[dekhn]

Everybody at Google wanted to add their specific feature to TF (gets visibility, users, research cred). Unfortunately, many different teams added multiple incompatible features that didn't compose. TF1 also had some serious problems where it was fundamentally designed around C++, without an understanding that most people wanted to work in Python. With TF2 a lot of stuff got redesigned, making many examples on the web stop working. There are too m any ways to parallelize your computation (along any of the dimensions), and they change too frequently.

But I think the writing was on the wall when the folks building ML Pathways hit some performance problems and realized that Jax made it much, much easier for them to express the computations they wanted and see them run quickly on TPUs (DeepMind had also concluded this). Once Jeff saw that Jax was making stuff that ran faster than TF (for his pet projects) the writing was on the wall.