AI Safety and Theoretical Computer Science, Scott Aaronson (University of Texas at Austin)

Abstract:
Progress on AI safety and alignment, like the current AI revolution more generally, has been almost entirely empirical.  In this talk,
however, I'll survey a few areas where I think theoretical computer science can contribute to AI safety, including:
- How can we robustly watermark the outputs of Large Language Models and other generative AI systems, to help identify academic cheating, deepfakes, and AI-enabled fraud?  I'll explain my proposal and its basic mathematical properties, as well as what remains to be done.
- Can one insert undetectable cryptographic backdoors into neural nets, for good or ill?  In what senses can those backdoors also be unremovable?  How robust are they against fine-tuning?
- Should we expect neural nets to be "generically" interpretable?  I'll discuss a beautiful formalization of that question due to Paul Christiano, along with some initial progress on it, and an unexpected connection to quantum computing.