What The Heck Is AI Anyway?

AI is everywhere. We all know what it is. According to wikipedia

That’s Not Helpful

I understand what each of those words mean, but the way they are assembled doesn’t make a whole lot of sense to me.

I was speaking with a friend about some new AI features coming soon to hearing aids. After a bit, he bravely mentioned that he really didn’t understand this AI stuff and suggested others may feel this way and maybe I could help.

Intelligence

Let’s start here. Many of us have intelligence. Show of hands…

Here’s a simple example. I touch the hot burner on the stove. I feel pain and pull my finger away. I subconsciously decide not to do that again. That’s intelligence in action.

Computers

Artificial intelligence involves computers. We know that because wikipedia told us.

In my day, a computer was a big heavy box with wires and cables and lights and maybe a keyboard (remember punch cards? tape?)

Today, they are a little tougher to recognize. I have a computer in my ear right now. It’s in my 1 remaining hearing aid. Read here about how I jumped in the pool and destroyed one hearing aid. This demonstrates the opposite of intelligence.

A computer by itself is worthless. It doesn’t do anything. Don’t believe me? Delete all the software from your computer, including wiping the hard-drive. The lights won’t even come on.

Software

Software, or computer programs, make a computer do fancy things. Computers have a base program, called an operating system, which tells the computer how to act and how to interact with peripherals – things like printers and external drives.

Using the computer and the operating system, programmers create computer programs that perform cool functions. Think of a computer program as a comprehensive set of rules. I began my career as a programmer at an insurance company. I had to write a program to pay commissions to sales folks.

My program had to find who the salesperson was, what they sold, how much it cost, and calculate a commission. There were a lot of rules. The more they sold, the more they got paid so I had to write a rule that once they sold more than X dollars of insurance, increase the commission by Y amount.

A computer is very good at reading and writing data, calculating quickly, and performing repetitive tasks, but a programmer has to tell it every single thing to do. It’s useless on its own.

Self-Driving

Commissions are really dull, so let’s jump to self-driving. I’ve read a bit. It’s fascinating.

I watched a video yesterday on BMW’s diver assistance package. It can enable the car to change lanes when we put the blinker on. No need to bother with the pesky task of turning the steering wheel. Think of the time and effort this saves.

Let’s think about what some BMW software developer needed to do to create this function. Lots of rules, right?

First, the car needs to see if there is a lane available. Imagine hitting the blinker while in a tunnel. You’d smash into the wall. Better write a rule for this.

Also, we better make sure no one is approaching in that lane. The car needs to look at its sensor or camera and see if it’s safe to move into the lane.

It also needs to look ahead. We don’t want to move into a lane where a vehicle ahead is stopped or slowing down. Another rule or 2.

That’s Complicated

We haven’t even gotten to AI yet. For now, we’re just talking about the explicit rules that we need to put into our computer program to create a lane change feature. I’m sure there are a lot more rules, but you get the idea. All the things that a person needs to do to change lanes need to be analyzed and documented and specifically written into very clear rules for the computer to follow.

Why Do We Need AI?

If we can write rules to make a complex feature like lane change, why do we need AI? We could continue to build more cool and more complex features like self-parking, and summon, to make the car come to us, we just need to write lots and lots of rules.

Cut Out The Middle-Man

Kind of a cliche in business, but cutting out the middle-man means streamlining. In the case of AI and computers, we’re the middle-man.

Wouldn’t it be cool if we didn’t have to do all the hard work of gathering all the requirements and writing rules. Computers are great at repetitive tasks. Couldn’t they do this?

Example, Please

Let’s go back to self-driving and the millions of rules we’d need to write to have a car handle everything we do as drivers. We’d have to analyze every situation, stop signs, traffic lights, merging…think of all the different things that could happen and write a rule for every one.

AI, or machine-learning, is the capability for the computer to create these rules on its own.

To do this, the computer needs to have sensors like cameras and radar that act as eyes do for people. Then it needs to have a framework (a more general set of rules) for processing and organizing the information it gathers and to organize it into the same set of rules that we would have written.

Tesla

It’s no surprise that Tesla has quite a head start here. There are a lot of Teslas on the road. I read an article a while back that Tesla’s share of the US automarket had dropped significantly – down to only 49%. They have a lot of cars on the road.

Teslas also have a lot of cameras. Tesla made a decision to use cameras over radar and lidar for its self-driving features. Turns out this was a good idea as it now has over 780 million miles of driving data.

Who Cares About Driving Data?

You may have heard the term “training an AI model”. Essentially, we need to send the computer to school. It needs to learn all the things we learn in order for it to have “intelligence”. This is a bit of a stretch. Our car doesn’t need to learn Shakespeare, or German, but it does need to learn everything about traffic signs, roads, intersections…

Tesla feeds its 780 million miles of driving data into its AI model to teach it what happens in different traffic situations.

How to Handle an Intersection

So Tesla’s AI model can absorb all of this driving data and see how it should handle a 4 way stop. It’s going to have to write a bunch of rules. Slow down as it approaches, come to a complete stop, note what other cars are at the intersection and when they arrived, go when it’s my turn. And probably a few more.

The AI model is feeling pretty good about its ability to handle an intersection so it goes out on a road-test.

It comes to an intersection, slows down, stops, waits its turn, and then proceeds slowly through the intersection. Just then, a kid who’s texting on his phone barrels through the intersection and t-bones it.

What happened? We followed all the rules?

Learning

And here’s the big deal with AI. Our vehicle, and its AI model, now have a new experience. Another driver who doesn’t follow the rule and stop.

Our model needs to learn that it also needs to look for trouble. A car going too fast to stop is trouble. Our model learns how fast is too fast, and it learns to stay in its position until the danger is past.

Those 780 million miles of driving data provide lots of situations that our AI model can lean from.

When is it Done Learning?

Never.

At some point, car manufacturers, The National Highway Traffic Safety Administration, and consumers will become comfortable and will allow/adopt self-driving cars, but our cars will continue to watch, experience and learn, and most importantly, will adapt its rules and behaviors to make better decisions.

That Sounds Good

Doesn’t it? Computers don’t forget. I’d like to think that I won’t dive into the pool again with my hearing aids on, but I probably will at some point. But once a computer has a rule, it’s there forever or until it is superseded by a better rule.

Better?

Better for whom?

This is the big question.

Watch the movie Subservience. It’s about an AI Robot built to take care of the home. The robot “learns” some things that a good ole’ fashioned programmer probably would not have written rules for – like to kill someone who threatens your home.

Once the computers cut out the middle-man and start writing their own rules, we may find that we don’t like the rules.

Wrap-Up

For the amount that the phrase AI is tossed around, there is amazingly little information that helps most of us conceptualize what it really means. Definitions, like the one from wikipedia are common, and while technically correct, really don’t help.

I thought I’d take a stab at explaining how I understand AI and relating it to something like driving that we’re all familiar with.

I hope this helps. Let me know.