Most active commenters
  • LPisGood(4)

←back to thread

AI agents: Less capability, more reliability, please

(www.sergey.fyi)
419 points serjester | 27 comments | 31 Mar 25 14:45 UTC | HN request time: 0.419s | source | bottom
1. joshdavham ◴[31 Mar 25 15:35 UTC] No.43536242[source]▶
My rule of thumb has thus far been: if I’m gonna allow AI to write any bit of code for me, then I must, at a bare minimum, be able to understand that code.

There’s no way I could do what some of these “vibe coders” are doing where they allow AI to write code for them that they don’t even understand.

replies(4): >>43536457 #>>43536559 #>>43537637 #>>43538600 #
2. kevmo314 ◴[31 Mar 25 15:56 UTC] No.43536457[source]▶
That's only true as long as you want to modify said code. If it meets your bar for reliability then you won't need to understand it, much like how we don't really need to read/understand compiled assembly code so we largely trust the compiler.

A lot of these vibe coders just have a much lower bar for reliability than you.

replies(2): >>43536626 #>>43538484 #
3. AlexandrB ◴[31 Mar 25 16:04 UTC] No.43536559[source]▶
I think there's a lot of code that gets written that's either disposable or effectively "write only" in that no one is expected to maintain it. I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain.

Basically, what's worse? "Vibes" code that no one understands or a cascade of 20 spreadsheets that no one understands? At least with the "vibes" code you can stick it in git and have some semblance of sane revision control and change tracking.

replies(6): >>43536629 #>>43536711 #>>43536898 #>>43537300 #>>43537330 #>>43537477 #
4. fourside ◴[31 Mar 25 16:09 UTC] No.43536626[source]▶
How do you know if it meets your bar for reliability if you don’t understand the output? I don’t know that the analogy to a compiler is apples to apples. A compiler isn’t producing an answer based on statistically generating something that should look like the right answer.
replies(1): >>43536977 #
5. Centigonal ◴[31 Mar 25 16:09 UTC] No.43536629[source]▶
> I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain.

I think this is a great use case for AI, but the analyst still needs to understand what the code that is output does. There are a lot of ways to transform data that result in inaccurate or misleading results.

replies(1): >>43536694 #
6. LPisGood ◴[31 Mar 25 16:15 UTC] No.43536694{3}[source]▶
Vibe coders focus on writing tests, and verifying function/correctness. It’s not like they don’t read _any_ of the code. They get the vibes, but ignore the details.
replies(3): >>43537314 #>>43537895 #>>43541938 #
7. pton_xd ◴[31 Mar 25 16:16 UTC] No.43536711[source]▶
> I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain.

That sort of makes sense, but then again... if you run some analysis code and it spits out a few plots, how do you know what you're looking at is correct if you have no idea what the code is doing?

replies(1): >>43537346 #
8. liveoneggs ◴[31 Mar 25 16:32 UTC] No.43536898[source]▶
two wrongs don't make a right
9. kevmo314 ◴[31 Mar 25 16:38 UTC] No.43536977{3}[source]▶
The premise for vibe coding is that it's generating the entire app or site. If the app does what you want then it's meeting the bar.
10. cube00 ◴[31 Mar 25 17:11 UTC] No.43537300[source]▶
> I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain

Considering the hallucinations we've all seen I don't know how they can be comfortable using AI generated data analysis to drive the future direction of the business.

11. cube00 ◴[31 Mar 25 17:12 UTC] No.43537314{4}[source]▶
> Vibe coders focus on writing tests

From the boasting I've seen, Vibe coders are also using AI to slop out their tests as well.

replies(2): >>43537365 #>>43540179 #
12. inetknght ◴[31 Mar 25 17:13 UTC] No.43537330[source]▶
> what's worse? "Vibes" code that no one understands or a cascade of 20 spreadsheets that no one understands? At least with the "vibes" code you can stick it in git and have some semblance of sane revision control and change tracking.

You can for spreadsheets too.

13. kibwen ◴[31 Mar 25 17:15 UTC] No.43537346{3}[source]▶
> how do you know what you're looking at is correct if you have no idea what the code is doing?

