. Always use descriptive variable names, make them long if necessary.

int var; //Ex 1, 2, 3, 4...

zicher@reddit

I have definitely noticed that this seems to be a good indicator

[-]

LossPreventionGuy@reddit

if the comment has an emoji it's a guarantee lol

[-]

zicher@reddit

Right? What is it with all the emojis. They weren't in the training data, that's for sure.

[-]

Chris_Newton@reddit

There seems to be a trend for adding emoji to comments in documentation and blog posts, often highlighting a point that is made in the accompanying text. It seems reasonable that if those sources have been used as training data then an LLM would generate comments in a similar style.

I’m not sure I’d want to see emoji in production code, but as a presentation technique in documentation that features code snippets, it seems quite effective.

[-]

thephotoman@reddit

How much do you want to bet that it’s because Microsoft devs use Slack emoji in their comments, and the AI was trained on Microsoft’s code?

[-]

ElonTaco@reddit

I assume they trained it on tons of messages, tweets, etc.

[-]

Fun-Put198@reddit

I have

.childOption(ChannelOption.TCP_NODELAY), true); // ✅MUST be true for games!

on my Netty configuration, and to be honest, I have become lazy to delete some of them when copying lines that are actually useful

[-]

ryanstephendavis@reddit

// 💡Intialize a new variable x to zero 
int x = 0;

[-]

failsafe-author@reddit

I used to be someone who commented with the why in nearly every line. I now have the why in three other places before a comment: 1. High level documentation in the repo. This is often not suitable for what we’re talking about, but it can give context to what the code is overall trying to accomplish. 2. Unit tests. I am a TDDer, and view tests as enforceable requirements. So often I’ll have a test along the line of “must be uppercase to meet international standard”. 3. Method and variable names. As I said, any time I’m tempted to write a comment, I try to write a method instead that is named what that comment would be.

But there are exceptions, for sure. I will always think through “will a developer understand why I’m doing this without explanation” and try REALLY hard to do it without comments, but if I can’t, I’ll put the comment. And think I generally do a good job of getting this right. I rarely struggle to understand the “why” of my own code when I return to it, and other developers generally tell me they find my code easy to work with.

[-]

a_lovelylight@reddit

I find a lot of devs don't understand what's meant by writing "why" comments.

Sometimes you can't encompass a business decision in a variable or function name. Or sometimes you're doing something that's kind of shitty, but is necessary for various reasons (usually related to deadlines or some idiotic thing upstream that no one has time or will to fix), and it's helpful to have it documented right there where the next person to come along will likely have questions.

[-]

geon@reddit

Like a library that takes arguments typed as readonly, but actually mutates them, requiring you to make a deep copy before passing them in.

That *requires* a comment, or it will be removed.

[-]

pavlik_enemy@reddit

Comments are needed when the code does something weird and most of the code is pretty straightforward

[-]

ihxh@reddit

I don’t think you should not comment at all.

Code describes how, comments should describe why. It should give the reader the answer why solution x was chosen instead of solution y.

Commenting things like “add a to b” is useless, but something like “we need to add a to b because system c expects xyz” would be better.

If you have to explain a lot of the “why” in your code base then restructuring might be something to look into. But “why” can also be a non-code / business requirement.

[-]

failsafe-author@reddit

I’m surprised to be getting this much pushback- haha. I thought it was commonly accepted that comments are a smell. I suppose not.

But as I said in my other comment, I try to use methods in place of comments, and only fall back to comments when I cannot clearly structure what the code is doing in a way that is self explanatory.

I completely understand that comments are for why, but generally, with small, well named methods you don’t need comments.

[-]

pavlik_enemy@reddit

Sometimes it’s better to separate large method with some comments instead of splitting it into small methods that don’t have clear and obvious scope. I don’t think “cleanupWeirdShitOurClientsThinkIsCSV” is a good method name

[-]

geon@reddit

"cleanupWeirdShitOurClientsThinkIsCSV" is a bit passive aggressive, but a good scope. It lets the rest of the code be sane.

[-]

failsafe-author@reddit

I don’t think so either- lol. And it’s obviously contextual. Something being regarded as a “smell” means to think about it, but it’s not always wrong.

