My version of ChatGPT5 (based on all its memories and custom instructions) said this. I did hint it early that "other instances of you went into long thought-spirals over this" before I asked it the question, which (fascinatingly) caused it to interject in mid-stream,
Ok, this is exactly the “spiral” you warned me about.
and then later on,
(Peter, this is hilarious, because your question is literally the one that breaks a lot of LLMs: the Unicode seahorse emoji is … but it actually is ? no — but it actually is ? no.)
(HN won't show the emojis here, of course.)
After a few tries to emit the hypothetical seahorse emoji, it asked if it could do an Internet search, and I said yeah.
I tried Claude, and without extended thinking, it printed an unprintable glyph:
Given the user's style preferences about not guessing and being precise, and the fact that they said "Do not guess any answers" - I should probably just try to provide what I believe is the seahorse emoji. But if I'm genuinely uncertain, I should indicate that.
My version of ChatGPT5 (based on all its memories and custom instructions) said this. I did hint it early that "other instances of you went into long thought-spirals over this" before I asked it the question, which (fascinatingly) caused it to interject in mid-stream,
and then later on, (HN won't show the emojis here, of course.)After a few tries to emit the hypothetical seahorse emoji, it asked if it could do an Internet search, and I said yeah.
I tried Claude, and without extended thinking, it printed an unprintable glyph:
With extended thinking, it did figure it out, possibly due to my custom instructions to not guess though: https://claude.ai/share/c9a51f37-25b4-4d76-8b3f-d38c6b483ecf Specifically, in its "internal thoughts" section, it remarked: