AI for Coding: Why Most Developers Are Getting It Wrong(ksred.com)
ksred.com
AI for Coding: Why Most Developers Are Getting It Wrong
https://www.ksred.com/ai-for-coding-why-most-developers-are-getting-it-wrong-and-how-to-get-it-right/
29 comments
The divide described in the article is just nonsense. Maybe it's more along the lines of those who focus on quality versus those who focus on shipping. In the end, we want both, but which aspect is more important to either party is what I get from a good faith read of the article.
If you care most about quality then you realize that AI coding assistants are fundamentally unable to provide that since they fundamentally rehash what they have been trained on, which is, fundamentally, mediocre code, on average. It must be, because it's hard (impossible?) to measure, so can't be filtered out of the training data. Quality focus engineers (the "craft" folks) don't consider that enough. If all you want is ship, you don't care.
And that'a fair, it's a matter of personal priorities, and customers value these things differently too, so there is a place in the market for either. But please stop phrasing this in terms of loving syntax puzzles or such nonsense.
If you care most about quality then you realize that AI coding assistants are fundamentally unable to provide that since they fundamentally rehash what they have been trained on, which is, fundamentally, mediocre code, on average. It must be, because it's hard (impossible?) to measure, so can't be filtered out of the training data. Quality focus engineers (the "craft" folks) don't consider that enough. If all you want is ship, you don't care.
And that'a fair, it's a matter of personal priorities, and customers value these things differently too, so there is a place in the market for either. But please stop phrasing this in terms of loving syntax puzzles or such nonsense.
I find it useful for some things, and not for others. I've gotten pretty good at judging what both Claude and I are going to do well, and what each of us will make a right mess of. Happiness lies in the middle ground.
It's not about solving puzzles for us, craft-focused developers. It's about caring about the clarity of our writing, for the same reason a writer wouldn't want an AI agent writing their book.
A writer wouldn't want an AI agent writing their book because the end product is their writing. Readers actually care about the words and prose. That's what they are paying for after all. Users of a software program meanwhile want it to work as promised, that's it. The syntax, language, design patterns used, how elegant the code is etc. are all irrelevant. If an AI agent can write "better" code (in terms of meeting that promise) than a human programmer than that is objectively the right way forward.
This can apply to anything. Why care about how well a shovel is built if all you care about is a hole in the ground. Why care about the quality of the hole in the ground if all you need is laying foundation that can hold a house… Why care about a neat kitchen, if all you care about is a cooked dish. At every layer there are multiple sets of end users. Code is used by programmers and businesses as the source of truth and automation. Just like shovels are used by people who dig with them, not just customers who need holes.
If your point is nobody will ever need to read the code, there's a reason why truly self-driving cars aren't happening yet. We will need human intervention as a failsafe, probably for a while. And humans have been known to care deeply about way lesser things than reducing friction for handling a failure contingency.
If your point is nobody will ever need to read the code, there's a reason why truly self-driving cars aren't happening yet. We will need human intervention as a failsafe, probably for a while. And humans have been known to care deeply about way lesser things than reducing friction for handling a failure contingency.
The code being written really is the end product. Code is unreasonably effective because it is an efficient representation of an idea.
Dismissing all the other concerns it addresses except what the user thinks and how much money it makes is not only irresponsible and short-sighted, but an objectively inferior product. In one word: fraud.
Dismissing all the other concerns it addresses except what the user thinks and how much money it makes is not only irresponsible and short-sighted, but an objectively inferior product. In one word: fraud.
It’s the personal experience of watching Claude Code fly off the rails. Could I spend even more on Claude and attempt to get it to half ass the review? I suppose. But it still my ass on the line when it all goes sideways.
I don’t think it’s a craft thing, I think it’s a speed thing. Senior engineers are faster in their existing workflows.
I’ve found ai tools most valuable for:
1. Quick “how to do x in y” language
2. Large scale refactorings that are mostly mechanical.
This still takes a bit of guidance to get the right output (and breaking down the refactoring an into multiple steps). But it does speed things up when I would touch 40-some files. I still review all the code.
I’ve found ai tools most valuable for:
1. Quick “how to do x in y” language
2. Large scale refactorings that are mostly mechanical.
This still takes a bit of guidance to get the right output (and breaking down the refactoring an into multiple steps). But it does speed things up when I would touch 40-some files. I still review all the code.
I would like to see this five-hour vibe coded platform and the commit history. Just curious.
I agree, to the point where I think this site should outright ban pro AI posts that don't show the code.
That would END all this spam because they NEVER show the code, and we all know why.
That would END all this spam because they NEVER show the code, and we all know why.
Why?
