Neil Girdhar

> I think assigning this a dedicated error code is a good idea. Finding a good compromise for everyone may be hard, and with e.g. --disable-error-code=typevar-abstract people will be able...

This is an excellent proposal. Some thoughts: If you're redesigning anyway, you may want to take a look at my [efax](https://github.com/NeilGirdhar/efax) library, which was designed with machine learning in mind....

> we would not be able to use any Jax code as we don't want to add another dependency. Of course! I only show it as illustration of an alternative...

> I don't believe that having to modify the shape parameters of a distribution is the common use of distributions. I am not saying this is not useful, but that...

What is the process for this redesign? Is there somewhere the interface is being discussed? I'd like to follow along the discussion if possible. I'm not sure what the usual...

> I think it would be more challenging than it's worth to share innards with the univariate distributions. I agree with you about the innards. Still, if you're going to...

> I'm not Jukka, but that sounds like a big project. It means code throughout the whole type checker needs to have a concept of types that are different depending...

What does merging the dataclasses consist of? Are flax dataclasses going to be inheriting the mapping interface?

> I think we should start to move towards allowing mutability Sorry, but why would you do that? > The merging of dataclasses is taking much longer than originally anticipated....

> Many users find it difficult to deal with frozen dataclasses/dicts. At the end of the day Python is not a functional language and partially making it behave like one...