It’s going to get even more dangerous over time because the LLMs are coming out of the uncanny valley but still have subtle problems, they are just getting harder to see. I just did a bunch of AI training at my company and on the one hand, some of it was already out of date, and on the other hand, the language used to describe it gave it a lot more credit than it deserved.
People are already thinking it can do things it really can’t. Like think or analyze.
I’ve been in this cycle since the first time I’ve interacted with an LLM or AI coding system where at first it looks impressive and I’m not sure what its limits are and then I slam into a wall that makes me realize in horror that it’s capabilities are far less than it seemed at first, then improvements come out and I’ll repeat the whole process because the previous wall seems to be dealt with and it becomes hard to argue with the people who are gung ho for AI.
Fun story from this week, we had a chore for the frontend to refresh to a new version of the UI framework. Fairly simple task, so off to a junior developer. Within a couple hours there was a merge request ready to go. Ok, a fairly normal amount of time to change version and at least do a sniff test and find nothing changed so I go in assuming I’ll look at a few version bumps, maybe one or two tweaks… I see the junior dev was proposing over 1,000 lines of code to be added… WTF…
I crack it open and there was just a firehose of css rules, all marked ‘!important’. Looking at one examlpe, it repeated the same classifier with the same exact bunch of rules 5 times in a row. It was like it found every possible derived css class combination with tag and defined !important CSS for most everything about it.
So I find out that the junior dev asked it to rebase and it did what he expected, just change some version and went. He tried it and due to a framework change, one element was misaligned by a little bit. So he gave the feedback to the LLM and tried again… and it failed, and he tried again and it failed and after 5 rounds, it finally got the element aligned and hit ‘merge request’. For fun I opened up his proposed change and just so much was just a bit dodgy css wise because it screwed with so much stuff, but the junior dev only concerned himself with the page as it opened.
So I said screw it, I’ll do it myself, and added the singular rule that was needed to adapt to the framework change, making it overall about a 5 line change including versioning and such.
Depressingly, I suspect an executive would consider me far less productive because I only did 5 lines of change and the junior dev would have done thousands…
Depressingly, I suspect an executive would consider me far less productive because I only did 5 lines of change and the junior dev would have done thousands…
Probably the very same execs that use phrases like “Do more with less!” and then completely miss the point when true efficiency stares them in the face.
Yup. AI might, and I mean might, be a force multiplier for senior developers but this whole “Everyone can just let AI code” is bullshit that will lead to a giant mess of unmanageable code when your developers your maturing don’t understand the underlying code or good software architecture.
Oh and also I’m right there with you with the fuck that whole ELOC bullshit as a metric.
In my experience, the bigger the codebase gets, the more confounded LLM gets at trying to make coherent changes. So LLM projects start on shaky ground and just get worse because they can’t maintain the stuff they themselves generated.
I’ve seen what LLM can do and it is certainly interesting and can do some stuff, but the vast majority of my experience is someone who had not coded before “vibing” themselves into a corner and demanding help to dig them out. A bit irritating because while before we could reasonably prioritize requests to do stuff because management understood making something from nothing was real work, now management says “they aren’t asking you to make something, just help them fix something that already exists, should be easy!”
On the ELOC metric, for a long time I pointed out how disastrous I must be because my contribution to a project I was on was about -10,000 lines of code by the time I went to something else.
That last sentence hits me where it hurts.
Depressingly, I suspect an executive would consider me far less productive because I only did 5 lines of change and the junior dev would have done thousands…
The first question I ask about any analytics requests is what you’re trying to do with the results, what business question you want to answer. The second is how the analytics question relates to that business question.
It’s very easy to ask a question, look for a way to measure the answer, find something you can measure and start looking for the best way to ask for that measure. It implicitly assumes that “I need to measure an answer -> I can measure this -> this is the answer”, but as you point out, this isn’t a valid implication.
My worst enemy is the sentiment “If you can’t measure it, you can’t manage it.” Curse every MBA that repeats it like a mantra, in the name of the Stocks and the Shareholder Value and the Holy KPI.
Yes, some measures can be valuable indicators, if contextualised correctly, but not everything that has to be managed can be measured effectively. To grasp for measures anyway twists your sight away from the actual, non-measured facts.
