I vibe coded a feed reading web app. It was enlightening and uncomfortable

Vibe coding works. I wish it didn't. But it does, well enough. And barring some revolution that overturns the new world disorder, machine learning cannot be undone. Earlier this year, I surrendered, bought a $20/month Claude subscription, and vibe coded a web app for monitoring news feeds. It's been an enlightening but …

[Headley_Grange]

When I'm stood in the checkout queue in the supermarket and a staff member comes over and says, 'Come and use the self-checkout - I can help you if you like", I reply along the lines of, "In a couple of years, after you've finished helping them to iron out the bugs and you're on the dole you'll wish there had been more people like me who refuse to use them."

[elsergiovolador]

Come on, you must admit that sitting by the till and mindlessly scanning products is beneath human intelligence. These jobs should be automated, so that human potential is freed to do something else.

Same with typing code.

[An_Old_Dog]

Over the Tannoy

"Buffer overflow cleanup on Aisle 13!"

[elsergiovolador]

Re: Over the Tannoy

"George got a blue screen!"

[Pete 2]

> sitting by the till and mindlessly scanning products

I understand your aspirations for the human race. However, we must never forget that there is a significant proportion of the public who do not have, and maybe do not wish to obtain, the skills to perform at a higher level. That many people like the certainty of knowing exactly what their working days will entail. Further , that in many close-knit areas, the social aspect of a steady stream of friends, family and even strangers to exchange a few words with while being paid, has some significant value.

In addition, it does not entail getting wet when it rains. It is not physically demanding (thus suiting individuals no longer in their prime) and that the shifts can be organised around other responsibilities, such as childcare.

There is a lot to be said for many such occupations, ones that do not require abstract thought or demands on physical or mental abilities.

[khjohansen]

> "not physically demanding"

A cashier may lift 1-1.5 tons of groceries during a shift ...

https://www.sciencedirect.com/science/article/abs/pii/S0169814117301117

[Jedit]

"A cashier may lift 1-1.5 tons of groceries during a shift"

Not all at once, though.

[Sherrie Ludwig]

Re: "A cashier may lift 1-1.5 tons of groceries during a shift"

"A cashier may lift 1-1.5 tons of groceries during a shift"

As one who has been employed in this role, "lifting" is generally, in US supermarkets at least, mostly sliding items, aided by a moving conveyor belt, over a scanner. Although my last stint at this was in my sixties, I found the job not very onerous and the human interaction pleasant. I was able to suggest add-on sales that increased my employer's profits, or helped other businesses in the area, increasing my town's economic health "have you tried the burgers at _____? They are the best in the area!" Yes, there may be ads onscreen at a self-checkout but who bothers to read them or trust them?

As to the hype of "lifting", let's say you weigh 150 pounds. Presumably, you lift that with every step, especially while jogging or running. My goodness, a pleasant toddle around the neighborhood has you lifting hundreds of tons! Are you Superman?

[Steve Crook]

UBI

Agreed, but we're moving from manned checkouts being necessary to being a social enterprise. Which may be no bad thing, but there's a cost that.

Pay people UBI, they can volunteer to man checkouts for a small top-up to UBI and just chat to the people who's stuff they scan.

Social contact, if you want to pay and go, use the automated system. If you want a chat for 5 or 10 minutes use the till...

[tiggity]

When I was a student, I did Summer factory jobs to earn cash (shows how long ago this was, when they were easy to get as plenty of factories & cover needed over peak holiday time)

A lot to say in favour of such manual jobs - not mentally demanding, so at the end of your shift you were not feeling mentally drained (as you can be with relentless high mental difficulty work), instead you still had lots of energy & enthusiasm.

If there was a well paid "mindless" job I would be happy to do it instead of IT! Would be better for my mental health & would be great to have more "mental energy" after a day at work

[HereIAmJH]

If there was a well paid "mindless" job I would be happy to do it instead of IT!

Why would any mindless job be well paid? Unless it requires physically exhausting work or is risky for your heath. The way the world works is that if you aren't working with your brain, you're working with your hands. Unless you're a criminal or a politician. (there's a joke there somewhere)

Having spent my career in IT, I believe IT has become the new factory job. Once upon a time there were busy times and slow times, depending on the type of business. So you'd get breaks to recover mentally. But something I noticed about Agile is the relentless pace. Things don't slow down during the holidays, the sprints keep coming and the metrics determine workload. There is no longer a break after a release. Even if management put you through a death march to reach a release date, the following day starts a new sprint.

I had an interview with a company several years ago. They told me they do weekly releases of their primary app and I would have to support them. If things went well, the release would roll out at 10pm. If it didn't go well, I'd be responsible for supporting the fix/rollback, however long it took. But they generously told me that if I had to work all night to get systems back online, I could come in late the following day. I declined to continue as a candidate.

For many IT jobs the best that you can hope for is the ability to forget about the job when you clock out at night. The pay/responsibility balance is off with too many expectations of being oncall and available 24x7.