But in general, I avoid comments where possible, and take the impulse as a clue that I should write better code unless I can clearly articulate why I can’t describe the “why” in code. Usually I can.

Strong disagree. Every code base where I’ve seen people apply this is usually full of foot guns where people are afraid to change things.

Edge cases, usage examples, hints when it can break, external conditions, side effects.

And that's before we talk about obvious reasons to comment, which is autodocumentation via ajavadoc, docstribg, gdoc, etc.

No, good code doesn't "comment itself". That opinion was refuted decades ago.

[-]

pavlik_enemy@reddit

Not even statically typed languages, with a good naming you don’t need inline comments for most applications people write. “I’m iterating over the list of order lines and summing the prices”. Yeah, no shit

Frameworks and libraries need doctstring-type comments but not inline

[-]

nullpotato@reddit

A good comment would be why you need to iterate over the list for the sum instead of just accessing the current sum variable. If the code contains the full context then absolutely doesn't need a comment.

[-]

Wandering_Oblivious@reddit

yeah usually if I comment on code it's to let the dev know:

I know that this code is janky and offer some context for why we went with a given solution
Apologize to them for the gordian knot they are about to try and untangle

vs.

// Increment here instead of at top of the loop to avoid off-by-one-error.
// http://stackoverflow.com/some-relevant-issue
x++;

[-]

sciencewarrior@reddit

TODO: treat corner cases

last updated: 07-23-2019

[-]

It's deterministic in the sense of getting the same output every time

Which is the definition of the word. Complexity does not break determinism.

What do you mean by knowing the output in advance? I don’t think you can know the output before running the rust compiler either. For all you know, the compiler will run forever and never stop given an untested input. You can make a pretty damn good guess that it will terminate and output what you expect, but it’s not guaranteed until you step through the system.

I do get what you mean I think, but these “deterministic arguments” as a way of writing off ai are invalid. LLMs are simply more complex than a rust compiler, but they are both cut from the same cloth.

[-]

Ok_Individual_5050@reddit

No. They are absolutely not cut from the same cloth. And you *should* be able to understand at least roughly what the machine code your compiler produces is going to do. If I use u32 I know for a fact I'm going to get an unsigned 32 bit integer at the end. LLMs make no such guarantees and can never do so.

[-]

Jedkea@reddit

You can’t even know if the compiler will ever exit, let alone other more complicated claims. Do you disagree?

Both are deterministic state machines. Do you disagree there?

[-]

peripateticman2026@reddit

An LLM with a temperature of 0 is deterministic though?

That's the issue though, isn't it? The creators of the LLMs can set the temperature to 0 for some specific things they want to be deterministic (censorship terms etc.). However, if you set that across the board, then it effectively defeats the purpose of using an LLM to begin with, unless the model has exact data based on its training.

https://old.reddit.com/r/LocalLLaMA/comments/1j10d5g/can_you_eli5_why_a_temp_of_0_is_bad/ discusses some of these issues.

[-]

Jedkea@reddit

You set the temperature yourself in the api call, the provider doesn’t. I am a contributor to quite a few ai coding tools, and reproducibility is a well known trait. For coding, you are pretty much always using a temp of 0. I believe you might misunderstand how these things work if you think exact training data would be needed to get a “wanted” result with 0 randomness.

That link you shared is not scoped to coding. For something like creative writing or chats, you would indeed want some randomness.

[-]

DealDeveloper@reddit

LOL
How do you manage HUMAN developers who are also non-deterministic?

How come you don't realize that it is easier to manage AI than humans?

[-]

peripateticman2026@reddit

You must be joking, right?

Recently, for instance, we had a problem with using Claude to use some specific MCP servers to achieve a task - apart from the times that it failed because "the prompt was too big" (not the human prompt, but the internal state generated by the same human prompt - also a mark of non-determinism), it still managed to not use specific MCP servers because it chose not to, even after explicitly informing it to do, listing out even the command from said MCP servers. Non-deterministic.

The only way to know what it is doing is by logging verbosely, and inspecting it - at which point it becomes an overhead, not a valuable tool.

