Tokenization doesn't explain this kind of a mistake. If you ask about "djsjc" you would get the proper answer of 5. The claim that this is performance trade-off does not hold.
I cannot edit the question, but would like to say that I'm extremely impressed by ChatGTP and entire question was an honest curiosity about the limitations of it. It is strange that many responses are about blaming my question and example as just wrong and not about the limitations of the ChatGPT model (admirable anyway).
I cannot edit the question, but would like to say that I'm extremely impressed by ChatGTP and entire question was an honest curiosity about the limitations of it. It is strange that many responses are about blaming my question and example as just wrong and not about the limitations of the ChatGPT model (admirable anyway).