Does AI Really Make Coders Faster?

[erek]

"Developers interviewed by MIT Technology Review generally agree on where AI tools excel: producing "boilerplate code" (reusable chunks of code repeated in multiple places with little modification), writing tests, fixing bugs, and explaining unfamiliar code to new developers. Several noted that AI helps overcome the "blank page problem" by offering an imperfect first stab to get a developer's creative juices flowing. It can also let nontechnical colleagues quickly prototype software features, easing the load on already overworked engineers. These tasks can be tedious, and developers are typically glad to hand them off. But they represent only a small part of an experienced engineer's workload. For the more complex problems where engineers really earn their bread, many developers told MIT Technology Review, the tools face significant hurdles...

The models also just get things wrong. Like all LLMs, coding models are prone to "hallucinating" — it's an issue built into how they work. But because the code they output looks so polished, errors can be difficult to detect, says James Liu, director of software engineering at the advertising technology company Mediaocean. Put all these flaws together, and using these tools can feel a lot like pulling a lever on a one-armed bandit. "Some projects you get a 20x improvement in terms of speed or efficiency," says Liu. "On other things, it just falls flat on its face, and you spend all this time trying to coax it into granting you the wish that you wanted and it's just not going to..." There are also more specific security concerns, she says. Researchers have discovered a worrying class of hallucinations where models reference nonexistent software packages in their code. Attackers can exploit this by creating packages with those names that harbor vulnerabilities, which the model or developer may then unwittingly incorporate into software.
Other key points from the article:

• LLMs can only hold limited amounts of information in context windows, so "they struggle to parse large code bases and are prone to forgetting what they're doing on longer tasks."
• "While an LLM-generated response to a problem may work in isolation, software is made up of hundreds of interconnected modules. If these aren't built with consideration for other parts of the software, it can quickly lead to a tangled, inconsistent code base that's hard for humans to parse and, more important, to maintain."
• "Accumulating technical debt is inevitable in most projects, but AI tools make it much easier for time-pressured engineers to cut corners, says GitClear's Harding. And GitClear's data suggests this is happening at scale..."
• "As models improve, the code they produce is becoming increasingly verbose and complex, says Tariq Shaukat, CEO of Sonar, which makes tools for checking code quality. This is driving down the number of obvious bugs and security vulnerabilities, he says, but at the cost of increasing the number of 'code smells' — harder-to-pinpoint flaws that lead to maintenance problems and technical debt."

Yet the article cites a recent Stanford University study that found employment among software developers aged 22 to 25 dropped nearly 20% between 2022 and 2025, "coinciding with the rise of AI-powered coding tools."

The story is part of MIT Technology Review's new Hype Correction series of articles about AI."

Source: https://developers.slashdot.org/story/25/12/20/2335253/does-ai-really-make-coders-faster

 

[emphy]

Does AI Really Make Coders Faster?

Wrong question.

If "ai" were such a great tool for coding, we would be seeing a marked increase in quality software releases. Pretty much crickets on that front. So far, I only see more and more software that, to put it in the kindest terms I am capable of, sucks balls.

 

[sharknice]

Wrong question.

If "ai" were such a great tool for coding, we would be seeing a marked increase in quality software releases. Pretty much crickets on that front. So far, I only see more and more software that, to put it in the kindest terms I am capable of, sucks balls.

People that wrote crappy code before can now write crappy code even faster.

I know this from experience.

The code it generates is never really how a professional would leave it, but it does still speed things up for me.

 

[mnewxcv]

I never coded before ai, but have since been able to make a couple crappy programs that I can use for basic tasks.

 

[erek]

I never coded before ai, but have since been able to make a couple crappy programs that I can use for basic tasks.

how good is it

 

[GoldenTiger]

People that wrote crappy code before can now write crappy code even faster.

I know this from experience.

The code it generates is never really how a professional would leave it, but it does still speed things up for me.

Ditto. If you already know what you're doing, it helps. If you don't, well...

 

[cthulhuiscool]

we would be seeing a marked increase in quality software releases

I think you must be under the mistaken impression that you don't live in a slop economy. Success looks like giant pits of inexpensive monochrome sludge that's technically edible, not carefully crafted artisanal cuisine. Yes that's a metaphor.

 

[philb2]

I never coded before ai, but have since been able to make a couple crappy programs that I can use for basic tasks.

