Lesson #1: you don’t know where you fall in a statistical distribution.
No shit. Really? There is a 1% chance of failure in a game. If I play the game an infinite number of times, I will fail to win 1% of the time.
The alleged twist: if the nature of the game is that a single failure results in you no longer being able to participate in the game, you cannot play the game an infinite number of times.
Color my mind blown. I’m absolutely shocked that Murray Gell-Mann was able to grasp this concept so quickly. A friend of mine has a PhD in Statistics, and I remember once telling him that the probability of a fair coin tossed an infinite number of times and landing on heads approaches 50%. He grasped that concept immediately. I was impressed. We started a company to explain this.
I’m kidding. That didn’t happen because it would be stupid in every possible way.
Lesson #2: Cost benefit analysis is impossible if you don’t know where you are in the distribution.
Somehow, not knowing where you are in the range of possible outcomes means that you can’t know the range of possible outcomes.
The alleged twist: even when you know the range of possible outcomes, you apparently don’t.
I don’t even know where to start with this. I can’t even make fun of it properly.
When you’re gambling, the range of possible outcomes is well-defined for most people. You lose all your money, you break even, or you win. This model is incomplete because it doesn’t take into account human pathology. Some people cheat. Some people take out loans from shady people and lose everything and get a leg broken. Some people commit suicide because they lost a few hundred dollars.
Guess what, if you include human behavior in the model of Russian roulette, it’s also a broken model. People will cheat. People will chicken out. People will want to die and pull the trigger twice. That does not make a cost/benefit analysis “undefinable,” it just makes the range of possible outcomes larger than most people are comfortable with. That’s why most people choose to not play the game.
What Taleb does here (again) is conflate a legitimate criticism of frequentist statistical models with a critique of incomplete decision models.
But the critique of frequentist models is entirely sophomoric. It boils down to, “You can’t do something an infinite number of times. Therefore, everything we know is wrong.”
Fuck off, man. This gets addressed in every basic statistics intro. I’m not saying there aren’t problems with the idea. There are. But it’s the same problem as saying that the limit of x as y approaches infinity is 1. Is calculus fucked and completely broken because y never gets to infinity?
No. Not even a little.
Lesson #3: social sciences are broken.
Completely agree.
The alleged twist: They are broken because the basic concept of probability is broken.
The absolute comically stupid apex of all of this linguistic garbage is that Taleb uses the social sciences to prove how little his argument makes sense.
He can’t tell the difference between a bad model and a broken system for creating models. He proudly claims that probability is broken because the models that people create to estimate outcomes are bad. And he uses the worst possible collection of models for the most complex systems that we know of as proof that the system is broken.
There are absolutely legitimate criticisms not only of frequentist statistics, but of inductive logic in general. It’s only a conversation that’s been going on since, I don’t know, Pythogras? Maybe before?
There is nothing new or insightful here. Taleb is wrapping up and idea that everyone already knows in a word cluster so obtuse my 5 fifth-grade English teacher would destroy as some deep new insight.
But hey, he posted it on Medium, so it must be awesome.
Give me a break.