kave wins

Update:

I'm currently not finding much evidence of any phase change for the 64 batch

There is a stark decrease on 400, but I don't know if that would stick around if I increased the epoch to 800. 

To speed things up I also reconfigured how training works, so now I'm verifying we still get a phase transition with the 128 batch size. Then after that I think I'm going to increase the batch size instead of decreasing it, to make the results come in quicker. Because the above plot took 9 hours of training to make.

Henceforth I will no longer bet on the market

Betting heavily on myself here not due to epistemology, but to subsidize people to bet differently from me, so I can hear their arguments.