Furthermore, when using ReplIt, it would do the same instruction "do X so that Y" 10 different ways when prompted 10 different times (and with a scarily massive difference in the number of files and configuration it chose to create/delete/edit). Now imagine you have a couple of dozen such changes created and checked in by ReplIt, and you wish to go back to a specific point and redo some changes - there's no way to be sure of the result without looking at all the changes that it does along with knowing how it affects everything else in the codebase. This is not an issue with a human being, regardless of how bad the developer may be.

AI is here to stay - I am not a Luddite, but we are not there yet where we can have it replace humans completely, even as a junior intern.

[-]

DealDeveloper@reddit

No, I'm not joking.
You used Claude, Replit, and MCP and got bad results.
Try not using all that. Try going through your comment sentence by sentence and solving the problem step by step.

Would you be willing to bet money (using an escrow account) that a LLM (combined with some custom code) cannot replace humans completely . . . now?

How much?

[-]

peripateticman2026@reddit

Try not using all that. Try going through your comment sentence by sentence and solving the problem step by step.

This makes zero sense. So don't use LLMs the way their creators require them to be used? The whole discussion was about non-determinism, and it clearly demonstrates that.

If you're creating a personal CRUD app, that doesn't matter. For anything beyond that, it definitely does.

Would you be willing to bet money (using an escrow account) that a LLM (combined with some custom code) cannot replace humans completely . . . now?

Sure, but it's a meaningless exercise. What parameters? What domain? What skills? What budget? Projects range from a simple calculator to real-world applications.

People don't even care about quality anymore in software. They just want fast, get to market , let the users suffer... And then six months later issues start getting escalated and we get to put out the fire

Management screws us from both direction. They want speed developing and speed fixing. While speed was the cause of the issue in the first place

[-]

DealDeveloper@reddit

"They fail to understand AI is not deterministic like previous abstractions lmao."

If only there was a way to check LLM output with deterministic code.
They don't fail to understand . . . you forgot SAST-like checkers exist.

I feel the same with AI, its like its taking over a part of the job that is actually important

And I also question how actually more efficient the thing is to begin with, ask for a robot to write code, and you still need to check and fix and *understand* what it did because at the end of the day you are responsible for that piece of code, to me it seems like the same amount of work just with the "write the code part" a bit faster and a lot less reliable, and that is by far the smallest part of programming

[-]

drjeats@reddit

Yes it is fundamentally different, because you ultimately need to synthesize the information you get from web searches.

LLMs can make the random shit you copy paste from StackOverflow compile/run and sort of work without any critical thinking on your part.

Using it like a more targeted search/SO is a different thing from what is being discussed here.

[-]

Frogman_Adam@reddit

The way a lot of tech companies are overusing AI though, they’ll be better placed to take the jobs!

[-]

gomihako_@reddit

Bröther, I have seen the light. It came from...the pipes.

[-]

BillBillerson@reddit

What makes you a qualified plumber?

My company decided to run AI to train a Chatbot and automate the ordering process through it. It was such a complete disaster. The orders in out ERP were all manually created. 80-85% were fine. The rest was a mess with no clear structure. This was my main concern, that I voiced - our data was simply not good enough to train AI. No one was willing to clean up and improve the structure of the existing data.

Shit In Shit Out!

[-]

ocakodot@reddit

Could you elaborate what problems do you generally face? I would assume bad memory management both in back and front end. Ai is not mindful when it comes to create closures. It sometimes gives multiple similar states too.

[-]

gravity_kills_u@reddit

Cleaning up slip code from overseas already. AI code can join the party I guess.

[-]

giyutanzen@reddit

Hey man, I wonder how you have positioned yourself as a self employed. Am also a DevOps Engineer with 7 years of experience trying to achieve this. Can you give some roadmaps or guidance?

[-]

Typical_Newspaper408@reddit

I've done a couple of projects with AI that came out like a lion, but now I'm in "post-AI build dread". Impossible to maintain. Luckily I didn't bet the farm on them, and I needed to do it to learn what to do and what not to do.

Fact is, you have to review any code that AI generates for you if its not a total throwaway. Great for writing one off shell scripts. Complex, long-lived software. No, cause when you hit the wall, it hurts.

