13 comments

[ 3.1 ms ] story [ 30.1 ms ] thread
With all due respect to the people who have studied this, I do not care. I'm relatively young, and yet the pain of typing has started to become unmistakeable and chipped away at my productivity, motivation, and sense of invincibility.

You could tell me AI coding makes me 50% slower. I'm taking it. I refuse to grind my wrists to dust.

> and yet the pain of typing has started to become unmistakeable

As in _physical pain_? If you haven't already, go to a doctor; that's not normal (it's somewhat common, but it is far from inevitable and there are things you can do about it).

If the task is large enough you will be faster w/ Claude Code, amp etc with current models.
I want to believe that this is true. But I don't.
No. Way. AI is much Faster than I will ever be at debugging, and idea generation.

It's not about working faster or slower, it's getting it right in the most efficient way possible.

I think it's because of the tasks they did. For run of the mill custom development projects AI certainly speeds things up a great lot. But when i tried to use it for comparatively hard tasks, i found it easier to do things by myself. It was nothing too fancy, just bitwise image manipulation - a custom code for pixel format conversion combined with image resizing that had to work a lot faster than ffmpeg's, but in a narrow set of conditions - no miracles here, it's not at all smarter than ffmpeg's, just specialised. 100% hand-made code took less time to build and was just as fast to execute. Of course, it used intrinsics.
In my experience, I don’t think hallucinations are a big problem anymore in terms of coding as long as you work within your domain of expertise.

The perception that AI tools make development faster is perhaps due to the part we spend a lot of time with thinking about how to write (like commenting) is solved instantly.

I think a lot of the delay is that it’s a new class of tool, and just like last gen IDE it takes a bit of getting used to and know where their strengths are, and know how to effectively fit it into your workflow.

The circumstances are exactly where AI is quite bad, and experienced humans very good - large, complex existing codebase, working on complex, possibly nuanced changes/fixes involving a lot of context, etc.
My personal experience is that in well and quickly autocompleted languages I know well (python) then I am roughly the same speed, I'd say...it's just a bit less annoying having AI do simple boilerplate for me, but sometimes annoying whenever I try to use it for larger refactoring where it gets style and structure incorrect.

In languages and libraries I know less well - vuejs+myriad of (especially) js libraries, I would say I'm much faster, especially as I delegate more style and structure to the AI.

Lots of HN discussion here - https://news.ycombinator.com/item?id=44522772 - it's an interesting paper but there are reasons we shouldn't make strong claims from it. I would say the most accurate reading would be "experienced developers but inexperienced Cursor users overestimate their immediate productivity gains".
Ironically, this fell off the HN front page without enough upvotes...neither hackernews.coffee nor Claude suggested it to me...
My personal experience: I shipped multiple features at work in the past 6 months that I simply wouldn’t have tried shipping otherwise, since my day job is mostly management. AI wrote maybe 80% of the code, I spent a bit of time rewriting some parts. No major bugs so far (ironically, the one big bug the team had to revert was done entirely by me)

I can guarantee I wouldn’t have shipped ANY of it, since it’d require focus blocks I simply don’t have on the job.

I’m also about to ship a Mac app that’s heavily vibecoded. I wouldn’t even try without AI, since I’m not a Swift developer.

Those aren’t “illusions” of performance. I imagine it’s hard to gauge every single scenario, and sensationalist takes like this research elicit an emotional response on the anti-AI crowd, but denying the impact is simply ignorance at this point…