This information was stated in his Mailchimp email update he sent out yesterday.
I've already read the first 12 chapters. I found it very practical for developers who have already worked with NNs. Basically the book discusses common problems and how to deal with them. Gold resource.
That's not true. Accuracy based on assuming an underlying binomial distribution has standard deviation sqrt( p (1 - p) / n ). Unless your accuracy is very near to 0 or 1, you're going to need more like a million samples to have a good chance of detecting a difference of 0.001 and be able to reject the hypothesis that it was not caused by statistical noise.
If the assumption is not an underlying binomial, can you explain why? I was hoping for more a lot more hard maths in this text.