[mindslight]

I think the algorithm is most likely worthwhile, but the article is terrible. Another instance of journalism adding a narrative meant to entertain but really just disinforms.

Also if I had to guess without studying the paper, the benefit of the sum of the weights being a power of two isn't just being efficient in memory or time, but rather efficient in use of the random input tape. Naively, if you want to generate a uniform number between 0-128 inclusive from an input of uniform bits, you consume 8 random bits and if they form a value within the range you return it. Otherwise (129-255) you throw all the bits away and start over. Using the leftover entropy seems like it should be possible somehow.