about 5- Where those numbers from? sound like interesting research
yoavgeva
sodapopcan
This resonates with me in my short experience with agentic programming (and actually, your latest blog post helped push me over the edge… I was gonna comment in the thread but then… uh… I didn’t, lol). For me the “speed up” has been getting over the resistance to diving into stuff I know is going to be a giant slog and has been great for learning new things (and fixing non-critical stuff in my editor setup I’ve just dealt with forever as well as developing some credo rules I’ve had in mind for years). But again on point with the topic, and I guess I’m just reiterating what I already said, but I think too many developers see this as a ticket to stop having to care about code. “Who cares so long as it works” is one of the great cop-out sayings of our industry for as long as I’ve been a part of it.
3
sodapopcan
I think we take the time put into this discussion that we wouldn’t be having if AI weren’t a thing as a data point 
1
mudasobwa
Creator of Cure
Wow. I was under impression I never receive a feedback on my posts simply because nobody bothers to read them.
Duh. A ticket is a great word. Like, cool, yeah, now it’s legit. But let’s be honest: did they ever care about the code beyond “it compiles, it passes primitive tests, what else do you want from me, your toxic reviewing magesty?”
2
sodapopcan
Yeah, I read it twice and was pretty upset at first, lol. Like, I was already starting to give in but it was the push I didn’t want but needed.
Yeah… this is exactly what I was trying to say: a large number of people (possibly most but I don’t have any data) working as a professional software developers do not care about code quality and never did. I’m not the best at expressing myself, which is partly why I dislike LLMs. Though I’ve been very successful with them over the past couple of weeks particularly when I describe how I want something implemented as opposed to describing the feature. I realize this is not how I’m “supposed” to do it, but it sure works well.
1
ziinc
Our codebase has around 7 active contributors concurrently working on it, AI usage is fine if it gets the job done faster.
Some things we have:
Extensive unit testing - All changes need test coverage if possible. Sloppy and lazy code or tests get sent back to the dev for reworking. Continuation of this would result in warnings.
Credo and dialyzer enforcement.
Unit testing against migration rollback scenarios.
Extensive integration tests across stack/service integration points.
Anything that is ui (JS + liveview, or js + api) requires a PhoenixTest against playwright.
A golden path for performing unit testing at each point of the application, so LLMs get them mostly right on the first try.
A shared AGENTS.md with coding style guide.
At the end of the day, the dev is responsible for the code quality that they put out, and they are held accountable for it. Saying that ‘Claude missed that oops’ is not an excuse.
6
tfwright
I have been mostly (~90%) vibe-coding a side project as an experiment and this is exactly what I’ve found. Whenever I look too closely at the slop it is just an uncanny valley of things that only mostly make sense, debugging it properly would require wasting all the time I saved by generating it. Once you reach a certain point, the only solution is to dig a bigger hole for the slop and shove it back down.
Also my experience. In my professional work, in order to maintain quality I mostly have ended up just using the AI as part smart rubber duck, part typing assistant. I am most grateful for it doing things like extracting an id from a screenshot a PM sent and adding it to a config, to take a recent example. Saved me the anger at the PM for not including the actual text in the ticket instead. Occasionally it’s useful for writing simple migration scripts. For actual code it kind of functions like a code generator that gives me an on demand template to start from. But the problem continues to be how others are using it, the many others who, as you say, don’t and have never cared that much about quality and now have an excuse and a fire hose.
5
sergio
Lately I’ve been feeling that any alleged speed increase is all paid back in full and then some when bugs happen and changes need to be made.
You notice the weird choices that didn’t seem weird initially, and subtle bugs that weren’t caught. Not to mention that you have to read the entire feature anyways to understand how it works and make the necessary updates.
I’m not really seeing turbo gains.
5
yoavgeva
From my experience only 10% from the developers of a team care from code-quality, the other just want to get the job done, so by adding deterministic tools that help code quality, you can reduce this, but still it never 100%, with AI you can actually do code review on their code which gave me good input how to split their issues and simplify and add rule set to to review by from Arch POV or other POV which are important to me, I have to say 90% of the time it really helped, and it actually easier to explain the issues.
I was a developer in the time that google/stackoverflow was starting and I remember as a junior back than that seniors always thought I was sloppy back than (they learned everything from books), which I agreed but with good mentorship I improved and learned a lot, some juniors care to improve, some don’t really care (they usually stay developer for a short while), if you want junior to develop as engineers it mean mentorship from the senior not only deterministic rule set, but real mentorship and see where they weak and how to improve, what you describing are issues that happen before and AI didn’t really change something, the slop was there before, now it’s just faster (which is annoying as hell), but with AI it easier to reject their code and tell him to refactor everything (feel good I won’t lie).
1
sodapopcan
Duuuuuude, don’t look at the code, man. You’re killin’ the vibes!
Your whole paragraph is on point there but ya, mostly how been using it as well. It’s also definitely been a great learning tool. I’ve been learning Flutter the past several days and I’m way farther along than I would have been if I’d had to start by learning Dart then getting into Flutter. Here I actually do just describe what I want and then take the time to learn what’s it’s done. Of course I have no way of knowing if it’s idomatic just from reading it, but it’s been a great starting point, and I’m less concerned about idomatic front end code (which is basically all it is). I’ve been mostly still writing all the backend code “by hand” (I still use and like snippets). Though it’s very Ash heavy which also comes with a lot of its own [deterministic] code generation.
2