Not to generalize too much, but the need to be right is sometimes a key aspect of an autistic person's personality. That's me, though some of it I'm sure is my step-father's insistence that I was wrong about everything. We don't need to unpack that here.
I've written a bunch about how disappointed I am with the coding output of LLM's, and I feel like it has really plateaued in the last six months. I find myself telling it to remember SOLID and DRY principles, shaking my fist at it like a grumpy staff engineer toward an intern. (Not my previous interns, y'all were awesome!) I tell the robots that I'm right, and they're wrong, a lot.
Lately, I've had to write a few super technical, almost legal docs, and figured I'd send them through Claude for polish. Keep in mind, I'm an author, I wrote a book once. It was complimented as being conversational, which feels good since it was a programming book. I have a style. I think people value that (or at least react emotionally to it, even if negative).
So Claude is good at tightening up the facts, totally. It sucks at tone. I mean, it only wants to be right. I have to constantly tell it, "Dude, I have to soften that because it's too abrasive." You know it responds... "You're right, I didn't consider the audience or the need to be aware of any sensitivities to the topic." Sounds exactly like a coding conversation.
When I was writing for my college newspaper, decades ago, my opinion column was about being right, and getting in some "zings" for good measure. That was decades before I was diagnosed with ASD, but it wouldn't have mattered. I needed to learn how to communicate effectively, and that took awhile.
This is where I say for the hundredth time that LLM's are word guessers that lack wisdom and judgment. It's that reason that AI is so easy to spot. If it's for writing in spoken language, it's trained on a lot of not-ideal speech. If it's for code, it's trained on a lot of garbage code. That's why human writers and software people are as important as ever. Humanity can't be algorithmically defined. After hundreds of years of science, we can't even define it ourselves.
I just hope we keep training younger folks to write code and prose. Especially code though. Software is bad enough now. The robots can help us make more software, but they're not at all good at making it better.
No comments yet.