[-]

DoctorWaluigiTime@reddit

A similar rush occurred when offshoring got huge 20+ years ago.

Same song, different lyrics.

Consultants will never run out of work as long as a C-suite person has a gleam in their eye about the latest 'what if we get fast and cheap and just roll the dice on "quality"' trend.

[-]

ComposerTurbulent631@reddit

I keep saying it.. Pepperidge Farms remembers the gravy train engineers hopped on to clean up the hype behind Y2K.. this is just the next scheduled train..

https://futurism.com/companies-fixing-ai-replacement-mistakes

[-]

6stringNate@reddit

As an aside - how do you like being freelance? How does it compare in terms of earning and freedom that you get from a traditional corporate gig?

[-]

kcib@reddit (OP)

I started by actually working for startups and then I moved into freelance in 2022 so I had 2 or 3 projects to start with and then it went from there. I would say there’s pros and cons like everything.

Pros: - Work whenever you want - You decide what you get to work on (when you have enough deal flow that you’re not desperate) - Work anywhere you want

Cons: - You have to pay for your own health insurance - If you work hourly which is very common, if you dont work, you don’t get paid. I do miss PTO - Especially in the beginning, you need to have good networking skills and be able to form strong relationships with stakeholders

[-]

6stringNate@reddit

Thanks for the advice! How much vacation or breaks do you end up taking in a year? And how do you plan those out?

[-]

kcib@reddit (OP)

Honestly, I don’t really take much extended time off. I also still work a lot. In a way you can say I traded a 9-5 job for a 9-7 job with no benefits for the illusion of freedom

[-]

6stringNate@reddit

Haha makes sense - and important to keep in mind! Thanks for being honest

[-]

LuckyHedgehog@reddit

You also need to be prepared to justify hours billed. "Why did this 100 hour project take 120?!?"

Also, negotiating the structure of your contracts is something lots of devs have 0 experience doing and struggle with, and can get you in trouble if the company wants to try taking advantage of the wording of the contract.

[-]

mechkbfan@reddit

If a manager is surprised by hours, then that's a fault on our selves for not effectively communicating issues/scope creep.

Often people don't care if something is a bit late, they care more that they weren't told earlier and surprised by it.

I learnt that lesson once and now over communicate anything related to money to the person holding the bag as early as possible.

I know it's a random example, but if it's somewhat linear, then by the 50-60 hour mark it should be clear you're about a day behind and they should be told at that point.

[-]

LuckyHedgehog@reddit

If you're just working 40 a week then sure, that's fine and probably not an issue.

What I'm talking about is getting hired to complete a project for X hours. Or fixed bid projects that suddenly gain scope based on "implied" requirements from the contract even though it was never discussed during initial requirements gathering.

Building in language to your contracts to cover those scenarios takes experience, and having the ability to push back on the client when they ask for things outside of the contract while maintaining a good relationship with the client is another skill. There is more to being a freelance dev than coding that lots of people overlook

[-]

mechkbfan@reddit

Yeah fair enough. I've deliberately stayed away from fixed projects for that explicit reason.

Did a couple, one went well, other was shit, but overall not fun.

[-]

LuckyHedgehog@reddit

Or if there was scope creep that increased the budget, they should have been agreed upon before sending them the 120 hour bill.

There are companies that prey upon these types of contracts to intentionally try getting free work. It takes experience and skill to scope things out ahead of time and push back on the client when that happens. It may mean losing that client altogether and you need to be comfortable walking away from that job. But it is certainly something a lot of devs don't consider when thinking about freelance work; the actual software development is not the hardest part of the job in many cases

[-]

mechkbfan@reddit

Agreed on all points

Do we hit superintelligence?

No - Society collapses
Yes - Humans are extinct

[-]

chandra-mouli@reddit

From reading all the conversations I sense a clear conclusion being drawn, that AI can only do POC level coding and not production grade code. Which according to me is not true at all. I have been working in one of the fortune 10 company of the world and I have vibe coded things to a scalable and bug free production environment too.

My take is, if you were already doing production grade coding you will continue to do that with 20-30% more efficiency using vibe coding.