We inadvertently let two Jr’s vibe code a project. They used their own Cursor subs and did not tell anyone. They are no longer with us. Now there is this project with no documentation, they removed the LLMs comments, and then ran a linter. I checked the git history. Nobody can make sense of the code base. We can’t even empirically show that it is cheaper to rewrite than start over, so we just stare at it everyday. Nobody can add any features. When you debug you end up in abstraction hell and work through the most buggy nonsense code.
They are no longer with us.
Hey, I’m annoyed by slop coding work as much as the next guy, but murder seems a bit much as a reaction…
Not to the shareholders.
I only want to punch the one guy in the throat. I am not that violent.
Have you tried to ask cursor to add some new features
We don’t even use Cursor. I have considered putting sections into Opus 4.6 and asking what bugs it can find for me. I have done that before with unquantifiable results.
Cursor uses opus 4.6 by default anyway.
Just open Claude code in your case, some thing
Yeah, the big thing is that management has no sense how little coding you actually do in a software engineering role. You spend so much more time understanding requirements, understanding how you can resolve roadblocks within your organization and understanding what the hell the code does that was previously written.
In particular, the last part is something that will most definitely take longer for vibecoded programs.
The code is often needlessly complex, because:- folks throw in additional features with no restraint,
- the AI will gladly generate a second implementation for stuff, you already solved in the codebase, and
- AI-generated code tends to just be noisy, because you need rigorous logical reasoning to find the most minimal solution.
But you also just don’t have human beings that made all the detail decisions and can tell you why they’re important. In vibecoded code, all of these detail decisions are accidental and only ‘proven’ in so far as the given accidental state that the code is in, happens to not explode in reality. If you need to tweak anything about it, you’re completely blind as to what’s actually important and what’s just in there, because the AI figured, it’s the most likely thing to autocomplete there.
Good article. Any company doing any of those examples deserves to die.
The companies that will pull ahead in the next 24 months are not the ones that adopt fastest. They are the ones whose judgment systems are mature enough that adoption does not break them.
Yeah, judgement doesn’t seem to be a high priority for the a.i. addled mind.
Yeah but not until the next financial quarter so it’s all good right?
Oh no i’m terrified to lose my “learning velocity.”
holy corporate word salad
Maybe enough that corporate types will find it compelling?
The real vibe coding is a bottle of beer and a lotta fuck it
There’s code over 20 years old still in use that I had written using that approach.
Who cares if you hard coded all of the if statements, shit works.
Shit never works.
So Ian Malcolm was right the whole time
I agree wholeheartedly with this article, however, it’s giving vibe-authored
The core problem was not simply the technology itself. It was the organizational inability to integrate AI into real workflows, learn from deployment and distinguish between a demo that worked and a system that delivered.
Yeah, it has that phrasing sometimes.
The missing Oxford comma… *twitch, twitch*
Forbes is just a bad source for news now.
This article is good, a rare exception in the current discourse around LLMs.
At first I thought vibe coding was just coding stuff for fun using whatever comes to your mind. Then I learned that it’s just letting ai code for you mostly and just copy paste the code.
Now I wonder if there are some cases of real vibe coding like my first assumption.
First one is called hobby programming.
Yeah, vibe coding is such a fun term, too bad it’s used for this purpose.
Sort of like pickup artists
Sort of like pickup artists
The danger here is that many people think that software is all about having code that seems to work when you try it. Those people have never been able to get past “Hello, World” in X for Dummies, so they don’t realize all the practical realities of software distribution that are very much more nuanced and complicated than just writing the code. They get their hands on some working code and wheeeee!!! Ship it!!!
A while back I compared LLMs to lightsabers - and pointed out how many amputees are found in the Galaxy far far away that has lightsabers.
Or should produce “correct results.”
Produce correct results even when encountering “edge cases.”
Not crash, even when encountering “edge cases.”
Work correctly in all deployment environments.
Work correctly after scope creep multiplies the feature set by 3x, 10x, 30x… yeah, successful projects experience that kind of expansion.
Work correctly after the operating environments shift under your feet - can the code be updated to work with the next version of Android? iOS? Windows? Linux? After “security updates” take away the infrastructure you were depending on for correct functioning?
Will it scale to 100 users? 10,000? 10,000,000?
