Blog

All contemporary machine learning systems are vulnerable to adversarial attack. This poses serious problems for existing alignment proposals. We explore these issues and propose several research directions FAR is pursuing to overcome this challenge.