7. AI au pair with Human Intelligence
1956. This is the official starting point of the race to create an artificial intelligence that can match the human one.. Some 20 mathematicians and scientists brainstormed for 8 weeks to outline a path forward. The hope was to be able to create a machine that would demonstrate a level of intelligence “au pair” with human intelligence.
Overall the mood was optimistic and that optimism pervaded the following 15 years as better and better programs were written. However, after the initial burst of results it seemed that AI work has ended in a roadblock. New approaches were needed, and AI faded away from mainstream “news”. In these 65 years we have seen waves of renewed interest followed by neglect. Each new wave, however, brought forward a bit more clarity on what should be tried out as well as what kind of intelligence should be expected. Even though the goal of an AI au pair with Human Intelligence, as expressed in this Megatrend, still fascinates imagination, we are now looking at something different.
As a matter of fact, when we say that we want to match human intelligence, the starting point should be an understanding of what “human intelligence” is. And the picture of human intelligence is quite fuzzy!
There are several types of human intelligence that have been identified (or classified, since it is not a partitioning that is based on some objective facts). Most experts in the field classify them into 9 types:
If we look at this list we can easily find types that are fitting a machine (software), like mathematical and logical capabilities, where we have several demonstration of capabilities to prove theorems (so going beyond pure calculus…). Actually, we have some theorems that have been proven by computers (and some mathematicians refuse to acknowledge the proof!). If you are interested in this area you may like this paper discussing “superhuman mathematics”.
Other areas, like “Spatial” involving the perception of 3D space and its implication, are now addressable by a machine in a very effective way, although leveraging on very little “intelligence”. As an example self-driving cars use a variety of sensors (LIDAR and image sensors) and lot of number crunching to create a 3D model of the surrounding, than apply some form of AI to make sense of it (like identifying objects and then deriving the probability of behaviour -a pole is unlikely to cross the road, not so for a stroller that may be pushed from the sidewalk to the right lane as the car is approaching, although unlikely, whereas a risk exists for a kid chasing a ball). This sort of “machine intelligence” does not exactly map onto human spatial intelligence that includes characteristics like orientation, something difficult for us (in principle – imagine being stranded in the middle of nowhere…) but quite straightforward for a machine having access to a GPS.
Other areas, like “Musical or Linguistics” have been considered “out of reach” for machines, but we have seen in these last few years examples of machines (software) that can create “music”, paintings and poetry to a level that can fool people (including experts). The difference with people, of course, is that a composer would enjoy what is doing and be proud of the result, sensations that are not present in a machine.
Other areas, like “Social Intelligence” would seem to be far from a machine. However, we have seen significant progress in social robotics that ends up in machine that can establish an empathic relation with people. Again, the machine has been programmed, or has been programmed to learn, to be empathic, it doesn’t feell empathy.
For sure the areas of “intra-personal” and Existential” intelligence do not make sense for a machine (at least so far, unless you are interested in science fiction).
So what is this Megatrend about? It is not about the essence of Intelligence (assuming we can reach an agreement in defining what it is), it is about “performance”. If a machine can perform as well as a human being in a broad variety of contexts and situations that would require “being smart”, thus allowing the replacement of a human with a machine, than we can say that AI is “au pair” with human intelligence, like:
- can we replace a poker partner with AI?
- can we replace a pilot with AI?
- can we replace a medical doctor with AI?
- can we replace a financial advisor with AI?
- can we replace a teacher with AI?
- can we replace a journalist with AI?
- can we replace a discussant with AI?
The list can go on and on. If you like to explore a bit more, click on the links and you’ll see what is already possible today in using AI to replace people.
However, here the crucial point may seem to be in the words: “a broad variety of contexts and situations”. All the above examples, and many more, seems to point to the fact that we have, today, the capability to replace a person with AI in a specific domain. It may not be done in practice, because of cost or because of some shortcomings, but it is obvious that in the near future the cost will go down and many of the shortcomings will be overcome. So we can say that performance achieved in a specific sector is not a proof of having reached a human like intelligence, only that we are able to develop expert systems in that domain.
Actually, this is not a satisfactory objections. Computer software is additive and can scale graciously (at a cost, of course). In other words if you have an AI system able to impersonate a pilot, you can extend that (if you want to) to impersonate a musician by adding the required software (or asking the self-learning engine to dedicate a day to become a proficient musician). Provided sufficient computer power (either centralised or distributed) and having AI software capable of delivering intelligence in a given sector you can pick up that capability and add it to another set of capabilities.
This Megatrend is not about a philosophical discussion on the equality of machine and human intelligence, it is about the availability, by the end of this decade of an artificial intelligence able to perform as well as humans in a broad variety of field. It is not suggesting that an AI creating music will “enjoy” the piece it has created, nor that an AI pilot will feel proud after a particularly tricky landing, just that AI did the job.
Taking into account that we are now moving towards open AI, that is the possibility to access AI functionality from any object, thus spreading out the intelligence in any ambient and, conversely, to have any object contributing to an emerging AI (by harvesting / sharing data and creating local intelligence) I am pretty confident that this Megatrend is a concrete possibility that will be implemented, step by step, over the coming years.