Here's a really good article- with illustrations- about something that theoretically can and so definitely will go wrong with "agentic coding" on any sufficiently large project. The Decoherence Trap

It warms my heart to read first-hand accounts of this starting to happen at real companies. I hope the financial AI crash holds off just a little longer to make sure they've all had enough time to dig themselves into holes past the point of no return.

Nightmare comment section. Just comment upon comment of 'what you're describing isn't true' and 'tell your agents to make fewer mistakes.'

Redlib link for those that would like to avoid Reddit: https://redlib.catsarch.com/r/ClaudeCode/comments/1xq2n3t/a_year_after_the_100_ai_mandate_nobody/

There’s a chap in the comments suggesting to stop worrying, you only need to keep to the spec. The sheer stupidity of that.. good god the stupidity.

This month management started quietly asking us to "write more of it ourselves" again.

I think... I think there's a chance that if I was in-person, and this happened, there would be violence.

This is the full post

A year ago leadership told us everything goes through agents now. Hand written code needed a reason. It felt great for about 6 months, we shipped more than we ever had.

Now the codebase is around 3x the size it was and nobody can explain how most of it works. The agents struggle too, they keep re-implementing helpers that already exist somewhere else because there are 4 versions of everything.

The part that actually hurts is who is left. The two people who used to read every PR and push back left in the spring, one quit, one was "restructured".

coderabbit still comments on every PR, but there's nobody left who can tell if a comment matters or not, so people just resolve them all.

This month management started quietly asking us to "write more of it ourselves" again. Except half of us haven't written anything non trivial in a year and it shows. I spent 2 days on something I would have done in an afternoon in 2024.

Not sure what I want from posting this.

Is anyone else at the point where the mandate is being walked back, and how did you get the understanding back?

(A) Humans are also not deterministic, in fact AI is much closer to determinism than human brains.

I would want to fire this guy so badly. Saying this out loud is such a clear indicator that you're a liability.

I’m fairly confident that the post is written by a human, but one that has used AI so much that they literally are using AI’isms like “The part that actually hurts” or “except … and it shows”.

That's what I'm afraid of for the rest of the development department. They have adopted AI to the core, and barely accept that my part can't be AI'd for a number of reasons. Once the LLM providers will have to actually charge costs+profits, they'll lose their whole toolbox, and I doubt they understand what they did to their code base.

Whole post read like programmers confused and found out the responsible one are in a very very minority, and the issue has been aired by literally everyone who are against vibecoding.

This is the point I've been making on this shit for 2 years - LLM code generation reduces the costs of code creation by adding massive costs in terms of reviewing and maintenability.

LLMs don't do software design, and that includes things like designing for maintenability.

Management was always like this long before AI.

They ask their assistants to do something and after a few years they lost touch with how anything works.

Everyone needs to read that post.

Yes, everyone. Even though it is on Reddit.

Maybe this is the AI game plan and why the big AI companies want the local models regulated.

Currently they're more or less using the follow the free sales tactic but with the difference that the customers are already used to paying, just by far not what is required to cover the operational costs.

They can soon crank up the price and companies will have to accept that. They're going to get extorted -- either pay or basically lose your product.

Amazing to see a Reddit thread with actual human replies that aren’t just snark or lame attempts at humor. I didn’t think those existed anymore.

I'm so tired

When people ask what I do for a living, my response is "software engineer". Not "programmer" or "coder", because the actual programming is only part of the job.

The big concerns for software engineers are reliability, maintainability, and scalability. Using LLM-generated code without reviewing and understanding it is antithetical to all three of those big concerns. Personally, I find that the amount of review and revising it takes doesn't end up saving me any time at all compared with me writing it myself.

Almost anybody can build a bridge given some materials. But there's a reason we have licensed engineers who review and sign plans for them. If your methed out neighbour built a bridge in an hour by himself, would you drive over it?

Let's store all pirated media in a virus that mutates our DNA to store this database, then infect the whole world with that virus. That way Disney literally couldn't take legal action against all of us.

Edit* nevermind Disney would get the lawyer that got epstine off and some how we would end up all becoming property of Disney.

I was just listening to a fitting song on the way to work this morning, by Memphis May Fire.

What goes around comes around
That's how the world turns
You'll get what's coming to you
Then maybe you'll learn
The price that you pay
For being two faced
Is sleeping in the bed that you made

I like the people who are all "when it gets too expensive, we'll just move on to self-hosting open weight models", as if that doesn't require having a fleet of GPUs running in the back room, or cloud compute in a data centre that will be charging a premium on cycles once everyone and their dog starts subscribing.

“Do it with Muse” ad that I got is ironic.

It’s also interesting how much handwritten “Claude speak” is on HN. You’d think that’s the primary source material … hmm.

I'm shocked! Not that shocked.

Wellity wellity wellity

I think the top hackernews comment is probably right and this post itself is AI engagement farming, the user has their account set to hidden, it's very pointedly feeding into a narrative and telling people what they want to hear, and the writing style is AI-ish even if it doesn't have the more obvious tells.

My actual God, the "people" in that thread.