In the past couple of weeks you might have seen that Google's AlphaGo software played the world Go champion four times. AlphaGo won three of those 4 times. Why are computers so smart?
One of the biggest reasons computers tend to be so smart is that they can gain more experience than a human being can in a lifetime. A chess player, for example, might play 2000 games in a lifetime if they're really dedicated. Maybe even 5000 if they started really young. But computers can play millions of games in a day. AlphaGo became as smart as it is because it played millions of times a day against itself.
There are more possible moves in a game of Go than there are atoms in the universe. So, there is no possible way for a computer to calculate all the possible moves and calculate the best moves. Therefore, Google had to come up with a way to make the software learn on it's own and make smart decisions.
The human mind is still pretty impressive because Lee Sedol was able to figure out how to trap AlphaGo just after playing it three times. Computers need to perform a task millions of times to learn from it, but this is proof that humans can learn the same amount from performing the task significantly fewer times.
But, even though computers need to perform a task millions of times to learn from it, it can do it quickly. Lee Sedol spent 5 hours playing each of the three games he played against AlphaGo. AlphaGo could have played hundreds, if not thousands, of games against itself in the same amount of time. So that makes up for the number of games computers have to play to learn something.
Computers also adapt to changes. Lee Sedol beat AlphaGo in the fourth game of Go, but I'm sure AlphaGo learned from that game and will have changed it's strategy to account for Lee Sedol strategies. Computers will not learn as quickly from just one game as humans do, but they will learn something. AlphaGo learned the probability of beating Lee Sedol in the next game it plays against him. AlphaGo also learned the strategy Lee Sedol used to beat it. Lee Sedol can, however, use a different version of the strategy which may seem foreign to the AI until it learns it. Humans are good at using analogies to relate something new to something they already know. Computers cannot do that very well. The dream of any AI expert is to make an AI that can learn one thing and learn everything else by creating analogies.
But, since that perfect AI doesn't exist yet, any modern AI will need to perform a task millions of times to learn to perform the task at a human level.