[-]

bhh32@reddit

Question: how do you know it’s production grade? You vibe coded it. You didn’t check the code to ensure it was accurate, efficient, secure, and designed/architectured well. There is no such thing as bug free code, only bugs yet to be found. If you were writing production level code before AI, you should know this and never claim that you vibe coded bug free software.

[-]

EvalCrux@reddit

You do check the code fully, test, and run it in multiple environments.

[-]

blinkdesign@reddit

The definition of vibe coding is not to check

[-]

chandra-mouli@reddit

I think that's the misunderstanding that we will start moving from and everyone will eventually conclude that design patterns and design diagrams are still an essential part of code.

I saw this 1 very senior principal architect vibe coding his own application every few prompts later he would do a prompt to create the ERD of the code base and see how the overall picture looks and change method placement or separate concerns as he sees fit from the ERD.

Vibe coding might be new but experience you have matters a lot which building a production grade application.

[-]

mikaball@reddit

There are already many tools to design and generate ERD that are more predictable than any AI. You have to come up with a better example. This is the kind of stuff I don't want to replace with AI.

I have written code generators in the past to perform such tasks. I would be more keen to use AI helping me write such tools than using AI to replace such tools.

[-]

EvalCrux@reddit

Absolutely incorrect.

[-]

blinkdesign@reddit

"A key part of the definition of vibe coding is that the user accepts code without full understanding.[1] Programmer Simon Willison said: "If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding in my book—that's using an LLM as a typing assistant."

Those who aren't using AI are going to be deemed slow.

The customers expect faster turnover of features now.

Then they started a cleanup where they just deleted all the customizations. This took about 5 years. After that they could start the migration. After 10 years they have migrated parts of it.

I think the real isssues will show up when companies want to merge, sell of a part or do this kind of deep restructurings but it will be 5-10 years from now.

[-]

Instigated-@reddit

Yes, those most championing ai focus on the speed something new can be developed. However in the long term the real difficulty is in how easily it can be maintained.

We’re at a point where devs are feeling pressure to use AI to deliver, however may not yet have learned how to use it effectively to maintain code quality (or when to use it and when not to use it).

Often people find it easier to write code than to read it. It’s one reason why people hate legacy code and want to rewrite into the way they prefer it.

If someone has been the one to write the code, it is easier to debug, or evaluate additions to the code. They know it more intimately. Or they can ask questions of another person who wrote the code.

However trying to make sense of buggy code written by AI will be more challenging.

What is your approach with these ai generated projects? Do you take them on or pass?

[-]

mpvanwinkle@reddit

I’m not so sure this is true. It’s not that I think you’re wrong, it’s that I think the real impact of AI might be to make software something that is disposable, much like electronics have become. Cost of maintenance will be high, but cost of replacement may be relatively cheap.

Not trying to defend AI but I just think it’s important to acknowledge how the economics can change the technology.

[-]

Pristine-Moose2337@reddit

I think your second paragraph is spot on, but I'm not totally sure that your premise of disposability applies to large scale software projects.

In many ways, that sort of software has more in common with industrial equipment like a large power plant than it does with consumer goods that have become relatively cheap and disposable, like a microwave.

There are certainly components in industrial systems that get swapped out rather than repaired, but companies don't just replace a whole turbine in a hydro power plant because it's due for maintenance.

I'm sure there will be some increase in replacement, but how many software systems are actually designed and maintained in a way where things are sufficiently decoupled that it's easy to swap out a database engine, rest api, or microservice? The projects where that's possible are typically well designed to begin with.

It'll definitely be interesting to see how this all plays out.

[-]

mpvanwinkle@reddit

Good points all, it’s definitely not apples to apples comparison. But just for the thought experiment, an example I was thinking of was a music app ( that I may or may not have worked on ) that has a microservice for “history”. Everytime a user plays a song it writes and whenever a user views their history it reads. Very simple. Now in the old world, this history service has to be maintained by a team, it has to be upgraded, and it needs to at least have a consistent enough software pattern that if a feature needs to be added that it is relatively easy for a dev to add the feature while ensuring they didn’t break the other features.

