On Fri, Jul 13, 2018 at 06:44:39PM +0200, Jonas Smedegaard wrote:
> I therefore believe there is no license violation, as long as the code
> is _possible_ to compile without non-free code (e.g. blobs to activate
> GPUs) - even if ridiculously expensive in either time or hardware.
>
> We have a practical problem in distributing such code, however, if every
> package release blocks our build daemons for 100+ years.
So if the Neural Net is self-training (e.g., the software plays Go
against itself for some huge number of GPU years), and the initial
weights, as well as the results of the self-training process, I would
claim that this would be GPL compliant.
This is equivalent of e2fsprogs distributing both the configure.ac and
configure file. The configure.ac file is the preferred form of
modification, and the configure file is distributed because creating
configure from configure.ac using a newer version of autoconf is not
guaranteed to result in a working (or properly working) configure file.
The question of whether we are obliged to rebuild the neural network
every time the package is compiled is a policy matter for Debian
(although I will note that no one forces package maintainers to
rebuild configure from configure.ac today) --- but it's not IMHO a
license compliance issue.
- Ted
