alt Hacker NewsIf you have a slight interest in originality of presentation, whatever website copy first comes out of Claude is not there yet. Its preoccupation with "no x, no y" and annoying runs of sentence fragments might have been a punchy rhetoric for a short while before it became a tired cliché, but by now it stylistically makes it look more like you're bragging on LinkedIn than presenting something fun and creative.
I mention it because it's immediately off-putting even though the project itself sounds fun enough. It tarnishes it with the impression that maybe you just don't care, but maybe you just aren't attuned to the style and its strong slop connotations.
I'm genuinely fascinated by the problem and I've wondered what in the training process causes the model to develop this peculiar style of writing. Is it because it's trained on old Medium slop? Someone else replied with a link to their SSH-based VPS, where the copy is absolutely saturated with the same annoying style.
Replies
> I've wondered what in the training process causes the model to develop this peculiar style of writing
I think this might just be inherent to the currently deployed LLM architecture. If the training process works at all, any model will come out of it with some (default) voice that it'll apply everywhere.
Human voices are not like that, both because one copywriter does not write a significant fraction of all texts in the world and because unlike LLMs they remember all of their own prior writing, which gives them an opportunity to get tired of their old tropes.
Yeah it kinda feels like a great, original meal but served on plasticware
I don't think Claude's writing style is peculiar, it is just that we are being flooded with it. Everything becomes stale with overuse.
> what in the training process causes
Read some early pre-mass-adoption LLM papers by the key scientists and researchers. The style is very similar...
It's an extremely pervasive misconception, but a model's voice is not really a matter of its training data, not by the time it reaches consumer eyes. The voice these models use are specifically selected for in RLHF, ie, curated by humans by telling it it did a good job when it uses this voice and telling it it did a bad job when it didn't use this voice.
Why do they specifically push for this voice? Because it is effective. This is "headline voice", the voice that gets people to endlessly click on Youtube clickbait and garbage articles. Most of the population gets a dopamine hit from this style of writing, making it legitimately addicting to them. Those who recognize it for what it is and find it insufferably grating are in a tiny, tiny minority.
I understand where you are getting from, thanks for posting this. I'll keep this in mind for my next project (or if I push changes to this one.) I also wanted this to appeal to non-hackers, and a big part of that was me wanting to show how simple it was to log in.