In Against Orthogonality and Stupid Monsters, Nick Land lays out his arguments against the orthogonality thesis, which as originally conceived by Nick Bostrom, states, “Intelligence and final goals are orthogonal axes along which possible agents can freely vary. In other words, more or less any level of intelligence could in principle be combined with more or less any final goal.” I will argue against Land’s objection to this thesis.
The argument Land presents in Against Orthogonaltity rests on the idea that a set of necessary sub-goals any sufficiently intelligent agent would possess; self-preservation, efficiency, resource acquisition, and creativity, known as Omohundro drives, exhausts the space of all possible goals. Land writes, “Nature has never generated a terminal value except through hypertrophy of an instrumental value.” He further explores this idea in Stupid Monsters, suggesting that humans only follow instrumental goals (which he conflates with “will-to-think”), as can be seen by their lack of commitment to the goal we typically assign evolutionary significance – reproduction. He writes,
So how ‘loyally’ does the human mind slave itself to gene-proliferation imperatives? Extremely flakily, evidently. The long absence of large, cognitively autonomous brains from the biological record — up until a few million years ago — strongly suggests that mind-slaving is a tough-to-impossible problem. The will-to-think essentially supplants ulterior directives, and can be reconciled to them only by the most extreme subtleties of instinctual cunning. Biology, which had total control over the engineering process of human minds, and an absolutely unambiguous selective criterion to work from, still struggles to ‘guide’ the resultant thought-processes in directions consistent with genetic proliferation, through the perpetual intervention of a fantastically complicated system of chemical arousal mechanisms, punishments, and rewards. The stark truth of the matter is that no human being on earth fully mobilizes their cognitive resources to maximize their number of off-spring.
This argument is baseless for a number of reasons.
Firstly, the environment humans find themselves in now does not resemble the environment which selective pressure encoded our goals. Land is like a computer scientist who trained an AI to perform well at the game of checkers wondering why it cannot beat him at chess. Evidence for this environmental mismatch can be seen by observing the discrepancy of birthrates in tribal vs highly industrialized societies.
Secondly, selection applies at levels other than the individual. As E.O. Wilson says, “In a group, selfish individuals beat altruistic individuals. But, groups of altruistic individuals beat groups of selfish individuals.” A species which
When Land says, “Nature has never generated a terminal value except through hypertrophy of an instrumental value.” He presupposes that values are only generated through selective pressure – which is not true – hardcoded value functions and gradient descent do not depend on selection. Furthermore, the position that the only terminal values are Omohundro drives only makes sense in an environment of sufficient selective pressure. Granted, many of the things we typically think of as human values disappear when the selective pressure is turned up; this is the Malthusian scenario as resource competition and selective pressure increase with population size. The peacock’s feathers only exist in absence of selective pressure at the level of competing species. Similarly, art, beauty, romantic love exist and were able to develop in the absence of selective pressure – after humans achieved a sufficient intelligence advantage relative to competing species.
Land seems to take the brutal competitive nature of the universe as a given, but there is evidence to suggest otherwise. In fact, it’s quite possible that we are the most intelligent species in the universe. I suspect Land would counter this by appealing to the group selection effects I mention above. Yes, humans may be the most intelligent species, but we are in competition with the techno-industrial system itself. But this does not mean that a smarter version of humans couldn’t recognize this problem (if Land thinks this is the case, it is the ultimate form of hubris) and chain the beast so to speak. Some humans keep lions and chimps as pets after all – and they get pleasure from this, but if the animals pose a real threat there is always a bullet waiting for them. That humans fail to recognize the threat of techno-capital is a problem for humans, not the orthogonality thesis.
Land takes the lack of reproductive maximization of individual humans as evidence that selection is not powerful enough to program terminal values other than Omohundro drives, but individual humans do not appear to maximize their Omohundro drives either. People commit suicide. This is likely not enough to convince Land that terminal drives other than inflated instrumental drives exist, but let me hope to convince you, dear reader. On your deathbed, you are given the choice between a “super ecstasy” pill which will give you 5 minutes of bliss unlike any that you have experienced and a “super nootropic” pill which will give you 5 minutes of John von Neumann level of intelligence. I urge you to take a few minutes to read the “quotes about von Neumann” section here, before making your decision. OK, got it? It’s clear that if you chose the ecstasy pill then you possess some value other than Omohundro drives. Have you out-smarted me by choosing the intelligence boost? I argue that choosing the nootropic proves the same point. In this moment, there is nothing ‘instrumental’ about the nootropic – you’re going to die. If you chose this option, it is likely because you have the itch of an unsolved problem somewhere in your brain, and you will get some satisfaction from solving it. That looks a lot like a non-instrumental value to me.
I hate writing, so gunna wrap it up here, thanks.