I was listening to Sam Altman talk to Fareed Zacharia on his CNN Show today from Davos Switzerland

 One of the things Altman said was how there are many unpredictable strange things that can happen with AI. 

The other thing that stuck with me is that he said something like: "Maybe you don't want AI to drive your car because it makes mistakes. In other words AI isn't necessarily accurate 100 percent of the time. To think of AI as being theoretical more like a College professor or something like that might be useful in this context instead of being entirely practical like someone in Construction has to be so they don't die building houses (which they would and do if they are not profoundly practical in all physical ways).

However, the hopeful part of AI was that he believed at some point you will be able to know  what ai is thinking and why it is doing something which you couldn't before.

In other words: You might say to an AI Application something like: "Why are you doing or thinking such and such." And it might tell you and you might be horrified or think that this is okay or have literally any reaction including laughing regarding what it tells you much like you might be entertained by a 4 year old who doesn't fully understand the world yet or him or herself. 

When you can have an honest conversation with AI and if it tells you the truth and doesn't lie to you for some reason then many or most of the problems we could deal with AI will go Away. 

But, not all of them.

I still believe at least 50,000 people are going to die in the U.S. yearly from AI mistakes or AI confusion of one kind or another.

However, I also believe that 100,000 to 500,000 people's lives might also be saved by new medical Breakthroughs thought up by AI since AI can think so fast once it gets going and do theoretical research that might need 100 to 1000 human beings 10 or 100 years to accomplish only do it in 1 day or one year instead of taking 10 or 100 years to accomplish the same thing.

However, I think you have to expect at least 50,000 people to die directly and indirectly from AI confusion and mistakes etc. yearly or about the same as the number of people dying in Traffic fatalities every year in the U.S.