Does it reaffirm the biases of the one who signs my paychecks? If so, then the code is correct.

replies(2): >>43537964 #>>43538069 #
14. kibwen ◴[31 Mar 25 17:16 UTC] No.43537365{5}[source]▶
Worry not, we can solve this by using AI to generate tests to test the tests.
replies(1): >>43540188 #
15. palmotea ◴[31 Mar 25 17:26 UTC] No.43537477[source]▶
> think there's a lot of code that gets written that's either disposable or effectively "write only" in that no one is expected to maintain it. I have friends who write a lot of this code for tasks like data analysis for retail and "vibe coding" isn't that crazy in such a domain.

> Basically, what's worse? "Vibes" code that no one understands or a cascade of 20 spreadsheets that no one understands?

Correction: it's a "cascade of 20 spreadsheets" that one person understood/understands.

Write only code still needs to work, and someone at some point needs to understand it well enough to know that it works.

16. SkyPuncher ◴[31 Mar 25 17:42 UTC] No.43537637[source]▶
Sure, but you're a professional software engineer, who I assume gets feedback and performance reviews based on the quality of your code.

There's always been a group of beginners that throws stuff together without fully understanding what it does. In the past, this would be copy n' paste from Stackoverflow. Now, that process is simply more automated.

replies(1): >>43541810 #
17. hooverd ◴[31 Mar 25 18:05 UTC] No.43537895{4}[source]▶
Huh. The whole promise of vibe coding is that you don't have to pay attention to the details.
replies(2): >>43538058 #>>43540165 #
18. usui ◴[31 Mar 25 18:14 UTC] No.43537964{4}[source]▶
LOL thanks for the laughs. But yes seriously though, most kinds of data analysis jobs several rungs down the ladder where the result is not in a critical path amount to reaffirming what upper people believe. Don't rock the boat.
19. namaria ◴[31 Mar 25 18:24 UTC] No.43538058{5}[source]▶
"You're programming wrong wrong" /s
20. AlexandrB ◴[31 Mar 25 18:24 UTC] No.43538069{4}[source]▶
Lol, that's definitely a factor. Actually plotting is the perfect example because python is really popular in the space and matplotlib sucks so much. While an analyst may not understand Python very well, they often understand the data itself through either previous projects or through other analysis tools. It's kind of like vibe coding a UI for a backend that's hand built.
21. joshdavham ◴[31 Mar 25 18:59 UTC] No.43538484[source]▶
This is an interesting point and it's certainly true with respect to most peoples' attitudes towards dependencies.

For example, while I feel the need to understand the code I wrote using pytorch, I don't generally feel the need to totally grok how pytorch works.

22. __MatrixMan__ ◴[31 Mar 25 19:09 UTC] No.43538600[source]▶
I think there are times where it's ok to treat a function like a black box--cases where anything that makes the test pass will do because the test is in fact an exhaustive evaluation of what that code needs to do.

We just need to be better about making it clear which code is that way and which is not.

23. LPisGood ◴[31 Mar 25 21:26 UTC] No.43540165{5}[source]▶
Yeah, of course. I don’t think what I described could possibly be misconstrued as someone paying attention to details.
24. LPisGood ◴[31 Mar 25 21:27 UTC] No.43540179{5}[source]▶
Yeah and tests are much easier to validate than functions.
25. LPisGood ◴[31 Mar 25 21:28 UTC] No.43540188{6}[source]▶
Testing the tests is pretty much the definition of being a vibe coder.
26. __jochen__ ◴[01 Apr 25 01:11 UTC] No.43541810[source]▶
There is also likely to be increased pressure in a SE job to produce more code. You'll find that if others use AI, it'll be hard to be a hold-out and hit fewer delivery milestones, and quality is hard to measure. People are rewarded for shipping, primarily (unless you're explicitly working on high reliability/assurance products).
27. achierius ◴[01 Apr 25 01:34 UTC] No.43541938{4}[source]▶
This seems optimistic -- prescriptive, rather than descriptive. From what I've observed, vibe coders largely do not include tests; they do verify functionality/correctness, but not rigorously or repeatably.