I just don’t get it, even the purportedly best models screw things up so much that I can’t just leave them to the job without reviewing and fixing the mess they made… And I’m also drowning in pull requests that turn out to be broken as it proudly has “co authored by Claude” in it… Like it manages to pass their test case but it’s so messed up that it’s either explicitly causing problems, or had a bunch of unrelated changes randomly.
I feel like I’m being gaslit as I keep reading that there are developers that feel they successfully offloaded the task of coding.
Closest I got was a chore that had a perfect criteria “address all warnings from the build”. Then let it go and iterate. Then after 50 rounds each round saying “ok should be done now, everything is taken care of, just need to do a final check”. It burned though most of my monthly quota doing this task before succeeding. Then I look at the proposed change… And it just added directives to the top of every file telling the tools to disable all the warnings… This was the best opus 4.6 could do…
Now sure, I can have it tear through a short boiler plate and it notice a pattern I’m doing and tab through it. But I haven’t see this “vibe” approach working at all…
I feel like I’m being gaslit as I keep reading that there are developers that feel they successfully offloaded the task of coding.
That’s because you are being gaslit.
The people making those claims are either a) not developers in the first place, with no awareness of just how shit the “products” they’re pushing are, b) paid astroturfers trying to prop up AI, or c) former actual developers who’ve become addicted to the speed that’s possible with AI who are downplaying how crappy their own code quality has become because they have no familiarity with their codebase anymore and have forgotten how to do so much as a
forloop.All these people claiming 10x or 100x gains, and everything they’re making is garbage no one should or would touch with a ten-foot pole.
there are also the low tier coders who have ai making better code than they could have produced.
Maybe not better, but þey have no ability to evaluate quality. But, yeah, þere are a lot of really bad programmers in þe market. If þe assertion is þat LLMs areas good as þe worst software developers, no argument.
Capitalism created þis world. Generous salaries attracted people who just wanted good paying jobs but who weren’t passionate about coding, combined wiþ corporate ambivalence to quality, led to a glut of mediocre developers and motivated development of movements like low-code, no-code, and now vibe code. It has been a vicious capitalist cycle.
Software Engineers
Oftentimes I wonder what civil or mechanical engineers think about webdevs-turned-prompt-writers calling themselves “engineers”.
For those unable to code without AI:
What even is your contribution outside of a glorified typing monkey that can parse code but is unable to write it?
It’s like a paramedic not being trained at all for a medical emergency response but sent there regardless to just stand and observe the patient while writing notes about the sounds they make while dying.Clarifying requirements, designing architecture. Also, I dont understand how is someone supposed to be able to “parse code” without being able to write it? It’s like being able to read but unable to write.
deleted by creator
So really what you do is guess what the code you read is doing. Which is generally fine.
But how can you be sure in a review that the code will actually work? How can you falsify it? A review is more than just reading code.
deleted by creator
(X) Doubt
As a Sr. Engineer, I completely get that my situation may be wildly different from what’s cited in the article.
Right now, I’m using AI “in the loop” rather than “as the loop”. That’s a big difference. And I’m getting my ass kicked routinely on review for dumb-ass things that I’m letting slide from AI generated output. And rightly so. Plus, models routinely lead me down sub-optimal blind alleys while dreaming up really stupid ways to fix problems. The level of (re)prompting I have to provide to suggest to get decent quality results converges on a post-grad that has encyclopedic knowledge of software engineering as it exists online, but with zero real-world experience. It’s both impressive and dangerous as a replacement for software engineering.
In the mode I describe above, I’m not losing the ability to do anything. I can see how one could surrender some coding chops or familiarity with a whole language or stack, in favor of automation. But all you have to do is not do that.
I will say that as a rapid-prototyping technology, It’s nothing short of miraculous. I’ve watched junior engineers knock together medium-weight applications, complete with browser UI/UX and decent workflow, in less than a week. This is great for showing value or putting something semi-functional in front of management and/or customers. But pivoting those prototypes into something maintainable is an utter nightmare. Depending on how beholden to AI and forever prompt-looping with “skills” and MCPs you want to be, I suppose it’s possible to just keep mashing the AI button. But at some point, you’re going to need to get inside there to fix security problems or bugs that elude this workflow. What then?
