You control the red block with the arrow keys. You need to catch the green block, which moves around randomly. The tiny purple block will try to catch you. It starts out pretty dumb, but will become a worthy opponent in a minute or so.
So how does it work? The purple dot acts on five pieces of information: Its own position, its own velocity, your position relative to it, your velocity relative to it and the position of the green block relative to it. Each of the five pieces have an x and a y component. It multiplies each of the five x-components with a number and sum them to get its acclereration along the x-direction and likewise for its acceleration along the y-direction. Its maximal acceleration is slightly less than the red blocks', so you have the advantage. The five numbers needed to calculate where to go is called its weights. They start as small, totally random numbers. Thus, in the beginning the purple dot is prone to just getting stuck in a corner doing nothing.
So how does it learn? It uses a simple gradient descent algorithm. Actually there are twelve purple dots, you can just only see one of them. The others all have slightly different weights and are moving around unseen. Every second or so the distance from each purple dot to you is measured and saved along with the dot's weights. This data are then used to compute a gradient, the best way to change the weights in order to get closer to the red dot. The purple dots' weights are then shifted along this gradient and they are all moved to the position of the single visible dot again. With time these small movements will roughly optimize the weights, and you will experience the purple dot getting better and better at pursuing you.
Good luck!