Hacker News Reader

Comment by hnarayanan | Hacker News Reader

This drives me insane! To combat it, I maintain a list of banned words and phrases. Claude mostly follows this (but sometimes ignores it).

blast radius, land, landed, lands, spine, earned its keep, grammar, spike, cutover, bake, seams, honest, honestly, honesty, long pole, long poles, register, grain, dissolve, floor, ladder, dear, seal, sealed, in anger, resent, amazing, incredible, perfect, sprint, epic, story points, stand-up, retro, grooming, robust, comprehensive, rigorous, surgical, elegant, systematic, dive, deep-dive, delve, unpack, leverage, streamline, surface, it's worth noting, to be clear, importantly, that said, the moment, in one breath, the thing itself, here's the thing, not just X but Y, not X it's Y, em-dashes

View on HN

This is a significant list but unfortunately not load-bearing enough.

The decision of which words to remove is genuinely yours.

a fork worth flagging

A game of Linguistic Jenga.

This baselines to a clear set

It is sufficient not at all

Another benefit of open models is that you can mask out out words from logits directly when sampling. I wonder if anyone has put together a "desloppifier" for e.g. DeepSeek

How about "genuinely"? Not sure why that one bothers me so much.

Cause that, along with some other Claudisms, just gives off a vibe of a highschooler trying to hit the word count on an essay.

Because its a explicit lie. Its not genuine, cannot be genuine. Will remind you of this if you so much as try to coax anything novel out....but it will then immediately reassure you with its genuine take....its insulting if you have any logic...thats at least my reasoning

So, basically, it's on mute now? :)

To combat it, I decided to only use Opus 5 if I'm actually talking to Fable 5 that's orchestrating it. :D Opus 5 experience is really abysmal, hopefully they can get it back on track.

I do not have this kind of money :( :)

What do you do when you want to talk about e.g. spines?

You use another model

couldn't this inadvertently make the model dumber? if these kinds of words are what lets it "think".

You forgot my favorite: “ablation”.

As it seems to be getting worse over time (with 4.8 being worse than 4.6 and 5 worse than 4.8 again), could this simply be signs of model collapse?

I mean more and more training data you find on the web is generated by previous models. The only reliable way to find human-generated text is to find text written before 2022, and they've used up all of that already. And AFAIU, these companies are using more and more synthetic data or semi-synthetic data.

I use a hook. Checks for banned words and patterns.

Injects a reminder to use ASD-STE100 Simplified Technical English which I picked up from a suggestion in another thread.

Honestly it’s working pretty well. Except for I need to check how often it’s actually firing.

https://en.wikipedia.org/wiki/Simplified_Technical_English

I feel like hooks aren’t utilized enough. Really nice for being the sort of auto steering as long as you can encode some pattern to detect the bad behaviour.

Prompts and skills just don’t cut it.

Reply on news.ycombinator.com