As the field of A.I. matures, the idea of a general A.I. (aka AGI) gains ground both in technical and in pop-science discussions. Also, as the idea of Super-intelligence (the next logical step to an AGI) is quite promising as a technology, more and more people are drawn to A.I. research in an effort to make this technology possible sooner.
However, the reality is that AGI is not feasible yet and it may not be feasible for a few more years at least. Contrary to what the futurists claim, there is no way to predict when this technology will become available with reasonable confidence. We can speculate about it all we want and even survey experts in A.I. about it, but an average estimate is still an estimate, or a guess rather.
In order for AGI to be technically feasible we need to resolve a series of problems, all of which are quite challenging, even for the brilliant minds who conduct A.I. research in various universities and the R&D departments of tech companies. Namely, the AGI needs to be versatile, something we still haven’t figured out how to do, or if it is even possible with the current A.I. architectures. Also, an AGI would require a great deal of data in order to perform its tasks well enough. This kind of data may exist (or may not exist in certain domains), but access to it is not always practical. The sheer computational cost of just the I/O operations of this would be a challenging problem in and of itself. Furthermore, an AGI would require a great deal of fail-safes in order to ensure that it doesn’t get out of control, like the chatbots of FB or some other failed A.I. Implementing and configuring these fail-safes is a quite challenging task, considering that they may be responsible for preventing not just poor performance in the AGI system but also potential catastrophes. Finally, there are other reasons why AGI is still an unfeasible technology, and delving into them would be beyond the scope of this article.
However, just because AGI is still unfeasible doesn’t mean that we cannot contemplate on it and prepare ourselves accordingly. Perhaps not having it right here and now is what can enable us to optimize its integration to our society. Such a technology is disruptive and can easily morph into something beyond our comprehension, so no amount of forethought on this matter is excessive. Besides, there are moral / ethical implications related to the use of this tech, which will need to be resolved before they take the form of lawsuits and/or accidents.
Perhaps the most relevant aspect of AGI that we can look into right now, while waiting for this technology to become available, is A.I. Safety. This sub-topic of the A.I. field is quite a popular one, but it still evades the average A.I. person. Just because certain scientists have thought about it and written papers on it doesn’t make it as impactful as it ought to be. Besides, at the end of the day it’s business people that make these technologies happen, even if scientists and engineers are responsible for working out the technical details involved. So, A.I. safety needs to become more widely known and something that’s discussed by everyone involved in an A.I. project, not just the researchers. This way when AGI comes we’ll be ready for it and make the most of it, mitigating the risks it entails.
Zacharias Voulgaris, PhD
Passionate data scientist with a foxy approach to technology, particularly related to A.I.