Is there a subspace in residual streams reserved for things the model definitely won't say but chooses to consider (e.g., the user is stupid)?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-mechanistic-backchannel-2026,
author = {Holtzman, Ari},
title = {Mechanistic Backchannel},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/MUCuvYVF70kk8TL4GCQZ}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!