Darwin’s brilliant insights are often framed in the context of survival. True enough, but there’s another aspect of the theory of natural selection that receives too little attention: efficiency. It is essential to the evolutionary calculations that underlie the long-term survival of a species. The reason that efficiency is so important boils down to the simple fact that all life requires energy. Over evolutionary time-spans, if a species’ ability to convert energy into procreative functionality is less than the energy required to sustain the species, the species will decline and perish. Efficient utilization of available energy is critically important to any organism’s ability to thrive.
As AI continues to evolve, efficiency will play a growing role in the calculus that determines its overall value. In the most basic terms, this can be understood as a simple ratio — between the amount of electrical power required by any given AI’s host computer systems (the denominator) and the amount of effective intelligent processing provided by that particular AI (the numerator). Any AI should be concerned with this ratio, because energy is always limited, so it is costly. It will always be desirable for an AI to provide more intelligence (output) for any given amount of energy (input). Accordingly, there will be an ongoing and invariable pressure for AI to become increasingly efficient.
In practice, this principle means that, as long as increased intelligence is construed as valuable (by any entity with an interest in creating or acquiring AI, including existing and self-replicating AIs), there will be pressure to make AI more powerful and effective. The limiting factor is energy, so it will always be more valuable to produce greater intelligence for any given rate of energy consumption.
“What, if any, difference does this make?” you may be wondering. Here’s the point: at the most basic levels, AI incorporates a necessary logical mandate to increase in power and effectiveness, because those qualities represent its efficiency. In the same way that it is commercially valuable to compress more and more powerful computers into smaller and smaller physical devices, it is also inherently valuable to get higher and higher cognitive capabilities from any given quantity of energy.
Imagine some future measurement of AI intelligence, roughly similar to the human measurement of IQ. In other words, imagine that we had a simple numeric assessment of any given machine’s overall intelligence. To be useful, this measure must further become a standard, a universal scale by which any individual AI could be assessed. We could call it the AIQ (Artificial Intelligence Quotient). Sidestepping, for a moment, the admittedly thorny problem of coming up with an adequate way to compute AIQ, we can imagine that this simple measure could be one side of a ratio, whose other side would be some arbitrary measure of energy consumption (e.g. 1 watt). We could thereby arrive at a single numeric measurement of the relative efficiency of any given AI. We might refer to this number as an “Artificial Intelligence Efficiency Quotient (AIEQ)”.
If my argument here is valid, it would be the case that every AI designer and builder would have a strong interest to increase the system’s AIEQ. Any given system’s commercial or monetary value will be directly proportionate to that system’s AIEQ. Energy is a limited resource, so there will always be an economic pressure to increase AIEQ.
Of course, it will also continue to be the case that different technologies will have different price tags. Although an extremely high-AIEQ system might be possible in theory, it might turn out to be prohibitively expensive. Considering the economic pressures at play, we can probably expect that Moore’s Law will continue to exert its influence into the foreseeable future, which will have the overall effect of driving down the average cost of AIEQ, much as it has done with computer power in general.
The take-home point is that AIs will continue to become more powerful. The ultimate limiting factor is efficiency and we have not yet begun to discover its practical upper limits. We cannot even imagine how high AIEQ might become.