fp.

Have a look at this recent Scrabble video where Claude plays semi reasonably and ChatGPT goes crazy https://youtu.be/8opLB1D_RYY (skip to 6:50 for the insanity)

skerit•1h ago

Neat. Is it a single under-trained token in GPT-5.2? Or is something else going on?

magicalhippo•41m ago

Based on their tokenizer tool[1], for GPT 5.x "geschniegelt" is tokenized into three tokens:

  (ges)(chn)(iegelt)

[1]: https://platform.openai.com/tokenizer

WatchDog•35m ago

Perhaps, the word does have it's own token, " geschniegelt"(geschniegelt with a space in front of it), is token 192786 in the tokenizer that GPT-5 apparently uses.

https://raw.githubusercontent.com/niieani/gpt-tokenizer/refs...

joaomacp•1h ago

Maybe it's getting confused by the expression "geschniegelt und gestriegelt", seeing both as geschniegelt and getting confused

WatchDog•8m ago

I tried this in chatgpt, asking " geschniegelt" on a 5.2 instant temp chat, and got some interesting results.

Sometimes it would reply with the correct definition of geschniegelt, the description would sometimes be in German, sometimes in English.

Most of the time it would give me a definition for a different German word "Geil".

For whatever reason, the most interesting results I got were via my work's m365 copilot interface, where it would give me random word descriptions in Hebrew[0] and Arabic[1].

[0]: https://pastebin.com/raw/h108gr9t

[1]: https://pastebin.com/raw/BFAbtVQN