Say you want a bot to identify what is in a picture, is it a dog or a cat? Well, this task is very easy for humans, even for little baby humans but it’s impossible for humans to explain it to bots in bot-language how to do it!
Because humans simply know what is a dog and a cat! A human can say in words the different characteristics of the two but bots don’t understand words.
Now, a builder that builds at random, and a teacher that tests not teaches, and the students that don’t learn and are what they are- in theory, shouldn’t work. But, in practice, It does!
Partly because in every iteration, the builder-house-slaughterhouse keeps the best and discards the rest.
The teacher bot isn’t overseeing an old tiny school of a dozen or so students; rather, an infinite warehouse of thousands of students.
And how many times the test-build, test-build series is repeated?
As many as necessary (infinitely)!
Eventually, a student bot emerges that can barely identify between dogs or cats. Say, it scores less than 20% grades in the test.
As this bot is copied and changed, slowly, the average test-score rises and the grade required to sustain the test rises higher and higher.
Eventually, from the infinite student warehouse (slaughterhouse), a student bot will emerge who can tell a dog from a cat from a photo it has never seen before, pretty well! (attains scores more than 95%)
But how does the student bot does this? Neither the teacher bot knows, nor the builder bot, nor the human overseer can understand. Most importantly, not even the student bot itself can tell how it can recognize the dog from a cat.
After several successful iterations, the changes in the wirings of the bot’s brain is incredibly complicated. And while an individual line of code may be understood, and clusters of code, general purposed, vaguely grasped, the whole is beyond!
Nonetheless, it Works!
But, this is frustrating, especially because the student bot is very good at exactly the kind of questions it has been taught to. It’s great with photos, but with videos, inverted photos, things that are not dogs or cats, it gets baffled.
Since the teacher bot can not teach, the human overseer can give the teacher bot more questions and get the tests long including questions that bot still isn’t good with answering.
This is important, this is actually why companies are obsessed with collecting data.
More data means longer tests, which means better bots.
So when you get the “Are you human?” tests on google, you are not only proving that you are not a bot, but are also helping humans build the tests that can tell a bot what are lakes, mountains, traffic signals, horses from humans, etc.
Now you understand what could the recently asked traffic questions could be useful for building? (AI-bot driven vehicles)
There is another kind of test that builds itself- i.e. the tests on the human!
For example, say hypothetically a human overseer wants users to keep watching his website videos for as long as possible.
Well, it is easy to measure how long a user stays online to watch a video.
Thus, a teacher bot gives each student bot a bunch of users to track how long they watch videos online and what kinds of videos they like.
The student bots do their best to show the users what they would like based on their previously watched videos.
The longer the time spent on watching videos, the higher the score the bot acquires.
A million cycles later, a student bot emerges that is pretty good at keeping the users watching. At least compared to what the humans could do.
But when people ask how does the Netflix algorithm select videos? There isn’t a great answer other than pointing to the bot and the user-data it had access to and most vitally, how the human overseer directs the teacher bot to score tests. Because that’s what the student bot is trying to be good at to survive.
But what the bot is thinking? Or how it thinks is not really knowable!
As our algorithmic buddies are everywhere and not going anywhere.
All that is knowable is the successful student bot gets to be the algorithm because it is 0.001% better than the previous bot at the test that human decides.
So everywhere on the internet, behind the screen are tests to increase user-interaction, set prices just right to maximize revenue, or pick the posts from all your friends you like the most, articles people share the most, or whatever.
What’s testable is teachable, and a student bot will graduate from the warehouse to become the algorithm of its domain. At least for a little while, till the time a better bot overtakes it.
Today we are increasingly in a position only to guide the bots to test, and we need to get comfortable with that.