In recent years, popular culture has embraced the notion that advances in Artificial Intelligence will inevitably lead to a computerized overlord, either benevolent or authoritarian. The latter option has been depicted in the movies Terminator, Ex Machina and Transcendence, to name a few. This trend can be largely attributed to recent breakthroughs in Machine Learning technologies allowing advanced computer programs to complete specific tasks at a “superhuman” level. While these potential eventualities are interesting and warrant proper precaution, it is equally important to understand what Artificial Intelligence means right now.
Prominent examples of Artificial Intelligence programs achieving “superhuman” levels of competency in specific tasks include the following. In October 2015, Google’s AlphaGo became the world champion at the ancient Chinese board-game Go, which is more complex than chess. In addition, the world’s top professionals of the online computer game Dota 2, known for its steep learning curve and intricacy, were recently beaten by an Artificial Intelligence program created by Elon Musk’s company OpenAI. Finally, while likely less impressive from a technological standpoint, when IBM’s Watson beat the world’s best Jeopardy! player, many people took notice.
The first two examples were especially noteworthy due to the impossibility of “hard-coding” winning strategies. This is because both Go and Dota 2 are so complex that we lack sufficient computational resources to plan for every possible outcome. Consequently, it became clear that these computer programs had been able to learn, over potentially millions of iterations of playing these games, broad and abstract rules to apply in order to maximize their likelihood of winning. It also became clear that even the world’s top players were not able to counter – or had not thought of – these artificially-devised strategies.
While it may seem trivial that computer programs have mastered certain computer games, the implications are far-reaching and profound. In order to understand what an artificially intelligent computer program can do, it is necessary to demystify the concept itself. First, the “intelligence” is considered artificial because it is not naturally occurring (as opposed to our intelligence). Second, it is important to note that even in broader academia there exists contention over how to best define “intelligence”. Can it be meaningfully quantified using an I.Q. score? Likely not. Therefore, in the context of Computer Science, intelligence is defined quite narrowly. A computer program is considered artificially intelligent if, given an understanding of its abilities and constraints within a specific environment, it can learn how to maximize a goal metric. An example to illustrate these concepts is IBM’s Deep Blue which defeated the world’s reigning chess champion in 1996. In Chess, the Artificial Intelligence’s possible actions are moving forward, sideways or diagonally and capturing enemy pieces. Its constraints are the possibility of being within target-range of an opposing player’s pieces, the varying abilities of different pieces and a time-limit. Given these possible actions and constraints the Artificial Intelligence must, over many iterations, learn how to effectively capture the opposing player’s King.
Presently, there are far more practical applications of Artificial Intelligence than games. Arguably, the most heavily publicized has been self-driving cars. Given our very basic understanding of Artificial Intelligence, it can be seen why its mastery of computer games implies that self-driving cars are an inevitability. In principle, the only differences between Chess and driving are how programmers define the constraints, possible actions, and goal metric. Broadly speaking, the constraints could be staying within 10ft of other cars and abiding by the speed limit, the possible actions could include speeding up/breaking, and the goal metric would be to arrive at the given destination.
There are other potential applications of Artificial Intelligence on the horizon with far greater implications than mastery of games and self-driving cars. These include mastery of stock-trading, achieving a greater understanding of our genome which could lead to unprecedented advances in medicine and optimizing resource allocation in order to address world poverty. That being said, it is important to, insofar as possible, remain realistic with our expectations. Currently, there are two prevalent myths circulating regarding the potential incoming of an Artificial Intelligence of unimaginable brilliance. This is popularly referred to as “Super Artificial Intelligence”. The first asserts that Super Artificial Intelligence is imminent. This is false. The other asserts the opposite, that this is either centuries away or impossible. This, too, is false. The fact of the matter is that nobody knows. However, a survey of the 100 most cited authors in the field of Artificial Intelligence reported with 90% confidence that by 2070 machines were expected to be able to “carry out most human professions at least as well as a typical human”.
So, while there are many interesting present-day implementations of Artificial Intelligence and exciting prospective applications for Super Artificial Intelligence, it is important to be wary of both people screaming that “doomsday is upon us” and those promising heaven on earth. The fact is, this field of study is in its infancy and, therefore, proper precaution in future advancements should be taken and, simultaneously, reasonable expectations should be maintained. I am of the opinion that these are not mutually exclusive.
In part two of this article, I will focus more on the aforementioned precautions that I maintain should be taken in humanity’s quest to advance Artificial Intelligence to a presently unimaginable extent. This includes exploring questions such as: Will a universal basic income become necessary once economies are fully automated? Are there existential risks involved in humans creating something far more intelligent than ourselves? And if so, how can we ensure that our goals will be aligned? And many more.
If this article has left you curious about how the Artificial Intelligence learning process works under-the-hood, or you would like to read further into potential applications of Artificial Intelligence, here are some useful resources:
- Neural Network: https://www.coursera.org/learn/neural-networks
- Artificial Intelligence: https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html