Hacker News
points by zzzeek 4 days ago | 0 comments
you can totally tell Claude to take some AI technical writing (which is, to be clear, writing where you explicitly need to not sound creative or interesting, and you need to maintain very strict "voice" [a term I only even know because claude explained it to me] so this is not a creative endeavor) and make it sound "less AI". At least working with opus 5.5, textual output already has a lot fewer of the colons and "load bearing" of a few months ago, and if a PR does have that kind of lingo I can just have Claude rewrite it "like a human" and it does great.
[–] citelao 4 days ago
OP: I'm not surprised you/other people are able to get better output than me.

And I agree that "less AI" edit passes reduce a lot of the hallmarks, but I think these models truly struggle with expressing complicated ideas. Have Opus been good at that for you?

[–] zahlman 4 days ago
> I'm not surprised you/other people are able to get better output than me.

I think people should find this sort of result surprising by default. Like, presumably you don't know why your output isn't as good, or you'd fix it. So doesn't that lack of understanding concern you?

[–] citelao 4 days ago
> So doesn't that lack of understanding concern you?

Good question! It mostly doesn’t concern me. I believe these models have a huge “latent space,” so it’s not surprising to me that other people, with the right harnesses and prompts, can get them to produce more-valuable content.

I spent a lot of time calibrating how I use these models to produce valuable, good code, right?

But as I alluded with my question above, I haven’t seen a ton of evidence that these models have human-level argumentation in that latent space. If my assumption there is wrong, then yes, I’d have to reexamine my understanding of these models.

[–] zahlman 3 days ago
> I spent a lot of time calibrating how I use these models to produce valuable, good code, right?

Ah, fair, there's opportunity cost to that. May as well prioritize what you actually care about, yeah. Personally I'm fine with skimming through the prose in the default style, at least as long as the LLM isn't getting stuck in its own invented jargon. It'd be different if I were publishing that output as project documentation, but I don't even really want to try that.

> I haven’t seen a ton of evidence that these models have human-level argumentation in that latent space.

Not sure what you mean by argumentation. I'd agree that they don't seem to have flashes of insight or make intuitive leaps in the same way. But they clearly can follow through the logical steps of reasoning very thoroughly, once they know what to reason about. For a technology that seemed to struggle with understanding negation a couple years back, it's seemingly miraculous.