I thought the discussion he had on the performance benefits of Toplink was interesting. I remember having a similar discussion a while ago where we considering using stateless sessions for our readonly web services. This isn't quite as good as the Toplink design since even on our write operations we typically read data we don't intend to modify to do various types of validation.
I was little surprised he supports use case optimal queries which from what I can gather is using sql to create DTOs or value objects and returning them from your repositories. I would have thought he would only recommend that in cases of extreme performance problems. He did say if you are doing it a lot then it is probably a code smell.
He also has a section about Type hierarchies and how you should limit use of them. I tried implementing a Type hierarchy in NHibernate and it didn't go well for me. He had some good ideas about how to hide the types inside the aggregate and how to not let them leak out to be used by other aggregates.
I was also surprised he didn't give an example of how to test a repository with a mock. He did give an example of how to test with a real database and with a stub though.