" Presumably the machine learning model has in some sense discovered Newtonian mechanics using the training data we fed it, since this is surely the most compact way to predict the position of the planets far into the future. "

To me, this seems to be an entirely unrealistic presumption (also true for any of its parallels; not just when it is strictly about the position of planets). Even the claim that NM is "surely the most compact [...]" is questionable, given that obviously we know from history that there had been models able to predict just the position of stars since ancient times, and in this hypothetical situation where we somehow have knowledge of the position of planets (maybe through developments in telescopic technology) there is no reason to assume analogous models with the ancient ones with stars couldn't apply, thus NM would not be specifically needed to be part of what the machine was calculating.

Furthermore, I have some issue with the author's sense that the machine calculating something is somehow calculating it in a manner which inherently allows for the calculation to be translatable in many ways. While a human thinker inevitably thinks in ways which are open to translation and adaptation, this is true because as humans we do not think in a set way: any thinking pattern or collections of such patterns can - in theory - consist of a vast number of different neural connections and variations. Only as a finished mental product can it seem to have a very set meaning. For example, if we ask a child if their food was nice, they may say "yes, it was", and we would have that statement as something meaning something set, but we would never actually be aware of the set neural coding of that reply, for the simple reason that there isn't just one.

For a machine, on the other hand, a calculation is inherently an output on a non-translatable, set basis. Which is another way of saying that the machine does not think. This problem isn't likely to be solved by just coding a machine in such a way that it could have many different possible "connections" when its output would be the same, cause with humans this happens naturally, and one can suspect that human thinking itself is in a way just a byproduct of something not tied to actual thinking but the sense of existence. Which is, again, another way of saying that a machine is not alive. Personally, I think AI in the way it is currently imagined, is not possible. Perhaps some hybrid of machine-dna may produce a type of AI, but it would again be due to the DNA forcing a sense of existence and it would still take very impressive work to use that to advance Ai itself; I think it can be used to study DNA itself, though, through the machine's interaction with it.