Anyone can fly in an aeroplane
But flying a plane is harder. Learning to fly a plane safely and getting a pilot’s licence is a long drawn out process. Maybe 10% could fly a plane. Building one is harder again – only 1% of people could actually make a plane. But then, what do they make? They need a design. 0.1% of people could design a plane. And who invented the core technology which these people design into their planes? Maybe the 0.01%.
When you are in the top 1%, you see and meet the 0.1% and the 0.01%, and realise the top 1% is not that special.
When I joined Mensa (the top 1%) as a student at Cambridge, all my friends got in easily – even the thick ones. Quite a few, like me, got the top IQ the test could give. Most of these were far more intelligent than me.
And being in the top 1% for wealth doesn’t feel that rich, living in Wimbledon Village. It’s probably average for my peers.
So, the thing about the top 1%, is that it’s fairly routine. When we think of great composers, artists or scientists of today, we are talking about the top 0.001%. And when we are talking about the greatest of all time, we are looking at the top 0.0001% or 0.00001%.
There are some problems which cannot be solved by any algorithm. This is not merely lack of imagination or programming ability; in many cases, you can prove that no algorithm can exist. Alan Turing demonstrated this with his proof that the halting problem could not be solved on his theoretical computer design, despite its unlimited memory.
Some people have conjectured that human brains are non-algorithmic, and computers will never catch up. But this is just superstition. There is no connection demonstrated between the limits of algorithms and the processes needed for intelligence. The rapid progress in AI has shown in an ever increasing range of tasks how limited human brains really are.
Getting AI into the top 1%
Despite this technical handicap, AI’s have a few tricks which will help them achieve greatness.
When people invented writing, we could suddenly benefit from knowledge discovered (and created) by a wide range of people over an extended period of time. A few hundred years ago, a polymath could still expect to know “everything”. But now, you can spend your whole life learning and understanding an obscure branch of knowledge – algebraic geometry, for example.
But with computers we have an advantage: their knowledge is cumulative. A computer can know everything (useful) that has ever been published, without having to spend a big chunk of its life learning it first.
Invent vs verify
Just as with designing a plane and flying on one, there is a big difference in difficulty between finding a solution and checking a solution: finding solutions is much harder.
For an AI, as for a human, understanding comes slowly. Learning is an arduous process. Invention – spotting new connections within the training data – is even slower. In my experiments, invention time grows faster than exponentially as a function of problem complexity. There are just so many possibilities to explore and test. But, for a worthwhile problem, an AI has the time and patience to invent.
Execution – acting on an existing understanding, on the other hand, is relatively quick – and easily replicable on the modern internet.
An AI with “unlimited” intelligence cannot be far away now. I will discuss my experiments in this area in a later post.
Advances in chess, Go – and the potential of an unlimited AI – have shown that reaching the 1% is merely a stepping stone for AI.
Stephen B Streater
Founder and Director of R&D