What happens when “threat actors” actively target the system?
What happens when your methods / development processes aren’t compliant with new government regulations?
Are you ready for IP lawsuits, whether deserved or not?
That is a pretty high bar compared to being correct. XD
Not copy-paste. Let the ai do it for you. …… else how would we get these entertaining stories of idiots letting ai delete their production database
Move fast, break things LOL
There are a lot of folks saying that Bluesky’s recent outages were due to the vast amounts of vibe coding in their systems. It was days of not working.
As an “I wonder” exercise… say that BlueSky wasn’t vibe coded, but instead was done “the old fashioned way” with 20x as many people taking 10x as long to produce the same product. Over that 10x as long timeframe, would they have experienced less or more total downtime with traditionally coded software? Not theoretically perfect software, the actual stuff that “professionals” building social media sites write?
Also, if they have staffed up with the same number of people as were traditionally required, can those people respond to and correct issues slower or faster than a traditional team?
LLMs are powerful tools, which have evolved fairly dramatically in the area of software devleopment across the last 12 months. I suspect as people learn to use them properly, safely, appropriately, they are going to prove out to be quite useful. In the meantime, there will be mistakes made…
There was an article a bit ago explaining that most AI companies are making a 95% loss. You know, spending 100, receiving 5 loss. All that debt is going to mean the price for AI is about 20 times lower than it needs to be just to break even. The software teams that came to rely on AI to save costs will soon enough find themselves on the hook for this mountain of debt. Enshitification is real. Enshitification is coming. AI will not stay cheap, convenient and free of advertising.
People forget this. Yes it has real use in very narrow contexts, yes it may get slightly better, but right now they are JUL getting the kids addicted to vapes and it is drawing ungodly amounts of power and electricity to do so.
The meat you eat has more of an impact on the environment than electricity from AI usage
Two things can both be wrong. And removing something that’s been in place for millennia and deeply embedded in the culture is likely to be more challenging than eliminating something that is still more planned than actually materialized.
Whataboutism doesn’t work here.
Plus the basis of generalized use is founded upon willful mass copytheft.
And which vehicle you go to work in dwarfs them both. Two things can be true at the same time.
Three things here:
-
right nowthey’re basically discovering what are real uses and what’s frivolous non-value add uses.
-
at least as used for software development, the past 12 months didn’t get slightly better, it got dramatically night to day better.
-
simultaneously, some pretty significant advances have been made at reducing costs of delivering value. I think this is hitting hardest in basic chatbot areas, getting the simple answers cheaper - in programming not as clearly, yeah it’s getting the simple answers cheaper there too, but it’s also succeeding at getting much more complex answers that just weren’t possible even a few months ago - those answers cost more, but they’re also worth more… will be interesting to see where this all shakes out.
Yeah, they are running loss leader stuff, yeah it’s going to go up in price when they figure out what its worth to people, because things aren’t priced at what they cost to make or deliver, things are priced at what people are willing to pay. The players with the deep pockets are jockeying for control of future markets, they’re investing their existing wealth in future power. Let’s hope the winners are slightly less ghoulish than our Oil barons.
Let’s hope the winners are slightly less ghoulish than our Oil barons.
What a foolish hope!
$200,000,000,000 debt.
Who well pay it?
You talk like gravity doesn’t exist!You’re wrong if you think that it won’t be heavily reliant AI customers like software companies who spend five years removing codewriting skills from their workforce and building up technical debt in their codebase because no one has to understand it in those five years and there’s a lot of subtle, hard to spot bugs that got through code review because humans simply don’t make those kinds of errors and no one ever had to spot one in their life before claude came along.
Did you think that enshitification wouldn’t affect the product? Yesterday’s computers and cars were easy to disassemble to replace parts. Now it’s much, much harder, and it’s very common to void your warranty if you do that. Today’s ai generated code is easy to tinker with and you can do what you like with your end product. Why would it stay that way? Why wouldn’t they engineer it to make that harder? It’s not difficult to make code confusing by changing variable names. I could fuck up your codebase for humans by simply swapping names like productSKU and customerID, let alone writing obfuscated code for any purpose whatsoever and with whatever variable names I like.
Some software companies are outsourcing their talent to AI behemoths with mountains of debt to recoup. Guess who’s going to pay the debt! And what’s the point of such a company in the long run? Why are you speedrunning paying to replace yourself?