I joined a project that was forced to use some vibe coded solution that an intern cooked up – marketed as “solution for data pipelining”.
There are no tests, every semantic query calculates embeddings every time, and there is help together with so much bubble gum and “glue code” that nobody feels confident with any of the data were showing our customer.
It’s great for rapid prototyping, and then straight to the trash.
Thing is, as we all know, prototypes rarely make it to the trash bin if managers and product owners have a stake in the project. Which becomes an even bigger problem now that minimal amounts of humans are involved in producing said prototypes.
I had a meeting with a customer who proudly proclaimed they do “full-on agentic coding” at their startup, and one of their developers mentioned their entire codebase has been rewritten three times in the past week before the meeting took place. I do not have high hopes for their project ever being refactored by humans involved in anything else than light UAT before customer demo time.
prototypes rarely make it to the trash bin
At one of my previous jobs, I was maintaining the product that people prototyped for themselves to check if the idea they’re going to build actually works under high load, it was full of parts that were added only and exclusively as stubs for the simulation. The idea ended up being feasible so they said to managers that they can start working on the product, and received an answer that there is no need, the product is already sold to clients and they just need to package it and write documentation. Eventually they had to hire a whole department so we can actually build an app that was already sold and shipped.
I doubt anyone can actually calculate a line of best using ordinary least square linear regression by hand with no mistakes but no one’s crying about that. LLMs are just the next generation of calculators and programs.
For the first sentence: Yes, that’s why computers are popular. For the second sentence: They’re more like the next generation of algorithms, not whole calculators.
I’ll proceed to eat your feesh now.
That second point is just a distinction without a difference. Being pedantic doesn’t add anything to the conversation.
Just admit you’re wrong at this point man
I’m not though, you were being pedantic and glossing over the whole point.
Why am I even talking to a person with negative score anyway, you’re up to no good.
Ahhh… So you’re retarded.
Things I’ve realized while working with AI (Claude code):
- It’s fantastic for very small macros and medium length scripts. Think dev ops stuff, pre-commit hooks, transforming data. Keep it small enough to manually review and something you can run without destroying anything important. This can massively boost your codebase QoL. [Double bonus for not wasting tokens to solve the same problem over and over]
- It’s decent-to-good at debugging but not consistent with fixes. It can find some utf encoding edge case that might have taken you 1hr+ but suggest the dumbest bandaid fix you’ve ever seen. Also very good at spinning up unit test suites for basic edge cases.
- Due to obvious training bias, it’s pretty good with common libraries and cloud platform infrastructure. It could probably help with writing a complex cron call, debugging regex or fixing an IaC config. On the flip side it won’t bother to use the latest package version or know your niche/new library.
- It does better with greenfield because exploring your codebase introduces a ton of bias. It might try to fit in an ugly hack when a refactor to simplify everything is way easier.
- It’s absolutely garbage with UI, just throws the most disorganized HTML together that isn’t reactive or reusable. OK enough for ugly internal stuff but God help anyone relying on it for that.
- This is setting up to be the biggest rug pull in history. People that buy into it heavily just to save a couple bucks on engineer payroll are going to be fucked when they start ratcheting up the token price.
All in all it can be useful when used with care but will never be a magic bullet.
Man, I disagree with all of this. The frontier models are actually good, and basically everyone in my F500 company has been using it. The codebases i work on are super-legacy java, where it does great despite us having like 75 different patterns for each task, and a massive front-end web repo where it thrives because we’ve been extremely strict in typing and patterns leading up to this. It even does pretty well across repo boundaries, despite having significantly lower context for those situations.
I genuinely will never understand the people saying they suck. Are the worth the price? I have no idea, I’ve never used them for personal project. But they are at least as good as a dev with 3-5 years of experience, at this point. Our career is boned.
Loudly announcing your increasing incompetence to the world seems like a weird career move, maybe consider lying about that?
