INTERPRETABILITY FOR DUMMIES >be humanity >build AI that’s about to be way smarter than us >have no idea what’s going on inside it >can only see what it says and does >it passes every safety test >but no way to tell if it’s actually safe or just pretending >need a way to look inside its head >call it interpretability >spend years getting nowhere >2023: OpenAI can only figure out what about 2% of an AI’s parts do >AI keeps getting bigger, more confusing, more capable >doomers say we’re…