Like the built-in type defaulting rules these plugins can propose candidates
to resolve ambiguous type variables.

Machine learning and other large APIs like those for game engines introduce
new numeric types and other complex typed APIs. The built-in defaulting
mechanism isn't powerful enough to resolve ambiguous types in these cases forcing
users to specify minutia that they might not even know how to do. There is
an example defaulting plugin linked in the documentation. Applications include
defaulting the device a computation executes on, if a gradient should be
computed for a tensor, or the size of a tensor.

See https://github.com/ghc-proposals/ghc-proposals/pull/396 for details.