The Alignment Problem: Machine Learning and Human Values

In 1935, 12-year-old Walter Pitts escapes bullies by hiding in a Detroit public library, where he becomes engrossed for three days in a complex book on formal logic. After identifying errors in the book, he writes to one of the book’s authors, Bertrand Russell, to communicate those mistakes. As a reply, Pitts receives an invitation to be Russell’s doctoral student at Cambridge, which he declines due to his young age. Three years later, Pitts runs away from home to attend Bertrand Russell’s public lecture in Chicago.

At the lecture, a teenage Pitts meets Jerry Lettvin, forming an inseparable bond despite the difference in their interests (Pitts’s focus is on logic, while Lettvin’s is on poetry and medicine). Pitts, who lacks formal education, informally audits classes at the University of Chicago and even points out errors in famous logician Rudolf Carnap’s book, earning his respect. By 1941, Pitts meets neurologist Warren McCulloch through Lettvin. Pitts and McCulloch start to collaborate, combining their interests in logic and neurology into a groundbreaking paper on neural networks. The paper, however, has only minimal initial impact. The paper suggests that the brain’s functions can be modeled through propositional logic This work sets the foundation for the development of neural network theory.