Meta and Google open sources PyTorch and Tensorflow so people can hopefully make one better than the other.
That’s a bit optimistic at this point… PyTorch is basically the Windows of machine learning libraries – it’s not particularly great on its own merits, because a lot of core features (XLA, jit compiling) are clearly added as afterthoughts and have a lot of very apparent issues, but everyone uses it because everyone uses it. It is a perfect expression of why maybe “move fast and break things” isn’t such a great philosophy for important libraries.
That’s a bit optimistic at this point… PyTorch is basically the Windows of machine learning libraries – it’s not particularly great on its own merits, because a lot of core features (XLA, jit compiling) are clearly added as afterthoughts and have a lot of very apparent issues, but everyone uses it because everyone uses it. It is a perfect expression of why maybe “move fast and break things” isn’t such a great philosophy for important libraries.