alt Hacker NewsI think I strongly disagree with your first sentence. Strongly. I've worked with enough people with fine coding skills and terrible English to know that the two are not an uncommon combination. Last year, I worked with someone whose first language was something from China - no idea which flavour - and they would often pass their written comms through an LLM. I could tell when they had done so by the onset of sudden grammatical correctness. But I don't think they were doing this with the intention of deceiving, merely of communicating clearly and correctly.
Now, I often find Claude's idea of idiomatic prose to be a bit load-bearingly seam-hitting as it lands not this point, but THAT one, but it is probably better than something hacked out by a person who confuses tenses, cases, pronouns and when they can enverb a noun. So I tend towards giving people the benefit of the doubt on this.
Lastly, if you think LLM-written software is full of security vulnerabilities and bugs, I have terrible news for you about the state of human-written code. The fact that what we do is often better than nothing at all, is no big recommendation.
> I've worked with enough people with fine coding skills and terrible English
What about people with fine English and terrible coding skills?
What about people with both terrible English and terrible coding skills?
> Now, I often find Claude's idea of idiomatic prose to be a bit load-bearingly seam-hitting as it lands not this point, but THAT one, but it is probably better than something hacked out by a person who confuses tenses, cases, pronouns and when they can enverb a noun. So I tend towards giving people the benefit of the doubt on this.
Personally, I find AI writing as insufferable as anyone else (though Anthropic's is particularly bad, maybe just due to my familiarity with it), but I wouldn't judge individuals for using it to make communication more readable, rephrase what they mean etc. If anything, any difficulty in reading my prose would support that.
> Lastly, if you think LLM-written software is full of security vulnerabilities and bugs, I have terrible news for you about the state of human-written code. The fact that what we do is often better than nothing at all, is no big recommendation.
I guess a lot depends on how you use the LLMs (I bet horrible coding skills coincide with horribly lazy and problematic usage of LLMs for development too), but I wonder how humans actually stack up to the slop-machine when it comes to how good or bad the code they produce is on average, since I'm sure that SOTA model code by now tends towards the upper end of that.
In other words, if a person wants to limit themselves to avoiding LLM code over fears of it being bad, they would also have to avoid any any all code written by people that essentially produced the problematic training dataset, as well as any programmers with the same capabilities (or lack thereof). You'd basically have to avoid using a lot/most of the software out there if that's your quality standard.