NeurIPS reviewers flag Claude-like writing styles and non-standard phrases like 'continuation top-20' in submissions

It seems to me that AI writing has changed, at least on technical topics. One year ago every AI was far too verbose always. But recent Opus outputs are sometimes not verbose enough. "Stacatto" for sure. Reminds me of this "split personality" point from @nostalgebraist

@sirbayes I also worry about false positives. A lot of text looks AI written but it might not be. I used to like using em dashes, but I use them less now, for fear that my writing could look AI generated.

"forward model" and "lexemes" and "continuation top-20" and "clause-terminal" read as really non-standard phrases to me

@sirbayes Terse? I struggle to get Claude to shut up and stick to facts. I hate it’s writing and summaries

@sirbayes I find even these orbit points unreadable. Fixed point would be completely alien, I imagine.

@sirbayes Yeah. I felt that 3/4 of the papers I reviewed were clearly written by Claude

@sirbayes Is the fixed point research progress or research stagnation?

@andrewgwils @sirbayes Everyone hates this. It won’t matter, everything is getting molded into this and you have to comply. So if you don’t write like AI, then it doesn’t matter it gets rejected.

@andrewgwils @sirbayes Everything it’s folded into the algorithms so fast, there is no novelty for anyone.

@sirbayes Also a lot of terminology borrowed from applied statistics papers that essentially adds nothing to the narrative, but it’s introduced as analysis rigor. 3/4 of my assigned papers had that.

Indeed I believe there's already evidence that human language use has shifted (both towards and and away from various LLM characteristics).

As far as whether the fixed point (of both humans and machines training on each other) will be readable to humans: I just hope that at that point, some humans will still care about the answer to this question.

@sirbayes Easy to win in this environment if one does writing oneself - reviewers will like it just because it doesn't sound like slop

@andrewgwils @sirbayes Yeah I really dislike how much the 'is this slop' muscle is growing in my brain. Feels like a costly heuristic to have to apply to like everything I read, but I haven't quite figured out the alternative...

@sirbayes I’ve noticed that style in my own drafts too, but to me it’s a red flag. It usually pops up when the model is hallucinating narrative around my points or lacking substance. It uses that staccato tone to circle the point without actually making it.

@sirbayes If you do a good job reviewing, then it will be valuable training data. Scientific publications that speak the truth are possibly our most valuable text.

@sirbayes This is mostly Opus 4.8 and they can be aligned via prompt to not do that.

But the gist is:

The model is “100% certain it must be uncertain”.

And yeah - GPT PRO writes the better papers as of now.

@andrewgwils @sirbayes When they thought about what this potentially could be, and they sold everything on these ‘Image your own future’ philosophies, they had no clue how to implement any of this.

@ArthurConmy hard agree, and after reading tons of this stuff i am starting to think they are very semantically dense in ways we possibly don't totally get.

@sirbayes I'm wondering if the fixed point is even the right way to think about it. Model collapse seems more like a slow fade: readable but empty, just the average of averages.