Reaction
Concerns about relying on which ideas reach pretraining for future AI safety
A poster argues that relying on which ideas make it into pretraining is too fragile to underpin future AI safety.
TLDR
A poster warns that relying on which ideas reach pretraining to align future superintelligences is dangerously fragile, likening the approach to witchcraft rather than engineering. A reply jokes, “the lightcone shall be dominated by obscure 2010 SEO slop.”
Combined views
296
2 Sources, first seen ago