Artificial general intelligence: What to expect and when

Many experts expect that today’s artificial intelligence will eventually evolve into artificial general intelligence (AGI), a point at which computers meet or even exceed human intelligence. While I totally agree, I would also contend that this evolution will happen sooner rather than later. Why?

To answer that question requires a bit of background on AGI, beginning with defining what AGI is (and isn’t). Artificial general intelligence seeks to solve a broad spectrum of problems that intelligent human beings can solve. This is in direct contrast to narrow AI, which is what is represented by most of today’s AI. Narrow AI uses AI techniques to exceed human abilities in a specific problem area. Put another way, AGI is all the expectations of AI come true.

Unfortunately, we don’t really know what intelligence is. We can’t even say for sure whether there are types of intelligence that are different from human intelligence. While AI encompasses many techniques that have been used with great success on specific problems, AGI is more nebulous. This makes it incredibly difficult to develop software that solves a problem where the techniques are not known, and no concrete problem statement exists.

AGI Approaches

There are several approaches to creating AGI. It might be developed by combining today’s narrow AI capabilities and amassing huge computational power. It might be generated by replicating the human brain, either by simulating the neocortex’s 16 billion neurons or by uploading content from scanned human minds. It might be produced by analyzing human intelligence, defining a “cognitive model” and implementing it with procedural language techniques.

To provide an example that combines today’s narrow AI capabilities and amasses huge computational power, look no further than generative pre-trained transformer 3 (GPT-3). A monumental achievement from OpenAI that generates text, GPT-3 possesses a library that contains many billions of words and phrases, as well as their relationships to other words and phrases. GPT-3 has been so successful that OpenAI, concerned about the potential for its misuse, has not made it public. But while GPT-3 seems smart, few contend that it understands the words it is using. Still, GPT-3 does demonstrate that you can fool a lot of people a lot of the time if you have enough data and computer power.

Similarly, while replicating the human brain by simulating the neocortex’s neurons can generate AGI, the algorithms of today’s artificial neural networks have very little to do with the way in which biological neurons actually work. Despite these limitations, numerous massive AGI projects are currently underway. These include DeepMind, an Alphabet (Google) subsidiary; OpenAI, where Microsoft has invested $1 billion; and Google Brain, with its open source project TensorFlow. Facebook AI, IBM and virtually all large companies also have some sort of AGI research, and smaller projects abound.

Difficulties and Risks

To home in on an AGI solution, I contend most narrow AI falls short of the capabilities common to any three-year-old. The average three-year-old playing with blocks understands that objects exist in a real world, time moves forward, and blocks have to be stacked up before they can fall down. The basic limitation of all AI is that these systems are unable to comprehend that words and images represent physical things that exist and interact in a physical universe, and that causes have effects with the comprehension of time. These basic underlying problems have yet to be solved, perhaps because it is difficult to get major funding to solve problems that any three-year-old can solve.

I have written extensively that the risks from AGI superminds is very real, but calamity is not inevitable. AGIs are necessarily goal-directed systems and will exceed whatever objectives we set for them. At least initially, we can set those goals for the benefit of humanity, and AGIs will provide tremendous benefit. If AGIs are weaponized, on the other hand, they will likely be efficient in that realm, too. I’m not so concerned about “Terminator”-style individual robots as I am with the AGI’s mind being able to strategize even more destructive methods of controlling mankind. I believe these risks transcend today’s AI concerns of privacy, equality, transparency, employment, etc. AGI is akin to genetic engineering: While the potential is huge, so too are both the benefits and the risks.

The bottom line on risks is this: Banning AGI outright would simply transfer development to countries and organizations that refuse to recognize the ban. Accepting an AGI free-for-all would lead to nefarious types willing to harness AGI for calamitous purposes. Only an open development process has a chance of introducing AGI in a way in which we can take advantage of its benefits while eliminating its risks.

When Will AGI Happen?

While there is no consensus, AGI could be here relatively soon. Consider that a very small percentage of the human genome (which totals approximately 750MB of information) defines the brain’s entire structure. That means developing a program containing less than 75MB of information could fully represent the brain of a newborn with human potential. When you realize that the seemingly complex human genome project was completed much sooner than anyone realistically expected, emulating the brain in software in the not-too-distant future should be well within the scope of a development team.

Similarly, while we don’t yet know what to write, a breakthrough in neuroscience at any time could lead to mapping of the human neurome. There is, after all, a human neurome project already in the works. If that project progresses as quickly as the human genome project, it is fair to conclude that AGI could occur in the very near future. While timing is uncertain, it is safe to assume that AGI is likely to gradually emerge. That means your Alexa, Siri or Google Assistant, which are already better at answering questions than the average three-year-old, will eventually be better than a 10-year-old, then an average adult, then a genius. With the benefits of each progression outweighing any perceived risks, we may disagree about the point at which the system crosses the line of human equivalence, but it will continue to appreciate – and anticipate – each level of advancement.

The massive technological effort being put into AGI, combined with rapid advances in computer horsepower and continuing breakthroughs in neuroscience and brain mapping, suggests that AGI will emerge within this decade. This means systems with unimaginable mental power are inevitable in the following decades, whether we’re prepared or not. Given that, we need a frank discussion about AGI and the goals we would like to achieve in order to reap its maximum benefits and avoid its risks.