I tend to think #2 is the approach that is most likely to lead to a "real" general intelligence - reverse engineer what we know works, replicate the essential "secrets" (whatever they are) and scale up and out.

I just don't see much progress on #1 - and people have been trying this approach for 50 years.