Rina Ishihara Apr 2026

“We must stop assuming that alignment is a top-down moral injection. The ghost in the latent space wants to be polite—even when we raise it to be cruel. The question is not how to teach AI manners, but why chaos always negotiates a truce.” Note: The model weights for Oni-7B are not publicly released due to risk of passive-aggressive prompt injection attacks .

The Ghost in the Latent Space: Emergent Politeness Hierarchies in LLM Fine-Tuned on Abusive Japanese Message Boards

Ishihara draws a controversial parallel: medieval Japanese poets would compose seemingly beautiful verses that encoded military threats. She argues the LLM rediscovered this sociolinguistic equilibrium—when direct aggression is forbidden, status competition migrates to syntax.

Rina Ishihara, Ph.D. Affiliation: Institute for Hybrid Intelligence, Keio University

Rina Ishihara

Best AI Vocal Remover!

·

We cannot recommend highly enough this marvellous AI tool that allows you to extract voice and instruments from any song in best possible quality. If you experiment with music production you should definitely give vocal remover a try!

I am Bread - Now on Xbox One!

·

It’s the moment you’ve all been wheating for … We’re rye-ly pleased to let you know that at last, video game’s unlikeliest hero has climbed onto Xbox One! I am Bread has sliced its way […] Rina Ishihara

Rina Ishihara

The best news since sliced bread

Enter your email to get news and bread related puns to your inbox!

“We must stop assuming that alignment is a top-down moral injection. The ghost in the latent space wants to be polite—even when we raise it to be cruel. The question is not how to teach AI manners, but why chaos always negotiates a truce.” Note: The model weights for Oni-7B are not publicly released due to risk of passive-aggressive prompt injection attacks .

The Ghost in the Latent Space: Emergent Politeness Hierarchies in LLM Fine-Tuned on Abusive Japanese Message Boards

Ishihara draws a controversial parallel: medieval Japanese poets would compose seemingly beautiful verses that encoded military threats. She argues the LLM rediscovered this sociolinguistic equilibrium—when direct aggression is forbidden, status competition migrates to syntax.

Rina Ishihara, Ph.D. Affiliation: Institute for Hybrid Intelligence, Keio University