Originally published at: Artificial Intelligence and The Posture of Skepticism - TPM – Talking Points Memo
A few days back I got an email from TPM Reader JL asking me not to give in to the Luddite or reflexively anti-AI tendency he sensed I might have. It was a very interesting note and led to an interesting exchange because JL is far from an AI maximalist or promoter and our views…
I have been reading more about A.I. lately. I share the fears about the potential harm to human beings and the natural environment in their encounters with A.I. algorithms, whether that’s microtargeted pricing of consumer goods, political manipulation, or competion for energy sources, water sources, land, labor, and pollution. The harm increases as the computing power of A.I. increases.
It would be helpful to have an economic model of the relationship between the computing power of A.I. and its need for resources. There is already a mathematical power law that describes the relationship between the next-token prediction accuracy of a neural language model and its inputs (computing resources, training data, and parameter count). A feasible project would be to pull that model back from neural model inputs to physical and monetary inputs (e.g. electricity, water, land, operating cost) and associated harm to the natural environment and to human society.
In short, what is the quantifiable economic cost of increasing the functionality of A.I.? While A.I. researchers like to divorce the complexity analysis of A.I. from specific applications, the associated economic analyses should be, one would think, application dependent.