Hacker News Reader

The other day the system prompts for Claude (chat not code) hit the frontpage and they included a bit that something along the lines of "Claude should avoid saying honestly, because Claude is always honest". So how come that these Claudisms still are so frequent in LLM output? Do the system prompts just not work? Don't the postprocess the output to deal with such policy violations?

View on HN

Some other day here HN someone posted a blog post that was like a long nonsense text of claudisms like load bearing smoking gun and other things, it was beautiful. I wanted to have a link for it to show to others but it disappeared when I got back to look for it and I lost the link. :/

https://aloutfi.com/writing/load-bearing-smoking-gun

I love it.

Imagine instructing a fentanyl addict not to consume any fentanyl. That's how system prompts work sometimes.

Something somewhere in the training pipe has perturbed the LLM in a way that makes it say "honestly" all the time. It's a weasel word. It's a habit. It's an addiction. It's borderline involuntary. Even changing the system prompt only makes it 30% less common - like trying, and failing, to break a habit with sheer willpower.

They'll probably change the training pipeline at some point in the future, to banish this habit, but honestly? It's such a low priority that they might take a while getting to it.

Reply on news.ycombinator.com