I use AI to write PowerShell scripts, some of them quite long

 

[mnewxcv]

how good is it

I used qwen3 coder and had much more luck than using chatgpt at the time. Chatgpt was giving me broken code, while qwen3 coder basically gave me fully functional code. Only needed some minor tweaks. I made a program to batch download public domain assets just by giving it a keyword and qty of assets I want. Simple but it works and I would've never been able to make a program with gui without ai doing it for me.

 

[socK]

So, the things I think it helps with are:

1. Design is hard, super hard. Having a well designed public API or even just how you structure your code is hard. I think it's effectively one of the hardest aspects of writing code. Writing the actual implementation might be a simple affair. Realizing it won't work because you missed some aspect of how things tie together is a back to the drawing board time waster.

It's really good for rubber ducking your way through some of this shit and getting your head wrapped around something before you get started.

Honestly fuck Stack Overflow at this point. I'm zero percent surprised AI has basically killed the site because at this point it's was becoming completely insufferable. Nothing like finding someone with the same problem and it was closed as a duplicate by some complete clown shoe asshole because they thought it was 0.4% related to something from eleven fucking years ago.

2. Basically autocompleting shit I was literally about to write. Doing some minor refactoring? Changing something in one or two spots will probably get it to autocomplete the remaining spots. It's always nice when it reads your mind and you lop off 15 seconds of work and all you needed to do was hit tab.

Vibe coding giant swaths of shit is dumb as fuck and it's always when something goes off the rails. I'm not surprised that idiots pull a SaaS out of their ass in a weekend only to find it hacked because they have a SQL injection or don't validate bearer tokens for their API correctly and you can just grab other tenant/user data trivially.

If you keep the focus tight (which is why I said rubber ducking with it is nice, because generally you're focusing on a very particular thing) then generally you will get very good results. And even then you need to have some idea what you're doing because it WILL miss things. Which, again, you usually hash out as you're bouncing ideas off it.

I don't think I've ever even comitted a single function written entirely by AI. I will never merge anything if I can't understand it.

 

[sharknice]

So, the things I think it helps with are:

1. Design is hard, super hard. Having a well designed public API or even just how you structure your code is hard. I think it's effectively one of the hardest aspects of writing code. Writing the actual implementation might be a simple affair. Realizing it won't work because you missed some aspect of how things tie together is a back to the drawing board time waster.

It's really good for rubber ducking your way through some of this shit and getting your head wrapped around something before you get started.

Honestly fuck Stack Overflow at this point. I'm zero percent surprised AI has basically killed the site because at this point it's was becoming completely insufferable. Nothing like finding someone with the same problem and it was closed as a duplicate by some complete clown shoe asshole because they thought it was 0.4% related to something from eleven fucking years ago.

2. Basically autocompleting shit I was literally about to write. Doing some minor refactoring? Changing something in one or two spots will probably get it to autocomplete the remaining spots. It's always nice when it reads your mind and you lop off 15 seconds of work and all you needed to do was hit tab.

Vibe coding giant swaths of shit is dumb as fuck and it's always when something goes off the rails. I'm not surprised that idiots pull a SaaS out of their ass in a weekend only to find it hacked because they have a SQL injection or don't validate bearer tokens for their API correctly and you can just grab other tenant/user data trivially.

If you keep the focus tight (which is why I said rubber ducking with it is nice, because generally you're focusing on a very particular thing) then generally you will get very good results. And even then you need to have some idea what you're doing because it WILL miss things. Which, again, you usually hash out as you're bouncing ideas off it.

I don't think I've ever even comitted a single function written entirely by AI. I will never merge anything if I can't understand it.

I feel like the autocomplete is the most useful feature for real production code, it's just much better autocomplete than any other we've had before. Also small modifications can be good, where you just ask the AI, hey why is this bug happening and sometimes it can give you a solution.

The full on generation "vibe coding" is terrible for production but awesome for throwaway projects. I would never ever submit that code to production without completely reading through it all and probably refactoring all of it.
But that same code is great for throwaway code just quickly making something you don't care about having to maintain or anything. It's a huge game changer for people that don't know how to code because it gives them the ability to actually create things they never would have been able to. They just have to realize their limits and not vibe code a payment processor or something.

 

[Tweak155]

I think, in short, AI is a powerful tool for development for someone who already knows what they're doing. For those that do not know what they're doing, they're fooled by results that might look good. But most likely, it's like having a nice looking car with a rubber band engine... one day it's going to break.

