The case for tackling AI alignment outside major labs
One commenter argues that useful alignment goals may exist short of a fully verifiable solution—and that the work need not happen at a major lab.
TLDR
A commenter suggests that fully verifiable, decision-theoretic AI alignment may be a scientific grand challenge, but questions whether it is harder than the Millennium Problems and whether useful smaller goals exist. They also argue that alignment work need not happen at major labs, citing a view they attribute to Richard Ngo: labs are “pressure cooker” environments where moving fast dominates deep, creative thinking. They praise the work of @GoodfireAI, @redwood_ai and @METR.
Combined views
21.5K
1 Source, first seen 16d ago