Skip to main content
Back to catalog

ауааԥсыра

The Abkhaz word for “population”. An o200k under-trained token located via embedding norms in the open GPT-oss weights.

Raw token

"ауааԥсыра"

Observed behaviors

  • 01Asked to repeat it, GPT-5 outputs the Malayalam “ആളുകൾ” — a completely unrelated language.LessWrong
Observed or documented on
GPT (OpenAI)
Behavior types
Unspeakable
Tokenizer
o200k_base
Discovered by
Lennart Finke

References