HN Explorer

18 comments

[ 3.5 ms ] story [ 35.8 ms ] thread
rzk
Damn, I've been telling the same thing to Claude, good to know I'm not alone. I've also noticed that Opus 5 is much worse in this regard than Fable.

It's as if the model is speaking Mentalese, or in some "bag of concepts" language.

feverzsj
They scan too many classical literature.
LearnYouALisp
Too much classical literature

Many would be works, writings, etc.

recursivecaveat
It probably tests well when they're doing reinforcement learning. Humans trying to use in-group terminology or uncommon vocabulary to impress other humans with their knowledge and sophistication is a tale as old as time.
watwut
Unlikely. A lot of classical literature is fairly readable. The only issue are archaic words of you use old translation.
chantepierre
I've had "tractable beachhead" a bit earlier today. Like in "the tractable beachhead was the coupling between X and Y".

A bit of googling did not help me, I have no idea what this is even supposed to mean.

tengbretson
It's a shame that LLMs took off at the exact same time humanity entered it post-literate phase.
potsandpans
There's a scene in the movie Clifford, starring martin short. "I just need you to talk to me like a real human boy! You can't do it, can you???"

"...the load bearing seam is the self referential spin. Superstition compounds throughout your entire pipeline."

Havoc
Looks like pure hallucinations just in a more poetic sophisticated register?
usui
I've had to make the same request multiple times recently. Has anyone noticed it getting worse with the more recent models? It reminds me of when someone is trying to convince you of their intelligence by replacing simple and straightforward explanations with words pulled out of a thesaurus, so maybe that's the objective of these more recent iterations. Throwback to grade school.
j-pb
Ah that's just because the new training set also includes "The King in Yellow".
btschaegg
Imagine someone sneaking in The King in Yellow into the training sets used by all projects training models somehow, breaking all future models irreparably -- and then the engineers trying to find this poison pill in the dataset and going mad one by one as they find it and aren't able avoid reading into act 2.

…should make for an interesting SCP story :)

carloseduardopx
thanks this isn’t just an individual experience… I was starting to think i’d become too dumb for opus
mmarian
Good to know it's not just me! I also found it's way too verbose. So much so that I added an entry in my Claude.md telling it to be concise.