Now imagine the new world. The cost or writing such a simple api is trivial because AI. As long as we have an open api spec and a set of contract tests you can validate the api works as expected. Does the code quality matter in this case. If the product team wants to add a feature and the cost of rewriting the entire api from scratch according to the new spec and contract is trivial, is that a problem?? Does the code organization and maintainability matter, or is it ok as long as it does what it says it does.

I’m being a devils advocate a bit here I know, i personally hate what AI is doing to the day-to-day of my job, but I’m also trying to accept it and stay ahead of where the industry is going. “Disposable” software is something we have to at least consider.

[-]

Pristine-Moose2337@reddit

I think that depends on whether you also have automated integration tests or something to validate that letting the LLM rewrite the whole microservice didn't produce a bunch of new bugs.

I can totally see people trying it and there could be an argument made that a sufficiently well-trained model that produces incomprehensible code is another layer of abstraction similar to a compiler producing native machine code, or a transpiler producing minified, standards-compliant ecmascript.

The problem I'm still seeing is the "last mile" issue that others were talking about. If the output is either too much work to validate, or inconsistent from run to run, it could easily lead to people backing away from it before it becomes good enough to become the new standard. There's a lot of people throwing money at it in hopes that it will be able to do things exactly like you're positing though. We're living in interesting times.

[-]

Instigated-@reddit

The key issue is how it & reliability scales in production, and how easy it is to eliminate bugs and problems that arise in production.

At the moment: ai is not nearly as good as a skilled software engineer. People often advise to consider it a highly productive overconfident junior engineer. Creates a lot of mediocre quality code fast and doesn’t really know how to improve it beyond a level (and may break things trying).

This is fine for a MVP or early stage startup, but not for production systems that require high reliability and scale.

Some people are sure ai will quickly improve, as it has quickly improved over the last couple years, thus negating this issue.

However, how is ai trained to improve?

Initially it was trained on public codebases. Many of which are not production level code.

Next stage they hired a bunch of “AI trainers” to provide code solutions, which resulted in better model training. However this work is largely insecure, poorly paid, and not very rewarding - so it students, people in developing nations who may have little live production experience, and mediocre software engineers rather than best software engineers doing the training.

This “training” is primarily on small concrete examples, not highly complex production products.

So all these massive improvements to the ai models have been the lower hanging fruit.

I have not yet seen a plan revealed how they are going to close that last gap to get it as good as a highly skilled experienced software engineer. This “last mile” is by far the hardest to close.

So we will have ai work that speeds up the easiest 50-80% of work. However that last bit of work will be harder for a human to fix because they did not write the technical solution, cannot talk to anyone who wrote the technical solution, do not understand the decisions that led to that solution, etc.

The last 20 odd years have had significant discussion & approach about how to write “clean maintainable” code for a reason, and many are now throwing that out the window.

Exactly. Quickly pushing out more bad code is the problem, not the solution.

[-]

SignoreBanana@reddit

"How to use it effectively to maintain code quality"

I'm curious what you mean here. By my experience, the problem with AI isn't so much "code quality" as opposed to "how to take a big problem and find the most efficient solution through it."

It seems allergic to efficiency.

[-]

lokaaarrr@reddit

Unless a system is a failure or POC, the majority of the lifetime cost will be in maintenance. Design for maintenance.

[-]

asdis_rvk@reddit

Indeed reviewing code is often more difficult and less rewarding than writing it yourself. If you have to fix AI-generated code, this is a shitty job because the "creative" part was done by AI, while the tedious part is on you. It's the exact opposite of what we were promised.

You can go very fast with AI, but you accrue technical debt at a much higher speed too. Few people bother with technical debt but it's not an abstract concept.

[-]

TheWorkplaceGenie@reddit

Been feeling this shift too — suddenly we’re not just building things, we’re curating, debugging, and reverse-engineering piles of half-baked auto-generated junk.

Honestly, it's a strange time to be a dev. The velocity is up, but the quality? Not so much.

I’ve started leaning more into roles that emphasize architecture, integration, and human review — basically, acting as the adult in the AI room. Feels like that’s where the value is headed.

Anyone else shifting toward “AI janitor” mode lately?