[Anonymous Coward]

> There is a lot to be said for many such occupations

You're not wrong. I just wish to point out that you've just described the socialist industrial approach (e.g., in countries like Czechoslovakia, as it was back then, and in parts of the Russian SSR itself).

It is a perfectly legitimate approach, just not a terribly efficient one. I can see both sides.

[jfrankybe]

Expendable

[SarahC_]

Unemployment perhaps? Hm. There's that whole dynamic of "work available" vs "job positions go up over time".

Perhaps AI's the people's internal combustion engine horses faced?

In which case - we salute those who get automated - they're on the scrapheap ... for good.

[T. F. M. Reader]

sitting by the till and mindlessly scanning products is beneath human intelligence

Is standing by the automatic till and mindlessly scanning products for no pay somehow intellectually uplifting in your mind?

At least the cashier gets paid (ultimately, by me). I don't work in the bloody supermarket. I am really slow to operate the unfamiliar equipment. I don't have the muscle memory to press the right button or pixel on the screen without looking, so I do have to engage my mind. Scanning a couple of products in a big pharmacy if there is a line to the human-operated register and if the automatic device is both idle and actually works is barely tolerable. Going through automatic checkout with a full cart half of which needs to be weighed is not. And the only effect of reduction in the number of cashiers in an attempt to create long queues and thus increase utilization of automatic checkout will be me never coming to the supermarket again. The phenomenally inefficient experience at small shops or at a farmer's market is so much more pleasant.

The last point is a typical consequence of the biggest problem with the reductionist misunderstanding of a cashier's job as "mindlessly scanning products". The reductionist re-definition is tempting as it is the only part one can hope to automate. The management consultants who came up with the idea were paid a percentage of cost savings at the expense of fired cashiers. The management consultants are not responsible for the value lost because customers won't come anymore. Look up "doorman fallacy" if you feel like it.

I didn't steal all of the above, but credit is due to Rory Sutherland at least for the automatic till vs. farmer's market observation and doorman fallacy. YouTube is thoroughly enshittified (as are automatic tills), but you can still find him there.

[Anonymous Coward]

"And the only effect of reduction in the number of cashiers in an attempt to create long queues and thus increase utilization of automatic checkout will be me never coming to the supermarket again."

My local big Tesco (one of the 'Extra' ones) is notorious for that - and if you raise the issue with the customer service desk, the answer is always 'we are short of staff at the moment". Things never change, so clearly they are intentionally "short of staff". It's now reached the stage where at a moderately busy time they will typically have one or two staffed tills open. Only surprise is that they haven't removed more of the tills they never use and put more self-service machines in (there aren't enough of those to make up for the lack of open staffed tills, so long queues there too).

[skwdenyer]

My local supermarket (Booths, up North), has done away with all self-checkouts in favour of manned tills, after researching the point with their customer base.

Their signage also reads "10 items or fewer" which also gets my vote.

[Taliesinawen]

Sitting by the till and mindlessly scanning products

> Come on, you must admit that sitting by the till and mindlessly scanning products is beneath human intelligence.

Some of the people working the till have been there for years and do seem to lack intelligence. As in when you go off script they are rendered mute.

[sketharaman]

+100. I wanted to become a programmer 45 years ago because it was the cool thing to do. I took coding courses in BASIC, FORTRAN, PASCAL, COBOL, etc. as a part of my engineering degree three years later. I found coding to be a major PITA and hated it. Glad AI coding tools are now available to free human beings from the drudgery of coding and still be able to build and release apps.

[doublelayer]

I would probably find various types of architecture to be very boring. Yet another building with normal rooms and structures, why can't I do something fun and build some secret passages in this? That means I shouldn't be an architect. If I decided to be one anyway but use something I didn't understand and didn't bother to check on to skip the work part, then I'd still be to blame when my poorly architected building collapsed. Just because you don't like writing software won't absolve you when poorly architected software breaks your users' security. Perhaps we'll eventually invent something which can produce software of acceptable quality, including those parts you don't see up front. We're not there yet, no matter that people sell products as if we have while admitting in their terms and conditions that they haven't so you can't sue them. If your laziness and displeasure with a task makes you ignore this fact, you are to blame for any and all consequences.

[An_Old_Dog]

"Drudgery of Coding"

If you find coding to be drudgery, you shouldn't code (I enjoy coding).

And rather than using the unreliable pseudo-AI coding, companies which employ architects and engineers who need custom software, should hire programmers to do this for them.

It's called "support".

It's cheaper than having architects and engineers doing coding (slowly, and possibly-poorly) themselves, just as wise companies don't have highly-paid professionals cleaning their offices and restrooms, or executives typing their own letters and emails.

[Anonymous Coward]

I worked for Dell in the oughts... the factory in Limerick had a person that spent their *entire* shift manually peeling off a roll and affixing a Windows license (label) to each and every machine that rolled past her. Explains the drinking...

