Why can computers play chess or Go better than any human but struggle with walking or seeing? This is Moravec’s paradox.
Here’s a list of things that humans can do. Not everything we can do, just some things we sometimes do:
- Solve complex algebra equations
- Beat a grandmaster in a game of Chess
- Make money investing in the stock market
- Walk on two legs
- Recognise a face
- Talk with someone
The first three are markers of intelligence (well, a particular type of intelligence). Only some people are able to do them. The last three are possible for most toddlers, and we are not usually impressed when they do so.
Back in the first heyday of artificial intelligence, there was a lot of optimism about the rate of progress. In a fairly short time, computers could already do complicated math and play chess faster and better than people. AI researchers had already solved many of the problems that were a marker of intelligence in people – the hard stuff – so surely a truly intelligent machine was not far behind?
Well, it turns out that it’s actually really difficult to teach a robot to walk on two legs, or train an algorithm that can simulate a convincing conversation. It’s not impossible, as the last few years of AI have demonstrated, but it’s several orders of magnitude more complicated than something like chess. To put it another way, for a machine hard is easy and easy is hard.
This apparent paradox was described by Hans Moravec, who noted that living creatures have been perfecting walking and looking for millions of years, but only got around to algebra and chess in the last couple of thousand years. Abstract thought itself is something we’ve only been doing for a few hundred thousand years now. Tasks that are actually really difficult – recognising a face, for example – have become automatic and intuitive to us… but not to AI. Traditional markers of intelligence were not a good predictor of what is difficult to develop from scratch.