PERSON
Paul Christiano
The alignment researcher who put a probability on civilizational catastrophe and kept working anyway—principal architect of reinforcement learning from human feedback, inventor of iterated amplification and AI safety via debate, and the clearest honest accountant of what could go wrong as AI systems grow more capable than the people who build them.
There is a particular kind of mind that becomes most useful precisely at the moment everyone else is either celebrating or panicking, and Paul Christiano has one of them. While the public conversation about artificial intelligence oscillates between utopian rapture and apocalyptic dread, Christiano has spent more than a decade doing something far less theatrical and far more consequential: he has tried to write down, with mathematical precision, exactly what could go wrong and exactly what we might do about it. The technique now called
reinforcement learning from human feedback—which quietly powers nearly every AI assistant you have spoken to—descends in significant part from his early insight that we could teach machines what we want without ever fully specifying it. That single idea, that we could align systems to human preference rather than to a brittle hand-written objective, reshaped the field. Yet Christiano