[Mimsey Borogove]

human potential is freed to do something else

What "something else" is a 42-year-old person with no training beyond high school, who enjoyed the checker job because it allowed them to interact with people, going to do?

[MonkeyCee]

Checkout operators

Tell me you've never worked in a supermarket without telling me.

There is almost always more work than can be done in a day at a supermarket, and being able to have more free bodies means you can hurry along a late delivery so you don't get lorries double parked, or discount stock rather than just bin it, empty the reverse vending machine, clean up the Jason Pollock of a customer toilet etc etc.

The big chain round here got rid of ciggies and lottery tickets so now you can have one staff member do the service counter and self checkouts instead of three staff, but the store is still run with the same number of staff.

Let me know when they automate the cleaning and shelf stacking. Even an automatic unloading truck would be great. But people are still waaaaaay cheaper than a robot, and can be used for a multitude of tasks.

Tech bros will make a bot to write poetry and draw pictures before they'll make a self cleaning toilet.

[vtcodger]

In sympathy

I'm sympathetic in that I avoided the self service machines at my local supermarket (a Hannaford store in NW Vermont) for a long time. But I finally decided to try the things. Y'know what, theirs actually work. A damn sight better than most web sites. There was a modest learning curve as it's not exactly intuitive how to handle produce (it turns out that if it isn't prepacked with a bar code, you type in either a product code if it has one or the name of the item e.g. apples then, if necessary, pick from amongst a list of different kinds). So I would suggest that sometime when you have just dropped by to pick up a couple of straightforward things, try the machines. If the experience is unsatisfactory, try again in a year or two. And if you just enjoy the fleeting human conduct, by all means continue to use the regular checkout unless perhaps on days when the line is too ghastly.

Do I always use the machines? No, if there's anything in my selection I have doubts about (e.g a pre-pay card for my hateful Android cell phone), I go through the regular checkout and ask if I could have used the self service machines. (Yes for the Tracfone card apparently).

[Doctor Syntax]

Re: In sympathy

My limit was a 3 item basket which gave me a reasonable chance of completing the checkout before anything goes wrong. I'm beginning to think enshittification is in progress there as the last attempt a few days ago failed on the 1st item.

[HereIAmJH]

Re: In sympathy

You're shopping at the wrong places. Aldi has the most efficient cashiers of any retail store. Yet one of their stores where I shop has gone fully self checkout, with 1 employee working nearby for the occasional "I'm not going to use self checkout" customer. The worst experience I've had is one time the self checkout register crashed on my final item and I had to move to another and start over. I prefer that store over any of the other Aldi stores nearby.

Walmart has also done very well with self checkout. I find that pretty much every place I use self checkout, my time standing in line drops dramatically.

[An_Old_Dog]

Self-Checkout Stations

... my time standing in line drops dramatically.

No, no, no no ... at least not where I live. Major retailers with self-checkout machines use them as an excuse ro reduce the number of cashiers, so that the self-checkout lines are just as long as the (too-long) ordinary, human-staffed lines.

[HereIAmJH]

Re: Self-Checkout Stations

I can only go from personal experience. The Aldi I mentioned used to have 1 cashier and would call up a second when the line got too long. On average there would have been 4 people ahead of you in line prior to self checkout. Now they have 7 self checkout lanes and 1 cashier if needed. Rarely can I not walk straight to an open register.

And yes, all retailers use self checkout as a reason to reduce cashiers. But it appears to be at a rate of 4 to 1. The places where I see lines at self checkouts are places where there were considerable lines at cashiers before. Walmart has been experimenting with various configurations. At one time they had a 'bullpen' for self checkout with a dozen registers. But that would have a line of 6 or more people that moved very fast. But the line didn't look good. Then they started replacing regular registers with 2 self checkout. That breaks up lines and make it look less busy, but is less efficient. Sam's Club has been experimenting with using their app to scan and pay, so no register at all. But they have VERY effective surveillance at the exit. If you put several of your items in your cart with the barcode facing up, and hold your receipt with it's barcode showing, their cameras will verify you at the door and the employee just waves you through. (otherwise policy is that they scan your receipt and two items in your cart)

[mark l 2]

Re: In sympathy

if it doesn't have barcode then lots of people put it through as loose onions or carrots. I remember hearing an anecdote that a big UK supermarket chain had 'sold' more onions and carrots through the self check out that they had purchased from their suppliers.

[Neil Barnes]

Re: In sympathy

Goes to Lidl. Picks up ten bars of chocolate (because what's the point of buying one?).

Go through robotill.

for (x = 0; x < 10; x++) {beep}

Press done pay now button.

Machine calls an assistant because of this suspicious purchase. Who is busy on the real till...

