[RobKohr]

AI loves to sell the words they are saying like a QVC salesperson with a knife set.

It doesn't just cut cans, it cuts tomatos too. You would think you have to sharpen it, but you don't.

Not just this, but that.

Sounds nothing like a normal succenct engineer.

[danudey]

I had Claude do a "short, succinct summary" of two bugs I found in someone else's Python code: one was missing parentheses around subtraction (x = a - b - c instead of x = a - (b - c)) and one was `SomeException("asdf")` instead of `raise SomeException("asdf")`.

Both explanations were a paragraph of text, each about six lines long, which I replaced with a single sentence each.

[throwitaway222]

When it spits out 4 pages of explanations, when I'm asking it what's better postgres locks or redis locks (two dbs I already have in my project), I sometimes tell it to keep it shorter than a few sentences. We don't need an essay for every decision.

[asveikau]

I think I wrote this in an HN comment before, but knowing what I do about linguistics and people, I think some of the AI writing styles will be re-enforced and organically more popular even for human authored writing over the next few years. We're all being unconsciously exposed to those language structures and they will feel comfortable, normalized, and we'll be inclined to repeat them.

I don't like it necessarily, but that's the way it works. Repeat exposure creates that.

[twoodfin]

Unfortunately, I suspect this phenomenon will not be limited to rhetorical style.

To pick a proximate example: It’s often the case that the assertive, confident LLM LinkedIn-ese is obscuring an invalid argument.

It would be surprising if repeat exposure to nonsense across billions of people—and indeed the particular nonsense embedded in the LLM world models—didn’t have some real bad effects in aggregate on humanity’s ability to cogitate reliably.