As the article points out, there is a huge market for selling these tools. And this is a forum of many developers.
When there is no actual code or full explanation of how the vibe coding process works, it seems like a straight up ad with no useful info.
That is my feeling at least. And I'm even open to these tools - my current assumption is that I am just bad at them. I'd like to see more examples of how to deliver on all these promises.
When there is no actual code or full explanation of how the vibe coding process works, it seems like a straight up ad with no useful info.
That is my feeling at least. And I'm even open to these tools - my current assumption is that I am just bad at them. I'd like to see more examples of how to deliver on all these promises.
It might be shilling, but it might also be someone just putting out low effort content to get traffic
I was actually referring more to what could be wrong with the code that they would want to hide?
I was actually referring more to what could be wrong with the code that they would want to hide?
Because the claims are either a lie or the code is atrocious. Showing the commit logs and a video would make it a lot harder to hide those facts.
I'm wary of these "I don't use an IDE anymore, just chat with claude-code".
It's like playing a game of blindfold navigation and you need to direct somebody to get somewhere by shouting commands "go left, no, no go right, my right".
I do use claude-code, but you can't ditch the code editor because you got to tweak things here and there. It gets very time consuming if you rely on claude for small stuff also. Merely the roundtrip to apply some small changes through Claude can take longer than just go and edit the file.
It's like playing a game of blindfold navigation and you need to direct somebody to get somewhere by shouting commands "go left, no, no go right, my right".
I do use claude-code, but you can't ditch the code editor because you got to tweak things here and there. It gets very time consuming if you rely on claude for small stuff also. Merely the roundtrip to apply some small changes through Claude can take longer than just go and edit the file.
Developers hate AI coding tools for the same reasons musicians dislike AI music and artists dislike AI art. Its the 'craft' problem as you say.
But an artist who values the end result over the craft is hardly an artist at all. They're a merchant at heart. The art is the product, and what excites them is shipping product.
For an artist at heart, however, the process is the product. Lucky for them, its about to become a lot more valuable.
But an artist who values the end result over the craft is hardly an artist at all. They're a merchant at heart. The art is the product, and what excites them is shipping product.
For an artist at heart, however, the process is the product. Lucky for them, its about to become a lot more valuable.
It's all about competition. Musicians dislike AI music because they're being outcompeted by AI. Same with artists. And, ironically, same with programmers.
AI programming isn't going to improve the state of the craft. Developers hate AI coding tools for the same reason they hated Dreamweaver back in the 00s: the generated code was crap. You'd spend more time "fixing" the generated code than you'd spend writing it from scratch.
What it is going to do is finally kill this obsession the tech industry has with moving fast and breaking things. We can't compete with AI on speed and breaking things. It's just not humanly possible. It's going to force the entire industry to find other metrics to compete on. I hope that's going to be quality (performance, uptime, and reliability), but, then again, I'm an optimist in this regard.
It's the same with music and art, too. AI isn't going to replace musicians and artists, but it is going to make them compete on different metrics than they're used to.
AI programming isn't going to improve the state of the craft. Developers hate AI coding tools for the same reason they hated Dreamweaver back in the 00s: the generated code was crap. You'd spend more time "fixing" the generated code than you'd spend writing it from scratch.
What it is going to do is finally kill this obsession the tech industry has with moving fast and breaking things. We can't compete with AI on speed and breaking things. It's just not humanly possible. It's going to force the entire industry to find other metrics to compete on. I hope that's going to be quality (performance, uptime, and reliability), but, then again, I'm an optimist in this regard.
It's the same with music and art, too. AI isn't going to replace musicians and artists, but it is going to make them compete on different metrics than they're used to.
That’s not quite right.
The tools are going to get better, and it’s going to lift everyone while doing so. But it will vary across task.
The bottom tier of engineers will have improvements to their code.
The top engineers will move faster. Top engineers will still be great at what they do.
—-
The same thing is happening in music. Everyone gets lifted but on different dimensions.
The tools are going to get better, and it’s going to lift everyone while doing so. But it will vary across task.
The bottom tier of engineers will have improvements to their code.
The top engineers will move faster. Top engineers will still be great at what they do.
—-
The same thing is happening in music. Everyone gets lifted but on different dimensions.
Unless there's a fundamental change in how the technology works, it'll never get to the point you're talking about. We're still waiting on our self-driving cars, remember.
Also, when it comes to music, who are the people in the rock and roll hall of fame? Those who can play the guitar real good? Or those who crafted songs that people love? AI will never be able to do the latter. But, AI can play power chords better than any session guitarist. That session guitarist is going to have to learn how to compete in another way than his guitar playing skill.