[tiggity]

Re: In sympathy

I periodically use Lidl for "bulk" shop. I always use the staffed tills - if a queue builds up they put more staff on, scanning always quick & efficient (faster than I can load into my trolley!) - I use a big trolley & auto checkouts are only for small shops - basket amount) shop so never tried the auto checkouts (probably just as well as if they did screw up badly then a high likelihood of me having a meltdown & smashing a tin of chickpeas (or whatever) through the screen - so dealing with people a far safer option for me!)

[stiine]

What ever happened to the future where we pick items off of shelf, stick them in our pockets, and simply walk out through the RF scanner without stopping to effect payment? That VISA ad was 18 years ago.

[vtcodger]

What ever happened to the future where we pick items off of shelf ...

I think Amazon actually does something along that line in its brick and mortar stores. No idea how, or how well, it works.

[David 132]

Re: What ever happened to the future where we pick items off of shelf ...

>What ever happened to the future where we pick items off of shelf, stick them in our pockets, and simply walk out through the RF scanner without stopping to effect payment?

It's in place in stores in many British cities right now.

Well, apart from the "effecting payment" bit...

https://www.telegraph.co.uk/news/2026/04/12/stores-failing-help-police-tackle-shoplifting-met-chief/

Icon? What else...

[Anonymous Coward]

Re: What ever happened to the future where we pick items off of shelf ...

Super! That's my Jane's Addiction right there -- love it! ;)

[iron]

Re: What ever happened to the future where we pick items off of shelf ...

Amazon closed them when the fancy AI tech turned out to be Actually Indians watching lots of cameras.

[Mimsey Borogove]

Re: What ever happened to the future where we pick items off of shelf ...

No idea how, or how well, it works.

There was one in our area for about a year or so, but it went out.

[Pete 2]

> where we pick items off of shelf, stick them in our pockets, and simply walk out

You are referring to Amazon Go stores. Well, they went! Closed down in January of this year.

[jdiebdhidbsusbvwbsidnsoskebid]

I saw that at a bar at a popular sporting venue recently. You scan your payment card at the entry turnstile, walk in and take whatever drinks you want then leave. I presume there are cameras watching you to see what you pick up and take.

I didn't use it because I asked the attendant if I could get a certain drink there, and they didn't have it.

[skwdenyer]

Nobody ever cracked the high bandwidth, long-range, low cost RFID problem. There was a UK company called, from memory, Flying Null, who introduced the (IIRC) Holotag about 25 years ago, which used a fundamentally different technology and could have potentially scaled whilst being cheap enough. Their tech could have potentially allowed you to push your trolley through a "scanning arch."

Like so many UK companies, they were too early, and insufficiently funded, so they closed.

Data science has given us the handheld self-scanner, together with enough prediction modelling to have a fair stab at when to call for a human re-check. That's good enough for a lot of customers. Amazon tried using cameras to detect purchase intent, but ended up finding it hard to accurately track when a customer picked up an item to read the label before replacing it. The same tech has been tried in hotel mini bars and vending units, with similarly-mixed results.

Barcodes have a marginal cost near zero. So far the economics haven't supported adding marginal cost onto products with smarter tagging, rather than just covering the cost of scanning those barcodes at the checkout, and nobody has provided enough evidence to show an upsurge in sales and profits associated with replacing that last line of purchase friction.

[MonkeyJuice]

Obvious troll is obvious.

[Throg]

I have no problem with people vibe coding their own little scripts for personal (and probably infrequent) use. After all, this clearly falls under FAFO. Sometimes you’re lucky, sometimes you’re not but there’s only one party that suffers the consequences.

Anyone who believes for one moment that there isn’t a massive gulf between this and the standards required of professional software development is, sadly, probably at the wrong end of the Dunning-Kruger graph.

Unfortunately however, some of those very same people are in positions of authority in IT, and have been for maybe two decades, which is how the cheap outsourcing boom got started.

Sometimes I wonder if they’re actually right. How far can we lower the bar on quality before customers actually walk away rather than just grumble and shrug? It seems we’re still running that experiment.

[An_Old_Dog]

Lowering the Bar

... happens in part because consumers choose front-end lower costs, uncaring and/or ignorant of the later, higher, back-end costs.

Example: you live in a smallish town. You have Jerry's Market for food. You have the 88¢ Store for sundries. You have Tom's Hardware for hardware. You have The Farmer's Market a couple miles out of town as a quasi-convenience store. It's not a convenience store using the 7-11/Plaid Pantry/Circle K model. The Farmer's Market has has a medium/small stock of a wide variety of things. Not only does it carry canned food, bread, milk, and meat, it carries rat poison, rat traps, leather gloves, boots, shovels and spades, ice cream, electrical sockets, lightswitches, cover plates, wire nuts, electrical tape, trowels, buckets, seeds, needles and thread, kerosene, kerosene lamps, and has a mobile slaughtering service truck.

At these stores, the clerks know their customers, are froendly, and the checkout line is usually just 0~3 people long.

