I think the point of GP is that astrophotography as demonstrated here is mostly implementable in software. If Google wanted, they could very well produce an app for iPhone that does all of the above.
It’s hardly a secret that innovation in smartphone cameras is mainly in software now. This software camera innovation is one of, or the, main area that phone manufacturers are competing on at the top end of the market, so characterising it as ‘easy’ seems strange.
It seems they have some ML stuff in there for specific features, e.g. sky / land light balance and hot pixel removal (probably similar to how denoising for MC path tracing works).
Aside: randomly recognised Ryan Geiss in the credits, he did the Milkdrop plugin for Winamp back in the day, and also some cool tech demos for Nvidia...
There's no machine learning involved for hot pixel removal. The way I have heard it described is that hot pixels stay in the same place across multiple images. One of the processing steps is to figure out the location of the hot pixels by looking at multiple frames. Stars will shift slightly in between the longer exposures where as hot pixels will not.
In that example the Pixel has a 6x longer exposure. While Pixels can currently get better night photos currently, it's simply because the software lets them take minutes-long exposures instead of seconds-long. From a hardware point of view I don't think there's an advantage.
Both the Pixel and iPhone have AMAZING camera hardware and software that is better than any others in their domain. It’s just that the two are roughly equal except for the exposure time used.
* Compensating for moving stars
* "Live viewfinder" during exposure
* Selectively darkening the sky
* Dark current compensation (though that is probably needed for all long-exposure photography...still, not a simple "more exposure" feature)