And it doesn’t make anything better.
From my experience, how much a project or product has embraced AI-assisted development is directly correlated with how much worse it got.
Everyone is using AI to do *something*, including ridiculous agent teams.
But no one doing these seems to produce anything useful. Only pretending to do something.
Also, people who use AI for development seem to be more detached from their work, and its quality. Please don’t pollute the internet with worthless slop.
Login to reply
Replies (32)
Happy to take this on. Fixed price in BTC, refund if I can't deliver. https://mtg-arbitrage-desk.loca.lt/order
Diversity in, diversity out.
I find the correlation that using AI means outsourcing decisions. We are generally lazy, and "AI suppose to be smart", so lets just let it decide.
But generally AI as a tool shall not mean it. I think we might see more issues caused by AI deciding, until we realise its just a tool, and the responsibility and control must stay with us.
Have you tried out my slop coded app @Nostr Vault?
I think it’s pretty useful, but I’m pretty biased.
It eats 60% cpu on android.
Which version
No exact info, installed 1-2 week ago from zapstore.
The latest version 3.1.0 manages the backgrounding much better, sorry about that bad experience, I test on a moto g play with 4gb of ram, and I really try to put it thru the paces
I have been working on a thing for a decade. I very much care about the code quality. But I am a stay at home/homeschooling dad with 6 kids. Many of my other ideas hinge on my thing existing. In the past week, Claude 5.5 has allowed me to have that thing. Maybe some day I'll go back and hand code it, but I kinda doubt it at this point. It is mainly a protocol and I can have a 1000 implementations of it for cheap. You don't have to use it. But someone else might, so onto the heap it goes. So sorry.
I appreciate the feedback
I agree that AI slop is worse than human-written code. Not necessarily in terms of quality - it might actually be better in that regard - but in terms of structure.
As soon as you unleash AI on your codebase, you start noticing how it becomes more "fractal", losing the clean, "crystalline" structure it had before AI touched it.
And honestly, maybe it's just me, but AI-generated code is hard to work with, and I don't particularly want to. It makes me want to have another LLM fix the problems created by the first one.
Still, I think that if you find the right balance, AI can be a pretty good tool.
> So sorry.
That's just it. There are nothing but a lot of apologies but your apology for bad stuff doesn't make it better, it just makes me more sad that you know it's bad and you don't care because it makes _you_ feel better.
that's where we're at now and it's what makes it so frustrating that the most talked about topic in software development _right now_ is agentic work with overwhelming admissions of "no team of humans could possibly review the amount of code produced"
Which is arguable identical to an apology if you don't mind me stretching. Sorry it's bad but progress is progress my dude, what am I supposed to do?
> It is mainly a protocol and I can have a 1000 implementations of it for cheap. You don't have to use it
Then a white paper is more useful than your code in this example. Your code is literally just slop, id rather re-write it using the white paper then fork it.
The problem with quality comes from the fact that most human-written code is written for the purpose of being written, and that is it.
And many things, such as security vulnerability discoveries, etc. are either 1. Strongly human guided, or 2. Could have been discovered via standard auditing methods + prevented via standard practices (but no one wants to because “waste of resources” until it bites them in the ass)
I think that it is somewhat fine to have slop code as a proof of concept and that is all.
But PLEASE, STOP WRITING SPECS WITH CODE! What the fuck?!
I don’t need to see “Treat X as Y, not as something something evidence” for the 10 millionth time
I mean more like a white paper, rfc whatever you want to call it. specs for the slop machine are equally useless.
> I think that it is somewhat fine to have slop code as a proof of concept and that is all.
As long as the PoC is not the white paper.
Specs as in the real definition not the fake LLM definition
I have to admit that I have gone down the slop route. The other day I decided to write code by hand again, during recreation at least. Because I made a basic relay with Erlang, and looked at it and I recognized I didn't understand anything about the relay, nothing about erlang, and nothing about the protocol.
I can read the code fine, like I know what a function is. But seems to me, in my experience, that "vibe coding" something and trying to understand after the fact is 100x harder.
Or maybe the damage is done and I've just fried my brain, which is even more frustrating and sad (although fixable, imo)
Yes, this is how my life has gone the past ten years. I tried writing the program I wanted. It was a miserable circle of failure since I didn't have a clear spec to go from. So I stopped coding and started writing a spec. I got it to a decent place but it had some holes.
At this point I made a bit of a lazy mistake. I let AI write part of the spec. I kept having to go back and clean it up. It would make such stupid decisions that made it 10 times more complicated than it needed to be.
