This is quite a surprise. Only recently, Loon collaborated with seasoned reinforcement learning researchers to build more efficient navigation controllers for the balloons, titled "Autonomous navigation of stratospheric balloons using reinforcement learning" [1]. Although, it would appear that much of the gains come from a simulator that they've been developing for a while.
Of course. My point was that it appeared like they were actively working on developing new technologies. Usually, a project that is expected to shut down would mostly lose momentum months before. Although, it is possible that this work was done months ago and only came out now due to the slower Nature review cycles.
[1]: https://www.nature.com/articles/s41586-020-2939-8