Hacker News Reader

Comment by victor9000 | Hacker News Reader

This gets at the absolute torrent of LLM diarrhea that people are adding to PRs these days. The worst of it seems to come from junior and first time senior devs who think more is more when it comes to LoC. PR review has become a nightmare at my work where juniors are now producing these magnificent PRs with dynamic programming, esoteric caching, database triggers, you name it. People are using LLMs to produce code far beyond their abilities, wisdom, or understanding, producing an absolute clusterfuck of bugs and edge cases. Anyone else dealing with something similar? How are you handling it?

View on HN

How is that different from junior devs writing bad code previously? The more things change, the more things stay the same.

You handle it by teaching them how to write good code.

And if they refuse to learn, then they get bad performance reviews and get let go.

I've had junior devs come in with all sorts of bad habits, from only using single-letter variable names and zero commenting, to thinking global variables should be used for everything, to writing object-oriented monstrosities with seven layers of unnecessary abstractions instead of a simple function.

Bad LLM-generated code? It's just one more category of bad code, and you treat it the same as all the rest. Explain why it's wrong and how to redo it.

Or if you want to fix it at scale, identify the common bad patterns and make avoiding them part of your company's onboarding/orientation/first-week-training for new devs.

> How is that different from junior devs writing bad code previously?

Because if it's bad, at least it's simple. Meaning simple to review, quickly correct and move on.

Like I said, with "object-oriented mostrosities", it's not like it was always simple before either.

And if you know a solution should be 50 lines and they've given you 500, it's not like you have to read it all -- you can quickly figure out what approach they're using and discuss the approach they should be using instead.

Maybe fight fire with fire. Feed the ChatGPT PR to ChatGPT and ask it to do a review, paste that as the comment. It will even do the markdown for you!

My company simply prohibits any AI generated code. Seems to work rather well.

My employer went all in and pays for both enterprise subscriptions (github Copilot+ chatgpt enterprise, which is just a company branded version of the regular interface)

We've even been getting "prompt engineering" meeting invites of 3+ hours to get an introduction into their usage. 100-150 participants each time I joined

It's amazing how much they're valuing it. from my experience it's usually a negative productivity multiplier (x0.7 vs x1 without either)

Sounds like they plan to lay half of you off once metrics stabilize.

How is this enforced? I'm not saying it isn't a good idea, just that it seems like it would be tricky to enforce. Separately, it could result in employees uploading code to a non-privacy-respecting AI, whereas if employees were allowed to use a particular AI then the company could better control privacy/security concerns.

When it gives me a complicated list expression or regex or something I like to ask ChatGPT to find a simpler way of doing the same thing, and it usually gives me something simpler that still works. Of course, you do have to ask, rather than simply copy-paste its output right into an editor, which is probably one step too many for some.

Why not deal with people who create problems like this the same way as one would have done four years ago?

If they're not doing their job, then why do they still have one?

In the code review, can't you simply say, "This is too complicated for what you're trying to do -- please simplify"?

Not quite the same but might be more relevant depending on context: "If you can't articulate what it does then please rewrite it such that you can."

Are they dealing with complexity which isn't there in order to appear smarter?

IMO they're just impressed the AI came to a conclusion that actually runs and aren't skilled enough to recognize there's a simpler way to do it.

Maybe time for some pair programming?

No.

That would be interesting though. What happens when one programmer attempts to chatgpt in pair programming. It's almost like they're already pair programming, just not with you!

They are welcome to do so, but not on company time. We do not find those tools useful at all, because we are generally hired to write new stuff and ChatGPT or other tools are useless when there are no good examples to steal from (e.g. darker corners of AWS that people don't bother to offer solutions for) or when there is a known bug or there are only partial workarounds available for it.

Don't you want programmers to familiarize themselves now to prepare for the time when it does? Claude 3.5 sonnet is getting close.

1954: "Don't you want power plant operators to familiarise themselves with this tokamak now to prepare for when it actually works?"

(Practical nuclear fusion has 10 years away in 1954, and it's still 10 years away now. I suspect in practice LLMs are in a similar space; everyone seems to be fixated on the near, supposedly inevitable, future where they are actually useful.)

AI and VR follow the same hype cycles. They always leave a trail of wasted money, energy, and resources. Only this time Gen AI is leaving a rolling coal-like trail of bullshit that will take some time to clean up. I am patiently waiting for the VC money to run out.

I work for clients that do not allow this shit, because their security teams and lawyers won't have it. But... they have in-house "AI ambassadors" (your typical useless middle managers, BAs, project managers, etc.) who see promoting AI as a survival strategy. On the business side these orgs are leaking data, internal comms, and PII like a sieve, but the software side is free of AI. For now.

Reply on news.ycombinator.com