Ask Lemmy

Vibe coding has its place (prototypes, simple webui), and use of ai for debugging, adding simple features can be effective. Even for more complex stuff it's getting better all the time; but one needs to keep an eye on what it's doing and refactor when things go off the rails (and ai can be a good tool for refactoring).
But ai messages that just make it look like someone is contributing are infuriating. I've responded by booking a meeting with the person who pasted ai slop to review the message because 'I want you to go over that and explain it to us'. And then when they try and start going over irrelevant crap I ask why it's relevant. And when they respond that actually it's from chatgpt and they don't know, if they don't immediately indicate how they realize now it was useless to paste stuff like that, I make sure they know my opinion (in a nice way). I haven't had the same person do it twice.

I have limited dev experience, but most Devs at my company utilise AI heavily for generic code generation.

For example, writing out the core functionality, and then having the AI propagate variables, error flags etc.

I like your point that it really should be a "Copilot". (Although ew MSoft.)

I think it's important to continue to experiment with new tools as they improve, and try to use them responsibly.

AI always produces extremely generic, slightly verbose phrases though. So it's important to generate in small enough chunks that you can use the structure it creates and then add the specificity and accuracy for your purpose. And cut out the waste.

enjoy the ride, look for a new job, so you're not surprised when the buisness dies due to declining quality or a customer lawsuit suing for damages.

I work in consulting big corporate, it's AI all the way down.

I work in a GenAI team building mostly LLM powered applications. I use LLMs to work, they can be useful if you pick thr the largest model available. with smaller models it's just a waste of time.

My boss is very technical (theoretically) but he sold his brain to LLMs and do hard vibe coding of whole apps by himself and the pushes the on the team to do bugfixing...

Fun fuct, he always downsize estimates, since "he can do in 1/5 of the time".

yep, welcome!

Murder.

Have you been to one of the workshops? It CAN be very useful when used right, setting up an MCP server pointing at internal docs/best practices made a huge difference etc.

Any criticism of AI you give will have a lot more weight if it comes from a base of knowledge. That means learning how it does (and doesn't) work so you can critique it without coming across as "that anti AI guy".

Make sure you go through everyone else's PRs carefully and pull out all the stupid AI stuff, it's great fun ripping them to shreds sometimes.

The best case for AI is that can heroicly solve problems that shouldn't exist in a first place in a well-managed company.

It can structure unstructured data, that should be structured. It can do simple coding where aboslutly no coder could be reached. It can sum up the red tape and produce red tape to meet complience requirement that were where created for the sake of the red tape alone.

If the work is important, it shoudn't be done by AI. If its busy work, it shouldn't exist in a first place. No one would need AI to summarize anything if everything was writen clearly.

And thats best case scenario when you don't count halucynations, errors, edge cases, company specyfic context - things that results in task taking more time after all the "oh it produced something and it looks good at a glance" cloud fall down.

If your company is getting a lot of value from AI than your company is trash.

Agentic use of AI didn't really work well enough until December of last year. The models and tools just improve that fast. Codex/claude (or opencode with the same top models) is what you'd need for it.

You still need to plan and define clear specifications for the model. Spend 80% of your time planning and breaking down the job into steps and it'll be pretty self-going from there.

Of course, this works best for common frameworks and solved problems or logical problems. React/node developers can easily 10x their output, and get it done better than they would by hand.

I'm working more with empirical development, so most of my time goes into studying environments and adapting to it. I get most benefit out of having agents read through logs and figure out what happened. It gets it right maybe half the time, but it's a good rubber ducky even when it goes wrong. I'd say it 2-3xes my output. But I can probably improve my usage, too.

But yeah, code review is where it hurts. If it's slop, it just takes so many rounds to get it right. Even when it's good, it's just so much code to review.

Tell him that vibecode is barely good enough for a prototype (a demo) and watch if he fires you. Either way is a good outcome.

My advice is to stick to making specific observations as not to sound hyperbolic.

Yet at the same time what is needed is a counter-narrative to "AI can do the job". My observation is the industry cycles between agile and waterfall development and people (who are the loudest) are using AI to do waterfall development. This is a bad idea for the same reasons waterfall development is bad: trying to write a spec that covers all situations is counterproductive.

The alternative I see is that we borrow from Agile and treat AI as a pair developer that you outsource small (or largely repetitive) tasks to. This is not vibe coding nor is it TDD as you are still actively developing all levels of the code, but your feedback loop (OODA) is kept short.

Switch employer.

(hey ich bin bei dir das kann nicht einfach jeder trivial aber aus diversen Gründen habe ich das bereits 2x in den letzten jahren)

You should try it yourself. See what it can, see what it can’t instead of just arguing with your boss about something you don’t really know anything about

And then you can actually bring the facts to your boss.

Give it a honest shot, learn the limits, help your team establishing best practices. Depending on the language, framework and vertical it can be mor e or less autonomous.

It always needs good specs, clear steps, and lots of testing. Some of which can be automated by the AI itself. Make sure you write yourself or strictly review (am besten im vieraugenprinzip) the code that touches anything related to money or safety.

Do better than the AI

I think you're gonna have to kill them