A young but earnest Zen student approached his teacher, and asked the Zen Master:
“If I work very hard and diligent how long will it take for me to find Zen.”
The Master thought about this, then replied, “Ten years.”
The student then said, “But what if I work very, very hard and really apply myself to learn fast – How long then ?”
Replied the Master, “Well, twenty years.”
“But, if I really, really work at it. How long then ?” asked the student.
“Thirty years,” replied the Master.
“But, I do not understand,” said the disappointed student.
“At each time that I say I will work harder, you say it will take me longer. Why do you say that ?”
Replied the Master,” When you have one eye on the goal, you only have one eye on the path.”

Zen Parable

In a world obsessed with objectives, what can we learn from not having any? Much more than you think, according AI researcher, Kenneth Stanley.

Stanley’s findings come from an online application called Picbreeder. Picbreeder allows users to ‘breed’ an image by branching through variations of it. Each time a new parent gets selected, and the process repeats. Users can publish any interesting images they manage to find. Others in turn may use these as a starting point for further breeding. The experiment has produced a variety of interesting images; a remarkable feat given that the initial image was simple blob.

What’s fascinating is how these images get discovered - by not trying to discover them. Almost every interesting image had a parent which in no way resembled the final product. There was no way to tell it would end up producing that final image. You had to ride the goalless path, and see where you’d end up.

Many of the interesting images get produced in 70 to 90 generations. Peanuts for a computer to crunch through. Could traditional objective based machine learning reproduce these results? Stanley’s team tried their best. Even after 30000 generations, each time the experiment was a dismal failure. The machine images looked nothing like the remarkable images discovered by humans in a fraction of the attempts.

A similar project to Picbreeder, the Living Image Project, relied on voting instead of branching. Users would vote for a the variation that they wanted to be the new parent, and the majority would decide. Yet despite far more activity, the images produced by the Living Image Project were boring in contrast to Picbreeder.

Some observations and notes:

Stanley’s findings have many general insights applicable to all areas of our world and lives. It has given me plenty to think about in my own. Perhaps this is something that Zen masters have always known and respected.