Reaction
The debate over predicting AI thoughts through mechanistic interpretability
One post calls concept-linked neuron findings preliminary to AI alignment; another disputes whether AI thoughts can be predicted.
TLDR
One poster argues that finding neurons associated with concepts such as the Eiffel Tower is only an early step toward understanding AI thought well enough to aid alignment. Another agrees with the broader criticism but rejects the proposed goal of predicting AI thoughts in advance, pointing to their skepticism about reading minds with fMRI.
Combined views
155
4 Sources, first seen ago
