> I don't think that GA4GH is literally using CORBA.
It's not literally CORBA, but people who spent time implementing literal CORBA in a bioinformatics context (for instance, the original author of https://github.com/bioperl/bioperl-corba-server) have noted that the design pattern and discussions followed by the GA4GH are pretty similar to those had in the EBI when they attempted to unify everything using CORBA.
> The data model is not implicit, it's explicit (there is a schema).
There is a schema, but the semantics of the data model are encoded in the comments of the schema. Without hooking into some kind of ontological basis it doesn't seem possible to avoid this.
> As for the semantic web, every bio project I've seen which adopted the semantic web ultimately failed -- the semantic web seems like a great idea, but attempts to fully implement to the point where it's useful for research always fail.
I'm aware of at least one group in the GA4GH that uses RDF internally, then converts into the custom schemas produced by the group in order to maintain compatibility with the top-down designs of the project. I believe this is the phenotype group. These are the people who are most interested what the author of the linked page is describing, they have decided to use the technology you believe is doomed to fail. But, they aren't failing. As far as I can tell from their presentations they are one of two or three groups in the project that have produced a functioning system.
It's very easy so say that hard things are impossible. This tends to keep them that way. I doubt we have any other viable option for building large distributed knowledge systems. The fact that these don't exist does not mean they are impossible to construct, but simply that no one has managed to do so yet. People leveled the same kinds of arguments against neural networks up until a few years ago, saying that they were a nice idea but destined to fail because they were too hard.
> So I think they're focusing on areas where they are likely to succeed (collection and processing of large amounts of raw and derived data using pretty conventional processes, but at a much larger scale, with a solid authentication and access mechanism).
The scales we're talking about are not even an order of magnitude above that which existing techniques allow. So I agree that they will succeed insofar as they simply adopt these existing community-driven standards and slap access control on top. However, in terms of generating new data models for genomics, I'm not so convinced that the centralized design and API-based approach which they are taking will work. I guess we will have to meet back here in a few years and see what happened.