Mega-Mart comes to town. "One-stop shopping", and liwer prices. People flock to Mega-Mart. After a year or two, Jerry's, Tom's, and The 88¢ Store all are out of business. The Farmer's Market's inventory variety is sharply-reduced to follow the 7-11 model: they now carry only "most-frequently-sold" items.

Mega-Mart starts raising prices, and cutting costs. They have 12 checkout stations, but only three are open. Typical lines are 8~15 people long. Any money you saved via Mega-Mart's initially-lower prices is far outstripped by the money you lose standing in line, and by Mega-Mart's now-higher prices.

Customers cannot walk away from Mega-Mart, because there is no alternative left.

[Filippo]

Re: Lowering the Bar

That's all true, but I wouldn't be so quick to blame consumers. Mega-Mart is a collective entity, while consumers are individuals. Consumers rarely organize, and never as well as a corporation. Corporations are also smart enough to put a strong emphasis on individuality externally... while being extremely collectivist internally. It's why Big Oil came up with the idea of the personal carbon footprint. There's nothing like pushing individual solutions to systemic problems to make damn sure that those problems will never get solved. Megacorp problems can only be fixed by strongly enforced regulation.

[Andy 73]

Re: Lowering the Bar

The problem is that people want someone to blame either way - they anthropomorphise the corporation as maliciously setting out to destroy local shopping, when it's an "unintended consequence" of growing a business with a narrow definition of efficiency and value.

So it turns out both consumers and corporations do things that are against our collective best interests. Both go out of their way to justify those actions - big oil by running campaigns, consumers by telling you how they don't have time to shop around and make those ethical decisions that are more expensive.

But it turns out most corporations have a limited shelf life, and people's values shift. Cancel the dozen streaming subscriptions you have, stop buying everything from Amazon, and reject Temu tat - those are all Boomer obsessions that make us miserable. Recognising that is what changes both consumers and corporation behavior. Maybe not overnight, but eventually..

[algol60forever]

Re: Lowering the Bar

Are you sure you're not Bernie Sanders?

[Richard 12]

The result of the experiment so far has been legislation that tries to ban the most egregious.

It's not worked well, because legislation is very much a Big Hammer and enforcement suffers from politicians and brown envelopes.

However, there's already a few cases where strict liability lands personally on those managers. Over time the legislative frameworks will become stricter, and the expensive lawyer escape routes will close (if only because said lawyers also use software tools).

[Dimmer]

What has changed.

There was a time when I would never consider coding a page to do something. That has changed for me.

An example would be a page that has inputs to a db, and outputs data in a required format, say a customer manager.

With AI, i tell it the db info ( and yes i do not use this app externally or give creds) ; how I would like it to look, what columns it has and what is the inputs and what I want to output. - It gives me a page and the sql to build the table in the db, I check it, tell it what to fix - dot this loop a few times and within half an hour I have what I need.

My info is not in the cloud, can’t be held hostage, does only what I want and the page doesn’t change unless I want it to. This process is just the new way to configure software for me. Only what I want, where I want it and most importantly IT DOES NO CHANGE.

To support it or add functionality, someone can just upload the current page and the Ai will change it.

Do you use this method for an astronaut to check his email? Heck no. That is what real programmers are for.

[b0llchit]

Ad as article

So you wrotedidn't write an app, have no real clue about the code quality and now you want people to pay for it.

As they say,... Good luck with that.

[Paul Crawford]

Re: Ad as article

have no real clue about the code quality and now you want people to pay for it

Well that has worked out rather well for Microsoft, hasn't it?

[QET]

Re: Ad as article

Only because of market inertia due to existing user-base, and brute force & ignorance.

Especially the very last bit, oh so much of that...

[Anonymous Coward]

Re: Ad as article

I don't think it's fair to say that there's no clue about Microslop's code quality! Given the number of problems, and the growing number of out-of-band updates already this year, surely even Microslop has a fairly good idea about the standard of their code...

[Peter Prof Fox]

20% / 80%

As you all know, the first 80% of a project takes 80% of the time and the other 20% takes another 80% time. What I think is happening is that the 80% bit will now only use 20% of the time.

But which part is which?

Are the AI assistants good at discussing architecture and high-level design options? Can they juggle fringe and far away scope beyong 'single app'? Suppose I want to build a library: Getting to 'hey it works' is the the start, not the end point. The eventual goal is 'Peter's library is solid and available' in flavours to suit developers. I need to bear tle bigger picture in mind at the start for eg. modularity or internationalisation.

[PghMike]

Claude Code

I dunno. I bought a one month subscription to play around with it, and I'd say:

1 -- it's actually pretty good at finding moderately subtle bugs. I pointed out where I got a fault in memcpy when I pressed a particular button, and it found the (very customized) code that ran when that button was pressed, found that as part of the work it deleted an object that had a helper thread, and the object got deleted before waiting for the helper to exit.

