Vibe Coding & the Value of Code

I vibe coded last night.

I think I need to go take another shower.

I didn't even read the code, let alone check it. It produced the output I wanted so (for now) I'm assuming it works as it should.... 🥴

Joking aside though, i didn't do anything complex or amazing. I asked Cursor to write me some code to clean the output from vmstat and merge it with a CSV for measuring power usage. Then plot it. I could have done it in Excel. In fact I started off doing it in Excel but it was faster (and more fun!) to vibe code some Python to do it instead. 

Thing is, the value of that code, even to me, has diminished. I could have written it by hand in an hour, and if I had done that, I would have stuck it on Github under an MIT license so the 1 other person in the world who cares about doing this might find it and use it. (Probably someone in Michael Ringenburg's team)

But it was 3minutes of prompting. And I'd have to put my name to the output. So it will probably just get binned once I have my data.

And that's a problem. At least I think it is.

We'll use a lot more energy regenerating that code again and again. More human time revalidating it rather than knowing something is tested and reliable.

There's a popular opinion that devs will be swimming in high paid work fixing LLM generated code in a year or two. What if we just decide we won't bother maintaining code anymore and will just rewrite it because that's "cheaper"?

All this coming from a confirmed AI skeptic... oh dear...

UPDATE

I just read some of the code it generated 🤮

Now I'll freely admit I'm a bit old school, i've spent my life working in environments where mistakes either cost lives (Aerospace) or lots of money (Finance), so lots of defensive programming is the norm, checking and handling of gracefully exiting in unknown or bad scenarios has to happen.

Even so this stuff is just BAD.

Seriously poor choices of data structures, reading the whole file into memory when it explicitly prompted not to so it can handle large files. Pointless comments everywhere making the code harder to read.

But it does "work". I manually checked one file (out of many) and the output looks correct.

So, while I may hate the code it wrote with my engineering hat on.... with my CEO/Founder hat on I really can't justify "fixing" that code or manually writing better code.

I am so conflicted on this!!

Some interesting comments on this on LinkedIn

I vibe coded last night. | Hamza M.

I vibe coded last night. I think I need to go take another shower. I didn't even read the code, let alone check it. It produced the output I wanted so (for now) I'm assuming it works as it should.... 🥴 Joking aside though, i didn't do anything complex or amazing. I asked Cursor to write me some code to clean the output from vmstat and merge it with a CSV for measuring power usage. Then plot it. I could have done it in Excel. In fact I started off doing it in Excel but it was faster (and more fun!) to vibe code some Python to do it instead. Thing is, the value of that code, even to me, has diminished. I could have written it by hand in an hour, and if I had done that, I would have stuck it on Github under an MIT license so the 1 other person in the world who cares about doing this might find it and use it. (Probably someone in Michael Ringenburg's team) But it was 3minutes of prompting. And I'd have to put my name to the output. So it will probably just get binned once I have my data. And that's a problem. At least I think it is. We'll use a lot more energy regenerating that code again and again. More human time revalidating it rather than knowing something is tested and reliable. There's a popular opinion that devs will be swimming in high paid work fixing LLM generated code in a year or two. What if we just decide we won't bother maintaining code anymore and will just rewrite it because that's "cheaper"? All this coming from a confirmed AI skeptic... oh dear... | 19 comments on LinkedIn

LinkedInHamza M.