There will be an AI crash and “consolidation”, meaning a switch to monopolies or near monopolies. Some companies are shedding institutional knowledge and programming skill like it was waste water. Once dependence comes, value extraction will follow it like disease follows unvaccinated infection.
There is already $200bn in debt and growing rapidly. The shareholders aren’t going to be paying it. The ai customers are.
Today’s ai generated code is easy to tinker with and you can do what you like with your end product. Why would it stay that way? Why wouldn’t they engineer it to make that harder? It’s not difficult to make code confusing by changing variable names.
Code obfuscators have existed for decades - they are rarely used in practice, and 10 years back when a vendor provided me a driver in obfuscated code I explained to them: “If we don’t get real source code, we won’t be buying your products.” The non-obfuscated code was in my in-box the next morning.
A year ago, the AI engines couldn’t code anything too complicated, successfully. It had to be assembled from “human sized chunks” or it just wouldn’t work.
I notice in a code review I’m doing just this morning, the AI is now managing chunk sizes that are annoyingly large, and doing it successfully. At this point, I’m having to apply push-back pressure, not to keep the code working, but to keep it manageable. The same kind of pressure has been necessary for management of most human developers / development teams for decades.
Enshittification wins most successfully in “free tier” products, people who care enough to pay for something do get influence of the products provided - sometimes. Your counter example of automobiles is a good one, along with appliances, etc. The industrial makers of these products have enshittified our legislatures with rules, regulations and laws which protect their industries and enable them to keep colluding to push overpriced under-durable garbage at us with no real alternatives. We need to push back on government for that, that’s the level where the impediments to customer influence exist.
Why are you speedrunning paying to replace yourself?
I’m old enough to qualify for the next buyout offer, if there is one. Speedrunning “the new tools” is what I have done for 35 years, it has always served me well in the past. Maybe this one backfires? Not my personal problem if it does - disposal of the elderly from the workforce is a tale as old as time, that’s what retirement accounts are for.
The shareholders aren’t going to be paying it. The ai customers are.
It’s much more likely that the banks and their insurers will be left holding the bag, and they’ll then be bailed out by the taxpayers.
There’s already negative ROI at even the current loss-leader prices.
-
Over that 10x as long timeframe, would they have experienced less or more total downtime with traditionally coded software?
i have a homework for you: if you ask professional chef how to keep the cheese on pizza, are they going to tell you to use some glue? once you figure out an answer to that, you should be able to answer your original question.
The term you’re looking for is “cowboy coding.”
Leave me out of this, partner.
I resemble that remark - rode herd on a whole passel 'o C back in the early 90s.
What’s it called when I get real high and code something that I can’t even figure out the next day if it was genius or insanity?
There are some comments in the code I’ve written saying “before you attempt to modify this module, it’d be wise to get a barge pole.”
Ballmer Peak?
I don’t think copy/paste is involved. With vibe coding, the AI agent typically has access to your repo/files directly!
That’s even more dumb and dangerous imo.
No one had the cultural standing to say this looks great, and we are not putting it into production.
Can someone in your organization look at a slick prototype and say “no” without career risk? If the answer is no, vibe coding becomes a one-way ratchet.
This is definitely the feeling at my company. “How fast is AI letting you ship” is the only question management & executive are asking.
the resulting ambiguity will be filled by whoever moves fastest, which is rarely whoever should be deciding.
There’s capitalism!
This is definitely the feeling at my company. “How fast is AI letting you ship” is the only question management & executive are asking.
Slightly faster, but with a way higher upkeep cost. And it might delete your companys database or have a customer-data leak.
Slightly faster, but with a way higher upkeep cost. And it might delete your companys database or have a customer-data leak.
You had me at “slightly faster”, no need to keep selling it. Sold!
- The Average CEO, this year.
These golden eggs are being laid so slowly! There’s got to be a faster way!
Yes! What if we grind up the golden goose and spit it back out as a fine goose paste?! It can’t hurt to try it. (This is sarcasm, as well.)
Can someone in your organization look at a slick prototype and say “no” without career risk? If the answer is no
You have toxic leadership and we have just handed them a mini-gatling-gun with which to shoot everyone’s feet off.