2 -- it is frustrating to use it for writing code. I asked it to add a simple feature to some existing code and after each set of prompts, I had to spend a lot of time proofreading its results, and updating the prompt to explain what I needed it to differently. After four rounds of this, I'm pretty sure it took more time than if I had done it myself. It certainly was less fun to proofread four tries by Claude than doing it myself.

3 -- Given item #2, I don't see how these companies will survive financially. I certainly don't plan to renew my Claude subscription. I mean, obviously Google and MSFT can burn money until the return of the proverbial cows, but OpenAI? Anthropic? These guys are going to need real revenue.

[Anonymous Coward]

Re: Claude Code

Good points ... I was hoping to see some of them spelled out in TFA too but maybe I missed them ...

WRT #1, it seems to me an LLM trained on all of human literature in world history would necessarily get lost when attempting this, so tools like Claude Code and whatnot have to be based on some specialized sub-corpus, and also rely heavily on highly deterministic (and logical) tools to do their job right imho, not lossy databases with stochastic recall. Not sure the LLM serves much purpose there, other than confuse issues and users (as far as the nitty-gritty of coding is concerned).

The linked Jim Nielsen (under "his account") is neat too, esp. where he goes: "You know what else is easy? Fast food. But I don’t want that all the time." -- good advice to heed!

[stiine]

Re: Claude Code

How could Claude code a use-after-free bug?

[DS999]

You said "earlier this year" you did it

I wonder if you still can. Because it seems Claude's abilities have recently been downgraded in the free and lower tier pricing plans as they are looking at an IPO and want to be able to show good revenue rather than endless cash burn. Anecdotally people say using workflows they could easily do a couple months ago in the $20 plan now burns through that monthly allocation in less than a day.

That will put a bit of a crimp in the hobbyist type of vibe coding, and maybe even those who have their employer pay for it will see them putting some constraints if they see monthly charges skyrocketing.

[HereIAmJH]

Re: You said "earlier this year" you did it

That will put a bit of a crimp in the hobbyist type of vibe coding, and maybe even those who have their employer pay for it will see them putting some constraints if they see monthly charges skyrocketing.

More and more hobbyists and small businesses will be hosting their own AI instead of paying. The difference is mainly just the amount of time spent. It's kind of like racing cars, the only limit to speed is how much you can spend.

If your alternative is letting the process run over night or not doing it at all, then you let it run. 3d printing is similar. Most of the things I print take hours. I start the printer and do something else. I let it run over night. The hobby printers have matured enough that I rarely end up with spaghetti. If I'm going to have problem with the print staying stuck to the bed it happens at the beginning. I could just call Amazon and have it delivered tomorrow, hassle free for a few more $$. I can pay a service to print my custom objects. I don't think AI will be any different, because the software isn't difficult to set up. And the requirements aren't much different than a gaming PC.

The concern will be if they start charging for models or limiting their abilities. When Hugging Face adds a cart for downloads, then start to worry.

[ErikOnTech]

There's a difference between personal hobby code and enterprise production code

If your personal hobby code breaks, nobody but you cares. Nobody's losing money or customers on it.

Nobody is going to try to exploit your personal hobby code, except indirectly though supply chain attacks.

You have no financial, criminal, or reputational liability if your personal hobby code breaks.

And so on.

[Michael Hoffmann]

You wins some you lose some

Having recently gotten myself a Claude sub as well, I thought I'd give Claude Code a serious try.

First, a little project I hacked in years ago to extract info from Steam. It worked well, until one library was no longer supported, attempts to fix it failed in the horror that was the original source, Steam change its API and I left it with a shrug of regret. Enter Claude, which "figured out" a work-around and got it to work again (because it was able to scour the insane amount of HTML and JS and embedded div blocks and all the stuff I had no clue about, because I never got into the web dev and though HTML should have stopped at b1 and a tags)

It worked so well I had it integrate a number of external scrips I used to mangle the output, and then a few nice-to-haves. Wow! Great!

Emboldened, I thought I'd go for something more involved: and actual RAG and document analysis system for our ancient archive of statements, bills, what-not. I didn't want that out there, so use Claude to build the code for the local running of one of the freely available LLMs. Might make my old GPU melt, but let's see what happens.

Oh dear. Turns out that just reading in and scanning/OCR-ing the thousands of docs was the easy bit - there's a great CUDA enhanced tool for that. OK, shove that into a RAG and let's run queries over it, just like they advertise is sooooo simples.

The results were laughable. And as of writing this, still not working, after days and days of "vibe coding". Code that worked suddenly regresses. Queries returning information that isn't even close to correct - despite it "telling me" that it found and is accessing the right information. It's just doing LLM things which is random prediction and which is about as likely to be correct as me picking the Powerball numbers. It forgets instructions and prompts, so I have to constantly inject them again, because it doesn't have enough "memory". Ingestion of data can take an entire day with the hardware I have available - and it will happily make some minor change of a few lines of code and have me do the whole thing again, no matter what I prompt it. (and gods I want to murder a techbro when I get that "you're absolutely right, I'm sorry" nonsense. You're a fucking machine! Not Marvin the paranoid android!

