Unverified50% confidenceFactExact time
The miscounting of the character 'r' stems from the tokenizer splitting 'strawberry' into a single token, so the model cannot directly access its internal character composition.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLLMs split text into tokens for processing, and 'strawberry' is often segmented into subwords such as 'straw' and 'berry,' causing models to easily make errors when counting at the letter level.72% similarUnverifiedThe GPT-4o tokenizer turns 'How many r's are in strawberry?' into 8 tokens: How | ·many | ·r | 's | ·are | ·in | ·strawberry | ?, where 'strawberry' is token #101830.62% similarUnverifiedEncoded with o200k_base, 'I love strawberry' becomes [40, 3047, 101830].52% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1022421API
curl https://kongchang.com/api/v1/knowledge/claims/1022421MCP
get_claim(id=1022421)