Rafael Rayes

Why Don't Zebras Have Lasers?

A zebra with lasers would certainly be a formidable animal, eliminating its predators with ease thanks to this evolutionary advantage.

But for an ability to be unlocked through natural selection, every step of its evolution has to be advantageous.

When you build a laser, you first put together several parts that do nothing on their own, until you add the battery and the laser turns on. Getting to a working laser takes many other steps that have no evolutionary advantage at all, and so they would never be selected. The laser would never be completed, even though it would be an advantage if it were.

The same concept applies to neural networks. Behaviors that are not advantageous at first are discarded, even if later on they could reach some extraordinary goal.

Society does the same thing with individuals. Far-fetched ideas are often discarded because they don’t work yet, and that way we lose countless brilliant minds and inventions.

It is possible to build the opposite into a neural network. You can make divergent behaviors keep being encouraged, even when they are not advantageous, in the hope that something better than the ordinary will emerge.

An example: imagine a neural network being trained to play a racing game. The default behavior that will quickly emerge in the first generations is to run. Now imagine that one instance of this AI tries to build a car at the start of the race. Because it can’t reach a good score right away, it would be discarded in the very first generations. But if it kept going, it is possible that in later generations the car would be ready, and it would win the race by a mile.

Translated from a note I wrote in Portuguese in February 2024.