Verisforge

The talk argues for evidence-based AI safety and distinguishes near-term risks, such as AI-assisted cyberattacks, from speculative claims about imminent loss of control. Clark Barrett presents automated reasoning and formal verification as ways to check code, while explaining that proofs depend on assumptions and specifications. Michael Koken Derfer describes methods for validating AI systems before and after deployment, drawing on aircraft safety research, adaptive stress testing, and runtime monitoring. In the discussion, the speakers consider independent evaluation, regulation, security, and the limits of benchmarks and formal methods.