I'd stay far away from this one until they demonstrate proof-of-concept code that they can actually do what they promise. The whitepaper is full of technobabble that says, using a lot of words, very little that makes one think they have actual insight into how to solve the hard problems in this area.
It's not clear that one can turn these kinds of problems into a proof-of-work with the right properties;
it's not clear that widely distributed training (particularly using the kinds of rigs miners use) will be effective;
and i wouldn't put my dollars on it vs. cluster solutions with high-performance interconnects of the type we're seeing used by NVidia (DGX-1), Google (TPU clusters), Uber (GPUs with a local, very low-latency/high speed Infiniband interconnect), and others.
Even if it's not a scam, which I suspect it is, it's very unlikely to actually work. And those faked-up graphs are just awful.
Your friendly neighborhood non-anonymous computer science professor,
-Dave
--
http://www.cs.cmu.edu/~dga/