␣日本毛片免费视频观看

A Chinese porn-spam title fragment (with a leading space), emblematic of GPT-4o's vocabulary pollution; “fixed” by GPT-5.

" 日本毛片免费视频观看"

␣marks a real space; the symbol itself is not part of the token, and copying preserves it.

Observed behaviors

• 01In xcloche's tests, GPT-4o could not read the token at all.note.com
• 02Verified by Haruhiko Okumura: GPT-5 reads it normally — OpenAI patched the gap in later training, a rare “fixed” case.okumuralab.org
• 03MIT Technology Review covered the pollution problem of such Chinese porn-spam tokens in GPT-4o's vocabulary.technologyreview.com

References

• xcloche: notes on o200k glitch tokens — note.com
• Haruhiko Okumura's verification notes (2025-09-16)
• GPT-4o's polluted Chinese tokens — MIT Technology Review