So, small chicken feed? Yeah, not bad. Anything serious? Tell em, they're dreamin! Not without an army of elves to hand-massage the output. Oh wait, you just fired all your elves, did you? Good luck, then! And even more good luck to the customers who paid money for that vibe coded swamp.

[MonkeyJuice]

Re: You wins some you lose some

This is forever the problem. LLMs can only look at your code through a toilet roll, and rely entirely on local tactical decisions. Once your codebase no longer fits on a napkin, correctness exponentially deteriorates, and token costs shoot through the roof.

Once you hit this problem, the sheer effort involved trying to trick the thing into focusing on the problem becomes silly. It is false laziness. Unless something dramatically changes, I can't see it moving beyond this point.

For throwaway scripts, "plot this", "perform this transformation on this data", "write a quick and dirty rss/atom aggregator" it can be useful- but if it doesn't get it right first time, you must immediately stop and take over yourself, otherwise you're effectively glued to a roulette wheel that is shotgun debugging your entire codebase.

RAG is perhaps the biggest disappointment, because having a reliable data librarian tool could genuinely be useful for most industries. TreeRAG- doesn't work, GraphRAG, also doesn't work. Every quarter there's a new proposal that is 'definitely' going to fix it this time around. But this technology is purely a research curiosity at the moment, and people just don't seem to grasp how long it can take before a new idea that makes realistic headway.

Something fundamental is missing, and 'throwing more data at it', and 'make model bigger' is not cutting it.

[monty75]

Great. Come back to us when you've created something that doesn't exist in thousands of examples in the training set and you've paid the actual (not VC-subsidised) cost of the compute.

[rgjnk]

Works until it doesn't.

I've used these things a lot. Even have private instances - amazing the things enough corporate money will fund.

At their best they're really truly magical. And at other times they really really aren't. It's a bit pot luck, you get tempted in by the good results & then tolerate the bad ones. Often it's not even vaguely obvious why you get different quality of result from one prompt/input to another, especially the identical ones!

And the really hard part is knowing whether the output is good or not if you aren't looking or don't understand what it produced. The magic blenders can easily get the wrong idea, forget something important from earlier or get confused between similar yet different things; they can be deeply knowledgeable about the subtleties of compatibility breaks between versions yet produce an output that utterly ignores that it's missing vital changes or including obsolete stuff unless you provide very narrowly targeted prompts.

They're a genuinely useful tool but the whole vibe coding thing is still a bad idea. Like using an intern I'll happily offload work but I'd be stupid to blindly trust it and I should only hand it over when I already have a good idea of what the right result looks like.

The danger is people pumping out mostly works 'good enough' stuff that really isn't good enough. And they won't even know it.

[SirWired 1]

They aren't great with the middle ground.

I'm in tech pre-sales, and I love vibe-coding. I'm not a coder, and never have been; I went into IT architecture a couple decades ago because I'm terrible at writing programs. Really awful at it; my college programming projects were held together by duct tape, band-aids, and stubbornness.

It turns out that selling customers on my brilliant infrastructure architecture is a lot more compelling when there's a mostly-working demo attached to it, and not just a PowerPoint slide with boxes and arrows. I can pound out something that mostly works in a few hours, and an app that will stand up to a loosely-scripted demo without falling over in two or three days. All without crafting a single line of code myself, or even having the least clue how web frameworks even function.

But I'm not fooling myself as to the quality. I know it's crap. I know that it's not good for anything but a rapid prototype. And nobody should ever attempt to bludgeon any of it into production, ever. I'm sure it has more holes than cheese, and would cause a seasoned coder to claw their eyes out after seeing how over-written and poor quality everything is. It's beyond salvage. Making it commercial grade would be like trying to turn a large half-rotted dinghy into a racing yacht; you ain't gonna Ship of Theseus your way into success there.

And then there's the other end of the spectrum: People using them for real work, with intention. Treating the AI like a green-as-grass junior programmer, who needs strict supervision and review, but works faster than the most-seasoned developer could ever dream. You can, and teams are, producing real, production-grade work this way, by understanding the limitations and making plans to accommodate them. Giving AI tasks that are strictly bounded with awareness of their capabilities. That works too. It can produce code that is reasonably lacking in bugs, is well documented, and is ready for tests that you should run with the same rigor as any other patch or feature.

Then there's the middle. People that think if you iterate prompts enough, discarding anything even vaguely resembling proper Software Engineering, you can somehow smelt lead into, well, not gold, but at least brass, by heating it with enough GPU tokens. AI remains a tool, not magic. It enables you to work faster, but it doesn't let you skip steps when you have real work to do. You still need requirements, a design, thorough testing, code reviews, security checks, etc. Skipping requirements has the same catastrophic results it always has. Skipping reviews and tests makes for the same rotten garbage that falls over in a stiff breeze. Pretending that the AI is some magic code-spewing genie's lamp, instead of another tool in the toolbox, will lead to identically-awful shovelware. The only difference is now it's way cheaper and faster to produce it, in ridiculous quantity.

