The Slowdown Hypothesis

Abstract

The so-called singularity hypothesis embraces the most ambitious goal of Artificial Intelligence: the possibility of constructing human-like intelligent systems. The intriguing addition is that once this goal is achieved, it would not be too difficult to surpass human intelligence. While we believe that none of the philosophical objections against strong AI are really compelling, we are skeptical about a singularity scenario associated with the achievement of human-like systems. Several reflections on the recent history of neuroscience and AI, in fact, seem to suggest that the trend is going in the opposite direction.

Eliezer Yudkowsky on Plebe and Perconti’s “The Slowdown Hypothesis”

The hypothesis presented for a curve of diminishing returns of optimization power in versus intelligence out is incompatible with the historical case of natural selection, in which it did not take a hundred times as long to go from Australopithecus to humans as it did to go from the first brainstto Australopithecus, but rather the reverse. Many people have postulated logarithmic returns or other such diminishing returns to intelligence. They are easy to postulate.

It is much harder to make them fit the observed facts of either the evolution of intelligence (for talk about diminishing returns to brain size, genome size, or optimization pressure on the brain) or the history of technology (for talk about diminishing returns to knowledge or intelligence). Specifically exponential theories of progress are probably wrong, of course; Moore's Law has already broken down. But the historical cases we've observed are for roughly constant input processes producing increasing (though not always exponential!) outputs. Constant evolutionary pressure has produced, not exponential, but increasing outputs from hominid intelligence. A fourfold increase in hominid brains has not produced exponential returns, but to characterize the resulting returns as sublinear seems rather odd. In a nuclear pile, neutron multiplication factors are strictly linear—each neutron giving rise to 1.0006 output neutrons on average, for example—and the resulting pile of neutrons sparking other fissions would produce an exponential meltdown if not for external braking processes such as cadmium rods. For the novel phenomenon of recursively self-improving intelligence, where AI intelligence in is a direct function of AI intelligence out, rather than the AI intelligence being produced by a constant external optimization pressure such as human programmers, to fail to go FOOM once a threshold level of intelligence is reached, we need all these observed curves to exhibit a sudden sharp turnaround the moment they are past the level of human intelligence, and produce extremely sharply diminishing curves of intelligence-out versus optimization power in. Simply put, nobody has ever devised a realistic model of optimization power in versus optimization power out which both accounts for the observed curves of hominid history and human technology, which fails to exhibit an intelligence explosion once intelligences are designing new intelligences and a feedback loop is added from design intelligence to output intelligence. In fact, nobody has ever tried to develop such a model, and all attempts to postulate the lack of an intelligence explosion have done so by making up models which either completely ignore the new feedback loop and simply project normal economic growth out into the indefinite future without considering that AIs creating AIs might be in any way qualitatively different from a world of humans making external gadgets without tinkering with brain designs; or which simply ignore the observed parameters of evolutionary history and technological history in favor of making up plausible-sounding mathematical models in isolation which would have vastly mispredicted the observed course of history over the last ten million or ten thousand years, predicting observed diminishing returns rather than increasing ones. This paper falls into the second class.