Comment by enraged_camel
11 hours ago
The reason LLMs use the same cadence and cliches is that their RLHF does not emphasize writing well, the way it emphasizes, say, coding. And the reason for that should be pretty obvious: coding is where the big money is at.
So the author may be correct, but for a different reason: unless writing starts being very valuable as a profession, it's unlikely the labs will spend significant resources making their AI models better at it.
No, they use "the same cadence and cliches" because they inflate a short and ambiguous prompt into long and specific prose by making statistical assumptions about what best fills in the gaps. It's not a training problem, it's an information theory problem, and it's not really surmountable.
Any given model will always have some distinct implicit voice that its biased towards for that infill content, and so a popular model will always become exhaustingly common, painfully familiar, and cliche. Users can use more elaborate prompts that shift the voice away from the most normative and towards some other nodes, but they need to put in special effort for that, and what people-at-scale specifically want from these tools is to put in very little effort, so we can expect that overwhelming number of casual and naive users will always be generating cliche slop with them.
Code escapes this problem not because of training but because it specifically benefits from cliche (boilerplate, patterns, etc) and so an model whose code "voice" reflects your own taste as a coder (or your toolchain's taste as a vibecoder) is going to feel like productive output rather than slop. But it's still cliche.
No. Reinforcement Learning is doing a lot here. Anyone who played with these models before the Davinci intstruct-tuning (completion) era can tell you the same. In some ways, SOTA models have gotten better at writing, but the neuroticism of instruct-tuning has still not been resolved.
"The reason LLMs use the same cadence and cliches is that their RLHF does not emphasize writing well, the way it emphasizes, say, coding. And the reason for that should be pretty obvious: coding is where the big money is at."
Software can be checked for being 'written well' by compilers / linters etc. There is no equivalent for well-written natural prose. Spelling and grammar checkers haven't a clue about prose semantics.
They can use human writers and also LLM-based multi-step scoring systems with analytical guidelines.
They just haven't got to it yet. It's not a high priority but eventually they will make a solid effort on easier style guidance and more humanization. They will probably drop the most worn ou constructs like "not X, not Y, but Z" that have become tells and target a lot of known issues in reinforcement training.
It's slightly weird how confident writers are that it won't get improved.
What's interesting is that we still read AI's output and that's still the primary interface between a human and the AI. Implicitly that means that writing is important and I think we're already starting to see this with the complaints around how Claude "talks"
Claude "talk" is particularly annoying if you're using Claude to perform a job function (write code, analyze a spreadsheet, etc.) but average people reading an ad, Facebook post, etc. don't really care from what I can tell. In fact, I think most of them have no idea what Claude "talk" is, even if some of them notice common patterns that have emerged recently.