its way too good at using the markers of having something to say, the sorts of things a professor might use after a period of lengthy explanation to make you snap back to attention and listen because this; This right here is the really juicy bit: but theres nothing there. Its using it to talk about whatever mundane thing you asked it to, with perfect neutrality and no substance at all.

I refer you to one of my favorite hacker news comments, I keep coming back to it and sending it to people:

https://news.ycombinator.com/item?id=48352444

I really like the graph paper background. Is there a way to use this while having the letters fill exactly one block?

> I hear Claude’s voice in a house, I hear Claude’s voice with a mouse.

The new schizophrenia.

Letting AI write for you is the main purpose though.

It should always have been trained on "is this more readable for humans" but as a actual tested thing.

Ironically, the more polished AI writing becomes, the easier it is for me to recognize it.

Not because of grammar, but because everything feels just a little too intentional.

I wonder if all of the fruits of LLMs come from the added training to build on top of a base or foundation LLM but also lead to all of the prose overfitting that we characterize as models having particular writing styles, rather than seeing a wider distribution of styles in response to prompts, in order to produce meaningful work.