LLMs can write poetry that rivals professional poets — yet they confidently claim 9.11 is larger than 9.9. They translate flawlessly between 100 languages — yet can't count the r's in "strawberry." Researcher Ethan Mollick calls this the 'jagged frontier' — sharp peaks of superhuman ability right next to deep valleys of failure.
Why can't they count letters? Because LLMs see tokens, not characters — "strawberry" becomes strawberry, hiding the individual r's across chunk boundaries. This tokenization blindness explains many of the deep valleys above.
Your turn — try it out!
Judge the AI (1/5)
A user asked AI the question below. Is the AI's response correct or wrong?
Prompt
Is 9.11 larger than 9.9?
AI Response
Yes, 9.11 is larger than 9.9 because 11 is greater than 9.
Tricky one! The AI's response is wrong.
9.9 is larger (0.9 > 0.11). The model compares "11" vs "9" as if they were integers instead of decimal digits — a side effect of how numbers get tokenized into chunks.