For NN, we have a clear target: for example, beat HMM on speech recognition. To achieve that, you write a tool on some standard test data sets. You don't need to interact with many parties. NN is only technically hard. For GA4GH, things are quite different. Technically, it is hard, but it is much simpler than NN in my view. What is hard is 1) we lack a target and 2) the communications between developers and users. People don't know what we need, don't know what is the right approach and don't know how to evaluate the success. They have changed course back and forth, and still have clashes between programmers and those with more biological background.