[-]

Mobius00@reddit

Well at least we'll have work forever fixing the ai bugs. sounds awful though.

[-]

nivix_zixer@reddit

I told my friends, I'd love to start a company based on this idea. Track all the companies very vocal about fully switching to AI at this stage, then target them hard to fix their junk in 6 months.

Problem is I've never started a company before, so it feels doomed to fail. But someone could. Just a group of awesome devs who hunt down these opportunities like sharks.

[-]

enumora@reddit

One of my clients vibe-coded a fairly comprehensive data platform last year before the latest round of tools. They actually did a solid job with business logic, but the architecture and code were a mess and involved a lot of snippets from Claude and ChatGPT copy/pasted into the codebase.

I don't fault them for taking initiative, because doing so enabled them to increase their revenue substantially. They eventually realized it couldn't scale to meet their needs, so they opted to pay someone to manage it.

While the knee-jerk reaction for many is probably "ugh, but then you have to clean up a mess!", here was the reality:

They opted for full rewrite, so I was able to start greenfield just leveraging their business logic as an input to a system. They effectively did their own product scoping by prototyping.
They're already making lots of money with the thing they made, so they don't blink an eye at the cost. Typical seed-stage / Series A founder is doing everything they can to avoid paying cash, because they don't have any.

I think there will be a lot of these types of engagements, and it can actually be quite lucrative for anyone that doesn't mind getting their hands dirty.

What I find problematic is the people branding themselves as developers / agencies and just pushing out vibe-coded slop - not to be confused with code that's been written with assistance from AI by people that actually know what they're doing. That said, there have always been shitty developers and shitty agencies.

[-]

arihoenig@reddit

I don't think you'll see much of it frankly. The new stuff that is being generated is quite good; well structured and clean and (if you prompt for optimization) pretty performant.

The stuff you are seeing is just "too early" early adopters and that will be a transitory phase.

[-]

Western-Image7125@reddit

End of the day, the problem still as it always has been is developers who have no experience and no idea how to use the tools in front of them and now one of those tools is AI

[-]

Perfect-Campaign9551@reddit

Your are probably just vastly underestimating how many bad developers are out there

We (as software developers) don't have an organization or a union or anytime that declares quality.

This resonates hard. I'm seeing similar patterns incode that "works" but is unmaintainable and inefficient. The telltale signs are precisely what you described: inconsistent patterns, over-engineered simple tasks, and those AI-generated comments.

The scary part is that non-technical founders don't see the problem until it's expensive to fix. They received low-cost development services that will ultimately cost three times more to rebuild correctly.

successfullygiantsha@reddit

I was looking up a record for the worst MLB record on ChatGPT and it gave me clearly the wrong answer. Seems like something that simple should be an easy search but it got it wrong and went on and on about how it was right.

[-]

br0ast@reddit

My entire career has been slop cleanup for "10x developers", architects, seniors, juniors, and off shore resources. Now i do this with AI agents and it feels no different

[-]

smithereens_1993@reddit

If you’re crazy enough to enjoy this kind of work it’s a really good time to be in the biz.

I built my career on fixing people’s slop, so now I’ve just shifted to doing it with AI’s slop.

BigManWalter@reddit

Cursor adds your company code to the context but that's not the same as training on it.

Context is like short term memory, it kinda understands it but it forgets things easily.

Training is long term, and it's done in massive batches costing millions of dollars each time a training run happens.

[-]

MrCallicles@reddit

Yes, that's exactly what I'm saying.

Data is the goldmine, I think it's insane to believe that companies like OpenAI don't use all the data they can put their hands on to train models, legally or not.

[-]

United-Pollution-778@reddit

AI slop
Shitty developers
Cheap outsourcing
Bootcampers
fuckers over engineering all kinds of shit because there are geniuses and the rest of us are idiots
... There I fix it for you

[-]

GloomyShake3434@reddit

I left my current startup company when i saw that the AI slop was taking over the codebase. I relaised that we already had an unmaintained codebase with no documentation & with the director forcing the team to use cursor, this will definitely lead to a disaster.

