MFML 009 - What is AI?

so what is ai well it depends on whom you ask ai was named pretty long time ago in the 50s

artificial intelligence and back then it kind of meant something else in regular usage and i wonder sometimes whether today's

calculators would have seemed like ai to folks in the 50s sometimes i like to joke that that version of ai was whatever would impress would have impressed those folks and

whenever that was achieved you just move the goal posts away a little bit but in that old school sense it is a sort of super set

and machine learning would be in there somewhere where machine learning is let's remind ourselves thing labeling using examples but the way that ai is talked about today

is actually something else it's more of a subset to machine learning so let's have a look at some ai tasks and see what they have in common there are tasks like figuring out

whether an image has a cat in it or not having a natural sounding human sounding conversation playing games that you don't know the rules for that you have to use

intuition to figure out the rules well when you the human do these tasks you are taking in information through your senses

and you get the answer as if by magic you don't even know how you know that's a cat you don't know how you know how to move that joystick in that game

you just know it you do something with these pixels and you don't know what you do now think about trying to solve a task like this the traditional way

you have to really think hard brainstorm what to do with every pixel and then give the computer those instructions by hand here's how you take the pixels for this photograph and here's how you figure out whether it has

a cat in it or not now what recipe are you going to write and as you're thinking up that recipe are you sure it's still going to work for that situation

it's pretty hard to write these rules you need complicated instructions to do this task and you have the benefit of eons of evolution your brain just does this you have no idea how it does it

and the task is easy for you and generating the examples and checking whether the task is done correctly that's really easy for you but you don't know how you do the task

so how can you express that to a computer you can't solve it the old way ai the way we think of it today

is about succeeding at those complicated tasks that programmers cannot write instructions for by hand and you need super flexible algorithms

neural networks and that is part of a class of stuff called deep learning it's part of machine learning and so when people say ai today they

tend to mean deep learning that's the way that it's used so solving these really complicated tasks that you couldn't solve a different way except by teaching the computer with examples

so this is about automating the ineffable you can't say how that task should be done you can't solve it the way that you do the calorie prediction now i also want to point out something

really powerful here when we as humans communicate with one another when we are trying to get another human to do something for us we have two modes of communication

available to us direct instructions or hey look at a bunch of examples and you figure it out

and we use both as the situation demands before machine learning we didn't have the ability to communicate that way with computers all we could do is give the instructions

directly so that's like a huge part of our natural communication that we want to use gagged think of this as an un-gagging

now as a programmer you can communicate with the computer two different ways so don't think of this as something sci-fi and

robotsy with a mind of its own think of it as unlocking a second way to get computers to do stuff for you that's what we're about and this is

really powerful because it's a whole class of tasks that you can automate that you just couldn't automate before