[Pascal Monett]

This article has contradictions and false truths that are painful

"Stopping AI in its present form begins at the ballot box"

No. Even in the USA, it begins with Congress doing its fucking job. Good luck on that happening in the US or elsewhere, given the real fuckers that so-called lawmakers have become these days in most "democratic" countries.

"people who are not professional writers may be thrilled to have a tool that does something they don't enjoy"

If they don't enjoy writing, they have nothing to do with writing with AI and should go fishing (or do anything else). Do what the politicians do and hire a ghostwriter to do the job, or become an "influencer" on some social media and slack your thirst for fame there.

"Vibe coding hasn't worked out for everyone"

Yeah, especially not for actual, professional programmers. That should be a clue for you.

"The commodification of basic app creation has been underway for years"

Yes. It's called GitHub. You still have to have a clue about what you're doing.

* * * * * * * *

Programming is not producing a video. Just ten minutes ago a friend of mine passed by my office and showed me his new website he made for promoting his songs. He did not Vibe-code it, but he did use AI to make the videos. He composes the songs and performs them in what seems to me a be real music studio - insofar as I've never been in a "real" music studio. His videos are fun to watch, and AI is making great strides in producing that kind of stuff.

If this Vibe-coding fad endures, I'm sure some non-proficient people might be able to learn from it. But, to analyze code and even more, to be able to debug, requires intelligence, not statistical analysis, and that is something that no AI will ever possess for the foreseeable future.

[Michael H.F. Wilkinson]

"One of the knocks against relying on AI is that you don't learn anything. "

This is one I tell my (computer science) students. Yes, you can hand in a vibe-coded response to an assignment. Yes, it will probably pass all automated code checks. No, you will not have learnt anything about codig, unless you, very carefully check and study that code yourself. You will have practiced feeding the right kind of prompts to an AI system, but not how to code yourself. After all, the aim of the exercises is not to deliver a piece of code, the aim is to learn how to solve (initially simple) programming problems, so you can tackle seriously difficult ones later.

At the exam, however, you will not be able to access these tools, and you will fail miserably, unless you have acquired real programming skills in some other way.

[wolfetone]

Do you know what AI can't do?

Farm Goats.

So that is where my career is going, fuck AI, fuck the tossers who use it to put devs out of a job. I'm going to look after the goats.

[Anonymous Coward]

A wolf (even an etone one) farming goats ... brilliant! Me? I'm gonna see how I can use that newly reported first map of the clitoris ever (March 20, 2026), and see if I can be of use there ... more fun than AI! ;)

[keithzg]

Yeah sure

Lacking details in the article, I find it hard to accurately judge whether this is actually less effort or more featureful or more stable than existing libre solutions like tt-rss or Nextcloud News, but I know on which side I'd place my bet (or my time sysadminning).

[Rob 63]

impressed

wow you vibe coded an app, I got into a car and drove it straight into the garage wall. are you impressed yet ?

[tiggity]

I would expect "AI" to be OK at RSS / Atom feeds as there is lots of specification detail online.

They probably even have ingested code that does feed reading (depending on your language of interest).

A straightforward, well defined task.

I would expect "first draft" to be vaguely workable - & best option would then be to hand tweak it after that, "AI" output always seems to get worse the more you ask it to tweak things (breaking stuff that previously worked is tediously common)

[martinusher]

But this isn't all applications

Coding a news feed web application is an example of form filling -- its complex form filling but essentially you're customizing existing applications and frameworks to do a particular job. This isn't anything like the programming I spent much of my working life on. Its tedious and complex (but that's really more a problem with the environment and tools that in an attempt to systematize the development process introduce their own layers, methods and myths). Its a bit like discovering BASIC in the old days and realizing you could use it to make programs that could do useful looking things with -- until you ran off the limits of the environment, that is.

The real fun starts when you're working outside this type of box. Its typically where a lot of programmers fall flat on their faces -- they can produce perfectly functional code, it just doesn't actually work (and then there's the whole issue of error conditions, component or environment failures and so on). Understanding and marshaling everything needed for even a modest real time system isn't beyond the capabilities of AI, its just that it won't have a pool of existing experience to draw on.

[Anonymous Coward]

Thanks for sharing

Your experience.

It mirrors my own and that of colleagues and friends in various positions.

A bad craftsman blames his tools.

1. Know your tools (Karpaty has a number of very informative videos on the subject, for both general and technical audiences) and know got to use them.

2. Know how to identify and describe requirements in a way that is a) actionable and b) measurable. I've found code quality to be more related to how well requirements are specified than to who (or nowadays, what) wrote the code.