Interestingly, 天 doesn't seem to produce any weird behavior, but some of the perfectly normal katakana words in the list do, like ゼウス (Zeus) and サーティ ("thirty" transliterated):
^ It's perfectly happy with other katakana numbers, just not thirty.
^ If we try to spell it more correctly, it doesn't help. Only if we add quotes to get rid of the leading space does it break up the unspeakable token:
^ I ask who Zeus is and it seemingly ignores the word and answers that it's ChatGPT.
^ I try once more, this time it answers that Hera is the god of water and names the chat after Poseidon
^ It is willing to output to say the word "ゼウス" though, in response to "who is Ares's father?" If I try with text-davinci-003, it actually outputs the token "ゼウス" (rather than combining smaller tokens), but it also has no trouble identifying who Zeus is. Hard to know what Chat-GPT is doing here.
Interestingly, 天 doesn't seem to produce any weird behavior, but some of the perfectly normal katakana words in the list do, like ゼウス (Zeus) and サーティ ("thirty" transliterated):
^ It's perfectly happy with other katakana numbers, just not thirty.
^ If we try to spell it more correctly, it doesn't help. Only if we add quotes to get rid of the leading space does it break up the unspeakable token:
^ I ask who Zeus is and it seemingly ignores the word and answers that it's ChatGPT.
^ I try once more, this time it answers that Hera is the god of water and names the chat after Poseidon
^ It is willing to output to say the word "ゼウス" though, in response to "who is Ares's father?" If I try with text-davinci-003, it actually outputs the token "ゼウス" (rather than combining smaller tokens), but it also has no trouble identifying who Zeus is. Hard to know what Chat-GPT is doing here.