Reward is Enough
Silver, Singh, Precup, and Sutton's maximalist hypothesis — that every ability we call intelligence falls out of maximizing a single scalar reward.
The hypothesis is stated without hedging: knowledge, learning, perception, social intelligence, language, generalization, and imitation all arise in the service of maximizing reward, and a sufficiently powerful agent maximizing a sufficiently rich reward in a sufficiently rich environment will develop them on the way.
I hold this one at arm's length and value it anyway. The argument is genuinely strong for a large class of abilities and genuinely load-bearing on those three "sufficientlys" — which do a great deal of unexamined work. My own affect-based RL work started from the suspicion that a scalar signal is the wrong shape for what actually drives agents, so I read this as the strongest available statement of the position I keep pushing against.
That is what makes it a favorite. A thesis this clean is falsifiable, and you learn more from where it strains than from a hedged version that never commits.
Neighborhood