Finding the Answer

Marcia,

This is one of those times that I truly hope that I am wrong. Navigating the potential risks vs. rewards of AI will be challenging. It would be easier if there are some basic tasks that can only be accomplished by humans. It would be easier if there are some fundamental reasons that AI can’t accomplish certain tasks which will ensure humans have a role in our joint human/machine future. I remain unconvinced that future AI/robotics will have such limitations. However, disagreements on this topic are to be expected; even the experts don’t agree! It may not yet be possible to resolve this debate and determine whether AI will eventually outperform humans at all relevant tasks, but it is possible to assemble key questions and observations from the discussion to guide our continuing exploration of the topic.

General Observations. There will be more than “one” AI. Different companies and different countries may each build their own version or “species” of AI. Similarly, different “individuals” of each species may be created by different users. These AI “individuals” may cooperate or compete with each other to perform key tasks. These tasks may help or hinder humanity. That complexity is why it is necessary to evaluate the overall impact of the entire AI “community” on all of humanity; a daunting task. As such, it is less relevant to worry about whether a single AI can perform every human task better than all humans; we need to worry about whether there are multiple AIs that can each perform one or more tasks better than humans such that, when combined, the collection outperforms all humans at all tasks.

Key Steps.The complexity and rapid advancement of AI make predicting its future very challenging. It is unrealistic for a novice to predict the future capabilities of AI from current trends in the underlying hardware and software; even experts struggle to do that. Novices can, however, make reasonable predictions about the future capabilities of AI based on the current trends in the capabilities themselves, at least for the near-term. A rough overall prediction for future AI capabilities can be generated by following the current trends for AI performance in a few key tasks. My own personal list of tasks are:

- Strategy Games: With AlphaGo beating the #1 Go player in the world in 2017, AI demonstrated superhuman strategy capabilities for situations with simple, well-defined rules.

- Machine Vision: Machine imagers (digital cameras) are generally on par with the human eye while image processing is not yet at the human level for general applications, but it is advancing rapidly. 

- Autonomous Driving: Continuous improvements have been achieved in driving aids (adaptive cruise, lane departure) while true autonomous driving is still years away. Full self-driving cars will likely require true artificial general intelligence (AGI).

- Robot Maneuvering: The Atlas robot by Boston Dynamics has demonstrated nearly human level maneuverability and its capabilities are rapidly advancing. 

- Turing Test: AI can seem convincingly human for simple tasks or narrow applications, but has not yet been convincingly human in a general environment. 

In summary, AI/robotics has achieved superhuman capabilities in narrow, well-defined tasks and is rapidly improving its performance in all of the general tasks that I have evaluated. Rapid improvements and new applications are expected to continue for AI over the near term.

The Key Step. Generally speaking, AI can outperform humans at a few specific, narrow tasks, but is still clearly lagging humans at general tasks such as autonomous driving. Artificial general intelligence, AGI, represents the key step in AI development towards unlocking the capability to perform general tasks better than humans. A machine with true AGI would be able to drive you to court, act as your lawyer in court, and then beat you at chess while driving you home! When will this threshold be achieved? Like all things AI, opinions are varied. There is, however, a general consensus that it will likely occur sometime in the next several decades. The true potential of AI may finally be illuminated when AGI is achieved. If you do nothing else, watch for this event. Both humanity and machines will experience rapid, fundamental changes when it occurs. 

Fundamental Differences. In addition to generally disrupting society, achieving AGI may also answer the very question that we are debating: will machines eventually outperform humans at everything? Achieving AGI may illuminate fundamental, permanent differences between human intelligence and artificial intelligence. When the current differences between human intelligence and AI are discussed, people are most often really describing the differences between general intelligence and narrow intelligence. If both humans and machines have achieved general intelligence, then the remaining differences may indeed be a sign of true, fundamental differences. Signs of these differences may also appear earlier as AI is continuously developed towards AGI. One major sign of a fundamental difference would be the demonstration that the human mind is more than just a complicated biological circuit. If something beyond purely physical processes is required, spiritual or otherwise, then a purely physical AI may not be able to replicate human mental capabilities. Similarly, maybe the human brain is simply a biological computer, but it is too complex to be fully replicated. Maybe the human brain needs the entire human body, including senses and muscles, to operate properly so its functions can’t be replicated unless the entire body is included. Perhaps machines won’t be able to handle multiple forms of motivation, like emotions. The list of potential limitations is a long one, but to date, I have not identified any proven fundamental limitations. 

Human Advancement. One key question that has completely been neglected is: are humans done advancing? Shouldn’t we compare the ultimate potential of machines to the ultimate potential of humans? What is our ultimate potential? Genetically, human beings change much slower than machines advance. Pure genetic advancements, our “hardware”, will not keep up with machine advancements. What about our “software”? Humans are continuously learning new information and skills. Will that keep up with the advancement of machines? Humans are also continuously developing new tools to further amplify our capabilities. Maybe humans can maintain a performance advantage over machines when all of our tools are included? Not likely. We have been following a megatrend of automation since the dawn of civilization; tools are continuously improved to help us better perform tasks, but the tools are eventually improved to the point that they are automated and can actually perform the entire task. What about the advancement of the understanding of ourselves, what it means to be human and live a good life? As I current see it, that is our only hope to maintain human superiority in key activities and avoid catastrophic consequences. Humanity must, as a whole, grow to understand that we are more than just biological machines, that more is not always better, and that we can have truly meaningful lives without AI. If we don’t realize that we are more than just biological machines, we will essentially be treated as if we are, we will be outcompeted, and quite likely we will be replaced by a more effective machine.