I agree with everything except for the jab at (paid) open access, which looks quite gratuitous to me. It may be true that "authors are willing to pay more to get their articles published in more prestigious journals. So, the more exciting the findings a journal publishes, the more references, the higher the impact the journal, the more submissions they get, the more money they make". But this is also true of non-open-access journals. Journals live off their prestige, and before paid open access was a thing, publishers still wanted to have high-impact journals so that university libraries would subscribe to them. It was a very distorted market (as it is now), but a market nevertheless.

While I often bash for-profit journals for being parasites that do little actual work and profit from withholding access to science that should be public, and for this I would open a bottle of champagne if they disappeared, I don't think journals have much to do with this particular problem of incentivization of bad science. Journals just respond to the demand of publishing more, and shallower, papers. That demand comes from hypercompetitivity in academia, where researchers need to fight for scarce positions and scraps of funding, often paired with too much bureaucratization (selection processes that look at "objective" and "verifiable" metrics like number of papers published at a given impact factor quartile, etc., instead of just asking a bunch of neutral experts whether the person is doing good research, which may be more opaque but also much more meaningful).

As evidence that journals are not the problem in this particular case, in fields like machine learning, where publication happens mostly in arXiv and conferences that don't charge for publishing or reading papers, the problems pointed out in the post also exist. Published models that only beat previous ones because they were lucky with random seeds or data splits are widespread.