[grokys]

My issue with AI-generated OSS contributions is:

If an AI improves developer productivity so much, why would maintainers of an OSS project want unknown contributors to sit in between the maintainer and the LLM? They'd be typing these queries into Claude Code themselves. To quote my colleague:

> We do not need a middleman to talk to AI models. We are not bottlenecked by coding.

[chenzhekl]

maybe you are not bottlnecked by coding. but there is high probability that you will be bottlenecked by verifying the correctness of LLM-generated code.

[Bridged7756]

Crazy how this doesn't register in people's heads. Has the real bottleneck ever been code written and not the review of code and everything involved? Understanding the nuance and implications behind design decisions; strategy.

In any REAL, workload, with good processes, code review makes speed of code generated a moot point. You still move as fast as you can review the code, and no, I won't debate that you can rely on LLMs, a deterministic language predictor, to determine the correctness of code; in the context of the business, and technical implications.

[grokys]

That is indeed the point I was making.

[amelius]

Where is the real bottleneck, if I may ask?

[solid_fuel]

> verifying the correctness of LLM-generated code

It's... pretty clear in the original conversation.

[saulpw]

I find that people who write "may I ask" are often/usually bad-faith arguers under cover of being polite.

[solid_fuel]

That's a good rule of thumb, it seems that way more often than not.

[notnullorvoid]

If you are a responsible maintainer you need to verify the correctness of the contribution wether you used an LLM to generate it or wether someone else did.

Having someone else be the AI-middlemen, just introduces additional complexity and confusion.

[gus_massa]

I'm almost not using AI, but a possible scenario is that the contributor spend like 20 hours in total.

Something like using the AI to get an initial bad version, make some tweaks to the prompt, make some manual fixes, ask the AI to fox something else, noticing some new related feature and asking the AI to add it, making some benchmarks and deciding to remove a small feature, or perhaps deciding between two similar implementations, add a few more manual fixes here and there, run the extended version of the automatic test and find a weird bug in the unusual setup, make a few fixes with the AI and manually. So after 20 hours of work, the final version has only 50 lines that have been rewriten like 5 times each. Now the mantainer can review only the final version in 1 hour or so.

This is very different to spending 5 minutes asking the AI to write a patch, that has 1000 lines that does not even compile and sending it to the maintainer without looking at it.

[gwbas1c]

I'm finding that AI, when successful, gives me 2-3x speedup. It's not the kind of thing I can give high-level instructions to like I can to a human.

I suspect the people who claim that AI works by only giving it high-level instructions are mostly working on "mindless" projects where a developer in the weeds wouldn't need to think very much.

[mexicocitinluez]

> If an AI improves developer productivity so much,

You're not suggesting the only metric of productivity is lines of code are you? And that the only benefit of using LLMs is for generating code you're too lazy to type yourself?

[eddd-ddde]

This reminds me of the critique of certain kinds of art.

"It's so easy, I could have done that myself"

Well yeah, but you didn't.