4o spirals incredibly when asked to make a prolog quine. For an added bonus, ask it to "read it aloud" via the "..." menu - it will read the text, and then descend into absolute word salad when trying to read the code. Fascinating stuff.
Very neat! A lot of small LLM's have a similar failure mode where they get stuck and repeat a token / get stuck in a 2-3 token loop until they hit the max message size cutoff. Very ironic that it's about a quine.
I think the funnier part is how it keeps pretending like it does that on purpose and saying things like "just kidding", "Alright, for real this time", "okay… Enough stalling"
IIRC this is what drove Bing Sydney insane - it had a filter on top that added emojis, and its output was fed back to it, which meant it was constantly out of distribution.
I love how it says "stop" multiple times after outputting the dragon emoji, as if it's actually getting annoyed and angry at it's own lm_head that keeps printing the wrong thing
GPT-5 Instant (no thinking) spirals wildly. Poor bot