“Key research bottlenecks for Jared Kaplan are honesty, self-correction, and the ability for AI systems to retrospectively identify, fix, and learn from their own mistakes.”