AI - What Is Superintelligence AI? Is Artificial Superintelligence Possible?



In its most common use, the phrase "superintelligence" refers to any degree of intelligence that at least equals, if not always exceeds, human intellect, in a broad sense.

Though computer intelligence has long outperformed natural human cognitive capacity in specific tasks—for example, a calculator's ability to swiftly interpret algorithms—these are not often considered examples of superintelligence in the strict sense due to their limited functional range.

In this sense, superintelligence would necessitate, in addition to artificial mastery of specific theoretical tasks, some kind of additional mastery of what has traditionally been referred to as practical intelligence: a generalized sense of how to subsume particulars into universal categories that are in some way worthwhile.

To this day, no such generalized superintelligence has manifested, and hence all discussions of superintelligence remain speculative to some degree.

Whereas traditional theories of superintelligence have been limited to theoretical metaphysics and theology, recent advancements in computer science and biotechnology have opened up the prospect of superintelligence being materialized.

Although the timing of such evolution is hotly discussed, a rising body of evidence implies that material superintelligence is both possible and likely.

If this hypothesis is proved right, it will very certainly be the result of advances in one of two major areas of AI research

  1. Bioengineering 
  2. Computer science

The former involves efforts to not only map out and manipulate the human DNA, but also to exactly copy the human brain electronically through full brain emulation, also known as mind uploading.

The first of these bioengineering efforts is not new, with eugenics programs reaching back to the seventeenth century at the very least.

Despite the major ethical and legal issues that always emerge as a result of such efforts, the discovery of DNA in the twentieth century, together with advances in genome mapping, has rekindled interest in eugenics.

Much of this study is aimed at gaining a better understanding of the human brain's genetic composition in order to manipulate DNA code in the direction of superhuman intelligence.

Uploading is a somewhat different, but still biologically based, approach to superintelligence that aims to map out neural networks in order to successfully transfer human intelligence onto computer interfaces.

  • The brains of insects and tiny animals are micro-dissected and then scanned for thorough computer analysis in this relatively new area of study.
  • The underlying premise of whole brain emulation is that if the brain's structure is better known and mapped, it may be able to copy it with or without organic brain tissue.

Despite the fast growth of both genetic mapping and whole brain emulation, both techniques have significant limits, making it less likely that any of these biological approaches will be the first to attain superintelligence.

The genetic alteration of the human genome, for example, is constrained by generational constraints.

Even if it were now feasible to artificially boost cognitive functioning by modifying the DNA of a human embryo (which is still a long way off), it would take an entire generation for the changed embryo to evolve into a fully fledged, superintelligent human person.

This would also imply that there are no legal or moral barriers to manipulating the human DNA, which is far from the fact.

Even the comparatively minor genetic manipulation of human embryos carried done by a Chinese physician as recently as November 2018 sparked international outrage (Ramzy and Wee 2019).

Whole brain emulation, on the other hand, is still a long way off, owing to biotechnology's limits.

Given the current medical technology, the extreme levels of accuracy necessary at every step of the uploading process are impossible to achieve.

Science and technology currently lack the capacity to dissect and scan human brain tissue with sufficient precision to produce full brain simulation results.

Furthermore, even if such first steps are feasible, researchers would face significant challenges in analyzing and digitally replicating the human brain using cutting-edge computer technology.

Many analysts believe that such constraints will be overcome, although the timeline for such realizations is unknown.

Apart from biotechnology, the area of AI, which is strictly defined as any type of nonorganic (particularly computer-based) intelligence, is the second major path to superintelligence.

Of course, the work of creating a superintelligent AI from the ground up is complicated by a number of elements, not all of which are purely logistical in nature, such as processing speed, hardware/software design, finance, and so on.

In addition to such practical challenges, there is a significant philosophical issue: human programmers are unable to know, and so cannot program, that which is superior to their own intelligence.

Much contemporary research on computer learning and interest in the notion of a seed AI is motivated in part by this worry.

Any machine capable of changing reactions to stimuli based on an examination of how well it performs in relation to a predetermined objective is defined as the latter.

Importantly, the concept of a seed AI entails not only the capacity to change its replies by extending its base of content knowledge (stored information), but also the ability to change the structure of its programming to better fit a specific job (Bostrom 2017, 29).

Indeed, it is this latter capability that would give a seed AI what Nick Bostrom refers to as "recursive self-improvement," or the ability to evolve iteratively (Bostrom 2017, 29).

This would eliminate the requirement for programmers to have an a priori vision of super intelligence since the seed AI would constantly enhance its own programming, with each more intelligent iteration writing a superior version of itself (beyond the human level).

Such a machine would undoubtedly cast doubt on the conventional philosophical assumption that robots are incapable of self-awareness.

This perspective's proponents may be traced all the way back to Descartes, but they also include more current thinkers like John Haugeland and John Searle.

Machine intelligence, in this perspective, is defined as the successful correlation of inputs with outputs according to a predefined program.

