The rate of progress in AI in the last few years has been amazing. Self-driving cars are seemingly around the corner, with leading car manufacturers, along with Tesla, already selling cars with unprecedented levels of autonomous driving. One can talk to their phones as well as virtual home assistants to direct them to do various things. Computers are now better than humans at strategic complex games such as Chess and Go. Enormous computing, human and financial resources are pouring into AI. In light of all these developments, it is not surprising that the natural question of a human-equal, or perhaps even superior, AI is in people’s minds.
Such visions have been around for decades, widely popularized in Stanley Kubrick epic 1968 film masterpiece “2001: A Space Odyssey”. One of the principal protagonists in the film was HAL 9000, a superior form computer-based AI which was to assist the crew in accomplishing their mission, though its action did not quite go as planned.
The impression of HAL 9000 was certainly of superhuman capabilities, simultaneously holding in its electronic brain all operational aspects of running the ship, together with abilities to engage in deep focused conversations at the same time. But those impressions were rather superficial, for HAL 9000 was just an imaginary film concept.
There are many deeper question how HAL 9000 or another kind of such superior AI would look like. It is natural to assume they would be at a genius level, or even above, but in terms of what? Would it be the breadth of their information, with seeming abilities to answer all kinds of questions from multitudes of human knowledge? Or would it be in superior reasoning skills, creating with ease deep and complex arguments about any subject?
Such questions are also applicable to superior human intelligence, at genius levels. For it is far from clear how we define genius and how they distinguish themselves from e.g. people with very high IQ who are very capable in achieving exceptional test scores. This is a question for which we have surprisingly few answers. In this post, we are going to delve into genius distinctions.
It is natural to assume that superior AI would have exceptional capabilities in terms of the breadth of their knowledge, as that would be a form of computer scaling problem. In simple terms, such AI would have at its disposal tremendous memory and storage resources that can be simply scaled up. In terms of present computing technology, think of a collection of computers with the most and the fastest RAM and other forms of secondary storage such as e.g. SSDs, connected by the fastest network we can implement.
Consider Google’s entire machine production resources for search, never disclosed, but assumed to be in the range of 100,000 servers. Those numbers include replication for supporting very fast responses for search query loads in tens of thousands queries per second (qps). And Google has been storing their search index in RAM for more than a decade, since 2003. Let us assume a single search cluster of 10,000 machines with 128GB of RAM each. The entire Web is around 100 billion pages and that number has been relatively stable for years. At 10kB of index data per page, we get that such a cluster would hold an index of the entire Web in RAM.
Consider human brain, known to have around 86 billion neurons. A single precision floating point number occupies 4 bytes, so it would take about 350GB just to store simple states of such a collection of neurons. The number of synapses in human brain is much higher, about 10,000X or four orders of magnitude more. It would take a datacenter to store such amount of data for weights of synaptic connections in RAM. A datacenter would also provide connectivity among all those neuron and synapses representations, though it is not clear how such connectivity would contrast to human brain.
Our hypothetical AI datacenter would compare in very rough terms to human brain in terms of (very) simple neural representations. It would also have enough memory to hold all of the human knowledge on the Web, which is a lot, and certainly much more that any individual can ever achieve, even at genius level.
But how important is breadth for genius level intelligence? That is a key question that has not been considered much. It has been well known that breadth and diversity of interests were hallmarks of the greatest minds such as Albert Einstein, Nikola Tesla, Isaac Newton, Leonardo da Vinci and others. Newton famously said “even if I was able to see further than others, it was only because I stood on the shoulders of giants”, acknowledging the vital importance of the previous knowledge in great discoveries.
Consider, for instance, Archimedes, one of the greatest minds who ever lived. He was able to came up with amazing concepts and discoveries in the context of the knowledge available at his time, which was rather limited. He was the first to calculate area under parabola, as a precursor of calculus. It took over 1800 years for Newton and Leibniz to independently discover what we know today as calculus.
Newton discovered classical mechanics. Maxwell discovered classical electromagnetics. Planck originated quantum physics. Cantor created classical set theory. Turing and Gödel built upon it and made tremendous breakthroughs in theory of undecidability and mathematical logic. Google was built on foundations of the now famous PageRank algorithm, which was an application of the concept of eigenvectors and eigenvalues from Linear Algebra , known for hundreds of years.
All of those great inventions and discoveries were built in the context and upon foundations of previous knowledge of our civilization. All the geniuses who made their fantastic discoveries had key abilities to leverage knowledge available to them and add to it. This is why breadth is of great importance for exceptional intelligence, both human and artificial.
But note that geniuses certainly had great abilities to know and use lots of knowledge from diverse disciplines but not even they were able to work with anything close to breadths spanning all of available knowledge, for that is physically impossible. There is no genius, and never was, who was the best mathematician, politician, economist, artist and philosopher at the same time who attained all of the knowledge available to us. In that respect, computers and AI are very different, as their scale already allows us to hold all of our knowledge.
At this point we arrive at the first new surprising feature of future superior AI – that it will be capable of directly using and leveraging virtually all of of knowledge, which is almost immeasurably more than even the smartest geniuses. This first realization almost immediately poses another question – how much better and smarter could such AI be for this breadth advantage even if its reasoning capabilities were only similar to human, never mind superior? For all the geniuses who were able to draw inspiration from multiple disciplines, this superior AI would be in principle be capable to draw upon all the disciplines and all of the knowledge.
Consider again PageRank, which was in essence a basic application of eigenvalues which would be readily understood by a mathematics freshman. This is not to say that Pagerank is not an epochal achievement, on the contrary, its simplicity is its greatest strength demonstrating the immense power of applying important concepts across diverse disciplines. Of course, one of the principal strengths of geniuses was their unique analogies, resemblances and similarities where others could not.
This very important question of what truly constitutes superior intelligence has not bee considered at all and it is here in this post where we want to start the process of addressing it. It is far from the only such question, there are others which we will be looking into.