@madhu_shrieks@sciences.social
First, kudos to the author for an engaging presentation. Second, there are several elements that appear misleading.

Caveat: I don't have access to JStor and can't seem to find the original paper on the Am Scientist website.

In the two person interaction, I understand the wealth dynamics to work like this. Each person wagers a fraction f of their wealth w1, w2 and the actual wager is f*min{w1,w2}. WLOG we can set w1+w2=1, so person 1's wealth follows the process:

w1=w1+p*f*min{w1,1-w1}

where p takes on the values -1 and 1 with equal probability. The right hand side has a zero expected value so 1's expected wealth does not change over time. So the statement that "the rich wind up with all the wealth" in this game is inaccurate. What does happen is that *someone* winds up with almost the wealth with high probability; the probabilities are determined by the initial wealth ratio w1/(w1+w2) because we know expected wealth is constant over time.

Here I am commenting on the statistical process; as a model of economic activity, what it shows is that even gambling with a zero expected loss is a loser for a risk averse person, as should be obvious. A fair bet still represents a garbling (in the Blackwell sense) of wealth and hence reduces expected utility for concave utility functions.

This is why there is a risk premium: risk averse people should only gamble when there is a positive expected return.