3 ms·

"This is very unfortunate and I'm sorry to hear that the author has been excluded and is suffering to this extent." I can't stop seeing this LLM verbiage every

"This is very unfortunate and I'm sorry to hear that the author has been excluded and is suffering to this extent." I can't stop seeing this LLM verbiage everywhere I look. I feel like once you recognize the repeated syntax that got RLHF'd into all of these models you never stop seeing it. Maybe everyone is learning those patterns from reading AI-generated language now too.

nemomarx 1y ago
I think an llm would have put a comma in the middle of that? I wonder why no one's tuned their model to make minor errors that way or leave a few things uncapitalized to fake a human touch
- freehorse 1y ago
  You can do it after, remove all commas and newlines. Make it all a huge continuous blob of text that is hard to read. Definitely not LLM written. A lot of this here imo is just llm paranoia. I do not think the style of the author is different than their pre-llm posts. One em dash is not proof of anything imo as long as the text in general does not smell as much like llm-generated. I could be wrong because I do not interact much with llms lately, but it does not seem that to me.
  - nemomarx 1y ago
    It's more like 4 or 5 uses of it, and some of the phrasing choices also have that vibe. Everyone's threshold for the smell will be different, of course. It seems to me like it's more effort to write something and then have an llm clean it up instead of just posting it so I mostly don't understand the behavior to start with. Why are we going through this new effort?
    - freehorse 1y ago
      Until around a year or so ago, I could see and experience that I give some text of mine to an LLM and instantly it may often (not always) feel more structured and well written. I would think that using LLMs for "improving" my style made my text better. I no longer think it actually does. For one, the feeling that everything has the same style is unnerving. And we are getting more sensitive in distinguishing this kind of style, and together with the parts of it that are not actually as nice, or they are too much like polite "corpo" speech, and these stand out much more. Who would have thought that losing individual styles would make us feel so bored.
mathiaspoint 1y ago
Personally I was expecting it to get more diverse since there are a number of companies producing the models but it seems to have converged? Are they all cheating off OpenAI Is there some fundamental linguistic reason for that? It seems super strange to me. At the very least I thought they'd try to RL other personalities to make it harder to catch. The older base models definitely could handle that.
- freehorse 1y ago
  Yeah I noticed the same, in the beginning they seemed to have different "personalities" as in ways of interacting with users. Maybe it is just a design related thing, same as all commercial websites ending up being the same, as the users know what to expect and how to interact.
morleytj 1y ago
Very snarky, but okay, haha. Do you not get that sort of sense at all from the above structure? I'm curious if it's just in my head and what the perspectives of others are.

nemomarx 1y ago

I think an llm would have put a comma in the middle of that? I wonder why no one's tuned their model to make minor errors that way or leave a few things uncapitalized to fake a human touch

freehorse 1y ago
You can do it after, remove all commas and newlines. Make it all a huge continuous blob of text that is hard to read. Definitely not LLM written. A lot of this here imo is just llm paranoia. I do not think the style of the author is different than their pre-llm posts. One em dash is not proof of anything imo as long as the text in general does not smell as much like llm-generated. I could be wrong because I do not interact much with llms lately, but it does not seem that to me.
- nemomarx 1y ago
  It's more like 4 or 5 uses of it, and some of the phrasing choices also have that vibe. Everyone's threshold for the smell will be different, of course. It seems to me like it's more effort to write something and then have an llm clean it up instead of just posting it so I mostly don't understand the behavior to start with. Why are we going through this new effort?
  - freehorse 1y ago
    Until around a year or so ago, I could see and experience that I give some text of mine to an LLM and instantly it may often (not always) feel more structured and well written. I would think that using LLMs for "improving" my style made my text better. I no longer think it actually does. For one, the feeling that everything has the same style is unnerving. And we are getting more sensitive in distinguishing this kind of style, and together with the parts of it that are not actually as nice, or they are too much like polite "corpo" speech, and these stand out much more. Who would have thought that losing individual styles would make us feel so bored.

mathiaspoint 1y ago

Personally I was expecting it to get more diverse since there are a number of companies producing the models but it seems to have converged? Are they all cheating off OpenAI Is there some fundamental linguistic reason for that? It seems super strange to me. At the very least I thought they'd try to RL other personalities to make it harder to catch. The older base models definitely could handle that.

freehorse 1y ago
Yeah I noticed the same, in the beginning they seemed to have different "personalities" as in ways of interacting with users. Maybe it is just a design related thing, same as all commercial websites ending up being the same, as the users know what to expect and how to interact.

morleytj 1y ago

Very snarky, but okay, haha. Do you not get that sort of sense at all from the above structure? I'm curious if it's just in my head and what the perspectives of others are.