I think it's important to point out that this is their second attempt at making a hedge fund with crowdsourced investment strategies. Their first attempt failed because they would pay out rewards (bitcoins) based on the performance in a historical benchmark. This made the whole machine learning competition susceptible to intentional overfitting (common problem in the HF world). They outline the problem in their white paper: https://numer.ai/whitepaper.pdf
More details: https://numer.ai/whitepaper.pdf
Really interesting though i'm not sure I agree it changes the incentive structure of the competition - more players means you need to bid more aggressively to see any of the prize pool. It does make it harder to game their private validation dataset which is likely the goal here.
The real brilliance of numer.ai is that for not even the salary of one SF data scientist they've managed to employ an army of 'em.