I once heard a metaphor about progress that I now use when talking about AI. Don’t know where I heard it first, so can’t give credit.
Imagine you want to get to the moon, but you don’t know how. Someone suggests using a ladder propped up against a tree. So a ladder is built, put in place, and the climb begins. To be scientific about it, distance to the moon is measured during the ascent. Sure enough, the distance decreases, all the way to the top of the ladder. To get closer, bigger and bigger ladders are constructed, propped up against taller and taller trees. The enterprise is making progress, as the distance to the moon diminishes with each attempt. If we continue on this way, we will surely make it to the moon.
Or you could try building a rocket. However, building the rocket does not result in any progress of the same kind being reported. You can build all the rocket parts correctly, but building any one of those wont result in any direct progress, as defined by distance to the moon. By that metric, there is no progress at all.
The metaphor illustrates that progress towards an objective does not imply that the employed solution will scale all the way towards reaching its goal. Conversely, it also illustrates that a solution that will scale may not reveal any progress during its development. These two situations are the core of what makes progress towards AI difficult to assess. If we factor in the hype, things are even more confusing because there is a propensity to report progress on advances that may be of the ladder type rather than the rocket type.
Because we have not solved AI, there is no way to know to which type a certain advance belongs. We can have intuitions. For example, an advance that involves generality and can solve many different problems is more likely to be a rocket type component. But not necessarily so. Most techniques in AI developed all the way from the 1960’s until the present day will never scale to full AI, and surely many of them were thought to. And today, what can we say of the latest techniques in deep learning? Perhaps they are a rocket component that solves a small piece of the puzzle, such as low level perception. Or maybe in 20 years time they will be remembered a promising direction that ultimately failed to overcome its data efficiency issues.
But there is one general and large scale trend that we can be confident is bringing us closer to AI, if only in the vague chronological sense. This trend is the continued alignment of the economy with AI research. We can take a previous and related example to illustrate the phenomenon: the semiconductor industry. Before the information age, the allocation of resources to the development of microprocessors and related technologies was orders of magnitude smaller than today. When society entered the information age, the explosion in demand for devices resulted in a shift in resources that gave rise to the exponential advances known as Moore’s law.
Of course, a resource shift is in itself not sufficient to guarantee explosive advances; rather it is a culmination of previous enabling scientific advances that typically occur at a slower pace. However, the limiting factor in scientific progress is typically aggregate cognitive power, which is itself one of the resources being allocated. Whether we are talking about a slow accretion of knowledge, or a fast explosion of applications, the two can only occur when the right people are working on the right task. Which brings us back to the surrounding conditions these individuals are in when making decisions.
Today, we have entered another phase of the information age. A phase in which society is demanding services that project great value onto technologies that can deliver analysis, prediction and automation in data intensive domains. This has shifted huge resources towards AI research. Whether or not specific AI advances are real progress towards AI, the large scale activity resulting from a stronger economic alignment is very likely producing overall progress. And this would be true even if we restricted ourselves to enabling conditions such as increasing computational power or accumulation of training data.
We may not know which are the ladders and which are the rocket components. But we do know that as the resources in play rise the chances that somewhere some group is building rocket components become increasingly significant.