

I think the idea is that the cost of producing standardized hardware is lower than the cost of producing a custom version without that codec just for the Raspberry Pi Foundation. The Raspberry Pi Foundation was not interested in that codec, so they didn’t buy a license. Separately, as a special agreement, they then allowed the few interested users to get a personal license directly from the IP owner. Sounds like a great solution to me.
Not sure if the same reasoning applies to BMW, though.
Adjusting the weights is how training works.
If you are thinking about adjusting the number of weights, that’s also a technique. The issue is that (afaik) it can only be trained through genetic algorithms, which take longer than backpropagation.
(This information might be incomplete/incorrect.)