“During his research at Google Brain in 2016, François Chollet found that gradient descent was incapable of discovering generalizable algorithms for reasoning problems, instead defaulting to overfit pattern matching.”