As a result, robots differ from humans in type, the latter being characterized only by conscious self-awareness.

Humans are supposed to comprehend the activities they execute, but robots are thought to carry out functions mindlessly—that is, without knowing how they work.

Should it be able to construct a successful seed AI, this core idea would be forced to be challenged.

The seed AI would demonstrate a level of self-awareness and autonomy not readily explained by the Cartesian philosophical paradigm by upgrading its own programming in ways that surprise and defy the forecasts of its human programmers.

Indeed, although it is still speculative (for the time being), the increasingly possible result of superintelligent AI poses a slew of moral and legal dilemmas that have sparked a lot of philosophical discussion in this subject.

The main worries are about the human species' security in the case of what Bostrom refers to as a "intelligence explosion"—that is, the creation of a seed AI followed by a possibly exponential growth in intellect (Bostrom 2017).

One of the key problems is the inherently unexpected character of such a result.

Humans will not be able to totally foresee how superintelligent AI would act due to the autonomy entailed by superintelligence in a definitional sense.

Even in the few cases of specialized superintelligence that humans have been able to construct and study so far—for example, robots that have surpassed humans in strategic games like chess and Go—human forecasts for AI have shown to be very unreliable.

For many critics, such unpredictability is a significant indicator that, should more generic types of superintelligent AI emerge, humans would swiftly lose their capacity to manage them (Kissinger 2018).

Of all, such a loss of control does not automatically imply an adversarial relationship between humans and superintelligence.

Indeed, although most of the literature on superintelligence portrays this relationship as adversarial, some new work claims that this perspective reveals a prejudice against machines that is particularly prevalent in Western cultures (Knight 2014).

Nonetheless, there are compelling grounds to believe that superintelligent AI would at the very least consider human goals as incompatible with their own, and may even regard humans as existential dangers.

For example, computer scientist Steve Omohundro has claimed that even a relatively basic kind of superintelligent AI like a chess bot would have motive to want the extinction of humanity as a whole—and may be able to build the tools to do it (Omohundro 2014).

Similarly, Bostrom has claimed that a superintelligence explosion would most certainly result in, if not the extinction of the human race, then at the very least a gloomy future (Bostrom 2017).

Whatever the benefits of such theories, the great uncertainty entailed by superintelligence is obvious.

If there is one point of agreement in this large and diverse literature, it is that if AI research is to continue, the global community must take great care to protect its interests.

Hardened determinists who claim that technological advancement is so tightly connected to inflexible market forces that it is simply impossible to change its pace or direction in any major manner may find this statement contentious.

According to this determinist viewpoint, if AI can deliver cost-cutting solutions for industry and commerce (as it has already started to do), its growth will proceed into the realm of superintelligence, regardless of any unexpected negative repercussions.

Many skeptics argue that growing societal awareness of the potential risks of AI, as well as thorough political monitoring of its development, are necessary counterpoints to such viewpoints.

Bostrom highlights various examples of effective worldwide cooperation in science and technology as crucial precedents that challenge the determinist approach, including CERN, the Human Genome Project, and the International Space Station (Bostrom 2017, 253).

To this, one may add examples from the worldwide environmental movement, which began in the 1960s and 1970s and has imposed significant restrictions on pollution committed in the name of uncontrolled capitalism (Feenberg 2006).

Given the speculative nature of superintelligence research, it is hard to predict what the future holds.

However, if superintelligence poses an existential danger to human existence, caution would dictate that a worldwide collaborative strategy rather than a free market approach to AI be used.

~ Jai Krishna Ponnappan

Find Jai on Twitter | LinkedIn | Instagram

You may also want to read more about Artificial Intelligence here.

See also: 

Berserkers; Bostrom, Nick; de Garis, Hugo; General and Narrow AI; Goertzel, Ben; Kurzweil, Ray; Moravec, Hans; Musk, Elon; Technological Singularity; Yudkowsky, Eliezer.

References & Further Reading:

  • Bostrom, Nick. 2017. Superintelligence: Paths, Dangers, Strategies. Oxford, UK: Oxford University Press.
  • Feenberg, Andrew. 2006. “Environmentalism and the Politics of Technology.” In Questioning Technology, 45–73. New York: Routledge.
  • Kissinger, Henry. 2018. “How the Enlightenment Ends.” The Atlantic, June 2018.
  • Knight, Heather. 2014. How Humans Respond to Robots: Building Public Policy Through Good Design. Washington, DC: The Project on Civilian Robotics. Brookings Institution.
  • Omohundro, Steve. 2014. “Autonomous Technology and the Greater Human Good.” Journal of Experimental & Theoretical Artificial Intelligence 26, no. 3: 303–15.
  • Ramzy, Austin, and Sui-Lee Wee. 2019. “Scientist Who Edited Babies’ Genes Is Likely to Face Charges in China.” The New York Times, January 21, 2019

What Is Artificial General Intelligence?

Artificial General Intelligence (AGI) is defined as the software representation of generalized human cognitive capacities that enables the ...