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!

A small company this level of cooked and immature leadership may not have the resources to recover long term from the damage. Even if the bubble pops on them in a year.

I’d start looking for alternatives now, before it becomes urgent.

Let him “vibe-code” himself into a problem, then tell him you can’t fix the mess he done did.

I am making similar experiences, but is is not as bad as you are describing it yet. We have a new member in the team who is not a developer by himself, but he has gotten the task to make our way of working more professional (we are mainly scientists and not primarily software engineers, so that’s a good thing).

His first task was to create programming guidelines and standards. He created 8 pages of LLM generated text and example non sense code. He honestly put a lot of effort in it, but of course there are a lot of things in it that are wrong. But the worst thing is the wall of text. You are nailing it - it is my task now to go through this whole thing and extract the relevant information. It sucks. And I am afraid that soon I will need to review more and more low quality MRs generated by people who have little experience in programming.

I had a manager who pushed AI a lot. When he left, all the pressure to use it seemed to die down. So maybe it’s just a couple of people creating this environment and if you can get away with avoiding them it’s better.

The problem with AI code we saw is that often no human has actually looked at it. During reviews you won’t check every line and you’ll have to trust much of the code that seems to do obvious things. But that assumes it was written by a human you also trust. When that human hasn’t reviewed the code either, you end up with code no one in the company has seen (and may not even know how it works).

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.

Managers using influence to inflate their importance is common. It sounds like they are hoping to head off innovation to protect themselves.

The second aspect is their harnessing of useful idiots. I think critical or practicing objective thinking is being pushed aside at the time when needed most. You probs have to tread carefully and maybe influence people to use their own minds not that of desperate managers. Alternatively set the useful idiots off in competition and watch the fireworks. Not helpful but it is fun.

There always some people are always looking for an easy fix, a way to rise quick, to be a yesman. But for the most they just wanna do a good job. Maybe appeal to the rationally of those people. Trust your brain and skill rather than cult practises.

I haven’t had a similar experience yet, but maybe some if your collegues feel the same way? You could write a letter stating your concerns and let anyone sign who agrees and then send it to your manager. Also, I’d like to add that under German law AI Output can not be copyrighted. You can only claim coownership or something. Maybe that could be interedting to your managers?

Two words: malicious compliance.

We’re using LLMs at the company I work at and it seems very useful in many cases but sometimes it still doesn’t work. I’m a bit worried about the aspect of the code rotting by LLMs generating stuff based on existing code.

My mindset has shifted a bit, now I’m more focused on making stuff easy to find and easy to figure out patterns to use so that the codebase becomes easier to work with. There’s some horrible code in the project and the LLM absolutely sucks balls at it but if it’s a clean routine job such as making a table with update dialogs and actions to manipulate the data the success rate is >95%.

So yeah, don’t trust it, treat it like a junior dev that got straight As in school and has never considered security. Code reviews are now where it’s at.

It’s hard to be a contrarian in these kinds of positions (I’ve been there, and it didn’t end well), so I wouldn’t be too outspoken, but at the same time, try to innocently point out the issues with approaches like this. I would just try to point out the flaws in this approach, the same that we would for any other kind of programming fad - without making it seem like it’s an agenda, of course.

For example, any time teams are looking for feedback - code review, retrospectives, etc. - just point out the flaws on why vibe coding is a bad idea and bring it up casually when the time comes. It doesn’t hurt to be honest as long as you don’t come off as being an ass about it.

Endure the next year or so, until it pops and there will be a massive need for senior devs for fixing the slop machine.

Let me give you a little parable.

There once was a juggler, who could juggle with three balls all day. Then someone from the audience threw a fourth ball, and he kept going. Someone threw a glass, then a flaming torch, and he kept going, occasionally burning his hands. Seeing he could do it, someone throws a machete, and the juggler almost never cuts his fingers keeping all those things in the air. A chainsaw gets added, and an open bottle of bleach, and occasionally the juggler gets his hair caught or spills some bleach, but he keeps going. As he keeps going, people keep adding more and more things. Eventually it’s too much, and it all comes crashing down, killing the juggler and several members of the audience and destroying all the objects in the air.

On the next street corner, a juggler stands with three balls. Someone from the audience throws in a fourth. He steps aside and lets it fall to the floor, happily juggling three balls.

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.

Remind them that copyright cannot be enforced on anything AI-written.