They are usually private datasets and details of the training are not usually disclosed. There could also be duplicated effort by every company competing in the field.

There is also the side that it would make it harder for startups to compete as the progress is made on training on large datasets using huge clusters of machines. Such datasets and compute power is impossible to acquire for small companies.

So maybe we need to start funding the acquisition of large public datasets in academia?

The fact that Stanford University creates patents from publicly funded research then turns around and sells them to non manufacturing entities makes me wary of pumping any new public money into research at universities. They're clearly corrupt to the core now.

I haven't heard of this before. This is pretty interesting. Do you have a source?

There was recently an episode of Planet Money about one instance of it:

http://www.npr.org/sections/money/2016/06/10/481597112/episo...

Or somehow incentivize companies to follow an academic model.