Self-Goodharting

Adversarially fragile fixed-weight models are therefore misaligned

Fixed-weight AI systems need categorical boundaries to pursue goals, but high-dimensional concept spaces admit…