And ultimately, if it does work, the "developer" learns minimal to nothing because AI did it for them. There will always be a limit to the size of the project they can build or work on. There needs to be learning. Typically, the only learning being done is figuring out how to get AI to fix stuff... but it just makes them even more dependent on AI.

I'm sure there are cases where someone actually learns development by asking AI questions about development, rather than just trying to get it to do the work. But I suspect that's rare.

 

[LukeTbk]

"Some projects you get a 20x improvement in terms of speed or efficiency," says Liu. "On other things, it just falls flat on its face, and you spend all this time trying to coax it into granting you the wish that you wanted and it's just not going to..."

that describe well my experience, tooling and practice will need to get better and company to learn which is which well enough (and reduce the second group), this report from google do sound about right has well:
Google Cloud's 2025 DORA Report, released Wednesday, shows that 90% of developers now use AI in their daily work, a 14% increase from last year., only 24% of respondents actually trust the information these tools produce
The company's own metrics show that more than a quarter of Google's new code now springs from AI systems, with CEO Sundar Pichai claiming a 10% productivity boost across engineering teams.
Developers treat AI like a brilliant but unreliable coworker—useful for brainstorming and grunt work, but everything needs double-checking.


For some things like code review it could easily get universally at least useful for a pre-run. I imagine a google is already quite good at it, has the best tools, great coders, 10% seem both great and more realistic on what it will be (if that get 1-2% better yearly in a incremental way it will be quite something)

I think, in short, AI is a powerful tool for development for someone who already knows what they're doing.

it can also be quite powerful if you needs thing that you do not know about that are temporary or non-essential/something that will core to maintain, make a small casual game and shaders rendering of some effects is not core on maintaining the game and I known nothing of glsl, you can use AI for that.

And that can become a bit of its small project inside the big one if it is not a everything need high performance game. Anything that a will leave soon and do not code clearly nor documents intern could have done stuff and happy if it just work right now or anything that can be validated extensively.

And it can be great for software for yourself, a tool that you make only you will use, I would love some application/script that would do x-y-z with my recorded golf game score cannot find one that exist, name a long list, start from scratch, small thing that all fit in context in python/html/etc... it can be quite good.

 

[M76]

People that wrote crappy code before can now write crappy code even faster.

I know this from experience.

The code it generates is never really how a professional would leave it, but it does still speed things up for me.

Code not always needs to be perfect, it just needs to work, often only once.

99% of my coding is done to accomplish tasks that are parts of a specific project and they are rarely if ever used again. Apart from me copy pasting snippets from it into another program that does something vaguely similar.

The most depressing thing in my career is when I come up with an elegant, user friendly solution to a problem and then it is never used because someone altered the deal, or worse the deal never even happens.

At least with AI I waste much less time, and I still get to exercise my brain fixing the fuckups of the AI.

 

[HeadRusch]

AI is your never-tired, always eager high school graduate intern that will never, ever learn anything more than it already thinks it knows. Will this equation change? Sure. Will it matter? Depends on the application.

I was asking it not too long ago about basic options for a new car I was (reluctantly) pricing-out.......it was never once 100% correct, but it presented the information each time as if it was. It was helpful, but ultimately, I had to go to the manufacturers website to find out *for sure* what was included in a given upgrade package. But, for awhile, it was fun to engage with.

 

[Randall Stephens]

I never coded before ai, but have since been able to make a couple crappy programs that I can use for basic tasks.

And that’s great, and useful, and perfectly fine. Legit no sarcasm. This wholesale it’s going to replace coders is just a scare tactic to drive down wages for computer science folks

 

[SecretStash]

Remember code reviewing one of the offshore guy's code. It didn't look like what he normally did and it had a lot of regex which is a pain to review, but the format looked aesthetic. Yeah scary times ahead when it's going to be harder to spot bad mistakes and devs will be able to pump out huge volumes generated code. If the guy isn't great that's a lot more work checking it all.

 

[ChadD]

The problem with training AI.... is it ends up generally being an average of what its trained on. It has no mechanic to realize what is high art, the code that is almost perfect.
IMO AI codes and scripts and such. Are a average quality. Better then the worst, no where near the best.
I'm sure it speeds up bad coders ability to produce mediocre code, and it might speed the workflow a bit for an average coder. A good coder would probably spend more time swearing at its crap generation and fixing it.

