There must be linear features in LLMs that correspond to 'this word/phrase/sound is funny' and it must be somewhat specific to a specific reader being modeled. But I think 'this word sounds funny' is low saliency enough that LLMs don't model all the reader groups they're exposed to. Can we get a sense of which personas are tracked in more or less depth by looking for such features in LLMs?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-phunny-phonetics-2026,
author = {Holtzman, Ari},
title = {Phunny Phonetics},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/RgOP0ot1IOWSrBZ6RuIN}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!