Also, when it comes to music, who are the people in the rock and roll hall of fame? Those who can play the guitar real good? Or those who crafted songs that people love? AI will never be able to do the latter. But, AI can play power chords better than any session guitarist. That session guitarist is going to have to learn how to compete in another way than his guitar playing skill.
I tend to agree, but at the end of the day I am providing for my family first and foremost.
This means having to knock out tasks each sprint, whether they tickle my fancy or not. If I can offload that work to the AI “agent”, then so be it.
I don’t feel the need to make my vocation a core part of my identity, so the time savings is worth more than elegantly crafted code or whatever other intrinsic value comes from a hand crafted solution.
This means having to knock out tasks each sprint, whether they tickle my fancy or not. If I can offload that work to the AI “agent”, then so be it.
I don’t feel the need to make my vocation a core part of my identity, so the time savings is worth more than elegantly crafted code or whatever other intrinsic value comes from a hand crafted solution.
What will happen to your salary in four years if that process works (which I doubt, but let's assume it)?
This is similar to the rage from 2010-2022 when developers, often at the behest of their employers, enthusiastically promoted the idea than everyone needs to learn how to code.
This is similar to the rage from 2010-2022 when developers, often at the behest of their employers, enthusiastically promoted the idea than everyone needs to learn how to code.
I’ll cross that bridge when I come to it. For now, I think the time horizon is long enough that I can make it to retirement.
Why is this work so hard for you to do in the first place?
“Craft” vs “delivery” is a false dichotomy. Someone who is “delivering” needs to understand what they are delivering! How can you support vibe-coded cruft that you don’t understand? This is why tools like TypeScript exist; TS by itself adds little in terms of functionality or in “deliverables”, but it makes the code more _understandable_. It’s worth the time to invest in making sure your types are correct because it makes you think about constraints. There are no shortcuts.
>Most developers hate AI coding tools
Not sure about that. If you go online and post anything disparaging AI you will definitely get a lot of support and +1's but my opinion is that there's a lot of noise from non-developers in these.
As in make a blog post "AI deleted my code" or "AI is worthless and slows me down, here's a very singular and specific example where this is true" and you'll be re-shared across the web with downvotes to the reasonable followup questions of "Just revert?" or "Don't use AI in that specific scenario?".
In the real world talking to peers i hear; "it's really good for getting out the boilerplate", "it's great for refactors", "it's pretty good at writing tests", "it one shot the UI implementation". Etc.
As in i hear measured praise and measured criticism for it that you just don't seem to get online. I guess that's true of any topic but AI and programming is something i know enough about to see the unreasonableness of these extremes.
Not sure about that. If you go online and post anything disparaging AI you will definitely get a lot of support and +1's but my opinion is that there's a lot of noise from non-developers in these.
As in make a blog post "AI deleted my code" or "AI is worthless and slows me down, here's a very singular and specific example where this is true" and you'll be re-shared across the web with downvotes to the reasonable followup questions of "Just revert?" or "Don't use AI in that specific scenario?".
In the real world talking to peers i hear; "it's really good for getting out the boilerplate", "it's great for refactors", "it's pretty good at writing tests", "it one shot the UI implementation". Etc.
As in i hear measured praise and measured criticism for it that you just don't seem to get online. I guess that's true of any topic but AI and programming is something i know enough about to see the unreasonableness of these extremes.
I was hoping that AI would lead to a golden age of testing, with devs focusing on making the best tests possible to determine if the AIs writing prod code were correct. Devs leading AI to correct code.
Unfortunately, expecting the AIs to make poop and the devs to wipe their butts only came half-true.
Unfortunately, expecting the AIs to make poop and the devs to wipe their butts only came half-true.
Have fun maintaining a slop blob.
It's not just about the craft. It's about understability, explainability and accountability.
It's not just about the craft. It's about understability, explainability and accountability.
Slop blobs happen without AI tooling though, unless everyone in the org is completely dedicated to their craft and on the same skill level.
> Craft-focused developers enjoy the process of writing code…
> Delivery-focused developers care about shipping products.
What the hell am I, then? I’m a craft-focused developer who cares about shipping products, and I like AI.
I find that AI doesn’t reduce my enjoyment of the craft. It reduces my need to type code with my fingers. I’m able to spend more time thinking above the code and put more time into getting a good design instead of getting something that just works. I’m not doing 100% AI generated code, though. I still find that there are lots of fiddly refactors that AI is too clumsy for.
I get the sense that most programmers don’t like to refactor. They enjoy the process of getting something working, but don’t like the process of making it good. To me, refactoring is the craft. That’s the part that I love. AI gives me more time for that.