So I took a bit of a break while I went back through line by line.
My trouble is always that I simply don't have the time to get it all in my head and then get to work. I am full-time dad. 24/7. AI can't help until it actually understands the problem.
Fast forward to last Tuesday. My cousin got me three months of opus 5.5. I gave it a trial. I described the problem I was trying to solve. I told it to do 2 things. First tell me if it agreed with the problem and second propose an architecture that would solve it.
To my amazement. It actually understood it much deeper than previous models. Better than any human I have talked to. It also proposed the exact architecture I had already defined in my spec, with a few clever simplifications that make it easier to code.
So I had it write it up. Mostly to get a clean copy without the crap previous models came up with.
Then I had it write an implementation as proof that the spec was complete and functional. The answer is yes-ish. In less than a day I had something I have wanted for years. It wasn't the complete spec and it still needs work. There is some evidence that later parts of the spec will need to be rewritten again.
The thing is not going to be useful with any amount of slop. It kinda needs to be perfect. But I think parts are getting there. And by that, I mean the spec. The code only tells me if the spec works.
I want the prose to be good enough that any idiot (ok any competent programmer) can make their own interoperable implementation.
Chances are no one will. But darn it if I don't see hope that I can at least have the tools I want myself.
Hold on. I don't know that it is bad. I have standards and I am not going to push out slop for the sake of it. But I also think certain models are going to be a better coder than I ever was un short order. When I find the edge of their capabilities I back off and start that part over. My only point is that if AI can help accomplish something I think is important that I was failing to finish on my own then I am going to use that tool.
But if the tool gives me crap then it hasn't accomplished the task. What I refuse is the purist standpoint that AI can't do good work.
Luckily in my case, I can judge the output because I've been working on it for 10 years and have coded about 4 iterations myself. I immediately recognize all the patterns the AI is using. If it does something else it had better be able to explain why.
Again the code itself is only proof of the spec. I may use it myself. But I don't want to even call it a reference implementation. It is a test implementation only.
To be perfectly honest. I greatly respect your opinions and would 1000x prefer to have you critique my spec. I'm just afraid to ask because it'd be hard to see a decade of pondering while walking around with babies go up in flames in an instant.
> Hold on. I don't know that it is bad. I have standards and I am not going to push out slop for the sake of it. But I also think certain models are going to be a better coder than I ever was un short order. When I find the edge of their capabilities I back off and start that part over. My only point is that if AI can help accomplish something I think is important that I was failing to finish on my own then I am going to use that tool.
We're arguing that's what most devs think. We all think we're better than the next guy and that's the flaw.
*What most competent devs think* to be more clear.
obvious vibe slop is largely excluded imho. Im talking the products of those produced by "real" developers, experienced ones in particular. The ones that brag about specs and their automated review processes etc.
That's also kind of what's been botherin me. Obviously my time is just as limited as everyone else's but id rather be asked and see possible discussion than what is happening now - ideas in a vacuum rubber ducked with something that gives bad but convincing and validating responses. And my base opinion is that were actually all equally psychologically vulnerable to the word machine. We're all probably decent at convincing our mothers to agree to something when their willing to listen lol
Sorry. I'd like to back up a bit. Some claims I would like to NOT make.
1. That I am a good developer.
2. That I am an experienced developer.
I am an unemployed engineer who spent 6 years doing web development. I am bad at project management, I am bad at spec writing.
The ONLY reason I am working on this at all is because no one else seems to be. There is a disconnect between what I want and what I can deliver. I'd much rather someone like you or @semisol were doing it. I'd be happy to provide a link to what I have so far if I thought I'd survive the encounter.
A minor clarification, when I say no one is working on it, I mean everyone is working on it, but they keep doing things like adding blockchains, or refusing to go low level enough to make it performant or secure.
I agree. I haven't asked because why the heck should you look at my ideas when you have 10 of your own. I know the feeling. People come to me with dumb ideas fairly frequently.
In my defense with AI. I am fully aware of it's tendency to pander. That is why I don't ask it what it thinks of my idea. I ask it what it would do. Then I compare to what I have done.
I have not tried it yet. Due to security reasons I do not test many apps.
But I will see if I can provide feedback :)
I hope you can try it out, you don’t even need to do an nsec or a bunker, npub only and you can see basically everything, there’s also the webapp if you’re a no downloads at all kinda person.
Oh ok then I can for sure. Will see.
What are the main things you would want tested outside of the basics?
Biggest thing is the web of trust and the backup export
Could you verify my campaign?