They would need it to interoperate with their existing stack. They chose Kotlin where they reasonably could have chosen Swift and then worked to extend it to work with their existing stack.
Judging by what I would argue has been a lukewarm reception to Swift for Tensorflow, I think it would a very hard sell to convince people within Google to support Swift as a first-class language in the rest of their ecosystem.
They could for example start by adopting Swift for server-side programming for Google cloud and they seem to be putting no effort into that, at least public-facing effort.
May 2017 - https://www.theverge.com/2017/5/17/15654988/google-jet-brain...
You could use Kotlin for years before Google officially endorsed it...given that Android Studio is made by JetBrains and they made it trivial to use. JetBrains did 100% of the heavy lifting, and Google just conceded eventually.
Swift for Tensorflow hasn't gotten much traction because the overlap between programmers who know swift and machine learning engineers is super tiny. Making the core language differentiable sounds very promising though .
There was probably several months or more of debate and work within Google by the time of the announcement, but it seems reasonable that during those 18 months they would have also considered and rejected Swift.
In any case, open source Swift was not ready for prime time at time of release.
I think it's going to be hard to get data-scientists using anything other than Python any time soon, but I'm hopeful it will take off at some point, since it's really painful to work with scripting languages once you get used to a decent type system.
It's also a super interesting project to follow. The stuff they are doing with automatic differentiation, and the stuff they are talking about with metaprogramming is truly fascinating stuff, with all kinds of applications outside of machine learning. I really hope they manage to get it off the ground.
Still given the growth of that industry, it's hard to imagine that Python can bear the burden of the increasingly complex work there forever.
Once you get to the modelling step all of the data is in typed arrays.
Static types start to shine once the project grows and the new type annotation syntax works great for that.
I've tried doing machine learning in C++, Java, Scala and JavaScript and all of them are a pain compared to python.
Python would be perfect if I could annotate parts of my code and compile them. (Pretty much what's commonly done with cython or numba jit.
In my experience, doing this kind of task in an weak-typed or duck-typed language is already tedious. Function signatures don't tell me enough detail to know exactly what I'm working with, so I end up spending a lot of time digging through documentation or print debugging. Why not let the compiler, or the function signatures do this work for you?
If they want to make an Apple only version of Tensorflow, so be it, but don't expect it to take off on Linux and Windows research labs.
Edit: or on Ubuntu just download the toolchain.
But I would assume a lot of people just want a notebook to do their work in. Does it really matter if it’s running in a container or natively? Most python projects run inside a virtual environment anyway, which is comparable if not more poorly encapsulated.
Also the cloud use-case is expanding. Environment basically doesn’t matter at all in that case.
Yes it matters, because many organisations get development stations managed by IT, and there are very strict compliance regulations regarding where data processing takes place.
Kotlin was open sourced around 2012 and people were using it in Android apps before Google officially started supporting it in 2016.
Swift open source 2015/12/3: https://developer.apple.com/swift/blog/?id=34