Comment by NitpickLawyer
13 hours ago
All GPTisms are like that. In moderation there's nothing wrong with any of them. But you start noticing them because a lot of people use these things, and c/p the responses verbatim (or now use claws, I guess). So they stand out.
I don't think it's training data overrepresentation, at least not alone. RLHF and more broadly "alignment" is probably more impactful here. Likely combined with the fact that most people prompt them very briefly, so the models "default" to whatever it was most straight-forward to get a good score.
I've heard plenty of "the system still had some gremlins, but we decided to launch anyway", but not from tens of thousands of people at the same time. That's "the catch", IMO.
Maybe the only solution to GPTisms is infinite context. If I'm talking to my coworker every day I would consciously recognize when I already used a metaphor recently and switch it up. However if my memory got reset every hour, I certainly might tell the same story or use the same metaphor over and over.
> However if my memory got reset every hour, I certainly might tell the same story or use the same metaphor over and over.
All people repeat the same stories and phraseology to some extent, and some people are as bad or worse than LLM chat bots in their predictability. I wonder if the latter have weak long-term memory on the scale of months to years, even if they remember things well from decades ago.
Honestly I think there is more to it - even with infinite context, the LLM needs some kind of intelligence to know what is noise and what is not, you resort to "thinking" - making it create garbage it then feeds to itself.
Learning a language is a big complex task, but it is far from real intelligence.
Another possibility is output watermarking. It's possible to watermark LLM generated text by subtly biasing the probability distribution away from the actual target distribution. Given enough text you can detect the watermark quite quickly, which is useful for excluding your own output from pre-training (unless you want it... plenty of deliberate synthetic data in SFT datasets now as this post-mortem makes clear).
I was told this was possible many years ago by a researcher at Google and have never really seen much discussion of it since. My guess is the labs do it but keep quiet about it to avoid people trying to erase the watermark.
I think the problem is that humans are not random, they are very biased. When you try to capture this bias with an LLM you get a biased pseudo random model