AI Diagnostics
                     AI Diagnostics
I think the key question people should ask 
when a company claims to be using AI is, 
"What's the downstream behavior you are 
trying to optimize?" That will tell you what 
they're really up to.
In the case of any given application, it may 
not be obvious. The most useful way to 
think about machine learning may be as a 
set of tools to do diagnostics. 
Know More-AI Diagnostics
What makes it hard to see this is that we think 
of diagnostics as something that only happens 
after we have made a decision. You go to the 
doctor, and she first makes some kind of 
diagnosis, then tells you what she thinks you 
should do about it. But in fact the two are not 
separate.
 When your doctor examines you, she is 
collecting data so that she can later refine 
her diagnosis and tell you what you should 
do about it.
This is true whether the diagnosis is made by a human or a computer, 
and whether it is based on symptoms or genetic screening or anything 
else. For example, we might notice that people who have a certain gene 
are more likely to have some disease than are other people, and use 
that as a test for the disease. 
That's what all medical tests do: they look 
for population-level differences between 
people with and without some disease or 
condition.
In some sense, the dream of AI researchers 
is to be able to tell, "If you do X, then Y will 
happen." The dream is a regulatory agency 
that can make an accurate prediction 
about the world.
It's not just in software. We'd like to know 
what happens if we raise interest rates or 
take a new drug. If we could be certain of 
the outcome, it would be easier to choose 
our actions.
But this doesn't work for complex systems. 
If you ask someone who eats ice cream 
every day whether they'll get a brain 
freeze if they eat ice cream today, they'll 
say yes. There's no mystery there: they've 
seen it happen before and expect it will 
happen again..