There is a very understandable reason why most people underestimate AI.
We see improvements only in gradual increments and are unable to comprehend the exponential complexity behind some systems.
One of the best examples for me to demonstrate this is the Rice on a chess field Problem:
Imagine a chessboard and that you have a grain of rice on the first field. For every field after the first, you always have to double the amount of rice…
This is a pretty common explanation that demonstrates the idea of exponential growth.
At the end of the chessboard you would need more rice then there is currently in the world.
But why is AI exponential?
AI gets also misunderstood because when we think about learning we mostly think of human learning.
But AI has not only the possibility to scale with data, it has also the ability to scale with itself.
Take the game of chess, the AI could learn from a human player for instance.
That is how most people picture how it works.
Of course, you can also give the first AI player the same AI as the second player as well. Now the same AI that wins also loses and learns from both experiences…
Now remember that time for AI is different – some games between AI´s are done in under 5 minutes because of how fast the moves of each AI player would be. If you account for that, then an AI can still learn 288 Games in one day. – Which would be quite impressive and better than any human in the world.
But still not enough with that, because you can make 1000 AI´s play each other. Now you have 288.000 Games per day the same AI learns from. And every improvement gets copied in every other version of that said AI.
Self Drive AI´s right now for automatic driving can train in a simulator and drive 65000 miles a day – that’s what normal people drive in 5 years. There is just no way a human could ever compete in regard of driving experience.
The last step where AI is scaling is in the sense that a new AI is improving other AI´s.
Let’s say we stay with the driving AI for automatically driving a car. When it is perfected, we can take that same AI and have a starting point for the next forklift for example.
AI can just multiply all the parameters that are effective in learning new skills:
- the dataset it works with
- the experience it gets
- the time it costs to do everything
Just imagine how much we could learn if the best experts in a certain field could solve complex problems every minute and we could just copy that experience to everybody else in the field – That’s AI learning.