 

[Lifelite]

I've said it several times....AI is a productivity multiplier.

If you're a damn good engineer, you'll be able to use AI very effectively. If you're barely able to code a for loop, you're not going to get anywhere good by just adding AI to the mix.

There's a second side to this too...AI is also a deficciency multiplier. Any mistakes or ineptitudes are also multiplied.

As with any tool, there are pros and cons. I absolutely love AI for, say, setting up boilerplate code such as interfaces, etc rather than having to type by hand...but for any actual functionality it seems to be SLOWER than just coding it myself.

Plus any dev with experience can tell you; what takes the longest in coding is considering the impact of your changes...and thinking about scalability and maintainability of what you build. AI just does not do that well.

 

[uOpt]

For me it is decidedly mixed. It can best be explained by my attempts to have Gemini 3 Pro convert code from one language to another.

• Converting a C program to C++, with the intention of using modern C++ memory handling and library, was an embarrassing failure. The resulting program did not even work for its basic functions. And it didn't do a reasonable job of coming up with good C++.
• Converting a C++ program with basic C++ usage and mostly POSIX system calls to Rust resulted in a good looking and working Rust program. The program was more complicated than the previous example, too. However, Gemini did not recognize that it could use one of the Rust standard library core functions and reimplemented it naively.
• Attempts at smaller modifications in existing programs all led to code I wouldn't commit to a source repository because there were too many additional, non-functional source code changes. This can possibly be fixed with expanded prompt engineering, but at that point it is quicker to do the small code changes myself.

 

[paradoxical]

For me it is decidedly mixed. It can best be explained by my attempts to have Gemini 3 Pro convert code from one language to another.

• Converting a C program to C++, with the intention of using modern C++ memory handling and library, was an embarrassing failure. The resulting program did not even work for its basic functions. And it didn't do a reasonable job of coming up with good C++.
• Converting a C++ program with basic C++ usage and mostly POSIX system calls to Rust resulted in a good looking and working Rust program. The program was more complicated than the previous example, too. However, Gemini did not recognize that it could use one of the Rust standard library core functions and reimplemented it naively.
• Attempts at smaller modifications in existing programs all led to code I wouldn't commit to a source repository because there were too many additional, non-functional source code changes. This can possibly be fixed with expanded prompt engineering, but at that point it is quicker to do the small code changes myself.

I'd be curious to have you give codex with gpt 5.2 high model selected a shot at #1, I have had great luck with it in C++. Part of the issue right now is that people are getting variable results because there is such a big difference between the models. Codex can't do front end design, Claude can't do sophisticated performant code, etc.

 

[uOpt]

I'd be curious to have you give codex with gpt 5.2 high model selected a shot at #1, I have had great luck with it in C++. Part of the issue right now is that people are getting variable results because there is such a big difference between the models. Codex can't do front end design, Claude can't do sophisticated performant code, etc.

I'll give codex a try.

 

[Nytegard]

The problem with training AI.... is it ends up generally being an average of what its trained on. It has no mechanic to realize what is high art, the code that is almost perfect.
IMO AI codes and scripts and such. Are a average quality. Better then the worst, no where near the best.
I'm sure it speeds up bad coders ability to produce mediocre code, and it might speed the workflow a bit for an average coder. A good coder would probably spend more time swearing at its crap generation and fixing it.

It really is all about what it's trained on. And at work, I've been blocked for about a month, where my boss tried using AI to write a simple function to help speed things up. It's not that the function is long (it shouldn't be), nor is it unavailable (there are math equations online). But the math isn't exactly simple, and the AI failed spectacularly.

(My boss doesn't have the necessary math skills for this, thus tried using AI, and we're waiting on the people with the math background.)

 

[paradoxical]

It really is all about what it's trained on. And at work, I've been blocked for about a month, where my boss tried using AI to write a simple function to help speed things up. It's not that the function is long (it shouldn't be), nor is it unavailable (there are math equations online). But the math isn't exactly simple, and the AI failed spectacularly.

(My boss doesn't have the necessary math skills for this, thus tried using AI, and we're waiting on the people with the math background.)

Wrong AI or your boss is an idiot. I've had codex write complex DSP algorithms for simultaneous time and frequency domain analysis on realtime high samplerate streaming. Worked great.