vibe coding is a ponzi scheme of tech debt

[exchgr@mastodon.world]

vibe coding is a ponzi scheme of tech debt

[angermcs@hachyderm.io]

@exchgr Do you want to fix this debt, or double it for the next engineer?

[tartley@fosstodon.org]

@exchgr This is so frakking good.

[swelljoe@mas.to]

@exchgr this might be true (despite my astonishment about how well the best models actually work), but in my experience a lot of human developers are also running a Ponzi scheme of tech debt, hopping from job to job leaving a mess in their wake. I suspect AI will accelerate that. Good developers will write good code faster. Bad developers will write bad code faster.

[rubenerd@bsd.network]

@swelljoe @exchgr

> Good developers will write good code faster

The evidence doesn't really support this.

https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

[swelljoe@mas.to]

@rubenerd @exchgr yeah, I read all the studies and felt quite smug about AI, as well. And, then I actually built some stuff and found some bugs with help from current gen frontier models, and my priors were upended.

I'm not saying I like it, as it's going to cause a tremendous amount of disruption, and not in a good way, given who holds every leadership position in government and industry right now. But, I simply can't pretend it doesn't work, anymore, because I've seen it with my own eyes.

[swelljoe@mas.to]

@rubenerd @exchgr I was dragged into it by my employer, as I like having health insurance, but it works. Over the holidays I built a bunch of stuff (an absurd amount of stuff), more working code than I've ever written in such a short time in my life. I wanted a weather app without ads, so I built one in a couple of hours. https://wthr.lol/ (And, if you're curious about code quality, it's here: https://github.com/swelljoe/wthr.lol )

I've used it to find bugs in huge projects and build from scratch.

[swelljoe@mas.to]

@rubenerd @exchgr I could have built the weather app in...maybe 3x-4x the time it took? But, I probably wouldn't have, because I didn't want it three or four times more than that amount of effort.

I've got two other much larger projects that haven't really launched yet, that have also taken remarkably less time than I would have required doing it myself.

I hate being "rah rah AI", but I'm not going to lie on the internet about it when I know it's gotten really good at writing code.

[swelljoe@mas.to]

@rubenerd @exchgr also, note that weather app has all the things I probably wouldn't have bothered doing on a small throwaway project because they're tedious: Good test coverage, every static analysis option possible, nice packaging and automatic deployment to the server when tagging a release via Github actions. All that would have taken me days to write by hand, even though that's exactly the kind of stuff I do every day at work, versus a few minutes of prompting and occasionally correcting.

[exchgr@mastodon.world]

@swelljoe @rubenerd counterpoint: i was trying claude code with opus 4.6 today at work for the first time on a hefty refactor, and it was…fairly slow and tedious to use. while it certainly can use tools to analyze and spit out quite a volume of code fairly quickly, i’m also accounting for time lost coercing it to correct its own errors, of which there were more than a few, and which a less experienced dev would certainly have missed

[exchgr@mastodon.world]

@swelljoe @rubenerd in the end it did what i wanted, but it took all afternoon and a bit into the evening to do what i probably could’ve done in a couple-few hours in intellij

[exchgr@mastodon.world]

@swelljoe @rubenerd there was a constant feeling of “it’s so close!” and then every time i’d read the diff as i was committing, i’d spot another area where it had missed something, or done it in a weird way

[swelljoe@mas.to]

@exchgr @rubenerd yeah, I've had that experience as well, though less often as I get accustomed to its limits and ways to encourage sane behavior the first time. It's not smart, but it has infinite patience. If you give it clear success criteria it'll hammer on it until it's right, and it can even do it without hand-holding, if you give it a sandbox (I have a couple of VMs, one Rocky and one Ubuntu, just for letting LLMs run wild) and run it with --dangerously-skip-permissions or --yolo, etc.