When we first started programming the Connection Machine, we figured out actually what was wrong with Amdahl's Law, which is Amdahl's Law always assumed that you broke up the program so that different parts of the program ran on different parts of the computer. And of course there's only so many ways to break up a task, and you can't break it up in an even way. But the more interesting way to program the Connection Machine was to break up the data and kind of do the same thing to all the data. So we worked out this program paradigm, which we called Data Parallel Programming to contrast it with Controlled Parallel Programming, which is breaking up the program. And basically it's now what's called MapReduce. The basic idea is that you have lots of pieces of data in lots of different processors. So the first operation you do is you do a Map operation. You apply the same function to every piece of data. And then what you... so it might be say a filter function. So let's say that I'm trying to count the number of things that are greater than five. So the first function is: I map the test if it's greater than five across all the pieces of data. Now I want to count them, that's a reduction. So somehow you bring all that information together into one number. And so I might add them up, I might count them, I might average them. And so, actually, because we knew we wanted a program and I designed the Connection Machine so that it had special hardware for doing those two functions... It had special hardware for mapping a program across everything, so you could broadcast out the same program to everything. And then it had a special network for reducing. So that was built into the hardware of the Connection Machine, MapReduce. It was basically a MapReduce machine. And in my thesis, I called that map and reduce Alpha and Beta, that you could... basically, you'd reduce over any associative function, like add or average. And you could map any function. You could just repeat it on each of the data elements. And so that paradigm for programming, MapReduce, is actually used... actually it got popularised, probably through Sergey Brin, who was a Connection Machine programmer. Google used it to do their machines, even when they went on and didn't use Connection Machines, they still used that MapReduce paradigm and then some people that spun out of Google created this thing called Hadoop, and so now that's basically the way an awful lot of parallel programming is done. But it's all kind of derived from the actual hardware of the way that the Connection Machine worked.