There is no need whatsoever for the ML field to go back. I also feel the cat is out of the bag for ML, and there is no turning back into a system that Nature promotes even if they are trying to.
When I started in NLP as an undergrad, almost all the papers I cared about were free to read from the ACL Anthology. As I shifted into broader machine learning, arXiv became my new home. Nature's closed Machine Intelligence would be a step back. I am glad that researchers and scientists want to keep research open and accessible.
P.S. It's already weird that Nature has DeepMind papers. At least in that case DeepMind have a formalized agreement / enough clout that they can host the paper themselves but even the hint of ambiguity can kill knowledge propagation at many institutions. An example would be "Mastering the Game of Go without Human Knowledge" / "Mastering the game of Go with deep neural networks and tree search".
Nature (the multidisciplinary journal) is a special case.
If you get article published in Nature you are not killing the knowledge propagation, you are enhancing it to the scientific community at large. It has maintained so much prestige that getting published there is type of signaling that still has value.
The old way of scientific publishing is going to shrink to almost nothing, but there is room for a handful of high-impact, high prestige publications like Nature to keep going. They would work like compendiums of best research in science.
The main thing is that they typically don't allow you to publish it outside of your website. I also don't know what the terms of this new journal are going to be.
"Authors wishing to publish open access retrospectively should contact the individual journal editorial office, and will need to sign a new open access licence to publish form (LTP) and pay the journal's article processing charge."
However, they don't disclose how much money it costs to convert one's work to open access, and whether this option is available for all Nature journals, the main Nature journal, or just the specialized subject journals, such as Nature Machine Intelligence.
Similarly, when I researched machine learning topics for my master thesis (around 2001), citeseer.org proved to be an absolute treasure trove. I did not check out anything beyond machine learning and general computer science topics, but for these at least open access papers have quite a bit of tradition by now.
(I hope Dietterich verifies the authenticity of these singatures. Someone could just add these as a prank.)