I’ve been trying full on vibe coding the last few weeks to see where it’s at. I don’t feel any less exhausted at the end of the day.
So ai sharing time. I was at a ai hands on for a monitoring company. One guy was going around showing off this thing. Its a dashboard and he made it by telling the ai to look at a type of data and then asking it to find out what the data was about and then asking it to put together a dash based on the datas applications. The result was a real nice dash. So I say. But do you understand the dashboard. Could you use it. He was like. Well presumably people who know about the particular data will be able to use it. So the only thing he actually knows about the dashboard is its pretty.
alias what-the-fuck-is-in-this-room=lsIt might be longer but it’s definitely more memorable to type.
alias REPORT!=ls honestly not shure if the bang would work.
They absolutely save you wall-time, huge amounts of it. At least if you’re competent.
Not a single study representative of normal tasks has shown that to be the case.
I’m absolutely fine with moving 10 times faster than
othersI was moving before, and people thinking they’re doing better than me at the same time. So I’m not complaining.I could be wrong, but everything I see in my own worklife suggests that I’m not. The performance increase is immense. I genuinely cannot understand why the results seem to disagree with this. I’ve always been a bit ahead of the curve, so perhaps that’s the reason now as well.
I’m happy for you. From what I’ve seen, that observation correlates highly with one or both of 1. a poor evaluation of “productivity”, 2. rather low productivity to begin with.
The first category is by far the worst, as they think they’re being extremely productive. But the burden is simply shifted onto someone else. No doubt that the authors of each of these vulnerability reports considered themselves to be immensely productive: https://gist.github.com/bagder/07f7581f6e3d78ef37dfbfc81fd1d1cd
Producing 10 times more garbage, is a hefty net negative.
Could be our team composition that’s the magic sauce: we’re 4 seniors with 20+ years of experience from software development each and just one who’s more like a junior. We could be doing this like artisans of pre-2025 but we’re in a bit of a hurry, being a startup. Also, using AI seems to give us quite a bit more free time.
Remains to be seen, I guess, whether it’s garbage we’re producing. But we have completed in 6 months what I would have assumed would’ve taken 2 years and a bigger team.
Guy at work did a whole mini project with just LLMs and prompting. I asked him some questions about how it works and some implementation details, and he had no idea. Great. I’m going to have to maintain this thing, probably.
A guy vibe coded something and said to incorporate it into my work.
Now it was a feature that people had asked for so I had to try it out.
It failed 75% of the things it was supposed to do and for the other what usually was a near instant interactive task took 5 minutes. I kicked it back saying he needed to fix the problems and improve performance. The end of the next day he answered that the infrastructure must be broken because the AI couldn’t get results and the performance problem was just the nature of the things the software had to interact with. I say “he”, but pretty sure his comment was LLM generated, long and useless.
But it was the impetus to get that function done now, as his “substantiative” work meant we could technically provide a customer request, lower priority as it may have been. So I spent a morning implementing the same thing it did but the old fashioned way, 100% worked and the unavoidably slow thing took less than a second.
If you’re technical, spec kit helps with this a lot.
You define the spec, and declare everything important about how it needs to work.
Then let the ai loose on developing the code. Then I have another ai check for accuracy to the spec, and have it walk me through the code at a high level.
I understand it about as well as I do most projects where there is a team working on something.
Works great if you are already a competent coder. If I was new to coding though, I don’t think I would be learning a ton
I feel like most of the time saved is by skipping the part where you learn stuff. Like, the AI fills out how you do a left join with this ORM library. Cool. Now I don’t know how to do that.
You know how a lot of managers are annoying and don’t know anything about how shit works? That’s down the road using LLMs like this.
I don’t disagree. That’s why I feel for new developers. I’ve been doing this for 25 years, and I have some stuff still to learn, but not all that much on a vibe coded saas app.
If I had to learn like this when I was new though, I probably would have never learned anything
I feel this way to.
Just use a LLM to maintain it. Duh.
If you ask AI to create the shell of a class for you with 3 variables (two ints and one Vector as you defined elsewhere in your project), that’s different than saying “make me Facebook 2” and expecting everything to go well.
Ethics of current LLMs in a still-capitalist-materially society aside, the problem with AI is that management wants the robots to do all the work so their jobs become easier and their stocks go up. But that shit just doesn’t DO that. LLMs don’t think. They are statistical autocomplete. Programming is not trivial, no matter how much it may seem that way because we have trivialized as much of it as possible to make it useful.
“They are statistical autocomplete.” nice
But doesn’t pressing the up arrow 10 times (and reading each command) waste both time and effort compared to just pressing 2 buttons?
Stretching an additional finger, as well as locating an additional key, could be considered extra effort.
Sure but both quantified and qualified, those 4 actions are less effort than the 20 actions of key pressing and reading the terminal.
Yes, which is a perfect comparison
except instead of
lsit runsrm -rf /-y-frfr
Didn’t steam actually do that for a bit on some systems when you tried to uninstall?
Why did you prompt for that?
To test the --no-preserve-root safeguard. It’s like playing Russian roulette
…pressing the ‘up’ arrow ten times in a shell might let you avoid typing ‘ls’
I feel seen. But in my defense, that directory was like, 6 levels deep
Ctrl r my friend
Let me introduce you to my savior: ctrl + r
And a better bash with fuzzy find: fish or zsh.
fzf is where it’s at
Yeah, that what you can type
^R lsinstead oflsto save time!Pretty sure I use
ls <dir>more than lsAlso, use
mcand you’ll never need ls
Thought it was going to be ctrl+r
I didn’t know that one for years.
(Searches command history.)
Yeah, pair that with fzf!
Yes, and you can make a door by driving a car through a wall. The outcome may be less than optimal though.
That’s a doorway, not a door