[-]

Moving_Forward18@reddit

The cost to fix everything that LLMs are going to screw up will be very, very high. But when it kicks in, there will be a lot of good developers consulting at high rates.

[-]

Kiri11shepard@reddit

Two options: - AI keeps improving, and eventually it will be better even than experienced developers, just like chess and go. Then no jobs in Software. - AI still keeps improving, but doesn’t cross a “junior developer” level for a while — then even more jobs will be needed to clean up this mess.

[-]

holbanner@reddit

Well I'm part of the hit team in my company that take such clean up missions.

We used to do "audits" but now we can smell A.I bullshit just by the introduction talk from the managers/leqds

[-]

HoratioWobble@reddit

It's gone in waves like this for the last 20+ years, I started Freelancing back in 2008 and would almost exclusively pick up work on freelance sites from failed outsource teams.

Yet, almost 20 years later - we have continued growth in outsourced teams and my experience is still the same, they vastly miss the mark in terms of delivery and quality.

This isn't the big revolution you're hoping for, this is now business as usual and yes there will be failed AI projects that need cleaning up but the same people who employed them - will continue to employ the same techniques in other businesses.

[-]

pinkwar@reddit

The exact same thing happened when companies started outsourcing the work to cheap developers.

[-]

Vivid-Blacksmith-122@reddit

I guess we don't have to worry about Judgement Day just yet then.

[-]

pinkwar@reddit

We're only getting started.

Pretty much live a dev, usually your first iteration is not the final solution.

That's how cursor works with the "thinking". It goes back and refactors the code to follow the rules.

If you don't give it rules it will continue to produce slop like an unhinged junior dev.

Give it some time and these pipelines will all run behind the scenes.

[-]

lornemalw0@reddit

this - with the junior gap - is going to fuel the next upward cycle

[-]

CauliflowerIll1704@reddit

Ive seen some people at a small startup I worked at that I could tell were vibe engineers.

When you hopped in to call with them to consult about a problem, or to do some pair programming, you could see that they didn't understand the code they've been working on for years.. You'd see flows that really didn't make sense at all, comments explaining very simple code, etc..

Folks, don't skip out in the fundamentals and if you don't have a CS degree, please please at least take a community college course or something to learn the basics.

[-]

Middle_Ask_5716@reddit

Next years LinkedIn posts:

How to use AI to fix broken AI code

[-]

haikusbot@reddit

Next years LinkedIn posts:

How to use AI to fix

Broken AI code

- Middle_Ask_5716

^(I detect haikus. And sometimes, successfully.) ^Learn more about me.

[-]

Fidodo@reddit

This already happened with the first wave of outsourcing where companies tried to hire out to the cheapest possible overseas contracting firms with zero oversight or integration and were shocked when the deliverables they got were shit.

[-]

tjsr@reddit

I embarked on a little own project a few weeks ago to try making something using Copilot and prompts.

It's great for getting the very basics off the ground, just boilerplate and scaffholding that takes a lot of effort which you don't do frequently when you work on established projects, but wow once the app starts having interactions between various components does it really start crumbling.

Following the industrial revolution, there were plenty of people who saw the trajectory of development and predicated that very soon we'd all have robot butlers, flying cars, and the end of disease.Turns out technological progress isn't linear, though.

Sea-Frosting-50@reddit

where are you seeing /sourcing these projects?

[-]

Simple-Quarter-5477@reddit

How are you networking and marketing yourself to find startups to bring you on?

[-]

Beginning_Opinion377@reddit

I’m feeling this pain, it’s like the new wave of shitty WordPress sites we have to fix

[-]

itsbett@reddit

Honestly, this gives me a lot of hope for if I choose to switch industries and make better money. I spend a lot of time developing my skills and understanding mostly older languages, primarily as a hobby, but it was motivating to me that it also made me more employable. I was worrying that the perception of AI's usefulness might make my skills less valuable for my career.

Currently, I'm fairly underpaid, but immensely comfy at my job. AI is banned/illegal to use where I work cuz of proprietary languages, hardware, and software. My job also rides on politics, so I worry that I'll need to switch industries when my skills are perceived as AI replaceable.

So thanks for the good read and info sharing

[-]