site banner

Friday Fun Thread for October 3, 2025

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

2
Jump in the discussion.

No email address required.

Why are LLMs so bad at wordplay? I asked them "What is the favorite sex toy of Linda Load and Diana Doll?", explicitly stating that it's a wordplay riddle, and none of them got it right.

Yeah LLMs are notoriously weak at anagrams due to how tokenizing works. Here's a fun little demo of how LLAMA's tokenizer breaks up an input sequence: https://belladoreai.github.io/llama-tokenizer-js/example-demo/build/

Hence the famous example of LLMs failing at "how many times does the letter R appear in the word strawberry" until the training data for newer models was contaminated with the answer.

Some domains they struggle with wordplay for tokenization reasons, especially for matters like rhyming, counting syllables, so on. Dunno if this is one of them.

This may be an artifact of the LLMs you're using being trained to avoid the idea space you're working around, although like SubstantialFrivolity I'll admit I don't know the 'right' answer, either. Nevoria, which is trained on a lot of sexual chat, gave ""A Peter Puffer" and "A Rod Rocket", which doesn't seem awful even if I'd have gone with something like "Fred Fluffer", but I get the feeling that there's some specific domain knowledge that I'm missing.

After staring at it for fifteen minutes, I realized that both names are an anagram for "anal dildo" and it's only the rule to be charitable that makes me call this a wordplay.

In fairness I have no idea what your riddle is getting at either. But the answer to all the "why are they bad at X" questions is "because they aren't actually intelligent".