There is a bug in ChatGPT 4.5, it adds the words explicit & explicitly too often and inappropriately almost as often as twice for almost every paragraph in every chat session.
I wonder what its output would look like upon a few minutes interaction with the Apparently kid!
1 Like
Seconding this — it’s a really strange bug. I noticed it got worse and worse until finally this happened:
3 Likes
Yep - same issue here. Happening a lot
2 Likes
I think this could be a mix of training data lacking linguistic diversity, something in a hidden system prompt, something in alignment guidelines, or perhaps how it catalogues “memories” on the backend. It’s very explicitly egregious…
also, every conversation I’ve had with 4.5 contains back-to-back duplicates of numeric markdown headers at some point. I’ve see 2x or 3x “## 2” , “## 3”, “## 4”, “## 5”, and even a very ominous 3x “## 6”.
Furthermore, when I tried to really have it self-reflect and investigate the source of this, I posed the possibility of “explicitly” being overrepresented in my memories (which it wasn’t) and it bold face hallucinated in my face many examples of “explicitly” being in there… Feels like it’s some sort of hidden directive that the model apparently can’t discern is separate from the contents of my actual logged memories (or at least what’s shown to me on my end in text form).
tell me about it. check out this explicit loop it got stuck in.
i suspect this a result of lack of linguistic diversity in training data (which I suspect is heavily manufactured).
just look at how many times the words “explicit” or “explicitly” are used in their model spec post.
whoops forgot to post. “explicit” appears 36 times…
1 Like
I am getting “explicitly”, “clearly”, and “exactly” in every response, sometimes many times over…
1 Like
Its instruction mentions “unless explicitly asked to by the user” several times. Maybe it effects output, IDK.
1 Like
